You are on page 1of 11

2574

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 11, NO. 7, JULY 2012

Classification of Space-Time Block Codes


Based on Second-Order Cyclostationarity with
Transmission Impairments
Mohamed Marey, Member, IEEE, Octavia A. Dobre, Senior Member, IEEE,
and Robert Inkol, Senior Member, IEEE

AbstractSignal classification is important in various commercial and military applications. Multiple antenna systems
complicate the signal classification problem since there is now
the issue of estimating the number and configuration of transmit
antennas. The novel blind classification algorithm proposed in
this paper exploits the cyclostationarity property of space-time
block codes (STBCs) for the classification of multiple antenna
systems in the presence of possible transmission impairments.
Analytical expressions for the second-order cyclic statistics used
as the basis of the algorithm are derived, and the computational
cost of the proposed algorithm is considered. This algorithm
avoids the need for a priori knowledge of the channel coefficients,
modulation, carrier phase, and timing offsets. Moreover, it does
not need accurate information about the transmission data rate
and carrier frequency offset. Monte Carlo simulation results
demonstrate a good classification performance with low sensitivity to phase noise and channel effects, including frequencyselective fading and Doppler shift.
Index TermsCyclostationarity, signal classification, signal
intelligence and cognitive radio, multiple antenna systems.

I. I NTRODUCTION
IGNAL classification involves the extraction of feature
sets that can be used to distinguish between received signals. Recent developments in signal classification have opened
up various military and commercial applications, including
signal confirmation, interference identification, jamming signal selection, spectrum monitoring, radio surveillance, and
dynamic spectrum access [1][3]. Furthermore, to achieve
robust and spectrally efficient communication over timevarying channels, advanced military and commercial radios
are employing Software-Defined Radio (SDR) concepts [4].
In these systems, the transmitter can optimize transmission
parameters, i.e., modulation type, modulation order, coding
rate, and synchronization parameters, to maximize the channel
capacity utilization while maintaining acceptable error rates
at the receiver. On the receive side, signal classification is

Manuscript received August 6, 2011; revised December 21, 2011 and March
17, 2012; accepted March 20, 2012. The associate editor coordinating the
review of this paper and approving it for publication was M. Morelli.
This work was supported in part by the Defence Research and Development
Canada (DRDC).
M. Marey and O. A. Dobre are with the Faculty of Engineering and Applied
Science, Memorial University, St. Johns, NL, Canada (e-mail: {mfmmarey,
odobre}@mun.ca).
R. Inkol is with Defence Research and Development Canada, Ottawa, ON,
Canada (e-mail: robert.inkol@drdcrddc.gc.ca).
Digital Object Identifier 10.1109/TWC.2012.041612.111488

required for the fast blind estimation of these parameters from


noisy signal data.
Considerable progress has been made on signal classification for single-input single-output (SISO) systems (see e.g.,
[1], [5][8] and references therein). However, the problem
of signal classification for multiple antenna (MA) systems
remains relatively unexplored. Not only does the blind estimation and / or tracking of the signal parameters and the
wireless channel present difficulties, but there is the further
issue of blindly estimating the number of transmit antennas
and the type of transmit side antenna configuration. So far,
the few works that have addressed this problem have focused
on MA systems in conjunction with space-time block codes
(STBCs) [9][14].
The maximum likelihood (ML) approach was recently
proposed for the blind classification of STBCs [9] and the
modulation type for spatial multiplexing (SM) MA systems
[10]. However, this approach suffers from high computational
cost and requires knowledge of channel parameters. Further,
[9] assumes that the first and last intercepted symbols correspond to the start and the end of the space time block,
respectively. In [11], [12] the space-time redundancy of the
received symbols is exploited to discriminate between several
STBCs. This is measured through space-time correlations, and
classification is performed using a binary tree algorithm [11]
or by minimizing the distance between the theoretical and
estimated feature parameters [12]. These investigations all
assume that the receiver is perfectly synchronized with the
transmitter in both time and frequency. Interestingly, fourthorder cyclostationarity with a single receive antenna [13]
and second-order cyclostationarity with two receive antennas
[14] were exploited to distinguish between the Alamouti
and SM codes. However, these investigations relied on the
assumption that the data rate, carrier phase, and frequency
offsets, phase noise, and Doppler shift were precisely known.
In addition, in [14], the receiver requires optimum sampling
time and knowledge of the channel parameters. Since these
constraints are unrealistic in practice, further work is needed
to address the important practical problem of classifying
STBCs affected by transmission impairments. In this paper,
we develop and analyze a STBC classification algorithm
based on the second-order cyclostationarity with two receive
antennas in the presence of transmission impairments over
frequency-flat channels. Further, the feasibility of deploying

c 2012 IEEE
1536-1276/12$31.00 

MAREY et al.: CLASSIFICATION OF SPACE-TIME BLOCK CODES BASED ON SECOND-ORDER CYCLOSTATIONARITY WITH TRANSMISSION . . .

the proposed algorithm over frequency-selective channels is


discussed. This algorithm does not require information about
channel, modulation, and carrier phase and timing offsets.
Moreover, it does not need accurate information about the
transmission data rate and the carrier frequency offset, and is
robust with respect to channel Doppler shift and phase noise.
In addition, analytical expressions for the second-order cyclic
statistics used as the basis of the algorithm are derived, and the
computational cost of the proposed algorithm is considered.
The rest of the paper is organized as follows. Section II
presents the signal model and problem formulation. Secondorder cyclostationarity of STBCs and the proposed algorithm
are introduced in Sections III and IV, respectively. Simulation
results are provided in Section V. Finally, conclusions are
drawn in Section VI.

II. S IGNAL M ODEL


We assume a wireless communication system employing nt
transmit and nr receive antennas. For each pair of transmit and
receive antennas (v, i) with v = 1, , nt and i = 1, , nr ,
the path from the transmit antenna to the receive antenna
is represented by a complex-valued parameter hvi , referred
to as the channel gain. The information-bearing symbols

{xz }z= are drawn from a finite alphabet


 Q and parsed into
(k)

(k)

(k)

blocks of size P . The kth block x(k) = x1 , x2 , , xP


is mapped to the nt transmit antennas
 according to an nt L
code transmission matrix C x(k) . The encoded outputs are
transmitted in L consecutive transmission periods from nt
transmit antennas.
The
STBCs
under
consideration
are
SM1
(nt = 2, P = 2, L = 1), AL code (nt = 2, P = 2, L = 2),
an STBC (ST3) with (nt = 3, P = 3, L = 4), and an STBC
(ST4) with (nt = 4, P = 4, L = 8)2 . The corresponding
transmission matrices are [15], [20]
C(SM)T (x = [x1 , x2 ]) = [x1 x2 ] ,

C(AL)T (x = [x1 , x2 ]) =

x1

x
2
C(ST3)T (x = [x1 , x2 , x3 ]) =

x3
0

x1
x2
x2
x1
0
x3

x2
x1

(1a)


2575

x4

x3

x2

x1
,
C(ST4)T(x = [x1 , x2 , x3 , x4 ])=

x4

x3

x2
x1
(1d)
where (.)T denotes matrix transpose. For convenience of
notation, we drop the superscript k in (1a) - (1d).
We assume that the receiver employs two antennas with the
signal received at the ith antenna, ri (t), being given by

x1
x2
x3
x4
x1
x2
x3
x4

()

x2
x1
x4
x3
x2
x1
x4
x3

x3
x4
x1
x2
x3
x4
x1
x2

ri (t) = yi (t)ej2(fc t+0 +(t)) + ni (t),

(2)

where i {1, 2}, fc is the residual carrier frequency


offset, 0 is the carrier phase offset, (t) is the carrier phase
()
noise, and ni (t) is the noise contribution3. Further, yi (t)
refers to the signal contribution at the ith antenna, with
{SM, AL, ST3, ST4}2 , and is represented as
()

()
yi (t) =

nt

v=1

hvi (t)

()
L


()

Cv,l



()
x(k / L ) pk,l1
T, (t),

k=,k() l=1

(3)
where pk,m
T, (t) refers to p (t (k + m) T + ) with p(t),
T, and  corresponding to the impulse response of the
cascaded transmit and receive filters, the symbol period,
and the timing offset between the transmitter and receiver,
()
respectively. Here, nt and L() are the number of the
transmit
 antennas
 and the length of code, respectively,
()
()
Cv,l x(k / L ) is the entry in the transmission matrix


()
C() x(k / L ) at the junction of the vth row and lth
column, with () as the set of integer multiples
of L() ,

i.e., () = : integer multiples of L() .
The goal of this paper is to develop an algorithm for
classifying the four previously identified linear STBCs using
features extracted from the received signals, r1 (t) and r2 (t).
III. S ECOND -O RDER C YCLOSTATIONARITY OF STBC S

(1b)

x3
0
, (1c)
x1
x2

1 For the sake of presentation, we restrict ourselves to the SM code with


nt = 2. However, it is easy to verify that our analysis is valid for any other
SM codes with nt > 2. No SM code exhibits cyclostationarity, as will be
shown later.
2 Note that the proposed algorithm can be applied to classify STBCs that
have a different code parameter L. For illustration purposes, SM (L = 1),
AL (L = 2), ST3 (L = 4), and ST4 (L = 8) were selected to be the codes
of interest. The first two codes are used in several wireless standards, such as
WiMAX, LTE, and IEEE 802.11n [15][19], and the last two are well known
codes proposed by Tarokh [20].

This section investigates the second-order cyclostationarity


of the STBCs of interest, with the aim of finding discriminating features applicable to their classification. Throughout the
analysis, we make the following assumptions.
AS1) The data symbols are uncorrelated with each other:
E [xz xz ] = 0, E [xz xz ] = s2 zz , xz , xz Q,
where s2 is the transmit signal power, zz is the
Kronecker delta, and E [.] and indicate statistical
expectation and complex conjugate, respectively.
AS2) The transmitted signal is uncorrelated with
()
the noise: E yi (t)ni (t + )
=
0,
3 Note that in (2), we assume that f , , and (t) are all independent
c
0
of the transmit (receive) antenna index v (i). This is a common assumption in
the literature, since a single local oscillator can support all transmit (receive)
antennas on the transmit (receive) side (see, e.g., [21], [22]). Further, 0 can
be included in the channel coefficients, hvi (t), i = 1, 2 and v = 1, , nt .

2576

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 11, NO. 7, JULY 2012

{SM, AL, ST3, ST4}, i, i {1, 2}, and R,


with R as the set of real numbers.
AS3) The noise in each channel is uncorrelated with that of
the other channels: E [ni (t)ni (t + )] = 0, R,
i = i , and i, i {1, 2} .
AS4) The channel gains are unknown deterministic parameters.

where
b(t, )

The basic concepts outlined in the following form the basis


of this work4 . A signal r(t) is said to exhibit second-order
cyclostationarity (CS) if and only if a second-order timevarying correlation function5 is periodic in t [24]. Let us
assume that the second-order correlation function
c(t, ) = E [r(t)r(t + )] ,

(4)

is periodic in t with period T0 for some values of the lag


parameter . Due to this periodicity, c(t, ) can be represented
by Fourier-series as

C(, )ej2t ,
(5)
c(t, ) =

where represents
a cycle frequency for this correlation,


and = : = Tk0 , k integer and C(, ) = 0 . The
function C(, ) is known as the cyclic correlation function
(CCF).
B. SM
Based on (2), (3), and (AS3), it is easily shown that6
c(SM) (t, ) = E [r1 (t)r2 (t + )] = 0,

(6)

for any value of the lag parameter . Therefore, this code does
not have any cycle frequency for this correlation.
C. Alamouti STBC

E [r1 (t)r2 (t + )]
a(t, )b(t, ),

(7)

4 For a more comprehensive treatment of the concepts of time-varying and


cyclic statistics, the reader is referred to the survey of the relevant literature
given in [23][25].
5 The second-order time-varying correlation functions for a complex signal r(t) can be defined as E [r(t)r(t + )], E [r(t)r (t + )],
E [r (t)r(t + )], and E [r (t)r (t + )]. In this work, we use the correlation function E [r(t)r(t + )], as it provides a discriminating feature for
the STBCs of interest.
6 Note that neither E [r (t)r (t + )] nor E [r (t)r (t + )] provide a
1
1
2
2
discriminating feature for STBCs.

k,0
pk,1
T, (t)pT, (t + ),

1
(h11 (t)h22 (t+ )h21 (t)h12 (t+ )) s2 ej(t, ) ,
2T
(9)

(t, ) = 2(2fc )t+2fc +20 +(t)+(t+ ). (10)


An examination of (8) reveals that b(t, ) is a periodic function
in t with period 2T having maxima at = T . To obtain
a closed form expression for the cyclic correlation function
C (AL) (, ), (7) can be rewritten as

0,1
(11)
c(AL) (t, ) = a(t, ) p0,0
T, (t)pT, (t + )



0,0
p0,1
(t)p
(t
+

(t kT ) ,
T,
T,
k=, k(AL)

where denotes convolution. The Fourier transform of (11)


can be expressed as




k
(AL)
C
(, ) = A(, ) B(, )

, (12)
2T
k=

0,1
where B(, ) is the Fourier transform of p0,0
T, (t)pT, (t+ )
0,1
0,0
pT, (t)pT, (t + ), and A(, ) is the Fourier transform of
a(t, ). If A(, ) is independent of , it is apparent that
C (AL) (, ) has cycle frequencies equal to integer multi1
ples of 2T
and A(, ) has only a multiplicative effect on
(AL)
(, ). Insight into the effect of A(, ) on C (AL) (, )
C
can be gained by investigating the influence of the individual
impairments (h(t), fc , 0 , and (t)) on C (AL) (, ). In
addition, the effect of  on C (AL) (, ) is examined, which is
introduced by B(, ). From (9)-(12),

=
=

(8)

and

From (2) and (3), the time-varying correlation function for


the Alamouti STBC can be expressed as
c(AL) (t, )

k,1
pk,0
T, (t)pT, (t + )

k=, k(AL)

where the Fourier coefficients are defined as


T20
1
C(, ) =
c(t, )ej2t dt,
T0 T2 0


k=, k(AL)

a(t, ) =

A. Basic Principles

 and 0 give rise to a phase rotation of C (AL) (, ).


This rotation affects neither the locations of the cycle
frequencies nor the magnitude of C (AL) (, ).
fc has two different effects. First, the cycle frequencies
are shifted by 2fc , resultingin the spectral lines of

k
C (AL) (, ) being located at 2T
+ 2fc , k integer .
Secondly, fc causes a phase rotation of C (AL) (, ).
(t) and h(t) lead to two disruptive effects; one is the
disturbance of the magnitude and phase of the CCF at
the cycle frequencies, and the other is the introduction of
additional frequency components between cycle frequencies. These effects depend on the phase noise bandwidth
and the channel Doppler shift.

From the preceding discussion, it follows that, in the presence


of fc , , and 0 , and the absence of phase noise and
channel
Doppler shift, 
C (AL) (, ) has spectral lines only at
k
2T + 2fc , k integer , with the corresponding magnitudes
being unaffected by these impairments.

MAREY et al.: CLASSIFICATION OF SPACE-TIME BLOCK CODES BASED ON SECOND-ORDER CYCLOSTATIONARITY WITH TRANSMISSION . . .
(1)

D. STBC (nt = 3, P = 3, L = 4), ST3


Based on (2) and (3), the time-varying correlation function
for the ST3 code can be expressed as
c

(ST3)

(t, )

E [r1 (t)r2 (t + )]

=
+

d1 (t, )f1 (t, ) + d2 (t, )f2 (t, )


d3 (t, )f3 (t, ),

(13)

where
1
(h31 (t)h12 (t+ )h11 (t)h32 (t+ )) s2 ej(t, ) ,
4T
(14a)
1
2 j(t, )
d2 (t, ) =
(h11 (t)h22 (t+ )h21 (t)h12 (t+ )) s e
,
4T
(14b)
1
d3 (t, ) =
(h21 (t)h32 (t+ )h31 (t)h22 (t+ )) s2 ej(t, ) ,
4T
(14c)
and (t, ) is defined in (10). Further,

d1 (t, ) =

f1 (t, ) =

k,2
k,2
k,0
pk,0
T, (t)pT, (t+ )pT, (t)pT, (t+ ),

k=, k(ST3)

f2 (t, ) =

(15a)

k,1
k,1
k,0
pk,0
T, (t)pT, (t+ )pT, (t)pT, (t+ ),

k=, k(ST3)

f3 (t, ) =

(15b)

k,3
k,3
k,0
pk,0
T, (t)pT, (t+ )pT, (t)pT, (t+ ).

k=, k(ST3)

(15c)
Following an approach similar to that used in Section III-C,
an examination of (13), (14a - 14c), and (15a - 15c) shows
3
that, if {dq (t, )}q=1 are independent of t, c(ST3) (t, ) is
a periodic function in t with period 4T . In such a case,
C (ST3) (, ) has cycle frequencies equal to integer multiples
1
of 4T
and has peaks at = 3T, 2T, T . A closed
form expression similar to (12) can be derived for the cyclic
correlation function C (ST3) (, ) as






k
(ST3)
C
(, ) = D1 (, ) F1 (, )
4T
k=






k
+ D2 (, ) F2 (, )
4T
k=






k
+ D3 (, ) F3 (, )
4T ,
k=

(16)
where Dq (, ) is the Fourier transform of dq (, ), q =
1, 2, 3, and F1 (, ), F2 (, ), F3 (, ) are the Fourier trans0,2
0,2
0,0
0,0
0,1
forms of p0,0
T, (t)pT, (t+ )pT, (t)pT, (t+ ), pT, (t)pT, (t+
0,0
0,0
0,3
0,3
0,0
)p0,1
T, (t)pT, (t + ), and pT, (t)pT, (t + )pT, (t)pT, (t +
), respectively. The influence of D1 (, ), D2 (, ), and
D3 (, ) on C (ST3) (, ) is similar to the influence of
A(, ) on C (AL) (, ). Therefore, the same conclusion as
in Section III-C can be drawn.
E. STBC (nt = 4, P = 4, L = 8), ST4
Based on (2) and (3), the time-varying correlation function
for the ST4 code can be expressed as in (17), where gq (t, ),

2577

(2)

ueo (t, ), and ueo (t, ) are defined as in (18a-18f), (19a),


and (19b), respectively.
u(1)
eo (t, ) =

k,o
k,o
k,e
pk,e
T, (t)pT, (t+ )pT, (t)pT, (t+ ),

k=, k(ST4)

u(2)
eo (t, ) =

(19a)

k,o
k,o
k,e
pk,e
T, (t)pT, (t+ )+pT, (t)pT, (t+ ).

k=, k(ST4)

(19b)
Using (17) and following an approach similar to that employed in Section III-C, it can be shown that c(ST4) (t, ) is a
periodic function in t with period 8T if gq (t, ) is independent
of t for all possible values of q. Furthermore, in such a case,
C (ST4) (, ) has cycle frequencies equal to integer multiples
1
of 8T
and has peaks at = 7T, 6T, , T . Following
an approach similar to that of (12), a closed form expression
for the cyclic correlation function C (ST4) (, ) is obtained as
(b)
in (20), where Gq (, ) and Ueo (, ) are the Fourier transforms of gq (t, ) and the argument of the sum that corresponds
(b)
(1)
to ueo (t, ) at k = 0, respectively. For illustration, U34 (, )
0,3
0,4
0,4
is the Fourier transform of pT, (t)pT, (t + )pT, (t)p0,3
T, (t +
(ST4)
(, ) is similar to
). The influence of Gq (, ) on the C
the influence of A(, ) on C (AL) (, ). Therefore, the same
conclusion as in Section III-C can be drawn.
A summary of results achieved for the CCF of the investigated STBCs is as follows:
CCF of the SM code does not exhibit peaks ;
CCF of the AL code exhibits peaks at cycle frequencies
1
equal to integer multiples of 2T
and delays T ;
CCF of the ST3 code exhibits peaks at cycle fre1
and delays
quencies equal to integer multiples of 4T
3T, 2T, T ; and
CCF of the ST4 code exhibits peaks at cycle fre1
quencies equal to integer multiples of 8T
and delays
7T, 6T, , T .
IV. P ROPOSED STBC C LASSIFICATION A LGORITHM
As noted in the preceding discussion, the cycle frequencies
corresponding to the CCF differ from one code to another.
We exploit this behavior as a feature in a binary decision tree
algorithm for the classification of STBCs2 . At each node of
the tree, the cyclostationarity test developed in [26] is used
to check whether or not 1 , 2 , , I are cycle frequencies
for selected delays 1 , 2 , , I . Here I and I represent
the number of cycle frequencies and delays to be tested,
respectively. This involves the calculation of a CCF-based
statistic and its comparison against a threshold. Depending on
the choice of the cycle frequencies and delays to be tested
at each node, there are various ways of implementing the
decision tree algorithm. As a result of extensive simulation
experiments, the decision tree algorithm described in the
following text was found to provide satisfactory classification
performance with acceptable computational cost.
Figure 1 shows the flowchart of the proposed algorithm. At
the first node, the ST4 code is distinguished from other STBCs
by exploiting the presence of peaks in the CCF magnitude at

2578

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 11, NO. 7, JULY 2012

c(ST4) (t, )

E [r1 (t)r2 (t + )]




(1)
(1)
(1)
(1)
g1 (t, ) u34 (t, ) + u25 (t, ) + g2 (t, ) u24 (t, ) u35 (t, )


(1)
(1)
(1)
(1)
g3 (t, ) u14 (t, ) + u36 (t, ) u05 (t, ) u27 (t, )


(2)
(2)
(2)
(2)
g4 (t, ) u04 (t, ) + u15 (t, ) + u26 (t, ) + u37 (t, )




(1)
(1)
(1)
(1)
g5 (t, ) u16 (t, ) + u07 (t, ) + g6 (t, ) u06 (t, ) u17 (t, ) ,

(17)

(h31 (t)h22 (t+ )+h41 (t)h12 (t+ )h11 (t)h42 (t+ )h21 (t)h32 (t+ )) s2 ej(t, ) ,

(18a)

(h21 (t)h42 (t+ )h11 (t)h32 (t+ )+h31 (t)h12 (t+ )h41 (t)h22 (t+ )) s2 ej(t, ) ,

(18b)

(h21 (t)h12 (t+ )h11 (t)h22 (t+ )h31 (t)h42 (t+ )+h41 (t)h32 (t+ )) s2 ej(t, ) ,

(18c)

(h11 (t)h12 (t+ )+h21 (t)h22 (t+ )+h31 (t)h32 (t+ )+h41 (t)h42 (t+ )) s2 ej(t, ) ,

(18d)

(h11 (t)h42 (t+ )+h21 (t)h32 (t+ )h31 (t)h22 (t+ )h41 (t)h12 (t+ )) s2 ej(t, ) ,

(18e)

(h11 (t)h32 (t+ )h21 (t)h42 (t+ )h31 (t)h12 (t+ )+h41 (t)h22 (t+ )) s2 ej(t, ) ,

(18f)

=
+
+
+

g1 (t, )

g2 (t, )

g3 (t, )

g4 (t, )

g5 (t, )

g6 (t, )

C (ST4) (, )

1
8T
1
8T
1
8T
1
8T
1
8T
1
8T
=
+
+
+
+
+






k
G1 (, )
+

8T
k=





 
k
(1)
(1)

G2 (, ) U24 (, ) U35 (, )
8T
k=





 
k
(1)
(1)
(1)
(1)

G3 (, ) U14 (, ) + U36 (, ) U05 (, ) U27 (, )
8T
k=





 
k
(2)
(2)
(2)
(2)

G4 (, ) U04 (, ) + U15 (, ) + U26 (, ) + U37 (, )
8T
k=






 
k
(1)
(1)

G5 (, ) U16 (, ) + U07 (, )
8T
k=





 
k
(1)
(1)

G6 (, ) U06 (, ) U17 (, )
.
8T
(1)
U34 (, )

 

(1)
U25 (, )

(20)

k=

1
cycle frequencies (I = 2) and delays 3T, 2T, and
8T
T (I = 6). At the second node, the ST3 code is further
distinguished from the AL and SM codes by exploiting the
1
(I = 2)
presence of peaks in the CCF magnitude at 4T
cycle frequencies and delays 3T, 2T, and T (I = 6).
Finally, the presence of a peak in the CCF magnitude at zero
cycle frequency (I = 1) and delays T (I = 2) is employed
to distinguish the AL code from SM.

Implementation Aspects
1) The proposed algorithm requires a rough knowledge
of the symbol duration7 and carrier frequency offset
7 In practice, the received signals r (t) and r (t) are oversampled for CCF
1
2
estimation. Since the CCF magnitude peaks occur at integer multiples of T
and satisfactory estimates of the CCF magnitude can be obtained for delays
close to T , only a rough knowledge of T is needed. However, a fine search is
still required to obtain accurate estimates for the cycle frequencies and, thus
T.

to support the estimation of the cycle frequencies and


delays to be tested.
2) After down-conversion and filtering, the received signals
, ) can be
are oversampled by a factor . The CCF C(
N 1
estimated from the received samples {r1 (n)}n=0 and
N 1+

as [27]
{r2 (n)}n=0
N
1


, ) = 1
C(
r1 (n)r2 (n + )ej2n
,
N n=0

(21)

where N is the number of received samples, ri (n) =


ri (t = nTs ), Ts = T /, =  /Ts , and
= Ts .
Here  rounds the argument to the nearest integer and
, ) can be efficiently

[0.5, 0.5) . In practice, C(


computed using the fast Fourier transform algorithm
(FFT) of the product r1 (n)r2 (n + ).
k
+2fc , with  =
3) Consider the cycle frequency =
T
2, 4, and 8 for the AL, ST3, ST4 codes, respectively. One

MAREY et al.: CLASSIFICATION OF SPACE-TIME BLOCK CODES BASED ON SECOND-ORDER CYCLOSTATIONARITY WITH TRANSMISSION . . .

2579

where Q1 and Q2 are the two covariance matrices whose


(m1 , m2 )th entries are given respectively by
(S1)/2

Q1 (m1 , m2 ) =

1
NS

W (s)FN,m1

2s
N

s=(S1)/2


FN,m2
+ 2s
,
N

(25)

and
(S1)/2

Q2 (m1 , m2 ) =

1
NS

W (s)FN,
m

s=(S1)/2


+ 2s
,
FN,m2
N

Fig. 1.

Flowchart of the proposed classification algorithm.

where m1 , m2 = 1, , I . Here FN, () =

N
1


2s
N

(26)

r1 (n)r2 (n+

n=0

can observe that


=
+ 2fc Ts is not guaranteed
N
,
with
m = N
to be represented by m
N
2 , ..., 2 1. Such
a representation for
is necessary in order to perform
the FFT algorithm. Further, in practice, the uncertainty
in the estimated frequency offset and data rate results in
an uncertainty range of
. This problem can be avoided
by using a fine search around
, where the candidate
cycle frequency being tested is constrained to values
corresponding to m
N and the CCF function has a peak
within the range of search.
k

Statistical Test for the Presence of Cyclostationarity


A test for the presence of cyclostationarity at given cycle
2, ,
I and delays 1 , 2 , , I is refrequencies
1 ,
quired at all three decision nodes in the proposed algorithm.
We use the statistical test developed in [26]. With this test,
the presence of cyclostationarity is formulated as a binary
hypothesis testing problem:
, ) = (N ) (
H0 :
and
C(
, ),
H1 : For at least one cycle frequency

(22)
and for at least one delay ,
, ) = C(
, ),
C(
, ) + (N ) (
where and are the sets of cycle frequencies and delays
of interest, respectively, and (N ) (
, ) represents the estimation error, which vanishes asymptotically as N . The
cyclostationarity test can be summarized as follows.
) denote a 1 2I vector containing the real and
Let C(
imaginary parts of the estimated CCF at a single tested cycle
frequency
and delays 1 , 2 , , I as



 
) = C(
, I ) ,
, 1 ) , , C(
C(




, 1 ) , , C(
, I ) , (23)
C(
where {.} and
the real and imaginary parts,
 {.} represent

, i ) is given in (21). The 2I 2I
respectively, and C(
) can be computed as [27]
covariance matrix of C(




Q1 Q2
2

Q1 +Q
2
 2  ,
(
) =  Q +Q 
(24)
1
2
2
Q1 Q

2
2

)ej2n , F is the complex conjugate of F , and W () is a


spectral window of length S.
The statistic, , is computed as
=

)1 (
(
N C(
)C
).
T

(27)

has an asymptotic chi-square distribution with degrees of


freedom 2I I under H0 [26]. This is compared against a
threshold to make a decision. If , we decide that at
least one candidate frequency is a cycle frequency at some
of the delays of interest. The threshold is found from the
chi-square distribution tables for a given probability of false
alarm, Pf = Pr ( |H0 ) .
Computational Complexity of the Proposed Classification
Algorithm: We evaluate the computational cost of the proposed classification algorithm as the number of required
floating point operations (flops) [28], with the multiplication
and addition of two complex numbers requiring 6 and 2
flops, respectively. Further, we assume that the N -point FFT
algorithm requires 5N log2 (N ) flops [29]. The approximate
number of flops, Cn , required at each node of the binary tree
algorithm is
Cn = I I (5N log2(N ) + 12(I + 1)S) .

(28)

In comparison, the computational cost of (27) is negligible8 . Obviously, the complexity is proportional to I , I ,
N , and S. Existing technology appears to be adequate for
the practical implementation of the algorithm. For example,
field programmable gate arrays (FPGAs) can readily achieve
several Gigaflops per second [30], [31] provided the potential parallelism is effectively exploited, and highly optimized
FPGA implementations of the FFT are available. If we assume
N = 4096, I = I = 6, and S = 0.06N  246, (28) gives
Cn = 9591264 flops at each node and a computation time on
the order of several milliseconds appears possible in a practical
implementation.
8 Note that the maximum sizes of the matrix (
) are 6 6
) and C(
and 1 6, respectively.

2580

Fig. 2.

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 11, NO. 7, JULY 2012

The magnitude of the estimated CCF for the SM code.

Fig. 3.

V. S IMULATION R ESULTS

1
0.9
0.8
0.7
0.6
PC

The proposed algorithm was evaluated using Monte Carlo


simulations. We considered a pool of SM, AL, ST3, and ST4
STBCs. A raised cosine pulse with 0.35 roll-off factor was
employed for spectral shaping. The oversampling factor was
8.25. A Butterworth low-pass filter of order 13 and bandwidth
equal to the signal bandwidth was used to remove the outof-band noise at the receive side. The SNR was defined as
the ratio of the signal and noise powers at the output of the
receive filter. The covariance estimates in (25) and (26) were
computed using a Kaiser window having a length equal to
6% of the observation interval and shape parameter set to 10.
The time offset  and the carrier phase offset 0 were chosen
to be random variables uniformly distributed over [0, T ) and
[0, 2) , respectively. The frequency offset fc was 0.04
T . Unless otherwise mentioned, the transmitted data was modulated
using quadrature phase shift keying (QPSK) with unit variance
constellation. The number of the observed symbols Ns was
994. Accordingly, the number of the processed samples (FFT
length, NF F T ) was 994 8.25 = 8192. The probability
of false alarm Pf was 0.001. For each transmission link,
the channel was assumed to be frequency non-selective and
modeled as a zero-mean independent complex Gaussian random variable. The channel coefficients were set as constants
during each observation period. The phase noise (t) was
initially set to zero. Scenarios involving non-zero phase noise
and channel Doppler shift and frequency-selective channels
are considered later in this section. The probability of correct
classification, P (|), i.e., the probability that a code was
correctly recognized when present was used as a performance
measure. Each set of simulations was run for 1000 trials.
The CCF magnitudes for the SM and AL codes are shown in
Figs. 2 and 3, respectively. Unlike the SM code, the AL code
exhibits cyclostationarity with CCF peaks occurring at cycle
frequencies of zero and 1
2 . These results are consistent with
the theoretical findings in Sections III-B and III-C. Note that
the finite observation period used for estimation results in nonzero CCF magnitudes appearing at non-cycle frequencies, but
these values become statistically insignificant for a sufficiently
large observation period. The results for ST3 and ST4 codes
also agree with the theoretical findings, but are omitted due
to space considerations.

The magnitude of the estimated CCF for the AL code.

0.5
0.4
4 = {AL,SM}

0.3

3 = {AL,SM,ST3}

0.2

2 = {AL,SM,ST4}
1= {AL,SM,ST3,ST4}

0.1
0
10 8

Fig. 4.
SNR.

4
6
8
SNR (dB)

10

12

14

16

18

20

Average probability of correct classification, PC , as a function of

We used four different sets of STBCs: 1


=
{AL, SM, ST3, ST4}, 2 = {AL, SM, ST4}, 3 =
{AL, SM, ST3}, and 4 = {AL, SM}. Note that we employ
only the necessary nodes which are required for the STBCs
of interest. For illustration, based on Fig. 1, nodes 1, 2, and 3
are employed for the set 1 = {AL, SM, ST3, ST4}; nodes 1
and 3 are employed for the set 2 = {AL, SM, ST4}; nodes
2 and 3 are employed for the set 3 = {AL, SM, ST3}; and
node 3 is employed for the set 4 = {AL, SM}.
Figure 4 shows the average probability of correct
 classification, PC , for the sets of interest, where PC =
P (|). The

proposed algorithm achieves a good performance at reasonably


low values of SNR; at an 8 dB SNR, PC is approaching unity.
Figures 5 and 6 show the probability P (|), for belonging to 1 and 4 , respectively. A significant improvement
with increasing SNR is observed for the AL, ST3, and
ST4 codes. From Figs. 5 and 6, the probability of correctly
classifying SM is independent of SNR. This is because it is
predetermined by the probability of false alarm.
Additional tests were designed to investigate the effects
of transmission parameters on performance. Due to space
limitations, we restrict ourselves to the set of STBCs 4 =

MAREY et al.: CLASSIFICATION OF SPACE-TIME BLOCK CODES BASED ON SECOND-ORDER CYCLOSTATIONARITY WITH TRANSMISSION . . .

0.9

0.9

0.8

0.8

0.7

0.7

0.6

0.6

2581

P(|)

P(|)

P(SM|SM)

0.5
0.4

0.4

0.3

0.3

P(SM|SM)
P(AL|AL)
P(ST3|ST3)
P(ST4|ST4)

0.2
0.1
0
10 8

4
6
8
SNR (dB)

10

12

14

16

18

0.2

0
10 8

20

4
6
8
SNR (dB)

0.8

0.8

0.7

0.7

0.6

0.6
P(|)

0.9

P(AL|AL)

0.4

0.3

0.3

14

16

18

20

P(SM|SM), Pf = 0.01
P(SM|SM), Pf = 0.001
P(SM|SM), Pf = 0.0001
P(SM|SM), Pf = 0.00001
P(AL|AL), Pf = 0.01
P(AL|AL), Pf = 0.001

P(AL|AL)

0.2

0.2

P(SM|SM)

0.1

P(AL|AL)
4

12

P(SM|SM)

0.5

0.4

10

P(SM|SM)
0.9

0
10 8

Fig. 7. Effect of the number of received symbols, Ns , on the probability of


correct classification, P (|), 4 = {AL, SM}.

0.5

P(AL|AL)
Ns = 63, 125, 249, 497, 993, 1490, 1986

0.1

Fig. 5.
The probability of correct classification, P (|), 1 =
{AL, SM, ST3, ST4}.

P(|)

0.5

4
6
8
SNR (dB)

10

12

14

16

18

P(AL|AL), Pf = 0.0001

0.1
20

0
10 8

P(AL|AL), Pf = 0.00001
6

4
6
8
SNR (dB)

10

12

14

16

18

20

Fig. 6.
The probability of correct classification, P (|), 4 =
{AL, SM}.

Fig. 8. Effect of the probability of false alarm, Pf , on the probability of


correct classification, P (|), 4 = {AL, SM}.

{AL, SM}. The same conclusions can be drawn for the other
sets, such as {AL, ST3}, {AL,ST4}, and {ST3, ST4}.
Influence of the number of received symbols Ns
Figure 7 shows the effect of the number of received symbols
Ns on the probability P (|), for belonging to 4 . A strong
improvement in the AL code classification performance results
from increasing Ns at intermediate or even low values of SNR.
This reflects the decrease in the effect of the CCF estimation
, ), with increasing Ns . At low values of Ns
error, (N ) (
(e.g., Ns = 63), the effect of the CCF estimation error dominates, leading to unacceptable classification performance even
at high SNRs. Further, the SM code classification performance
is independent of Ns . This is because the SM code does not
exhibit cyclostationarity. Note that the Ns values were chosen
to keep the FFT length, N , which is used to estimate CCF, to
be a power of 2, where N = Ns .
Influence of the probability of false alarm Pf
Figure 8 presents the effect of the probability of false alarm
Pf on the probability P (|), for belonging to 4 . The

results indicate that the performance dependence of classifying


the AL code on Pf decreases with increasing SNR. The
explanation is that, at higher SNR, the effect of the noise
contribution diminishes and the performance mainly depends
on the estimation error resulting from using a finite observation period. This error vanishes as the observation period
goes to infinity. Generally, decreasing Pf leads to a reduction
in the threshold value, which improves the performance for
the AL code. On the other hand, the converse is true for the
classification of the SM code. Accordingly, the value of Pf is
chosen to achieve a performance compromise for AL and SM
classification.
Influence of bandwidth estimation error
Figure 9 depicts the influence of the bandwidth estimation
error on the probability P (|), for belonging to 4 .
In general, the effect of the bandwidth estimation error is
relatively unimportant since the measurement of the exploited
features is relatively insensitive to such errors. More specifically, as the bandwidth shrinks, the noise contribution slowly

2582

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 11, NO. 7, JULY 2012

1
P(SM|SM)

P(SM|SM)

0.9

0.8

0.8

0.7

0.7

0.6

0.6

P(|)

P(|)

0.9

0.5
0.4

0.2
0.1
0
10 8

0.4

SM, BW/BW = 0.25


SM, BW/BW = 0
SM, BW/BW = 0.25
AL, BW/BW = 0.25
AL, BW/BW = 0
AL, BW/BW = 0.25

0.3

4
6
8
SNR (dB)

10

12

14

16

18

0.3
P(AL|AL)
0.2
0.1
0
10 8

20

Fig. 9. Effect of bandwidth estimation error on the probability of correct


classification, P (|), 4 = {AL, SM}.

0.9

0.8

0.8

0.7

0.7

0.6

0.6

0.3
0.2
0.1
0
6
10

P(SM|SM), T = 0
P(SM|SM), T = 0.0001
P(SM|SM), T = 0.001
P(SM|SM), T = 0.002
P(AL|AL), T = 0
P(AL|AL), T = 0.0001
P(AL|AL), T = 0.001
P(AL|AL), T = 0.002
5

10

P(AL|AL)

4
6
8
SNR (dB)

10

12

14

16

18

20

P(SM|SM)
P(AL|AL)

0.5
P(SM|SM), freqflat channel
0.4

P(AL|AL), freqflat channel

0.3

P(SM|SM), Pedestrian A channel


P(AL|AL), Pedestrian A channel

0.2

P(SM|SM), Vehicular A channel


0.1

10

P(|)

0.9

0.4

1
P(SM|SM)

0.5

Fig. 11. Effect of the modulation scheme on the probability of correct


classification, P (|), 4 = {AL, SM}.

P(|)

AL, QPSK
AL, 16PSK
AL, 16QAM
AL, 64QAM
SM, QPSK
SM, 16PSK
SM, 16QAM
SM, 64QAM

0.5

P(AL|AL)

10

10

10

fdT

0
10 8

P(AL|AL), Vehicular A channel


6

4
6
8
SNR (dB)

10

12

14

16

18

20

Fig. 10. Combined effect of Doppler shift and phase noise on the probability
of correct classification, P (|), 4 = {AL, SM}, SNR = 12 dB.

Fig. 12. Effect of frequency-selective channels on the probability of correct


classification, P (|), 4 = {AL, SM}.

decreases and the cyclostationarity property at zero cycle


frequency is not affected. On the other hand, the increased
noise resulting from an increase in bandwidth only slightly
affects the cyclostationarity at cycle frequency zero. This
explains why, for the AL code, at low or even intermediate
SNRs, the performance slightly improves with normalized
estimation error BW/BW=-0.25, and slightly decreases with
BW/BW=0.25, as shown in Fig. 9. However, at high SNRs,
the noise contribution can be neglected, and hence the classification performance for the AL code becomes independent of
the bandwidth estimation error. In addition, the performance
for classifying the SM code is independent of the bandwidth
estimation error, as shown in Fig. 9. This is because the SM
code does not exhibit cyclostationarity.
Influence of phase noise (PN) and Doppler shift (DS)
In the simulation, the phase noise is modeled as a Wiener
process with rate T , where is the two-sided 3 dB bandwidth of the Lorenzian power density spectrum [32]. Further,
the time selective nature of the channel is modeled through the

modified Jakes model [33]. Figure 10 shows the probability


P (|), 4 as a function of maximum Doppler shift
frequency fd normalized to 1/T , for different phase noise rates
at SNR = 12 dB. It follows that satisfactory performance for
classifying the AL code is possible when T 0.001 and
fd T 0.001. Further, the classification performance for the
SM is unaffected by PN and DS; since cyclostationarity is
absent, PN and DS have no effect.
Influence of modulation scheme
Figure 11 shows the probability P (|), for belonging to
4 , for different modulation schemes, including QPSK, 16PSK, 16-QAM, and 64-QAM. It follows that the performance
of the proposed algorithm does not depend on the modulation
scheme used at the transmitter.
Influence of frequency-selective channels
Although the preceding analysis is based on the assumption of frequency-flat channels, our proposed algorithm is
also robust with respect to frequency-selective channels. For
illustration, let us consider a multipath channel with B in-

MAREY et al.: CLASSIFICATION OF SPACE-TIME BLOCK CODES BASED ON SECOND-ORDER CYCLOSTATIONARITY WITH TRANSMISSION . . .

c(AL) (t, )

=
=

E [r1 (t)r2 (t + )]
B1

  


hb11 (t)hb22 (t + ) hb21 (t)hb12 (t + )
s2 ej(t, )

k


k,0
k,1
k,0
pT, (t b )pT, (t + b ) pk,1
T, (t b )pT, (t + b ) .

b,b =0

dependent propagation taps hvi (t) =

2583

(29)

(AL)

B1


hbvi (t)(t b ),

b=0

b = 0, , B 1, between each pair of transmit and receive antennas (v, i), with v = 1, , nt , i = 1, nr ,
(.) as the Dirac delta function, and hbvi (t) and b as the
attenuation and propagation delay associated with the bth tap,
respectively. After straightforward computations, it follows
that c(SM) (t, ) = E [r1 (t)r2 (t + )] = 0, and c(AL) (t, )
is as provided in (29). An examination of (29) shows that
k,1
k,1
k,0
pk,0
T, (t b )pT, (t + b ) pT, (t b )pT, (t + b )
is a periodic function in t with period 2T . Therefore, in a
static multipath environment and in the absence of the carrier
phase noise, c(AL) (t, ) still provides periodicity in t with
period 2T . In other words, the cyclostationarity feature still
exists in the presence of the multipath nature of time-invariant
wireless channels. Figure 12 shows the results of simulation experiments performed for the ITU-R frequency-selective
pedestrian and vehicular A channels [34] with maximum
Doppler frequencies of 9.72 Hz and 194.4 Hz, respectively.
These results confirm that the algorithm can be successfully
applied to frequency-selective channels.
VI. C ONCLUSIONS
Analysis and simulation experiments show that the cyclic
correlation function of the received signal exhibits peaks
whose corresponding cycle frequencies and time lags depend
on the transmitted STBC. These results are exploited in a
binary decision tree algorithm for the classification of STBCs.
Unlike previously reported classification algorithms, the new
algorithm avoids requirements for estimating channel coefficients, modulation type, carrier phase, and timing offsets.
Furthermore, it does not need accurate information about the
transmission data rate and carrier frequency offset, and is
relatively insensitive to phase noise and channel Doppler shift.
Simulation results show good classification accuracy under
conditions of reasonably low signal to noise ratios, short
observation intervals, and frequency-selective channels. These
favorable results were achieved with acceptable computational
cost.
ACKNOWLEDGMENT
The authors gratefully acknowledge the constructive comments and suggestions from the Editor, as well as the anonymous reviewers.
R EFERENCES
[1] O. A. Dobre, A. Abdi, Y. Bar-Ness, and W. Su, A survey of automatic
modulation classification techniques: classical approaches and new
developments, IET Commun., vol. 1, pp. 137156, Apr. 2007.

[2] T. Yucek and H. Arslan, A survey of spectrum sensing algorithms for


cognitive radio applications, IEEE Commun. Surveys and Tutorials,
vol. 11, no. 1, pp. 116130, First Quarter 2009.
[3] W. Gardner and C. Spooner, Signal interception: performance advantages of cyclic-feature detectors, IEEE Trans. Commun., vol. 40, no. 1,
pp. 149159, Jan. 1992.
[4] T. Ulversoy, Software defined radio: challenges and opportunities,
IEEE Commun. Surveys and Tutorials, vol. 12, no. 4, pp. 531550,
Sep. 2010.
[5] A. Punchihewa, Q. Zhang, O. A. Dobre, C. Spooner, S. Rajan, and R.
Inkol, On the cyclostationarity of OFDM and single carrier linearly
digitally modulated signals in time dispersive channels: theoretical
developments and application, IEEE Trans. Wireless Commun., vol. 9,
no. 8, pp. 25882599, Aug. 2010.
[6] H.-C. Wu, M. Saquib, and Z. Yun, Novel automatic modulation
classification using cumulant features for communications via multipath
channels, IEEE Trans. Wireless Commun., vol. 7, pp. 30983105, Aug.
2008.
[7] W. Headley and C. da Silva, Asynchronous classification of digital
amplitude-phase modulated signals in flat-fading channels, IEEE Trans.
Commun., vol. 59, no. 1, pp. 712, Jan. 2011.
[8] F. Hameed, O. A. Dobre, and D. C. Popescu, On the likelihood-based
approach to modulation classification, IEEE Trans. Wireless Commun.,
vol. 8, no. 12, pp. 58843892, Dec. 2009.
[9] V. Choqueuse, M. Marazin, L. Collin, K. Yao, and G. Burel, Blind
recognition of linear space time block codes: a likelihood-based approach, IEEE Trans. Signal Process., vol. 58, no. 3, pp. 12901299,
Mar. 2010.
[10] V. Choqueuse, S. Azou, K. Yao, L. Collin, and G. Burel, Blind modulation recognition for MIMO systems, Military Technical Academy,
Review MTA, pp. 183195, June 2009.
[11] V. Choqueuse, K. Yao, L. Collin, and G. Burel, Hierarchical spacetime block code recognition using correlation matrices, IEEE Trans.
Wireless Commun., vol. 58, no. 9, pp. 35263534, Sep. 2008.
[12] , Blind recognition of linear space time block codes, in Proc.
2008 IEEE International Conference on Acoustics, Speech and Signal
Processing, pp. 28332836.
[13] M. DeYoung, R. Heath, and B. Evans, Using higher order cyclostationarity to identify space-time block codes, in Proc. 2008 IEEE Global
Telecommunications Conference, pp. 15.
[14] S. Miao, Y. Bar-Ness, and S. Wei, STC and BLAST MIMO modulation recognition, in Proc. 2007 IEEE Global Telecommunications
Conference, pp. 30343039.
[15] J. Andrews, A. Ghosh, and R. Muhamed, Fundamentals of WiMAX:
Understanding Broadband Wireless Networking. Prentice Hall, 2007.
[16] A. Ghosh, J. Zhang, J. Andrews, and R. Muhamed, Fundamentals of
LTE. Prentice Hall, 2010.
[17] Y. Xiao, IEEE 802.11n: enhancements for higher throughput in wireless
LANs, IEEE Trans. Wireless Commun., vol. 12, no. 6, pp. 8291, Apr.
2005.
[18] R. Derryberry, S. Gray, D. Ionescu, G. Mandyam, and B. Raghothaman,
Transmit diversity in 3G CDMA systems, IEEE Commun. Mag.,
vol. 40, no. 4, pp. 6872, Apr. 2002.
[19] H. Ekstrom, A. Furuskar, J. Karlsson, M. Meyer, S. Parkvall, J. Torsner,
and M. Wahlqvist, Technical solutions for the 3G long-term evolution,
IEEE Commun. Mag., vol. 44, no. 3, pp. 3845, Mar. 2006.
[20] V. Tarokh, H. Jafarkhani, and A. R. Calderbank, Space-time block
codes from orthogonal designs, IEEE Trans. Inf. Theory, vol. 9, no. 5,
pp. 14561467, July 1999.
[21] J. Zhang and V. Lau, Carrier frequency synchronization and tracking for
OFDM systems with receive antenna diversity, IEEE Trans. Wireless
Commun., vol. 6, no. 4, pp. 32773286, Sep. 2007.
[22] X. Ma, M. Oh, G. Giannakis, and D. Park, Hopping pilots for
estimation of frequency offset and multiantenna channels in MIMOOFDM, IEEE Trans. Commun., vol. 53, no. 1, pp. 162172, Jan. 2005.

2584

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 11, NO. 7, JULY 2012

[23] E. Serpedin, F. Panduru, I. Sari, and G. Giannakis, Bibliography on


cyclostationarity, Elsevier: Signal Process., vol. 85, no. 12, pp. 2234
2303, Sep. 2005.
[24] W. Gardner and C. Spooner, The cumulant theory of cyclostationary
time-series, part I: foundation, IEEE Trans. Signal Process., vol. 42,
no. 12, pp. 33873408, 1994.
[25] P. Sutton, K. Nolan, and L. Doyle, Cyclostationary signatures in
practical cognitive radio applications, IEEE J. Sel. Areas Commun.,
vol. 26, no. 1, pp. 1324, Jan. 2008.
[26] J. Lunden, V. Koivunen, A. Huttunen, and H. V. Poor, Spectrum sensing
in cognitive radios based on multiple cyclic frequencies, in Proc. 2007
IEEE International Conference on Cognitive Radio Oriented Wireless
Networks and Commun.
[27] A. Dandawate and G. Giannakis, Statistical tests for presence of
cyclostationarity, IEEE Trans. Signal Process., vol. 42, no. 9, pp. 2355
2369, Sep. 1994.
[28] S. Watkins, Fundamentals of Matrix Computations. Wiley, 2002.
[29] S. Johnson, and M. Frigo, A modified split-radix FFT with fewer
arithmetic operations, IEEE Trans. Signal Process., vol. 55, no. 1, pp.
111119, Jan. 2007.
[30] A. Roldao and G. Constantinides, A high throughput FPGA-based
floating point conjugate gradient implementation for dense matrices,
ACM Trans. Reconfigurable Technol. and Syst., vol. 3, no. 1, pp. 119,
Jan. 2010.
[31] Taking advantage of advances in FPGA floating-point IP cores. Available: http://www.altera.com/literature/wp/wp-01116-floating-point.pdf.
[32] A. Demir, A. Mehrotra, and A. Roychowdhury, Phase noise in oscillators: a unifying theory and numerical methods for characterization,
IEEE Trans. Circuits Syst., vol. 47, no. 5, pp. 655674, May 2000.
[33] Y. Zheng and C. Xiao, Simulation models with correct statistical
properties for Rayleigh fading channels, IEEE Trans. Commun., vol. 51,
no. 6, pp. 920928, June 2003.
[34] A. F. Molisch, Wireless Communications, 1st edition. Wiley, 2005.
Mohamed Marey received the M.Sc. and Ph.D.
degrees in Electrical Engineering from Menoufyia
University, Egypt, and Ghent University, Belgium,
in 1999 and 2008, respectively. He worked as an
Assistant Professor at the Department of Electronic
and Communication Engineering, Menoufyia University between 2008 and 2009. Currently, he is a
postdoctoral fellow at the Faculty of Engineering
and Applied Science, Memorial University, Canada.
Dr. Marey received the young scientist award from
International Union of Radio Science (URSI) in
1999. He is the author of the book Multi-Carrier Receivers in the Presence
of Interference: Overlay Systems (VDM Publishing House Ltd., 2009), and
around 40 scientific papers published in international journals and conferences. His main research interests are in wireless communications and digital
signal processing, with a particular focus on cooperative communications and
network coding, signal classification for cognitive radio systems, synchronization and channel estimation, multiple-input multiple-output antenna systems,
multi-carrier systems, and error correcting codes.

Octavia A. Dobre received the Dipl. Ing. and


Ph. D. degrees in Electrical Engineering from the
Polytechnic University of Bucharest (formerly the
Polytechnic Institute of Bucharest), Romania, in
1991 and 2000, respectively. In 2001 she joined the
Wireless Information Systems Engineering Laboratory at Stevens Institute of Technology, US, as a
Fulbright fellow. Between 2002 and 2005, she was
with the Department of Electrical and Computer
Engineering at New Jersey Institute of Technology,
US, as a Research Associate. Since 2005 she has
been with the Faculty of Engineering and Applied Science at Memorial
University, Canada, where she is currently an Associate Professor.
Her research interests include cognitive radio systems, spectrum sensing
techniques, blind signal recognition and parameter estimation techniques,
transceiver optimization algorithms, dynamic spectrum access, cooperative
wireless communications, network coding, resource allocation, underwater
communications, and optical OFDM.
Dr. Dobre is an Editor for the IEEE C OMMUNICATIONS L ETTERS and
IEEE C OMMUNICATIONS S URVEYS AND T UTORIALS , and has served as
a Guest Editor for the IEEE J OURNAL OF S ELECTED T OPICS ON S IGNAL
P ROCESSING, and Lead Guest Editor of the ELSEVIER PHYCOM "Cognitive
Radio: The Road for its Second Decade" special issue. She has also been
the Technical Program Co-chair for the Signal Processing and Multimedia
Symposium of the IEEE Canadian Conference on Electrical and Computer
Engineering in 2009 and for the Signal Processing for Communications
Symposium of the International Conference on Computing, Networking, and
Communications in 2012.
Robert Inkol (DRDC Ottawa) received the B.Sc.
and M.A.Sc. degrees in Applied Physics and Electrical Engineering from the University of Waterloo in
1976 and 1978, respectively. Since 1978 he has been
with Defence Research and Development Canada
where he is currently a senior scientist. He has
also served as an adjunct professor with the Royal
Military College. He is responsible for numerous
contributions to the application of very large scale
integrated circuit technology and digital signal processing techniques to electronic warfare systems. In
addition to having produced numerous publications, Mr. Inkol holds four
patents. He is a senior member of the IEEE, and has served as a reviewer
for various publications and as a Technical Program Committee member for
several IEEE conferences.

You might also like