926

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 6, NO. 3, MARCH 2007

Exact BER for M -QAM with MRC and Imperfect

Channel Estimation in Rician Fading Channels
Yao Ma, Member, IEEE, Robert Schober, Member, IEEE, and Dongbo Zhang, Student Member, IEEE

Abstract In this paper, we study the effect of imperfect

channel estimation (ICE) on the performance of M -level quadrature amplitude modulation (M -QAM) with maximum ratio
combining (MRC) and pilot-symbol assisted modulation (PSAM)
in generalized Rician fading channels. By expressing the bit
error rate (BER) of MRC diversity M -QAM in terms of the
distribution of new decision variables, we derive novel, exact,
and easy-to-evaluate BER expressions for diversity M -QAM
with channel estimation errors. Our results include versatile
system and fading channel parameters (e.g., arbitrary spatial and
temporal correlation patterns among the diversity branches), and
are valid for arbitrary linear channel estimators and square and
rectangular M -QAM with different constellation sizes. In addition, we evaluate the performance of minimum mean-squared
error (MMSE)- and sinc-interpolator-based channel estimators
with PSAM, and provide some new insights into the performance
of M -QAM with PSAM in generalized fading channels.
Index Terms Bit error rate, channel estimation error, M -level
quadrature amplitude modulation, maximum ratio combining,
Rician channels, correlated fading.

I. I NTRODUCTION
-LEVEL quadrature amplitude modulation (M -QAM)
provides high spectral efficiency and is thus widely
employed in many wireless communication applications [1]
[3]. The accurate bit error rate (BER) evaluation of the M QAM receivers in various fading conditions has attracted a lot
of interest in the literature [4][14]. In practice, the channel
estimates cannot be perfect in fading channels, and thus the
adverse effect of imperfect channel estimation (ICE) on the
BER performance must be taken into account in the system
design. In [4], for pilot symbol assisted modulation (PSAM),
an approximate BER for M -QAM was obtained for a Rayleigh
fading channel. In [5], the BER for the minimum meansquared-error (MMSE)-based channel estimator for a nondiversity Rician fading channel was derived. In [6], the exact
BERs for pulse amplitude modulation (PAM) and QAM signals in a maximum ratio combining (MRC) diversity Rayleigh
fading channel and a non-diversity Rician fading channel
with ICE were obtained, respectively. In [9], [10] a symbol
error rate (SER) formula involving a two-fold integral was

Manuscript received April 20, 2005; revised August 17, 2005; accepted
November 11, 2005. The associate editor coordinating the review of this paper
and approving it for publication was S. Zhou. This paper has been presented
in part at IEEE Wireless Commun. and Networking Conf. (WCNC), New
Orleans, March 2005.
Yao Ma and Dongbo Zhang are with the Department of ECpE, Iowa State
University, Ames, IA, 50011 (e-mail: {mayao, dbzhang}@iastate.edu).
Robert Schober is with the the Dept. of Electrical & Computer Engineering,
University of British Columbia, 2356 Main Mall, Vancouver, BC, V6T 1Z4,
Canada (e-mail: rschober@ece.ubc.ca).
Digital Object Identifier 10.1109/TWC.2007.05243.

developed for arbitrary two-dimensional signalling formats in

a non-diversity Rician fading channel. In [11], the authors
provided a BER expression (in the form of a two-fold integral
of special functions) for 16-QAM with MRC and ICE in
a Rayleigh fading channel with independent and identically
distributed (i.i.d.) diversity branches. Recently, closed-form
BER expressions for MRC diversity square M -QAM with
PSAM in i.i.d. Rayleigh and Rician fading channels were
derived in [12] and [13], respectively.
In many practical scenarios, the receiver diversity branches
have non-equal signal gains, and they may even be arbitrarily correlated. To the best of our knowledge, an accurate
performance result for diversity M -QAM with PSAM valid
for correlated fading channels (including arbitrary spatial and
temporal fading correlation and noise correlation) has not been
obtained yet, despite its theoretical importance and practical
relevance.
In this paper, by expressing the BER of the M -QAM MRC
receiver in terms of the distribution of new decision variables
(DVs), we derive exact and easy-to-compute analytical expressions for the BER of square and rectangular M -QAM
with ICE. Our results are valid for arbitrary linear channel
estimators, generalized Rician and Rayleigh fading channels,
and account for various system and fading channel parameters,
including frequency-offset and arbitrary fading spectra. Using
our new result, we evaluate the performance of MMSE- and
sinc-interpolator-based channel estimators with PSAM, and
provide some new insights into the performance of M -QAM
with ICE.
The remainder of this paper is organized as follows. The
signal model for the M -QAM MRC receiver with ICE is introduced in Section II. A BER analysis framework is proposed
in Section III, and the procedure to evaluate the distribution
of the new decision variables is elaborated in Section IV.
The effects of various system parameters on the BER of M QAM are discussed using numerical examples in Section V.
In Section VI, we draw some conclusions.
Throughout this paper, we use superscripts , T , H, 1 to
represent the scalar conjugate, vector (or matrix) transpose,
conjugate transpose, and matrix inversion, respectively. We
use E[x], (x), and (x) to denote the expected value, real
part, and imaginary part of x, respectively; and det(A) is
the determinant of matrix A. IL represents the L L identity
matrix, and 0MN and 1MN are M N all-zero and all-one
matrices, respectively. diag(x1 , . . . , xL ) is a diagonal matrix
that is formed by setting x1 , . . . , xL on its main diagonal.

c 2007 IEEE
1536-1276/07$25.00 

MA et al.: EXACT BER FOR M-QAM WITH MRC AND IMPERFECT CHANNEL ESTIMATION IN RICIAN FADING CHANNELS

II. S YSTEM AND C HANNEL M ODEL

A. Signal Model
We use (I, J)-QAM to denote the modulation with I
and J signal levels in the horizontal and vertical directions,
respectively, where I, J {2n }, n = {1, 2, . . .}. The total
number of signal
M is given by M = I J. When
levels
I = J, the ( M , M )-QAM is the square M -QAM; and
when I = J, the (I, J)-QAM is called the rectangular QAM
[3], [8]. We denote the data signals received in the ith symbol
interval over all L diversity branches by

(1)
y(i) = Pd c(i)d(i) + n(i)
where P
d is the transmission power of data symbols and
d(i)
=
Ea (h + jv ) is an M -QAM symbol, with j =

1, h {I + 1, . . . , 1, 1, . . . , I 1}, and v
{J + 1, . . . , 1, 1, . . . , J 1}. Ea is used to normalize the
data symbol energy to unity, i.e., Ed = E[|d(i)|2 ] = 1. For
3
square M -QAM Ea = 2(M1)
; and for rectangular (I, J)3
QAM Ea = I 2 +J 2 2 [8]. c(i) = [c1 (i), . . . , cL (i)]T is
the channel-coefficient vector for the L branches. In Rician
fading channels, c(i) can be written as c(i) = c (i) +
T
cf (i), where c (i) = E[c(i)] = [c1 (i), . . . , cL (i)] and
cf (i) = [cf1 (i), . . . , cfL (i)]T are the line-of-sight (LOS) and
diffuse components of c(i), respectively. At the lth branch,
cl (i) = |cl (i)|ej2(fdl +fol )iTs +l , where fdl is the Doppler
shift of the LOS component, fol is the frequency offset, l
is a random residual phase, and Ts is the symbol duration.
The Rician factor is defined as Kl = |cl (i)|2 /c2l , where
c2l = E[|cfl (i)|2 ] is the variance of the diffuse fading
component of the lth branch. The normalized temporal channel
c (n) =
correlation coefficient at the lth branch is defined as R
l

2
E[cfl (i)cfl (i n)]/cl . For Clarkes fading spectrum,
c (n) = J0 (2nBf Ts )ej2fol nTs ; and for the Gaussian
R
l
l
c (n) = exp((nBf Ts )2 )ej2fol nTs ,
fading spectrum, R
l
l
where J0 (x) is the zeroth order Bessel function and Bfl is
the Doppler fading bandwidth of the diffuse component.
The additive background noise vector, n(i)
=
[n1 (i), . . . , nL (i)]T , is a zero-mean circularly symmetric
complex Gaussian process, with average power
E[|nl (i)|2 ] = N0 , l = 1, . . . , L. In (1), n(i) may
be temporally and spatially correlated. The average bit
signal-to-noise ratio (SNR) of the lth branch is given by
b,l = Pd E[|cl (i)|2 ]/N0 = Pd (|cl (i)|2 + c2l )/N0 . Without
loss of generality, we assume Pd = 1 throughout this paper
unless otherwise stated.
B. PSAM-based Channel Estimators
For a PSAM-based channel estimator (PSAM-CE), a pilot
symbol is inserted into the data stream every P symbol
intervals. To estimate the channel coefficient vector c(i) for
the desired symbol d(i), F pilot symbols are employed which
may be written as an F 1 vector dPS = [d(i P F1 +
ioff ), . . . , d(i P + ioff ), d(i + ioff ), . . . , d(i + P (F2 1) + ioff)]T ,
where F1 and F2 (with F1 + F2 = F ) are the numbers of pilot
symbols on the left and right side of d(i), respectively, and
ioff (ioff = 1, 2, . . . , P 1) is the offset of the desired symbol
d(i) to the closest pilot symbol on its right side. The received

927

signals at the pilot symbols positions for estimating channel

c(i) may be written as an F L 1 vector yPS , expressed in a
compact form as
yPS = (diag(dPS ) IL )cPS + nPS

(2)

where denotes the Kronecker product [15], cPS = [cT (i

P F1 + ioff), . . . , cT (i + ioff ), . . . , cT (i + P (F2 1) + ioff)]T and
nPS = [nT (i P F1 + ioff ), . . . , nT (i + ioff ), . . . , nT (i + P (F2
1) + ioff )]T are the channel gain and noise components of all
the L branches at the pilot symbols positions, respectively.

Without loss of generality, we assume dPS = PPS 1L1 ,

where PPS is the transmitted power of the pilot symbols, and
may be different from Pd , the transmitted power
of the data
symbols. Thus, equation (2) simplifies to yPS = PPS cPS + nPS .
The channel estimate for c(i), denoted as c(i), is given by
T T
] is an L F L
c(i) = WyPS , where W = [w1T , . . . , wL
channel estimator filter matrix. cl (i) = wl yPS is true, where
wl is the filter (lth row of W) to estimate the channel gain
cl (i).
Although the analysis of this paper is applicable to any
linear channel estimator W, we outline two types of channel
estimators below which have attracted particular research
interest.
1) Sinc-CE: For the sinc-interpolator-based CE with
rectangular window, we have Wsinc = [hsinc (P F1 +
ioff ), . . . , hsinc (ioff ), . . . , hsinc (P (F2 1) + ioff )] IL , where
hsinc (x) = sin(x)/(x). A Hamming window may also be
employed. Our numerical results (shown in Section V) show
that the relative performance of rectangular and Hamming
windows depends on the operating channel and system parameters.
2) MMSE-CE: We derive the linear MMSE channel estimator (MMSE-CE) for the PSAM in generalized Rician fading
channels below. A detailed exposition here is necessary to
elaborate the relation between the channel estimator and the
BER performance of M -QAM. By using the Wiener filter
[16], the MMSE-CE is given by Wmmse = Rc,yPS R1
yPS , where
H
Rc,yPS = E[c(i)yPS
] is the cross-correlation matrix between
c(i) and yPS (taking into account the LOS component), and
H
RyPS = E[yPS yPS
] the auto-correlation matrix of yPS . After
some manipulations, we can show that

PPS [Rc (P F1 ioff ), . . . , Rc (ioff ),
Rc,yPS =
. . . , Rc (P (F2 1) ioff )] (3)
and RyPS is given by (4) shown at the top of the next
page, where Rc (m) = E[c(i)cH (i m)] = c (m) +
c (i)H
c (i m) is the spatial-temporal correlation matrix
of {c(i)}i=0,1,... (including the LOS components). c (m) =
E[cf (i)cH
f (i m)] is the channel spatial-temporal covariance
matrix. Assuming the noise n(i) is a zero-mean process,
n (m) = E[n(i)nH (i m)] is defined as the noise spatialtemporal covariance matrix. For i.i.d. white noise, we obtain
that n (0) = N0 IL , and n (m) = 0LL , when m = 0. In
summary, our channel model includes arbitrary joint spatialtemporal fading correlation, noise correlation, and Rician
fading. Thus, it is much more general than the i.i.d. Rayleigh
and Rician fading models studied in [11][13].

928

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 6, NO. 3, MARCH 2007

H
PPS RH
c (P ) + n (P )
PPS Rc (0) + n (0)

PPS Rc (0) + n (0)

PPS Rc (P ) + n (P )
..
.

RyPS =

...
...
..
.

PPS Rc (P (F 1)) + n (P (F 1))

b1 = 1
b2 = 1

b2 = 1

~
B1

~
B2

~
B2

3a

3a

11

10

00

01

Fig. 1. Decision boundaries and bit-symbol mapping (b1 , b2 ) for 4-PAM.

2 denote the bit decision boundaries for b1 and b2 , respectively.
1 and B
B
b1 = 1
b2 = 1

b2 = 1

~
B2

~
B3

7a
111

~
B1

~
B3

~
B2

~
B3

~
B3

5a

3a

3a

5a

7a

110

100

101

001

000

010

011

b3 = 1

b3 = 1

b3 = 1

Fig. 2. Decision boundaries and bit-symbol mapping (b1 , b2 , b3 ) for 81 , B

2 , and B
3 denote the bit decision boundaries for b1 , b2 , and b3 ,
PAM. B
respectively.

III. E RROR P ROBABILITY A NALYSIS

We derive the BER for horizontal and vertical pulse amplitude modulation (PAM) first, and then extend the result to
both square and rectangular M -QAM.
A. I-PAM

Consider horizontal I-PAM with d(i) = h Ea , where

Ea = I 231 . For I = 4, as shown in Fig. 1, the two bits
b1 b2 are Gray-coded and mapped to four possible transmitted
R = (D)
be the real part of D.
The error
symbols. Let D
probabilities for b1 and b2 are, respectively, given by
P4,H (1)

1
R < 0|h = 1}]
R < 0|h = 3} + P {D
[P {D
2

P4,H (2) =
1
R < 2a|h = 3}
R < 2a|h = 3} P {D
[P {D
2
R < 2a|h = 1} + P {D
R > 2a|h = 1}]
+P {D

(4)

. . . PPS Rc (0) + n (0)

where a (as shown in Fig. 1) is given by a = Ea , and

P {A|B} denotes the conditional probability of event A given
event B.

I = (D)

For vertical 4-PAM with d(i) = jv Ea , let D

The BERs for b1 , b2 can be
be the imaginary part of D.
I < 0|v = 3} + P {D
I <
obtained as P4,V (1) = 12 [P {D
1

0|v = 1}], and P4,V (2) = 2 [P {DI < 2a|v = 3}

I < 2a|v = 3} +P {D
I < 2a|v = 1} + P {D
I >
P {D
2a|v = 1}] For horizontal and vertical PAM, respectively, let
us define the conditional BERs as
PH (n |n )
PV (n |n )

R < an |d(i) = an }
= P {D
I < an |d(i) = jan }
= P {D

(5)
(6)

where n is an integer specifying the decision boundary. Thus,

we obtain P4,H (1) = 12 [PH (0|3) + PH (0|1)] and P4,H (2) =
1

2 [PH (2|3) PH (2|3)+ PH (2|1)+ 1 PH (2|1)]. Here, for

R > 2a|h =
deriving P4,H (2) we used the equality that P {D

1} = 1 P {DR < 2a|h = 1} = 1 PH (2|1). We can

express the average BERs for horizontal and vertical 4-PAM,
respectively, as
P4,H

=
=

C. Symbol Detection
(i) to detect d(i),
By using the estimated channel vector c

the complex decision variable

(DV)
for
MRC
L
isL given by2 D =
H
H
c (i)y(i)/[
c (i)
c(i)] = l=1 cl (i)yl (i)/ l=1 |
cl (i)| . The

transmitted symbol d(i) can be recovered by comparing D

with the horizontal and vertical QAM decision boundaries [7],
[8], [11].

P4,V

1
[P4,H (1) + P4,H (2)]
2
N4

1
[C4 +
wn PH (n |n )]
4
n=1

(7)

N4

1
[C4 +
wn PV (n |n )]
4
n=1

(8)

where C4 = 1, N4 = 6, and the coefficients wn , n , and n

are given in Table I (a).
For horizontal 8-PAM, the bit-symbol mapping is shown by
Fig. 2. Using the definition in (5), we can express the BERs
for the bits b1 , b2 , and b3 , respectively, as
P8,H (1) =

P8,H (2) =

P8,H (3) =

1
[PH (0|1) + PH (0|3) + PH (0|5)
4
+PH (0|7)],

(9)

1
[2 + PH (4|5) PH (4|5)
4
+PH (4|7) PH (4|7) + PH (4|3)
+PH (4|1) PH (4|3) PH (4|1)] (10)
1
2 + PH (6|7) PH (2|7) + PH (2|7)
4
PH (6|7) PH (6|5) + PH (2|5)
PH (2|5) + PH (6|5) PH (6|3)
+PH (2|3) PH (2|3) + PH (6|3)
PH (2|1) + PH (6|1) + PH (2|1)

PH (6|1)

(11)

MA et al.: EXACT BER FOR M-QAM WITH MRC AND IMPERFECT CHANNEL ESTIMATION IN RICIAN FADING CHANNELS

The average BER for horizontal 8-PAM can be obtained as

P8,H

=
=

1
[P8,H (1) + P8,H (2) + P8,H (3)]
3
N8

1
[C8 +
wn PH (n |n )]
12
n=1

For rectangular (I,J)-QAM (with M = I J),conditioned

on the transmitted symbol d(i) = (h + jv ) Ea , let us
define the conditional BERs for the horizontal and vertical
components of QAM as
PH (n |h , v )


R < n Ea |d(i) = (h + jv ) Ea }
= P {D
PV (n |h , v )


I < n Ea |d(i) = (h + jv ) Ea }
= P {D
Note that for Gray-coded (I,J)-QAM, the decision boundaries for horizontal signals (real part of d(i)) are independent
of the vertical signal levels (imaginary part of d(i)), and vice
versa, see e.g. [7], [8]. Thus, we can directly extend the BER
result for PAM obtained in Subsection III-A to the case of
M -QAM, and obtain
2
PI,H =
[CI
I log2 I
NI
J
1

+
wn PH (n |n , J 2m + 1)] (13)
J m=1 n=1

(14)

where C4 = 1, C8 = 4, N4 = 6, and N8 = 28. When I (or

J) = 4 and 8, the values of {wn , n , n } have to be selected
from Table I (a) and (b), respectively. The average BER is
given by
PM =

1
[(log2 I)PI,H + (log2 J)PJ,V ]
log2 (I J)

n
n
n
wn

(12)

B. M -QAM

2
[CJ
J log2 J
NJ
I
1

+
wn PV (n |I 2m + 1, n )]
I m=1 n=1

TABLE I
C OEFFICIENTS FOR BER CALCULATION FOR 4-PAM AND 8-PAM.

(a) 4-PAM

where C8 = 4 and N8 = 28, and the wn , n , and n are given

in Table I (b), respectively. Similarly, using the symmetry, it
follows that the average BER for vertical 8-PAM is P8,V =
N8
1
n

n=1 wn PV (n |n )]. The results for 2 -PAM (for

12 [C8 +
n 4) can be obtained using a similar procedure, but are
omitted here due to space limitations.
We underscore that when both c(i) and
c(i) are circularly
symmetric complex Gaussian processes as usual, we have that
PI,H = PI,V , for I = 2n , n {1, 2, 3, . . .}. However, when
either c(i) or
c(i) are not circularly symmetric (e.g. for an
improper process [17]), the detection error probabilities for the
real and imaginary parts of d(i) may not be equal, i.e., PI,H =
PI,V . Therefore, our framework is valid for both proper and
improper fading processes.

PJ,V =

929

(15)

We emphasize that for non-Gray-coded bit-mappings, our

results are applicable with slight modifications. In detail,

1
2
3
1

2
-2
3
-1

3
-2
1
1

4
2
1
-1

5
0
3
1

6
0
1
1

(b) 8-PAM
n
1
2
3
4
5
6
7
8
9
10
11
12
13
14

n
0
0
0
0
4
-4
4
-4
-4
-4
4
4
6
2

n
1
3
5
7
5
5
7
7
3
1
3
1
7
7

wn
1
1
1
1
1
-1
1
-1
1
1
-1
-1
1
-1

n
15
16
17
18
19
20
21
22
23
24
25
26
27
28

n
-2
-6
6
2
-2
-6
6
2
-2
-6
2
6
-2
-6

n
7
7
5
5
5
5
3
3
3
3
1
1
1
1

wn
-1
-1
-1
1
-1
1
-1
1
-1
1
-1
1
1
-1

equation (13) has to be rewritten as

J

1
2
PI,H =
CI (m)
I log2 I J m=1
+

NI

wn (m)PH (n (m)|n (m), J 2m + 1) (16)

n=1

where CI (m), wn (m), n (m), and n (m) are now dependent

on the signal level m (m = 1, . . . J) in the vertical direction,
and we just need to tabulate them for each m, following the
same procedure as for generating Table I. Similar steps can be
taken to re-evaluate PJ,V . Thus, our new result is applicable
to arbitrary symbol-bit mappings.
When both c(i) and c(i) are circularly symmetric Gaussian
processes, we obtain that PH (n |n , m ) = PV (n |m , n ).
For Gray-coded square M -QAM, (15) simplifies to

4

PM =
CM
M log2 M

M N
M

1
+
wn PH (n |n , M 2m + 1) (17)
M m=1 n=1
When M = 16 and 64, the values of {wn , n , n } have
to be selected from Table I (a) and Table I (b), respectively.
Next, we re-formulate the conditional BERs PH (n |h , v )
and PV (n |h , v ) for QAM in terms of the distribution of
some new DVs. PH (n |h ) and PV (n |v ) for PAM can be
obtained as special cases.
We note that in [8] the BER of M -QAM is obtained by
extending the result of I-PAM to QAM. The result in [8] is
valid for the case of perfect channel state information (CSI)

930

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 6, NO. 3, MARCH 2007

and a non-fading channel, and is based on a recursive relation

(valid for Gray-coded bit mapping) between the BER and the
signal amplitude levels. For the case of ICE the SNR penalty
is a function of the signal amplitude, and thus such a recursive
relation [8] is no longer applicable, and the BER analysis
method in [8] cannot be straightforwardly extended to the
ICE case. In comparison, we used a different framework by
evaluating the distribution of some new decision variables, and
our result includes the effects of both generalized fading and
ICE.

IV. E VALUATING THE D ISTRIBUTIONS OF THE N EW DV S

We first obtain the cdfs of the new DVs for generalized
Rician fading channels, and then discuss special-case results
for Rayleigh fading and independent diversity branches.
A. Rician Fading Channels
We study the general case of arbitrary spatial and temporal
correlations between diversity branches and between different
symbol intervals.
To obtain the MGFs of DH (n |d) and DV (n |d), we
express them in Gaussian quadratic forms as
DH (n |d)
DV (n |d)

C. Reformulation of the BERs

By definition, PH
(n |h , v ) = P {DR < n Ea |d(i)
R =
= (h + jv ) E
}. Using the equalities D
a L
L

2
(
l (i)yl (i))/ l=1 |
cl (i)|
and
yl (i)
=
l=1 c
cl (i) Ea (h + jv ) + nl (i), we obtain PH (n |h , v ) =
P {DH (n |d(i)) < 0}, where DH (n |d(i)) is a new DV
defined as
 L


DH (n |d(i)) = 
cl (i)yl (i)
l=1
L


|
cl (i)|2
Ea n

(18)

l=1

Similarly, we define the conditional BER for the vertical signal

components as PV (n |h , v ) = P {DV (n |d(i)) < 0},
where
 L

L

cl (i)yl (i) Ea n
|
cl (i)|2
DV (n |d(i)) = 
l=1

l=1

Below, we drop the symbol index i of d(i) when no confusion

arises. To evaluate the cumulative distribution functions (cdfs)
P {DH (n |d) < 0} and P {DV (n |d) < 0}, we derive
the moment generating functions (MGFs) of the new DVs
DH (n |d) and DV (n |d), and then use the inverse Laplace
transform to obtain these cdfs.
We note that our approach for BER analysis is different
from those given in [7], [11][13]. In [7] and [11], several
BER expressions were obtained by evaluating the effective
SNRs in terms of estimation-error-rotated signal in-phase
and quadrature components and the decision boundaries. In
[12] and [13], closed-form BER formulas were obtained
by comparing the receiver output with decision boundaries
using a characteristic function method. The result in [7] is
applicable to square QAM in a non-diversity Rayleigh fading
channel, and the results in [12], [11] and [13] are applicable
to square QAM in i.i.d. Rayleigh and Rician fading channels,
respectively, c.f. [12, eqs. (9)-(12)], [11, eqs.(32)-(36)], and
[13, Table I]. In our paper, we evaluate the distribution of some
new decision variables given in terms of the complex signal
points and signal boundaries. Our exact result is valid for
both square and rectangular M -QAM, for arbitrarily correlated
Rician fading channels, and for all linear channel estimators.
The new result also accounts for non-Gray-coded bit mapping,
and both proper and improper fading processes. Thus, it is
much more general than the results given in [7], [11], [13],
and [12].

where

= vH (d)QH (n )v(d)
= vH (d)QV (n )v(d)


v(d)

QH (n )

QV (n )



=

c(i)
c(i)d + n(i)




n aIL
0.5IL

0.5IL
0LL

n aIL
0.5jIL

0.5jIL
0LL

Using a property of Gaussian quadratic forms [18], [19],

we obtain the MGF of DH (n |d) in Rician channels as
DH (n |d) (s) =

1
(d))
exp(
vH (d)[Q1
v (d)]1 v
H (n )s
det(I2L sv (d)QH (n ))

(19)

(d) and v (d) are the mean vector and the covariance
where v
(d) and
matrix of v(d), respectively. The procedure to find v
v (d) for generalized Rician fading and any linear channel
estimators is outlined below.
c (i) + cf (i), where
c (i) and cf (i) are the
Let c(i) =
LOS and diffusive component vectors of c(i), respectively.
The mean and the auto-covariance matrix of v(d), respectively, be given by
(d)
v

v (d)

T (i), Tc (i)d]T
[
 c

c,c
c,c d
2
H
c,c d c,c |d| + n (0)

(20)
(21)

where c,c = E[cf (i)cH

f (i)] is the cross-covariance matrix
between c(i) and c(i), c,c = E[cf (i)cH
f (i)], and c,c =
H
E[cf (i)cf (i)]. Note that c,c , c,c , and c,c completely
model the effect of ICE on the diffuse channel components,
and the signal correlation between different branches.
(d) for a given
Next, we show how to evaluate v (d) and v
channel estimator W. Let cPS = PS + cfPS , where PS and
cfPS are the vectors of the LOS and diffuse components of
PS , where y
PS =
c
PS , respectively. Then, yPS = PPS PS + y
PPS cfPS + nPS is a zero-mean vector. We can show that

c (i) =

PPS WPS
(22)
H
yPS cf (i)]
c,c = E[W

H
= WE[( PPS cfPS + nPS )cH
(23)
f (i)] = Wc,yPS


H
c,c = E[W( PPS cfPS + nPS )( PPS cfPS + nPS ) WH ]
=

WyPS WH

(24)

MA et al.: EXACT BER FOR M-QAM WITH MRC AND IMPERFECT CHANNEL ESTIMATION IN RICIAN FADING CHANNELS

H
where c,yPS = E[cf (i)
yPS
] is the cross-covariance matrix
H
PS
between c(i) and yPS , and yPS = E[
yPS y
] is the autocovariance matrix of yPS . They can be obtained by replacing
Rc (m) in (3) and (4) with c (m), respectively, as shown
by the equations given next and at the top of the next page,
respectively,

c,yPS =
PPS [c (P F1 ioff ), . . . , c (ioff ),

. . . , c (P (F2 1) ioff )]
A special case is discussed below. For the case of
white noise, independent diversity branches, and identical fading correlation functions at all the branches (i.e.,

R
c1 (n) = . . . = RcL (n) = Rc (n)), we have c,yPS =

c (P (F2 1)
PPS [Rc (P F1 ioff ), . . . , Rc (ioff ), . . . , R
2
2
ioff )] diag(c1 , . . . , cL ), and yPS = N0 IF L +

c (P (F 1))
c (P ) . . . R
1
R

c (P )
R
1
...,

PPS

..
.
.

.
.,
c (P (F 1))
R
...
1
diag(c21 , . . . , c2L )
Applying (20) (24) to (19), we can evaluate the MGF of
the new DV. The cdf P {DH (n |d) < 0} can be evaluated by
the inverse Laplace transform of the MGF, as shown by
P {DH (n |d) < 0} =
 c+j

DH (n |d) (s)
1

ds
2
js
cj

(25)

where c is a small real constant in the convergence region

[20]. The MGF of DV (n |d) (s) can be obtained by replacing
QH (n ) with QV (n ) in (19), and the cdf P {DV (n |d) <
0} can be obtained by replacing DH (n |d) (s) with
DV (n |d) (s) in (25).
Equation (25) can be evaluated by a Gauss-Chebyshev
quadrature (GCQ) formula [21],


N
1
(2n 1)
N

+R
P {DH (n |d) < 0} =
2N n=1
2N

where ()
= DH (n |d) (c jc tan(/2))(1 j tan(/2)),

and RN is a residual term which vanishes for N . The

GCQ formula with (15)(17) gives accurate BER results for
M -QAM in arbitrary Rician fading channels taking ICE into
account. Using the results in [22], a closed-form expression
for (25) can be obtained, which is, however, omitted here due
to the space limitations.
B. Rayleigh Fading Channels
The results for arbitrarily correlated branches in Rayleigh
(d)
channels can be obtained using (19) and (25) by setting v
to a zero vector. Below, we consider independent diversity
branches.
1) I.N.D. Diversity: To gain more insight into the relation
between channel estimation accuracy and BER performance
of M -QAM with PSAM, we derive the BER formula for
independent but non-identically distributed (i.n.d.) diversity
Rayleigh channels. To our knowledge, an exact BER result

931

for MRC M -QAM in i.n.d. Rayleigh fading channels is not

available. Define the normalized correlation coefficient l
between cfl (i) and cfl (i) as
cfl (i)cfl (i)]/[cl cl ]
(26)
l = E[

where E[
cfl (i)cfl (i)] = [c,c ]l,l , cl = [c,c ]l,l , and l,l
is the (l, l)-th entryof matrix . Thus, we obtain that l =
1
H
H
cl [c,yPS W ]l,l / [WyPS W ]l,l , which is a function of
the estimation filter W (which may be non-MMSE-CE) and
various fading channel parameters.
For i.n.d. Rayleigh fading branches, the MGF for the DV
DH (n |d) is given by
DH (n |d) (s) =

L

l=1

det(I2 svl (d)QH,2 (n ))1



an
0.5

0.5
0

where QH,2 (n ) =
and vl (d) =

2

cl
l cl cl d
. Let the two eigenvalues of
l cl cl d c2l |d|2 + N0
vl (d)QH,2 (n ) be denoted by
l (n , d). Since the signals in all L branches are independent, all the poles of
+
L
DH (n |d) (s) are given by {1/
l (n , d), 1/l (n , d)}l=1 .
Using the result in the Appendix, we obtain

1
[(l,R h l,I v )acl cl n ac2l ]
2
1
c {a2 [c2l (1 2l,R )2v + c2l (1 2l,I )2h
2 l
2
2c,l
l,R l,I h v 2n cl cl

(l,R h l,I v ) + n2 c2l ] + N0 }1/2 (27)

l (n , d) =

where l,R = (l ) and l,I = (l ). For convenience

we sort the 2L eigenvalues in the ascending order so that
2L
{l (n , d)}L
l=1 are negative, and {l (n , d)}l=L+1 are positive. Assuming all the negative eigenvalues of v (d)QH (n )
are distinct, we obtain
P {DH (n |d) < 0} =
2L
L


l=1 m=1,m=l

l (n , d)
l (n , d) m (n , d)

(28)

The BER of M -QAM can be obtained by substituting (28)

into (15)(17), respectively.
Obviously, from (28) we observe that an error floor appears
if l (n , d) < 0 for l = 1, . . . , L (i.e., vl (d)QH,2 (n ) has
negative eigenvalues) when the noise power N0 0, or the
average SNR goes to infinity. We observe that when N0 = 0
and l = 1 for all l, it follows from (27) that
l (d, n ) = 0
for all l, which shows that an error floor does not occur in
this case.
2) I.I.D. Diversity: In the literature, a substantial part of the
research effort on the BER evaluation of M -QAM has been
devoted to i.i.d. Rayleigh fading channels, e.g., c.f. [6], [11],
[12]. Here, we derive results for this case based on our new
approach for the purpose of comparison. For i.i.d. branches,
we obtain that l = , c2l = c2 , and c2l = c2 for l =
1, . . . , L. In this case, v (d)QH (n ) has only two distinct
eigenvalues with each of them being repeated L times, and
we denote them as (n , d). Using a result in [23], the cdf

932

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 6, NO. 3, MARCH 2007

PPS c (0) + n (0)

PPS c (P ) + n (P )
..
.

yPS =

H
PPS H
c (P ) + n (P ) . . .
PPS c (0) + n (0)
...
..
.

PPS c (P (F 1)) + n (P (F 1))

L1

L + l 1
L
P {DH (n |d) < 0} = [A(n |d)]
l
l=0

(29)

(n ,d)
where A(n |d) = (n,d)
+ ( ,d) . By using (27), we
n
obtain (30), shown at the top of the next page.
Let us rewrite A(n |d) as A(n |d) = A(n |h , v ), where
d = a(h +jv ). Then, we obtain a closed-form average BER
expression for M -QAM with L-fold MRC as (31).
For comparison, the result given in [11, eq. (37)] is in
the form of a two-dimensional (2-D) integral of the 2L-order
parabolic cylinder function, and thus is numerically intensive
to evaluate. In [12, eqs. (9), (12)], an alternative closed-form
BER expression for M -QAM with PSAM was presented. Our
result in (31) and (30) explicitly shows the relation between l
and the BER, and thus provides more insight into the behavior
of PSAM, as will be discussed next in Subsection IV-B.3.
3) MMSE-CE: The MMSE channel estimator for PSAM
has attracted a lot of research interest in the literature, and
it provides a performance benchmark for other linear channel
estimators. For MMSE-CE with i.n.d. branches, we can show
that

H ] /
[Wmmse yPS Wmmse
l =
l,l
cl

H
=
[c,yPS 1
(32)
yPS c,yPS ]l,l /cl


2 2
2 2
|l | cl
|l | cl d
vl (d) =
(33)
|l |2 c2l d c2l |d|2 + N0

Obviously, l is real-valued for MMSE-CE. Using (27) we get

the eigenvalue pairs as

l (d, n ) =

1 2 2
1
cl al (h n ) l {c4l a2
2
2
[2h + 2l n2 + (1 2l )2v 22l n h ]
(34)
+c2l N0 }1/2

By substituting (34) into (28), we obtain a closed-form BER

expression for QAM with MMSE-CE in i.n.d. Rayleigh fading.
For MMSE-CE and i.i.d. Rayleigh fading channels, it
follows that the BER of M -QAM is given by (31), where
A(n |d) therein is replaced by
A(n |d) =

1 1
[ac (h n )]
2 2
[c2 a2 (2h + 2 n2 + (1 2 )2v
22 n h ) + N0 ]1/2

. . . PPS c (0) + n (0)

P {DH (n |d) < 0} can be obtained as

[1 A(n |d)]l

(35)

The study of the detection error floor of QAM with PSAM

is of practical interest. Let us consider the BER of MRC QAM

with MMSE-CE for the asymptotic case of high SNR. We note

that when h n > 0 and h n < 0, the conditional BERs
are given by P {DH (n |d) < 0} and 1 P {DH (n |d) <
0}, respectively. Thus, for brevity of presentation, we assume
h n > 0 and study P {DH (n |d) < 0} below.
For MMSE-CE the effect of ICE at the lth branch is
manifested by the fact that l < 1. When N0 = 0 and l < 1,
using (34) we observe that
l (d, n ) < 0 is true. Note that
the negative eigenvalues of vl QH,2 (n ) for N0 = 0 mean
that a detection error floor occurs. When N0 = 0 and l = 1
for all l, using (34) again we observe that
l (d, n ) = 0
2
(d,

)
=

a(

),
which
shows
that an error
and +
n
h
n
cl
l
floor does not occur in this case. Thus, {l }L
l=1 is closely
related to the performance of diversity QAM with PSAM. The
error floor for M -QAM with MMSE-CE can be analytical
evaluated by using (32)(35) with (28) or (29). We note
that the normalized residual MSE (which is proportional to
1|l |2 defined in this paper) of MMSE-CE in a non-diversity
Rayleigh channel has recently been studied in [24, Appendix].
It was shown that for a bandlimited fading process (e.g.,
the Jakes model) with Bf Ts P < 1/2 and infinite PSAM
interpolation order (F ), the normalized residual MSE
decreases proprotionally as the SNR increases, which suggests
that under these conditions an error floor does not occur.
Finally, we underscore that we have verified that for all
the square and rectangular M -QAM constellations obtained
in this paper, our new BER results, when simplified to the
case of perfect CSI, are numerically identical to the known
results given in [3], [8].
V. N UMERICAL E XAMPLES AND D ISCUSSIONS
In this section, we provide some numerical results for the
performance of the M -QAM MRC receiver in general Rician
fading channels taking into account the effects of ICE and
various system and channel parameters. We assume Pd = PPS
unless otherwise stated. To clearly illustrate the effects of
Doppler fading bandwidth and the Rician K factor, we also
assume Bf,l = Bf , fd,l = fd , and Kl = K (for all l) for all
the cases studied.
A. Verification of the BER Analysis
For correlated fading, the popular constant spatial correlation model [3] (with correlation coefficient c ) is assumed.
Forexample, for balanced
branches and L = 3, c (0) =

1 c c
c2 c 1 c . To show the effect of the phase vector
c c 1
= [1 , , L ] of the LOS components, we assume fd Ts =
0 and Bf Ts = 0.02. In our simulation, we assume Rc (n) =
Rc (0)J0 (2nBf Ts ), i.e., the spatial and temporal correlations
can be separated.

MA et al.: EXACT BER FOR M-QAM WITH MRC AND IMPERFECT CHANNEL ESTIMATION IN RICIAN FADING CHANNELS

A(n |d) =

[(R h I v )ac n ac]

1 1

(30)
2
2
2
2
2
2
2 2 a [c (1 R )v + c (1 I )2h 2c2 R I h v 2n cc (R h I v ) + n2 c2 ] + N0

PM

933

M N M

4
1

{C M +
wn [A(n |n , M 2m + 1)]L
log2 M
M m=1 n=1


L1

L+l1

[1 A(n |n , M 2m + 1)]l }
l

(31)

l=0

10

Perfect CSI (independent), simu.

MMSECE (independent), simu.
MMSECE (correlated), simu.
Analysis

= [0, 0, 0]

10

10

Bit error probability

Bit error probability

10

L=1

10

L=2

10

10

= [0, /2, ]

Perfect CSI, simu.

5

MMSECE, simu.

10
4

10

sincCE (Hamming), simu.

sincCE (Rect.), simu.
Analysis

L=3
6

10

10
15
Average bit SNR per branch (dB)

20

0.1

0.2

0.3

0.4
0.5
0.6
Spatial correlation coefficient

0.7

0.8

0.9

Fig. 3.
BER (averaged over all ioff s) vs. average bit SNR per branch
for 16-QAM MRC receivers with CSI and the MMSE-CE, respectively, in
independent and correlated Rician (K = 5 dB) fading channels. = 01L ,
L = {1, 2, 3}, P = 10, F1 = F2 = 5, Bf Ts = 0.02, fd Ts = 0, and
fo = 0.

Fig. 4. BER (averaged over all ioff s) vs. spatial correlation coefficient c
for 16-QAM MRC receivers with CSI, MMSE- and sinc-interpolator-based
(with rectangular and Hamming windows) channel estimators, respectively, in
Rician (K = 5 dB) fading channels. = [0, 0, 0] and [0, /2, ],
b = 10
dB, L = 3, P = 10, F1 = F2 = 5, Bf Ts = 0.02, fd Ts = 0, and fo = 0.

First, we present the simulated and calculated BER results

(averaged over all possible ioff s) for 16-QAM receivers with
L = {1, 2, 3} in i.i.d. and correlated Rician fading channels
in Fig. 3, and the results for the effect of c in Fig. 4, respectively. The simulation results (markers) verify the validity and
accuracy of our new BER analysis approach. Fig. 3 shows
that the correlated fading (with being an all-zero vector)
severely degrades the performance of MRC QAM in Rician
fading channels. As L increases, the degradation of BER for
MMSE-CE caused by correlated branches becomes larger.
Fig. 4 shows that as c increases, the performance of MRC
QAM with CSI, MMSE-CE, and sinc-CE becomes worse for
= [0, 0, 0], but becomes better than the independent fading
case for = [0, /2, ] (i.e., a non-cophased LOS vector).
This observation may be explained by the fact that the phase
vector significantly affects the MRC output Rician-K factor
and the SNR gain in correlated Rician fading channels, c.f.
[25].
Next, we present the BER results for the 32-QAM MRC
receiver in i.n.d. Rayleigh and Rician fading channels in
Fig. 5, and those for 16-QAM in i.i.d. Rayleigh and Rician
fading channels in Fig. 6. For independent Rician channels the
distribution of the LOS phase vector does not affect the BER
performance, and thus we set to a zero vector for all the
remaining numerical examples. In Fig. 5, for Bf Ts = 0.02,
QAM with MMSE-CE gives a performance close to that of

the perfect CSI case, and the SNR loss is within 2 dB for
all SNRs shown in the figure; while in Fig. 6, for a larger
Bf Ts (Bf Ts = 0.03) and a larger P the relevant SNR gap
becomes larger. Fig. 5 shows that the BER of sinc-CE with
Hamming window (sinc-Hamm-CE) is uniformly lower than
that of sinc-CE with rectangular window (sinc-rect-CE). On
the contrary, for the parameters studied in Fig. 6, QAM with
sinc-rect-CE performs significantly better than QAM with
sinc-Hamm-CE. In addition, the SNR gaps between sinc-CE
and MMSE-CE depend on the average SNR, P , F , Bf Ts ,
and other parameters. This observation demonstrates that the
relative performance between sinc-rect-CE and sinc-HammCE, and the performance gap between sinc-CE and MMSECE, crucially depend on the channel and system parameters.
B. Effects of Design Parameters
In the following, we show the effects of some design
parameters, including the pilot symbol insertion interval P ,
number of pilot symbols F , and the ratio PPS /Pd on the
performance of the MRC M -QAM.
We show the effect of P on BER of 16-QAM with MMSECE and sinc-CE in Fig. 7, assuming L = 2 and ioff = P/2 .
Our results show that the BER degrades slowly when P 14.
When P > 14, both the performances of MMSE-CE and sincHamm-CE start to degrade rapidly. Sinc-Hamm-CE may give

934

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 6, NO. 3, MARCH 2007

0

10

10

Perfect CSI, simu.

Perfect CSI
MMSECE
sincHammCE

MMSECE, simu.

32QAM
1

10

SincCE, Hamm
Win., simu.

10

SincCE, Rect.
Win, simu.
Analysis

16QAM
B T = 0.03, L=2

Bit error probability

Bit error probability

10
10

Dash lines: Rayleigh

Solid lines: K=5 dB

10

Rayleigh Fading

10

10

10

K = 5 dB

10

10

10
15
Average bit SNR at the first branch (dB)

20

10

25

Fig. 5. BER (averaged over all ioff s) vs. average bit SNR of the first branch
for 32-QAM MRC receivers with CSI, MMSE- and sinc-interpolator-based
channel estimators, respectively, in i.n.d. Rayleigh and Rician (K = 5 dB)
fading channels. L = 2, P = 10, F1 = F2 = 4, Bf Ts = 0.02, fd Ts =
0.01, and fo = 0. The SNR in the first branch is 2 dB larger than that in the
second branch.

16QAM
Dash lines: Rayleigh
Solid lines: K=5 dB
With channel gain power normalization

10

Perfect CSI, simu.

MMSECE, simu.
SincCE, Hamm
Win., simu.
SincCE, Rect.
Win, simu.
Analysis

16QAM

25

10

20

Fig. 7. BER versus the pilot symbol insertion interval P for MRC 16-QAM
with MMSE-CE and sinc-CE, respectively, in i.i.d. Rayleigh and Rician (K =
5 dB) fading channels.
b = 20 dB, L = 2, ioff = P/2, F1 = F2 = 6,
Bf Ts = 0.03, fd Ts = 0.015, fo = 0.

10

10

10
15
Pilot symbol interval P

Perfect CSI
MMSECE
sincHammCE
sincRectCE
F=20, P=10
BfT=0.03, L=3

SNR = 10 dB

10

Bit error probability

Bit error probability

10

Dash lines: Rayleigh

10

10

SNR = 20 dB
5

10

10

Solid lines: K=5 dB

6

10

10

10

10

10
15
Average bit SNR per branch (dB)

3
4
5
Power Ratio between Pilot and Data symbols PPS/Pd

20

Fig. 6. BER (averaged over all ioff s) vs. average bit SNR per branch for 16QAM MRC receivers with CSI, MMSE- and sinc-interpolator-based channel
estimators, respectively, in i.i.d. Rayleigh and Rician (K = 5 dB) fading
channels. L = 2, P = 15, F1 = F2 = 5, Bf Ts = 0.03, fd Ts = 0.015,
and fo = 0.

a performance very close to (though uniformly worse than)

that of MMSE-CE, especially for P = 12.
Next, we study the effect of PPS /Pd , the power ratio
between pilot and data symbols in Fig. 8. The total average
transmission power is given by Pt = P1 ( + (P 1))Pd ,
where = PPS /Pd . For a fair comparison between the cases
of PPS /Pd = 1 and PPS /Pd = 1, we let Pt to be identical
for both cases. When increases, the power allocated to the
data symbol Pd decreases which degrades the BER, while
at the same time the channel estimation accuracy becomes
better which improves the BER. Without power normalization, c(i) could be directly obtained as c(i) = WyPS . We
study the case with channel gain powernormalization where

c(i) is obtained from

c(i) = WyPS / . To evaluate the
corresponding BER, we need to modify (22)(24) as follows:

Fig. 8. BER vs. the power allocation ratio PPS /Pd for MRC 16-QAM
with MMSE-CE and sinc-CE (with power normalization), respectively, in
i.i.d. Rayleigh and Rician (K = 5 dB) fading channels.
b = {10, 20} dB,
L = 3, P = 10, ioff = 5, F1 = F2 = 10, Bf Ts = 0.03, fd Ts = 0.015,
fo = 0.

c (i) = W PPS PS / , c,c = Wc,yPS / , and c,c =

WyPS WH /.
Fig. 8 shows that PPS /Pd [2, 3] gives optimal BER
performance for MMSE-CE and sinc-CE, and compared to
the case of PPS /Pd = 1 the BERs can be reduced by a
factor of about 1.5 2.5. We note that for non-constant
modulus formats, such as M -QAM, the power normalization
of the channel estimate is necessary (except for the MMSE-CE
where the power normalization has been taken into account in
Wmmse ). Otherwise, the case of PPS /Pd = 1 (e.g., for sinc-CE)
may severely degrade the performance.
We study the effects of F2 and F on the BER of the
M -QAM MRC receivers with MMSE-CE and sinc-CE. We
present the BER vs. F2 for 32-QAM in i.i.d. Rician and
Rayleigh fading channels in Fig. 9 for F = 24. The cases of
F2 = 0 and F2 = 12 correspond to a prediction-based (thus

MA et al.: EXACT BER FOR M-QAM WITH MRC AND IMPERFECT CHANNEL ESTIMATION IN RICIAN FADING CHANNELS

935

10

VI. C ONCLUSIONS

Perfect CSI
MMSECE
sincRectCE

B T = 0.03, P=15
f

L=2, SNR=20 dB
1

32QAM
Dashed lines: Rayleigh Fading
Solid lines: K = 5 dB

Bit error probability

10

10

10

10

10

10

15

20

25

F2

Fig. 9. BER versus F2 for MRC 32-QAM with MMSE-CE and sinc-rectCE, respectively, in i.i.d. Rayleigh and Rician (K = 5 dB) fading channels.

b = 20 dB, L = 2, P = 15, ioff = 7, F = 24, Bf Ts = 0.03, fo = 0.

By expressing the BER of MRC diversity M -QAM in terms

of the distribution of new decision variables, we derived the
exact BER formulas of diversity square and rectangular M QAM with ICE for generalized Rician fading channels. Our
new result is valid for arbitrary linear channel estimators,
and unbalanced and correlated diversity branches. In addition,
we have discussed the relation between channel estimation
accuracy and BER floor, and evaluated the performance of
MRC M -QAM with PSAM MMSE-CE and sinc-CE using
numerical examples. Our results have shown that the relative
performance of Hamming window and rectangular window
for sinc-CE, and the performance gap between sinc-CE and
MMSE-CE, depend on the system and channel parameters.
These results provide valuable insights for the design of
PSAM-based channel estimation schemes for diversity QAM
in generalized channels.
R EFERENCES

10

Perfect CSI
MMSECE
sincHammCE
sincRectCE

Bf T = 0.03, P=15
K = 5 dB, L=2
2

Bit error probability

10

64QAM

10

10

10

16QAM
6

10

10

15
20
Number of pilot symbolsF

25

30

35

Fig. 10. BER versus F for MRC 16- and 64-QAM receivers with MMSECE and sinc-CE, respectively, in an i.i.d. Rician fading channel. K = 5 dB,

b = 20 dB, L = 2, ioff = 7, F1 = F2 = F/2, Bf Ts = 0.03, fo = 0.

one-sided) and a centered observation window, respectively.

Our result shows that the centered observation window (F1 =
F2 = 12) gives a better performance than the case of F2 = 12
for MMSE-CE and sinc-rect-CE, as expected. Furthermore,
MMSE-CE is less sensitive to a non-centered window than
sinc-CE.
The BER vs. F for 16- and 64-QAM with MMSE-CE
and sinc-CE is presented in Fig. 10. We observe that as F
increases, the BER performances of QAM with MMSE-CE
and sinc-CE improve monotonically. For a small to medium
F , sinc-Hamm-CE may be worse than sinc-rect-CE, while for
a larger F (e.g. F 15), sinc-Hamm-CE performs uniformly
better than the latter, and its performance converges to that
of the MMSE-CE. This observation confirms that the relative
performance of sinc-rect-CE and sinc-Hamm-CE depends on
the system parameters.

[1] L. Hanzo, W. Webb, and T. Keller, Single- and Multi-Carrier Quadrature Amplitude Modulation. New York: Wiley, 2000.
[2] G. L. Stuber, Principles of Mobile Communications. Norwell, MA:
Kluwer Academic Publishers, second ed., 2003. Third Printing.
[3] M. K. Simon and M.-S. Alouini, Digital Communications over Fading
Channels: A Unified Approach to Performance Analysis. John Wiley &
Sons, 2000.
[4] J. K. Cavers, An analysis of pilot symbol assisted modulation for
Rayleigh fading channels, IEEE Trans. Vehic. Technol., vol. 40,
pp. 686693, Nov. 1991.
[5] M. G. Shayesteh and A. Aghamohammadi, On the error probability
of linearly modulated signals on frequency-flat Ricean, Rayleigh and
AWGN channels, IEEE Trans. Commun., vol. 43, pp. 14541466,
Feb./Mar./Apr. 1995.
[6] S. K. Wilson and J. M. Cioffi, Probability density functions for analyzing multi-amplitude constellations in Rayleigh and Ricean channels,
IEEE Trans. Commun., vol. 47, pp. 380386, March 1999.
[7] X. Tang, M.-S. Alouini, and A. J. Goldsmith, Effect of channel
estimation error on M-QAM BER performance in Rayleigh fading,
IEEE Trans. Commun., vol. 47, pp. 18561864, Dec. 1999.
[8] K. Cho and D. Yoon, On the general BER expression of one- and twodimensional amplitude modulations, IEEE Trans. Commun., vol. 50,
pp. 10741080, July 2002.
[9] X. Dong and L. Xiao, Two-dimensional signaling in Ricean fading
with imperfect channel estimation, in Proc. IEEE Wireless Commun.
and Networking Conf. (WCNC), pp. 11591164, 21-25 Mar. 2004.
[10] X. Dong and L. Xiao, Symbol error probability of two-dimensional
signaling in Ricean fading with imperfect channel estimation, IEEE
Trans. Vehic. Technol., vol. 54, pp. 538549, Mar. 2005.
[11] L. Cao and N. C. Beaulieu, Exact error-rate analysis of diversity 16QAM with channel estimation error, IEEE Trans. Commun., vol. 52,
pp. 10191029, June 2004.
[12] B. Xia and J. Wang, Effect of channel-estimation error on QAM systems with antenna diversity, IEEE Trans. Commun., vol. 53, pp. 481
488, March 2005.
[13] L. Cao and N. C. Beaulieu, Closed-form BER results for MRC diversity
with channel estimation errors in Ricean fading channels, IEEE Trans.
Wireless Commun., vol. 4, pp. 14401447, July 2005.
[14] Y. Ma, R. Schober, and D. Zhang, Exact BERs for M -QAM with
MRC and channel estimation errors in Rician channels, in Proc. IEEE
Wireless Commun. and Networking Conf. (WCNC), vol. 2, pp. 967972,
Mar. 2005.
[15] R. A. Horn and C. R. Johnson, Topics in Matrix Analysis. Cambridge
University Press, 1991.
[16] S. Haykin, Adaptive Filter Theory. Englewood Cliffs, NJ: Prentice Hall,
3rd ed., 1996.
[17] F. D. Neeser and J. L. Massey, Proper complex random processes with
applications to information theory, IEEE Trans. Inform. Theory, vol. 39,
pp. 12931302, July 1993.
[18] G. L. Turin, The characteristic function of hermitian quadratic forms
in complex normal variables, Biometrika, vol. 47, pp. 199201, 1960.

936

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 6, NO. 3, MARCH 2007

[19] M. Schwartz, W. Bennett, and S. Stein, Communication Systems and

Techniques. McGraw Hill, 1966.
[20] C. W. Helstrom, Calculating error probabilities for intersymbol and
cochannel interference, IEEE Trans. Commun., vol. 34, pp. 430435,
May 1986.
[21] M. Abramowitz and I. A. Stegun, Handbook of Mathematical Functions
with Formulas, Graphs, and Mathematical Tables. New York: Dover,
9th ed., 1970.
[22] Y. Ma and T. J. Lim, Bit error probability for MDPSK and NCFSK
over arbitrary Rician fading channels, IEEE J. Sel. Areas Commun.,
vol. 18, pp. 21792189, Nov. 2000.
[23] J. G. Proakis, Digital Communications. McGraw-Hill, 3rd ed., 1995.
[24] X. Cai and G. B. Giannakis, Adaptive PSAM accounting for channel
estimation and prediction errors, IEEE Trans. Wireless Commun., vol. 4,
pp. 246256, Jan. 2005.
[25] Y. Ma, Impact of correlated diversity branches in Rician fading
channels, in Proc. IEEE Int. Conf. Commun. (ICC), vol. 1, (Seoul,
Korea), pp. 473477, May 16-20 2005.
[26] E. W. Weisstein, Eigenvalue., MathWorld: A Wolfram Web Resource.
URL: mathworld.wolfram.com/Eigenvalue.html.

A PPENDIX
D ERIVATION OF E QUATION (27)
For the Rayleigh fading with i.n.d. branches, let us
consider the eigenvalue pair
l (n , d) at the lth branch.
For simplicity,
we
suppress
the
subscript
l below. Let


c2
cc d
and QH,2 (n ) =
v (d) =
2
2
cc d c |d| + N0

an 0.5
, where d = (h + jv )a and = R + jI .
0.5
0
Here, R and I are the real and imaginary parts of ,
respectively.
It follows that v (d)QH,2 (n ) =


n ac2 + 0.5 cc d
0.5c2
.
n acc d + 0.5c2 |d|2 +0.5N0 0.5
cc d

c1 c2
For a matrix of the form
we obtain its two
c3 c4
eigenvalues as [26],


1
(36)
(c1 + c4 ) 4c2 c3 + (c1 c4 )2
=
2
Based on this observation, we obtain
(c1 c4 )2 = n2 a2 c4 c2 c2 [(d)]2 + j2n ac3 c (d),
4c2 c3 = 2n ac3 c d + c2 c2 |d|2 + c2 N0 , and c1 + c4 =
(d)cc n ac2 . Applying these equalities to (36), we
obtain the two eigenvalues of v QH,2 (n ) given in (27).

Yao Ma (S98, M01) received the B.Sc. degree

from Anhui University, and M.Sc degree from University of Science and Technology of China (USTC),
China, in 1993 and 1996, respectively, both in electrical engineering and information science; and the
Ph.D degree in Electrical Engineering from National
University of Singapore, in year 2000. His Ph.D thesis was on the diversity reception over fading channels and CDMA multiuser detection. From April
2000 to July 2001, he was a Member of Technical
Staff at the Centre for Wireless Communications,
Singapore. From July 2001 to July 2002, he was a Post-doctorate Fellow at
the ECE Department of the University of Toronto. Since August 2002, he
has been an Assistant Professor at the Electrical & Computer Engineering
Department of the Iowa State University.
His research interests include the digital communication over fading channels, estimation and multiuser detection, adaptive filtering, MIMO systems,
and UWB communication. He is currently an Editor for the IEEE Transactions
on Wireless Communications and an Associate Editor for IEEE Transactions
on Vehicular Technology.
Robert Schober (S98, M01) was born in Neuendettelsau, Germany, in 1971. He received the Diplom
(Univ.) and the Ph.D. degrees in electrical engineering from the University of ErlangenNurnberg
in 1997 and 2000, respectively. From May 2001
to April 2002 he was a Postdoctoral Fellow at
the University of Toronto, Canada, sponsored by
the German Academic Exchange Service (DAAD).
Since May 2002 he has been an Assistant Professor and Canada Research Chair (Tier II) in Wireless Communication at the University of British
Columbia (UBC), Vancouver, Canada. His research interests include cooperative diversity systems, equalization, UWB, MIMO systems, and spacetime
processing and coding.
In 2001, Dr. Schober was a corecipient of the best paper award of the
German Information Technology Society (ITG). In 2002, he received the
Heinz MaierLeibnitz Award of the German Science Foundation (DFG). He
also received the 2004 Innovations Award of the Vodafone Foundation for
Research in Mobile Communications. He served as a Guest Editor for the
IEEE Journal on Selected Areas in Communications (JSAC) and he is an
Editor for Detection, Equalization, and MIMO for the IEEE Transactions on
Communications.
Dongbo Zhang (S00) was born in Xian, China,
in October 1980. In July 2002, he received the B.E.
degree in electrical engineering from Xian Jiaotong
University, Xian, China. Since August 2002 he
has been a Ph.D candidate in the Department of
Electrical and Computer Engineering, Iowa State
University, Ames, Iowa. His research interests include fading channel generation and analysis for
wireless communications, multi-carrier multi-user
communications, MIMO systems and wireless sensor networks. Mr. Zhang is a member of the Phi
Kappa Phi National Honor Society.