Você está na página 1de 5

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 57, NO.

8, AUGUST 2009 2193

A Simple Bit Error Probability Analysis for


Square QAM in Rayleigh Fading with Channel Estimation
Hua Fu and Pooi Yuen Kam, Senior Member, IEEE

Abstract—A new approach is presented for analyzing the form in complex Gaussian random variables is less than zero.
bit error probability (BEP) of square, multilevel, quadrature However, evaluating the various coefficients leads to much
amplitude modulation over a nonselective Rayleigh fading chan- numerical computations. Approximate BEP’s with adaptive
nel, with imperfect channel estimation employing pilot-symbol-
assisted-modulation. It is much simpler and more powerful than PSAM schemes are given in [9]. The moment generating
those in the literature, and the average BEP is obtained by function (MGF) approach of [14] is used in [10], but the
calculating the BEP for each individual bit. The results are given results are not explicit. Using the PDF of the decision statistics
in simple, exact, closed-form expressions that do not require and the equivalent noise [11], the SER of a two-dimensional
any numerical integration. These expressions show explicitly the signal with channel estimation is analyzed in [12], which
behavior of the BEP as a function of various system parameters.
Three channel estimation schemes are investigated. It is shown requires two-fold numerical integral evaluation [12, eqs.(17)-
that existing channel estimation schemes using sinc interpolation (18)]. Thus, in general, the results available in the literature are
and Gaussian interpolation can be improved. not simple, and one cannot easily see the BEP behavior as an
Index Terms—QAM, BEP, MSB, LSB, channel estimation, explicit function of the system parameters. This paper provides
PSAM, Rayleigh fading. a new approach for computing the BEP, following the idea in
[15] and [16], i.e., the true fading and the estimated fading are
jointly complex Gaussian, and conditioned on the estimated
I. I NTRODUCTION fading, the true fading is Gaussian. Specifically, 16-QAM is

T HE effect of channel estimation error of pilot-symbol-


assisted-modulation (PSAM) scheme on the performance
of square, quadrature amplitude modulation (QAM) over
chosen as the example, and the explicit BEP’s for both the
individual bits and the average BEP of all the bits are derived.
The results are given in exact, simple, closed-form expressions
nonselective Rayleigh fading channel with additive, white, without requiring any numerical integration. The approach
Gaussian noise (AWGN) is a subject of considerable research is very general, and can easily be applied to any arbitrary
[1]−[12]. The optimum Wiener interpolator and an upper square QAM. It exploits the Gaussian statistics of the inphase
bound on symbol error rate (SER) for 16-QAM were derived and quadrature phase components of the channel estimation
in [1]. The solution requires the inversion of the measure- errors, and the fact that these errors can be combined with the
ment covariance matrix, which is computationally intensive in AWGN to form an effective additive Gaussian noise. Although
practice if done on-line and if the pilot number is high. To the studies in [6]−[12] consider more general scenarios, their
avoid this direct matrix inversion, Gaussian interpolation and results do not readily specialize to ours.
sinc interpolation are proposed in [2] and [3], respectively. We consider three channel estimation schemes, and since
Recently, the bit error probability (BEP), which is the most their physical insights are distinct from one another, the BEP
useful performance metric in most cases, is analyzed in is derived differently for each one. First, the optimal maxi-
[4]−[10] for various fading scenarios. Generally, the BEP is mum a posteriori probability (MAP) estimator is studied. To
calculated in [4] and [5] by averaging the conditional BEP reduce the computational load, the suboptimum Gaussian/sinc
for the AWGN channel over the fading represented by the interpolator is employed. However, unlike previous works, we
joint probability density function (PDF) of the true fading show that the Gaussian/sinc interpolator can be improved if
amplitude, the estimated fading amplitude, the true fading the interpolated fading estimate is weighted according to the
phase and the estimated fading phase [5, eqs.(25)-(26)]. This channel condition, and this forms our second scheme. The
approach, however, leads to a laborious derivation. Moreover, third scheme considered is the same as that in [4, 5]. In Section
the BEP derived in [4] and [5] is expressed as the weighted II, the system model and PSAM scheme are introduced. The
sum of a number of single and double integrals which require BEP analysis is presented in Section III. Section IV gives
numerical integration [4, eq.(41)], [5, eq.(19)]. The approach numerical and simulation results. Overhead ∼ and superscript
in [6]−[8] relies on making use of the derivation in [13, App. ∗ denote a complex quantity and its conjugate, respectively, E
B] concerned with the probability that a general quadratic is the ensemble average operator, δ is the Kronecker delta, and
[·]T and [·]H denote the matrix transposition and Hermitian
Paper approved by R. Mallik, the Editor for Diversity and Fading Channels transpose, respectively.
of the IEEE Communications Society. Manuscript received October 2, 2007;
revised April 11, 2008 and July 9, 2008.
The authors are with the Department of Electrical and Computer Engineer-
ing, National University of Singapore, Republic of Singapore 117583 (e-mail:
II. S IGNAL M ODEL A ND PSAM S CHEME
{elefh, elekampy}@nus.edu.sg). For square QAM transmission over nonselective, Rayleigh
This paper is presented in part at the 2006 IEEE Global Communications
Conference, San Francisco, USA, November 2006. fading channel with AWGN, the received signal during the kth
Digital Object Identifier 10.1109/TCOMM.2009.08.070507 symbol interval is r̃(k) = s̃(k)c̃(k) + ñ(k) [1]. Here, {s̃(k)}
0090-6778/09$25.00 
c 2009 IEEE

Authorized licensed use limited to: McGill University. Downloaded on October 4, 2009 at 06:36 from IEEE Xplore. Restrictions apply.
2194 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 57, NO. 8, AUGUST 2009

represents the QAM symbols, the fading process {c̃(k)} is a


sequence of zero-mean, complex, Gaussian random variables
with covariance E[c̃(k)c̃∗ (k − j)] = 2R(j), and {ñ(k)} is
zero-mean AWGN with covariance E[ñ(k)ñ∗ (j)] = N0 δkj .
The spectrum of the fading process is even around the carrier
[15, Table 1] so that the inphase component Re[c̃(k)] and the
quadrature phase component Im[c̃(k)] are independent with
the same covariance function R(j). Although {Re[c̃(k)]} is
independent of {Im[c̃(k)]}, the amplitude process {|c̃(k)|} is
dependent on the phase process {∠c̃(k)} [5]. This dependence
not only renders the BEP derived in [4] only approximate, but Fig. 1. Structure of PSAM system.
also makes the analysis in [5], which uses the same approach
as in [4], become more cumbersome [5, eqs.(16)-(19)]. ⎡ 5 ⎤
1+ 2αγb ρ(N ) ···
ρ[(L1 + L2 )N ]
We assume that {s̃(k)}, {c̃(k)} and {ñ(k)} are mutually ⎢
⎢ ..


⎢ 5 ⎥
independent and all symbol points are equally likely. The ⎢ ρ(N ) 1 + 2αγb · · · . ⎥
Ψ=⎢⎢⎢ ⎥
⎥.
average received signal-to-noise ratio (SNR) per bit is defined ⎢ .. .. .. .. ⎥

as γb = 2Eb R(0)/N0 . The fading correlation coefficient ⎢

. . . . ⎥

5
ρ(k) at the matched filter output over a k-symbol interval is ρ[(L1 + L2 )N ] ··· ··· 1 + 2αγb
defined as ρ(k) = R(k)/R(0). Without loss of generality, we Thus, the MAP estimate of the true fading gain c̃(iN + j),
choose 16-QAM as the example for our analysis. The signal denoted by
c̃(iN + j), is
c̃(iN + j) = wop T
z. The

constellation and bit mapping are shown in [4, Fig. 3], where conditional variance σc = E |c̃(iN + j) − c̃(iN + j)|2 z is
2

each signal point is represented by a 4-bit symbol (i1 , q1 , i2 , σc2 = 2R(0) − 2R(0) x Ψ−1 xT . Using the estimated fading
q2 ). The bits i1 and q1 are the most significant bits (MSB), gain
c̃(iN + j), each signal point in [4, Fig. 3] is multiplied
and the bits i2 and q2 the least significant bits (LSB) [4, Fig. by
c̃(iN + j), which amounts to rotating the I-Q-coordinate
5]. The average symbol energy is Es = 10d2 and the average system by the angle ∠
c̃(iN + j) and scaling it by |
c̃(iN + j)|
bit energy is Eb = 5d2 /2. In a PSAM system, known pilot to form a new signal constellation in the I’-Q’-coordinate
symbols are inserted periodically into the data sequence. system, as shown in Fig. 2. The minimum Euclidean distance
Fig. 1 shows the packet frame structure where each frame decision rule [1]−[12] assigns a received signal r̃(k) to the
contains N symbols, with the first symbol being the pilot. symbol in the I’-Q’-coordinate system if and only if r̃(k)
The parameters i and j represent the frame index and symbol is closer in Euclidean distance to this symbol than to any
index within each frame, respectively. Since all frames have other symbol. This decision rule leads to the new detection
the same length N , one has j = 0, 1, · · · , N − 1, where j = 0 boundaries which are determined by rotating the decision
is the pilot symbol. Thus, we have the time index k = iN + j. boundaries in [4, Fig. 3] by the angle ∠
c̃(iN + j) and scaling
As in [4], the fading estimate for the current frame is obtained them by |
c̃(iN + j)|, using
c̃(iN + j). Clearly, in the case of
from L1 + L2 + 1 nearest modulation-free pilot symbols imperfect channel estimation, the knowledge
c̃(iN + j) from
(i.e., L1 previous pilots, one current pilot and L2 subsequent channel estimation cannot pin down c̃(iN + j). Thus, we can
pilots). The known pilot symbol, s̃(k) = p̃, can be any express the true fading gain c̃(iN + j) as the sum of two
signal point in the signal constellation. Thus, the energy of components. One is the estimate
c̃(iN + j) which is known
the pilot symbol is given by |s̃(k)|2 = |p̃|2 = αd2 , where at the receiver due to channel estimation; the other is the
α ∈ {2, 10, 18}. With reference to Fig. 1, the channel fading estimation error ẽ(iN + j) = c̃(iN + j) −
c̃(iN + j), which
at the pilot symbol time (i − L1 )N, · · · , iN, · · · , (i + L2 )N is unknown and independent of
c̃(iN + j). This estimation
is given, after dividing the received signal by the known pilot error can be treated as an additional noise component.
1 )N ]
symbol, by z = c̃[(i − L1 )N ] + ñ[(i−L p̃ · · · c̃(iN ) +
T
ñ(iN )
· · · c̃[(i + L2 )N ] + ñ[(i+L 2 )N ]
. The task of channel III. BEP A NALYSIS
p̃ p̃
estimation is to recover accurately the true fading gain We first evaluate the BEP conditioned on the estimated
c̃(iN + j) for the jth data symbol in the ith packet frame fading gain
c̃(iN + j) = u. The final BEP is obtained by aver-
from measurements z. Using MAP criterion, since c̃(iN + j) aging this conditional BEP over the distribution of
c̃(iN + j).
and z are jointly complex Gaussian, the optimal estimate of Due to the symmetry of the signal constellation in Fig. 2, we
c̃(iN + j) is the conditional mean E[c̃(iN + j)|z], which need only compute the BEP for the inphase component. Since
is a linear minimum mean-square error (MMSE) estimate the MSB i1 and the LSB i2 are equally likely, the average
that is obtained by the Wiener interpolator [1, eq.(14)] with conditional BEP is Pc = 12 (Pc,MSB +Pc,LSB ). We begin with
interpolation coefficients given by wop = Ψ−1 xT [1, eq.(19)], computing Pc,MSB . We assume that i1 = 0. The case that i1 =
where, x = [ρ(L1 N + j) · · · ρ(j) · · · ρ(L2 N − j)] and 1 gives an identical result. It is seen from Fig. 2 that the MSB
i1 = 0 is associated with the symbols 0001, 0011, 0000, 0010,
0100, 0110, 0101 and 0111. Because of the symmetry of the
constellation, we only need consider the symbols 0001, 0011,
0000 and 0010. Conditioned on symbol 0001 being transmit-
ted and the channel estimate
c̃(iN + j) = u, the inphase

Authorized licensed use limited to: McGill University. Downloaded on October 4, 2009 at 06:36 from IEEE Xplore. Restrictions apply.
FU and KAM: A SIMPLE BIT ERROR PROBABILITY ANALYSIS FOR SQUARE QAM IN RAYLEIGH FADING WITH CHANNEL ESTIMATION 2195

   
(5|u|d)2 (|u|d)2
Q 1 2 2 N0 and Pc4,LSB = Q 1 2 2 N0 −
2 σc 10d + 2 2 σc 18d + 2
 
(5|u|d)2
Q 1 2 2 N0 , for the associated symbols 0010 and
2 σc 18d + 2
0011 being transmitted, respectively. The conditional BEP for
i2 = 1 is then equal to 12 (P c3,LSB + Pc4,LSB ). Consequently,
1
we obtain Pc,LSB = 4 Pc1,LSB + Pc2,LSB + Pc3,LSB +

Pc4,LSB .
Next, we average Pc over the distribution of
c̃(iN + j).
Note from Section II that
c̃(iN + j) = wop T
z is a weighted
linear sum of zero-mean, jointly complex Gaussian random
variables. Thus,
c̃(iN + j) is zero-mean, complex Gaussian
with variance R(0)wop T
xT . Moreover, the conditional BEP’s
Pc,MSB , Pc,LSB and Pc depend only on |
c̃(iN + j)|2 which
has a central Chi-square distribution with two degrees of
freedom [13, eq.(2.1-126)]. By averaging Pc,MSB and Pc,LSB
over |
c̃(iN + j)|2 , the final BEP’s for the MSB and the LSB
are given, respectively, by
Pe,MSB =
     
Fig. 2. Rotated signal constellation of 16-QAM. 1 1 1 1
1− + 1−
8 1 + 5+20Aγ
2Bγb
b 8 1 + 5+4Aγ
2Bγb
b

     
1 1 1 1
component of r̃(k) in the new constellation is conditionally + 1− + 1−
Gaussian [16] with mean |u|d and variance 12 σc2 10d2 + N0 /2. 8 1 + 5+36Aγ
18Bγb
b 8 1 + 5+20Aγ
18Bγb
b

Thus, noting that the decision boundary  is Q’, the conditional


 Pe,LSB =
(|u|d)2      
BEP is given by Pc1,MSB = Q 1 2 2 . Using 1 1 1 1
2 σc 10d +N0 /2
1− + 1−
similar
 arguments, we
 have the conditional BEP’s P2c2,MSB  = 4 1 + 5+20Aγ b 8 1 + 5+4Aγ b
(|u|d)2 (|u|3d) 2Bγb 2Bγb
Q 1 2 2 , P c3,MSB = Q 1 2 2 and      
2 σc 2d +N 0 /2  2 σc 18d +N0 /2
(|u|3d) 2 1 1 1 1
Pc4,MSB = Q , for symbols 0000, 0011, + 1− + 1−
8 1 + 5+20Aγ 8 1 + 5+4Aγ
1 2 2
2 σc 10d +N0 /2
b b
18Bγb 18Bγb
0010 being transmitted, respectively, where Q(x) is the Gaus-      
sian Q-function [13, eq.(2.1-97)]. Consequently, we have 1 1 1 1
Pc,MSB = 14 (Pc1,MSB + Pc2,MSB + Pc3,MSB + Pc4,MSB ). + 1− − 1−
8 1 + 5+36Aγ
2Bγb
b 8 1 + 5+20Aγ
50Bγb
b

Next, we compute Pc,LSB . Unlike the MSB i1 , the conditional   


BEP for the LSB i2 = 0 is different from that for i2 = 1. First, 1 1
− 1− (1)
assume that i2 = 0. Again, due to the symmetry, we only need 8 5+36Aγb
1 + 50Bγb
consider the associated symbols 0000 and 0001. Conditioned
on symbol 0000 being transmitted and
c̃(iN + j) = u, the where B = xwop and A = 1 − B. The average BEP for all
inphase component of r̃(k) is conditionally Gaussian with bits is Pe = 12 (Pe,MSB + Pe,LSB ). The closed-form results
mean |u|d and variance 12 σc2 2d2 + N0 /2. Now, the decision of Pe,MSB , Pe,LSB and Pe show the behavior of the BEP as
 an explicit function of the fading correlation coefficient ρ(k),
region for i2 = 0 is over the interval − 2d|u|, 2d|u| along the average received SNR per data bit γb , and the quantity
the I’ axis. Thus,  the conditional  BEP,  Pc1,LSB , is obtained
 α ∈ {2, 10, 18} which indicates the pilot used. No numerical
(3|u|d)2 (|u|d)2 integration is needed for their actual evaluation. Since the
as Pc1,LSB = Q 1 2 2 N0 +Q 1 2 2 N0 . Sim-
2 σc 2d + 2 2 σc 2d + 2
average data energy per frame is (N − 1)Es and the pilot
ilarly, the conditional BEP, Pc2,LSB , for symbol
 0001 being
 symbol energy in a frame is αd2 , the total energy over a frame
(3|u|d)2
transmitted is obtained as Pc2,LSB = Q 1 2 2 N0 + is Etotal = (N − 1)Es + αd2 . With 4 bits per data symbol and
2 σc 10d + 2
  N −1 data symbols in a frame, we define the average effective
(|u|d)2 10(N −1)Eb +αEb
Q 1 2 2 N0 . Accordingly, the conditional BEP for the data energy per bit as Eb,ef f = 4(N −1) =
Etotal
10(N −1)
2 σc 10d + 2
and the average effective received SNR per data bit as γb,ef f =
LSB i2 = 0 is equal to 12 (Pc1,LSB + Pc2,LSB ). The procedure 10(N −1)+α
for obtaining the conditional BEP for i2 = 1 is parallel to 10(N −1) γb . Thus, replacing γb by γb,ef f , we can obtain an
alternative set of BEP results which depend on γb,ef f instead
that followed in the case for i2 = 0. However, note that in
of γb .
this case the decision
  region for  i2 = 1 is over the interval Computing Ψ−1 in real-time to obtain wop is computa-
− ∞, −2d|u| ∪ 2d|u|, +∞ along the I’ axis. Thus, we tionally intensive if the number of pilot samples is large.
 
(|u|d)2 Thus, efficient but suboptimum interpolation approaches such
have the conditional BEP’s, Pc3,LSB = Q N0 −
1 2 2
2 σc 10d + 2 as the sinc and Gaussian interpolators [2, 3] are desirable.

Authorized licensed use limited to: McGill University. Downloaded on October 4, 2009 at 06:36 from IEEE Xplore. Restrictions apply.
2196 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 57, NO. 8, AUGUST 2009

0
10
We study here the BEP performance of these suboptimum
interpolators. Recall that the vector z is the channel measure- (a) (b)
(c)
ments through PSAM. Given z, the suboptimum interpolators (d) (e) (f)

first produce
 a new observation quantity ỹ = wT z, where −1
10
T
w = w[(i − L1 )N ] · · · w(iN ) · · · w[(i + L2 )N ] are real-
valued interpolation coefficients given by [3, eq.(3)] and [2,
eqs.(12)−(15)] for the sinc and low-order Gaussian interpola-

BEP
−2 (a) LSB, Gaussian, j=7
10
tors, respectively. The interpolation coefficients are fixed since (b) LSB, Sinc, j=7
(c) MSB, Gaussian, j=7
they do not depend on the channel parameters such as ρ(k) and (d) average, Wiener, j=7
(e) LSB, Wiener, j=7
γb . We want an estimate of c̃(iN +j) based on the observation (f) MSB, Sinc, j=7
(g) LSB, Wiener, j=1 (g)
ỹ. Since c̃(iN + j) and z are jointly complex Gaussian and −3
10
(h) average, Wiener, j=1
(i) MSB, Wiener, j=1 (i)
ỹ is a linear transformation of z, c̃(iN + j) and ỹ are also (k) MSB, Wiener, j=7
∗ simulation
(h)
(k)
jointly complex Gaussian, both with mean zero. The variance
of ỹ, Σ(0), can be obtained as Σ(0) = R(0)wT Ψw, and
the covariance coefficient between ỹ and c̃(iN + j), denoted
−4
10
 T
0 5 10 15
γb (dB)
20 25 30

wT R(L1 N +j) ··· R(L2 N −j)


as λ, can be obtained as λ = √ √ .
R(0) Σ(0)
Now, conditioned on ỹ = v,  c̃(iN + j) is conditionally Fig. 3. BEP performance comparison between the Wiener, sinc and Gaussian
interpolators.
Gaussian with mean u = λ R(0)/Σ(0)v and variance
ηc2 = R(0)(1 − λ2 ) [16]. This result states that, conditioned
⎡ ⎤ ⎡ ⎤
on ỹ = v, the conditionally Gaussian PDF with mean u and  
⎢  1 ⎥ ⎢  1 ⎥
variance ηc2 contains all the available information concerning 
⎣1 − 

⎦ + ⎣1 −  ⎦+
5+4γb (1−C 2 ) 5+20γb (1−C 2 )
the true fading gain c̃(iN + j). Thus, the optimal estimate 1+ √
2γb (2 D−C )
2 1+ √
2γb (2 D+C )
2

of c̃(iN + j) should be the value of c̃(iN + j) that has the ⎡ ⎤ ⎡ ⎤


 
highest conditional probability, namely, the conditional mean ⎢  1 ⎥ ⎢  1 ⎥
 
u = λ R(0)/Σ(0)ỹ, rather than ỹ itself. To derive the BEP, ⎣1 −  5+20γb (1−C 2 )
⎦ + ⎣1 −  5+20γb (1−C 2 )
⎦−
1+ √ 2 1+ √ 2
exactly the same procedure is used as that for the Wiener 2γb (2 D−C ) 2γb (3C−2 D )
⎡ ⎤ ⎡ ⎤
interpolator. Conditioned on ỹ = v, the receiver rotates the  
⎢  1 ⎥ ⎢  1 ⎥
signal constellation [4, Fig. 3] by the angle ∠u , and scales 
⎣1 − 

2 ) ⎦ + ⎣1 −  ⎦
5+36γb (1−C 2 )
it by the magnitude |u | to construct a new constellation, as 1 + 5+20γb (1−C
√ 2 1 + √ 2
2γb (3C+2 D) 2γb (3C−2 D )
shown in Fig. 2. It is easy to show that in this case the ⎡ ⎤⎫
   ⎪

MSB BEP Pe,MSB , the LSB BEP Pe,LSB and the average ⎢  1 ⎥

BEP Pe of all bits are identical to Pe,MSB , Pe,LSB and Pe , − ⎣1 − 
 ⎦ (2)
1+ 5+36γb (1−C 2 )
√ 2


respectively, except that A and B are now given, respectively, 2γb (3C+2 D )

by A = 1 − C 2 and B = C 2 , where C = √1D xw


Finally, we have the average BEP Pe = 12 (Pe,MSB

+
with D = wT Ψw. Hence, the BEP performance for the   
Pe,LSB ). Like Pe , we see that Pe depends on ρ(k), γb and
suboptimum interpolators depends not only on ρ(k) and γb ,
w. Since the optimal estimate of the true fading gain is not
 w.
but also on the interpolation coefficients
ỹ, we expect that Pe is worse than Pe .
Evaluating the conditional mean λ R(0)/Σ(0)ỹ requires
prior knowledge of the fading correlation cofficients and the
average received SNR. To further simplify the estimation IV. N UMERICAL A ND S IMULATION R ESULTS
procedure, we can merely choose ỹ as the estimated value This section presents numerical/simulation results to il-
of the true fading gain. This scheme was adopted in [4] and lustrate the BEP behavior for the three estimation schemes
[5]. Now, conditioned on ỹ = v, a new signal constellation is as a function of the system parameters, namely, the fading
formed through rotating [4, Fig. 3] by ∠v and scaling it by |v|. correlation coefficient, the average received SNR and the in-
From Fig. 2, we see that the decision boundary for the MSB terpolation coefficients. The mobile radio channel with Jake’s
i1 is Q’, and the decision rule depends on whether Re[r̃(k)] model is used, i.e., we assume ρ(k) = J0 (2πfd T ), where

was greater or less than zero. Thus, the BEP Pe,MSB for the J0 (·) is the zeroth-order Bessel function of the first kind,

MSB is identical to Pe,MSB . As for the LSB i2 , we note that and fd is the Doppler frequency. The sinc interpolator [3,
 for i2 = 0 and i2 = 1 is over the intervals
the decision region eq.(3)] with Hamming window and the second-order Gaussian
− 2d|v|, 2d|v| and − ∞, −2d|v| ∪ 2d|v|, +∞ of the I’ interpolator [2, eqs.(12b)−(12d)] are assumed. In order to
axis, respectively. Following the derivation procedure similar validate the analysis, computer simulation results are included.
to that in the second scheme, the BEP for the LSB is obtained The simulation results are shown as the star markers in the
as figures. The simulation results confirm the feasibility and
⎧⎡ ⎤ accuracy of the theoretical analysis.
⎪  In Fig. 3, we plot the BEP performance for the MSB
 1 ⎨⎢ 
 1 ⎥
Pe,LSB = ⎣1 −  ⎦+ and LSB with the Wiener, sinc and Gaussian interpolators
8⎪⎩ 1+ 5+4γb (1−C 2 )

2γb (2 D+C )
2
against γb with Doppler spread fd T = 0.02. For sinc/Gaussian

Authorized licensed use limited to: McGill University. Downloaded on October 4, 2009 at 06:36 from IEEE Xplore. Restrictions apply.
FU and KAM: A SIMPLE BIT ERROR PROBABILITY ANALYSIS FOR SQUARE QAM IN RAYLEIGH FADING WITH CHANNEL ESTIMATION 2197

0
10
Our extensive study of numerical results also shows that
(a) for a slower fading rate, the MAP estimator can perform well
for a wide range of values of the frame size. As the fading
(b)
rate increases, the BEP performance degrades increasingly
−1
10 (c) when the frame size increases. For a fixed fading rate and
a fixed average effective received SNR, the frame size has an
(d)
optimum value at which the average BEP of all data symbol
BEP

locations is minimized. It is seen from γb,ef f = 10(N −1)+α


10(N −1) γb
that for fixed α and γb,ef f , increasing the frame size N
−2
10 (a) α=10, Sinc implies that more energy is devoted to the data. This improves
(b) α=18, Sinc
(c) α=10, improved Sinc the BEP performance. On the other hand, as N increases,
(d) α=18, improved Sinc
∗ Simulation
ρ(k) decreases, which leads to the degradation of the channel
estimation accuracy. The tradeoff between these two effects
gives the optimum value of the frame size. Similar observation
−3
10
0 5 10 15 20 25 30
on the SER bound performance is presented in [1].
γb (dB)

R EFERENCES
Fig. 4. LSB BEP performance comparison for the sinc interpolator and the
improved sinc interpolator. [1] J. K. Cavers, “An analysis of pilot symbol assisted modulation for
Rayleigh fading channels," IEEE Trans. Veh. Technol., vol. 40, pp. 686-
693, Nov. 1991.
[2] S. Sampei and T. Sunaga, “Rayleigh fading compensation for QAM in
interpolation, the second estimation scheme is used. We as- land mobile radio communications," IEEE Trans. Veh. Technol., vol. 42,
pp. 137-146, May 1993.
sume that N = 15, α = 10 and L1 = L2 = 1. Two [3] Y. S. Kim, C. J. Kim, G. Y. Jeong, Y. J. Bang, H. K. Park, and S.
data symbols, j = 1 and j = 7 in the current frame, are S. Choi, “New Rayleigh fading channel estimator based on PSAM
considered. From Fig. 3, several conclusions can be drawn. channel sounding technique," in Proc. IEEE Int. Conf. Commun. (ICC),
Montreal, Canada, pp. 1518-1520, June 1997.
First, the MSB and the LSB have unequal BEP. This means [4] X. Y. Tang, M. S. Alouini, and A. J. Goldsmith, “Effect of channel
that the estimation errors affect the performance differently estimation error on M-QAM BER performance in Rayleigh fading,"
for different transmitted bits. Thus, using the MSB which has IEEE Trans. Commun., vol. 47, pp. 1856-1864, Dec. 1999.
[5] L. Z. Cao and N. C. Beaulieu, “Exact error-rate analysis of diversity
higher error protection to convey more important information 16-QAM with channel estimation error," IEEE Trans. Commun., vol.
can improve communication reliability. Second, the effect of 52, pp. 1019-1029, June 2004.
the estimation errors depends strongly on the data symbol [6] L. Z. Cao and N. C. Beaulieu, “Closed-form BER results for MRC
diversity with channel estimation errors in Ricean fading channels,"
location. Third, imperfect channel estimation has led to the IEEE Trans. Wireless Commun., vol. 4, pp. 1440-1447, July 2005.
expected irreducible error-rate floor. Finally, the figure clearly [7] L. Najafizadeh and C. Tellambura, “BER analysis of arbitrary QAM for
shows the performance degradation due to the suboptimum MRC diversity with imperfect channel estimation in generalized Ricean
fading channels," IEEE Trans. Veh. Technol., vol. 55, pp. 1239-1248,
sinc/Gaussian interpolators although they are computationally July 2006.
efficient. [8] B. Xia and J. Wang, “Effect of channel estimation error on QAM
As mentioned in Section III, with sinc/Gaussian interpola- systems with antenna diversity," IEEE Trans. Commun., vol. 53, pp.
481-488, Mar. 2005.
tion, the BEP performance Pe = 12 (Pe,MSB  
+ Pe,LSB ) of [9] X. Cai and G. B. Giannalis, “Adaptive PSAM accounting for channel
 1  
the third scheme is worse than Pe = 2 (Pe,MSB + Pe,LSB ) of estimation and prediction errors," IEEE Trans. Wireless Commun., vol.
the second scheme. The reason is that given ỹ, the quantity 4, pp. 246-256, Jan. 2005.
 [10] Y. Ma, R. Schober, and D. Zhang, “Exact BERs for M-QAM with
λ R(0)/Σ(0)ỹ rather than ỹ provides the optimal estimate MRC and channel estimation errors in Rician channels," in Proc. IEEE
of the true fading gain. This is illustrated in Fig. 4, with Wireless Commun. Networking Conf. (WCNC), New Orleans, USA, pp.
fd T = 0.02, N = 15, L1 = L2 = 1 and j = 1. The sinc 967-972, Mar. 2005.
  [11] S. K. Wilson and J. M. Cioffi, “Probability density functions for ana-
interpolator is used. Since Pe,MSB is identical to Pe,MSB , Fig. lyzing multi-amplitude constellations in Rayleigh and Ricean channels,"
 
4 gives only the comparison between Pe,LSB and Pe,LSB . To IEEE Trans. Commun., vol. 47, pp. 380-386, Mar. 1999.
further verify the analysis, the BEP Pe,LSB
is also computed [12] X. D. Dong and L. Xiao, “Symbol error probability of two-dimensional
signaling in Ricean fading with imperfect channel estimation," IEEE
by using [5, eqs.(14)−(19)], and it is found that the resulting Trans. Veh. Technol., vol. 54, pp. 538-549, Mar. 2005.
numerical plot overlaps with that obtained by using (2). The [13] J. G. Proakis, Digital Communications, 4th ed. New York: McGraw-Hill,
results in Fig. 4 confirm that the channel estimator employed 2001.
[14] M. K. Simon and M. S. Alouini, Digital Communication over Fading
in [2]−[5] can be improved if the interpolated fading estimate Channels, 2nd ed. New York: John Wiley & Sons, 2005.
is weighted according to the channel statistics ρ(k) and the [15] P. Y. Kam, “Bit error probabilities of MDPSK over the nonselective
SNR γb . We also note from the figure that increasing the pilot Rayleigh fading channel with diversity reception," IEEE Trans. Com-
mun., vol. 39, pp. 220-224, Feb. 1991.
power (measured by α) can help reduce the estimation inac- [16] P. Y. Kam, “Optimal detection of digital data over the nonselective
curacy of the channel fading gain due to the noise component, Rayleigh fading channel with diversity reception," IEEE Trans. Com-
and thus, improve the BEP performance. mun., vol. 39, pp. 214-219, Feb. 1991.

Authorized licensed use limited to: McGill University. Downloaded on October 4, 2009 at 06:36 from IEEE Xplore. Restrictions apply.

Você também pode gostar