Documente Academic
Documente Profesional
Documente Cultură
s
()
+ n, (1)
where r = [
1
=1
are
independent circular complex zero-mean Gaussian random
variables, with real and imaginary parts of variance /2. The
sequence {
()
=1
is independent and identically distributed,
with values drawn from an alphabet specic to the modula-
tion . Without loss of generality, we consider unit variance
constellations and dene the signal-to-noise ratio (SNR) as
=
2
/.
III. LB MODULATION CLASSIFICATION ALGORITHMS
A. The LB Approach to Modulation Classication
Under the assumption of statistically independent received
symbols, the likelihood function is given by
(r{
()
=1
, ) =
=
=1
()
1
exp
{
()
2
}
= ()
exp
{
1
r
s
()
2
}
,
(2)
where = [ ]
can be expressed as
()
ALRT
(r) =
{
()
=1
[(r{
()
=1
, )], (3)
while with the latter as
()
(Q)HLRT
(r) =
{
()
=1
[(r,
()
{
()
=1
)], (4)
where
{
()
=1
[] and
()
is the vector of the unknown parameter
estimates under hypothesis
as [8], [9]
2
()
ALRT-UB
(r) =
=1
=1
[
()
1
exp
{
()
,
2
}]
,
(5)
where
,
represents the symbol over the -th period, drawn from this
signal constellation. The criterion used to make a decision on
the modulation is [8], [9]
= arg max
ln{
()
ALRT-UB
(r)}, (6)
where
is the estimate of the modulation, and ln{} is the
natural logarithm. For the BPSK and QPSK case study, one
can easily nd the analytical closed-form expressions for the
PDFs
(BPSK)
ALRT-UB
(r) and
(QPSK)
ALRT-UB
(r) respectively as
(BPSK)
ALRT-UB
(r) =
=1
()
1
exp
{
1
(
2
+
2
)
}
cosh(2
1
Re{
}),
(7)
and
(QPSK)
ALRT-UB
(r) =
=1
()
1
exp
{
1
(
2
+
2
)
}
cosh(
2
1
Re{
})
cosh(
2
1
Im{
}),
(8)
where cosh denotes the hyperbolic cosine function, and Re{}
and Im{} are the real and imaginary parts, respectively.
Furthermore, under the assumption of equally likely hypothe-
ses and by following [8], one can show that the average
probability of correct classication, dened as
BPSK,QPSK
=
2
1
=BPSK,QPSK
()
, with
()
BPSK,QPSK
1
2
(BPSK)
(BPSK)
(QPSK)
(QPSK)
,
where () represents the Q-function, and
()
and
()
are
the mean and variance of the variable
2
In this case, the average over all possible sequences of symbols reduces
to the average over all possible values of each symbol.
(
) = ln
{
1
2
2
=1
[
1
exp
{
(BPSK)
,
2
}]}
ln
{
1
4
4
=1
[
1
exp
{
(QPSK)
,
2
}]}
given the hypothesis
is true, = BPSK,QPSK.
C. The HLRT-based MC Algorithm with Signal Amplitude,
Phase, and Noise Power as Unknown Parameters
With the HLRT approach, MLE of the unknown parameters
are used to calculate the PDF of r in (4). By using the
joint parameter estimation method [17, Ch. 7], one can nd
the analytical closed form expressions for the ML estimators
of the unknown signal amplitude, noise power, and phase
respectively as
()
= Re
{
(s
()
r)(r
s
()
)
}
/s
()
2
, (9)
()
=
1
(r
2
r
s
()
2
)/s
()
2
), (10)
()
=
2
ln
(
s
()
r
r
s
()
)
, (11)
where the superscript denotes Hermitian transpose. It is
noteworthy that the ML estimators depend on the symbol
sequence s
()
. As such, these cannot be directly estimated in
an MC problem, in which the symbols are unknown. However,
by using (2), (4) and (9)-(11), one can obtain the expression
for the PDF of r under
as
()
HLRT
(r) =
(
r
2
)
=1
1
(1
()2
, (12)
where
()
= r
s
()
/(r s
()
), = 1,
, is
the correlation coefcient between r and s
()
, with s
()
as
a sequence of symbols drawn from the signal constellation
of modulation . With (12) calculated under all hypotheses,
the modulation is identied by correspondingly using (6). As
one can easily notice, the complexity to compute the PDF
of r under
is of the order of (
), and increases
with the number of symbols, (exponential increment) and
modulation order,
.
D. QHLRT-based MC Algorithm with Signal Amplitude,
Phase, and Noise Power as Unknown Parameters
With the QHLRT, NDA non-MLE of the unknown para-
meters are used to calculate the PDF of r in (4). NDA
estimators of the unknown parameters are of particular interest
for MC, as they do not rely on the presence of pilot symbols.
As the likelihood function is multimodal [18], [19], this cannot
provide a unique solution, and one resorts to statistic-based
estimators. When MoM estimates are employed, we refer to
the QHLRT-based algorithm as to QHLRT-MoM. The MoM
estimators of signal amplitude, , and noise power, for PSK
signals are respectively given by [6], [20]
()
=
(
2
2
,21
,42
)
1/4
, (13)
HAMEED et al.: ON THE LIKELIHOOD-BASED APPROACH TO MODULATION CLASSIFICATION 5887
and
()
=
,21
(
2
2
,21
,42
)
1/2
, (14)
where
,21
and
,42
are the estimates of the second-order/
one-conjugate and fourth-order/ two-conjugate moments of
the PSK received signal, given respectively by
,21
=
=1
2
and
,42
=
1
=1
4
. Note that
(13)-(14) hold regardless the PSK modulation order, and
the signal amplitude and noise power MoM estimators do
not actually depend on the hypothesis
when recognizing
PSK signals of different modulation orders. The MoM phase
estimator for PSK signals is given by [18, Ch. 5]
()
=
1
arg
(
=1
)
, (15)
and depends on the hypothesis
()
QHLRT-MoM
(r) =
=1
=1
(
()
)
1
exp
{
(
()
)
1
()
()
,
2
}
.
(16)
With (16) calculated under all hypotheses, the modulation is
identied by correspondingly using (6). One can easily notice
that the complexity to compute (16) is of order of (
),
and, apparently, much lower when compared to HLRT. Note
that NDA estimators of the unknown parameters for QAM
signals (including MoM estimators) can be used with the
QHLRT algorithm. For such estimators one can see, e.g., [6],
[18, Ch. 5], [20], [21].
IV. AN UPPER BOUND ON THE PERFORMANCE OF
QHLRT-BASED MC ALGORITHMS
A. Proposed Upper Bound on the Performance of QHLRT-
based MC Algorithms
Results in [3] and [7] show that an improved classication
performance is attained with more accurate parameter esti-
mates. For unbiased estimates, minimum variance estimates
which reach the CRB provide the best accuracy, and, accord-
ingly, lead to the best classication performance. An upper
bound on the performance of QHLRT-based algorithms which
employ unbiased NDA joint estimates of the unknown param-
eters is thus achieved for variances equal to the CRBs, and
calculation of the CRBs is crucial for the performance bound.
When considering estimates which are also normally dis-
tributed, we refer to this bound as the QHLRT-UB. With signal
amplitude, phase, and noise power as unknown parameters,
the QHLRT-UB is obtained by using
()
(, CRB
( )),
()
(, CRB
(
)), and,
()
(, CRB
( )) in (16),
with (,
2
) denoting a normal distribution with mean
and variance
2
, and CRB
( ), CRB
(
), and CRB
( ) as
the CRB of
()
,
()
,
()
, respectively. The modulation is
identied by correspondingly applying (6).
30 20 10 0 10 20 30
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
SNR, (dB)
g
1
(
)
,
g
2
(
)
g
2
()
g
1
()
Fig. 1. Numerical evaluation of the functions
1
() and
2
() in (18)
and (19), respectively, versus SNR.
B. The QHLRT-UB for BPSK and QPSK Modulation Classi-
cation
As already mentioned, the QHLRT-UB requires calculation
of the CRBs of NDA joint estimates of the unknown param-
eters. This is presented below for the case study of BPSK
and QPSK modulations ( =BPSK,QPSK), with unknown
signal amplitude, phase, and noise power. We refer interested
readers to [17, Ch. 3] for complete procedure details. The
Fisher information matrix for BPSK signals is given by (see
Appendix for derivations)
J
BPSK
=
2
1
()
1
()
0
1
() 0
1
2
2
1
() 0
0
2
2
2
()
,
(17)
where
1
() =
exp{}
2
exp{
2
/2}
cosh(
2)
, (18)
and
2
() =
exp{}
exp{
2
/2}
cosh(
2)
. (19)
The functions
1
() and
2
() are calculated through
numerical integration and plotted versus the SNR, , in Fig. 1.
They both decrease monotonically from one towards zero as
the SNR increases. By taking the inverse of the FIM matrix
in (17), one obtains the expressions for the CRBs of joint
estimates of , , and , respectively as
CRB
BPSK
( ) =
2
2
1 2
1
()
1 (2 + 1)
1
()
, (20)
3
CRB
BPSK
(
) =
2
1
1
()
1 (2 + 1)
1
()
, (21)
3
It can be shown that the CRB for the estimate of the noise power equals
4
/4 times the CRB for the fourth-order/ two-conjugate normalized cumulant
used for MC in [13] (which assumes perfect knowledge of ), taking into
account that the noise is complex-valued.
5888 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 12, DECEMBER 2009
and
CRB
BPSK
( ) =
1
2
1
1
2
()
. (22)
As one can easily notice, the CRBs depends on and .
Furthermore, the CRBs of the estimates of and also
depend on the actual value of the parameter, whereas the CRB
of the estimate of does not. It should be noted that the phase
CRB is actually decoupled from those of amplitude and noise
power. For QPSK signals, one can similarly show that
J
QPSK
() =
2
1
(/2)
1
(/2)
0
1
(/2) 0
1
2
2
1
(/2) 0
0
2
2
(1 + )
2
(/2)
.
(23)
By taking the inverse of the FIM matrix in (23), the
expressions for the CRBs of joint estimates of , , and
are respectively obtained as
CRB
QPSK
( ) =
2
2
1 2
1
(/2)
1 (2 + 1)
1
(/2)
, (24)
CRB
QPSK
(
) =
2
1
1
(/2)
1 (2 + 1)
1
(/2)
, (25)
and
CRB
QPSK
( ) =
1
2
1
1 ( + 1)
2
(/2)
. (26)
The above remarks on the CRBs in case of BPSK hold for
QPSK, as well.
V. NUMERICAL RESULTS
Numerical results for the CRBs of NDA joint estimates of
signal amplitude, phase, and noise power are subsequently
reported, along with simulation results for MoM estimates
of these parameters. Classication results for QHLRT-MoM,
QHLRT-UB, and ALRT-UB are presented, and a link between
QHLRT-MoM and QHLRT-UB is explained. A comparison of
the classication performance of the QHLRT-MoM, qLLR [4],
[5] and fourth-order cumulant based algorithms [12], [13] is
performed. The applicability of the QHLRT-MoM to MC in
Ricean block fading channel is additionally shown.
A. Simulation Setup
Unless otherwise mentioned, the signal amplitude, , is
set to 1, and the phase, , is uniformly distributed over the
[/4, /4) range and xed over the duration of the data
sequence. The number of symbols processed at the receive-
side is set to = 100. The number of Monte Carlo trials used
to compute the variance of MoM estimates of the unknown
parameters and the probability to correctly decide on the
modulation ,
()
)
/
2
i =QPSK
i =BPSK
Fig. 2. CRB
( )/
2
versus SNR, for =BPSK,QPSK.
10 5 0 5 10
10
0
10
1
10
2
SNR, (dB)
K
C
R
B
i
(
N
)
/
N
2
i =QPSK
i =BPSK
Fig. 3. CRB
(
)/
2
versus SNR, for =BPSK,QPSK.
B. CRBs of Non-Data Aided Joint Estimates of the Unknown
Parameters
By using (20)-(22) and (24)-(26), results for
CRB
( )/
2
, CRB
(
)/
2
, and CRB
( ),
= BPSK, QPSK, are plotted versus SNR in Figs. 2,
3, and 4, respectively. As expected, all curves decrease as
the SNR increases, and results for BPSK are below or close
to those for QPSK. In addition, the gap between BPSK and
QPSK increases with a decrease in SNR.
C. MoM Estimates of the Unknown Parameters
The variance of MoM estimates of and for the QPSK
signal, along with corresponding CRBs, is plotted versus
SNR in Fig. 5. The variance is lower bounded by the CRB
above a certain SNR, when MoM estimates are unbiased;
such a range is shown in Fig. 5. As expected, the variance
decreases as the SNR, , increases; this is relatively close to
corresponding CRB, especially for signal amplitude. Although
results for MoM estimates are presented here only for and
for the QPSK signal, we should mention that a similar
behavior has been observed for . Moreover, this remains
valid for BPSK, with the remark that the SNR above which
the estimates are unbiased is much lower when compared with
HAMEED et al.: ON THE LIKELIHOOD-BASED APPROACH TO MODULATION CLASSIFICATION 5889
10 5 0 5 10
10
2
10
1
10
0
10
1
10
2
10
3
10
4
SNR, (dB)
K
C
R
B
i
(
)
i =QPSK
i =BPSK
Fig. 4. CRB
a
n
d
N
Var
QPSK
(
N)
CRB
QPSK
(
N)
Var
QPSK
( )
CRB
QPSK
( )
Var and CRB for
N
Var and CRB for
Fig. 5. Variance of MoM estimates of the signal amplitude and noise power,
and the corresponding CRBs versus SNR, for QPSK and = 100.
QPSK. Simulations have been also carried out to study the
distribution of MoM estimates of the unknown parameters
for BPSK and QPSK signals. A normal distribution occurs at
higher SNR, whereas this distribution does not hold at lower
SNR. For example, MoM estimates of the phase approach
a uniform distribution as the SNR decreases. This is in
agreement with results obtained for the variance of MoM
estimates of the phase, which approaches
2
/48 as the SNR
decrease. This represents the variance of a uniform distributed
random variable, with values in the [/4, /4) range.
D. QHLRT-MoM, QHLRT-UB, and ALRT-UB
Simulation results for the QHLRT-MoM algorithm, as well
as for the QHLRT-UB and ALRT-UB are presented in Fig. 6,
when recognizing BPSK and QPSK signals. The average prob-
ability of correct classication,
BPSK,QPSK
, is plotted versus
SNR, for a range of values of practical interest (above 0.8).
Note that classication results for the QHLRT-MoM reect
the behavior of MoM estimates of all unknown parameters
for both BPSK and QPSK signals. For the
BPSK,QPSK
range
shown in Fig. 6, both QHLRT-UB and ALRT-UB provide
upper bounds on the classication performance of QHLRT-
8 6 4 2 0 2 4 6 8 10
0.8
0.82
0.84
0.86
0.88
0.9
0.92
0.94
0.96
0.98
1
SNR, (dB)
P
c
c
B
P
S
K
,
Q
P
S
K
ALRTUB
QHLRTUB
QHLRTMoM
Fig. 6. Performance of QHLRT-MoM, and QHLRT-UB and ALRT-UB, when
recognizing BPSK and QPSK with = 100.
TABLE II
VARIANCE AND CRB OF THE PARAMETER ESTIMATES AT 3.5 DB AND
2 DB SNR, RESPECTIVELY, FOR QPSK AND = 100.
Var
QPSK
0.0041 0.0111 0.009
CRB
QPSK
0.0044 0.01442 0.009
MoM. As expected, QHLRT-UB yields a tighter bound than
ALRT-UB. Although the latter is a very optimistic upper
bound, it is still of interest, as providing the best performance
that can be achieved with any MC algorithm, including LB
algorithms.
A link between the performance of the QHLRT-MoM and
QHLRT-UB is shown, as follows. According to Fig. 6, an
average probability of correct classication of almost one,
BPSK,QPSK
value at
around 2 dB SNR. Apparently, achieving a probability of cor-
rect classication for QPSK of almost 1,
(QPSKQPSK)
1,
requires higher SNR when compared to correctly identifying
BPSK with the same accuracy and number of symbols. The
performance is limited for QPSK, in which case 3.5 dB SNR
is required to attain
(QPSKQPSK)
1 with QHLRT-MoM,
whereas only 2 dB is needed with QHLRT-UB. The variances
and CRBs corresponding to the MoM estimates of the signal
amplitude, , phase, , and noise power,
, for which this
performance is achieved, are given in the rst and second row
of Table II. One can easily notice the similarity of the values
of the variance and CRB for each parameter, which explains
similar results for the QHLRT-MoM and QHLRT-UB at the
aforementioned SNRs.
E. Comparison of the Performance of QHLRT-MoM, qLLR,
and Fourth-order Cumulant Based Algorithms
Classication results achieved with the QHLRT-MoM are
compared with those for the qLLR and fourth-order cu-
mulant based algorithms proposed in [4], [5] and [12],
[13], respectively. Classication of BPSK versus QPSK and
16-QAM versus V.29 is considered. The qLLR is derived as
5890 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 12, DECEMBER 2009
2 1 0 1 2 3 4 5
0.75
0.8
0.85
0.9
0.95
1
SNR, (dB)
P
B
P
S
K
,
Q
P
S
K
c
c
No residual
channel
effect
QHLRTMoM
qLLR
cr,42
| cr,40|
Residual
channel effect,
h
2
=0.2
Residual
channel effect,
h
2
=0.1
=1
4
is compared against a threshold to make a decision whether
the modulation is
-PSK or
-PSK (
>
) at each
node [4]. For QAM classication, the qLLR is used for binary
hypothesis testing problems, and applied to distinguishing 16-
QAM and V.29 with = 100 symbols [5]. Proposed statistics
are
4
=
=1
4
,
4
=
=1
(see footnote
4
), and
4
= 0.0135
4
0.0246
4
. Thresholds for decision making
are set similarly to the PSK case. In [12], [13], fourth-order
cumulant based features are proposed to classify linear digital
modulations. In [12], a hierarchical classier is developed to
identify a large pool of modulations, such as -ary PSK,
QAM, and ASK, V.29, V.32, and V.29c. The magnitude of
the fourth-order/ zero-conjugate normalized cumulant,
,40
,
is employed for the classication of 16-QAM versus V.29,
as well as BPSK versus QPSK. In [13], the fourth-order/
two-conjugate normalized cumulant,
,42
, is used to classify
a reduced pool of modulations when compared to [12]; the
algorithm is not applicable to -ary PSK, with > 4
5
.
In [12], [13], normalization of the fourth-order cumulants
is performed to the second power of the second-order/ one-
conjugate cumulant of the signal component. The test statistics
are estimated and compared against thresholds for decision
making. The thresholds are set as the arithmetic average of
the test statistic means under hypotheses considered in the
ascending order of these means.
Results for BPSK versus QPSK achieved with the QHLRT-
MoM and algorithms proposed in [4] and [12], [13] are
4
One can easily notice that these statistics represent magnitudes of signal
moments multiplied by the number of symbols, .
5
An extension to the time dispersive channel is performed in [13], while
only the additive white Gaussian noise channel is considered in [4], [5],
[12], and this work. Nevertheless, the effect of a residual channel on the
classication performance of the investigated algorithms is shown here.
presented here, with signal amplitude, phase, and noise power
as unknown parameters. Neither qLLR nor fourth-order cumu-
lant based algorithms require phase estimation; both require
estimation of signal amplitude. In addition to the aforemen-
tioned unknown parameters, the effect of a residual channel
is investigated. A three-tap channel [1
1
2
] is considered,
with
1
and
2
as zero-mean independent Gaussian random
variables, each with variance
2
4
. With the QHLRT-MoM, the average probability of correct
classication approaches 1 as the SNR exceeds 20 dB. In
addition to not exhibiting saturation, QHLRT-MoM outper-
forms both qLLR and fourth-order cumulant based algorithms
for average probabilities of correct classication of interest
(above 0.8).
F. QHLRT-MoM Under Fading Conditions
Performance of the QHLRT-MoM under fading conditions
is presented in Fig. 9. The average probability of correctly
classing BPSK and QPSK,
BPSK,QPSK
, is plotted versus
the Rice factor, . The values = 0 and = correspond
to the cases of Rayleigh fading and no fading, respectively.
Simulation results show that QHLRT-MoM yields a reasonable
performance for the entire range of values.
VI. CONCLUSION
In this paper we investigate likelihood-based algorithms for
linear digital modulation classication. Our results indicate
that the HLRT is very computationally demanding, while
the QHLRT algorithm offers a reasonable alternative. An
upper bound on classication performance of QHLRT-based
HAMEED et al.: ON THE LIKELIHOOD-BASED APPROACH TO MODULATION CLASSIFICATION 5891
0 5 10 15 20 25
0.5
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
SNR, (dB)
P
1
6
Q
A
M
,
V
.
2
9
c
c
QHLRT-MoM
qLLR, t4
cr,42
| cr,40|
qLLR, p4
qLLR, q4
Fig. 8. Performance comparison between QHLRT-MoM, qLLR, and
fourth-order cumulant based algorithms when recognizing 16-QAM and
V.29, with K=100.
infinity 0 5 10 15 20
0.75
0.8
0.85
0.9
0.95
1
R (dB)
P
B
P
S
K
,
Q
P
S
K
c
c
K=100, SNR=5 dB
K=100, SNR=3 dB
Fig. 9. Performance of QHLRT-MoM when discriminating BPSK and QPSK
in Ricean fading, with = 100.
algorithms is proposed (QHLRT-UB), for the case when unbi-
ased and normally distributed non-data aided estimates of the
unknown parameters are available. This requires calculation
of the CRBs of non-data aided parameter estimates. As a
case study we investigate classication of BPSK and QPSK
modulations with signal amplitude, phase, and noise power
as unknown parameters. The CRBs of joint non-data aided
parameter estimates are derived for this case, and further
employed to develop the QHLRT-UB. MoM estimates of the
unknown parameters are explored and used with the QHLRT,
leading to the QHLRT-MoM. According to simulation results,
when the MoM estimates are unbiased, their variance is
relatively close to corresponding CRBs, and the QHLRT-MoM
provides a reasonable performance. A comparison between
the QHLRT-MoM and qLLR algorithms reveals that although
both provide a similar performance when discriminating PSK
signals with no residual channel effects, the QHLRT-MoM
is more robust to this model mismatch. On the other hand,
the fourth-order/ two-conjugate cumulant based algorithm is
the most robust under such conditions. When compared with
qLLR and fourth-order cumulant based algorithms, QHLRT
can be successfully applied to distinguish QAM modulations.
In addition, QHLRT provides an acceptable performance un-
der fading conditions. We note that QHLRT displays good
performance with reasonable complexity, and appears to be
a promising candidate algorithm. Hence, we intend to further
explore the QHLRT algorithm in future work, by extending
the case study presented here to a larger pool of modulations,
more unknown parameters, and more complex propagation
environments.
APPENDIX
DERIVATION OF (17)
The derivation of the rst element of the Fisher In-
formation Matrix for the BPSK case, [J
BPSK
()]
1,1
=
[
2
ln
BPSK
(r)/
2
], is presented here. Other ele-
ments of the matrix can be similarly derived. The like-
lihood function
BPSK
(r) is computed by averaging
(r{
(BPSK)
=1
, ) over the constellation points correpond-
ing to the BPSK modulation; this is given by (7). By using (7),
one can easily show that
2
ln
BPSK
(r)
2
= 2
1
+4
2
=1
(
,
cos +
,
sin)
2
sech
2
(2
1
(
,
cos +
,
sin)),
(27)
where
,
and
,
are the in-phase and quadrature compo-
nents of the signal samples
, = 1, , , respectively.
Futher, by applying expectation w.r.t
BPSK
(r) to (27), one
can write
[J
BPSK
()]
1,1
=
2
ln
BPSK
(r)
2
BPSK
(r)r
=
1
+
2
, (28)
where
1
= [2
1
] = 2
1
and
2
= [
=1
(
)]
with
(
) = 4
2
(
,
cos +
,
sin )
2
sech
2
(2
1
(
,
cos +
,
sin)).
By using that {
=1
are independent and identically dis-
tributed random variables, and the expectation is linear, one
can drop the dependency on , and write
2
as
2
= [()]. (29)
The expectation in (29) is w.r.t
BPSK
(), in which = 1
and the dependency on is dropped. By using that
()
+
,
+
,
, with
,
and
,
as the in-phase
and quadrature components of
, respectively, and by
dropping the dependency on and changing variables, one
can show that (29) can be written as
2
= 4
1
3
exp{
1
2
}
exp{
1
(
)
2
}
(
()
, , ), (30)
where (
()
, , ) =
=
(
()
+
)
2
exp{
1
(
()
+
)
2
}
cosh(2
1
(
()
+
))
.
5892 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 12, DECEMBER 2009
By using that the integrand in the expression of (
()
, , )
is an even function of
()
+
2
=
2
exp{}
2
exp{
2
/2}
cosh(
2)
. (31)
Finally, by replacing results for
1
and
2
in (28), it is
straightforward to obtain the expression for [J
BPSK
()]
1,1
as in (17). Derivation of the FIM matrix for QPSK can
be similarly carried out. Note that results obtained here for
[J
]
,
, , = 1, 2 are consistent with those reported in [22],
where the CRB of NDA estimates of SNR is derived. In
addition, numerical results achieved for [J
]
3,3
by using the
analytical expression derived here are consistent with those
reported in [23].
ACKNOWLEDGMENT
The authors would like to thank the anonymous reviewers
for their constructive comments on the paper.
REFERENCES
[1] O. A. Dobre and F. Hameed, Likelihood-based algorithms for linear
digital modulation classication in fading channels," in Proc. 19
th
IEEE Canadian Conf. Electrical Computer Engineering CCECE 2006,
Ottawa, Ontario, Canada, May 2006, pp. 1347-1350.
[2] , On performance bounds for parameter estimation and modulation
classication," in Proc. IEEE Sarnoff Symposium, 2007, pp. 1-5.
[3] O. A. Dobre, A. Abdi, Y. Bar-Ness, and W. Su, A survey of automatic
modulation classication techniques: classical approaches and new
trends," IET Commun., vol. 1, pp. 137-156, 2007.
[4] C. Y. Huang and A. Polydoros, Likelihood methods for MPSK mod-
ulation classication," IEEE Trans. Commun., vol. 43, pp. 1493-1504,
1995.
[5] C. Long, K. Chugg, and A. Polydoros, Further results in likelihood
classication of QAM signals," in Proc. IEEE MILCOM, 1994, pp. 57-
61.
[6] A. Abdi, O. A. Dobre, R. Choudhry, Y. Bar-Ness, and W. Su, Modu-
lation classication in fading channels using antenna arrays," in Proc.
IEEE MILCOM, 2004, pp. 211-217.
[7] O. A. Dobre, A. Abdi, Y. Bar-Ness, and W. Su, Blind modulation
classication: a concept whose time has come," in Proc. IEEE Sarnoff
Symposium, 2005, pp. 223-228.
[8] W. Wei and J. Mendel, Maximum-likelihood classication for digital
amplitude-phase modulations," IEEE Trans. Commun., vol. 48, pp. 189-
193, 2000.
[9] J. A. Sills, Maximum-likelihood modulation classication for
PSK/QAM," in Proc. IEEE MILCOM, 1999, pp. 217-220.
[10] P. Panagiotou, A. Anastasopoulos, and A. Polydoros, Likelihood ratio
tests for modulation classication," in Proc. IEEE MILCOM, 2000, pp.
670-674.
[11] O. A. Dobre, J. Zarzoso, Y. Bar-Ness, and W. Su, On the classication
of linearly modulated signals in fading channels," in CD, Conf. Inform.
Sciences Systems (CISS), Princeton University, 2004.
[12] A. Swami and B. M. Sadler, Hierarchical digital modulation classi-
cation using cumulants," IEEE Trans. Commun., vol. 48, pp. 416-429,
2000.
[13] H.-C. Wu, M. Saquib, and Z. Yun, Novel automatic modulation
classication using cumulant features for communications via multipath
channels," IEEE Trans. Wireless Commun., vol. 7, pp. 3098-3105, Aug.
2008.
[14] C. M. Spooner, On the utility of sixth-order cyclic cumulants for RF
signal classication," in Proc. IEEE ASILOMAR, 2001, pp. 890-897.
[15] D. Boudreau, C. Dubuc, F. Patenaude, M. Dufour, J. Lodge, and
R. Inkol, A fast automatic modulation recognition algorithm and its
implementation in a spectrum monitoring application," in Proc. IEEE
MILCOM, 2000, pp. 732-736.
[16] H. L. Van Trees, Detection, Estimation and Modulation Theory. New
York: Wiley, 2001.
[17] S. M. Kay, Fundamentals of Statistical Signal Processing, Estimation
Theory, Vol. 1. Englewood Cliffs, NJ: Prentice Hall, 1993.
[18] U. Mengali and A. N. DAndrea, Synchornization Techniques for Digital
Receivers. New York: Plenum, 1997.
[19] T. R. Benedict, The joint estimation of signal and noise from sum
envelope," IEEE Trans. Inf. Theory, vol. IT-13, pp. 447-454, 1967.
[20] D. R. Pauluzzi and N. C. Beaulieu, A comparison of SNR estimation
techniques for the AWGN channel," IEEE Trans. Commun., vol. 48, pp.
1681-1691, 2000.
[21] H. Xu, Z. Li, and H. Zheng, A non-data aided SNR estimation
algorithm for QAM signals," in Proc. ICASSP, 2004, pp. 999-1003.
[22] N. S. Alagha, Cramer-Rao bounds of SNR estimates for BPSK and
QPSK modulated signals," IEEE Commun. Lett., vol. 5, pp. 10-12, 2001.
[23] W. G. Cowley, Phase and frequency estimation for PSK packets:
bounds and algorithms," IEEE Trans. Commun., vol. 44, pp. 26-28,
1996.
Fahed Hameed received his Bachelors Degree in
Electrical Engineering from National University of
Sciences and Technology, Pakistan in 2003 and Mas-
ters Degree in Electrical Engineering from Memo-
rial University of Newfoundland, Canada, in 2006.
Currently, he is working in the Telecoms O&M de-
partment of Qatar Electric Supply Company (Kahra-
maa). His research interests include blind parameter
estimation, modulation classication, and multihop
cellular networks.
Octavia A. Dobre received the Engineering
Diploma and Ph. D. degrees in Electrical Engi-
neering from Politehnica University of Bucharest
(formerly the Polytechnic Institute of Bucharest),
Romania, in 1991 and 2000, respectively. In 2001
she joined the Wireless Information Systems Engi-
neering Laboratory at Stevens Institute of Technol-
ogy in Hoboken, NJ, as a Fulbright fellow. Between
2002 and 2005, she was with the Department of
Electrical and Computer Engineering at New Jersey
Institute of Technology (NJIT) in Newark, NJ, as
a Research Associate. Since 2005 she has been an Assistant Professor with
the Faculty of Engineering and Applied Science at Memorial University of
Newfoundland, Canada. Her current research interests include blind modula-
tion classication and parameter estimation, cognitive radio, multiple antenna
systems, multicarrier modulation techniques, cyclostationarity applications in
communications and signal processing, and resource allocation in emerging
wireless networks. Dr. Dobre is an Associate Editor for the IEEE COMMUNI-
CATIONS LETTERS, and has served as the Chair of the Signal Processing and
Multimedia Symposium of the 2009 IEEE Canadian Conference on Electrical
and Computer Engineering (CCECE).
Dimitrie C. Popescu received the Engineering
Diploma and M.S. degrees in 1991 from the Poly-
technic Institute of Bucharest, Romania, and the
Ph.D. degree from Rutgers University in 2002, all
in Electrical Engineering. His research interests are
in the areas of wireless communications, digital
signal processing, and control theory. He is cur-
rently an Assistant Professor in the Department of
Electrical and Computer Engineering, Old Dominion
University. Between 2002 and 2006 he was with the
Department of Electrical and Computer Engineering,
the University of Texas at San Antonio. He has also worked for AT&T Labs
in Florham Park, New Jersey, on signal processing algorithms for speech
enhancement, and for Telcordia Technologies in Red Bank, New Jersey, on
wideband CDMA systems. He has served as technical program chair for the
vehicular communications track of the IEEE VTC 2009 Fall, nance chair for
the IEEE MSC 2008, and technical program committee member for the IEEE
GLOBECOM, WCNC 2006, and VTC conferences. His work on interference
avoidance and dispersive channels was awarded second prize in the AT&T
Student Research Symposium in 1999. He is the co-author of a monograph
book on interference avoidance methods for wireless systems published in
2004 by Kluwer Academic Publishers.