www.ietdl.

org
Published in IET Communications
Received on 22nd February 2010
Revised on 15th December 2010
doi: 10.1049/iet-com.2010.0667

ISSN 1751-8628

Spatial modulation: optimal detector asymptotic

performance and multiple-stage detection
N.R. Naidoo H.J. Xu T. Al-Mumit Quazi
School of Electrical, Electronic and Computer Engineering, University of KwaZulu-Natal, Durban 4041,
Republic of South Africa
E-mail: nigelreece.naidoo@gmail.com

Abstract: Spatial modulation (SM) is a recent multiple-input multiple-output transmission technique, which entirely avoids interchannel interference as well as the need for transmit antenna synchronisation. The rst objective of this study is to present
an asymptotic bound to quantify the average bit error rate (BER) performance of M-ary quadrature amplitude modulation
(M-QAM) SM with optimal-based detection over independent and identically distributed Rayleigh at fading channels. The
analytical frameworks are validated by Monte Carlo simulation results, which show that the derived lower bounds are
increasingly tight for large signal-to-noise ratio values. The second objective is to introduce a novel SM detection scheme,
termed multiple-stage (MS) detection. Performance and complexity comparisons with existing SM detectors show two main
benets of MS detection: near-optimal BER performance and up to a 35% reduction in receiver complexity as compared to
the maximum likelihood-based SM detector.

Introduction

Multiple-input multiple-output (MIMO) technology is integral

to the development of current and future wireless
communication systems. In recent years, MIMO has been
considered as one of the core techniques for improving data
throughput, link reliability and spectral efciency [1 3].
Open loop MIMO techniques can be divided into two main
categories namely, spatial diversity and spatial multiplexing
schemes. Spatial diversity schemes, space time coding
[4, 5] for example, improve link reliability by transmitting
multiple redundant copies of data to a receiver over
independent channels. Spatial multiplexing schemes, such as
vertical-bell layered space-time architecture (V-BLAST) [6],
divide the input data stream into several sub-streams. Each
sub-stream is then encoded independently and transmitted in
parallel via a specic antenna, thereby increasing the
capacity. As a result, spatial multiplexing schemes are
ideally suited to next generation wireless communication
systems, where high data rates are required. However, spatial
multiplexing schemes experience several limitations such as
high inter-channel interference (ICI) because of the coupling
of multiple symbols in space and time, the need for interantenna synchronisation (IAS) to ensure the simultaneous
transmission of data and high complexity maximum
likelihood (ML) detection algorithms [7, 8].
Spatial modulation (SM) is a novel transmission approach
proposed by Mesleh et al. [9, 10], which is not prone to the
aforementioned spatial multiplexing limitations. The basic
idea behind SM is to map a block of data bits to an MQAM/M-ary phase shift keying (M-PSK) symbol and
transmit antenna index. The modulated symbol is then
transmitted over a wireless channel via a single antenna as
1368
& The Institution of Engineering and Technology 2011

specied by the transmit antenna index. Hence, SM

efciently avoids the ICI and IAS limitations of V-BLAST,
since only a single transmit antenna is active during
transmission. At the receiver, a maximum-receive ratio
combining (MRRC) SM detector is used to estimate both
transmit antenna index and modulated symbol. In [9] it was
shown that for the same spectral efciency SM can achieve
similar performance to V-BLAST while maintaining a low
complexity detection algorithm. However, Jeganathan et al.
[11] demonstrated that the original MRRC SM detector is
sub-optimal and therefore proposed an ML-based optimal
SM detection scheme.
1.1

Recent work

Recent studies have led to the development of several novel

SM schemes. Relevant examples to mention are: trelliscoded spatial modulation in [12], where trellis-coded
modulation is applied to the antenna constellation points of
SM in order to improve the performance over spatially
correlated channels; soft output ML detector in [13], where
a soft-output ML detector for SM orthogonal frequency
division multiplexing systems is introduced and shown to
outperform the conventional hard decision-based SM
detector and time orthogonal signal design-assisted SM in
[14], which is a novel SM-based scheme that offers transmit
diversity gains and better performance over correlated
channels when compared to conventional SM.
1.2

Motivation and contributions

The ML-based optimal SM detector is analysed in [11], where

the average bit error rate (BER) is evaluated in closed form.
IET Commun., 2011, Vol. 5, Iss. 10, pp. 13681376
doi: 10.1049/iet-com.2010.0667

www.ietdl.org
However, the theoretical performance bound [11, Eq. (8)] is
inadequate since it can only predict the performance of
binary phase shift keying (BPSK) SM congurations. The
ever increasing need for high rate data transmission cannot
be met using a low rate digital modulation technique like
BPSK in conjunction with SM. In order to increase the data
rate, it is necessary to combine SM with a more spectrally
efcient technique such as M-QAM. To date, there has
been no literature reporting on the theoretical performance
of M-QAM spatial modulation with optimal detection (SMOD) in fading channel conditions. In this paper, we derive
a closed form expression to quantify the average BER
performance of square M-QAM SM-OD in independent and
identically distributed (i.i.d) Rayleigh at fading channels.
It is shown that the analytical bound closely predicts BER
performance in the high signal-to-noise ratio (SNR) region.
The optimal SM detector is based on the ML technique,
which can become extremely complex with high order
modulation schemes and a moderate number of transmit
antennas. It will be shown subsequently that for high order
(M 16) M-QAM SM congurations, M-QAM SM-OD
increases receiver complexity by 125% as compared to
the sub-optimal SM detector. This result motivates for an
investigation into an alternate SM detection scheme, termed
multiple-stage (MS) detection. The proposed detector is a
hybrid scheme which combines the operation of both suboptimal and optimal SM detectors. As a result, the MS
detector inherits the desirable properties of its constituent
detectors namely, the low complexity of sub-optimal
detection and the high performance of optimal detection.
Note that the MS detector is proposed specically for a four
transmit multiple receive antenna SM architecture with high
order M-QAM, however, MS detection is a generic concept
that can be applied to any SM conguration.
1.3

Organisation

The rest of the paper is organised as follows:

Section 2 reviews the SM transmission approach and current
SM detection schemes. In Section 3, an analytical approach to
the BER analysis of M-QAM SM-OD in i.i.d Rayleigh at
fading channels is presented. Section 4 introduces the MS

detection scheme, including a detailed description of the two

stage detection process. Section 5 rst validates the analytical
frameworks developed in Section 3 and then presents a
performance comparison between the various SM detection
schemes. In Section 6, the various SM detection schemes are
analysed in terms of computational complexity and nally
Section 7 concludes the paper.
1.4

Notation

Bold italics upper/lower case symbols denote matrices/

vectors, while regular letters represent scalar quantities. We
use (.)T, (.)H, E[.], |.| and .F for transpose, Hermitian,
expectation, Euclidean norm and Frobenius norm operators,
respectively.

2
2.1

Spatial modulation
Transmission

The Nt Nr M-QAM SM system model is shown in Fig. 1

(redrawn from [9]), where Nt and Nr denote the number of
transmit and receive antennas, respectively. Initially, the
SM mapper assigns an r log2(MNt) bit binary input to a
transmit antenna index j and M-QAM symbol xq . The
assignment of j and xq are dened by the SM mapping
table, which is known at both transmitter and receiver. A
sample SM mapping table for 3 bits/s/Hz SM transmission
is shown in Fig. 2 of [9]. The SM mapper output can be
expressed as [11]
jth position

xjq = [0 0 . . . xq . . . 0]T


1st position
Ntth position

(1)

where xjq is the Nt dimensional signal vector, xq denotes the

qth symbol from an M-ary constellation with E[|xq |2 ] = 1
and q [ [1:M ].
In SM, the modulated symbol xq is transmitted via a single
antenna j. As a result, the signal vector (1) has Nt 2 1 zero

Fig. 1 SM system model

Fig. 2 MS detector operation

IET Commun., 2011, Vol. 5, Iss. 10, pp. 13681376
doi: 10.1049/iet-com.2010.0667

1369

& The Institution of Engineering and Technology 2011

www.ietdl.org
entries corresponding to the dormant transmit antennas and
a single non-zero entry xq at the jth position corresponding
to the active transmit antenna. After the mapping process,
signal vector xjq is then transmitted over the Nr Nt MIMO
channel H and experiences Nr dimensional additive white
Gaussian noise (AWGN) n = [n1 , n2 . . . nNr ]T . The received
signal vector is given by [11]
y=


pHxjq + n

(2)

where y = [y1 y2 . . . yNr ]T represents the received signal vector

and p is the average SNR at each receive antenna. Note that
both H and n have i.i.d entries according to the complex
Gaussian distribution CN (0, 1).
Assuming transmission from the jth antenna, the received
signal vector can also be represented by [11, Eq. (1)]
y=


phj xq + n

Sub-optimal detection

The MRRC-based sub-optimal SM detector decouples the

transmit antenna and symbol estimation processes. The
transmit antenna index is estimated rst followed by symbol
estimation. In general, these two estimation processes are
interdependent and their subsequent decoupling during
detection leads to reduced performance. In [11] Jeganathan
et al. showed that the sub-optimal SM detector [9] is only
applicable in constrained channel conditions, where a
constrained channel is obtained by normalising the channel
prior to transmission such that hj 2F = c for all j and
constant c. Since the sub-optimal detector only operates
under some articial channel assumptions, a modied
version of this scheme is suggested in this paper. The
modied sub-optimal detection rules can be applied in
conventional channel conditions as follows
zj =

hH
j y
hj F

(4)

where j [ [1:Nt ] and (4) is obtained from [9, Eq. (4)] by

introducing the factor hj F .
As in [9, Eq. (6)], the transmit antenna index is estimated
by
j = argmax |zj |

(5)

where j represents the estimated transmit antenna index.

Assuming correct transmit antenna index estimation, the
combined symbol based on MRRC is given by
x = hj F zj = hH
j y

(6)

The ML criterion is then applied in order to obtain a symbol

1370
& The Institution of Engineering and Technology 2011

xq = argmax
q


phj xq 2F 2Re{xxq }

(7)

where xq represents the symbol estimate and q [ [1:M ].

2.3

Optimal detection

The ML-based optimal SM detector fully exploits the

advantages of SM by performing a joint detection of
transmit antenna index and symbol. The optimal detection
rule can be applied as follows [11, Eq.(4)]
[j, xq ] = argmax pg (y|xjq , H)
j,q


= argmin [ pgjq 2F 2Re{yH gjq }]

(3)

where hj denotes the jth column of H.

At the receiver, the SM detector obtains estimates of the
transmit antenna index and modulated symbol. These
estimates are then input to the SM de-mapper, which uses
the SM mapping table to perform a reverse mapping
process in order to recover an estimate of the original r bit
binary input. Note that the SM detection process assumes
perfect channel state information at the receiver.
2.2

estimate [4, Eq. (6)]

(8)

j,q

where gjq = hj xq , j [ [1:Nt ] and q [ [1:M ]. j and xq

represent the estimated transmit antenna index and
symbol based onML, respectively, and pY (y|xjq , H) =
pN r exp (y pHxjq 2F ) is the probability density
function (PDF) of the received signal vector y conditioned
on both xjq and H.

3 Asymptotic performance analysis of

M-QAM SM-OD
In this section, we derive an asymptotic performance bound
for M-QAM SM-OD in i.i.d Rayleigh at fading channel
conditions. The SM detector is responsible for the
estimation of two quantities: the active transmit antenna
index and transmitted symbol. As a result, SM performance
depends on the error rates of these two processes. The
subsequent analysis follows the same approach as [10],
where independent transmit antenna index and symbol
estimation processes are assumed. Let Pa denote the bit
error probability of transmit antenna index estimation given
that the symbol is perfectly detected and Pd be the bit error
probability of symbol estimation given that the transmit
antenna index is perfectly detected. The overall probability
of bit error is then bounded by
Pe 1 Pc = Pa + Pd Pa Pd

(9)

where Pc = (1 Pa )(1 Pd ) is the probability of correct SM

bits.
In general, it is incorrect to assume independent estimation
processes, since the optimal SM detector performs a joint
detection of transmit antenna index and symbol. The
assumption of independent estimation processes represents
an ideal scenario, therefore (9) corresponds to the best
performance (lower bound) achievable by an M-QAM SMOD scheme. In what follows, the BER of each estimation
process is considered separately.
3.1

Analytical BER of symbol estimation

In SM, only one antenna is active during transmission.

Therefore, at any instant the Nt Nr SM model may be
viewed as a 1 Nr single-input multiple-output
conguration. Assuming perfect transmit antenna detection,
the SM detector applies the ML criterion for symbol
IET Commun., 2011, Vol. 5, Iss. 10, pp. 13681376
doi: 10.1049/iet-com.2010.0667

www.ietdl.org
The PEP conditioned on channel matrix H is given by
(refer to Appendix for derivation)

detection as follows
xq = argmax
q


pgjq 2F 2Re{yH gjq }

(10)

where xq represents the estimated symbol based on ML,

q [ [1:M ], p is the average SNR at each receive antenna,
gjq = hj xq , hj denotes the jth column of channel matrix H,
xq refers to the qth symbol from an M-QAM constellation
and y is the received signal vector.
In [4] Alamouti shows that (10) is equivalent to an Nr
branch MRRC of the received signal followed by a regular
symbol-by-symbol detection. The average symbol error rate
(SER) of square M-QAM with MRRC reception over
Rayleigh fading channels given by [15, Eq. (15)]
 
Nr

Nr
a 1
2
a
1
SER(p) =

c 2 bp + 2
2 bp + 1
+(1 a)

c1 


Si
bp + Si

i=1

Nr

2c1

i=c

Si
bp + Si



P(xjq  xjq |H) = P(y phj xq F , y phj xq F |H)

= Q( k )
(14)
1
2
 et /2 dt and k is dened as [18, Eq. (5)]
where Q(x) = x 1
2p
r 2
p
an
k = (hj xq hj xq )2F =
2
n=1
2N

where an  N (0, s2a ) with s2a = p2 |xq |2 .

The random variable k is the summation of 2Nr squared
i.i.d Gaussian random variables. Hence, k has a central
chi-square distribution with 2Nr degrees of freedom and
PDF given by [17, p. 41]
2

pk (v) =

Nr 

(11)

where a = (1 1/ M ), b 3/(M 2 1), m = log2 (M ),
Si = 2 sin2 ui , ui = ip/4n, Nr is the number of receive
antennas and c is the number of summations. It is shown in
[16] that applying c . 10 results in a 0.0015, 0.0025 and
0.0029 dB difference between the simulated and theoretical
SER for the 4, 16 and 64 QAM congurations, respectively.
Pd is then derived from the approximate relationship as
follows [17]
SER
Pd
m

(12)

1
P(xjq  xjq ) =

P(xjq  xjq ) = mNa r

(17)


Nr 1 + w
[1 ma ]w
w

(18)




2
where ma = 12 1 1+sas2 .
a

N(j, j)P(xjq  xjq )

N (j, j)P(xjq  xjq )
Nt M

(13)

where P(xjq  xjq ) denotes the pairwise error probability

(PEP) of choosing signal vector xjq given that xjq was
transmitted and N(j, j) is the number of bits in error
between transmit antenna index j and estimated transmit
antenna index j.
IET Commun., 2011, Vol. 5, Iss. 10, pp. 13681376
doi: 10.1049/iet-com.2010.0667

Nr
Nt 
Nt N (j, j )ma
M 




Nr 1 Nr 1 + w
[1 ma ]w
w=0
w
Nt M
(19)

j=1 q=1 j=1


N
r 1

w=0

j=1 q=1 j=1


Q( v)pk (v)dv

Finally, substituting (18) into (13) yields

The bit error probability of transmit antenna index estimation

is derived using the same approach as [18]. Given that the
transmitted symbol is perfectly detected, the average BER
of transmit antenna index estimation is union bounded by
[17, pp. 261 262]:

Nt 
Nt
M 


(16)

The closed form solution for (17) is given in [19, Eq. (64)]

3.2 Analytical BER of transmit antenna index

estimation

Pa Ej

for v . 0

v=0

Pa

vNr 1 ev/2sa
(2s2a )Nr G(Nr )

1
where G(Nr ) = 0 t Nr 1 et dt is the gamma function.
Since k has a known distribution the PEP term in (13) can
be computed as follows [18, Eq. (6)]

where m = log2 (M ) is the number of bits per M-QAM

symbol.

(15)

MS detection

The optimal SM detector requires a brute force search among

all possible (Nt M ) pairs of transmit antenna index and
modulated symbol. As a consequence, the complexity of
SM-OD increases rapidly with modulation order M and the
number of transmit antennas Nt . One way to decrease the
complexity of SM-OD is to reduce the number of inputs to
the detector. The MS detection scheme achieves this
reduction in detector input by splitting the SM detection
process into two stages, as shown in Fig. 2. The rst stage
of MS uses the low complexity sub-optimal (MRRC) SM
detector to select the N (N Nt ) most probable estimates of
transmit antenna index, thereby reducing the number of
inputs pairs from Nt M to NM. The reduced input set is then
sent to the optimal (ML) SM detector in the second stage of
MS, where a nal estimate of transmit antenna index (jMS )
and modulated symbol (xq MS ) is obtained. Since the NM
inputs pairs to the second stage are the most probable
estimates of transmit antenna index and modulated symbol,
1371

& The Institution of Engineering and Technology 2011

www.ietdl.org
the high performance inherent to SM-OD is still maintained.
A mathematical description of the MS detection scheme is
presented next.

denote the nal transmit antenna index and symbol

estimates, respectively.

4.1

First stage: sub-optimal SM detector

In the rst stage of MS, the N most probable estimates of

transmit antenna index are determined based on the suboptimal antenna detection algorithm. The trade-off between
complexity and performance when choosing a suitable
value for N is discussed in Section 6. The sub-optimal
antenna detection rules can be applied to the received signal
vector as follows
hH
j y
hj F

zj =

(20)

where y is the received signal vector, hj denotes the jth

column of channel matrix H and j [ [1:Nt ].
Based on (20) the vector z can be dened as
z = [z1 z2 . . . . . . zNt ]T

(21)

Assuming noise free transmission, the active transmit antenna

will correspond to the index of the maximum absolute value
element in (21) [9]. However, the aforementioned condition is
not always satised in non-ideal conditions, thus leading to
incorrect transmit antenna estimation. Depending on the
severity of the noise, the active transmit antenna may
correspond to the index of the second or even Nt th largest
absolute value element in (21). In view of this fact, the rst
stage of MS selects the N most probable estimates of
transmit antenna index as opposed to a single estimate. The
N transmit antenna estimates and their associated M-QAM
constellations are then input to the second stage, which
applies the optimal SM detection rule to obtain a nal
estimate of transmit antenna index and modulated symbol.
The N estimates of transmit antenna index are obtained by
evaluating
j = argmaxN {|zj |}

(22)

where the argmaxN operator selects the N values of j which

j
maximise |zj | and arranges them in a vector j.
Hence, vector j comprises N estimates of transmit antenna
index and can be represented by
j = [ j1 . . . . . . jN ]

Simulation results

The aim of this section is to rst validate the analytical

frameworks developed Section 3 and then provide a
performance comparison between the various SM detection
schemes. Monte Carlo simulations are performed over an
i.i.d Rayleigh at fading channel with AWGN, where the
average BER is plotted against the average SNR at each
receive antenna. The parameters regarding the AWGN and
channel are consistent with those dened in Section 2.1.
The following were assumed during simulation: Gray coded
M-QAM constellation; full channel knowledge at receiver;
transmit and receive antennas are separated wide enough to
avoid correlation and the total transmit power is the same
for all transmissions.
5.1 Analytical and simulated BER of symbol
estimation
Fig. 3 shows the analytical and simulated average BER of the
symbol estimation process for 2 4 M-QAM SM-OD. The
relation in (12) assumes that one symbol error is equivalent
to one bit error. However, this relation is generally incorrect
at low SNR values, where a single symbol error may
correspond to multiple bit errors. This explains why the
analytical BER underestimates the simulated BER at low
SNR values. For high SNR values, the probability of a
symbol error with multiple bit errors is negligible. As a
result, both analytical and simulated BER show a close
match in the high SNR region.
5.2 Analytical and simulated BER of antenna
estimation
Fig. 4 shows the analytical and simulated average BER of the
transmit antenna estimation process for 2 4 M-QAM SMOD. The analytical bounds are tight over the entire range of
considered SNR, therefore (19) does not change the
direction of the inequality in (9).

(23)

where j1 corresponds to the most probable transmit antenna

index estimate and jN the least probable estimate.
4.2

Second stage: optimal SM detector

The second stage of MS detection utilises the optimal SM

detector to obtain a nal estimate of transmit antenna index
and modulated symbol as follows


[jMS , xq MS ] = argmax pgjlq 2F 2Re{yH gjlq }

(24)

jl ,q

where p is the average SNR at each receive antenna,

gjlq = hjl xq with l [ [1:N ] and q [ [1:M ]. jMS and xq MS
1372
& The Institution of Engineering and Technology 2011

Fig. 3 2 4 M-QAM SM analytical and simulated Pd comparison

IET Commun., 2011, Vol. 5, Iss. 10, pp. 13681376
doi: 10.1049/iet-com.2010.0667

www.ietdl.org

Fig. 4 2 4 M-QAM SM analytical and simulated Pa comparison

5.3 Analytical and simulated BER of M-QAM

SM-OD
The analytical and simulated average BER for 2 4 M-QAM
SM-OD is shown in Fig. 5. In all results, it is clear that the
theoretical lower bounds closely predict BER performance
in the high SNR region.
5.4 Performance comparison of various
SM detection scheme
Fig. 6 shows the BER performance of 4 4 16-QAM SM
using the various SM detection schemes discussed in this
paper. The rst notable observation is that the sub-optimal
SM detector exhibits reduced performance (4 dB at a BER
of 1025) as compared to SM-OD. It can also be seen that the
N 1 MS detector achieves the same performance as suboptimal detection. The rst stage of N 1 MS detection
computes a single transmit antenna index estimate by
evaluating (20) and (22). These are equivalent to the suboptimal antenna detection rules given by (4) and (5). In the
second stage, the N 1 MS detector computes (24) in order
to obtain a symbol estimate. It can be shown that (24) is
equivalent to the sub-optimal detection rule (7). Hence, the

Fig. 5 2 4 M-QAM SM analytical and simulated Pe comparison

IET Commun., 2011, Vol. 5, Iss. 10, pp. 13681376
doi: 10.1049/iet-com.2010.0667

Fig. 6 BER Performance of 4 4 16-QAM SM with various SM

detectors

N 1 MS detector reduces to sub-optimal SM detection,

since both schemes use equivalent rules for detection. The
MS detector with N 2 or 3 achieves virtually the same
BER performance as SM-OD. The reason for this behaviour
is that the MS detection scheme utilises an optimal detector
in its second stage. This optimal detector operates on a
restricted input of most probable transmit antenna and
symbol estimates. As a result, the MS detector achieves
optimal performance at a reduced complexity.
The average BER performance of 4 4 64-QAM SM
using the various SM detection schemes is shown in Fig. 7.
Similar to the results in Fig. 6, the N 1 MS detector
performs as the sub-optimal detector and the N 2 or 3
MS detector performs the same as the SM-OD scheme.

Complexity analysis

The complexity of sub-optimal, optimal and MS detection

schemes are compared in this section. The complexity
computation is similar to the analysis performed in [9, 20],

Fig. 7 BER Performance of 4 4 64-QAM SM with various SM

detectors
1373

& The Institution of Engineering and Technology 2011

www.ietdl.org
where only addition and multiplication of complex numbers
are considered as operations.

term of (8) has complexity [11]

dopt
6.1

Sub-optimal SM detector

The sub-optimal SM detection scheme comprises transmit

antenna and symbol estimation processes. The antenna
detection rules are given by (4) and (5). It was shown in [9]
that the numerator of (4) requires Nr complex
multiplications and (Nr 1) complex additions. The
Frobenius norm in the denominator of (4) is obtained by
rst multiplying a vector of length Nr with its complex
conjugate and then taking the square root. This results in Nr
complex multiplications and zero complex additions. (4) is
evaluated for j [ [1:Nt ] and therefore needs a total of
2Nt Nr + Nt (Nr 1) complex operations. The absolute value
operation in (5) requires one complex multiplication and
zero complex additions. (5) is also evaluated for j [ [1:Nt ]
and thus requires Nt complex operations [20]. Summing the
complexity contributions of (4) and (5) yields the total
complexity for antenna detection as

dant = 2Nt Nr + Nt (Nr 1) + Nt = 3Nt Nr

(25)

The complexity of signal combining in (6) is negligible since

zj and hj F have been calculated previously in (4).
Furthermore, the multiplication of zj by hj F is a noncomplex operation that does not contribute to the overall
complexity.
The complexity of symbol estimation is obtained by
analysing (7). As in [11], the rst term is simplied by
splitting hj xq 2F = hj 2F |xq |2 . The complexity of hj 2F is
ignored since hj F has been computed previously in (4)
and the squaring of hj F is a non-complex operation. The
calculation of |xq |2 involves one complex multiplication and
zero complex additions. This operation is performed for
q [ [1:M ] and therefore needs M complex
Note
 operations.
that the complexity of computing
phj 2F |xq |2 is not
considered, since it requires the multiplication of noncomplex quantities. In the second term of (7), x xq is
evaluated for q [ [1:M ] and therefore requires M complex
multiplications and zero complex additions. Thus, the
complexity of symbol detection can be expressed as

dsymbol = 2M

(26)

Hence, the total computational complexity of sub-optimal SM

detection is given by

dsubopt = dant + dsymbol = 3Nt Nr + 2M

6.2

(27)

Optimal SM detector

The complexity of SM-OD is obtained by analysing the ML

based detection rule given by (8). As in [11], the rst term
is simplied by splitting hj xq 2F = hj 2F |xq |2 . The
Frobenius norm operation hj 2F requires Nr complex
multiplications and zero complex additions. hj 2F is
evaluated for j [ [1:Nt ] and therefore needs Nt Nr complex
multiplications and zero complex additions. The
computation of |xq |2 requires 1 complex multiplication and
zero complex additions. This operation is performed for
q [ [1:M ]
and
therefore
involves
M
complex
multiplications and zero complex additions. Hence, the rst
1374
& The Institution of Engineering and Technology 2011

term1

= Nt Nr + M

(28)

The complexity of term two in (8) is dependent on the

calculation of (yH hj )xq . The yH hj operation is performed for
j [ [1:Nt ] and therefore needs Nt Nr complex multiplications
and Nt (Nr 1) complex additions. It can be shown that the
multiplication of yH hj by xq requires Nt M complex
multiplications [11] and zero complex additions. Hence, the
second term has complexity

dopt

term2

= Nt (2Nr + M 1)

(29)

Therefore, the total computational complexity of SM-OD is

given by

dopt = dopt
6.3

term1

+ dopt

term2

= Nt (3Nr + M 1) + M (30)

MS detector

The MS scheme consists of two stages. The rst stage of MS

detection utilises the sub-optimal antenna detection rules
given by (20) and (22). As shown in (25), the complexity
of the rst stage given by

dFirst

Stage

= dant = 3Nt Nr

(31)

In the second stage, the MS detector applies the optimal

SM detection rule (24). Term one of (24) is simplied by
splitting hjl xq 2F = hjl 2F |xq |2 [11]. The calculation of
|xq |2 requires one complex multiplication and zero complex
additions. |xq |2 is evaluated for q [ [1:M ] and therefore
term one needs M complex operations. Note that the
computation of hjl 2F is not considered, since hjl F has
been computed previously in (20). Furthermore, the
squaring of hjl F is a non-complex operation that does not
contribute to the overall complexity. The complexity of
term two in (24) can be obtained from (29) by replacing Nt
with N and this yields N(2Nr + M 1).Therefore, the
computational complexity of the second stage is given by

dsecond

stage

= M + N(2Nr + M 1)

(32)

Summing the complexity contributions of the rst and

second stages yields the total complexity of MS detection as

dMS = dfirst

stage

+ dsecond

stage

= 3Nt Nr + N(2Nr + M 1) + M

(33)

Fig. 8 shows the complexity of the various SM detection

schemes for xed 4 4 antenna conguration and variable
M-QAM constellation size. It can be seen that optimal, suboptimal and MS detection schemes have comparable
complexity for low order M-QAM (M 4). However, the
optimal detector requires signicantly more complex
operations with high order M-QAM (M 16). For example,
at M 128 the sub-optimal SM detection requires 304
complex operations. For the same modulation order, the
optimal detector requires 684 complex operations (125%
increase in complexity). Therefore, SM-OD is not practical
for high order M-QAM SM systems. In Section 5.4,
simulation results demonstrated that both N 2 and N 3
MS detectors achieve similar performance to SM-OD.
IET Commun., 2011, Vol. 5, Iss. 10, pp. 13681376
doi: 10.1049/iet-com.2010.0667

www.ietdl.org
9

Fig. 8 Complexity of sub-optimal, optimal and MS detection

schemes for the 4 4 M-QAM SM conguration
Table 1

Complexity of sub-optimal, optimal and MS detectors

for the 4 4 M-QAM SM configuration

4 4 SM
sub-optimal
optimal
MS N 2
MS N 3

16 QAM

64 QAM

128 QAM

80
124
110
133

176
364
254
325

304
684
446
581

However, the N 2 MS detector is less complex as compared

to N 3 MS and SM-OD. Generally, it can be stated that for
any 4 Nr M-QAM SM system with M 16 the N 2 MS
detector offers optimal performance at signicantly reduced
complexity as compared to SM-OD. To quantify our
ndings numerical complexity results are shown in Table 1.
As compared to SM-OD, the N 2 MS detector reduces
complexity by 11, 30 and 35% for the 16, 64 and 128
QAM constellations, respectively. On the other hand, suboptimal detection has lower complexity than N 2 MS
detection but this is offset by reduced performance, as
shown in Section 5.4.

Conclusion

A theoretical performance bound to quantify the average BER

of M-QAM SM-OD over i.i.d Rayleigh at fading channels
was derived in this paper. It was demonstrated that the
theoretical performance bound closely estimates the
simulated BER in the high SNR region, thereby validating
analytical frameworks. A novel SM detector for high order
M-QAM SM schemes (M 16) was also proposed in this
paper. Complexity analyses and Monte Carlo simulations
have shown that the MS detector with N 2 is optimal for
the 4 Nr SM conguration, since it attains the same BER
performance and up to a 35% reduction in receiver
complexity as compared to SM-OD.

Acknowledgment

The nancial assistance of the National Research Foundation

(NRF) towards this research is hereby acknowledged. Opinions
expressed and conclusions arrived at, are those of the author
and are not necessarily to be attributed to the NRF.
IET Commun., 2011, Vol. 5, Iss. 10, pp. 13681376
doi: 10.1049/iet-com.2010.0667

References

1 Telatar, E.: Capacity of multi-antenna Gaussian channels, Eur. Trans.

Telecommun., 1999, 10, (6), pp. 558595
2 Foschini, G.J., Gans, M.J.: On limits of wireless communications in a
fading environment when using multiple antennas, Wirel. Pers.
Commun., 1998, 1, (6), pp. 311 335
3 Ran, R., Yang, J., Kim, D.: Multimode precoder design for STBC with
limited feedback in MIMO based wireless communication system,
IEICE Electron. Express, 2005, 1, (2), pp. 222229
4 Alamouti, S.: A simple transmit diversity technique for wireless
communications, IEEE J. Sel. Areas Commun., 1998, 16, (8),
pp. 1451 1458
5 Tarokh, V., Seshadri, N., Calderbank, A.: Space-time codes for high
data rate wireless communication: performance criterion and code
construction, IEEE Trans. Inf. Theory, 1998, 44, (2), pp. 744 765
6 Wolniansky, P., Foschini, G., Golden, G., Valenzuela, R.: V-BLAST:
an architecture for realizing very high data rates over the richscattering wireless channel. Proc. URSI Int. Symp. on Signals,
Systems and Electronics, Pisa, 1998, pp. 295 300
7 Goldsmith, A., Jafar, S., Jindal, N., Vishwanath, S.: Capacity limits of
MIMO channels, IEEE J. Sel. Areas Commun., 2003, 21, (5),
pp. 684702
8 Damen, M., Abdi, A., Kaveh, M.: On the effect of correlated fading on
several space-time coding and detection schemes. Proc. IEEE Vehicular
Technol. Conf., Atlantic City, NJ, 2001, pp. 13 16
9 Mesleh, R., Haas, H., Ahn, C.W., Yun, S.: Spatial modulation a new
low complexity spectral efciency enhancing technique. Proc.
ChinaCOM, Beijing, 2006, pp. 1 5
10 Mesleh, R., Haas, H., Sinanovic, S., Ahn, C.W., Yun, S.: Spatial
modulation, IEEE Trans. Veh. Technol., 2008, 57, (4), pp. 2228 2241
11 Jeganathan, J., Ghrayeb, A., Szczecinski, L.: Spatial modulation:
optimal detection and performance analysis, IEEE Commun. Lett.,
2008, 12, (8), pp. 545547
12 Mesleh, R., Stefan, I., Haas, H., Grant, P.M.: On the performance of
trellis coded spatial modulation. Proc. Int. ITG Workshop on Smart
Antennas, Berlin, 2009, pp. 235241
13 Hwang, S., Jeon, S., Lee, S., Seo, J.: Soft-output ML detector for
OFDM spatial modulation systems, IEICE Electron. Express, 2009,
6, (19), pp. 14261431
14 Renzo, M.D., Haas, H.: Performance comparison of different spatial
modulation schemes in correlated fading channels. Proc. IEEE ICC.
Conf., Cape Town, 2010, pp. 1 6
15 Xu, H.: Symbol error probability for generalized selection combining
reception of M-QAM, SAIEE Afr. Res. J., 2009, 100, (3), pp. 68 71
16 Al-Shahrani, I.: Performance of M-QAM over generalized mobile
fading channels using MRC diversity. M.S thesis King Saud
University, Riya, Saudi Arabia, February 2007
17 Proakis, J.G.: Digital communications (McGraw-Hill, New York,
2001, 4th edn.)
18 Jeganathan, J., Ghrayeb, A., Szczecinski, L., Ceron, A.: Space shift
keying modulation for MIMO channels, IEEE Trans. Wirel.
Commun., 2009, 8, (7), pp. 36923703
19 Alouini, M.S., Goldsmith, A.J.: A unied approach for calculating error
rates of linearly modulated signals over generalized fading channels,
IEEE Trans. Commun., 1999, 47, (9), pp. 13241334
20 Mesleh, R.: Spatial modulation: a spatial multiplexing technique for
efcient wireless data transmission. Ph.D thesis Jacobs University,
Bremen, Germany, June 2007
21 Beauregard, F.: Linear algebra (Prentice-Hall, New York, 1995, 3rd
edn.)
22 Goldsmith, A.J.: Wireless communications (Cambridge University
Press, New York, 2005, 1st edn.)

10

Appendix

The received signal vector is given by

y=


phj xq + n

(34)

The derivation of (14) is as follows





P(xjq  xjq |H) = P y phj xq F , y phj xq F
(35)
1375

& The Institution of Engineering and Technology 2011

www.ietdl.org
Substituting (34) in (35) yields:
 

P(xjq  xjq |H) = P  phj xq + n phjxq F



,  phj xq + n phj xq F
 


= P ( phj xq phjxq ) + nF , nF
(36)

Fig. 9 Noise projection

The triangle inequality states that [21]





 phj xq phj xq F nF ( phj xq phj xq )


+ nF  phj xq phj xq F + nF
Hence
 


P(xjq  xjq |H) = P  phj xq phj xq F nF , nF




( phj xq phj xq )F
= P nF .
2
(37)
Therefore, the conditional PEP (37) is equivalent to the
probability

that
 noise vector n lies closer to the vector
( phj xq phj xq ) than to the origin. This probability is
dependent solely on the projectionof n onto
 the line
joining the origin and the point phj xq phj xq [22,
pp. 138 139], as shown in Fig. 9 [redrawn from 22].
The projection of n onto this one dimensional line is a one
dimensional Gaussian random variable n distributed
according to N (0, 1/2).

1376
& The Institution of Engineering and Technology 2011

Thus



d
P(xjq  xjq |H) = P n .
2


1
1
v2
exp 2 dv
= 
2s
2ps2 d/2
 
d
=Q
2s

(38)



where d = ( phj xq phj xq )F
Substituting s = 12 in (38) yields



d
p
P(xjq  xjq |H) = Q  = Q
(h x hj xq )F (39)
2 j q
2


Let k = p2 (hj xq hj xq )2F

Then

P(xjq  xjq |H) = Q( k )

(40)

IET Commun., 2011, Vol. 5, Iss. 10, pp. 13681376

doi: 10.1049/iet-com.2010.0667