1050

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 60, NO. 3, MARCH 2012

On the Product of Independent Complex Gaussians

Nicholas ODonoughue, Member, IEEE, and Jos M. F. Moura, Fellow, IEEE

AbstractIn this paper, we derive the joint (amplitude, phase)

distribution of the product of two independent non-zero-mean
Complex Gaussian random variables. We call this new distribution the complex Double Gaussian distribution. This probability
distribution function (PDF) is a doubly infinite summation over
modified Bessel functions of the first and second kind. We analyze
the behavior of this sum and show that the number of terms needed
for accuracy is dependent upon the Rician -factors of the two
input variables. We derive an upper bound on the truncation error
and use this to present an adaptive computational approach that
selects the minimum number of terms required for accuracy. We
also present the PDF for the special case where either one or both
of the input complex Gaussian random variables is zero-mean. We
demonstrate the relevance of our results by deriving the optimal
NeymanPearson detector for a time reversal detection scheme
and computing the receiver operating characteristics through
Monte Carlo simulations, and by computing the symbol error
probability (SEP) for a single-channel
-ary phase-shift-keying
(M-PSK) communication system.
Index TermsComplex double Gaussian, complex Gaussian,
detection, time reversal, probability distribution of product of
Gaussian variables.

I. INTRODUCTION

HIS paper presents the probability distribution function

(PDF) for the product of two non-zero-mean complex
Gaussian random variables. This pdf is not known. We refer to
it as the complex Double Gaussian PDF. This PDF is useful
in many practical applications, for example, in communication
systems, the keyhole or pinhole channel model proposed in [1]
and [2] describes a system where both the transmitter and the
receiver are surrounded by multipath scattering and all communication between them passes through a single waveguide,
such as the corner of a building. In this scenario, the channel is
the product of two complex Gaussian random variables.
Another class of applications where this distribution can
apply is in time reversal detection [3][5]. In time reversal,
a source first probes the channel. The receiver collects this
response, performs time reversal and energy normalization,
and then transmits the time-reversed signal back through the
channel to the original source. If the channel is random, then
the aggregate channel becomes the product of two complex

Manuscript received June 18, 2011; revised October 19, 2011; accepted October 31, 2011. Date of publication November 23, 2011; date of current version
February 10, 2012. The associate editor coordinating the review of this manuscript and approving it for publication was Prof. Joseph Tabrikian. The work of
N. ODonoughue was partially supported by a National Defense Science and
Engineering Graduate Fellowship.
The authors are with the Department of Electrical and Computer Engineering, Carnegie Mellon University, Pittsburgh, PA 15217 USA (e-mail:
nodonoug@ece.cmu.edu; moura@ece.cmu.edu).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TSP.2011.2177264

Gaussian random variables. The optimal design and analysis of

detectors for this class of problems requires knowledge of the
probability distribution function of the product of two complex
Gaussian random variables to derive the likelihood ratio test
and to compute the corresponding rates of detection and false
alarm.
A third example arises when studying the error performance
of -ary phase-shift-keying (MPSK) communication systems.
The linear combiner output for a single-channel system can be
expressed by the product of two complex Gaussian random variables [6]. Knowledge of this distribution allows us to analyze the
symbol error probability (SEP) for this class of communication
system.
This paper addresses the problem of two independent nonzero-mean complex Gaussian random variables
and and
their product . We derive the joint (amplitude and phase) probability distribution of
, where
is the amplitude and
is the principal value of the phase of defined in
the interval
, i.e.,
. We show that
is computed via a doubly-infinite summation, whose terms include modified Bessel functions of the first and second kind [7].
We call this new distribution the complex Double Gaussian distribution. We analyze this result and show the conditions under
which the infinite summation can be truncated; we also show
that the number of terms needed for a high degree of accuracy
depend upon the statistics of the two input variables. We apply
our results to deriving the optimal detector and characterizing
the performance of a time reversal detection scheme, as well
as to perform the error analysis for an M-PSK communication
system.
In Section II, we review the prior work and outline the
contributions of our work. We provide a detailed formulation
of the problem in Section III. The main result is presented
in Section IV, and we provide two example applications in
Section V. Finally, we discuss our results in the conclusion,
Section VI.
Notation: We represent random variables by capital letters,
, while lower-case letters, , denote specific instances of a
random variable. Bold capital letters, , denote matrices, while
lower-case bold letters, , denote vectors. The symbol
describes the parameters and distribution of a random variable.
For example,
states that
is a Gaussian (or
Normal) random variable with mean and variance . The
letter denotes
. The functions
and
are the
modified Bessel functions of the first and second kind, respectively, with order and argument . The Bessel functions are
defined and discussed in great detail in [7].
II. PRIOR WORK
The product of two complex Gaussian random variables is
a problem that has been studied to some degree in the past,

1053-587X/$26.00 2011 IEEE

ODONOUGHUE AND MOURA: ON THE PRODUCT OF INDEPENDENT COMPLEX GAUSSIANS

although none of the prior efforts yielded the PDF for the
problem considered here: the product of two independent complex Gaussian random variables with non-zero expectation.
Recently, Mallik derived the joint (real, imaginary) characteristic function of the inner product of two independent complex
Gaussians random vectors [6]. Inversion of this characteristic
function, however, must be performed numerically, which precludes the analytic derivation of a PDF. The reference handbook
[8], written by Simon, provides solutions for a large selection of
products involving real Gaussian random variables, as well as
Rician, Rayleigh, and chi-squared random variables. Reference
[8] can be used to construct the marginal PDF of the amplitude
of the product of complex Gaussian random variables, but not
the joint PDF of the amplitude and phase.
We derive the joint PDF of two independent complex
Gaussian random variables. We coin the term complex double
Gaussian to refer to this distribution. We also present the
complex double Gaussian PDF for the special cases where one
or both inputs are zero-mean. The applications considered in
Section V, both the design and analysis of the optimal detector
for a time reversal detection scheme and the error analysis
of an M-PSK communication system, serve to illustrate the
applicability of these results.
III. PROBLEM FORMULATION
In this paper, we consider the product
of two independent complex Gaussian random variables
and . We wish
to compute the distribution of . Since is complex, this will
be a bivariate distribution in amplitude
and phase
.
First, however, we review the complex Gaussian distribution [9], [10]. Given a complex Gaussian random variable X,
with non-zero expectation, the amplitude follows a marginal
Rician distribution [8], while the phase , when conditioned on
the amplitude, follows the Tikhonov distribution [11]. For completeness, we provide the following prior results. For the complex Gaussian random variable
with mean value
and
variance , the joint distribution of and is [10]
(1)
The marginal distribution of the amplitude is found by integrating (1) over the interval
, which yields the Rician
distribution [8]
(2)

1051

is similarly defined. The interested reader is referred to [10],

where the complex Gaussian distribution is discussed in great
detail.
For the remainder of this paper, we make use of the Rician
-factor [12], which is used extensively in the analysis of Rician fading channels [13][15]. The -factor is defined as the
ratio of the power between the mean square and the variance.
In Rician fading channels, this corresponds to the ratio of the
power between the line-of-sight and the multipath components.
Here, we use a similar definition:
(5)
Recall that and are the mean and standard deviation, respectively. The -factor modulates the randomness in the underlying distribution. At its lower limit,
, the distribution of
becomes Rayleigh. As increases, the distribution becomes
dominated by the mean value , and the random variable approaches a deterministic value at
.
In the next section, we compute the complex distribution of
in polar format, i.e.,
. We will evaluate
this distribution for one general case and two important special
cases: a) and are non-zero mean; b) is non-zero mean,
and
is zero mean; and c)
and
are both zero mean. In
addition to the complex distribution of , we will also derive
the marginal distribution of its amplitude:
. This result
will be presented for all three scenarios.
IV. DOUBLE GAUSSIAN PDF
In this section, we outline our results and proofs. Our primary contribution is the complex distribution of
. In
Section IV-A, we outline the joint distribution (amplitude and
phase) of for all three scenarios. In Section IV-B, we outline
the marginal distribution of the amplitude
for all three scenarios. When necessary, we discuss the convergence of infinite
summations.
A. Joint Distributions
This section focuses on the joint distribution of the amplitude
and phase of . We discuss the three relevant scenarios: a) both
inputs are non-zero mean; b) one input is zero mean; and c) both
inputs are zero mean.
Theorem 4.1 (Product of two non-zero-mean comand
plex Gaussian variables): If
are independent random variables, then
is characterized by the polar distribution:
the product

where
(3)
is the modified Bessel function of the first kind,
and
with order and argument [7]. The marginal distribution of
the phase, given the amplitude, follows the Tikhonov distribution [11]. This is computed by dividing (1) by (2):
(4)

(6)
where
Proof: See Appendix A.
A number of comments regarding (6):

1052

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 60, NO. 3, MARCH 2012

We call this result the complex Double Gaussian distribution

and use the shorthand
to refer to it:
(7)
It is not immediate from (6) what the first two moments are.
However, since the inputs and are independent, computation of those moments is trivial:
(8)
(9)
Likewise, higher order moments of are obtained from the first
two moments of
and .
Clearly, by symmetry of the roles of and as factors of ,
the expression (6) is symmetric with respect to the parameters
and . The infinite summations over and arise from
of
the Rician -factors and . These terms express the relative
strength of the deterministic components of and , and it is
their presence that leads to the double sum in (6). If one of or
is zero, the corresponding sum across or , respectively, becomes superfluous and the PDF reduces to a single summation
term (see Section IV-A3). Lemma 4.2 shows that the number of
summation terms necessary for a high degree of accuracy is reand .
lated to the values of
We plot an example of (6) in Fig. 1(a) for the case where
, and
. This plot
was generated using the finite sum approximation described in
(12) with
(100 terms). Fig. 1(b) shows a Monte
Carlo simulation of the same scenario, and confirms the validity
of our derived result. The plot exhibits a cardioid shape, this is
because we are plotting the polar form of the two-dimensional
. A simple conversion to Cartesian coordiPDF
nates would remove the hole that appears at the origin. Furthermore, the notch at
in Fig. 1(b) is an artifact from the
method used to compute the 2-D histogram, and can be ignored.
We can see from these plots that, in terms of phase, the energy
, this is a reasonable result, since the
is centered around
and phase is additwo inputs were centered at
tive under multiplication. Within the angular range and ,
the energy is centered between
and
. The long
tail matches the behavior of Rician distributions. We discuss the
accuracy of these plots below.
1) Real-Imaginary Notation: Through a simple transformation of variables, we can express the joint PDF of the real
and imaginary
parts of :

(10)
, and

where,
.

Fig. 1. Plot of the two-dimensional PDF

, described in (6), for
and
. (a) Analytical result. (b) Monte Carlo simgiven in (13) versus
for various
ulation. (c) Plot of the maximum error
summation lengths .

ODONOUGHUE AND MOURA: ON THE PRODUCT OF INDEPENDENT COMPLEX GAUSSIANS

1053

TABLE OF UPPER BOUNDS FOR

Fig. 2. The infinite summation over two variables (n and p) is depicted graphically. The finite summation approximation of (12) is depicted in the lower-left
region, while the four remaining regions represent error terms.

2) Convergence Analysis: We note that (6) is a doubly-infinite summation of modified Bessel functions of the first
and second
kinds [7]. To evaluate this summation, we
truncate the series. For simplicity, we define the general summation term
such that (6) can be written
(11)
Define the partial sum:
(12)
and the error term
(13)
The numbers of computed terms
and in (12) are user se. This
lected parameters. We simplify by stating that
allows us to break the error term into four distinct summation
regions (see Fig. 2) and to rewrite it as .
Lemma IV.2 (Upper Bound on (13)): For sufficiently large
for
number of terms , the truncation error
in (12) is bounded by

TABLE I
, ACCORDING TO (13) WHEN

the upper bound on the error, according to (13). When the error
is guaranteed to be below some desired threshold, the computation is terminated. A set of sample values for (13) is given
in Table I and plotted in Fig. 1(c). Error values below
are
and
are
assumed to be 0. In both Table I and Fig. 1(c),
expressed in decibel scale:
(15)
and are swept across the range
6 dB 6 dB . The -axis
of Fig. 1(c) is given in decibels, while the -axis plots linear
clearly increases as
and
values on a log scale. The error
increase, necessitating more summation terms in order to
maintain a given upper bound on the error . This confirms that
and that cause the infinite summations to arise in (6).
it is
In other words, the stronger the deterministic components of
and are, the more summations terms are needed for accurate
computation.
From Table I, the number of terms needed for an accuracy of
is
(25 terms). Also, using this table, we can
say that each pixel in Fig. 1(a) is accurate to within an error of
.
As described in Appendix B, this error bound is not always
, and the
valid. Depending on the parameters
that is considered, smaller numbers of terms
amplitude
may produce erroneous bounds. This stems from the decay rate
and
,
bounds applied and occurs when the parameters
which are defined in (88) and (91), respectively, are greater than
1. If this error is encountered, then
should be incremented,
and their values retested.
,
3) One or More Inputs Are Zero Mean: If we let
then the joint distribution
in (6) reduces to

(14)
where
is the
th summation term, explicitly defined in
(81),
is the upper bound on the decay rate of
as
increases, defined in (88),
is the upper bound on the decay
as increases, defined in (91), and
is the upper
rate of
along the line
, defined
bound on the decay rate of
in (98).
Proof: See Appendix B.
The error upper bound (14) allows us to iteratively compute (6)
for increasing numbers of terms . At each stage, we compute

(16)
If, in addition to
distribution

, we allow
in (6) reduces to

, then the joint

(17)

1054

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 60, NO. 3, MARCH 2012

B. Marginal Distribution
In addition to the joint distribution
, we present
the marginal distribution of the amplitude
. This
result is useful in many applications, and is a direct extension of
the joint distribution. We present the results for all three target
scenarios and note that prior work corroborates all three of the
derivations in this section. This section is provided for completeness, tying these prior results into the contribution presented in Section IV-A.
Theorem IV.3 (Product of two non-zero-mean complex Gaussian variables): If
and
are independent random variables, then
the amplitude
is characterized by the distribution

(19)
Proof: We first note that
Fig. 3. Plot of the 2-dimensional PDF
, and
.

can be rewritten

, depicted in (16), when

These results can be quickly verified by first noting the

approximation to a modified Bessel function of the first kind
for small arguments
and non-negative integer orders
[7, Sec. (10.30.1)]:

(20)
We can, thus, begin with the joint distribution of
(61) and apply the transformation

and

in

(21)
which results in the distribution

(18)
(22)
for (16) and then
and then taking the limit of (6) first as
again as
for (17).
We first note that, while the general joint PDF in (6) is a
doubly-infinite summation, this simplifies to an infinite summation over one term when one of the inputs is zero mean, as in
(16), and reduces further to a closed form solution when both
inputs are zero mean, as in (17).
We note that both PDFs are independent of
. Thus,
is uniformly distributed and is independent of
in both scenarios. This result is intuitive, since the input is uniformly distributed across phase (as well as , in the second case). Thus,
when
and
are added, the result is also uniformly distributed over
. We plot an example of (16) in Fig. 3 for
the case where
, and
. The plot
shows circular symmetry about the origin, caused by the independence of
. This is in stark contrast to Fig. 1(a), where a
strong dependence on
is seen. Further, we note that the energy in Fig. 3 is concentrated between
and
,
as opposed to the window of
to
for Fig. 1(a).
The difference comes from the fact that
was reduced from 1
to 0. A reduction of
from 1 to 0 (as seen in the case where
both and are 0) would further reduce the radius of the expectation.

Next, we integrate over

, recall the identity in
(70) and convert both of the modified Bessel functions to their
infinite summation representations

(23)
We follow the same approach as in (74) and arrive at the solution

(24)

This result appears in [8] under the product of independent Rician random variables. See Fig. 4(a) for a plot of (19) for various
values of , when
is fixed to 1. The solid lines correspond
to the analytical results computed from (19), while the dotted
lines follow a weighted histogram from Monte Carlo trials. For

ODONOUGHUE AND MOURA: ON THE PRODUCT OF INDEPENDENT COMPLEX GAUSSIANS

TABLE II
TABLE OF REPRESENTATIVE VALUES FOR

1055

, ACCORDING TO (27) WHEN

We approximate (19) with the partial sum

(26)
and define the truncation error
(27)
Lemma IV.4 (Upper Bound on (27)): For sufficiently large
for
in (26)
number of terms , the truncation error
is upper bounded by

(28)
where the summation term is defined as
(29)
Fig. 4. (a) Plot of the probability distribution
computed using (19) for
is set to 1. These plots are compared against the
various values of , when
. The solid lines plot the analytical results, while
Rician distribution
the dotted lines follow a weighted histogram from Monte Carlo trials used to
given in (27) versus
,
verify our results. (b) Plot of the maximum error
.
for various number of terms

and the various decay rate bounds are defined as

(30)
(31)

this test,
. As the Rician -factor
increases,
the PDF approaches the Rician distribution
. This trend
makes sense. As
approaches , the variance
approaches
0 (since
is held constant at 1). The result is that the random
variable becomes less random and starts to behave as the deterministic variable
. In this limiting case
.
1) Convergence Analysis: The result (19) is an infinite summation over the terms
. In order to evaluate it,
we truncate the summations. We define the general term of the
summation in
, such that (19) can be written
(25)

(32)
For brevity, we have used the term
the definitions above to signify the quantity

and

in

(33)
Proof: Repeat the steps taken in Appendix B, replacing (6)
with (19).
A set of sample values for (27) is given in Table II and Fig. 4(b).
We see from the table that
(225 terms) are required for

1056

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 60, NO. 3, MARCH 2012

A. Time Reversal Detection

For our first example, we consider the detection of a target in
the presence of clutter using a single transmitter and a single receiver. We will utilize a time reversal detection strategy similar
to the ones outlined in [3][5] with one notable difference. In the
detection system we consider here, the time reversal mirror will
not communicate with the detector. Thus, the detection system
must operate without knowledge of the result of the forward
transmission. The meaning of this statement will be clarified
below.
We define the frequency samples
. We
model the target as a point target with a deterministic response
. The clutter is drawn from a zero-mean complex Gaussian
distribution with power spectral density
. This channel
model is discussed in detail in [16]. We transmit the probing
signal
and write the response
(36)
Fig. 5. Plot of the probability distribution
computed using (34) for
when
is set to 0. These plots are compared against the
various values of
. The solid lines plot the analytical results, while
Rayleigh distribution
the dotted lines follow a weighted histogram from Monte Carlo trials used to
verify our results.

the error bound

. This is much larger than in Lemma
IV.2, which required only 25 terms. This increase is because
the error here represents the integral of the error in Lemma IV.2
across phase.
2) One or More Inputs Are Zero Mean: If we let
,
then the marginal distribution
in (19) reduces to

is the clutter response, and

represents additive noise. For simplicity,
we use a white probing signal
where

(37)
for some transmit power . The time reversal probing signal
is generated using a scaled, phase-conjugated version of the received signal
(38)
where

(39)

(34)
If, in addition, we let
in (19) reduces to

, then the marginal distribution

(35)
These results appear in [8], the former under the product of independent Rician and Rayleigh random variables and the latter
under the product of independent Rayleigh random variables.
These equations can be verified by using the same approach as
in Section IV-A3.
See Fig. 5 for a plot of (34) for various values of . For this
test,
. As the Rician -factor
increases, the pdf tends
towards the Rayleigh distribution
.
V. APPLICATIONS
In this section, we present two example applications. The
first, in Section V-A, is a blind time reversal detection strategy
where the complex Double Gaussian distribution is used to
derive the likelihood ratio test and NeymanPearson detector.
The second, in Section V-B, is an
-ary phase-shift-keying
(M-PSK) communication system where the complex Double
Gaussian distribution is used to derive the SEP.

is the energy normalization factor defined by

We assume that is approximately deterministic, as was argued

in [17]. The time reversal probing signal
is distributed
as a complex Gaussian:
(40)
is then transmitted from the receiver back
The signal
to the source, where the received signal is
(41)
where
is the clutter channel for the second transmission,
and
is the noise signal for the second transmission. We
assume that the clutter and noise signals are independent of each
other and independent from one transmission to the next. If we
ignore the noise term, then (41) is distributed according to the
product of independent complex Gaussians:
(42)
, and
. At this stage, we set up the binary hypothesis test. The detectors in [3][5] were designed to use both
where

ODONOUGHUE AND MOURA: ON THE PRODUCT OF INDEPENDENT COMPLEX GAUSSIANS

1057

and
. In this application, however, we consider
the case where
is not available to the detector. In the null
hypothesis, the case where no target is present,
. In the
alternative hypothesis,
. Thus

(43)
1) Likelihood Ratio Test: The received signal
follows the complex double Gaussian distribution presented
in Section IV. Under
is distributed according
to the special case (17). Under
is distributed
according to the general result in Theorem IV.1. To compute
the likelihood ratio, we divide (6) by (17):

(44)

, and
. It would be possible to
construct the likelihood ratio numerically, through a Monte
Carlo simulation, but that approach would be much more
computationally complex, and would not yield the analytical
representation in (44). From the likelihood ratio , we can
construct the test [18]
where

(45)
for some threshold . Ideally, we would use the distribution
of the test statistic to determine the appropriate threshold
for some desired false alarm rate. However, its distribution is
unknown, so we must rely on Monte Carlo simulations to determine the appropriate threshold.
2) Monte Carlo Simulations: We simulated a scenario where
the clutter channel follows a Gaussian power spectral density
in the band 24 GHz, and the target (a point target) has a constant value
. We confine our transmit power to the
same level
and vary the total clutter power
. We sample
frequencies and conducted
Monte Carlo trials for each scenario.
Using these Monte Carlo trials, we were able to compute the
receiver operating characteristics for the detector presented in
(45), for varying levels of signal-to-noise ratio, defined as
SNR
We also define the signal-to-clutter ratio SCR
SCR

(46)
:
(47)

Fig. 6. Receiver operating characteristics for the likelihood ratio test (LRT)
given in (44) and the energy detector (ED) given in (48) for various SNR values
5 dB and (b) SCR 0 dB.
when (a) SCR

For comparison, we also computed the receiver operating characteristics of an energy detector (ED):
(48)
and plot the reFor the first test, we set the SCR
ceiver operating characteristics (
versus
) for both the
likelihood ratio test (LRT) and the energy detector (ED) for
SNR
and 10. The results are shown in Fig. 6(a). For
the second test, we set the SCR
and repeated the first
test. The results are shown in Fig. 6(b). In both plots, the solid
lines depict the results for the case where SNR
, and the
dotted lines depict the results for the case where SNR
. It

1058

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 60, NO. 3, MARCH 2012

is clear from these results that the LRT derived from the distributions we have presented in this paper outperforms the ED in
this application.
B. M-Ary Phase Shift Keying Communications
A second potential application is the error analysis for
M-PSK communication systems (a system where the
constellation points are equally spread about the unit circle). We
consider a system similar to the one analyzed in [6], with the
exception that we limit the number of diversity branches to
and, since there is only one receiver, do not consider
Maximal Ratio Combining. Given the transmission of a symbol
, we write the received signal as
(49)
where

is the channels complex random gain and

is an additive noise term. The symbol belongs
to an M-PSK constellation , given by
,
where
(50)

and
is the transmit energy. In order to estimate the channel
gain, we first transmit pilot symbols:
(51)
and write the received signal vector
(52)
with the IID noise vector
estimate of is given by

. The least-squares

(53)
pilot symbols each data symbol
is transFollowing the
mitted, and the received data signal
is run through a linear
combiner
(54)
This correlation output is distributed according to the complex
Double Gaussian distribution:
(55)
We can compute the SEP with [6]
(56)
where the product

is distributed
. We compute this probability

with the integral

(57)

Fig. 7. Symbol Error Probability against SNR for an M-PSK Communication

(Binary PSK),
(quadrature PSK), and
.
System with
Solid lines show analytical results computed from the complex Double Gaussian
distribution, while dotted lines show average error results from Monte Carlo
simulations.

These analytical results are shown as the solid lines in Fig. 7,

which plots the probability of error against SNR. To verify these
analytical results, we compute the error probability term in (57)
by simulating the transmission of
pilot symbols and
100 data symbols through a channel
. This experiment is repeated for
independent Monte Carlo trials.
These are shown as the open circles in Fig. 7. We define the
SNR:
SNR

(58)

in (57), we evaluate the PDF

To compute the integral
over a grid of 50 range bins and 100 phase
intervals and numerically integrate over the region of interest.
At each range bin, we use the error bound in (14) to determine
the minimum number of terms
in order to guarantee a
truncation error of
. There is a strong correlation
between the Monte Carlo (circular marks) and analytical (solid
lines) results.
In order to study the effects of the number of terms
on the
analytical computation of symbol error probability, we repeated
the simulation with
25, 100, 225, 400, and 625 terms.
The results are shown in Fig. 8 for the case where the number
of constellation points is
. It is clear from Fig. 8 that,
for the range of SNR and parameters considered,
is sufficient for high accuracy. However, if we are interested
in higher noise scenarios, then fewer terms are required. Fig. 8
also shows that, for example, at SNR
terms
is sufficient. This dependence on the parameters illustrates the
value of the error bound in (14), as it can be used to determine
the appropriate number of terms needed adaptively.

ODONOUGHUE AND MOURA: ON THE PRODUCT OF INDEPENDENT COMPLEX GAUSSIANS

1059

Recall that the desired distribution is

. As
an intermediate step, we will compute the joint distribution
, using Bayes theorem [19]:
(60)
and then perform the transformation of random variables
. Since
and
are independent, the joint distribution
is simply the product of two Rician
distributions:
(61)
In order to compute the distribution of
conditioned on
and , we convolve two Tikhonov distributions circularly over
the interval
[19]:

Fig. 8. Symbol Error Probability versus SNR for an M-PSK Communication

constellation points. Solid lines show analytical results
System with
computed from the complex Double Gaussian distribution with varying number
, while dotted lines show average error
of terms
results from Monte Carlo simulations.

(62)
We make the substitution
(63)

VI. DISCUSSION AND CONCLUSION

We derived a new distribution, which we refer to as the complex double Gaussian distribution, to describe the product of two
independent complex Gaussian random variables. We also derived the special cases where one or more of the inputs is zero
mean. We analyzed the convergence of this result, which contains a doubly infinite summation, and derived an upper bound
on the truncation error. We showed that the number of terms
needed for accurate results varies with the Rician -factors (the
ratio of mean-squared to variance) for both inputs. We adaptively compute the PDF by specifying a desired accuracy and
use the error bound to adaptively determine how many terms
are needed. We applied this result to derive the optimal filter,
in a NeymanPearson sense, for a time reversal based detection
system, wherein the received signal is distributed according to
the complex Double Gaussian distribution and verified that the
detector outperforms the standard energy detector for that scenario. We also presented a single-channel MPSK communication system for which the linear combiner output follows the
complex Double Gaussian distribution. Using this fact, we derived the SEP and verified our results using Monte Carlo trials.
This distribution is a novel result and is useful in a wide array
of applications.

(64)
where
(65)
The integral solution for (64) can be found in [20][22]. We
substitute (61) and (64) into (60):
(66)
Next, we define the nonlinear transformation
(67)
which results in the distribution
(68)
noting that is now a function of . We integrate out the dummy
variable :

APPENDIX A
PROOF OF THEOREM 4.1

(69)

To begin, we recall the polar representations

and
. The product
is written

We first note the identity [22, Sec., (8.445)]

(59)

where
the principal value of

, and
, defined over the interval

is
.

(70)

1060

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 60, NO. 3, MARCH 2012

for integer values . We recall the definition for in (65) and

define
and
according to (3). Thus, the modified Bessel
function
can be written

Now, we reorder the summation indices to place

and isolate all of the terms dependent upon :

on the inside,

(71)
(78)

We make use of the trinomial expansion, which can be easily

derived from the binomial expansion:
(72)

We recall the Bessel function identity presented in (70); this

leads to the solution

Thus

(73)
We insert the expanded form of

into (69):

(79)
where

is defined by

(80)

(74)
If we make the substitution
gral portion of (74) becomes

, then the inte-

APPENDIX B
PROOF OF LEMMA 4.2
We recall the definition of the summation term in (11) and the
PDF in (6). From those equations,
can be written

(75)
From the hyperbolic cosine identity [22, Sec. (1.311.3)] and the
integral solution [22, Sec. (3.337.1)], (75) reduces to

(81)
We also recall the truncation error in (13); we can thus write
with

(76)
where
with order
reduces to

is the modified Bessel function of the second kind

and argument [7]. From (76), we see that (74)

(82)
(77)

for large

(83)

ODONOUGHUE AND MOURA: ON THE PRODUCT OF INDEPENDENT COMPLEX GAUSSIANS

We assume that the number of terms

is sufficiently large so
that for all four error regions (see Fig. 2), we can apply the
approximation [7]. Thus,

1061

B. Error Region 2
We turn to error region 2 (see Fig. 2) and note that its structure is similar, so the approach will be the same. We begin by
, this time as
defining the ratio between successive terms of
increases:

(84)

A. Error Region 1
We turn first to Error Region 1 (see Fig. 2), and wish to show
that the infinite summation over is upper bounded by the geometric series:

(90)
We apply (87) to define an upper bound

, where

(85)
In order to define
terms of
, as

, we look at the ratio between successive

increases

(91)
(the simThis equation is monotonically decreasing with
plification yields a polynomial with principal order of
),
thus
. Thus, we have the upper bound for error
region 2:
(92)

(86)
Note that we negated the order of the modified Bessel function of the second kind. This operation is permitted because
[7]. For the ratio of modified Bessel functions, we cite the inequality [23, Sec. (1.13)]:
(87)
This can be used to create a relaxation
which is defined as

such that

As with

, this bound is only valid when

C. Error Region 3
For error region 3 (see Fig. 2), we have the bounds
and
. The inner summation is similar to that
in error region 1. Thus, we begin with (89), and adjust the summation limits:

(93)
Recalling their definitions in (84) and (88), we have
(94)

(88)
This equation is monotonically decreasing with (simplification yields a polynomial with principal order of
). Thus, we
can provide an upper bound with the smallest value of in the
series,
; thus:
. This leads to the upper bound
given in (85), as long as
In Table I, we demonstrate
that this is not always true for small N. However, the values for
which this is not valid are also the values of
for which the
error bound is large. Therefore, whenever
, we can
say that
should be incremented. Inserting (85) into the full
summation yields the bound:
(89)

(95)
We define
mation, as

as the ratio between successive terms of the sumincreases:

(96)

We note that
is a monotonically decreasing function for
large (as increases, the ratio
approaches 1
and becomes a noncontributing factor). For this reason, we can
say that
(97)

1062

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 60, NO. 3, MARCH 2012

This is a monotonically decreasing function of

upper-bounded by its limiting value (at
):

This allows us to define the upper bound

(98)
For large , the ratio
write

is approximately 1, thus, we can

(109)
We can see that
. This leads to the upper bound on the
summation for error region 3:

(99)
This is a monotonically decreasing function of and can thus be
upper-bounded by its limiting value (at
):
. This
leads to the upper bound on the summation for error region 3:
(100)

and can be

(110)

E. Synthesis
We take the bounds calculated in (89), (92), (100), and (110),
and insert them into the error term defined in (13), to obtain the
upper bound on the error

Once again, this bound is only valid when

, which we
can easily show to be true for all that satisfy the inequality
(101)
(111)

D. Error Region 4
Finally, we turn to error region 4 (see Fig. 2) and note its
similarity to error region 2. We start with (92) and adjust the
summation limits
(102)
Recalling their definitions in (84) and (91), we have
(103)
(104)
We define the ratio
(105)
is a monotonically decreasing function for
We note that
large (as increases, the ratio
approaches 1
and becomes a noncontributing factor). For this reason, we can
say that
(106)
This allows us to place an upper bound on the ratio
(107)
For large , the ratio

is approximately 1; we can write

(108)

REFERENCES
[1] D. Chizhik, G. Foschini, M. Gans, and R. Valenzuela, Keyholes,
correlations, and capacities of multielement transmit and receive
antennas, IEEE Trans. Wireless Commun., vol. 1, no. 2, pp. 361368,
Apr. 2002.
[2] D. Gesbert, H. Bolcskei, D. Gore, and A. Paulraj, Outdoor MIMO
wireless channels: Models and performance prediction, IEEE Trans.
Commun., vol. 50, no. 12, pp. 19261934, Dec. 2002.
[3] J. M. F. Moura and Y. Jin, Detection by time reversal: Single antenna,
IEEE Trans. Signal Process., vol. 55, no. 1, pp. 187201, Jan. 2007.
[4] Y. Jin and J. M. F. Moura, Time-reversal detection using antenna arrays, IEEE Trans. Signal Process., vol. 57, no. 4, pp. 13961414, Apr.
2009.
[5] N. ODonoughue, J. Harley, and J. M. F. Moura, Detection of targets embedded in multipath clutter with time reversal, in Proc. IEEE
Int. Conf. Acoust., Speech, Signal Process. (ICASSP), May 2011, pp.
38683871.
[6] R. Mallik, Distribution of inner product of two complex Gaussian
vectors and its application to MPSK performance, in Proc. IEEE Int.
Conf. Commun. (ICC), May 2008, pp. 46164620.
[7] NIST Handbook of Mathematical Functions, F. W. Oliver, D. W.
Lozier, R. F. Boisvert, and C. W. Clark, Eds. Cambridge, U.K.:
Cambridge Univ. Press, 2010.
[8] M. Simon, Probability Distributions Involving Gaussian Random Variables. Norwell, MA: Kluwer Int., 2002.
[9] B. Picinbono, Second-order complex random vectors and normal
distributions, IEEE Trans. Signal Process., vol. 44, no. 10, pp.
26372640, Oct. 1996.
[10] P. J. Schreier and L. L. Scharf, Statistical Signal Processing of Complex-Valued Data. Cambridge, U.K.: Cambridge Univ. Press, 2010.
[11] H. Fu and P. Y. Kam, Exact phase noise model and its application
to linear minimum variance estimation of frequency and phase of a
noisy sinusoid, in Proc. IEEE 19th Int. Symp. Personal, Indoor, Mobile Radio Commun. (PIMRC), Sep. 2008, pp. 15.
[12] S. O. Rice, Statistical properties of a sine wave plus random noise,
Bell Syst. Tech. J., vol. 27, no. 1, pp. 107157, Jan. 1948.
[13] P. Soma, D. Baum, V. Erceg, R. Krishnamoorthy, and A. Paulraj,
Analysis and modeling of multiple-input multiple-output (MIMO)
radio channel based on outdoor measurements conducted at 2.5 GHz
for fixed BWA applications, in Proc. IEEE Int. Conf. Commun.
(ICC), 28 Apr.May 2, 2002, vol. 1, pp. 272276.

ODONOUGHUE AND MOURA: ON THE PRODUCT OF INDEPENDENT COMPLEX GAUSSIANS

[14] C. Lemoine, E. Amador, and P. Besnier, On the k-factor estimation

for Rician channel simulated in reverberation chamber, IEEE Trans.
Antennas Propag., vol. 59, no. 3, pp. 10031012, Mar. 2011.
[15] M. Hashim and S. Stavrou, Wind influence on radio waves propagating through vegetation at 1.8 GHz, IEEE Antennas Wireless
Propag. Lett., vol. 4, no. 1, pp. 143146, Jan. 2005.
[16] S. Kay, Optimal signal design for detection of Gaussian point targets in stationary Gaussian clutter/reverberation, IEEE J. Sel. Topics
Signal Process., vol. 1, no. 1, pp. 3141, Jun. 2007.
[17] Y. Jin, J. M. F. Moura, and N. ODonoughue, Time reversal in multiple-input multiple-output radar, IEEE J. Sel. Topics Signal Process.,
vol. 4, no. 1, pp. 210225, Feb. 2010.
[18] L. L. Scharf, Statistical Signal Processing. Reading, MA: AddisonWesley, 1991.
[19] A. Papoulis and S. Pillai, Probability, Random Variables and Stochastic
Processes, 4th ed. New York: McGraw-Hill, 2002.
[20] H. Van Trees, Detection, Estimation, and Modulation Theory, Part I:
Detection, Estimation, and Linear Modulation Theory. New York:
Wiley-Interscience, 1968.
[21] T. T. Kadota, Optimum reception of M-ary Gaussian signals in
Gaussian noise, Bell Syst. Tech. J., vol. 44, no. 9, pp. 21872197,
Nov. 1965.
[22] I. Gradshteyn and I. Ryzhik, Tables of Integrals, Series, and Products. New York: Academic, 1980.
[23] A. Laforgia and P. Natalini, Some inequalities for modified Bessel
functions, J. Inequalities Appl., vol. 2010, p. 10, 2010.

Nicholas ODonoughue (S03M12) received the

B.S. degree in computer engineering from Villanova
University, Villanova, PA, in 2006 and the M.S. and
Ph.D. degrees in electrical and computer engineering
from Carnegie Mellon University, Pittsburgh, PA, in
2008 and 2011, respectively.
His general research interests lie in the area of statistical signal processing, with applications in radar/
sonar and communications and a special interest in
the technique of time reversal signal processing.
Dr. ODonoughue is a recipient of the 2006 National Defense Science and Engineering Graduate (NDSEG) Fellowship, the
2006 Dean Robert D. Lynch Award from the Villanova University Engineering
Alumni Society, and the 2006 Computer Engineering Outstanding Student
Medallion from Villanova University. He has published more than 20 technical
journal and conference papers, including two that were chosen as Best Student
Paper. He is a member of several IEEE societies, Tau Beta Pi, and Eta Kappa Nu.

1063

Jos M. F. Moura (S71M75SM90F94)

received the E.E. degree from Instituto Superior
Tcnico (IST), Lisbon, Portugal, and the M.Sc., E.E.,
and D.Sc. degrees in Electrical Engineering and
Computer Science from the Massachusetts Institute
of Technology (MIT), Cambridge.
He is University Professor at Carnegie Mellon
University (CMU), Pittsburgh, PA, with the Departments of Electrical and Computer Engineering
and, by courtesy, BioMedical Engineering. He was
a Professor of IST and has been a visiting professor
at MIT. He is a founding co-director of the Center for Sensed Critical Infrastructures Research (CenSCIR) and manages a large education and research
program between CMU and Portugal, www.icti.cmu.edu. His research interests
include statistical and algebraic signal and image processing. He has published
over 450 technical journal and conference papers and has eight issued patents
with the USPTO.
Dr. Moura is the IEEE Division IX Director Elect (2011), to serve as
IEEE Division IX Director in 20122013. He was President of the IEEE
Signal Processing Society (SPS) (20082009), Editor-in-Chief for the IEEE
TRANSACTIONS IN SIGNAL PROCESSING, and was on the Editorial Board
of several other journals and steering and technical committees of several
conferences. He is a Fellow of the American Association for the Advancement
of Science (AAAS) and a corresponding member of the Academy of Sciences
of Portugal. He received several awards, including the IEEE SPS Technical
Achievement Award, the IEEE SPS Meritorious Service Award, the IEEE
Millennium Medal, the CMUs College of Engineering Outstanding Research
Award, and the Philip L. Dowd Fellowship Award for Contributions to Engineering Education. In 2010, he was elected University Professor at CMU.