Layered Space-Time Codes for Wireless Communications

Using Multiple Transmit Antennas

Da-shan Shiu and Joseph M. Kahn

University of California at Berkeley

Abstract Multiple-antenna systems provide very high nals. An LST code is a channel code that is designed and pro-
capacity compared to single antenna systems in a Rayleigh cessed according to the LST architecture. An LST code is
fading environment. Space-time codes are channel codes constructed by assembling 1-D constituent codes. With the
designed to exploit this high capacity for multiple-antenna use of interference suppression and interference cancellation
systems without requiring instantaneous channel knowledge at
at the receiver, these constituent codes can be separated and
the transmitter. A practical concern for high data rate space-time
then decoded using conventional decoding algorithm devel-
codes is their decoding complexity. The decoding complexity with
ML criterion can be prohibitively large. In this paper, we focus oped for 1-D constituent codes, leading to a much lower
on layered space-time (LST) codes. Two types of LST codes, the decoding complexity compared to ML decoding. Other pos-
horizontally-layered space-time (HLST) codes and the sible low-complexity decoding techniques include sequential
diagonally-layered space-time (DLST) codes, are presented. We decoding (SD) [7] and multistage decoding [4] [8].
analyze the performance of both types of LST codes under slow In this paper, we analyze the performance of both horizon-
and fast fading conditions. We conclude that, in a slow fading tally-layered space-time (HLST) codes and diagonally-lay-
environment, DLST codes have superior performance over
ered space-time (DLST) codes. We will show that DLST
HLST codes. The design criteria for DLST codes are proposed.
codes can achieve better performance compared to HLST
codes. From our analysis, we propose the design criteria, i.e.,
I. INTRODUCTION
the criteria to choose the constituent code, for DLST codes
Recent studies have explored the ultimate limit of multiple- based on the truncated multidimensional effective code length
antenna systems from the information-theoretic point of view (TMEL) and the truncated multidimensional product distance
[1] - [3]. Consider a multiple-antenna system that has n trans- (TMPD) of the constituent code.
mitting and m receiving antennas. It is shown that, if the nar- The remainder of this paper is organized as follows. In Sec-
rowband slow fading channel can be modeled as an n m tion II, the background of multiple-antenna systems and
matrix with i.i.d. complex Gaussian random entries, the space-time codes is reviewed. In Section III, we introduce the
average channel capacity of such a system is approximately LST architecture. In Section IV, the performances of HLST
min(n, m)-times higher than that of a single antenna system and DLST codes are analyzed. The analysis leads to the
for the same overall transmitting power. design criteria for DLST codes. We also present an example
In most applications, a major obstacle to utilizing this high DLST code. Concluding remarks can be found in Section V.
throughput is that the transmitter cannot have the instanta-
neous information about the fading channel. The transmitter II. BACKGROUND
thus must employ a channel code that can guarantee good per-
In this paper, we focus on single-user to single-user com-
formance with the majority of possible channel realizations.
munication using multiple antennas at both ends over narrow-
Such a channel code is inherently multi-spatial-dimensional
band flat-fading channels. We refer to a multiple-antenna
and thus is called a space-time code [4] [5].
system in which the transmitter has n transmitting antennas
Aside from the consideration of combatting channel uncer- and the receiver has m receiving antennas as an (n, m) system.
tainty, another practical consideration for space-time codes is A general space-time code can be described as follows. The
the decoding complexity. As stated above, the channel encoder first applies the space-time code to the input informa-
capacity of a multiple-antenna system is approximately pro- tion bits to generate an n-row (possibly semi-infinite) matrix
portional to min(n, m). This means that each channel usage on C. The matrix C represents the signal that is to be transmitted
average can convey proportional to min(n, m) bits of informa- by the transmitter. Specifically, the kth row, th column ele-
tion. The complexity of decoding such a high data-rate ment of C, denoted by c k , represents the signal to be trans-
channel code using the maximal-likelihood (ML) criterion can mitted by antenna k at time slot . We emphasize that there is
be prohibitively high even if min(n, m) is just moderately no mechanism, such as time-, frequency-, or code-division
large; thus, space-time codes that admit high performance, multiplexing, employed to ensure that the signals transmitted
low complexity suboptimal decoding algorithms are desirable. by different transmitting antennas are orthogonal upon recep-
The layered space-time (LST) architecture proposed by tion by the receiver. The signal received by the receiving
Foschini in [6] is a framework of processing space-time sig- antenna l during the time slot is denoted by r l . This
received signal r l contains a superposition of transmitted sig- Fig. 1(c). Using our codeword matrix notation, to construct
nals c k , k = 0, 1, , n 1 , and an AWGN component v l . the DLST codeword matrix C, the output symbols from con-
For a narrowband flat-fading channel, the gain connecting stituent coder 0 are used to fill the leftmost NW-SE diagonal
transmitting antenna k and receiving antenna l at time can be of C, and the output symbols from constituent coder 1 are
denoted by a complex number h l k . Define the vectors1 used to fill the next diagonal, and so on. The layered structure
c = ( c 1 c 2 c n )' , r = ( r 1 r 2 r m )', v = ( v 1 v 2 v m )' , of both HLST and DLST codeword matrices are shown in Fig.
and the channel matrix H, ( H kl ) = h l k . The discrete-time 2.
input-output relation of an (n, m) multiple-antenna system If the data rate of the constituent code maintains constant
over a narrowband flat-fading channel can be written in the regardless of n, the data rate of an LST code is obviously pro-
following vector notation: portional to n.
r = H c + v . (1) B. Decoding
To decode an LST code, the received signal is processed
The following terminology is used in this paper. The matrix
along both the spatial and temporal dimensions. Here, we
C, a coded matrix output of the transmitter encoder, is referred
introduce the spatial signal processing in the LST architecture.
to as a space-time codeword matrix. A space-time codeword
Focus on a given instance in time, say . The transmitted
matrix can be thought of as a serial concatenation of n-tuples,
n-tuple is c, and the received m-tuple is r = H c + v . The
and an n-tuple is composed of n symbols. Note that the first
received signal r is a superposition of transmitted coded
row of the matrix C is indexed as row zero, not row one. To
symbols scaled by the channel gain and corrupted by AWGN.
facilitate the comparison between multiple transmit-antenna
The task here is to determine the values of the n components
systems using space-time codes and single transmit-antenna
systems using conventional 1-D channel codes at equal
average transmit powers (total over all transmit antennas), the CC 0 c2 c1 c0
average energy of an n-tuple is E, regardless of the spatial
information
dimensionality n. bits CC 1

III. THE LAYERED SPACE-TIME ARCHITECTURE CC 2

A. Encoding
(a)
The encoding process is illustrated in Fig. 1. In Fig. 1(a),
the input information bit sequence is first demultiplexed into n c3 c2 c1 c0 to antenna 0
subsequences, and each subsequence is subsequently encoded
by a 1-D encoder. These 1-D channel codes are referred to as to antenna 1
the constituent codes (CC). The output of the constituent
coder k is a sequence of symbols s k , = 0, 1, . In the LST to antenna 2
architecture, the multi-spatial dimensional codeword matrix C
is constructed by assigning these symbols to the slots of C in a
(b) HLST
systematic fashion with the goal of reducing the receiver com-
plexity.
One intuitive assignment rule is to simply place s k at the c3 c0 to antenna 0
kth row, th column slot of C. With this assignment rule, the
c4 c1 to antenna 1
output coded symbol from constituent encoder k is always
transmitted using the transmit antenna k. This is illustrated in c5 c2 to antenna 2
Fig. 1(b). Under this assignment rule, the space-time code-
word matrix has an obvious horizontally layered structure.
Therefore, it is referred to as the horizontally layered space- (c) DLST
time (HLST) architecture. HLST was originally proposed by
Foschini in [6]. Another assignment rule, also proposed in [6], Fig. 1. LST code encoding process. Here, n = 3. Each
is the diagonally layered space-time (DLST) architecture. In square represents a symbol. (a) The incoming information bit
sequence is first demultiplexed into n subsequences. Each
DLST, instead of always sending the output symbols from a
subsequence is then encoded using a constituent code. (b)
constituent coder to a particular transmit antenna, they are fed In HLST, the coded symbols from constituent encoder k are
to the n transmitting antennas in turn. DLST is illustrated in transmitted by antenna k. (c) In DLST, the coded symbols
from a constituent encoder are transmitted by the n
1. The
transpose and conjugate transpose of x are denoted by x' , transmitting antennas in turn.
and x , respectively.
 component. Because R is upper triangular,
Encoder 0 1 2 3 ...
y k = ( R kk ) c k + v ' k +
C= (3)
Encoder 0 1 2 3 ... { constribution from c k + 1, c k + 2, , c T }.
n

Encoder 0 1 2 3 ... Note that in (3) only the interferences from cl, l < k, are sup-
pressed in yk. The interference term in (3) is canceled by using
(a) the available decisions c k + 1, c k + 2, , c n 1 . Assuming that
these decisions are all correct, the decision variable z k is
Encoder ... z k = ( R kk ) c k + v ' k , k = 0, 1, , n 1. (4)

Encoder ... The complexity of the spatial processing described above is

O( n 2 + nm ) per transmitted n-tuple. Other interference sup-
... pression criteria, such as the linear least square criterion, can
Encoder
also be employed instead [9].

0 0 0 3 3 3 ... IV. LAYERED SPACE-TIME CODES

C= 0 1 1 1 4 4 4 ... A. HLST
0 0 2 2 2 5 5 5 To decode an HLST codeword matrix C, the receiver first
extracts the decision variables for the symbols of the bottom-
most row of C. The resulting decision variable sequence,
(b) { z n 1 }, = 0, 1, , is then passed to a conventional 1-D
Fig. 2. The layered structure of an LST codeword matrix. (a)
decoder of the corresponding constituent code to produce the
HLST. (b) DLST. decisions on the symbols of this row. The receiver uses these
decisions to modify the received signal sequence {r}, and
then proceeds to decode row n 2, n 3, and so on. In short,
of c, i.e. c 0 , c 1 , , c n 1 , with the only available informa- the HLST codeword matrix C is decoded row by row, or layer
tion being r and H. by layer, from bottom to top.
Under the LST architecture, the decisions on the values of Consider the kth row of an HLST codeword matrix C. Let
these n components are made sequentially according to a pre- { c k } denote the actual transmitted symbol sequence on this
determined order. In this paper, without loss of generality, row, and { e k } denote a distinct possible transmitted symbol
given a fixed , c k is decided in descending order of k. Con- sequence. Conditioned on the channel realization H = {H0,
sider the symbol c n 1 , which is to be detected first. An inter- H1, }, the probability that the likelihood of transmitting
ference suppression operation is employed to extract a { e k } is higher than { c k } is
decision variable about c n 1 from the received signal r. This
decision variable, denoted by z n 1 , contains a low level of
Prob ( c k e k H ) = Q ---------- ( R kk ) 2 c k e k 2
E
interference from other transmitted symbols. The decision on 2N 0
c n 1 is then made based on z n 1 . Making use of the decision (5)
on c n 1 , the receiver can modify the received signal r by E
exp ---------- ( R kk ) 2 c k e k 2 ,
removing the contribution of c n 1 to it. This modifying oper- 4N 0
ation is referred to as interference cancellation. The process of
interference suppression, symbol value decision, and interfer- where the matrix R comes from the QR decomposition of H,

ence cancellation is repeated for the remaining symbols c n 2 , i.e., H = U R , and Q ( x ) = ( 2 ) 1 / 2 exp ( x 2 2 ) dx . In
c n 3 , , c 0 . x
(5), we make use of the Chernoff bound for the Q function,
Suppose that the interference from other transmitted sym- Q ( x ) exp ( x 2 ) . Prob ( c e H ) is the conditional
2 k k

bols is to be completely suppressed using linear operation. Let pairwise error probability between { c k } and { e k }. The
the QR decomposition of H be H = ( U R ) R , where average pairwise error probability can thus be upper-bounded
( U R ) is a unitary matrix and R is an upper triangular by taking the expected value of the right side of (5) over the
matrix. Left-multiply the received signal r by ( U R ) , distribution of ( R kk ) 2 , which is a chi-squared distribution
with 2(m k) degrees of freedom [10]. In a fast fading envi-
y = ( U R ) r = R c + v ' , (2) ronment, the ( R kk ) 2 are i.i.d. for distinct . The average pair-
wise error probability, Prob ( c k e k ) , can be upper bounded
where v ' = ( U R ) v is an m-tuple of i.i.d. AWGN noise
by: E ( m )
Prob ( c e ) 1 + c e 2 ---------
4N
-
Prob ( c k e k ) ( c, e ) 0

E E m (9)
E exp ---------
- ( R kk ) 2 c k e k 2 ( c e 2 ) ( m ) ---------- ( c, e ) ,

( c k, e k ) 4N 0 (6) ( c, e ) 4N 0
E ( m k )
1 + c k e k 2 ---------
= - , where ( c, e ) = { c e } . When the SNR is low, i.e.

4N 0
( c k, e k ) c e 2 E 4N 0 1 for all , using the approximation
( 1 + mx ) 1 ( 1 + x ) m for small mx, (9) can be approximated
where ( c k, e k ) = { c k e k } . In a slow fading environ-
by
ment, ( R kk ) 2 = R kk 2 for all , and Prob ( c k e k ) can be
upper bounded by: E 1 { c e 2 ( m ) }
1 + ---------
Prob ( c e ) 4N 0
-
E ( m k ) ( c, e )
Prob ( c k e k ) E 1 + ---------- R kk
4N 0
2
c k e k 2 (10)

(7) 1

( c, e )
c e 2 ( m )

= 1 + ----------
( m k ) E
= 1 + ---------- c e 2
E ,
. 4N 0
4N 0
where is an arbitrarily small positive number.
B. DLST
C. Comparison
A DLST codeword matrix is decoded diagonal by diagonal.
By comparing (7) and (9), we conclude that the perfor-
To illustrate this, consider the example DLST codeword
mance of HLST codes in slow fading environments is inferior
matrix in Fig. 2. The receiver first generates the decision vari-
to that of DLST codes. For an HLST code, the average pair-
ables for the symbols of the first diagonal of C, namely 0 ,
wise error probability of the bottommost row is inversely pro-
1 , and 2 . Based on the decision variables, this diagonal is
portional to the ( m n + 1 )th power of SNR. In contrast, in
decoded, and the decision is then fed back to remove the con-
DLST, the average pairwise diagonal error probability
tribution of this diagonal to the received signal. The receiver
between two diagonals c and e is inversely proportional to the
then continues to decode the next diagonal, and so on.
( ( c, e ) m )th power of SNR. Therefore, if constituent
Here we consider the probability of a diagonal decision codes of equivalent data rate and complexity are deployed, the
error. Consider the leftmost NW-SE diagonal of a DLST code- error probability of a DLST code in a slow fading environ-
word matrix. On this diagonal, the transmitted symbols are ment can be much lower than that of an HLST code.
c , = 0, 1, , n 1 . The probability that, under the DLST
decoding algorithm, the likelihood of a distinct diagonal e = D. Design Criteria for DLST Codes
{ e 00 e 11 e nn 11 } is higher than that of the transmitted diag- Define the truncated multi-dimensional effective length
onal c = { c 00 c 11 c nn 11 }, conditioned on the channel real- (TMEL) and the truncated multi-dimensional product distance
ization H = {H0, H1, }, can be derived by applying (3) (TMPD) between two distinct diagonals c and e as
noting that the interference term in (3) is zero for this diag-
onal: TMEL m and (11a)
( c, e )
E n1
Prob ( c e H ) = Q ---------- ( R ) 2 c e 2 TMPD c e 2 ( m ) . (11b)
2N 0 = 0 ( c, e )
(8)
At high SNR, the pairwise error probability between c and e is
E n1
exp ---------- ( R ) 2 c e 2 . approximated by Prob ( c e ) ( TMPD ) 1 ( E 4N 0 ) TMEL .
4N 0 = 0 The code design criterion is to maximize the minimum value
of Prob ( c e ) ( TMPD ) 1 ( E 4N 0 ) TMEL over all pairs of
Equation (8) applies in both fast and slow fading environ- distinct diagonals. If the exact operating SNR is not known
ments because the ( R ) 2 are i.i.d. for = 0, 1, , n 1. but can be assumed to be reasonably high, an approximate
The upper bound of the average pairwise error probability design criterion is to maximize the minimum two-tuple
is again obtained by taking the expected value of the right- (TMEL, TMPD) in dictionary order.
hand side of (8). When the SNR is high, At low SNR, the pairwise error probability is approximated
by (10). We define the exponent ( c, e ) c e 2 ( m )
to be the truncated multi-dimensional Euclidean distance
(TMED) between c and e. The code design criterion at low is only quadratic in the number of antennas, making it suitable
SNR is to maximize the minimum TMED between any pair of for systems that have a large number of antennas. Further-
distinct diagonals. more, the existing technology of 1-D codec can be effectively
leveraged.
E. Example
We analyzed the performance of both HLST and DLST
In this example, a (7, 3) RS code is used as a constituent codes. Our result indicates that in a slow fading environment,
code of a 7-D DLST code. The (7, 3) RS encoder maps three DLST codes have superior performance compared to HLST
8-ary input digits into seven 8-ary output digits. Each 8-ary codes. We proposed the design criteria for DLST codes based
digit selects a point (symbol) on the 8-PSK constellation on the following parameters of the constituent code: TMEL,
according to the Gray code mapping, and these seven constel- TMPD, and TMED.
lation points are used to fill the slots of a diagonal of the 7-D
DLST code. The minimum TMEL of this code given n = m = VI. REFERENCES
7 is 15.
[1] E. Telatar, Capacity of Multi-antenna Gaussian Channels, AT&T-
Monte-Carlo simulations are performed to obtain the per- Bell Labs Internal Tech. Memo., June 1995.
formance of this DLST code in slow fading environments
[2] G. J. Foschini and M. J. Gans, On Limits of Wireless Communication
given n = m = 7. Fig. 3 shows the average error probability in a Fading Environment When Using Multiple Antennas, accepted
assuming perfect decision feedback. Fig. 3 shows that, with an for publication in Wireless Personal Communications.
average SNR of 12 dB, it is possible to use this simple DLST
[3] G. J. Foschini, R. A. Valenzuela, Initial Estimation of Communication
code to achieve a data rate of 9 bits/s/Hz with an average diag- Efficiency of Indoor Wireless Channel, Wireless Networks 3, pp. 141-
onal detection error probability lower than 10-4. 154, 1997.
[4] V. Tarokh, N. Seshadri, A.R. Calderbank, Space-Time Codes for High
V. SUMMARY Data Rate Wireless Communication: Performance Criterion and Code
Construction, IEEE Trans. Inform. Theory, vol. 44, no. 2, March
In this paper we considered layered space-time codes. We
1998.
showed that, if the wireless channel is i.i.d. Rayleigh fading,
an (n, m) multiple-antenna system employing an LST code [5] J. Guey, M.P. Fitz, M.R. Bell and W. Kuo, Signal Design for transmit-
ter diversity wireless communication systems over Rayleigh fading
can achieve a throughput min(n, m) times higher than that of a channels, in Proc. IEEE VTC96, pp. 136-40.
single-antenna system for a given overall transmit power
[6] G. J. Foschini, Layered Space-Time Architecture For Wireless Com-
limit.
munication in a Fading Environment When Using Multi-Element
LST codes have two important characteristics. First, the Antennas, Bell Labs Technical Journal, Autumn 1996.
transmitter is not required to have the instantaneous channel [7] S. Lin, D. J. Costello, Error Control Coding: Fundamentals and Appli-
information to employ an LST code. Secondly, LST codeword cations, Prentice Hall, New Jersey, 1983.
matrices are constructed from one-spatial-dimensional con- [8] A.J. Calderbank, Multilevel Codes and Multistage Decoding, IEEE
stituent codewords. The decoding complexity of an LST code Trans. Commun., vol. 37, no. 3, pp.222-29, March 1993.
[9] D. Tse, Capacity Scaling in Multi-Antenna Systems, in preparation.
10-1 [10] R. J. Muirhead, Aspects of Multivariate Statistical Theory, John Wiley
& Sons, New York, 1982.
Average error probability

10-2

10-3

10-4

10-5
8 9 10 11 12 13
SNR E/N0 (dB)
Fig. 3. Simulation results of the error probability of a 7-D
DLST code using the (7, 3) RS as its constituent code in
slow fading environments. Here, n = m = 7.