Makings of The 4 × 4 Space-Time Block Code of Choice

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 6

2012 IEEE Wireless Communications and Networking Conference: PHY and Fundamentals

Makings of the 4 × 4 Space-Time Block Code of Choice


Sandipan Kundu, Dimitris A. Pados, Weifeng Su Rohan Grover
Dept. of Electrical Engineering Mirics
State University of New York at Buffalo, NY 14260 Needham, MA 02494
E-mail: {skundu,pados,weifeng}@buffalo.edu E-mail: rohan.grover@mirics.com
Abstract—We present a new 4 × 4 Hadamard-precoded Quasi- The rest of the paper is organized as follows. In section II,
Orthogonal Space-Time Block Code (QO-STBC) that enables the system model for the proposed codeword is discussed. In
reliability-based prioritized symbol detection using linear filters. section III, we discuss the proposed optimal and sub-optimal
Approximate block-error-rate analysis is carried out and used to
optimize the code rotation angles. For benchmarking purposes, detectors. Performance analysis for the sub-optimal detector
the maximum-likelihood (ML) code detector is also derived in shown section IV. Complexity and error rate performance
with complexity of the order of joint two-real-symbol decoding. comparisons are being done in section V. Some concluding
Numerical and simulation studies compare linear filter decoding remarks are given in section VI.
against ML decoding.
Index Terms—Maximum-likelihood detection, matched-filter II. S YSTEM M ODEL
(MF), minimum-mean-square-error (MMSE) filter, multi-input
multi-output (MIMO) communications, quasi-orthogonal space- Throughout the paper, we use the following notation: E{·}
time block codes (QO-STBC), space-time block codes (STBC). denotes statistical expectation; (B)−1 , kBk2F , BT , BH , denote
the inverse, Frobenius norm, transpose, and Hermitian of
I. I NTRODUCTION matrix B; IL , 0L is the identity and all zero matrix of size
L; (·)∗ is the complex conjugation operator.
Orthogonal space-time block codes (O-STBC) [1] attain
We consider Mt = 4 transmit antennas, Mr = 11 receive
full transmit diversity and enable easily realizable independent
antenna, and T = 4 time slots over which the space-time code
single-complex-symbol maximum-likelihood (ML) decoding.
is transmitted. We assume perfect channel state information
The first-in-line celebrated 2-transmit antenna Alamouti code
(CSI) at the receiver and no CSI at the transmitter. The
[2] has also rate one and ushered wireless communications in
independent symbols to be transmitted are formed by mapping
the era of space-time coded transmission. What would the 4-
the incoming bits onto a known unit energy signal constellation
transmit-antenna code of choice be? 2-transmit-antenna (and
A to form symbols z1 , · · · , z4 which are represented as a
higher) O-STBCs loose their rate-one attribute. The Jafarkhani
column vector z. The symbol vector z is precoded by the
[3] and Tirkkonen-Boariu-Hottinen [4] quasi-orthogonal (QO)
code is of rate one but at the expense of diversity loss and
Hadamard matrix· V of size¸ 4 followed by the diagonal rotation
I2 02
increased decoding complexity. Rate-one, full-diversity QO- matrix Pθ , 02 ejθ I2
operator to form the final precoded
STBCs for 4-transmit-antennas with joint two-complex-symbol symbol vector x with elements x1 , · · · , x4 , x = Pθ V
2 z where
and single-complex-symbol or joint two-real-symbol ML de- 2 is the energy normalization constant and θ is the rotation
coding were proposed in [5],[6] and [7]-[10], respectively. angle. The space-time block code of interest with symbols to
Arguably, the practical significance of a high-performing be transmitted has the structure
low-computational complexity 4 × 4 space-time code warrants  
further research. Recently, several general proposals appeared x1 x2 x3 x4
in the literature on linear (or partial-interference-cancelation  −x∗ x∗1 −x∗4 x∗3 
X= x2 x4 x1 x2  . (1)
3
(PIC)) receivers of space-time codes [11]-[13]. −x∗4 x∗3 −x∗2 x∗1
In this paper we develop a new Hadamard precoded 4 × 4
QO-STBC codeword that, unlike codewords from [7]-[10], fa- The received signal y of size T × Mr is given by
cilitates reliability-based linear-complexity sub-optimal decod- r
ing using minimum-mean-square-error (MMSE) and matched ρ
y= Xh + n (2)
filters (MF). Precoded codewords and linear receivers were also Mt
pursued by Sezgin and Oechtering [14] with resulting error where ρ is the SNR per receive antenna, h is the channel
rates that may be deemed unsatisfactory. A semi-closed-form vector with elements h1 , · · · , h4 and n is the noise vector.
approximate block-error-rate expression is also presented for The channel is assumed to quasi-static with Rayleigh faded
the MMSE-MF decoder and minimized to find the optimal coefficients. The elements of h and n, hi , nl , 1 ≤ i ≤ Mt ,
codeword rotation angle. As a side benchmarking result we 1 ≤ l ≤ T , respectively, are being modeled as independent
derive also the exact ML decoder for the proposed Hadamard- and identically distributed complex Gaussian random variables
precoded 4 × 4 QO-STBC which has joint two-real-symbol de- with zero mean and unit variance. The channel-equivalent
coding form. The diversity-product maximizing rotation angle
is found. Simulation results are provided to compare the error 1 For ease in presentation we set the number of receive antennas to one; the
rate of the proposed code under ML and MMSE-MF decoding. treatment is directly extendable to more than one receive antenna.

978-1-4673-0437-5/12/$31.00 ©2012 IEEE 618


representation of (2) as shown in [6], [14] (see also [15]) is Thus, ML decoding of (3) requires joint decoding of the real
given by and imaginary part of the two different symbols (i.e. two-real-
r symbol decoding) resulting in same decoding complexity as in
ρ
ye = HVz + ne (3) [7]-[10]. The proposed codeword achieves full-diversity under
4Mt
ML decoding. The proof is omitted herein due to lack of space.
where ye , ne are obtained by conjugating the second
B. MMSE-MF Decoding
 fourth element of y and n respectively, and H ,
and
h1 h2 h3 h4 Below we derive and describe in detail a two-stage linear-
 h∗2 −h∗1 h∗4 −h∗3 filter decoding approach where two symbols are MMSE de-
 h3 h4 h1
P is the equivalent channel matrix.
h2  θ
h∗4 −h∗3 h∗2 −h1∗ coded and removed and then the two remaining symbols are
The Hadamard precoding enables us to propose two low- MF decoded.
complexity detectors as discussed in the following section. Hadamard precoding bestows different signal-to-
interference-plus-noise ratio (SINR) at the output of the
III. D ECODERS MMSE filter to different symbols. The symbols having higher
In this section we derive and discuss the proposed ML SINR (reliability) are to be decoded first and removed from
joint two-real-symbol decoder and the sub-optimal MMSE-MF the linearized received signal, ye . The remaining symbols are
decoder. to be decoded at the second stage with simple channel MF
filters. The first step of the decoder is to determine which of
A. Joint two-real-symbol ML Decoding the symbols will have higher SINR and will be decoded at
With knowledge of the CSI and the fixed precoder, V, the the first stage with the MMSE filter. To answer the question,
ML decoder of z requires 4 independent joint two-real symbol we will calculate the post MMSE filtering SINR and we
decoding as opposed to joint decoding of 4 complex symbols: will show that a simple channel parameter can be used to
r determine the symbols having higher SINRs.
ρ
ẑM L = arg min kye − HVzk22 For the received signal model in (3) the MMSE filter (in
z∈A 4Mt matrix form) is given by
·
= arg min −2 Re(yeH HVz)
r ´−1
ρ ³ ρ
z∈A WL = (HV) I4 + VH HH HV .
r ¸ 4Mt 4Mt
ρ H H H
+ (z V H HVz) . (4)
4Mt The MMSE receiver output is
The cross-interference among the symbols in z can be under- H
ẑ = WL ye .
stood by simplifying the second term in (4):
· ¸ The post-filtering SINRs are
(a + b cos θ)I2 −jb sin θ I2
(zH VH HH HVz) = zH z 1
jb sin θ I2 (a − b cos θ)I2 SINRℓ = ·³ ´−1 ¸ − 1, (7)
· ¸ ρ H H
(a + b cos θ) −jb sin θ I4 + 4Mt V H HV
= zH z +
13 jb sin θ (a − b cos θ) 13 ℓ,ℓ
· ¸ ℓ = 1, · · · , 4.
(a + b cos θ) −jb sin θ
zH z (5)
24 jb sin θ (a − b cos θ) 24 Simplifying (7),
4 ³ ´2 ³ ´2
ρ
where z13 , [z1 z3 ]T , z24 , [z2 z4 ]T and a , |hi |2 , b , Mt a+ 1 − Mρt b
P
i=1 SINRℓ = ³ ´ − 1, ℓ = 1, 2, (8)
2Re(h1 h∗3 + h2 h∗4 ) are all known to the receiver. As the two ρ
M t a − b cos(θ) + 1
terms in (5) are identical, we will expand the first only to study ³ ´2 ³ ´2
ρ
the inter-symbol-interference: Mt a + 1 − Mρt b
· ¸ SINRℓ = ³ ´ − 1, ℓ = 3, 4. (9)
ρ
H (a + b cos θ) −jb sin θ a + b cos(θ) + 1
z13 z = |z1 |2 (a + b cos θ) + Mt
jb sin θ (a − b cos θ) 13
|z3 |2 (a − b cos θ) + 2b sin θ(z1R z3I − z1I z3R ). (6) Hence, by the Hadamard precoding, one symbol in each
pair (z1 , z3 ) and (z2 , z4 ) has higher SINR than the other.
As shown above, inter-symbol-interference exists between From (8), (9), we conclude that out of the 4 symbols in a
symbol pairs (z1 , z3 ) and (z2 , z4 ), while the two groups of codeword, 2 of the symbols have high SINR and the other
symbols are orthogonal to each other. Furthermore in each pair two have lower SINR, depending on the channel parameter
the real and complex part of the symbols do not interfere with b. If b ≥ 0, symbols z1 , z2 have higher SINR than z3 , z4
each other. For the pair (z1 , z3 ) interference exists between real and are decoded first. If b < 0, symbols z3 , z4 have higher
symbols (z1R , z3I ) and (z1I , z3R ); for the pair (z2 , z4 ) inter- SINR than z1 , z2 and are decoded first. The proposed MMSE-
ference exists between real symbols (z2R , z4I ) and (z2I , z4R ). MF decoder is summarized below where V[i,j] denotes

619
the submatix of V consisting of the i and j columns and for illustration purposes only. Other constellations can be easily
U(·) is the symbol decision operator after linear filtering. covered as well (for example phase-shift-keying). For a square
————————————————————————————– QAM constellation of size |A| = 2f where f is even, the
MMSE-MF Decoder
————————————————————————————–
1. Compute sign of b , 2Re(h1 h∗3 + h2 h∗4 ). instantaneous symbol error probability is
2. If b ≥ 0 p p
(i) (ẑ1 , ẑ2 ) = U(W H
q L ye );
PL1CSI
= 4µQ( δ(γ̃L1 + 1)) − 4µ2 Q2 ( δ(γ̃L1 + 1))
′ ρ
(ii) ye = ye − 4Mt (HV) [ẑ1 ẑ2 0 0]T ;
µ ¶ where µ = 1 − √1 , δ = 3
|A|−1 and Q(·) is the Gaussian
H ′ |A|
(iii) (ẑ3 , ẑ4 ) = U (V)H[3,4] H ye . Q-function. Using the notation
else
ηU τU ηU τU
" #
H
Z Z Z Z
(i) (ẑ3 , ẑ4 ) = U(W
q L ye );
′ ρ
ǫ1 − ǫ2 F (ε)dε , ǫ1 F (ε)dε − ǫ2 F (ε)dε (12)
(ii) ye = ye − 4Mt (HV) [0 0 ẑ3 ẑ4 ]T ; ηL τL ηL τL
µ ¶
H ′
(iii) (ẑ1 , ẑ2 ) = U (V)H[1,2] H ye . and integral representation of Q-function, we can further sim-
end plify to
———————————————————————————— " π π
#
4µ2
Z Z
4µ 2 4
IV. P ERFORMANCE A NALYSIS CSI
PL1 = − e−α1 (φ) dφ (13)
π 0 π 0
In this section, we provide approximate block-error-rate
analysis of the MMSE-MF decoder of the proposed codeword L1 +1)
where α1 (φ) , δ(γ̃
2sin2 φ . We now try to find the average
in (1) with the aim of finding the optimal (minimum block-
symbol error probability over all channel realizations
error-rate) rotation angle in Pθ .
We calculate the probability of error for each of the symbol
Z
CSI
PL1 = PL1 f (h)dh (14)
pairs (z1 , z3 ) and (z2 , z4 ) and then the error rate for the whole
codeword. The error rate is calculated stage-wise rather than
symbol-wise. As both the pairs have the same probability of where the f (h) is the probability density function of the
error due to the structure of the STBC, any pair of symbols can channel. Even numerical calculation of PL1 in (14) is hard as
be chosen to carry out the error rate analysis. We denote the it requires executing five-fold integration. The instantaneous
symbol pairs chosen to be detected at the first and second stage SINR parameter γ̃L1 in (11) of the MMSE filter output is a
by (s1 , s3 ) and (s2 , s4 ), respectively. The average probability function over two real random variables a, w, which turn are
of error of the pair of symbols (s1 , s2 ) is given by functions of the channel vector h. We derived (shown in the
Appendix) the joint distribution of a, w:
P(s1 ,s2 ) = (1 − PL1 )PL2 + (1 − PL2 )PL1 + PL1 PL2
1
= PL1 + PL2 − PL1 PL2 (10) fa,w (a, w) = (1 − w2 )a3 e−a , a ∈ (0, ∞), w ∈ (0, 1). (15)
4
where PL1 , PL2 are the average probability of error of the
Using (15) and (13) in (14),
symbols detected with linear MMSE in the first stage and the
MF detector in the second stage, respectively. Z 1 Z ∞
CSI
Symbol s1 is detected at the first stage with the MMSE filter PL1 = PL1 fa,w (a, w) dadw
0 0
receiver. The SINR at the output of the MMSE filter at the first " π
Z2Z1Z∞
π
Z4Z1Z∞# δ(γ̃ +1)
stage is 4µ 4µ2 − L1
= − e 2sin2 φ fa,w (a, w)da dw dφ
³ ´2 ³ ´2 π π
ρ
Mt a + 1 − Mρt b 0 0 0
π
0 0 0
π
γL1 = −1 " Z2Z1 Z4Z1 #
µΓ(4) µ2 Γ(4) c(φ)(1 − w2 )
³ ´
ρ
Mt a − |b| cos(θ) + 1 = − i4 dw dφ (16)
π π
h
2 0 0 0 0 G1 (w, φ)
a(1 − w ) 1 + w cos(θ)
≈ (ρ/4) + ,
1 − w cos(θ) 1 − w cos(θ) − δ
) 2
where c(φ) = e 2sin2 φ , G1 (w, φ) = 8sinδ(ρ)(1−w2 φ(1−w cos θ) + 1
a(1 − w2 ) △
≥ (ρ/4) + 1 = γ̃L1 + 1 (11) and Γ(·) is the gamma function. We can further integrate
1 − w cos(θ)
(16) with respect to the variable w, but the result is highly
where w , |b|/a(< 1). The approximation comes under the complicated and laborious and does not provide new insight
consideration ρ → ∞. on the analysis of the system. Hence, the average symbol error
Next, we calculate the channel instantaneous symbol error probability calculation is left at two-fold integration (from five-
probability at the output of the MMSE filter, denoted by fold integration initially).
CSI
PL1 . The instantaneous symbol error probability can be Next, we calculate the symbol error probability PL2 for the
found for many signal constellations in [16]. We focus on symbol s2 detected with the MF filter decoder at the second
square quadrature amplitude modulation (QAM) constellations stage. The error rate of the second stage PL2 , depends on the

620
error rate of the first stage PL1 , and is as follows: analysis and hence is omitted. Using (16) and (23) in (18) we
³ ¯ ´³ ´ obtain the upper-bound on the symbol error probability at the
PL2 = Pe ŝ2 6= s2 ¯ŝ1 = s1 1 − PL1
¯
second stage PL2 .
³ ¯ ´
+Pe ŝ2 6= s2 ¯ŝ1 6= s1 PL1 (17)
¯ ³ ¯ ´
• Calculate Pe ŝ2 6= s2 ¯ŝ1 6= s1 , Pe|L1 :
¯
³ ¯ ´³ ´
≤ Pe ŝ1 6= s2 ¯ŝ1 = s1 1 − PL1 + PL1 (18) We use the same signal model as (19). Under the assumption
¯
ŝ1 6= s1 , the resultant signal after the removal of the incorrectly
where ŝ1 and ŝ2 are the³ symbols ¯detected at´ first and second detected symbol of the first stage (ŝ1 ) from ye 12 is given by:
stage respectively. Pe ŝ2 6= s2 ¯ŝ1 = s1 is the average
¯
r
probability that the symbol detected at the second stage ŝ2 , E ρ
ye12 = ye12 − g1 ŝ1
is in error provided the ³ detected symbol
¯ ŝ1 ´at the first stage 4Mt
is correct. Similarly, Pe ŝ2 6= s2 ¯ŝ1 6= s1 is the average
r r
ρ ρ
¯
= g1 ∆s + g2 s2 + ne (24)
probability that the symbol detected at the second stage ŝ2 , is 4Mt 4Mt
in error provided the detected symbol ŝ1 in the first stage is
also in error. It is clear that (18) provides an upper-bound on where ∆s = (s1 − ŝ1 ). The matched filter g2 is used in
the symbol error probability for the second stage and it is a detecting the symbol s2 at the second stage. The channel
byproduct in calculation
³ of¯ (17). ´ instantaneous SINR is given by:
• Calculate Pe ŝ2 6= s2 ¯ŝ1 = s1 , Pe|L1 :
¯
q
ρ
We need to know the received signal under the above condition. E|g2H ( 4M t
g2 s2 )|2
E
The component of the received signal due to the symbols s1 , s2 γL2 = q
ρ
in noise is given by : E|g2H ( 4M t
g1 ∆s + ne )|2
r · ¸
ρ s1 ρ H 2
ye12 = HV[1,3] J + ne (19) 4Mt |g2 g2 |
4Mt s2 = ρ H 2 2 H
4Mt |g2 g1 | |∆s| + |g2 g2 |
where V[1,3] are the first and third columns of the Hadamard ρ H 2
matrix and J is the permutation matrix depends on the sign of 4Mt |g2 g2 |
≈ (25)
the channel parameter b. We represent the columns of HV[1,3] J ρ H 2 2 H
4Mt |g2 g1 | dmin + |g2 g2 |
by g1 and g2 . Under the assumption ŝ1 = s1 , the resultant
signal after the removal of the detected symbol of first stage (1 − w cos(θ))2 E
≈ , γ̃L2 (26)
(ŝ1 ) from ye12 is given by (w sin(θ) dmin )2
r r
C ρ ρ where we obtain the approximation in (25) by replacing the
ye12 = ye12 − g1 ŝ1 = g2 s2 + ne . (20)
4Mt 4Mt square of the absolute value of the error with the square of
The MF g2 is used for detecting the symbol s2 at the second the minimum distance, dmin of the signal constellation. The
stage. The channel instantaneous SINR is given by approximation is made under the notion that symbol error
occurs with highest probability with the nearest symbol in the
constellation. We used high SNR approximation (ρ → ∞) in
q
ρ
E|g2H ( 4M g2 s2 )|2 ρ
C
γL2 =
t
= (a)(1 − w cos(θ)). (21) (25) to get a simple intuitive result of (26). In essence, (26)
E|g2H ne |2 Mt tells us, given the symbol detected at the first stage is incorrect,
Using the notation (12), we simplify the channel instantaneous the SINR of symbol detected at the second stage with the
symbol error probability to MF detector is independent of the operating SNR. Intuitively
it shows vulnerability of MF detectors in presence of equal
energy interference. Using the notation (12), we simplify the
" Z π Z π
#
4 2 4 4
CSI
Pe|L1 = µ − µ2 e−α2 (φ) dφ (22) channel instantaneous symbol error probability to
π 0 π 0
" π π
#
4µ2
Z Z
C
δγL2 4µ 2 4
where α2 (φ) , .
We now try to find the average symbol
CSI
Pe|L1 = − e−α3 (φ) dφ (27)
2 sin2 φ π π
0 0
error probability at the second stage under (ŝ1 = s1 ), over all
the channel realizations δγ̃ E
Z 1Z ∞ where α3 (φ) , 2 sinL2
2 φ . We now try to find the average symbol
CSI
Pe|L1 = Pe|L1 fa,w (a, w) dadw error probability at the second stage under (ŝ1 6= s1 ), over all
0 0 channel realizations
π π
" Z2Z1Z∞ Z4Z1Z∞# C
4µ2 δγL2 Z 1Z ∞
4µ − CSI
= − e 2 sin2 φ fa,w (a, w)da dw dφ Pe|L1 = Pe|L1 fa,w (a, w) dadw
π π 0 0
0 0 0 0 0 0 π π
π π " Z2Z1Z∞ Z4Z1Z∞# E
Z Z1
2 Z4Z1 # 4µ 4µ2 δ γ̃L2
"
2 2 −
µΓ(4) µ Γ(4) (1 − w ) = − e 2 sin2 φ fa,w (a, w)da dw dφ
= − i4 dw dφ (23) π π
π π
h
G2 (w, φ) 0 0 0 0 0 0
0 0 0 0 π π
" Z2Z1 2 Z4Z1 #
cos θ) µΓ(4) µ Γ(4) (1 − w2 )
where G2 (w, φ) = δ(ρ)(1−w
8sin2 φ + 1. Similarly as before, = − h i dw dφ (28)
π π G3 (w, φ)
integrating further is cumbersome without any insight on 0 0 0 0

621
TABLE I
h i
δ(1−w cos θ)2
d2 w2 2 sin2 φ sin2 θ C OMPARISONS OF R EAL M ULTIPLICATIONS AND A DDITIONS
where G3 (w, φ) = e min . Using (16), (23),
and (28) we get an approximation of (17). We are able to QAM MMSE-MF of code in (1) Code in [8] ML Code in [5] ML
A RM RA RM RA RM RA
obtain semi-closed form approximate average block-error-rate 4 283 221 512 432 439 451
A
for two symbols in the codeword given by P(s 1 ,s2 )
by using 16 309 223 2048 1728 3319 4675
(16) and (17) in (10). Due to the symmetry of the codeword 64 309 223 8192 6912 43639 67651
transmission the other pair of symbols represented by (s3 , s4 )
have same approximated block-error-rate as the pair (s1 , s2 )
A
and is represented by P(s 3 ,s4 )
. Hence, the approximate average 85

Optimal Rotation Angles using (29)


block-error-rate for 4 symbols in the codeword in given by
80

A A A A
P(s 1 ,··· ,s4 )
= P(s 1 ,s2 )
+ P(s 3 ,s4 )
− P(s PA
1 ,s2 ) (s3 ,s4 )
. (29)
75

Rotation Angle in Degrees


Similarly, we can obtain an upper-bound in block-error-rate
70
for the two symbols in the codeword by using (16) and (18)
UB
in (10) and is denoted by P(s 1 ,s2 )
. Hence, the average upper- 65
bound block-error-rate for 4 symbols in the codeword is given
by 60

UB UB UB UB
P(s 1 ,··· ,s4 )
= P(s 1 ,s2 )
+ P(s 3 ,s4 )
− P(s P UB ,
1 ,s2 ) (s3 ,s4 )
(30) 55

UB
where P(s 3 ,s4 )
is the upper-bound in block-error-rate for the 50
10 15 20 25 30 35 40
SNR in dB
symbol pair (s3 , s4 ).
C
As a concluding remark, from (21) we see that γL2 >
ρ Fig. 1. Optimal rotation angle versus SNR for 4-QAM constellation.
Mt (a)(1 − cos(θ)), which shows full-diversity error perfor-
mance in Pe|L1 . Then, from (18) we can say that the error
performance of the second stage is dominated by PL1 . From (29) and the upper-bound (30) block-error-rate of the proposed
(11), we see that PL1 depends on w. For channel conditions for QO-STBC with MMSE-MF decoding for the 4-QAM signal
which w → 1, γL1 → 0 and the effectiveness of the MMSE- constellation. We observe that the approximate block-error-
MF receiver decreases. rate acts as tight upper-bound on the observed (simulated)
V. C OMPARISONS block-error-rate, since we started out the analysis using an
approximated lower-bound on the SINR of the first stage.
In Table I, we carried out a detailed complexity analysis
and calculate the number of real multiplications and additions, VI. C ONCLUSION
denoted by RM and RA respectively, needed by the ML We developed a new 4 × 4 QO-STBC based on Hadamard
implementation of [5], [8] and our developed proposed QO- precoding with ML-optimal joint two-real-symbol ML decod-
STBC with MMSE-MF decoding. In all cases, we consider one ing and highly efficient and effective sub-optimal MMSE-MF
receive antenna (Mr = 1). We can conclude that the proposed decoding. The diversity product maximized rotation angle was
code/decoder reduces decoding complexity significantly. obtained for the ML decoder. The block-error-rate minimized
For error-rate illustration purposes we consider the 4-QAM rotation angle were obtained for the MMSE-MF decoder. The
signal constellation example. For the proposed codeword in sub-optimal MMSE-MF decoder shows loss in bit-error-rate in
(1) with ML decoding, diversity-product maximizing rotation the order of 0.1 dB to the ML decoder with strikingly lower,
angle and optimal diversity-product value is 63.4◦ and 0.3343 state-of-the-art complexity (for example, 309 real multiplica-
respectively. Under MMSE-MF decoding, the block-error-rate tions and 223 real additions for 64-QAM).
minimizing rotation angle is shown in Fig. 1 as a function of
the SNR. A PPENDIX : P ROOF OF (15)
In Fig. 2, we compare the block-error-rate (BLKER) and We wish to identify the joint distribution of the channel
bit-error-rate (BER) performance of the proposed codeword parameters a and w. We define the new variables
(1) under MMSE-MF, and ML decoding and the codewords √
+ △
from [5], [8] with ML decoders. As expected the proposed u+ = u+R + juI = (1/ 2) · (h1 + h3 ) ∼ CN (0, 1),
code with ML decoding and the ML decoded code in [8] − △

u− = u−R + juI = (1/ 2) · (h1 − h3 ) ∼ CN (0, 1),
have the same error performance due to their near identical △ √
+
diversity-product value. The MMSE-MF decoder shows only v + = vR + jvI+ = (1/ 2) · (h2 + h4 ) ∼ CN (0, 1),
a 0.3 dB and 0.1 dB loss to the ML decoded codewords from − △ √
v − = vR + jvI− = (1/ 2) · (h2 − h4 ) ∼ CN (0, 1),
[5] and [8], respectively, for practical bit-error-rate values but
has a much lower decoding complexity as shown in Table I. where CN (0, 1) is the complex Gaussian distribution with
(±) (±) (±) (±)
Finally, in Fig. 3 we plot the approximate semi-closed form zero mean and unit variance and uR , uI , vR , and vI

622
10
−1 given by:
(1) (1) (2) (2)
fZ1 ,Z2 (z1 , z2 ) fZ1 ,Z2 (z1 , z2 )
fw,a (w, a) = (1) (1)
+ (2) (2)
(32)
10
−2
BLKER |J(z1 , z2 )| |J(z1 , z2 )|
(1) (1) (2) (2)
where z1 = ( 1+w 1−w 1−w
2 )a, z2 = ( 2 )a, z1 = ( 2 )a, z2 =
1+w
BER
( 2 )a and J(·, ·) are the real roots of the transformations
Error rate

−3
10
and the Jacobian matrix. Simplifying (32) we get a neat and
simplified pdf as
ML [8]
−4 ML [5] 1
(1 − w2 )a3 e−a , a ∈ (0, ∞), w ∈ (0, 1). (33)
10
MMSE−MF 63.4o fw,a (w, a) =
Proposed ML 4
MMSE−MF optimal angle

−5
10
11 12 13 14
SNR in dB
15 16 17 18
R EFERENCES
[1] V. Tarokh, H. Jafarkhani, and A. R. Calderbank, “Space-time block codes
Fig. 2. Block-error-rate and bit-error-rate versus SNR comparisons for 4- from orthogonal designs,” IEEE Trans. Inform. Theory, vol. 45, pp. 1456-
QAM constellation. 1467, May 1999.
[2] S. Alamouti, “A simple transmit diversity technique for wireless com-
munications,” IEEE J. Select. Areas Commun., vol. 16, pp. 1451-1458,
Aug. 1998.
10
−1
[3] H. Jafarkhani, “A quasi-orthogonal space-time block code,” IEEE Trans.
Approximate block−error−rate
Commun., vol. 49, pp. 1-4, Jan. 2001.
Upper−bouund block−error−rate [4] O. Tirkkonen, A. Boariu, and A. Hottinen, “Minimal non-orthogonality
Simulated block−error−rate rate 1 space-time block code for 3+ Tx antennas,” in Proc. IEEE 6th
Int. Symp. Spread-Spectrum Techniques and Applications (ISSSTA 2000),
10
−2 Parsippany, NJ, Sept. 2000, pp. 429-432.
[5] W. Su and X. G. Xia, “Signal constellations for quasi-orthogonal space-
Block−Error Rate

time block codes with full diversity,” IEEE Trans. Inform. Theory, vol.
50, pp. 2331-2347, Oct. 2004.
[6] N. Sharma and C. B. Papadias, “Improved quasi-orthogonal codes
−3 through constellation rotation,” IEEE Trans. Commun., vol. 51, pp. 332-
10
335, Mar. 2003.
[7] Z. A. Khan, B. S. Rajan, and M. H. Lee, “Rectangular co-ordinate inter-
leaved orthogonal designs,” in Proc. IEEE GLOBECOM, San Francisco,
CA, Dec. 2003, vol. 4, pp. 2004-2009.
−4
[8] C. Yuen, Y. L. Guan, and T. T. Tjhung, “Quasi-orthogonal STBC with
10
11 12 13 14 15 16 17 18 19 minimum decoding complexity,” IEEE Trans. Wireless Commun., vol. 4,
SNR in dB pp. 2089-2094, Sept. 2005.
[9] H. Wang, D. Wang, and X.-G. Xia, “On optimal quasi-orthogonal space-
Fig. 3. Upper-bound, approximated, and simulated block-error-rate with time block codes with minimum decoding complexity,” IEEE Trans.
MMSE-MF decoding versus SNR for the 4-QAM constellation (rotation angle Inform. Theory, vol. 55, pp. 1104 - 1130, Mar. 2009.
63.4◦ ). [10] W. Liu, M. Sellathurai, P. Xiao and J. Wei, “A New Restricted Full-
Rank Single-Symbol Decodable Design for Four Transmit Antennas”,
IEEE Signal Processing Lett., vol. 15, pp.765-768, Nov. 2008.
[11] Y. Shang and X.-G. Xia, “ Space-time block codes achieving full diversity
with linear receivers,” IEEE Trans. Inform. Theory, vol. 54, pp. 4528-
are real and imaginary parts of the corresponding complex 4547, Oct. 2008.
random variables where each of the real and imaginary part [12] G. S. Rajan and B. S. Rajan, “MMSE optimal algebraic space-time
codes,” IEEE Trans. Wireless Commun., vol.7, pp. 2468-2472, July 2008.
is an independent and identically distributed real Gaussian [13] X. Guo and X.-G. Xia, “On full diversity space-time block codes with
random variable with zero mean and 0.5 variance. Next, partial interference cancellation group decoding,” IEEE Trans. on Inform.
we define two new random variables to used to obtain a Theory, vol. 55, pp. 4366-4385, Oct. 2009.
△ [14] A. Sezgin and T. J. Oechtering, “Complete characterization of the
and w. Z1 = |u+ |2 + |v + |2 = u+2 +2
R + vR + uI
+2
+ vI+2 , equivalent MIMO channel for quasi-orthogonal space-time codes,” IEEE
△ −2 −2 −2 −2
Z2 = |u− |2 + |v − |2 = uR + vR + uI + vI . Clearly, Trans. Inform. Theory, vol. 54, pp. 3315-3326, July 2008.
[15] W. Su, S. N. Batalama, and D. A. Pados, “On orthogonal space-time
Z1 , Z2 are both independent and identically distributed with block codes and transceiver signal linearization,” IEEE Commun. Lett.,
generalized Chi-square distribution. The joint distribution of vol. 10, pp. 9193, Feb. 2006.
random variables Z1 and Z2 is given by [16] M. K. Simon and M.-S. Alouini, Digital Communication over Fading
Channels: A Unified Approach to Performance Analysis. New York:
Wiley, 2000.
fZ1 ,Z2 (z1 , z2 ) = z1 z2 e−(z1 +z2 ) , z1 ∈ (0, ∞), z2 ∈ (0, ∞).
(31)
We now aim to find the joint distribution of w, a from joint
probability distribution of Z1 , Z2 using linear transformation.
Simple transformation gives, w = |Z 1 −Z2 |
Z1 +Z2 , a = Z1 + Z2 . The
joint probability density function of random variables w, a is

623

You might also like