Anda di halaman 1dari 3

IEEE COMMUNICATIONS LETTERS, VOL. 12, NO.

8, AUGUST 2008 545

Spatial Modulation: Optimal Detection and Performance Analysis


Jeyadeepan Jeganathan, Ali Ghrayeb, and Leszek Szczecinski

AbstractIn this letter, we derive the optimal detector for the


so-called spatial modulation (SM) system introduced by Mesleh
et al. in [1]. The new detector performs significantly better than
the original ( 4 dB gain), and we support our results by deriving
a closed form expression for the average bit error probability.
As well, we show that SM with the optimal detector achieves
performance gains ( 1.5 3 dB) over popular multiple antenna
systems, making it an excellent candidate for future wireless
communication standards.
Index TermsAntenna modulation, maximum likelihood de-
tection, MIMO, spatial modulation.
Fig. 1. Spatial modulation system model.

I. I NTRODUCTION
Second, in order to support our results, we analyze the

U SING multiple antennas in wireless communications


allows unprecedented improvements over current sys-
tems. Large spectral efficiency is obtained by using trans-
performance of the SM system, and derive a closed form
expression for the bit error probability when real constellations
are used.
mission techniques designed for multiple input multiple out- Organization: Section II introduces the basic SM system
put (MIMO) systems, such as the vertical Bell Laboratories model, the mapper, and the original detector. In Section III, we
layered space-time (V-BLAST) architecture [2]. Due to inter- derive the optimal detector, and provide a comprehensive look
channel interference (ICI) caused by coupling multiple sym- at its complexity. An in depth analysis of SMs performance
bols in time and space, maximum likelihood (ML) detection is then provided in Section IV. Section V presents some
increases exponentially in complexity with the number of simulation results, and we conclude the paper in Section VI.
transmit antennas. Consequently, avoiding ICI greatly reduces
receiver complexity, and contributes in attaining performance II. S PATIAL M ODULATION
gains.
A. SM Transmission
The so-called spatial modulation (SM), introduced by
Mesleh et al. in [1], [3], is an effective means to remove The general system model is shown in Fig. 1, which
ICI and the need for precise time synchronization amongst consists of a MIMO wireless link with Nt transmit and
antennas. SM is a pragmatic approach for transmitting infor- Nr receive antennas.1 A random sequence of indepen-
mation, where the modulator uses well known amplitude/phase dent bits b enters the SM mapper, which groups m =
modulation (APM) techniques such as phase shift keying log2 (M Nt ) bits and maps them to a constellation vector
 T
(PSK) and quadrature amplitude modulation (QAM), but also x = x1 x2 xN t , where we assume a power con-
employs the antenna index to convey information. Only one straint of unity (i.e. Ex xH x = 1). In SM, only one antenna
antenna remains active during transmission so that ICI is remains active during transmission and hence, only one of the
avoided, and inter-antenna synchronization (IAS) is no longer xj in x is nonzero. The signal is transmitted over an Nr Nt
needed as in the case of V-BLAST, where all antennas transmit wireless channel H, and experiences an Nr dim additive T
at the same time. However, the detector in [1] is sub-optimal, white Gaussian (AWGN) noise = 1 2 Nr .

and hinders the full performance gains achievable by SM. The received signal is given by y = Hx + , where is the
Contribution: The contribution of this paper is twofold. average signal to noise ratio (SNR) at each receive antenna,
We first derive the optimal detector for SM, improving over and H and have independent and identically distributed (iid)
the sub-optimal detection rules suggested in [1]. Thanks to entries according to CN (0, 1).
our development, SM compares favorably to other transmis- As mentioned earlier, SM exploits the antenna index as
sion schemes, such as APM with maximum ratio combining an additional means to transmit information. The antenna
(MRC), and the well known V-BLAST. 1 The following notations are used throughout the paper. Italicized symbols
denote scalar values, while bold lower/upper case symbols denote vec-
Manuscript received May 13, 2008. The associate editor coordinating the
tors/matrices. We use ()T for transpose, ()H for conjugate transpose, and
review of this letter and approving it for publication was M. Uysal. This work 
for the binomial coefficient. We use || for absolute value of a scalar,

was supported in part by NSERC (Canada) under a Post Graduate Scholarship
(PGS), and by FQRNT (Quebec) under research grant F00803. and F for the Frobenius norm of a vector/matrix. We use CN , 2 for
J. Jeganathan and A. Ghrayeb are with the ECE Department the complex Gaussian distribution
  of a random variable, having independent
at Concordia University, Montreal, Canada (e-mail: {j jegana, Gaussian distributed N , 12 2 real and imaginary parts with mean and
2
aghrayeb}@ece.concordia.ca). variance 2 . We use P () for the probability of an event, pY () for the
L. Szczecinski is with INRS-EMT, Montreal, Canada (e-mail: probability density function (PDF) of a random variable Y, and Ex [] for
leszek@emt.inrs.ca). the statistical expectation with respect to x. We use Re {} for the real part
Digital Object Identifier 10.1109/LCOMM.2008.080739. of a complex variable, and X represents a constellation of size M .
1089-7798/08$25.00 
c 2008 IEEE
546 IEEE COMMUNICATIONS LETTERS, VOL. 12, NO. 8, AUGUST 2008

combined with the symbol index make up the SM mapper, PDF of y, conditioned on xjq and H. It can be seen that
which outputs a constellation vector of the following form: optimal detection requires a joint detection of the antenna
 T indices and symbols, as opposed to the scheme outlined in
0 0 xq 0 0 Section II-B, where the problem is decoupled.
xjq  ,
j th position

where j represents the activated antenna, and xq is the q th B. Complexity


symbol from the M -ary constellation X . Hence, only the j th To compare the complexity of the optimal detector with
antenna remains active during symbol transmission. Figure 2 the one given in Section II-B, we use the number of mul-
in [1] illustrates an example of the mapper and is omitted tiplications required in the detection process. The number
here due to space constraints. For example, in 3 bits/s/Hz of additions can be shown to have a similar value for both
transmission with Nt = 4 antennas, the information bits are detectors. The complexity of Meslehs SM detector is obtained
mapped to a 1 binary PSK (BPSK) symbol, and transmitted from [1] as2 SM = 2Nr Nt + Nt + f (M ), where the last term
on one of the four available antennas. The output of the depends on the type of demodulation assumed.
channel when xq is transmitted from the j th antenna is
Similar to [1], we analyze (4) to obtain the complexity
expressed as
of optimal SM detection. It can be shown that the first and
y = hj xq + , (1)
second terms in (4) result in Nr Nt + M and Nr Nt + Nt M
where hj denotes the j th column of H. multiplications, respectively (for the first term, we simplified
the computation by splitting hj xq 2F = hj 2F |xq |2 ). Hence,
B. SM Detection (Sub-Optimal) the total computational complexity for optimal SM detection
In [1], a sub-optimal detection rule based on MRC is given is given by SM,opt = 2Nr Nt + Nt M + M .
by We observe that the complexity involved between both
detectors is comparable. These results motivate the use of our
j = arg max zj (2) optimal SM detector in order to fully exploit SMs advantages.
j
q = D (zj ) , (3)
IV. P ERFORMANCE A NALYSIS
|
hH
j y |
where zj = , j and q represent the estimated antenna
hj 2F The performance of the SM system (with the optimal
and symbol index, respectively, and D is the constellation detector) is derived using the well known union bounding
demodulator function. Since the mapping is one to one, the technique [4, p. 261-262]. The average bit error rate (BER) in
demapper obtains an estimate of the transmitted bits by taking SM is bounded as
j and q as inputs.
We note that the simulation results of [1] could not be 
reproduced using the conventional channel assumptions of Pe,bit Ex N (q, q)P (xjq xjq )
j,q
Section II. The reason for this can be seen (in the noiseless
Nt M Nt 
M
case) by substituting (1) for y in (2). Therefore, zj reduces to N (q, q)P (xjq xjq )

|hH
k hj xq |
= , (5)
hk 2F
and, in order to detect the correct antenna index j=1 q=1 j=1 q=1
Nt M
|hH hj |
(i.e. k = j), we require hk 2 < 1. By invoking Cauchys where N (q, q) is the number of bits in error between the
k F
inequality to the left hand side, we find that hj F hk F symbol xq and xq , and P (xjq xjq ) denotes the pairwise
is a necessary condition for antenna detection without errors, error probability (PEP) of deciding on the constellation vector
which should materialize in the absence of noise. One way to xjq given that xjq is transmitted. By using (4), the PEP
ensure this condition is by normalizing the channel prior to conditioned on H is given by
2
transmission (i.e. hj F = c for all j, where c is a constant).  
We refer to these channels as constrained. P (xjq xjq | H) = P (djq > djq | H) = Q ,


We now derive the optimal detector for SM, where the need 2
where djq = gjq F 2Re yH gjq , and Q (x) =
for constrained channels is not necessary.  1 t2
2 dt. We define as
x 2 e
III. O PTIMAL D ETECTION
Nr

A. SM Detection (Optimal) 
gjq gjq F =
2 2
|A (n) + iB (n)| , (6)
2
Since the channel inputs are assumed equally likely, the n=1
optimal detector is based on the ML principle:
where i = 1 and
[jML , qML ] = arg max pY (y | xjq , H)   
R R I I R R I I
j,q A (n) = 2 hnj xq hnj xq hnj xq + hnj xq
  
2
= arg min gjq F 2Re yH gjq , (4) R I I R R I I R

j,q B (n) = 2 hnj xq + hnj xq hnj xq hnj xq .

where gjq = hj xq , 1 j
Nt , 1 q M , and
2 2 In [1], the complex multiplications involved with h 2 and || are
pY (y | xjq , H) = Nr exp y Hxjq F is the
j F
ignored, and results in a slightly different SM .
JEGANATHAN et al.: SPATIAL MODULATION: OPTIMAL DETECTION AND PERFORMANCE ANALYSIS 547

0
10
MRC (APM), Nt=1 bits/s/Hz transmission with Nr = 4 antennas. Figure 2 illus-
MRC (APM), Nt=1
trates the simulation results for both constrained (dotted line)
VBLAST, N =3
1
t
VBLAST, N =3
t
and conventional (solid line) channel assumptions (see Section
10
SM (Mesleh), Nt=4 II-B).
SM (Mesleh), Nt=4
SM (Optimal), Nt=4 For reference, we use two different transmission setups.
2
10
SM (Optimal), N =4
t The first one is APM, 8-QAM transmission with Nt = 1
(single antenna transmission) and M = 8. The second is
SM (Analytical), Nt=4

V-BLAST with BPSK modulation, Nt = 3, and ordered


Pe,bit

conventional H
constrained H
3
10 successive interference cancellation (OSIC) with the minimum
mean squared error (MMSE) receiver [6]. SM with BPSK
and Nt = 4 antennas is shown for both sub-optimal [1] and
4
10 optimal receivers (derived in Section III-A). We also plot the
Pe,bit bound of (8) for SM using BPSK modulation, where
0, q = q
Nt = 4, M = 2, 2 = 2 and N (q, q) = .
5
10 1, q = q
0 5 10
(dB)
15 20
Let us first consider the case of constrained channels (dotted
lines). As shown, the optimal SM detector gains 4 dB at
Fig. 2. BER performance of spatial modulation versus SNR, for m = 3 Pe,bit = 105 over the sub-optimal detector of [1]. Also,
bits/s/Hz transmission (Nr = 4).
(minor) gains are also evident over MRC and V-BLAST, which
is not the case with sub-optimal detection. For the conventional
The superscript R and I denote the real and imaginary part, channel model (solid line), it is shown that optimal SM
respectively, and hnj is the element of H in the nth row, provides performance improvements of 3 dB over APM, and
and j th column. The distribution of the random variable 1 dB over V-BLAST (at Pe,bit = 105 ). As well, we notice
in (6) is not easily obtained since A (n) and B (n) are not, that the derived BER bounds are relatively tight, and support
in general, independent. In this case, the performance can our simulation results.
be evaluated numerically. However, for symbols x drawn
from a real constellation  X , this independence is satisfied VI. C ONCLUSION
2Nr 2
 2

and (6) reduces to = n=1 n where n N 0, In this letter, we derive the optimal SM detector for which,
(|xq |2 +|xq |2 ) significant performance gains are observed over the detector
with 2 = 4 . Hence, is a chi-squared ran-
dom variable with 2Nr degrees of freedom and PDF p (v) in [1]. To support our results, we also derive a closed form
given in [4, p. 41]. expression bounding the average BER of SM when real
 The PEP can then be formulated as
constellations are used. The simulation results indicate that
P (xjq xjq ) = v=0 Q ( v) p (v) dv, which has a closed
form expression given in [5, Eq. (64)]. Thus, SM with optimal detection outperforms V-BLAST and APM
transmission, which makes it a promising candidate for low
N r 1
 
Nr 1 + k complexity transmission techniques.
P (xjq xjq ) = Nr
[1 ]k . (7)
k
k=0

 2 R EFERENCES

where = 12 1 1+ 2 . Plugging (7) into (5), we [1] R. Mesleh, H. Haas, C. W. Ahn, and S. Yun, Spatial modulationa new

obtain low complexity spectral efficiency enhancing technique, in Proc. Conf.
Comm. and Networking in China, Oct. 2006.
M  M Nr 1 Nr 1+k k
 Nt N (q, q)N
r
[1 ] [2] P. Wolniansky, G. Foschini, G. Golden, and R. Valenzuela, V-BLAST:
k=0 k
Pe,bit . an architecture for realizing very high data rates over the rich-scattering
q=1 q=1
M wireless channel, in Proc. ISSSE-98, Pisa, Italy.
[3] R. Mesleh, H. Haas, S. Sinanovic, C. W. Ahn, and S. Yun, Spatial
(8) modulation, IEEE Trans. Veh. Technol., vol. 57, no. 4, pp. 22282241,
July 2008.
V. S IMULATION R ESULTS [4] J. G. Proakis, Digital Communications, (4th ed.) New York: McGraw-
Hill, 2001.
In this section, we present some examples to compare [5] M.-S. Alouini and A. Goldsmith, A unified approach for calculating error
the optimal SM detector over the original detection scheme rates of linearly modulated signals over generalized fading channels,
IEEE Trans. Commun., vol. 47, no. 9, pp. 13241334, Sept. 1999.
[1]. We perform Monte Carlo simulations for 106 channel [6] R. Bohnke, D. Wubben, V. Kuhn, and K. D. Kammeyer, Reduced
realizations and plot the average BER performance versus complexity MMSE detection for BLAST architectures, in Proc. IEEE
(the average SNR per receive antenna). We target m = 3 Globecom03, San Francisco, California, USA, Dec. 2003.

Anda mungkin juga menyukai