You are on page 1of 4

IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 2, NO.

4, AUGUST 2013 383

On the Mutual Information and Precoding for


Spatial Modulation with Finite Alphabet
Xinrong Guan, Student Member, IEEE, Yueming Cai, Senior Member, IEEE, and Weiwei Yang, Member, IEEE

Abstract—In this letter, we investigate the effect of finite antenna selection


alphabet inputs on the performance of spatial modulation (SM)
for multiple-input-single-output (MISO) channels. The closed- p bits hm D m ,i v
source data ^h1 , h2 ,..., hM `
u
form expression of the mutual information is first derived in
an information-theoretic framework. With careful analysis of
the expression, we develop a precoding scheme to improve
the performance of SM. Utilizing the theory of traditional p  q bits
q bits
^ x1 , x2 ,..., xN `
 y

constellation design, the precoding coefficient is obtained by xi


maximizing the minimum Euclidean distance. Numerical results
demonstrate substantial difference in the mutual information symbol selection
between SM with finite alphabet inputs and SM with Gaussian
inputs, and show that our precoding scheme achieves significant Fig. 1. Space modulation with precoding, in which selecting antenna is
gains. equivalent to selecting channel.
Index Terms—Spatial modulation, finite alphabet, mutual in-
formation, precoding.
[8]. Rather, the inputs are usually drawn from finite discrete
I. I NTRODUCTION constellations such as pulse amplitude modulation (PAM),
quadrature amplitude modulation (QAM), and phase shift key-
S a low-complexity technique for implementation of
A multiple antenna systems, spatial modulation (SM)
achieves a spatial multiplexing gain without inter channel
ing (PSK) modulation. Furthermore, for SM, finite alphabet is
inherently able to work since the index of the transmit antenna
itself, which is used to convey part information, is discrete
interference (ICI) and requires no synchronization between and finite. Due to the constraint of finite alphabet inputs,
the transmit antennas [1]. Based on the uniqueness and ran- the capacity calculation would differ from that uses Gaussian
domness properties of the wireless channel, the index of the inputs. Therefore, SM with finite alphabet is a promising field
transmit antenna can be exploited as an additional dimension of practical significance, and this sparks our research.
to convey part of the information bits, which helps to realize In this letter, the mutual information is first derived in
higher data rates. Along its history, the principle of SM has an information-theoretic framework, and the impact of finite
been known in various forms, such as space shift keying alphabet inputs on the performance of SM is investigated. A
(SSK) modulation, orthogonal spatial division multiplexing loss on the mutual information is yielded due to the use of
(OSDM), and information-guided channel-hopping (IGCH) finite alphabet. Moreover, a comparison between SM and other
modulation [2]. Until recent works, much research interest has multiple-input-single-output (MISO) schemes is made. After-
been focused on the issues such as optimal receiver design wards, a specially designed precoding scheme is developed to
and error performance analysis [3-5]. In [3], the authors have improve the performance of SM with finite alphabet.
developed the maximum-likelihood (ML)-optimum receiver
for SM. In [4] and [5], the bit error probability of SM over II. S YSTEM M ODEL FOR SM WITH F INITE A LPHABET
Nakagami-m fading and generalized fading channels has been
Consider SM in a MISO system with M (M = 2p )
derived. However, the topic of computing the capacity in an
transmit antennas and one receive antenna. Let h ∈ H
information-theoretic framework has been rarely involved.
(H = {h1 , h2 , ..., hM }) be the link from the transmit antenna
An enlightening result has been proposed in [6], in which
to the receiver. We assume that the elements in H are
the capacity of SM over Rayleigh fading channels has been de-
independent from each other. As a condition for precoding,
rived. However, this work was based on continuous Gaussian
the channel state information (CSI) is completely known at
source signals. Although Gaussian inputs are optimum, they
the receiver and transmitter. A joint ML method is used for
can never be realized in a practical communication system [7]
detection. Let x ∈ X (X = {x1 , x2 , ..., xN }, N = 2q ) denote
Manuscript received December 15, 2012. The associate editor coordinating the symbol drawn from the conventional equiprobable discrete
the review of this letter and approving it for publication was H. Viswanathan. modulation constellation. And we assume a power constraint
This work is supported by the Important National Science & Technology of unity, i.e. E(|x|2 ) = 1, where |∗| denotes the modulus
Specific Project (Grant No. 2010ZX03006-002-04), and the NSF of China
(Grant No. 61001107). operator.
The authors are with the Department of Wireless Communications, As depicted in Fig. 1, the source data block is divided into
Institute of Communications Engineering, PLA University of Science two sub-blocks, which contains p bits and q bits, respectively.
and Technology, Nanjing, 210007, China (e-mail: ywygxr@yahoo.com.cn,
caiym@vip.sina.com, yww_1010@yahoo.com.cn). The first sub-block is used to select the antenna, i.e. select the
Digital Object Identifier 10.1109/WCL.2013.050613.120923 channel, whereas the second one is used to choose a symbol
2162-2337/13$31.00 
c 2013 IEEE
384 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 2, NO. 4, AUGUST 2013

in the N points modulation constellation. In other words, the substituting (6) and (7) into (5), the mutual information for
source information bits are mapped into two information car- SM with finite alphabet can be calculated by
rying units: one is the index of the selected transmit antenna, 1

M  N
i.e. m, and the other is the chosen symbol xi . With precoding, I (x, h; y) = log2 M N − MN Ev
 
m=1 i=1 
the signal xi transmitted by the mth antenna is multiplied by M  N   2 
the precoding coefficient αm,i . Each antenna will be selected log2 1 +  m2 ,i2
exp −γ dm,i + v − |v| 2

with the same probability, i.e. p (h = hm ) = 1/M, ∀m. By m2 =1 i2 =1


detecting hm and xi , the p + q bits source data can be (m2 ,i2 )=(m,i)
  
recovered at the receiver. f (γ)
(8)
III. M UTUAL I NFORMATION AND P RECODING in which (m2 , i2 ) = (m, i) denotes that the events m2 = m
and m2 = m do not occur simultaneously, and γ = 1/σ 2
A. Mutual information represents the transmitting signal to noise ratio (SNR). As
During the following process of deriving the mutual infor- (8) reveals, the mutual information of SM with finite al-
mation, precoding would not be taken into account for brevity. phabet is an increasing function of γ. However, we have
For example, the mth transmit antenna is selected to emit the f (γ) → 1 when γ → +∞, thus I (x, h; y) is upper bounded
symbol xi , thus the received signal will be expressed by by log2 M N . This implies that the finite alphabet input yields
a loss on the mutual information as compared to the Gaussian
y = hm xi + v (1) input cases.
where v is the complex additive white Gaussian noise On the other hand, the traditional MISO beamforming (BF)
(AWGN) at the receiver, with zero mean and variance of scheme and choosing the best-gain channel (CBC) scheme
σ 2 . Obviously, the inputs are discrete variables, whereas are also presented for comparison. Similar to (8), the mutual
the output is a continuous random variable since the ex- information for BF and CBC is written as
N
istence of the noise. Since every antenna is selected with I (x; y) = log2 N − N1 Ev
the same probability 1/M and the variable x is uniformly  
i=1  (9)
distributed, the received signal follows a Gaussian distribu- N 2
|A(xi −xi2 )+v| −|v|2
log2 exp − σ 2
tion with the conditional probability density function (PDF) i2 =1
p (y |h = hm , x = xi ), p (y |x = xi ) and the marginal PDF For BF, we have A = hwT , where h = [h1 , h2 , ..., hM ] and
p (y) given as w = h∗ / |h| is the beamforming weights vector. For CBC,
 
2
1 |y −hmxi | we have A = hB , where hB = arg max {|hm |} is the best-
p(y |h = hm , x = xi ) = exp − (2) hm
πσ 2 σ2 gain channel. Obviously, the maximum achievable mutual
  information for BF and CBC is both log2 N . Recalling that the
1  1 |y −hmxi |2
M
p(y |x = xi ) = exp − (3) maximum achievable mutual information for SM is log2 M N ,
M m=1 πσ 2 σ2 SM outperforms BF and CBC for high SNR region with the
  multiplexing gain up to log2 M .
1  1
M N 2
|y − hm xi |
p(y) = exp − (4)
M N m=1 i=1 πσ 2 σ2
B. Precoding
The mutual information between the input and output As concluded in last subsection, the upper bound of the
variables can be expressed by [6] mutual information for SM is log2 M N . However, this bound
I (x, h; y) = I (h; y |x ) + I (x; y) (5) may be not achievable in some extreme cases. For instance,
in which I (h; y |x ) stands for the conditional mutual infor- if it holds that hm = hm2 (∀m, m2 ), then we always have
mation between h and y when x is given , and I (x; y) stands dm2 = (hm −hm2 ) xi = 0, thus I (h;y |x ) can be rewritten
m
as

for the mutual information between x and y. Using (2), (3) 1 M  N 
M

and (4), I (h; y |x ) and I (x; y) are given by (6) and (7) I(h; y |x)= log2 M − Ev log2 exp(0) = 0 (10)
M N m=1i=1 m2 =1
1
M  N
I (h; y |x ) = log2 M − MN Correspondingly, the maximum value of I (x, h; y) is no
 m=1 i=1
  (6) longer log2 M N , but log2 N . This also implies that the infor-
M
|dm2 +v|2 −|v|2
Ev log2 exp − m σ2 mation bits conveyed on the index of the antenna is no longer
m2 =1 uniquely decodable and the multiplexing gain introduced by
1  
M N SM vanishes.
I (x; y) = log2 N − MN To avoid degrading the performance of SM such badly, the
⎡ m=1 i=1  2 ⎤

M N
exp −
| d 2 2 +v|
m ,i
m,i
(7)
uniqueness of the channels should be enlarged. This inspires
⎢ m2 =1 i2 =1
σ2
⎥ us to develop a precoding scheme with the assistance of
Ev ⎣log2  2  ⎦

M
exp −
| dm2 +v|
m
CSI. With precoding, the ith symbol transmitted by the mth
σ2
m2 =1 antenna is multiplied by the coefficient αm,i , so the process
in which dm 2 ,i2
= h x
m i −h x
m2 i2 , d m2
= (h m −h m2 ) x i , and of SM can be viewed as a mapping ψ given by
m,i m
Ev (∗) denotes the expectation operator with respect to v. By ψ : H × X → R where R= {αm,i hm xi |∀m, i } (11)
GUAN et al.: ON THE MUTUAL INFORMATION AND PRECODING FOR SPATIAL MODULATION WITH FINITE ALPHABET 385

where R can be seen as the received signal space for non- from antenna 1 0.8
from antenna 1
1 from antenna 2 from antenna 2
noise channel, in which there are totally M N points. The from antenna 3 0.6
from antenna 3
from antenna 4
Euclidean distance between any two of them is denoted by 0.4 from antenna 4

 m2 ,i2  0.5

dm,i  = |αm,i hm xi −αm2 ,i2 hm2 xi2 |. 0.2

0 0

Generally, it is quite difficult to solve the expectation with −0.2


−0.5
respect to v in (8). Therefore, the optimizing object for the −0.4

precoding can not be intuitively obtained from (8). Instead, a −1


−0.6

lower bound of the mutual information is presented in (12) −0.8

to give more insights into how the object function for the −1 −0.5 0 0.5 1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8
(a) without precoding (b) with precoding
precoding is derived and how the precoding coefficient αm,i
is designed. It should be noticed that the motivation of the Fig. 2. Signal space diagrams for SM, without precoding and with precoding,
finally proposed precoding scheme is to improve the lower respectively.
bound I (x, h; y)Low , and the numerical result demonstrates
that it is consistent with the object of improving I (x, h; y).
can be rewritten as  
Lower bound: The mutual information of SM with finite  2 ,i2 
max dmm,i 
alphabet can be lower bounded by αm,i   (16)

M  2 2
1
N
s.t.E |r| = h /M
I (x, h; y)Low =log2 M N−(log2 e−1)− MN
  m ,i m=1 
i=1
(12) This can be considered as a traditional constellation design

M 
N
|dm,i2 2 |
2
problem. Considering PAM, PSK and QAM constellations,
log2 exp − 2σ2
m2 =1 i2 =1 the last one provides the maximum dmin , within the same
power constraint [10, Table 3.2-1]. At this point, we rearrange
Proof: The mutual information in (8) can be rewritten as the M N points in the space R, to form an M N −QAM
  2 
I (x, h; y) = log2 M N − Ev log2 exp |v|
σ2 − MN 1 constellation. We denote a M N −QAM constellation by
     Q = {q1 , q2 , ...qMN }, with E(|q|2 ) = 1. Based on the above
M  N 
M 
N −|dm,i
m2 ,i2
+v |
2

Ev log2 exp σ2 analysis, the precoding coefficient αm,i is determined by


m=1 i=1 m2 =1 i2 =1
(13) q(m−1)N +i h
αm,i = √ (17)
in which the second term on the right-hand side equals log2 e. hm xi M
Since log (x) is a concave function, using Jensen’s inequality Since the CSI is also perfectly known at the receiver, αm,i
the lower bound of the third term can be expressed as can be determined independently during detection. Accord-
  M N  m ,i  
  −|dm,i 2 2
+v |
2 ingly, the ML detection is written as
 
Ev log2 exp σ2
 M mN 2 =1 i2 =1  m ,i 2 2 ĥm , x̂i = arg min (|y − αm,i hm xi |) (18)
(m,i)
  1  |dm,i2 2+v| +|v|
≤ log2 πσ2 v exp − σ2 dv (14) For QPSK, we take a 4 × 1 channel with the chan-
m2M=1 i2 N
=1   nel coefficients h1 = 0.7734+j0.3318 , h2 = 0.3187−j0.2670 ,
  1 |d m 2 ,i2 2
|
= log2 2 exp − m,i
2σ2
h3 = − 2.5723 − j0.1602 and h4 = 1.0631+j0.3732 for in-
m2 =1 i2 =1 stance. The corresponding diagram for space R without pre-
Arranging terms, the lower bound in (12) is obtained. It should coding is depicted as Fig.2 (a). But if we amplify and rotate
be noticed that a constant gap (log2 e−1) exists between the signal points, the diagram can be shown as Fig.2 (b). Ob-
I (x, h; y) and I(x,h; y)Low at low and high SNR regions, viously, the M N signal points in the space R are maximally
which is similar to that been observed in [9]. separated and the minimum distance dmin is increased.
Directly, the problem of maximizing I (x, h; y)Low is equiv-
IV. N UMERICAL R ESULTS
alent to the following problem
   All of the results demonstrated below are based on a Monte-

M  N M 
N
|dm2 ,i2 |
2

min log2 exp − m,i 2σ2 carlo simulation over 10000 independent channel realizations.
αm,i m=1 i=1
  m2 =1 i2 =1 (15) Fig. 3 depicts the mutual information comparison between
2 2
s.t. E |r| = h /M N-PSK inputs SM and Gaussian inputs SM [6]. In Gaussian
where r ∈ R, and ∗ denotes the Frobenius norm. With- inputs case, the mutual information always increases as the
out precoding, the average power of signal points in R SNR increases, whereas finite discrete inputs lead to a loss
2 on mutual information. But this performance gap becomes
equals h /M . To guarantee the average power
 remains
smaller and smaller as the modulation order and the number of
the same, we have the constraint E |r| = h 2 /M .
2
antennas increase. Moreover, it is very interesting to observe

Close observation shows 
that the minimum distance, i.e.
 m2 ,i2  that the mutual information with Gaussian inputs varies little
dmin = min dm,i  is the dominant term in (15), as M increases. On the contrary, it increases by 1 bit/s/Hz as
(m2 ,i2 )=(m,i)
thus it has an important impact on the performance of SM. M varies from 2 to 4 in the finite alphabet inputs scenario.
Therefore, we intend to design αm,i to provide the maximum The mutual information for BF, CBC and SM versus the
possible minimum distance, i.e. the optimizing problem in (15) SNR with BPSK and QPSK inputs is depicted in Fig. 4. It is
386 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 2, NO. 4, AUGUST 2013

10 3

7.25 SM
9 M=2 BF
Gaussian inputs 2.5
M=4 CBC
8 7.2
Mutual infomation (bits/s/Hz)

M=8 M=2 QPSK

Mutual infomation (bits/s/Hz)


7 24 M=4
2
16PSK
6

5 1.5

4
1
3

2
0.5
1 QPSK BPSK
0 0
−20 −16 −12 −8 −4 0 4 8 12 16 20 24 28 32 −20 −15 −10 −5 0 5 10 15 20
SNR(1/σ2,dB) 2
SNR(1/σ ,dB)

Fig. 3. Mutual information comparison between BF, CBC, and SM for Fig. 4. Mutual information comparison between N-PSK (N = 4, 16) inputs
M × 1 (M = 2, 4) channels. SM and Gaussian inputs SM for M × 1 (M = 2, 4, 8) channels.

3
mutual information
interesting to observe that there exists an intersection of curves lower bound
2.5
for BF, CBC and SM. For low-SNR, due to the obtained spatial without precoding

Mutual infomation (bits/s/Hz)


diversity gain, BF and CBC outperform SM. But it also clearly precoding in [8]
2 proposed precoding
shows that, for high SNR, SM outperforms BF and CBC
largely owing to the multiplexing gain that logarithmically 1.5
increases with the number of transmit antennas. For example,
the maximum mutual information for BF and CBC with BPSK 1

inputs is 1 bit/s/Hz, but the SM scheme achieves 2 bits/s/Hz M =2


0.5
and 3 bits/s/Hz for M = 2 and M = 4, respectively. This
observation is consistent with the results derived in (8) and 0
M =4
(9), i.e. SM has up to log2 M bits/s/Hz multiplexing gain.
In Fig. 5, a comparison has been made between our −0.5
−20 −15 −10 −5 0 5 10 15 20 25
proposed precoding scheme and that in [8] designed for SNR(1/σ2,dB)
two-user multiple access (MAC) channel. In [8], each user
is equipped with single antenna and finite input alphabets, Fig. 5. Mutual information and the lower bound with BPSK input for M ×
1 (M = 2, 4) channels.
thus its precoding scheme can also apply to our SM model.
The mutual information and the lower bound is plotted. It
clearly shows that our proposed precoding scheme can yield
improved performance. Moreover, it also can be observed that [3] J. Jeganathan, A. Ghrayeb, and L. Szczecinski, “Spatial modulation:
optimal detection and performance analysis,” IEEE Commun. Lett., vol.
as the number of antennas increases, the performance of the 12, no. 8, pp. 545–547, Aug. 2008.
precoding scheme in [8] degrades. As mentioned in Sec. III-B, [4] M. Di Renzo and H. Haas, “Bit error probability of space modulation
there is a constant gap (log2 e − 1) exists between I (x, h; y) over Nakagami-m fading: asymptotic analysis,” IEEE Commun. Lett., vol.
15, no. 10, pp. 1026–1028, Oct. 2011.
and I (x, h; y)Low at low and high SNR regions. [5] M. Di Renzo and H. Haas, “Bit error probability of spatial modulation
(SM-) MIMO over generalized fading channels,” IEEE Trans. Veh.
Technol., vol. 61, no. 3, pp. 1124–1144, Mar. 2012.
V. C ONCLUSIONS [6] Y. Yang and B. Jiao, “Information-guided channel-hopping for high data
As an entirely new physical layer transmission technique, rate wireless communication,” IEEE Commun. Lett., vol. 12, no. 4, pp.
225–227, Apr. 2008.
SM has received an upsurge of research interest recently. [7] A. Lozano, A. M. Tulino, and S. Verdu, “Optimum power allocation
In this letter, we have first investigated the effect of finite for parallel Gaussian channels with arbitrary input distributions,” IEEE
alphabet inputs on the mutual information of SM by deriving Trans. Inf. Theory, vol. 52, pp. 3033–3051, July 2006.
[8] J. Harshan and B. S. Rajan, “Finite signal-set capacity of two-user
the closed-form expressions. Moreover, a precoding scheme is Gaussian multiple access channel,” in Proc. 2008 ISIT, pp. 1203–1207.
proposed to improve the performance. [9] W. Zeng, C. Xiao, and J. Lu, “A low-complexity design of linear
precoding for MIMO channels with finite-alphabet inputs,” IEEE Wireless
Commun. Lett., vol. 1, no. 1, Feb. 2012.
R EFERENCES [10] J. Proakis and M. Salehi, Digital Communications, 5th edition. McGraw-
Hill, 2007.
[1] R. Mesleh, H. Haas, S. Sinanovic, C. W. Ahn, and S. Yun, “Spatial
modulation,” IEEE Trans. Veh. Technol., vol. 57, no. 4, pp. 2228–2241,
July 2008.
[2] M. D. Renzo, H. Haas, and P. M. Grant, “Spatial modulation for multiple-
antenna wireless systems—a survey,” IEEE Commun. Mag., vol. 49, no.
12, pp. 182–191, 2011.

You might also like