Praise for the Hardcover Edition of Simon's

Probability Distributions Involving Gaussian Random Variables

In my research in wireless communications, I have found that I can get all

the mathematical forms and material that I need from one source, whereas
before I had to consult five or six references to get the complete informa-
tion. There are a small number of reference works that have proven so in-
valuable that I have purchased a home copy in addition to my office copy.
This handbook is one of them.
--Dr. Norman C. Beaulieu, University of Alberta

Our research Group at the University of Canterbury is engaged in both

theoretical and experimental work on advanced wireless communications
systems. We have found that the book by Simon, Probability Distributions
Involving Gaussian Random Variables is an indispensable and often used
reference for this work, so much so that we have purchased copies for the
bookshelf in our laboratory. We do significant amounts of analysis involv-
ing Gaussian statistics and quadratic forms. The book is the one source that
brings almost all of the necessary background material together in a readily
usable form and is thus an extremely valuable reference.
-- Desmond P. Taylor, University of Canterbury

The Gaussian distribution and those derived from it are at the very core of
a huge number of problems in multi-disciplinary fields of engineering,
mathematics and science. All scientists and professional engineers face,
sooner or later, a problem whose solution involves processing, either ana-
lytically, or by simulation, or by a mixture of the two, random variables
whose distribution is related to Gaussian variables. The book, with its
comprehensive information in analytical, tabular, and graphical form, is an
invaluable tool for scientists and engineers.
-- Sergio Benedetto, Politecnico di Torino
 Praise for the Hardcover Edition of Simon's
Probability Distributions Involving Gaussian Random Variables

The fact that the book collected all these results (after double-checking
them) with unified notations and in an organized fashion made it very use-
ful for many of the graduate students and researchers doing analytical
types of investigations. Not only did I use it, I also recommend it to my
graduate students and many of our papers benefited from the formulas
available in the handbook. The handbook saves valuable time for all the
researchers that rely on it to simplify their results or write them in terms of
well-known functions that are available in popular mathematical packages
such as Matlab and Mathematica.
-- Mohamed-Slim Alouini, Texas A&M University at Qatar

The book Probability Distributions Involving Gaussian Random Variables

is a handy research reference in areas such as communication systems. I
have found the book useful for my own work, since it presents probability
distributions that are difficult to find elsewhere and that have non-obvious
derivations.
-- Gordon Stiiber, Georgia Institute of Technology

Dr. Simon's book is by far the most comprehensive compilation of Gaus-

sian probability distributions and their associated statistics - in fact to my
knowledge this is the only book of its type in publication now. This hand-
book is a much needed and most welcome entry to the toolbox of any en-
gineer or scientist. For those working in digital and wireless communica-
tions, signal and image processing, pattern recognition, and many other
technical fields, the text is absolutely essential.
-- Payman Arabshahi, University of Washington

The handbook Probability Distributions Involving Gaussian Random Vari-

ables by Marvin K. Simon is a wonderful resource for both theoreticians
and practitioners in that it contains a wealth of information summarized in
one book.
-- Costas N. Georghiades, Texas A&M University
 Praise for the Hardcover Edition of Simon's
Probability Distributions Involving Gaussian Random Variables

This is a unique book and an invaluable reference for engineers and scien-
tists in the fields of electrical engineering, mathematics and statistics.
There is no other single reference book that covers Gaussian and Gaussian-
related distributions of random variables that are encountered in research
work and practice. This is a reference book that I recommend to all my
graduate students working in the area of telecommunication system analy-
sis and design.
-- John Proakis, Professor Emeritus, Northeastern University
and Adjunct Professor, University of California, San Diego

As a researcher and teacher in the digital communication theory area for

the last forty years, I can state that Simon's book contains materials in his
ten chapters and appendices most relevant to researchers and developers of
modern communication systems. These sophisticated Gaussian related ex-
pressions are needed in the analysis, design, and optimization of modern
microwave and cellular communications. Many of these results are un-
available elsewhere, since these results were originated by Simon himself.
-- Kung Yao, UCLA

After I spent a few minutes trying to come up with my own derivation, I

decided to consult Probability Distributions Involving Gaussian Random
Variables, where I found the needed result within a few seconds. On this
and many other occasions, I have found Marvin Simon's book to be a
unique and extremely valuable resource for my own research. The book is
even more crucial for the graduate students who, unlike those in my gen-
eration, did not grow up with the Gaussian Q function, Marcum's Q-
function, Rayleigh and Rician distributions, chi-squared random variables,
and all the other important functions, distributions, and concepts that have
their roots in Gaussian random variables.
-- Michael B. Pursley, Clemson University
 Praise for the Hardcover Edition of Simon's
Probability Distributions Involving Gaussian Random Variables

Because the source of randomness and noise is usually modeled as Gaus-

sian, engineers and scientists often need the probability distributions of
functions of Gaussian random variables. Marvin Simon has written an ex-
tremely useful handbook for everyone working in these fields. It is by far
the most complete and easy-to-use collection of these characterizations of
functions of Gaussian random variables.

Marvin Simon has created an important and very popular book that is use-
ful to a very large community of engineers and scientists. This work is
also very carefully done with many corrections to earlier works that were
not easily available to this community.
-- Jim Omura, Gordon and Betty Moore Foundation

Probability Distributions Involving Gaussian Random Variables by

Marvin Simon is a unique and invaluable reference for researchers in
communications, signal processing, math, statistics, and many other scien-
tific fields. I have used this book in my own research and also recom-
mended it to colleagues in universities and in industry, as well as to gradu-
ate students.
-- Andrea Goldsmith, Stanford University

The reference book Probability Distributions Involving Gaussian Random

Variables, authored by Dr. Marvin Simon, has become, in a very short
time frame, one of the most useful aids to research in the field of digital
communications that has come out in many years. It has numerous results
that can save researchers in the field endless hours of work. It has replaced
various other well known resources because of its concentration of rele-
vant, timely, and easily accessible results.
-- Larry Milstein, UCSD
PROBABILITY DISTRIBUTIONS
INVOLVING'GAUSSIAN
RANDOM VARIABLES
A Handbook for Engineers and Scientists
THE SPRINGER INTERNATIONAL
SERIES IN ENGINEERING AND
COMPUTER SCIENCE
PROBABILITY DISTRIBUTIONS
INVOLVING GAUSSIAN
RANDOM VARIABLES
A Handbook for Engineers and Scientists

Marvin K. Simon
Principal Scientist
Jet Propulsion Laboratory
Pasadena, Califomia, U.S.A.

- Springer
Library of Congress Control Number: 2006933933

Simon, Marvin K.
Probability Distributions Involving Gaussian Random Variables
A Handbook for Engineers and Scientists/ by Marvin K. Simon

p. cm.

ISBN: 978-1-4020-7058-7(HC) ISBN: 978-0-387-34657-1 (SC) E- ISBN: 978-0-387-47694-0

Printed on acid-free paper

First Softcover Printing, 2006

Copyright O 2002 by Springer Science +Business Media, LLC
All rights reserved. This work may not be translated or copied in whole or in part without the written
permission of the publisher (Springer Science +Business Media, LLC, 233 Spring Street, New York, NY
10013, USA), except for brief excerpts in connection with reviews or scholarly analysis. Use in
connection with any form of information storage and retrieval, electronic adaptation, computer software,
or by similar or dissimilar methodology now known or hereafter developed is forbidden.
The use in this publication of trade names, trademarks, service marks and similar terms, even if they are
not identified as such, is not to be taken as an expression of opinion as to whether or not they are subject
to proprietary rights.
This book is dedicated to my wife Anita, daughter Brette, and son
Jeffrey who, through their devotion and dedication, allowed me
to continuously burn the midnight oil. A tedious project such
as this could have never come tofruition without the support,
understanding, and encouragement they provided me
all during the preparation of the manuscript.
 CONTENTS

A Brief Biography of Carl Friedrich Gauss xiii

Preface xxi

Acknowledgment xxiii

INTRODUCTION

1. BASIC DEFINITIONS AND NOTATION

2. FUNDAMENTAL ONE-DIMENSIONAL VARIABLES

A. Gaussian

B. Rayleigh

C. Rician

D. Central Chi-square

E. Noncentral Chi-square

F. Log-Normal

3. FUNDAMENTAL MULTIDIMENSIONAL VARIABLES

A. Gaussian
B. Rayleigh

C. Rician

D. Central Chi-Square

E. Noncentral Chi-Square

F. Log-Normal

4. DIFFERENCE OF CHI-SQUARE RANDOM VARIABLES

A. Independent Central Chi-Square (-) Central Chi-Square

B. Dependent Central Chi-Square (-) Central Chi-Square

C. Independent Noncentral Chi-Square (-) Central Chi-Square

D. Independent Central Chi-Square (-) Noncentral Chi-Square

E. Independent Noncentral Chi-Square (-) Noncentral Chi-Square

5. SUM OF CHI-SQUARE RANDOM VARIABLES

A. Independent Central Chi-Square (+) Central Chi-Square

B. Dependent Central Chi-Square (+) Central Chi-Square

C. Independent Noncentral Chi-Square (+) Central Chi-Square

D. Independent Noncentral Chi-Square (+) Noncentral Chi-Square

6. PRODUCTS OF RANDOM VARIABLES

A. Independent Gaussian (x) Gaussian (Both Have Zero Mean)

B. Dependent Gaussian (x) Gaussian (Both Have Zero Mean)

C. Independent Gaussian (x) Gaussian (One Has Zero Mean, Both Have
Identical Variance)

D. Independent Gaussian (x) Gaussian (Both Have Nonzero Mean and

Identical Variance)

viii
E. Independent Rayleigh (x) Rayleigh

F. Dependent Rayleigh (x) Rayleigh

G. Independent Rice (x) Rayleigh

H. Independent Rice (x) Rice

I. Dependent Rayleigh Products

7. RATIOS OF RANDOM VARIABLES

A. Independent Gaussian (+) Gaussian (Both Have Zero Mean)

B. Independent Gaussian (+) Gaussian (One Has Zero Mean)

C. Independent Gaussian (+) Gaussian (Both Have Nonzero Mean)

D. Dependent Gaussian (+) Gaussian (Both Have Zero Mean)

E. Dependent Gaussian (t)Gaussian (One Has Zero Mean)

F. Dependent Gaussian (+) Gaussian (Both Have Nonzero Mean)

G. Independent Gaussian (Zero Mean) (t)Rayleigh

H. Independent Gaussian (Zero Mean) (+) Rice

I. Independent Rayleigh (+) Rayleigh

J. Dependent Rayleigh (+) Rayleigh

K. Independent Rice (+) Rayleigh

L. Independent Rice (+) Rice

M. Dependent Rayleigh Ratios

8. MAXIMUM AND MINIMUM OF PAIRS OF RANDOM

VARIABLES
A. Independent Gaussian

B. Dependent Gaussian
C. Independent Rayleigh

D. Dependent Rayleigh

E. Independent Log-Normal

F. Dependent Log-Normal

9. QUADRATIC FORMS 89
A. Both Vectors Have Zero Mean 91

B. One or Both Vectors Have Nonzero Mean 92

C. A Reduced Quadratic Form Where the Vectors Have Different Numbers

of Dimensions 94

D. General Hermetian Quadratic Forms 96

10. OTHER MISCELLANEOUS FORMS

A. Independent Rayleigh (+) Rayleigh

B. Independent Gaussian (x) Rayleigh

C. Independent Gaussian (x) Rayleigh (+) Gaussian

D. Independent Gaussian (+) Rayleigh

E. General Products of Ratios of Independent Gaussians

APPENDIX A: ALTERNATIVE FORMS 103

1. ONE-DIMENSIONAL DISTRIBUTIONS AND

FUNCTIONS 103
A. The Gaussian Q-Function 103

B. The Marcum Q-Function 104

C. The Nuttall Q-Function 111

D. The Complementary Incomplete Gamma Function 113

2. TWO-DIMENSIONAL DISTRIBUTIONS AND
FUNCTIONS
A. The Gaussian Q-Function

APPENDIX B: lNTEGRALS INVOLVING Q-FUNCTIONS

1. THE GAUSSIAN Q-FUNCTION

A. Q-Function and x

B. Q-Function with Exponentials and x

C. Q-Function with Exponentials, Bessel Functions and x

2. THE FIRST-ORDER MARCUM Q-FUNCTION

A. Q-Function with One Linear Argument

B. Q-Function with One Linear Argument and Exponentials

C. Q-Function with One Linear Argument and x

D. Q-Function with One Linear Argument, Exponentials and Powers of x 121

E. Q-Function with One Linear Argument, Bessel Functions, Exponentials

and Powers of x 123

F. Product of Two Q-Functions with One Linear Argument 125

G. Q-Function with Two Linear Arguments and x 126

H. Q-Function with Two Linear Arguments, Exponentials and x 126

3. THE GENERALIZED (MTH-ORDER) MARCUM

Q-FUNCTION 127
A. Q-Function with One Linear Argument and Powers of x 127

B. Q-Function with One Linear Argument, Exponentials and Powers of x 127

C. Q-Function with One Linear Argument, Bessel Functions, Exponentials

and Powers of x 128
D. Q-Function with Two Linear Arguments, Exponentials and Powers of x 129

APPENDIX C: BOUNDS ON THE GAUSSIAN Q-FUNCTION

AND THE MARCUM Q-FUNCTION 131

1. THE GAUSSIAN Q-FUNCTION 131

2. THE MARCUM Q-FUNCTION 136

REFERENCES 139

ILLUSTRATIONS 143
 Johann Carl Friedrich Gauss
1777-1855

At the age of seven, Carl Friedrich Gauss started elementary school,

and his potential was noticed almost immediately. His teacher,
Buttner, and his assistant, Martin Bartels, were amazed when Gauss
summed the integers from 1 to 100 instantly by spotting that the sum
was 50 pairs of numbers each pair summing to 101.

In 1788 Gauss began his education at the Gymnasium with the help of
Buttner and Bartels, where he learnt High German and Latin. After
receiving a stipend from the Duke of Brunswick- Wolfenbuttel, Gauss
entered Brunswick Collegium Carolinum in 1792. At the academy
Gauss independently discovered Bode's law, the binomial theorem
and the arithmetic- geometric mean, as well as the law of quadratic
reciprocity and the prime number theorem.

In 1795 Gauss left Brunswick to study at Gottingen University.

Gauss's teacher there was Kaestner, whom Gauss often ridiculed. His
only known friend amongst the students was Farkas Bolyai. They met
in 1799 and corresponded with each other for many years.

Gauss left Gottingen in - the construction of1798 without a diploma,

but by this time he had made one of his most important discoveries a
regular 17-gon by ruler and compasses This was the most major
advance in this field since the time of Greek mathematics and was
published as Section VII of Gauss's famous work, Disquisitiones
Arithmeticae.
Gauss returned to Brunswick where he received a degree in 1799.
After the Duke of Brunswick had agreed to continue Gauss's stipend,
he requested that Gauss submit a doctoral dissertation to the
University of Helmstedt. He already knew Pfaff, who was chosen to
be his advisor. Gauss's dissertation was a discussion of the
fundamental theorem of algebra.

With his stipend to support him, Gauss did not need to find a job so
he devoted himself to research. He published the book Disquisitiones
Arithmeticae in the summer of 1801. There were seven sections, all but
the last section, referred to above, being devoted to number theory.

In June 1801, Zach, an astronomer whom Gauss had come to know

two or three years previously, published the orbital positions of
Ceres, a new "small planet" which was discovered by G Piazzi, an
Italian astronomer on 1 January, 1801. Unfortunately, Piazzi had only
been able to observe 9 degrees of its orbit before it disappeared
behind the Sun. Zach published several predictions of its position,
including one by Gauss which differed greatly from the others. When
Ceres was rediscovered by Zach on 7 December 1801 it was almost
exactly where Gauss had predicted. Although he did not disclose his
methods at the time, Gauss had used his least squares approximation
method.

In June 1802 Gauss visited Olbers who had discovered Pallas in

March of that year and Gauss investigated its orbit. Olbers requested
that Gauss be made director of the proposed new observatory in
Gottingen, but no action was taken. Gauss began corresponding with
Bessel, whom he did not meet until 1825, and with Sophie Germain.

Gauss married Johanna Ostoff on 9 October, 1805. Despite having a

happy personal life for the first time, his benefactor, the Duke of
Brunswick, was killed fighting for the Prussian army. In 1807 Gauss
left Brunswick to take up the position of director of the Gottingen
observatory.

Gauss arrived in Gottingen in late 1807. In 1808 his father died, and a
year later Gauss's wife Johanna died after giving birth to their second
son, who was to die soon after her. Gauss was shattered and wrote to
Olbers asking him give him a home for a few weeks,

to gather new strength in the arms of your friendship - strength for a life
which is only valuable because it belongs to my three small children.

xiv
Gauss was married for a second time the next year, to Minna the best
friend of Johanna, and although they had three children, this marriage
seemed to be one of convenience for Gauss.

Gauss's work never seemed to suffer from his personal tragedy. He

published his second book, Theoria motus corporum coelestium in
sectionibus conicis Solem arnbientium, in 1809, a major two-volume
treatise on the motion of celestial bodies. In the first volume he
discussed differential equations, conic sections and elliptic orbits,
while in the second volume, the main part of the work, he showed
how to estimate and then to refine the estimation of a planet's orbit.
Gauss's contributions to theoretical astronomy stopped after 1817,
although he went on making observations until the age of 70.

Much of Gauss's time was spent on a new observatory, completed in

1816, but he still found the time to work on other subjects. His
publications during this time include Disquisitiones generales circa
seriem infinitam, a rigorous treatment of series and an introduction of
the hypergeometric function, Methodus nova integralium valores per
approximationem inveniendi, a practical essay on approximate
integration, Bestimmung der Genauigkeit der Beobachtungen, a discussion
of statistical estimators, and Theoria attractionis corporum
sphaeroidicorum ellipticorum hornogeneorum methodus nova tractata. The
latter work was inspired by geodesic problems and was principally
concerned with potential theory. In fact, Gauss found himself more
and more interested in geodesy in the 1820's.

Gauss had been asked in 1818 to carry out a geodesic survey of the
state of Hanover to link up with the existing Danish grid. Gauss was
pleased to accept and took personal charge of the survey, making
measurements during the day and reducing them at night, using his
extraordinary mental capacity for calculations. He regularly wrote to
Schumacher, Olbers and Bessel, reporting on his progress and
discussing problems.

Because of the survey, Gauss invented the heliotrope which worked

by reflecting the Sun's rays using a design of mirrors and a small
telescope. However, inaccurate base lines were used for the survey
and an unsatisfactory network of triangles. Gauss often wondered if
he would have been better advised to have pursued some other
occupation but he published over 70 papers between 1820 and 1830.

In 1822 Gauss won the Copenhagen University Prize with Theoria

attractionis ... together with the idea of mapping one surface onto
another so that the two are similar in their smallest parts. This paper
was published in 1825 and led to the much later publication of
Untersuchungen iiber Gegenstande der Hoheren Geodasie (1843 and 1846).
The paper Theoria combinationis observationum erroribus minimis
obnoxiae (1823), with its supplement (1828), was devoted to
mathematical statistics, in particular to the least squares method.

From the early 1800's Gauss had an interest in the question of the
possible existence of a non-Euclidean geometry. He discussed this
topic at length with Farkas Bolyai and in his correspondence with
Gerling and Schumacher. In a book review in 1816 he discussed
proofs which deduced the axiom of parallels from the other Euclidean
axioms, suggesting that he believed in the existence of non-Euclidean
geometry, although he was rather vague. Gauss confided in
Schumacher, telling him that he believed his reputation would suffer
if he admitted in public that he believed in the existence of such a
geometry.

In 1831 Farkas Bolyai sent to Gauss his son JQnosBolyai's work on the
subject. Gauss replied

to praise it would mean to praise myself.

Again, a decade later, when he was informed of Lobachevsky's work

on the subject, he praised its "genuinely geometric" character, while in
a letter to Schumacher in 1846, states that he

had the same convictions for 54 years

indicating that he had known of the existence of a non-Euclidean
geometry since he was 15 years of age (this seems unlikely).

Gauss had a major interest in differential geometry, and published

many papers on the subject. Disquisitiones generales circa superficies
curva (1828) was his most renowned work in this field. In fact, this
paper rose from his geodesic interests, but it contained such
geometrical ideas as Gaussian curvature. The paper also includes
Gauss's famous theorema egregrium:

I f an area in E3 can be developed (i.e. mapped isometrically) into another area

of E3, the values of the Gaussian curvatures are identical in corresponding
points.

xvi
The period 1817-1832 was a particularly distressing time for Gauss.
He took in his sick mother in 1817, who stayed until her death in 1839,
while he was arguing with his wife and her family about whether
they should go to Berlin. He had been offered a position at Berlin
University and Minna and her family were keen to move there.
Gauss, however, never liked change and decided to stay in Gottingen.
In 1831 Gauss's second wife died after a long illness.

In 1831, Wilhelm Weber arrived in Gottingen as physics professor

filling Tobias Mayer's chair. Gauss had known Weber since 1828 and
supported his appointment. Gauss had worked on physics before
1831, publishing Uber ein neues allgemeines Grundgesetz der Mechanik,
which contained the principle of least constraint, and Principia
generalia theoriae figurae fluidorum in statu aequilibrii which discussed
forces of attraction. These papers were based on Gauss's potential
theory, which proved of great importance in his work on physics. He
later came to believe his potential theory and his method of least
squares provided vital links between science and nature.

In 1832, Gauss and Weber began investigating the theory of terrestrial

magnetism after Alexander von Humboldt attempted to obtain
Gauss's assistance in making a grid of magnetic observation points
around the Earth. Gauss was excited by this prospect and by 1840 he
had written three important papers on the subject: Intensitas vis
magneticae terrestris ad mensuram absolutam revocata (1832), Allgemeine
Theorie des Erdmagnetismus (1839) and Allgemeine Lehrsatze in Beziehung
auf die im verkehrten Verhaltnisse des Quadrats der Entfernung wirkenden
Anziehungs- und Abstossungskrafte (1840). These papers all dealt with
the current theories on terrestrial magnetism, including Poisson's
ideas, absolute measure for magnetic force and an empirical
definition of terrestrial magnetism. Dirichlet's principle was
mentioned without proof.

Allgemeine Theorie... showed that there can only be two poles in the
globe and went on to prove an important theorem, which concerned
the determination of the intensity of the horizontal component of the
magnetic force along with the angle of inclination. Gauss used the
Laplace equation to aid him with his calculations, and ended up
specifying a location for the magnetic South pole.

Humboldt had devised a calendar for observations of magnetic

declination. However, once Gauss's new magnetic observatory
(completed in 1833 - free of all magnetic metals) had been built, he
proceeded to alter many of Humboldt's procedures, not pleasing

xvi i
Humboldt greatly. However, Gauss's changes obtained more accurate
results with less effort.

Gauss and Weber achieved much in their six years together. They
discovered Kirchhoff's laws, as well as building a primitive telegraph
device which could send messages over a distance of 5000 ft.
However, this was just an enjoyable pastime for Gauss. He was more
interested in the task of establishing a world-wide net of magnetic
observation points. This occupation produced many concrete results.
The Magnetischer Verein and its journal were founded, and the atlas of
geomagnetism was published, while Gauss and Weber's own journal
in which their results were published ran from 1836 to 1841.

In 1837, Weber was forced to leave Gottingen when he became

involved in a political dispute and, from this time, Gauss's activity
gradually decreased. He still produced letters in response to fellow
scientists' discoveries usually remarking that he had known the
methods for years but had never felt the need to publish. Sometimes
he seemed extremely pleased with advances made by other
mathematicians, particularly that of Eisenstein and of Lobachevsky.

Gauss spent the years from 1845 to 1851 updating the Gottingen
University widow's fund. This work gave him practical experience in
financial matters, and he went on to make his fortune through shrewd
investments in bonds issued by private companies.

Two of Gauss's last doctoral students were Moritz Cantor and

Dedekind. Dedekind wrote a fine description of his supervisor
... usually he sat in a comfortable attitude, looking down, slightly stooped,
with hands folded above his lap. He spoke quite freely, very clearly, simply
and plainly: but when he wanted to emphasise a new viewpoint ... then he
lifted his head, turned to one of those sitting next to him, and gazed at him
with his beautiful, penetrating blue eyes during the emphatic speech. ... l f h e
proceeded from an explanation of principles to the development of
mathematical formulas, then he got up, and in a stately very upright posture
he wrote on a blackboard beside him in his peculiarly beautiful handwriting:
he always succeeded through economy and deliberate arrangement in making
do with a rather small space. For numerical examples, on whose careful
completion he placed special value, he brought along the requisite data on
little slips of paper.

Gauss presented his golden jubilee lecture in 1849, fifty years after his
diploma had been granted by Hemstedt University. It was

xviii
appropriately a variation on his dissertation of 1799. From the
mathematical community only Jacobi and Dirichlet were present, but
Gauss received many messages and honours.

From 1850 onwards Gauss's work was again of nearly all of a practical
nature although he did approve Riemann's doctoral thesis and heard
his probationary lecture. His last known scientific exchange was with
Gerling. He discussed a modified Foucalt pendulum in 1854. He was
also able to attend the opening of the new railway link between
Hanover and Gottingen, but this proved to be his last outing. His
health deteriorated slowly, and Gauss died in his sleep early in the
morning of 23 February, 1855.

Article by: J J O'Connor and E F Robertson

http://www-groups.dcs.st-and.ac.uk/-history/Mathematicians/Gauss.html

xix
 PREFACE

This book is intended for use by students, academicians and

practicing engineers who in the course of their daily study or research
have need for the probability distributions and associated statistics of
random variables that are themselves Gaussian or in various forms
derived from them. The format of the book is primarily that of a
handbook in that, for the most part, the results are merely presented
in their final form without derivation or discussion. As such the
reader must rely on the typographical accuracy of the documented
expressions, which the author has taken great pains to assure. Also
included at the end of the book are numerous curves illustrating the
behavior of a variety of the probability measures presented in
mathematical form.
The author wishes to acknowledge his many colleagues in industry and
academiafor the encouragement and support they provided for this
project without which it might never have gotten started.
 INTRODUCTION

There are certain reference works that engineers and scientists alike
find invaluable in their day-to-day work activities. Many of these
reference volumes are of a generic nature such as tables of integrals,
tables of series, handbooks of mathematical formulas and transforms,
etc. (see Refs. 1, 2, 3, and 4 for example), whereas others are
collections of technical papers and textbooks that directly relate to the
individual's specific field of specialty. Continuing along this train of
thought, there exists a great deal of valuable information that, in its
original form, was published in university and company reports and
as such the general public was in many cases not aware of its
existence. Even worse, today this archival material is no longer
available to the public in any form since its original source has
declared it to be out of print for quite some time now. Furthermore,
most of the authors of these works have long since retired or sadder
yet have passed on; however, the material contained in the
documents themselves has intrinsic value and is timeless in terms of
its value to today's practicing engineer or scientist. As time marches
on and new young engineers and scientists replace the old ones, the
passing of the torch must include a means by which this valuable
information be communicated to the new generation. Such is the
primary motivation behind this book, the more detailed objective
being described as follows.
One of the most important, from both the theoretical and
practical viewpoint, probability distributions is the Gaussian
distribution or as mathematicians prefer to call it the normal
distribution. Although the statistical characterization of the basic
Gaussian random variable (RV), e.g., its probability density function
(PDF), cumulative distribution function (CDF), and characteristic
function (CF) are well-known and widely documented in the
literature (e.g., [ 5 ] ) ,in dealing with the applications, one is quite often
in need of similar characterizations for arithmetic combinations, e.g.,
sums, differences, products, ratios of Gaussian RVs and also the
square of Gaussian RVs (so-called chi-square RVs.) Other applications
involve log-normal RVs and thus their statistical characterization is
2 PROBABILITY DISTRIBUTIONS INVOLVING GAUSSIAN RANDOM VARIABLES

also of interest. Still other applications involve the maximum or

minimum of two or more RVs. Although many published references
on the above subjects implicitly make use of such PDFs, CDFs, and
CFs, a publicly available compilation of these probability measures
for a wide variety of cases is difficult to find. Perhaps the most
comprehensive source for a large number of such probability
measures is a Stanford University report by J. Omura and T. Kailath
[6] that dates back more than three and a half decades. To those who
are aware of its existence, this report has become somewhat of a
classic reference on the subject and considering the fact that it was
prepared prior to the age of personal computers and desktop
publishing, it should be looked upon as a remarkable feat. Because
of this nature of preparation, however, its absolute reliability suffers
from the fact that it contains crucial typographical as well as other
miscellaneous errors. These errors were discovered by this author
when trying to apply the results in Omura and Kailath [6] to his own
research. Furthermore, although the report includes a list of
references from which the results contained therein were either
obtained or derived, many of these references are no longer available
(e.g., unpublished reports from companies which, in some instances,
no longer exist, books that are out of print, etc.). Thus, the ability to
fix the incorrect results is indeed hampered by the lack of availability
of the references.
With the above in mind, the author set out to find the correct
results and document them here for the benefit of: (a) those who have
Ref. 6 and have discovered the existence of the errors but didn't care
to take the time to correct them, (b) those who have Ref. 6 and have
not yet found or are unaware of the existence of the errors, or (c) those
who have never seen the report in the first place and would benefit
from having these results. In addition, for the results that are indeed
correct in Ref. 6, we have, in many instances, expressed them in a
form which has a more pleasing appearance and lends itself better to
the applications at hand. Furthermore, although the book does not
focus on any particular application, from time to time specific
applications and the location of their discussion in the literature will
be mentioned as evidence of the usefulness of the results contained
herein.
From the standpoint of the results themselves, the goal is to
express them, as far as possible, in terms of well-known (and
tabulated) functions. (In some instances, this requires an expression
in the form of a single (occasionally double) infinite series.) When
this goal is not achievable, as is more typical for CDFs than for PDFs,
a set of dashes will be used to indicate the lack of availability of such
results. While in keeping with the style of Ref. 6 this book has the
 INTRODUCTION 3
flavor of a reference handbook for researchers, it also serves as a
valuable companion to a college textbook used to teach a course on
probability and random processes.
 BASIC DEFINITIONS AND NOTATION

Throughout the book, a variable with a boldface type will be used to

denote a matrix or vector, the latter always assumed to be in column
format. Random variables are denoted by upper case letters whereas
the values that these variables assume are denoted by lower case
letters. Finally, p(.) is used to denote a PDF, P(.) is used to denote a
CDF, and Y(.) is used to denote a CF.
A real Gaussian random variable (RV) is defined as one having
the probability density function (PDF)

where XA E { X } is the statistical mean of X and o; is the variance of

X. The corresponding CDF, Px(x),and CF, Yx(a), are given by

where

is the Gaussian probability integral which is related to the

complementary error function by

: (3
Q(x)= -erfc -

and
6 PROBABILITY DISTRIBUTIONS INVOLVING GAUSSIAN RANDOM VARIABLES

with j = n. A complex Gaussian RV 2 = X, + jXI is one in which

X, and XI each have a PDF of the form in (1.1). The PDF of 2 is
represented by the joint PDF of X, and X,.
For the purpose of this book, a Gaussian vector X is a vector
whose components are independent Gaussian RVs with equal
variance oi but in general different statistical means. That is, for

we have

The class of Gaussian vectors described above shall be denoted by

Nl1(ft,0 2 ) .Thus, X E NIl(ff, 02)means that X is a vector having the
properties in (1.7).
Only Gaussian vectors of the same dimension are allowed to
be statistically independent. In particular, if x") E N,,(ff'l),o:) and
x'~)
E N, (XI2),
0:) are dependent Gaussian vectors, then for
 BASIC DEFINITIONS AND NOTATION 7
we have

That is, only components of x(')and x ( ~ ) identical subscripts

having
can be correlated. This dependence is summarized by the covariance
matrix for each pair of components

whose inverse is given by

Clearly if x(')and x ( ~ ) independent, then p = 0.

are
A Rayleigh RV, R, of order n is defined by the norm of a zero
mean Gaussian vector. That is, if X E N,(o, 02),then

A Rician RV, R, of order n is defined analogously by the norm of a

nonzero mean Gaussian vector, X E N, (R,0 2).I )
A chi-square RV is defined as the squared norm of a Gaussian
vector. If the Gaussian vector is zero mean, then

----

I The more common usage of the terms Rayleigh RV and Rician RV

refers to the specific case n=2. Here, for convenience, we use them in
their extended form. We also note that the Rayleigh distribution of
order m is equivalent to the Nakamagi-m distribution [7] when m is
integer. Likewise the Rician distribution of order 2 is equivalent to
the Nakagami-n distribution [7].
8 PROBABILITY DISTRIBUTIONS INVOLVING GAUSSIAN RANDOM VARIABLES

is referred to as a central chi-square RV with n degrees of freedom. For

n = 1, the central chi-square distribution simplifies to an exponential
distribution. Furthermore, for n # 1, the central chi-square distribution
is equivalent to the gamma distribution with parameter n when n is
integer. If the Gaussian vector is nonzero mean, then Y of (1.13) is
referred to as a noncenfral chi-square RV with n degrees of freedom and
noncenfralify parameter

Whereas the book will only provide results for the noncentral
moments of a RV, e.g., E { x k ] , k integer, the central moments, most
notably the variance, can be obtained from the relation

E{(X -~ ) =l }$(;)(-I)'-' F"'E{X'], k integer

i=O

where (r) k!
= -denotes a combinatorial coefficient.
i!( k - i)!
A log-normal RV is one whose logarithm has a Gaussian
distribution. That is y =lOX'10 is a log-normal RV when X is
characterized by the PDF of (1.1).
 FUNDAMENTAL ONE-DIMENSIONAL
VARIABLES

A. Gaussian

The PDF, CDF, and CF of a Gaussian RV X E N , , ( X , D ~are) given in

(1.1), (1.2), and (1.5) respectively with ox replaced by a . For X = 0,
the even moments of the components of X are given by

All odd moments are equal to zero.

6.Rayleigh

Since the square root always yields a positive quantity, then by

definition R = @ = 1x1, i.e., a single-sided Gaussian RV with PDF
and CDF given by

Also, the moments of R are given by

10 PROBABILITY DISTRIBUTIONS INVOLVING GAUSSIAN RANDOM VARIABLES

E { R ~=
J (202r'2 (F), k integer

Here R corresponds to a conventional Rayleigh RV with PDF and

CDF

Also, the moments of R are given by

3
E { R ~=) (202)*i2T 1+ - ,k integer
(

r(rn+
E { R ~=} (202)*i2
E) , k integer
( m - l)!
 FUNDAMENTAL ONE-DIMENSIONAL VARIABLES 11

E{R'} = (20')1'~
.(rn + y)
,k integer (2.13)
T(m+l/2)

C. Rician

where , ~ ; ( a ; P ; y is
) the confluent hypergeometric function [2] and
a= 1x1.

Here R corresvonds to a conventional Rician RV with varameter

where
12 PROBABILITY DISTRIBUTIONS INVOLVING GAUSSIAN RANDOM VARIABLES

x2 + a2
Q,(a.P) = li (2.20)

is the first-order Marcum Q-function [8].2

where

is the generalized (mth-order)Marcum Q-function [8].

( s)
+(-I)" exp - - C i!(m
(m+i-I)!
nl-l

i=o -i-I)!
(02)i]7
-
2ra

More often than not in the literature, the subscript "1" identifying
the order of the first-order Marcum Q-function is dropped from the
notation. We shall maintain its identity in this text to avoid possible
ambiguity with the two-dimensional Gaussian Q-function defined in
Eq. (A.37) of Appendix A.
E { R ~=} (202)li2
exp

k integer

D. Central Chi-square
1. n = l

Since the kth moment of a central chi-square RV with 2m degrees of

freedom is equal to the 2kth moment of a Rayleigh RV of order 2m,
then it is straight-forward to obtain the moments of Y from (2.10) as
14 PROBABILITY DISTRIBUTIONS INVOLVING GAUSSIAN RANDOM VARIABLES

E { Y ~ =} (202)*r ( m + k ) , k integer
( m - I)!

r(m+k+;)
E { Y ~ =} ( 2 0 2 ) *
i)
r(rn +
,k integer (2.39)

E. Noncentral Chi-square
 FUNDAMENTAL ONE-DIMENSIONAL VARIABLES 15

1
1 - 2jwo 1-2jwo2

Since the kth moment of a noncentral chi-square RV with 2rn degrees

of freedom is equal to the 2kth moment of a Rician RV of order 2rn,
then it is straightforward to obtain the moments of Y from (2.23) as

(
( 2)
(rntk-l)! 2:2)
E { Y ~ =} ( 2 0 2 ) lexp - ,F; rn+k;m;- , k integer (2.47)
(rn-I)!

k integer

F. Log-Normal

Let X E N ~ ( X0, 2 ) .Then the PDF of y = 1O*l0 is given by

where 5 = 10 / ln 10 and X (dB) and o2 (dB) correspond to the mean
and variance of IOlog,, y . The CF of y is not obtainable in closed
form but can be approximated by a Gauss-Hermite expansion as

where x,, are the zeros and Hx,, are the weight factors of the N,-order
Hermite polynomial and can be found in Table 25.10 of [2]. In
addition, the moments of y are given by

[F- F
~ { y *=} exp - X + - - o ,k integer
 FUNDAMENTAL MULTIDIMENSIONAL
VARIABLES

A. Gaussian

The joint PDF of a Gaussian vector X E ~ , , ( ? t , o ' ) covariance

with
matrix

is given as

where IM,~ denotes the determinant of M,. Letting

then the joint CF is given by

r

For the special case of n = 2,X = 0, and the covariance matrix of (1.10),
the joint moments are given as
18 PROBABILITY DISTRIBUTIONS INVOLVING GAUSSIAN RANDOM VARIABLES

10, k, + k2 odd

k, + k2 even

B. Rayleigh

Consider the pair of Rayleigh RVs of order n, R, = IIX("II,R2 = I x("I

defined from the underlying Gaussian vectors x'" E N~~( 0 , ~ ; and )
x E N ( 0 ) Recalling that only components of these vectors with
identical subscripts are correlated with covariance matrix as in (1.10),
then the joint PDF and CDF of R, and R2 are given by

The joint moments of R, and R2 are given by

where , F; (a,P;y ;x ) is the Gaussian hypergeometric function [2].

 FUNDAMENTAL MULTIDIMENSIONAL VARIABLES 19
1. n = 2

Using the alternative representation of the first-order Marcum Q-

function given in Eqs. (A.6) and (A.7) of Appendix A, the bivariate
Rayleigh CDF of (3.10) can be expressed in the form of a single
integral with finite limits as follows :

"
(I -p4)y2
[(i.2Y2 + 21p/lq
+ lpl(l- p 2 ) yY,(y2+ q2)sin 8
+c ) ( ~ Ido, (3.11)
+ 2'1 p l y Y o + p Z y 2 )

where I: 4 r;: / &oi and

20 PROBABILITY DISTRIBUTIONS INVOLVING GAUSSIAN RANDOM VARIABLES

C. Rician

Consider the pair of Rician RVs of order n , Rl = I x"'I , I$ = I x")I

defined from the underlying Gaussian vectors x")E ~ ~ ( f t " ' , and
~:)
(x('),
x"'E N~ 0;) where in addition IIX'l'II = IF"'II
= a . Again recalling
that only components of these vectors with identical subscripts are
correlated with covariance matrix as in (1.10), then the joint PDF of Rl
and R2 is given (for n > 2) by

For the special case of n = 2, the joint PDF is given by

where E~ is the Neumann factor, i.e., E, = 1, E~ = 2 for i > 0.

 FUNDAMENTAL MULTIDIMENSIONAL VARIABLES 21
D. Central Chi-square
Consider the pair of central chi-square RVs of order n ,
8= \IX")(~,
Y, = IP'2)I12
defined from the underlying Gaussian vectors
x"'E Nn( 0 , ~ : )and x("E N~(0.0;). Then, the joint PDF and CDF of
and Y, are given by

The joint moments of I: and Y, are given by

22 PROBABILITY DISTRIBUTIONS INVOLVING GAUSSIAN RANDOM VARIABLES

E. Noncentral Chi-square
Consider the pair of central chi-square RVs of order n l
Y; = IIx"' ~ ~~,
6= IIx")~~
defined from the underlying Gaussian vectors
x"' E N,(P, 0:) and x'~'E N~(X(l),0:). Thenl the joint PDF of Y; and
Y, is given (for n > 2) by

[
xexp - 1 [++,-
2(1- p2)
( 4+4- 2 ~ o P ~ ) ~ ~ ) ]
o:o;

For the special case of n = 2, the joint PDF is given by

 FUNDAMENTAL MULTIDIMENSIONAL VARIABLES 23

F. Log-Normal

Consider the pair of correlated log-normal RVs y, =lOX1'10 and

y, =10 Xl
,
O corresponding to the correlated Gaussian RVs
X, E N1($,o:) and X, E N1($,a:). Then the joint PDF of y, and y, is

Once again the joint CDF and also the joint CF are not available in
closed form. However, the joint moments can be determined from the
joint CDF of the corresponding Gaussian RVs. In particular,

Since from (3.4)

1 1
(al,
Yxl,x2 0.) = exp jwl$ + jw,$ - -0:~:- -w:o:
2 2
- pwlw201a2)
(3.22)
then
 DIFFERENCE OF CHI-SQUARE
RANDOM VARIABLES

A. Independent Central Chi-square (-) Central Chi-

Square
Define Y = - Y, where I: and Y, are independent central chi-square
distributed RVs with n, and n, degrees of freedom, respectively.

where K , ( x ) is the modified Bessel function of the second kind [2].3

-- - -
Note that K,(x)is defined for 0 < x I whereas I , ( x ) is defined
for I x I and is an even function of x .
 1
,exp
20,
- --
[
1
( 2 ; ) (rn -
111-1

) (2(m- 1) - i)!
i m I - i)!

)
20:
1
( 1 nl-l (2(m- 1) - i)!
i!~-
m1 - i)!

1 m-I i
(2(m- 1) - i)!
P, (y) = exp[ 2;,) -
:- ( r n - ~-)-
! [-~ ~ ? ~ ~ ~ ~ ~ ( i - ~ ) ! ( r n - ~ - i )
i-l
 [
P , ( ~ =) l - e x p -
2; l-
~-
1
-
](m-
nl-I i
( 2 ( m- 1) - i)!
~)![o~~o~~~~(i-~~!(m-~

[
1-exp -7 Y ) [- 01 ) [ L ) ~ - ' , ~ ~ ~
$ .)==+[
2 0 , 0,+ 0, =;, ,=, i - l)! a: + 0:
m-I-i

2 4
 (ml + m2 - 2 - i)!
r112-'
1
,exp
20, [)
- --
[ ,"
20; (m, - l)! a, + a,2 ) " " ~ i!(rn2-1-i)!

1 0: [)1"21~(ml+m2-2-i)!
(4.16)

[y ) (m, - I)! a: + a; i=, i!(m,- 1- i)!

q 1e x P --20:

Ill1
-1 i
,112
(m,+m2- 2 - i ) !
exp - --
1
(2:;) (rnl - 1 ) a )Z ( i - i)!(rn2- 1 - i)!

(
I-exp --2
1
) r n 2 -I ! (A) ,n2
,II,-I i (m, +m2- 2 - i ) !
i -m - I - i)!

B. Dependent Central Chi-square (-) Central Chi-

Square
 DIFFERENCE OF CHI-SQUARE RANDOM VARIABLES 29
To simplify the expressions, we introduce the parameters
112
[(of- o:)1+ 4o:of(l- p i ) ] of - 0:
y- = , a"=y-f (4.19)
o;o; (1- p 2 ) o;0;2 (1- p 2 )

Note that a+2 0 and a-2 0. Then,

C. Independent Noncentral Chi-square (-) Central
Chi-square
 DIFFERENCE OF CHI-SQUARE RANDOM VARIABLES 31
Define Y = 1.; - Y, where 1.; and Y, are independent noncentral and
central chi-square distributed RVs with n, and n, degrees of freedom,
respectively.
 1 (m,-l+i)!
i!(m2- 1 - i)! (m, - I)!

PY (Y)= -I - 2)
20,' 0, + 0,
[ (a,,+ l ) / 2

exp[+)exp[- 2(0~: O,'J

nt2-I-i
1
(-*)
~ [ ~ ) ( ~ l l l - ~ ~ ~ ~ n ~ 2 - l

i!(m2- 1- i)! 2' (*)I

where
 DIFFERENCE OF CHI-SQUARE RANDOM VARIABLES 33

is a generalization of the Marcum Q-function defined in [9, Eq. (86)]

and which has recursive properties [9, Eq. (87)]

Note that the special case Q,,,,,,~,(a,~)

is related to the generalized
(mth-order)Marcum Q-function of (2.24) by

Before concluding this section we point out that for the case
n, = n, = m, m odd, the PDF can be expressed in the form of an infinite
series in Whittaker functions [2] which themselves are expressed in
terms of the confluent hypergeometric function , 4(a;P;y ) . Because
of the absence of the functions in standard mathematical software
manipulation packages such as MathematicaG9, and the complexity of
the resulting expressions, their use is somewhat limited in practical
applications and thus the author has decided to omit these results for
this case. Nevertheless, the CF is still simple and given by

D. Independent Central Chi-square (-) Noncentral

Chi-square
Define Z = Y; - Y, where Y; and Y, are independent central and
noncentral chi-square distributed RVs with n, and n, degrees of
freedom, respectively. Then, the PDF of Z can be determined in
terms of the PDF of Y given in the previous section by
34 PROBABILITY DISTRIBUTIONS INVOLVING GAUSSIAN RANDOM VARIABLES

PZ ( 2 ) = PY ( - 2 ) a+a, (4.45)
a1-'a2
u;+u;

Similarly, the CDF of Z can be determined from the CDF of Y as

E. Independent Noncentral Chi-square (-)

Noncentral Chi-square
For this case, the results for the PDF can be expressed in the form of a
doubly infinite series in Whittaker functions. Once again, for the
reasons stated above, the author has decided to omit these results for
this case. Nevertheless, the CF for the generic case n, = m,,n, = m, is
still simple and given by
 SUM OF CHI-SQUARE RANDOM
VARIABLES

Define the RV Z2 = -Y,. Then the PDF of Z, is given by

pz2( z ) = pr, (-z), z 5 0. From the form of p y ( y ) for central chi-square
RVs, we observe that for n odd, the PDF of Z2 is given by the PDF of
Y, with y replaced by z and -0; substituted for a;. For n even, the
PDF of Z2 is given by the negative of the PDF of Y, with y replaced
by z and -0; substituted for 0;. From the form of P,(Y) for
noncentral chi-square RVs, we observe that in addition to the above
substitutions, -a; must be substituted for a;. For example, for Y, a
central chi-square RV with 2m2 degrees of freedom, the PDF of Z2 is
expressible as

that is, we use the expression for the PDF of Y, (which applies for
y 2 0) but substitute z for y, -0;for a;, and then take its negative
and apply it for z 5 0 . Similarly, for Y, a noncentral chi-square RV
with 2m2 degrees of freedom, the PDF of Z2 is expressible as
A. Independent Central Chi-square (+) Central Chi-
Square
Define now the RV Z = I: + Y, = I: - Z2. Also, for the results of Section
4A, define the notation

Then, it can be shown that the PDF of Z is given by

Note that since Z only takes on positive (or zero) values, the PDF of
Z is defined only for z 2 0 . Similarly, define the notation

Then, it can be shown from (5.4) that

+ p- ,;;( ~ ) ~ o ; + -2 u>
p ; ( ~ ) ~ ~ ;- 9 -;0,n, odd
PZk) = (5.6)
P; (~)l , ~ +-, , - P; (z)l,,+ - a ; , 2 0, n, even
 SUM OF CHI-SQUARE RANDOM VARIABLES 37
Before proceeding, the reader is cautioned that care must be
exercised in applying (5.4) and (5.6) since in some instances the
substitution 0;-+ -0; in the generic form of the PDF of Y might
result in functions with imaginary or undefined arguments. In these
instances, one is better off deriving the result for the chi-square sum
directly from a convolution of the individual chi-square RV PDFs
rather than from the result for the chi-square difference. In this same
regard, a closed-form result for the CDF might exist for the chi-square
sum RV even though it doesn't exist for the chi-square difference RV.

a,'- a:
(5.7)
 "l-' (2(m- 1) - i)!
1
p z ( z ) = q
exp --
(
III-I-i
1
--
2S:)(m-l)![o:"Jniz ilirn-l-i)!

(5.15)
+L
20; exp[- L) 2[Alrn
(2(m 1) - i)!
20; (m-I)! o2-o1
i!(m-1-i)!
-

(
7 --
1 III-Ii (2(m- I ) - i)!
P,(z) = l - e x p -2Sl)(m-l)!(,5Jiz~(i-q!(rn-l-i)!

1 - i
(2(m- I ) - i)!
(
-exp --
2);:
--
(
(rn - 1 ) i - i)!(rn- I - i)!

Note that when a; = o; = 0 2 ,then Z simply becomes a central chi-

square RV with 4m degrees of freedom with PDF, CDF, and CF
determined from (2.32), (2.33), and (2.34), respectively, with m
replaced by 2m. The moments of Z are given by
 SUM OF CHI-SQUARE RANDOM VARIABLES 39

Using (4.13) in (5.4), we immediately arrive at

which, of course, reduces to (5.11) when m = 1. The corresponding

expression for the CDF can be obtained by using (4.14) in (5.6) with
the result
Note that when 0;= 0: = a', then Z simply becomes a central chi-
square RV with 2(m + 1) degrees of freedom with PDF, CDF and CF
determined from (2.32), (2.33), and (2.34), respectively, with m
replaced by m + 1.

B. Dependent Central Chi-square (+) Central Chi-

Square

To simplify the expressions, we introduce the parameters

 SUM OF CHI-SQUARE RANDOM VARIABLES 41
Note that p+ 2 0 but P-' 10. Then,

k integer
(5.33)
k integer
(5.37)
 SUM OF CHI-SQUARE RANDOM VARIABLES 43
p ( z )= - - - - - - (5.42)

C. Independent Noncentral Chi-square (+) Central

Chi-square
Define Z = I.; + Y, where I.; and Y, are independent noncentral and
central chi-square distributed RVs with n, and n2 degrees of freedom,
respectively.
44 PROBABILITY DISTRIBUTIONS INVOLVING GAUSSIAN RANDOM VARIABLES

L(3)
p z ( 4 = 20:
2"f2
02
($1 (rn, +n~,-1)12

exp[- g) (5.48)

For 0; > o:, applying (5.4)and (5.6) to (4.35) and (4.36) gives

For the limited case of 0; = 0; = 0 2 ,one can use the series expansion
of the generalized Marcum Q-function, namely,
 1- Qk(a,p) = exp(- v)
SUM OF CHI-SQUARE RANDOM VARIABLES

$(!)I
45

in (5.51)to arrive at the results

These results could also have been immediately obtained by noting

that, for this limiting case, Z is simply a noncentral chi-square RV
with 2(m + 1) degrees of freedom and the value of a, is still obtained
from (1.14) since the addition of the central chi-square RV to does
not change this value.
For 0;< o:,the form of the PDF and CDF change with respect
to those given in (5.51) and (5.52). Still we can apply the series
expansion of the generalized Marcum Q-function to these equations
keeping in mind that now the arguments of the Marcum Q-function,
a and p, will be purely imaginary. Carrying out the algebra and
recalling that I,(-x) = (-1)' I,(x), we obtain

and

which can also be put into the form

D. Independent Noncentral Chi-square (+)
Noncentral Chi-square
Define Z = Y; + Y, where I: and Y, are independent noncentral chi-
square distributed RVs with n, and n, degrees of freedom,
respectively.

P,(z) = (R,) (
2 exp - 1
i- n ~ ( /n / 2 ++ i 20:
+ ~ ) [ ~ ( ~ ; ~ o f ) '
(5.61)
 SUM OF CHI-SQUARE RANDOM VARIABLES 47

Pz(z)=
(2)2"'2
( )=,r(rn2
a2
exp -2 +i+l)
2 4 ;=, ,=, i! l! ~ ( m+,1)
(5.64)

1
Y Z ( 4 =
(1- 2jwo:)"" (1- 2jwo:)"" 1 - 2jwo,
 PRODUCTS OF RANDOM VARIABLES~

A. Independent Gaussian (x) Gaussian (Both Have

Zero Mean)

Let x"'E N,,(0.0;) and x'~' E N,,(0.0;) be independent Gaussian

vectors. Then, the inner product of these vectors, namely,

has the following statistical properties.

1. n = l

4
A large number of the PDF and statistical moment results in this
section come from Ref. 5.
 i even
(6.8)
(0,k odd

px(x) = olo2( m- I)!

exp[-.L)e
0102i,O
+ i - [K)
i!( m- i - l)! olo2
2m+i
m-1-i

x<o
Px (4= - 1 (m+i-I)! [L) m-1-i-1

I-- ?
( m- I)! olo2 i=o i!(m- i - 1 - l)!
2m+i o,02
B. Dependent Gaussian (x) Gaussian (Both Have
Zero Mean)

Let x"'E N,,(O,O:) and x'~' E~ ~ 0 ~ ; () be

0 , dependent Gaussian
vectors. Then, the inner product of these vectors as defined in (6.1)
has the following statistical properties.
(1+ P)" X
m-1 m-i-1
+
(m i -I)!
(m-I)! 2"+'i!(m- i - 1 - l)!
 PRODUCTS OF RANDOM VARIABLES 53

C. Independent Gaussian (x) Gaussian (One Has

Zero Mean, Both Have Identical Variance)

Let x"'E (x'l',

N!, a') and x'~' E Y, (0,0 2 )be independent Gaussian
vectors. Then, the inner product of these vectors as defined in (6.1)
has the following statistical properties.

(-) i-l-r

,x<O
54 PROBABILITY DISTRIBUTIONS INVOLVING GAUSSIAN RANDOM VARIABLES

(m+i+l-I)!
i=o r=o 2'i!(m+ i - l)!

(m+i+1-l)!
2li!(m+ i - l)!
 PRODUCTS OF RANDOM VARIABLES 55
px (+ -- -- - - (6.38)

D. Independent Gaussian (x) Gaussian (Both Have

Nonzero Mean and Identical Variance)

Let x") E N,,( f f l 1 ) , 0 2 )and x(')E N,,(X('),4') be independent Gaussian

vectors. For this case, the results for the PDF of the inner product of
these vectors as defined in (6.1) can be expressed in the form of a
doubly infinite series in Whittaker functions. Thus, once again,
because of the complexity of the resulting expressions and their
somewhat limited use in practical applications, the author has
decided to omit these results for this case. Nevertheless, the CF is
still relatively simple and given by

A similar statement can be made for the corresponding dependent

case. Here the CF is given by

I
xexp -
02w2(a:+ a : ) - 2(jw + p o 2 0 2 ) x$"$"

E. Independent Rayleigh (x) Rayleigh

-2 j p d w )
2(1+ 0402
/=I

" 1 (6.41)
56 PROBABILITY DISTRIBUTIONS INVOLVING GAUSSIAN RANDOM VARIABLES

Let R, = IIx")II and R2 = I x("I

be Rayleigh RVs corresponding to
independent Gaussian vectors x") E N , (0.0;) and x ( ~E )N~~~ (0,o;).
Then the product RV R = R,R, has the following statistical properties.

E { R ~ =) ( 2 0 , ~ ~ ~ ) ~ ,k integer
X
 PRODUCTS O F RANDOM VARIABLES 57
4. nl,n2

r(T)
E{R'} = ( 2 o I o 2 ) '
(2) ,k integer

F. Dependent Rayleigh (x) Rayleigh

Let R1 = IIx'"II and Z$ = IIx'~)~~

be Rayleigh RVs corresponding to
dependent Gaussian vectors x") E N,,( 0 , ~ : and
) x'~'E N,,(o,0 ; ) . Then
the product RV R = R,R2 has the following statistical properties.
G. Independent Rice (x) Rayleigh

Let R, = I x(')I and R2 = I x(')I~ be Rice and Rayleigh RVs corresponding

to independent Gaussian vectors x E N ( , ) and
x ' ~E)N,12( 0 , o ; ) . Then the product RV R = R,R2 has the following
statistical properties.

r ) [ )
r
P,(r)=l-exp --
2a0 : [ --
i,o (i!)
) 20,
- Ki+l - , r 2 0
a,a2 6 1 0 2
(6.61)

; [ 2,) [; $)
E { R ~ =} ( 2 0 , a 2 ) ' ~ ' -+1 exp ---;- ,F; - + l ; l ; l ,k integer (6.62)
 PRODUCTS OF RANDOM VARIABLES 59

.(+)
.(+)
E{R*)=( 2 0 ~ o , ) ~

(2)
I(:)
(
exp -- k+n,.nl. 4
2 $ : ) " [ T 9 2 '2af)'(6.65)

k integer

H. Independent Rice (x) Rice

Let Rl = IIx")II and R2 = I x")I be Rician RVs corresponding to

independent Gaussian vectors ( X(')a
x(') E N 9 ) and
x EN ( 2 , a ) Then
. the product RV R = R,R, has the following
statistical properties.
I. Dependent Rayleigh Products

Consider the product RV U , = RlR, where Rl = IIx")II and R, = IIx(~)II

are
Rayleigh RVs corresponding to independent Gaussian vectors
x"' E N,,( 0 , ~ : and
) x(I' E N,,( 0 , ~ : ) Likewise
. consider the product RV
U2 = &R4 where R =x and R4 = I x"'I
are Rayleigh RVs
corresponding to independent Gaussian vectors x"' E N1,( 0 , o : ) and
x E N ( 00 ) . The Gaussian vectors x"' and x"' are dependent
and likewise the Gaussian vectors x ( ~and ) x ( ~are
) dependent with
correlation matrix as in (1.10). Then U , and U2 have the joint PDF.
11-1
1
Pu, .u, (u1 u2 ) =
9
2211-40:a,2r2(n12)
 RATIOS OF RANDOM VARIABLES=

A. Independent Gaussian (+) Gaussian (Both Have

Zero Mean)

Let X, E N, (0,o:) and X2 E N, (0,~:) be independent Gaussian RVs.

Then, the ratio of these RVs X = X2 / X, has the following statistical
properties.

(2
Y, ( w ) = exp - -1w)
1
B. Independent Gaussian (+) Gaussian (One Has
Zero Mean)

Let XI E Nl(XI,0:) and X2 E N, (0,~:) be independent Gaussian RVs.

A large number of the CDF results in this section come from Ref. 11
where they appear in their normalized (the RVs that form the ratio
have unit variance) form. Once again, a large number of the PDF and
statistical moment results come from Ref. 5.
62 PROBABILITY DISTRIBUTIONS INVOLVING GAUSSIAN RANDOM VARIABLES

Then, the ratio of these RVs X = X, / XI has the following statistical

properties

is a bivariate Gaussian integral that has been tabulated in Ref. 10.

C. Independent Gaussian (+) Gaussian (Both Have

Nonzero Mean)

Let XI E Nl ($,0 : ) and X2 E Nl ( y 2 ,0,') be independent Gaussian RVs.

Then, the ratio of these RVs X = X2 / XI has the following statistical
properties
 RATIOS OF RANDOM VARIABLES 63
p (X) = - -- --- (7.8)

D. Dependent Gaussian (+) Gaussian (Both Have

Zero Mean)

:1 (
y X ( w )= exp - jwp- (1 - p 2 ~ ~ w ~ ) ]

E. Dependent Gaussian (+) Gaussian (One Has

Zero Mean)
64 PROBABILITY DISTRIBUTIONS INVOLVING GAUSSIAN RANDOM VARIABLES

F. Dependent Gaussian (i) Gaussian (Both Have

Nonzero Mean)

G. Independent Gaussian (Zero Mean) (+) Rayleigh

I x")I
Let X E N , ( 0 , ~ : )and R = with x("E N,,(0.0;) be independent
Gaussian and Rayleigh RVs. Then, the ratio Z = X I R has the
following statistical properties.
 ( z2 1 E21
Yz(o)= l+-lwl exp ---l-lwl
66 PROBABILITY DISTRIBUTIONS INVOLVING GAUSSIAN RANDOM VARIABLES

H. Independent Gaussian (Zero Mean) (+) Rice

I x("I
Let X E N, ( 0 , ~ : )and R = with x(')E N,,(%'~),
0:)be independent
Gaussian and Rice RVs. Then, the ratio Z = X I R has the following
statistical properties.

1. n = l

1
-tan-'
7C is,)
--

where
I. Independent Rayleigh (i-) Rayleigh

Let Rl = IIx'l'll and R, = I x(~'I be independent Rayleigh RVs

68 PROBABILITY DISTRIBUTIONS INVOLVING GAUSSIAN RANDOM VARIABLES

corresponding to Gaussian vectors x")E N I l , ( 0 , o ? ) a n d

x N ( 0 0 ) . Then, the ratio R = 4 I R, has the following
statistical properties.
 (recursive form)
(7.48)

(m, + m, - l)!
PR( r ) = 1- (ml - I)! (m2- I)! z(
.
m
2i )
-1
,$
m , + i o , r + 0 2 j m ' t i , r > o (7.51)
70 PROBABILITY DISTRIBUTIONS INVOLVING GAUSSIAN RANDOM VARIABLES

8. 11, = 2m, + 1, n2 = 2m2

J. Dependent Rayleigh (+) Rayleigh

Let R, = IIx"'II and 4= /Ix"']~ be independent Rayleigh RVs

corresponding to Gaussian vectors x"' E and
~@(l),o:)

x E N , ( ~ , ) . Then, the ratio R = R2 I R, has the following

statistical properties.
 ( 1 - p 2 ) ~ ' 1 ' r 2 " " ( o+~o:)
2(2m - 1)![o,o2 r2
PR( r )= !"+I12 7 r20
+ o:)2- 4 p
[ ( m- l)!12[(o:r2

( I - p2)"-'(0:- o ; r 2 ) r
PR (r)l 2n1 = PA' (~11 -
21~~-2
J j i ( 2 m - 2)(m- 2)!
111-112 9

r 2 0 (recursive form)

r 2 0 (recursive form)
K. Independent Rice (+) Rayleigh

Let R, = I x(')I
and R2 = IIxi2)ll
be independent Rice and Rayleigh RVs
corresponding to Gaussian vectors x"' E N1,(0.o:) and
x EN ( 2 , 0 ) . Then, the ratio R = R2 1 R, has the following
statistical properties.

where
 RATIOS OF RANDOM VARIABLES 73

PR ( r )=
2 r ( m + 11 2)o~"'o,r2"'-1
exp --
(
1 a;o?r2
I
f i ( m - 1)!(c(r2+ 0i)11f+112 ~ ~ ; ) 1 6 [ m + 5 ; m ; 2 0 ~ ( 0 ~ r 2 + 0 ; )

'ex'[-
a,'
1-('
2 ( 4 r 2+ 0 : ) i=2
1
4;-I
2i-2 0fr2
i -l)(o?r2
i-l
74 PROBABILITY DISTRIBUTIONS INVOLVING GAUSSIAN RANDOM VARIABLES

6.Independent Rice (+) Rice

Let Rl = I x(~'I and R2 = I x"'I be independent Rice RVs corresponding

to Gaussian vectors x ( I ) E NrIl(R('),0 : ) and x'~'
E N,, (ff'2',0 : ) . Then,
the ratio R = R2 / Rl has the following statistical properties.
 RATIOS OF RANDOM VARIABLES 75
1. n, = n2 = 1

2 a l r + a 2 a , ( 0 2l o l ) - a 2 ( o l o 2 ) r
pR(r)=
n + 2( dmolr + o2' -4
alr- a, a,( 0 2 1 o , )+ a 2 ( 0 1 0 2 ) r , r 2 0
' d m

where

d "'
X-
dc"'
76 PROBABILITY DISTRIBUTIONS INVOLVING GAUSSIAN RANDOM VARIABLES

where

where a and p are defined in (7.84) and

with 6,, denoting the Kronecker delta; 6,, = 1 for 1 = 0, 6,, = 0

otherwise.

[ :[a ++-+ /Ioi ! ~(m, +m2+ i + l - I ) !

exp --
!(m
+i-1)!(m2
, +z-I)!
 RATIOS OF RANDOM VARIABLES 77

where

c , ( m , - l , m 2 -1;r) =

M. Dependent Rayleigh Ratios

Consider the ratio RV U , = R3 I R, where R, = IIx"'~~and R3 = are IIx'~'~~

Rayleigh RVs corresponding to independent Gaussian vectors
x"' E N,,( 0 , ~ : and
) X"' E N , (~0 , ~ : ) .Likewise consider the ratio RV
U2 = R, I R2 where R, = / l ~ ' ~ ' l l and R4 = I x("I are Rayleigh RVs
corresponding to independent Gaussian vectors x"' E N,,( 0 , ~ : and
)
x E N ( 0 o ) . The Gaussian vectors x"' and x"' are dependent
and likewise the Gaussian vectors x ( ~and
) x ( ~are
) dependent with
correlation matrix as in (1.10). Then U , and U2 have the joint PDF.

4(1- p2)2"(~I~2)"-i
Po,$0,( ~ 9 u1 2 ) =
T 2( n I 2)(1+ u:)"(l+u;)"
 MAXIMUM AND MINIMUM OF PAIRS OF
RANDOM VARIABLES

Finding the probability distributions of the maximum and minimum

of groups of RVs involves the study of order statistics. Since a
complete study of this subject on its own fills an entire textbook,
examples of which are Refs. 12 and 13, we shall restrict our attention
to the case of two RVs, in particular those who first and second-order
probability distributions are discussed in Sections 2 and 3.
We denote the maximum and minimum of two RVs XI and
X2 by X,,, = max(x,,x,) and X,,, = min(xl,x,), respectively. Since
P~{x,,,,,,
I x} = pr{xI Ix, X2I x) = PX,,,*
(x,x), then the CDF of Xmaxis
simply the joint CDF of XI and X2 evaluated at the same argument,
X , i.e.,

The PDF of X,,, can then be obtained by differentiating the CDF in

(8.1). Similarly, since PT(X,~, > X} = P~{x,
> x,X, > x} = 1 - Px,
(x)- Px2
(x)
+PX,,,*(x,x), then the CDF of Xmi,, which is equal to 1- pr{xmin > x), is
given by

Again, the PDF of Xmin can then be obtained by differentiating the

CDF in (8.2).
For XI and X2 independent, we have
A. Independent Gaussian

Let XI E N ~ ( ~ , , O and
: ) X2 E N,($,o;) be independent Gaussian RVs.
Then,

B. Dependent Gaussian
 MAXIMUM AND MINIMUM OF PAIRS OF RANDOM VARIABLES 81

(see Appendix A, Eq. (A.37)for definition of y ; p) ).

Qx,,x2 (x,

C. Independent Rayleigh

Let R = x and $ = I x(~'I be independent Rayleigh RVs

corresponding to Gaussian vectors x"'E N, ( 0 , ~ : )and x"'E N, ( 0 , ~ ; ) .
Then, the maximum R,,,,, = r n a x ( ~,4) , )
, and minimum R,,i. = r n i n ( ~,$
of the pair R,, R, have the following statistical properties.
 (20;)Li1+( 2 0 : ) ~ '-~ ,k integer (8.19)
0; + 0;

[
k12

{ R i= 0; + 0; i(l+ f), k integer (8.20)

r ) r2)[(
P R ~ X ('1 = 0;( m-
20,
exp --
20;
1-exp --
[
2r0 ;2 );=, = i!L 20;
[~)i]

+ a ; ( mr - l)! ) 20; exp --20; [

r 2 ) [1-exp --20:
r 2 ) i=o [
= li![ 20:
~ji],r20

( r ) = 2 - exp - -
PR,na5
2r2
0; ( )= i=,
I[ 20:
i! r2 li [
-exp -7
2r2
0 , )=L[L)'
;=, i! 2 0 ;
D. Dependent Rayleigh

where
 - Iplo,a2cos[m(6+ n / 2)]
-(lll-I)12

0:+ 21p10,02sin 6 + p20:

a: + 0;+ 21pI0,a~sin 6
h,(6)4
240;(1- p 2 )

and
 m"' 1
( m - I ) ! 277
I' (hl(0))-".f,,
-n
( 7 ,h;' (e))h(e)dO,r 2 0
(8.34)

+ o,2(m-1)!
r [ r 2 2 ) ' 1 ' -exp
20,
1 [ --r 2 2 ]
202

+-- m"' 1
( m - I)! 2 n -"
I'
(h,(6))-"'
fll (r,h;' (6))h(6)d& r 2 0
(8.36)

E. Independent Log-Normal
Let y , = 10X1"Oand y , = 1 0 ~ ~ be " ~ independent log-normal RVs
corresponding to Gaussian RVs XI E N, ( F , ,0 2 )and X2 E N, ($,0 2 ) .
Then, the maximum y = max(y,,y 2 ) and minimum Ymi.= m i n ( ,,~y 2 )
mdx

of the pair y ,,y2 have the following statistical properties.

 " Y = - (
10 log,, y - X,
, ( IOlog,, y - X,
02 ).Y20 (8.40)

F. Dependent Log-Normal
 10 log,, y - j7,

(8.43)
9

0 2

Q[
0,(10 log,, Y - x 2 ) - PO, (10 log,, Y - XI)
0102 4- I
 QUADRATIC FORMS

Let X and Y be complex Gaussian vectors each of dimension n. As

usual, the components of each vector are independent and have
identical variances, which for complex components are defined as

Furthermore, as before only components of X and Y with identical

subscripts can be correlated with complex cross-correlation defined
by

For convenience, we define pairs of complex components with the

same subscript by the vectors

and analogous to (1.10) we define the covariance matrix

We are interested in the statistics of the RV having the quadratic form

where
which can also be written in the vector form

Finally, we define the matrix P by

where from (9.4) and (9.7)we obtain

P, = AP, + C k Y 9 Py, = BPyy+ C * k Y

(9.9)
P, = C * P , + BP;p Pyx = CPyy+ A &

Also, for what follows it is convenient to define the parameters

which when using (9.9) evaluate to

with
 QUADRATIC FORMS 91

A. Both Vectors Have Zero Mean

(v,v2)" n + k - 1 (-d)"-k-l

P&) =
exp(v2d)C
k=, (v, +
,,)'I+* (
k k - , ! ,d<O
(9.15)
,d2O

A special case of (9.16) that is of practical interest (see [14], Appendix

B and [15], Appendix 9A, for example as applied to the study of the
performance of digital communication systems over fading channels)
corresponds to P,(O) = Pr{D S O ) . For this case, (9.16) reduces to
An alternative form for P,(O) which will appear as a special case of
the results to be presented in the next section for nonzero mean
vectors is given by

B. One or Both Vectors Have Nonzero Mean

When one or both of the two vectors have nonzero mean, finding
closed-form expressions for the PDF and CDF of the quadratic form in
(9.5) is difficult. However, for the particular case corresponding to
PD(0), it is possible to obtain such expressions [14], [15]. Two
different closed-forms versions will be provided, the first involving
the first-order Marcum Q-function and modified Bessel functions of
order O,l, ...,n - 1(as well as other elementary functions) and the
second involving only the zero-order modified Bessel function but
instead Marcum Q-functions of order 1,2,...,n . Also provided will be
a finite-limit integral form for P,(O) that is derived based on the
alternative representation of the Marcum (2-function discussed in
Appendix A and is convenient for numerical evaluation purposes.
The results are as follows:

where
 QUADRATIC FORMS 93

with

where
I
dB, a < b
94 PROBABILITY DISTRIBUTIONS INVOLVING GAUSSIAN RANDOM VARIABLES

Note that the form of PD(0)in (9.22) does not depend on the ratio a 1 b
and thus for the case where both vectors have zero mean, whereupon
a = b = 0, using the fact that Q,(0,0)= 1, we immediately obtain the
result in (9.18).

C. A Reduced Quadratic Form Where the Vectors

Have Different Numbers of Dimensions
We consider now a reduced version of the quadratic form in (9.5)
combined with (9.6),namely,

where the complex Gaussian vectors X and Y now have dimensions

n, and n,, respectively. Once again for the most general case where
one or both of the two vectors have nonzero mean, finding closed-
form expressions for the PDF and CDF of the quadratic form in (9.25)
is difficult. As in the previous section, however, it is possible to arrive
at closed-form expressions for P,(O) [16]. Analogous to (9.22) and
(9.23), the results are as follows:
 QUADRATIC FORMS 95

where

and because C = 0, a and b of (9.20) simplify to

Note that since yU,pyy> 0, then for either A < 0 or B < 0, the
arguments of the square roots in (9.30) are always positive.
D. General Hermetian Quadratic Forms
As an alternative to describing each component of the quadratic form
of (9.5) in a matrix form as in (9.7), the quadratic form itself can be
expressed in a matrix form as follows. Let

'H 0 . . 0 -
O H 0 0
Q = . 0 . . .
0
-O O O H-

Then, analogous to (9.7),

Note from (9.31) together with (9.7) that the matrix Q is Hermetian,
i.e., Q = Q*'.
Consider now a general quadratic form of the type in (9.32)
where Q is again Hermetian but not restricted to the diagonal form in
(9.31). Furthermore, denoting the complex elements of V by
X,, = X,,, + jX,, and q, = q,, + jqrl,,,assume that their real and imaginary
parts satisfy the relations

and
 QUADRATIC FORMS 97
Then, denoting the complex covariance matrix of V by
L = E{(V - V)(V - v)*~}
assumed to be nonsingular, the CF of D is
given by [17]

YD(w)= 11- j w ~ ~ lexp

-I {- V e T ~ I- - (I - ~ ~ L Q ) - ' ] B }
I[ (9.35)

where I is the identity matrix.

The quadratic form in the argument of the exponential of
(9.35) can be diagonalized by use of the transformation V = U,N-'U2F
where U, is the normalized modal matrix of L-I, N2 = UTL-IUI, and
U, is the normalized modal matrix of N-'UTTQU,N-'. Then,
alternatively to (9.35)

where the f,'s are the elements of F and the Ak's are the eigenvalues
of LQ.
A special case of (9.36) corresponds to X and Y having zero
mean in which case V = 0. For this case, the f,'s are all equal to zero
and thus (9.36) simplifies to

whose singularities consist solely of a finite number of finite-order

poles.6 Furthermore, the mean and variance of D are given by

E{D} = 5Ak, var D = EA:

211

In the most general case of nonzero mean X and Y vectors,

determining the PDF of D from the inverse Fourier transform of
YD(w)is difficult because of the complicated nature of the exponential

The appearance of poles in YD(a)when V = 0 occurs because of the

constraints placed on the real and imaginary parts of the components
of V as in (9.33) and (9.34).
factor in (9.36). However, for the zero mean vector case, once the
eigenvalues of LQ are found, p,(d) is readily determined from (9.37)
through the use of the residue theorem. For example, if all
eigenvalues have multiplicity one, then the partial fraction expansion
of (9.37) is given by

with corresponding PDF

where K+ is the set of k corresponding to positive eigenvalues and

K- is the set of all k corresponding to negative eigenvalues.
Furthermore,
 OTHER MISCELLANEOUS FORMS

A. lndependent Rayleigh (+) Rayleigh

Let R =x and R2 = IIx( ~ ) I I be independent Rayleigh RVs

corresponding to Gaussian vectors x(')E N, ( 0 ,0 2 )and x") E N2( 0 ,0 2 ) .
Then, the sum RV R = R, + R2 has the PDF

B. lndependent Gaussian (x) Rayleigh

Let X E N, ( 0 , ~ : and
) R= I x")I be independent Gaussian and
Rayleigh RVs where x ' ~E)N , ( o , D : ) . Then, the product RV Z = XR
has the PDF

C. lndependent Gaussian (x) Rayleigh (+) Gaussian

100 PROBABILITY DISTRIBUTIONS INVOLVING GAUSSIAN RANDOM VARIABLES

I x("I
Let XI E N , ( o , ~ : ) ,R = and X, E ~ ~ ( 0 . beo ~independent
)
Gaussian ,Rayleigh and Gaussian RVs and x"' E N, ( 0 , ~ : ) Then,
. the
RV Z = XIR + X, has the PDF

D. Independent Gaussian (+) Rayleigh

Let x E N,(o. o:) and R = I x'~'I

be independent Gaussian and
( :0) ., Then, the RV Z = X + R has the
Rayleigh RVs where x"' E ~ ~ 0
PDF

E. General Products of Ratios of lndependent

Gaussians

Let X, E N , (0,of),Y, E N, (0,o:), k = 1,2,...,n be mutually independent

Gaussian RVs. Define the RV

Then, the PDF of Z is as follows:

OTHER MISCELLANEOUS FORMS 101
 APPENDIX A: ALTERNATIVE FORMS

In this appendix, we present a number of alternative forms for the

Gaussian Q-function, Marcum Q-function and incomplete gamma
function. The motivation behind these forms comes from applications
where it is desirable to have an integral representation in which the
argument of the function is neither in the upper nor lower limit and
furthermore appears in the integrand as an argument of an
elementary function such as exponentials or trigonometrics.

1. One-Dimensional Distributions and

Functions

A. The Gaussian Q-Function

The Gaussian Q-function, Q(x), which is defined as the complement
(with respect to unity) of the CDF corresponding to a Gaussian RV,
has the canonical representation given in (1.3). In principle, this
representation suffers from two disadvantages. First, from a
computational standpoint, (1.3) requires truncation of the upper
infinite limit when using numerical integral evaluation or algorithmic
techniques. Second, the presence of the argument of the function as
the lower limit of the integral poses analytical difficulties when this
argument depends on other random parameters that ultimately
require statistical averaging over their probability distributions. Such
is the case, for example, when studying the performance of digital
communication systems over additive white Gaussian noise (AWGN)
channels that are also perturbed by random fading [15]. In such
instances, what would clearly be more desirable would be to have a
form for Q(x) wherein the argument of the function is neither in the
upper nor the lower limit of the integral and furthermore appears in
the integrand as the argument of an elementary function. Still more
desirable would be a form wherein the argument-independent limits
are finite and the integrand maintains its Gaussian nature.
Although the existence of such a form dates back to a 1972
classified report (which has since become unclassified) by Nuttall [911,
its popularity and recent extensive application primarily stem from its
reintroduction by Craig [20] who presented it as a by-product of his
work on the evaluation of the average probability of error
performance for the two-dimensional AWGN channel. Specifically,
the so-called Craig representation for Q(x) (but only for x 2 0) is
given by

The form in (A.l) is not readily obtainable by a change of variables

directly in (1.3). However, by first extending (1.3) to two dimensions
(x and y) where one of the dimensions (y) is integrated over the half
plane, a change of variables from rectangular to polar coordinates
readily produces (A.l). Furthermore, (A.l) can be obtained directly
by a straightforward change of variables of a standard known integral
involving Q(x), in particular, [I, Eq. (3.363.2)l. A description of both
of these derivations can be found in [15, Appendix 4A]. Yet another
derivation of (A.l) is given in Ref. 21 and is based on the fact that
since the product of two independent random variables, one of which
is a Rayleigh and the other a sinusoidal random process with random
phase, is a Gaussian random variable, determining the CDF of this
product variable is equivalent to evaluating the Gaussian Q-function.

B. The Marcum Q-Function

The first-order Marcum Q-function, ~ ( a , p )which

, is defined as the
complement (with respect to unity) of the CDF corresponding to a
normalized noncentral chi-square random variable with two degrees
of freedom has the canonical representation given in (2.20). It is of
interest to note that the complement (with respect to unity) of the

' The relation given there is actually for the complementary error
function which is related to the Gaussian Q-function by (1.4). Other
early traces of this alternative form can be found in the work of
Weinstein [I81 and Pawula [19].
first-order Marcum Q-function can be looked upon as a special case of
the incomplete Toronto function 122, pp. 227-2281 which finds its
roots in the radar literature and is defined by

T,(rn, n, r ) = 2r"-""1e-r2
JOB t"-"e-"I,, (2rf)dt

In particular, we have

Furthermore, as /3 + w, Q, (a,P) can be related to the Gaussian Q-

function as follows. Using the asymptotic (for large argument) form
of the zero-order modified Bessel function of the first kind, we get [9,
Eq. (A-27)]

Using integration by parts, it can also be shown that the first-

order Marcum Q-function has the series form

A comparison of (2.20) with (1.3) reveals that the canonical

form of the Marcum Q-function suffers from the same two
disadvantages as previously discussed for the Gaussian Q-function.
Once again from the standpoint of application, it would be desirable
to have an integral form of the Marcum Q-function in which the limits
are finite and the integrand possesses a Gaussian nature. Historically,
the roots of such an alternative form can be traced back to a 1965
report by Stein [23]; however, the more popular and available
disclosures from which a recent flurry of applications have stemmed
are found in Refs. 24/25. Specifically, it has been shown that the first-
order Marcum Q-function has the desirable representation
 [ [ (3.(3l]
xexp -- 1+2 - slnO+ - de,P>alO

An analogous development to that used in arriving at (A.6) yields the

result2

[ [ (3 (a')]
xexp -- 1+2 - sine+ - dO,a>P2O

The results in (A.6) and (A.7) can put in a form with a more
reduced integration interval. In particular, using the symmetry
properties of the trigonometric functions over the intervals (-z,O)
and (O,z),we obtain the alternative forms

At first glance it might appear from (A.7) that the Marcum-Q

function can exceed unity. However, the integral in (A.7) is always
less than or equal to zero. It should also be noted that the results in
(A.6) and (A.7) can also be obtained from the work of Pawula [26]
dealing with the relation between the Rice le-function and the
Marcum Q-function.
and

Before moving on to the generalized (mth order) Marcum Q-

function, we discuss another alternative, and in some sense simpler,
form of the first-order function that dispenses with the trigonometric
factor that precedes the exponential in the integrands of (A.6) and
(A.7) in favor of the sum of two purely exponential integrands each
still having the desired dependence on a or p as appropriate. This
relatively recent discovery due to Pawula [27] is given as

PrarO
and

Q,(a, + L/"
= 1 47~-" {exp[- $[I +)!(2 I)')!(
sin0 +

or equivalently in the reduced forms analogous to (A.8) and (A.9)

and

Since the first exponential integrand in each of (A.lO) through (A.13)

is identical to the exponential integrand in the corresponding
equations (A.6) through (A.9), we can look upon the second
exponential in the integrands of the former group of equations as
compensating for the lack of the trigonometric multiplying factor in
the integrands of the latter equation group.
The generalized ( mth-order) Marcum Q-function, Q,,,(a,P),
which is defined as the complement (with respect to unity) of the CDF
corresponding to a normalized noncentral chi-square random variable
with m + 1 degrees of freedom has the canonical representation given
in (2.24). Once again, it is of interest to note that the complement
(with respect to unity) of the generalized Marcum Q-function can be
looked upon as a special case of the incomplete Toronto function.
Specifically, analogous to (A.3),

Furthermore, as P + =, Q,,,(a,P) can be related to the generalized

Gaussian (2-function in the same manner as was done for the first-
order Marcum Q-function. Specifically, since the asymptotic (for
large argument) form of the kth-order modified Bessel function of the
first kind is independent of the order, then
 APPENDIX A: ALTERNATIVE FORMS 109

A comparison of (2.24) with (1.3) and (2.20) reveals that the canonical
form of the generalized Marcum Q-function suffers from the same
two disadvantages as previously discussed for the Gaussian Q-
function and the first-order Marcum Q-function. Once again from the
standpoint of application, it would be desirable to have an integral
form for the generalized Marcum Q-function in which the limits are
finite and the integrand possesses a Gaussian nature. The discovery
of such a form was made independently in Refs. 24 and 25 with the
following results:

xexp
[ ";
-- 1+2 - sine+
(ill]
- do, O + l a l P < l

and

xexp
[ --
( 3I)'):(
1+2 - sine+ - dB, O < P l a < l

that can be further simplified and separated into m odd and m even as
110 PROBABILITY DISTRIBUTIONS INVOLVING GAUSSIAN RANDOM VARIABLES

and

[.;! (3
xexp -- 1 + 2 - sine+
cr]]
- de,OIPla<l,meven
 APPENDIX A: ALTERNATIVE FORMS 111
We observe from (A.16) through (A.21) that a andB are restricted to
be unequal. The special case of a = P has the closed-form result [24]

Finally, we note that unlike the first-order Marcum Q-function,

additional simplification as in (A.8) and (A.9) are not readily possible
for the mth-order function.

C. The Nuttall Q-Function

The Nuttall Q-function [9] is a generalization of the Marcum Q-
function and is defined in (4.41) together with its recursive relation in
(4.42). Since the Nuttall Q-function is not considered to be a tabulated
function, it is desirable to express the recursion relation of (4.42) in a
form that allows solution in terms of the first-order Marcum Q-
function which has been tabulated [8] or equivalently in terms of the
generalized Marcum Q-function. The possibility of doing such when
m + n is odd was suggested in Ref. 9; however, the explicit solution
was not provided. Letting m = n + 2 k + 1 in which case
m + n = 2(n+ k ) + l and is therefore odd, an explicit expression for
(a,B) in terms of only the first-order Marcum Q-function and
Q,l+2ktl,,l
modified Bessel functions of the first kind was found by this author
[28]. Specifically,

where
112 PROBABILITY DISTRIBUTIONS INVOLVING GAUSSIAN RANDOM VARIABLES

is the generalized (kth -order) Laguerre polynomial [I] and &(p2)is

a polynomial of order k - 1 in P2,namely,
k-l

with

- 1 -
d, ( k ,1 ) 4 2k-1-J( )
k - 1 - j ( n + l + j)! (1-I)!
)
( n + k ) ! =2k-,-j ( - 1 j ) ! k + n
k-1- j
(A.26)

Note that for k = 0, l$"(x) = 1 and (A.23) reduces, as it should, to

(4.43). Also, the two weighted sums of Bessel functions in (A.23)
could, in principle, be combined into a single sum of weighted Bessel
functions; however, the additional notational complexity needed to
do this is probably not worth it. Finally, since exp[-(a2 + b2)l2]1,,(ab)
can be expressed in the form

(a,p) can be put into the form of say (A.16), namely,

then Qllt2k+l,l,

where g,,,(8;a,p) is a purely trigonometric function of 8.

Unfortunately, for rn + n even, as pointed out by Nuttall [9], a
solution for Qlll,ll(a,b),e.g., Q,,,,,,,
(a,b) with k an arbitrary integer,
requires having two fundamental functions, either Q,,o(a,b) and
b) or $,o(a, b)and Q,,(a,b ) , neither pair of which is related to
Q2,0(a,
the Marcum Q-function or other tabulated functions. Thus, it is not
b) in the form of (A.28).
possible to find a solution for Q,,,,,,(a,

D. The Complementary Incomplete Gamma

Function
The complementary incomplete gamma function, T(a, x), is defined
by

or equivalently, letting t = y 2 I 2,

T(a,x) = -I
1 -
y2"-'exp
2"-' Jr;

Two special cases of this function are of interest. First for a = 1, a

normalized version of (A.29) gives the exponential function, i.e.,

Second for a = 112, a normalized version of (A.30) gives the Gaussian

Q-function, i.e.,

Analogous to what was done for the Gaussian Q-function and

Marcum Q-function, the incomplete gamma function can be put into a
more desirable form where the limits of the integral are finite and the
integrand retains its Gaussian nature. Specifically, it can be shown
that T(a, x) has the form
114 PROBABILITY DISTRIBUTIONS INVOLVING GAUSSIAN RANDOM VARIABLES

or equivalently

Since from (A.l) we have

then equating (A.35) with (A.32) and using the form in (A.34) gives
yet another alternative form for the Gaussian Q-function, namely,

r(1/2,x2/2)- 1 coss x x2
Q(x)= 2 r j q -XI~
WE (
exp --
2s1n20)
d0 (A.36)

2. Two-Dimensional Distributions and

Functions

A. The Gaussian Q-Function

The two-dimensional Gaussian Q-function, ~ ( x~,; p )is, defined by

QkY;P) =
1
2nJI-p2
u2 + v2 - 2puv
1
dvdu

and is the complement of the two-dimensional CDF corresponding to

(A.37)

the PDF in (3.2) with proper change of notation. Using a geometric

approach similar to that used by Craig [20] in arriving at the
alternative form of the one-dimensional Gaussian Q-function in (A.l),
the author first derived an analogous form for ~ ( xy;,p) in the region
x 2 0,y 2 0 as [15, Eq. (4.7)]
 APPENDIX A: ALTERNATIVE FORMS 115

More recently, this author derived [28] another alternative form for
~ ( x , ~ ;inp this
) same region that dispensed with the trigonometric
factor that precedes the exponentials in the integrands of (A.38) and
furthermore resulted in an exponential argument that is precisely in
the same simple form as that in the Craig representation of ~ ( x ) .
Specifically:

x , ;p = J1
2n 0
tan-'
I
dl-p2xly
-
l-pxly ]ex(-

dl-p2y/x
+ ~2n~ ~ - l [ w ] ~ ~ ( -

Note that for p # 0 but y = x , (A.39) simplifies to

Also, for p = 0, we immediately get from (A.40) the Craig form for
Q 2(x), namely [25, Eq. (80)]

whereas for p = 1, (A.40) becomes the Craig form for ~ ( x. )

To arrive at alternative forms analogous to (A.39) for other

regions of x and y , we first note the following relations:

Since 1- px 1y and 1- py 1x can take on positive or negative values,

the arctangents in the upper limits of the integrals in (A.39) are
defined by tan-' X I Y = n(1- sgn Y )12 + (sgn Y )tan-' X 11~1.
Recognizing that because of the symmetry of the sinusoidal function,
the alternative form of the one-dimensional Gaussian Q-function in
(A.1) can also be written as

then substituting (A.45) and (A.39) in (A.42) - (A.44), we arrive at the

following:
 APPENDIX B: INTEGRALS INVOLVING
THE GAUSSIAN Q-FUNCTION AND THE
MARCUM Q-FUNCTION

The material in this appendix is largely taken from a number of public

documents (including the author's) 19, 15, 29, 301 some of which
again, because of their age and origin or specialization of subject, may
not be readily available or familiar to the average reader. It is for this
reason that this appendix is being included in the book. Not all
integrals involving these functions are included in the tabulation.
Rather, emphasis is placed on those that less commonly appear in the
published literature but are of interest in practical applications

1. The Gaussian Q-Function

Using the alternative form of the Gaussian Q-function as given in
(A.l), we can express integrals of the form ) - ~ ( o f i ) ~ ( y ) din
y terms of
the Laplace transform of g(y), namely, G(s) = bevg(y)dy,by4

or equivalently

4
Quite often one is interested in integrals of this type where g(y)
takes the form of a PDF.
The advantage of (B.l) or (B.2) is that the integration now has finite
limits and thus even if the Laplace transform is such that the integral
cannot be obtained in closed form, it can still be evaluated quite
simply by numerical integration. In this section of the appendix, we
focus on functions g ( x ) whose Laplace transform is known in closed
form and as such the integral can be expressed either in the finite
integral form of (B.l) or (B.2) or in closed form.

A. Q-Function and x

B. Q-Function with Exponenfials and x

e(2k)[ )
=
,, k
mb2
k

4(a2+ m b 2 )
, 2 0. m integer

a 2 0, m noninteger
 APPENDIX B: INTEGRALS INVOLVING Q-FUNCTIONS 119

C. Q-Function with Exponentials, Bessel Functions

and x

2. The First-Order Marcum Q-Function

A. Q-Function with One Linear Argument

B. Q-Function with One Linear Argument and

Exponentials
120 PROBABILITY DISTRIBUTIONS INVOLVING GAUSSIAN RANDOM VARIABLES

[exp(-p2x212)Q,(b,ax)dx =

[exp(-p2x21 2)Q,(a,b)& =

C. Q-Function with One Linear Argument and x

2 + b2 - a2c2
J c m x(b,
~ ,a ) d x =
2a2 2a (
Q,(b,ac) + -exp -
b2) (l3.14)
x [acl,(abc)+ bl,(abc)]

x [acl,(abc)+ bl,(abc)]

x [aclo(abc)+ bl,(abc)]

1 x[l - Q,( a ,b)]dx=

b2 - a2c2
2a2 2a (
Ql(b,ac) + -exp -
a2c;b2),,.l,)
x [acI0(abc)+ 61,(abc)]
 APPENDIX B: INTEGRALS INVOLVING Q-FUNCTIONS 121
D. Q-Function with One Linear Argument,
Exponentials and Powers of x

ICmx exp(--p2x2
/ 2)Ql (ax,b)dx =
1
P
exp(- ?)
p z ~ 2
QI (ac,b )

jocx exp(p2x2I 2)Q1(ax,b)dx =

Icmxexp(-p2x212)Q1(b,ux)dx =
122 PROBABILITY DISTRIBUTIONS INVOLVING GAUSSIAN RANDOM VARIABLES

jOc
x e x p ( - p 2 x 21 2)Q1(b,ar)dx =

loc
x e x p ( p 2 x 2I 2)Q1(b,ax)dx =

[x2"-I e ~ p ( - ~ ' x1' 2)Q1(b,m ) d x = 2"-I ( n - I ) !

P2"
where

and L,(x) is the Laguerre polynomial which is the special case of the
generalized Laguerre polynomial defined in (A.24) corresponding to
n = 0, i.e.,

E. Q-Function with One Linear Argument, Bessel

Functions, Exponentials and Powers of x

[ x e x p ( - p 2 x 21 2)10( c x ) ~( b, ,ax)& =
 [XI, ( c x ) ~(b,, ux)dx =
[x enp(-p2x2I 2)1,( b x ) ~(c,, a ) d x =

[x exp(-p2x21 2)1,( b x ) ~(ax,

, c)dr =

Note that in checking the consistency between (B.35) and (B.36), use
has been made of the relation

1
Q,(a,a )= -[I + e x p ( - a ' ) ~ ~ ( a ~ ) ] (B.37)
2

which holds even when a is imaginary and also the fact that
I, ( x ) = Io(-x).

I, (cx)Q,(b,ax)dx = (B.38)
C
&xi exp(-p2x2/ 2 ) 4 ( c x ) ~(ax,
, b)dx =

[x2 exp(-p2x2/ 2 ) 4 ( c x ) ~( b, ,m ) d x =

F. Product of Two Q-Functions with One Linear

Argument

[x exp(-p2x2 / 2)Q1(ax,b ) ~(cx,, d)dx =

abc
11 (B.42)
126 PROBABILITY DISTRIBUTIONS INVOLVING GAUSSIAN RANDOM VARIABLES

G. Q-Function with Two Linear Arguments and x

c2 bc2
xQ, (ax,bx)dx = -Q,(ac,bc) +
2 2(a2- b 2 )

(
xexp -- + a11(abc2)]
b2 c2)[b$(abc2)

-2Q1 ( cmin(a,b),c max(a,b))], a + b

c2
Joc XQ, (ax,ax)dx = -{1
4
+ exp(-02c2)[Io(a2c2)
+ I, (a2c2)l} (8.44)

H . Q-Function with Two Linear Arguments,

Exponentials and x
To simplify the notation, define

Then

% x exp(-p2x2l 2)Q,(ax,bx)& =
 [x exp(-p2x212)Q1(ax,bx)& = - (1l
2P2
):
+

3. The Generalized (mth-Order) Marcum Q-

Function

A. Q-Function with One Linear Argument and

Powers of x

[x~'~-'Q,,,
( b ,m ) d x =
2 ( n-1 );
4,, , 1 ( -- ,n2l (B.50)
a2"

B. Q-Function with One Linear Argument,

Exponentials and Powers of x
 4P ): {(y)
111-l

[ x exp(-p2x21 2)QI1
(ax,b)dx =

[[
x exp
a
2(a2+ p 2 ) =
,,
a2b2 1.1+x-
k! 2(a2+ p 2 )
- 2

=
1(b2).]
-
k! 2

C. Q-Function with One Linear Argument, Bessel

Functions, Exponenfials and Powers of x
D. Q-Function with Two Linear Arguments,
Exponentials and Powers of x

1 x e x p ( - p 2 x 2l 2 ) ~(ax,
, bx)k=

where r , s , and t are defined in (B.45).

l s-r f s-r\"'\

h xzk-'~ x ~ ( l z)Q,,, ~ bx)dx

- ~ (ax, x ~ =

For the special case k = m = 1, using the relations

and
130 PROBABILITY DISTRIBUTIONS INVOLVING GAUSSIAN RANDOM VARIABLES

the result in (B.60) reduces to the simple form in (B.48).

 APPENDIX C: BOUNDS ON THE
GAUSSIAN Q-FUNCTION AND THE
MARCUM Q-FUNCTION

To facilitate evaluations involving the Gaussian Q-function and the

Marcum Q-function, it is quite often useful to have upper and lower
bounds on these functions as an alternative to their exact integral
representations. In this appendix, we present a variety of such
bounds many of which are derived from the alternative forms of the
functions presented in Appendix B. Our interest here is not in
comparing the degree of tightness of these various bounds but rather
to merely present them to the reader allowing him or her to determine
the advantages and disadvantages of each form for his or her
particular application.

I.The Gaussian Q-Function

Perhaps the simplest upper bound on the Gaussian Q-function can be
immediately obtained from the alternative form given in (A.l) by
noting that the integrand is a monotonically increasing function of the
integration variable and thus the maximum of this integrand occurs at
8 = n 1 2 . Thus, upper bounding the integrand by its maximum value,
namely, exp(-x2 1 2), we immediately obtain

which is equivalent to the Chernoff bound on this function improved

by a factor of 1/2. This procedure for obtaining the upper bound can
be extended to provide tighter bounds [31]. In particular, we note that
n, ]
any monotonically increasing function f ( x ) in the interval [ x m i x,,,
132 PROBABILITY DISTRIBUTIONS INVOLVING GAUSSIAN RANDOM VARIABLES

can be upper bounded by a piecewise constant (staircase) function

j(x) defined as follows:

'f (x,), X,," = X, I X I XI

f (x,), X, < x 5 X,

where x, < x2 < ... < xN-, is a set of N - 1 monotonically increasing

values of x arbitrarily chosen in the interval (xmi,, x,,, ) . Since the
integrand of (A.l) is a monotonically increasing function of 0, then it
can be upper bounded in the interval (-nI 2, n I 2) as in (C.2) in which
case the integral, Q(x), is upper bounded by

where 0 = 0, < 0, < 0, < ... < ON-,< 8, = n I 2. Note that (C.3) is nothing
other than a particular form of Riemann sum that is ordinarily used to
approximate the integral of a function, and thus as N increases, the
tightness of the bound improves. Clearly, for the special case N = 1
whereby 0, = 0 and 9, = n 1 2 , (C.3) reduces to (C.l).
The two-dimensional Gaussian Q-function of (A.39) can
similarly be upper bounded. Consider first the case where 1- py 1 x
(or 1- px I y ) is positive. Then, as for the one-dimensional Gaussian
Q-function, the integrands of (A.39) are a monotonically increasing
function of 0 in their respective integration intervals in which case
the integral, ~ ( xy:,p), is upper bounded by
where

and Nil,Ni2 are the smallest integers such that

For the case where 1- py 1 x (or 1- px 1 y ) is negative, then in the

interval between n 1 2 and the upper limit, the integrands are a
monotonically decreasing function of 8 and can be upper bounded by
a downward staircase function analogous to (C.2). Hence, for this
case, defining Nit!,Ntl2as the largest integers such that

then the two-dimensional Gaussian Q-function can be upper bounded

by

1
Q(x,Y :p ) 5 -
[ L)+'
(8,- Bi-,) exp -
(R-4-l)

2n {=I 2 sin28, 2 n i=N,,, +2

Another interesting upper bound related to the ability of

separating the argument of the function can be obtained from the
alternative form in (A.l). In particular, we first note from this
equation that
 1 n/2
Q(J=)=-~
7r 0 (
exp --
2s': 0) ex'(-&)de (C.9)

Now applying Schwarz' inequality [32] to the integral in (C.9), we

have

A simple integral form of the deviation of the Gaussian Q-

function from the first term of its asymptotic series can be obtained
from the alternative form of this function in (A.36). First integrating
(A.36) gives

But the definite integral of the Gaussian Q-function can be found in

closed form using integration by parts. In particular,

Equating (C.ll) and (C.12) gives

where the first term corresponds to the leading term in the asymptotic
expansion of Q(x). Thus, the second term can be regarded as an
integral representation of the deviation of Q(x) from its
approximation by the first term of its asymptotic series. Furthermore,
since the second term is always negative for x > 0, then the first term
is also an upper bound on Q(x), i.e.,
 APPENDIX C:BOUNDS ON Q-FUNCTIONS 135
A lower bound on Q(x) can be obtained by first rewriting the
classical representation of Q(x) in (1.3) as

and then integrating by parts using 1u d v 1

= uv - v d u where

resulting in

Further integrating by parts the second term in (C.17) with

v=-exp(-$), dv=yexP(-$)dy
we obtain

Ignoring the third (integral) term in (C.19) , we obtain the desired

lower bound for x > 0

2. The Marcum Q-Function

 Simple upper and lower bounds on Q,(a,P) can be obtained in
the same manner that the upper bound on the Gaussian Q-function in
(C.l) was obtained from (A.l). In particular, for P > a 2 0, we observe
that the maximum and minimum of the integrand in (A.6) occurs for
8 = - 7 ~/ 2 and 8 = 7 ~1 2, respectively. Thus, replacing the integrand
by its maximum and minimum values leads to the upper and lower
"Chernoff-type" bounds

9
P+a
exp - ( ( " Y Y ] ~Q 1 ( a , P ) s Gfi e x p - (P - 4' ] (c-21)

which, in view of the fact that Q , ( o , ~ ) = e x p ( - ~12) ' are

asymptotically tight as a + 0.
For a > P 2 0, the integrand in (A.7) has a minimum at
8 = -n I 2 and a maximum at 8 = n / 2. Since the maximum of the
integrand, [P l ( a + B)]exp[-(a + B)' / 21, is always positive, the upper
bound obtained by replacing the integrand by this value would
exceed unity and hence be useless. On the other hand, the minimum
of the integrand, -[/3 l ( a - B)]exp[-(a - B)' / 21, is always negative.
Hence a lower Chernoff-typebound on Q,( a , P) is given bf

1- -X
a-P
a
- (a ] (a.P)
iPi2 5 Q1

The forms of the Marcum Q-function in (A.lO) and ( A . l l )

immediately allow obtaining tighter upper and lower bounds of this
function than those in (C.21) and (C.22). In particular, once again
recognizing that for P > a 2 0 the maximum and minimum of the first
exponential integrand in (A.lO) occurs for 8 = -n / 2 and 8 = n / 2 ,
respectively, and vice versa for the second exponential integrand,
then we immediately obtain

Clearly since Q, (a$) can never be negative, the lower bound is only
useful for values of the arguments that result in a non-negative value.
 APPENDIX C: BOUNDS ON Q-FUNCTIONS 137
Making a similar recognition in (A.11), then for a > /3 2 0 we obtain
the lower bound

{ [
1 - - 1 exp -
2
-
2
"'1 - '2"'I] 5 Ql (a.S) (C.24)

The advantage of the bounds is that they are simple and of

exponential-type; however, they may not necessarily be the tightest
bounds achievable over all ranges of their arguments. In this regard,
the following upper and lower bounds have also been reported [33].

For P > a , the bounds in (C.25) are tighter than those in (C.23).
However, for a > /3 the lower bound in (C.24) is very tight and better
than that in (C.25).
Upper and lower bounds on the mth-order Marcum Q-
function are not readily obtainable from the alternative forms in
(A.16) and (A.17). Nevertheless, it is possible [34] to obtain such
bounds by using the upper and lower bounds on the first-order
Marcum Q-function given in (C.23) together with the recursive
relation

The results are as follows:

Note that the first term in (C.27) represents the upper bound on the
first-order Marcum Q-function as given in (C.23). Similarly, the first
term (corresponding to n = 0) in (C.28) is the lower bound on the first-
order Marcum Q-function as given in (C.23). Also, for a = 0, (C.28)
becomes equal to the exact result for Q,(o,~)as given by

Finally, because of the monotonicity of the integrands in (A.12)

and (A.13), the technique used to upper bound the first- and second-
order Gaussian Q-functions in (C.3) and (C.4) can also used to upper
bound the first-order Marcum Q-function. In particular, the following
results are obtained:

and
 REFERENCES

[I] I. S. Gradshteyn and I. M. Rhyzhik, Table of Integrals, Series, and

Products, 5thed., San Diego, CA: Academic Press, 1994.
[2] M. Abramowitz and I. A. Stegun, Handbook of Mathematic Functions
with Formulas, Graphs, and Mathematical Tables, U. S. Department of
Commerce, National Bureau of Standards Applied Mathematics
Series #55 (reprinted by Dover Press), 9thprinting, 1970.
[3] G. E. Andrews, R. Askey and R. Roy, Special Functions, Cambridge
University Press.
[4] L. B. W. Jolley, Summation of Series, 2ndrevised ed., New York, NY:
Dover Press.
[5] K. S. Miller, Multidimensional Gaussian Distributions, New York,
NY: John Wiley & Sons, Inc., 1964.
[6] J. K. Omura and T. Kailath "Some Useful Probability Functions"
Tech. Rep. 7050-6, Stanford University, September 1965.
[7] M. Nakagami, "The m-distribution: a general formula of intensity
distribution of rapid fading," in Statistical Methods in Radio Wave
Propagation. Oxford: Pergamon Press, 1960, pp. 3-36.
[8] J. I. Marcum, Table of Q Functions, U.S. Air Force Project RAND
Research Memorandum M-339, ASTIA Document AD 1165451, Rand
Corporation, Santa Monica, CA, January 1,1950.
[9] A. H. Nuttall, "Some integrals involving the Q-function,"
Technical Report TR4297, Naval Underwater Systems Center,
Newport, RI, April 17,1972.
[lo] National Bureau of Standards, Tables of the Bivariate Normal
Distribution Function and Related Functions, Appl. Math. Series no. 50,
Washington, D.C.: U. S. Government Printing Office, Superintendant
of Documents, 1959.
[ I l l R. Price, "Some non-central F-distributions expressed in closed
form," Biometrika, vol. 51, issue 1/2, June 1964, pp. 107-122.
[I21 H. A. David, Ordev Statistics, New York: John Wiley & Sons, Inc.,
1981.
[13] B. C. Arnold, N. Balakrishnan, and H. N. Nagaraja, A First Course
in Order Statistics, New York: John Wiley & Sons, Inc., 1992.
140 PROBABILITY DISTRIBUTIONS INVOLVING GAUSSIAN RANDOM VARIABLES

1141 J. Proakis, Digital Communications, New York, NY: McGraw-Hill,

4th ed., 2001.
[15] M. K. Simon and M.-S. Alouini, Digital Communications over
Fading Channels: A Unified Approach to Performance Analysis, New
York, NY: John Wiley & Sons, Inc., 2000.
[16] M.K. Simon and M.-S. Alouini, "On the difference of two chi-
squared variates with application to outage probability computation,"
IEEE Trans. Commun., vol. 49, no. 11, November 2001, pp. 1946-1954.
Also see 38th Annual Allerton Conf. Rec., Monticello, IL, October 2000,
pp. 42-51.
[17] G. Turin, "The characteristic function of Hermetian quadratic
forms in complex normal variables," Biometrika, vol. 47, issue 1/2,
June 1960, pp. 199-201.
[I81 F. S. Weinstein, "Simplified relationships for the probability
distribution of the phase of a sine wave in narrow-band normal
noise," IEEE Trans. Inf. Theory, vol. IT-20, September 1974, pp.
658-661.
[19] R. F. Pawula, S. 0. Rice and J. H. Roberts, "Distribution of the
phase angle between two vectors perturbed by Gaussian noise," IEEE
Trans. Comm., vol. 30, August 1982, pp. 1828-1841.
[20] J. W. Craig, "A new, simple and exact result for calculating the
probability of error for two-dimensional signal constellations," IEEE
MILCOM'SI Con5 Rec., Boston, pp. 25.5.1-25.5.5.
[21] K. Lever, "New derivation of Craig's formula for the Gaussian
probability function," Electron. Lett., vol. 34, September 1998, pp.
1821-1822.
1221 J. I. Marcum and P. Swerling, "Studies of target detection by
pulsed radar," ZEEE Trans. Inf. Theory, vol. IT-6, April 1960.
[23] S. Stein, "The Q function and related integrals," Research Report
No. 467, Applied Research Laboratory, Sylvania Electronic Systems,
Waltham, MA, June 29,1965.
[24] C. W. Helstrom, Elements of Signal Detection and Estimation, Upper
Saddle River, NJ: Prentice Hall, 1995.
[25] M. K. Simon, "A new twist on the Marcum Q-function and its
application," IEEE Commun. Lett., vol. 2, February 1998, pp. 3941.
[26] R. F. Pawula, "Relations between the Rice Ie-function and
Marcum Q-function with applications to error rate calculations,"
Electron. Lett., vol. 31, September 28,1995, pp. 1717-1719.
[27] R. F. Pawula, "A new formula for MDPSK symbol error
probability," IEEE Commun. Lett., vol. 2, October 1998, pp. 271-272.
[28] M. K. Simon, "A simpler form of the Craig representation for the
two-dimensional Gaussian Q-function," IEEE Comm. Lett., vol. 6, no.
2, February 2002, pp. 49-51.
[29] A. H. Nuttall, "Some integrals involving the Q,-function,"
Technical Report TR4755, Naval Underwater Systems Center,
Newport, RI, May 15,1974.
[30] M.K. Simon and M.-S. Alouini, "Some new results for integrals
involving the generalized Marcum Q-function and their application to
performance evaluation over fading channels," to appear in IEEE
Transactions on Wireless Communications, 2002.
[31] M. Chiani, D. Dardari, "Improved exponential bounds and
approximation for the Q-function and related bounds on its inverse,"
to be published.
[32] A. Papoulis, Signal Analysis, New York: McGraw-Hill, Inc., 1977.
[33] M. Chiani, "Integral representation and bounds for Marcum Q-
function," Electron. Lett., March 1999, pp. 445446.
[34] M. K. Simon and M.-S. Alouini, "Exponential-type bounds on the
generalized Marcum Q-function with application to fading channel
error probability analysis," IEEE Trans. Commun., vol. 48, no. 3,
March 2000, pp. 359-366.
 0 1 2 r
3 4 5

Fig. 1. Rayleigh PDFs: (a) n=l, Eq. (2.2);(b) n=2, Eq. (2.5).
Fig. 2. Rayleigh: (a) n=l, Eq. (2.3);(b) n=2, Eq. (2.6).
Fig. 3. Rayleigh Moments: (a) n=l, Eq. (2.4); (b) n=2, Eq. (2.7).
Fig. 4. Rician PDFs: (a) n=l, Eq. (2.14); (b) n=2, Eq. (2.17).
 0 1 2 3 4 5
r

Fig. 5. Rician CDFs: (a) n=l, Eq. (2.15); (b) n=2, Eq. (2.18).
Fig. 6. Rician Moments: (a) n=l, Eq. (2.16);(b) n=2, Eq. (2.19).
Fig. 7. Central Chi-square PDFs: (a) n=2, Eq. (2.32); (b) n=3, Eq. (2.36);
(c) n=4, Eq. (2.32); (d) n=5, Eq. (2.36).
Fig. 7. cont'd.
Fig. 8. Central Chi-square CDFs: (a) n=2, Eq. (2.33); (b) n=4, Eq. (2.33).
Fig. 9. Central Chi-square Moments: (a) n=3, Eq. (2.39); (b) n=4,
Eq. (2.34).
Fig. 10. Noncentral Chi-square PDFs: (a) n=2, Eq. (2.44); (b) n=3,
Eq. (2.48); (c) n=4, Eq. (2.44); (d) n=5, Eq. (2.48).
Fig. 10.cont'd.
Fig. 11.Noncentral Chi-square CDFs: (a) n=2, Eq. (2.45); (b) n=4,
Eq. (2.45).
Fig. 12. Noncentral Chi-square Moments: (a) n=l, Eq. (2.43); (b) n=2,
Eq. (2.47).
Fig. 13. Independent Central Chi-square (-) Central Chi-square PDFs:
(a) n,=n2=l,equal variance, Eq. (4.1); (b) nl=n,=l, unequal variance,
Eq. (4.1);(c) nl=n2=2,equal variance, Eq. (4.4); (d) nl=n2=2,unequal
variance, Eq. (4.4).
Fig. 13. cont'd.
Fig. 14. Independent Central Chi-square (-) Central Chi-square CDFs:
(a) n,=n2=2,equal variance, Eq. (4.5); (b) n1=n2=2,unequal variance,
Eq. (4.5).

159
Fig. 15. Independent Central Chi-square (-) Central Chi-square PDFs:
(a) n,=4, n2=2,equal variance, Eq. (4.13); (b) n,=4, n2=2,unequal
variance, Eq. (4.13).
Fig. 16. Independent Central Chi-square (-) Central Chi-square CDFs:
(a) n1=4, n2=2,equal variance, Eq. (4.14); (b) n1=4,n,=2, unequal
variance, Eq. (4.14).
Fig. 17. Dependent Central Chi-square (-) Central Chi-square PDFs:
(a) n1=n2=1,equal variance, Eq. (4.20); (b) n,=n,=l, unequal variance,
Eq. (4.20); (c) n1=n2=2,equal variance, Eq. (4.23); (d) n,=n,=2, unequal
variance, Eq. (4.23).
Fig. 17. cont'd.
Fig. 18. Dependent Central Chi-square (-) Central Chi-square CDFs:
(a) nl=n,=2, equal variance, Eq. (4.24); (b) nl=n,=2, unequal variance,
Eq. (4.24).
Fig. 19. Independent Noncentral Chi-square (-) Central Chi-square
PDFs: (a) nl=n,=2, equal variance, Eq. (4.32); (b) nl=n,=2, unequal
variance, Eq. (4.32).
Fig. 20. Independent Noncentral Chi-square (-) Central Chi-square
CDFs: (a) nl=n,=2, equal variance, Eq. (4.33); (b) nl=n,=2, unequal
variance, Eq. (4.33).
Fig. 21. Independent Central Chi-square (+) Central Chi-square PDFs:
(a) nl=n2=1,unequal variance, Eq. (5.7); (b) nl=n2=2,unequal variance,
Eq. (5.11).
Fig. 22. Independent Central Chi-square (+) Central Chi-square
CDFs: (a) n,=n,=l, unequal variance, Eq. (5.8);(b) n,=n,=2, unequal
variance, Eq. (5.12).
Fig. 23. Independent Central Chi-square (+) Central Chi-square
Moments: (a) n,=n,=l, unequal variance, Eq. (5.10); (b) nl=n2=2,
unequal variance, Eq. (5.14).
Fig. 24. Dependent Central Chi-square (+) Central Chi-square PDFs:
(a) n,=n,=l, equal variance, Eq. (5.30); (b) n,=n,=l, unequal variance,
Eq. (5.30); (c) nl=n,=2, equal variance, Eq. (5.34); (d) nl=n,=2, unequal
variance, Eq. (5.34).
Fig. 24. cont'd.
Fig. 25. Dependent Central Chi-square (+) Central Chi-square CDFs:
(a) nl=n2=l,equal variance, Eq. (5.31); (b) nl=n2=l,unequal variance,
Eq. (5.31); (c) nl=n,=2, equal variance, Eq. (5.35); (d) n,=n2=2,unequal
variance, Eq. (5.35).
Fig. 25. cont'd.
Fig. 26. Dependent Central Chi-square (+) Central Chi-square
Moments: (a) n,=n,=l, Eq. (5.33); (b) n,=n,=2, Eq. (5.37).
Fig. 27. Independent Noncentral Chi-square (+) Central Chi-square
PDFs: (a) n,=n,=2, Eq. (5.45); (b) n,=4, n,=2, Eq. (5.48).
Fig. 28. Independent Noncentral Chi-square (+) Central Chi-square
CDFs: (a) n,=n2=2,Eq. (5.46); (b) n,=4, n,=2, Eq. (5.49).
Fig. 29. Independent Zero Mean Gaussian (x) Gaussian PDFs:
(a) n=1, Eq. (6.2); (b) n=2, Eq. (6.5).
Fig. 30. Dependent Zero Mean Gaussian (x) Gaussian PDFs:
(a) n=l, Eq. (6.15); (b) n=2, Eq. (6.18).
Fig. 31. Zero Mean Gaussian (x) Gaussian CDFs: (a) Independent,
n=2, Eq. (6.6); (b) Dependent, n=2, Eq. (6.19).
Fig. 32. Independent Rayleigh (x) Rayleigh PDFs: (a) n,=n,=2,
Eq. (6.45);(b) n,=n,=4, Eq. (6.48).
Fig. 33. Dependent Rayleigh (x) Rayleigh PDFs: (a) n=2, Eq. (6.54);
(b) n=4, Eq. (6.54).
Fig. 34. Independent Rayleigh (x) Rayleigh Moments: (a) n,=n,=2,
Eq. (6.47); (b) n,=n,=4, Eq. (6.50).
Fig. 35. Eependent Rayleigh (x) Rayleigh Moments: (a) n=2, Eq. (6.56);
(b) n=4, Eq. (6.56).
Fig. 36. Independent Rice (x) Rayleigh PDFs: (a) n,=n,=2, equal
variance, Eq. (6.59); (b) n,=n,=2, unequal variance, Eq. (6.59).
Fig. 37. Independent Rice (x) Rayleigh CDFs: (a) nl=n,=2, equal
variance, Eq. (6.60); (b) nl=n,=2, unequal variance, Eq. (6.60).
Fig. 38. Independent Rice (x) Rayleigh Moments: (a) nl=n2=2,
Eq. (6.61); (b) n1=n2=4,Eq. (6.64).
Fig. 39. Zero Mean Gaussian (t)Gaussian PDFs: (a) Independent,
Eq. (7.1); (b) Dependent, Eq. (7.9).
Fig. 40. Zero Mean Gaussian (+) Gaussian CDFs: (a) Independent,
Eq. (7.2); (b) Dependent, Eq. (7.10).
 I " ' ' ]

Fig. 41. Independent Zero Mean Gaussian (a)Rayleigh PDFs:

(a) n=2, Eq. (7.19); (b) n=3, Eq. (7.21).
Fig. 42. Independent Zero Mean Gaussian (+) Rayleigh CDFs:
(a) n=2, Eq. (7.20); (b) n=3, Eq. (7.22).

190
Fig. 43. Independent Zero Mean Gaussian (+) Rice PDFs:
(a) n=1, Eq. (7.28);(b) n=2, Eq. (7.30).
Fig. 44. Independent Zero Mean Gaussian (+) Rice CDFs: (a) n=l,
Eq. (7.29); (b) n=2, Eq. (7.31).
Fig. 45. Independent Rayleigh (+) Rayleigh PDFs: (a) n,=n,=l, Eq.
(7.37); (b) n,=n,=2, Eq. (7.44).
Fig. 46. Dependent Rayleigh (+) Rayleigh PDFs: (a) n=1, Eq. (7.56);
(b) n=2, Eq. (7.58).
Fig. 47. Independent Rayleigh (+) Rayleigh CDFs: (a) n,=n2=l,
Eq. (7.38);(b)n,=n2=2, Eq. (7.45).
Fig. 48. Dependent Rayleigh (t)Rayleigh CDFs: (a) n=l, Eq. (7.57);
(b) n=2, Eq. (7.59).
Fig. 49. Independent Rice (+) Rayleigh PDFs: (a) n,=n,=l, Eq. (7.67);
(b) n,=n,=2, Eq. (7.74).
Fig. 50. Independent Rice (+) Rayleigh CDFs: (a) nl=n,=l, Eq. (7.68);
(b) nl=n,=2, Eq. (7.75).
Fig. 51. Independent Rice (+) Rice PDFs: (a) n,=n,=2, equal noncen-
trality parameters, Eq. (7.82); (b) n,=n,=2, unequal noncentrality
parameters, Eq. (7.82).
Fig. 52. Independent Rice (+) Rice CDFs: (a) n,=n,=2, equal noncen-
trality parameters, Eq. (7.83);(b) n,=n,=2, unequal noncentrality
parameters, Eq. (7.83).