IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 45, NO.

4, MAY 1999 1121

Combined Array Processing and Space–Time Coding

Vahid Tarokh, Member, IEEE, Ayman Naguib, Member, IEEE, Nambi Seshadri, Senior Member, IEEE,
and A. Robert Calderbank, Fellow, IEEE

Abstract—The information capacity of wireless communication Seshadri and Winters [10] as a special case. These papers
systems may be increased dramatically by employing multiple approached transmit diversity from a signal processing point
transmit and receive antennas. The goal of system design is to of view. Space–time codes [11] combine signal processing at
exploit this capacity in a practical way.
An effective approach to increasing data rate over wireless the receiver with coding techniques appropriate to multiple
channels is to employ space–time coding techniques appropriate transmit antennas, and this approach provides significant gain
to multiple transmit antennas. These space–time codes introduce over [10] and [15]. Specific space–time codes designed for
temporal and spatial correlation into signals transmitted from two to four transmit antennas perform extremely well in slowly
different antennas, so as to provide diversity at the receiver,
and coding gain over an uncoded system. For large number of
varying fading environments (such as indoor transmission) and
transmit antennas and at high bandwidth efficiencies, the receiver come within 2–3 dB of the outage capacity as computed by
may become too complex whenever correlation across transmit Telatar [12] and independently by Foschini and Gans [4]. The
antennas is introduced. bandwidth efficiency of these codes are about three to four
This paper dramatically reduces encoding and decoding com- times that of the current systems. The most important contrib-
plexity by partitioning antennas at the transmitter into small
groups, and using individual space–time codes, called the com- utor to this performance is diversity, which we may think of
ponent codes, to transmit information from each group of anten- as providing some less attenuated replica of the transmitted
nas. At the receiver, an individual space–time code is decoded signal to the receiver. The space–time codes presented in [11]
by a novel linear processing technique that suppresses signals provide the best possible tradeoff between constellation size,
transmitted by other groups of antennas by treating them as data rate, diversity gain, and trellis complexity.
interference. A simple receiver structure is derived that provides
diversity and coding gain over uncoded systems. This combina- When the number of transmit antennas is fixed, decoding
tion of array processing at the receiver and coding techniques for complexity (measured by the number of trellis states in the
multiple transmit antennas can provide reliable and very high decoder) increases exponentially with transmission rate. This
data rate communication over narrowband wireless channels. A can be refined to some extent by designing space–time codes
refinement of this basic structure gives rise to a multilayered
with a multilevel structure and adopting multistage decoding as
space–time architecture that both generalizes and improves upon
the layered space–time architecture proposed by Foschini. described in [11]. For a moderate number of transmit antennas
(three to six), this method provides higher data rates while
Index Terms— Antenna arrays, array processing, diversity,
trellis codes, wireless communication channel.
reducing the decoding complexity. However, there is a penalty
to be paid for simplified decoding. Multistage decoding is
suboptimal, due in part to magnification in the error coefficient,
I. INTRODUCTION and this performance penalty means that alternative solutions
are needed for achieving very high data rates.
P HYSICAL limitations on wireless channels present a fun-
damental technical challenge to reliable communication.
Bandwidth limitations, propagation loss, time variance, noise,
In order to achieve very high data rates on narrowband
wireless channels, we need many antennas at both transmit-
interference, and multipath fading make the wireless channel ter and receiver. Consider a wireless communication system
a narrow pipe that does not easily accommodate the flow of employing transmit and receive antennas, where the
data. Further challenges come from power limitation as well subchannel between each transmit and receive antenna is
as size and speed of devices in wireless portables. quasi-static Rayleigh, flat, and independent of others. If is
Deploying multiple transmit antennas at both the base and fixed, then the capacity increases only logarithmically with
remote stations increases the capacity of wireless channels [4], [12]. On the other hand, if is fixed then intuition
[4], [12], and information theory provides measures of this suggests that there must come a point where adding more
increase. The standard approach to exploit this capacity is transmit antennas will not make much difference and this can
linear processing at the receiver [14] or extensions thereof be seen in the mathematics of outage capacity. For instance, if
[1]. Transmit diversity has been explored by Wittneben [15], , then Foschini and Gans [4] prove that the capacity of
and this proposal includes the delay diversity scheme of the aforementioned system is a random variable of the form
SNR , where is a random variable formed
Manuscript received June 15, 1997; revised February 1, 1998. The material
in this paper was presented in part at the Annual Allerton Conference by summing the squares of independent Gaussian random
on Communication, Control, and Computing, University of Illinois at Ur- variables with mean zero and variance one. This means that by
bana–Champaign, Urbana, IL, October 1997. the strong law of large number in distribution.
The authors are with AT&T Labs-Research, Florham Park, NJ 07932 USA.
Communicated by M. Honig, Associate Editor for Communications. Practically speaking, for , and the capacity
Publisher Item Identifier S 0018-9448(99)03166-1. is the familiar Gaussian capacity SNR per complex
0018–9448/99$10.00  1999 IEEE
1122 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 45, NO. 4, MAY 1999

dimension. Thus in the presence of one receive antenna, little systems with a given diversity gain. This combination of
can be gained in terms of outage capacity by using more than array processing at the receiver and coding techniques for
four transmit antennas. A similar argument shows that if there multiple transmit antennas provides reliable and very high
are two receive antennas, almost all the capacity increase can data rate communication over wireless channels. One obvious
be obtained using transmit antennas. These observations advantage of the group interference suppression method over
also follow from the capacity plots given by Telatar [12]. the architecture of Foschini [3] is that the number of receive
If increases and , then information theory [4], antennas can be less than the number of transmit antennas.
[12] shows that the capacity of the system increases at least We follow Foschini and Gans [4] in assuming that the
linearly as a function of . Thus it makes sense to increase the fading is flat and quasi-static, that is constant over a frame and
number of antennas at both receiver and transmitter in order to varying from one frame to another. Nevertheless, the extension
obtain higher capacities. Many antennas at both transmitter and to frequency-selective channels is straightforward. Similarly,
receiver create a multiple-input multiple-output system where results are presented for the spatio-temporal domain, but the
the number of degrees of freedom is given by the product of extension to frequency domain is likewise straightforward.
the number of transmit and receive antennas. The approach The outline of this paper is as follows. Section II presents a
pursued here is to employ a small fraction of these degrees mathematical model for multiple-antenna communication sys-
of freedom for improving the performance while using the tems and Section III introduces group interference suppression
remaining degrees of freedom to achieve higher data rates. method. This provides a method to maintain the effects of a
This leads to a combination of space–time coding at the group of desired signals while canceling the effects of a group
transmitter and array processing at the receiver which achieves of undesired signals. We then combine space–time coding
high data rates and operates close to the limits given by outage with this new signal processing method to arrive at combined
capacity. array processing and space–time coding. Section IV derives
Foschini considered this problem in [3]. He proposed a the receiver for combined array processing and space–time
multilayered structure that in principle can achieve a tight coding and analyzes its performance. Section V introduces
lower bound on the capacity. If transmit and receive the concept of multilayered space–time coded modulation and
antennas are used, then at the receiver the transmitted signal provides simulation results. Finally, some conclusions are
from transmit antenna is treated as the desired signal, while made in Section VI.
the signals transmitted from other transmit antennas are treated
as interference. Linear processing is then used to suppress II. THE COMMUNICATION MODEL
the interfering signals using receive antennas, providing a
We model a wireless communication system with antennas
diversity gain of one. Once the signal transmitted from antenna
at the base and antennas at the mobile station. Data is
is correctly detected, the signal transmitted from antenna
encoded using a channel code. The encoded data goes through
is treated as the desired signal while the signals transmitted
a serial-to-parallel converter and is divided into streams of
from transmit antennas are treated as interference.
data. Each stream of data is used as the input to a pulse shaper.
Linear processing is then applied to suppress the interfering
The output of each shaper is then modulated. At each time slot
signals using receive antennas. This provides a diversity
the output of modulator is a signal that is transmitted
gain of two. This process is repeated until all the transmitted
using transmit antenna (Tx antenna) for . We
signals are detected. Clearly, the dominant diversity in this
emphasize that the signals are transmitted simultaneously
architecture is one. In this light, long frames of data combined
each from a different transmit antenna and that all these signals
with powerful coding techniques are needed to achieve the
have the same transmission period . The signal at each
lower bound on outage capacity [3]. The diversity level can
receive antenna is a noisy superposition of the transmitted
be improved to albeit at the expense of losing half of
signals corrupted by Rayleigh or Rician fading (see Fig. 1).
the rate [3].
At the receiver, the demodulator computes a decision statis-
The approach described above is purely a signal processing
tic based on the signals received at each receive antenna
one. In this work, we depart from a purely signal process-
. The signal received by antenna at time
ing point of view and approach the problem from an array
is given by
signal processing combined with channel coding perspective.
Antennas at the transmitter are partitioned into small groups,
and individual space–time codes, called the component codes, (1)
are used to transmit information from each group of antennas.
At the receiver an individual space–time code is decoded by a The coefficient is the path gain from transmit antenna to
novel linear array processing technique, called the group inter- receive antenna and is the noise for the channel between
ference suppression method, that suppresses signals transmitted transmit antennas and receive antenna at time . The path
by other groups of antennas by treating them as interference. gains are modeled as samples of independent complex
This method involves only elementary linear algebra, and is Gaussian random variables with mean zero and variance
similar in spirit to techniques developed for multiuser detection per dimension. This is equivalent to the assumption that sig-
[8], [13] and to temporal and spatial techniques developed for nals transmitted from different antennas undergo independent
interference suppression [1], [14]. A simple receiver structure Rayleigh fades. The noise quantities
is derived that provides diversity and coding gain over uncoded are samples of independent complex Gaussian
TAROKH et al: COMBINED ARRAY PROCESSING AND SPACE–TIME CODING 1123

Fig. 1. Block diagram of the transmitter and the receiver.

random variable with mean zero and variance per The task is to recover the transmitted data at the receiver.
dimension. We further assume that are constant during a One approach is to jointly decode for the transmitted code-
frame and vary from one frame to another (quasi-static fading). word, but the difficulty here is decoding complexity. Indeed,
For any vector , let denote the transpose of . We can if we require a diversity of , where , then
now write (1) in the vector form given by the complexity of the trellis of is at least states
[11] and the complexity of the product code is at least
(2) states. This means that if is very large, the scheme may be
where too complex to implement.
The approach we advocate is group interference suppres-
(3) sion. The idea is to decode each code separately while
(4) suppressing signals from other component codes. This ap-
(5) proach has a much lower complexity but achieves a lower
diversity order than the full diversity order , which is the
and product of the numbers of transmit and receive antennas.
To describe the group interference suppression method,
without any loss of generality, we take and look to
decode . There are interfering signals. We assume
(6) that there are receive antennas and that the
.. .. .. .. ..
. . . . . receiver knows the matrix (channel state information). The
matrix

III. COMBINED ARRAY PROCESSING

AND SPACE–TIME CODING

A space–time product encoder accepts a block of input (7)

.. .. .. .. ..
bits in each time slot and these bits are divided into strings . . . . .
of lengths with
Consider the communication model given in Section II, with
antennas at the base station partitioned into groups has rank less than or equal to the number of its columns.
, respectively, comprising anten- Thus . Recall that the null space
nas with . Each block [6] of this matrix is the set of all row vectors such that
is then encoded by a space–time encoder . The output of . Furthermore
goes through a serial-to-parallel converter and provides
sequences of constellation symbols for which are
simultaneously transmitted from the antennas of the group .
This gives a total of sequences of constellation symbols that Since , it follows that
are transmitted simultaneously from antennas . We . Hence we can compute a (not necessarily unique)
view a space–time product encoder as a set of space–time set of orthonormal vectors
encoders called the component codes operating in parallel on
the same wireless communication channel, with each encoder
using transmit and receive antennas for . It
will be denoted by . in and, in fact, there are fast algorithms for computing
The partition of transmit antennas into small groups, and the . We let denote the
use of component space–time codes to transmit information matrix whose th row is . Clearly,
from each group of antennas fits naturally in the framework , where is the Hermitian of
of product codes [2]. Product codes are applied for interference and is the identity
suppression in [1], where Calderbank, Pottie, and Seshadri use matrix. We multiply both sides of (2) by to arrive at
the product code construction of first-order Reed–Muller codes
for a combination of coding gain and interference suppression. (8)
1124 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 45, NO. 4, MAY 1999

Let This means that following the execution of group interference

suppression, the decoder of treats the vector as the
received word at time and the matrix as the matrix of
path gains.
(9)
.. .. .. .. .. If the code is a space–time trellis code, then the branch
. . . . . metric for a branch labeled with at time is
.
Since is the all-zero matrix, (8) can be We now analyze the error performance of the system. The
written as next result formalizes the fact that the group interference
suppression method does indeed suppress the interference from
(10) the other component codes at no expense to the
where . Setting performance of the component code .

(11) Theorem 1: Consider a multiple-antenna wireless commu-

nication system with transmit and
(12)
receive antennas. Let denote a partition of
(13) transmit antennas into groups antennas, respec-
we arrive at the equation tively. Let denote a product space–time
encoder. At each time , a block of input bits arrive at
(14) the encoder of and these bits are divided into strings
with . Every block
This is an equation where all the signal streams out of
is then encoded by the encoder . For
antennas are suppressed.
the output of encoder is sent using the antennas
in group and all these transmissions are simultaneous. Let
IV. THE RECEIVER
denote the average transmit power out of
In this section, we compute the structure of the decoder antenna . Let denote the probability of error for using
for the code given that group interference suppression is group interference suppression method.
performed to suppress all the signal streams out of antennas Consider another wireless multiple antenna communication
. system with transmit and receive antennas. At
To this end, suppose that is given. The receiver each time , a block of bits arrive at the encoder of . The
computes a set of orthonormal vectors output of the encoder is sent using the transmit antennas
and all these transmissions are simultaneous. Suppose that the
average transmit power out of antenna is .
and the matrix as described in the previous section. Let denote the probability of error for this system. Then
Let and denote the th and th elements of .
. By definition and where and Proof: Suppose that is given. The receiver com-
are, respectively, the th and th columns of . The putes a set of orthonormal vectors
random variables and have zero means given .
Moreover,
and the matrix as described in the previous section.
Following group interference suppression the decoder of
(15) uses the decision statistics and the coefficients to decode
where is the Kronecker delta function given by . By (14)
if and if . We conclude that the (17)
elements of the matrix are inde-
pendent complex Gaussian random variables of variance The elements of the matrix were previously
per real dimension. Similarly, the components of the noise proved to be independent complex Gaussian random variables
vector are independent Gaussian random of variance per real dimension. Similarly, the components
variables of variance per real dimension. of the noise vector are independent Gaussian
Assuming that all the codewords of are equiprobable random variables of variance per real dimension.
and given that group interference suppression is performed, Consider the second communication scenario, where is
the proposed receiver for decides in favor of the codeword used with transmit and receive antennas and
the same transmit powers . The received word
for this scenario at time is of the form
if it minimizes the decision metric
(18)
(16) where was an random matrix having
independent and identically distributed (i.i.d.) elements with
TAROKH et al: COMBINED ARRAY PROCESSING AND SPACE–TIME CODING 1125

the same distribution as the elements of the matrix . The

size and distribution of the noise vectors are the same as .
Thus the probability of error is the same as the probability
of error for using group interference suppression.

V. MULTILAYERED SPACE–TIME CODED MODULATION

Theorem 1 reduces code design for a multiple-antenna
communication system with transmit and receive antennas
to that of designing codes for communication systems with
transmit and receive antennas Fig. 2. The 4-PSK constellation.
where and . Using this insight,
we may design a multilayered space–time coded-modulation
scheme. The idea behind such a system is multistage detection
and cancellation.
Suppose that is decoded correctly using combined ar-
ray processing and space–time coding. By Theorem 1, the
space–time code affords a diversity gain of
. After decoding , we may subtract the contribution
of these codewords to signals received at different antennas.
This gives a communication system with transmit and
receive antennas. We next use combined array processing
and space–time coding to decode . The space–time code
affords a diversity gain of . Proceeding in
this manner, we observe that by subtracting the contribution of
previously decoded code streams to the received
signals at different antennas, the space–time code affords
a diversity gain of .
We can choose space–time codes to provide
these diversity gains and such that the sequence

is an increasing sequence. Assuming there was no decoding

error in steps , then at decoding step , the
probability of error for the component code is by Theorem
1 equal to the probability of error for when employed in
a communication system using transmit and
receive antennas.
Since the diversity in each decoding stage is more than
that of the previous decoding stage , the transmit power
out of each antenna at level can be substantially less Fig. 3. 4-PSK space–time code, 32 states 2 bits/s/Hz.
than that of the previous layer. Thus the transmitter should
divide the available transmit power among different antennas Example 1: Here, four transmit and four receive antennas
in an unequal manner. Power allocation for this scenario are used. The transmission rate is 4 bits/s/Hz. Let denote
is straightforward. In fact, powers at different levels could the 32-state 4-PSK (see Fig. 2) space–time trellis code [11]
be allocated based on the diversity gains. In this way, the given in Fig. 3. We use the product code where
allocated powers may decrease geometrically in terms of the for transmission of 4 bits/s/Hz. At each time
diversity gains. Other approaches are also possible. slot, upon the arrival of the 4 bits of the input data, the first
At this point, it is advantageous to revisit the multilayered two bits are used as the input to the encoder of and the
architecture of Foschini [3]. The scheme of [3] is a case of encoded symbols are transmitted by antennas one and two.
multilayered space–time architecture described above where The second two bits are used by as the input to the encoder of
, for all . Thus our scheme generalizes that and the encoded signals are transmitted by antennas and
given by Foschini [3]. . We assume that the average powers radiated from antennas
We illustrate the construction of multilayered space–time and are equal but each is twice as much as the average
coded modulation schemes by some examples. power radiated from antennas and . At the decoder, group
1126 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 45, NO. 4, MAY 1999

Fig. 4. The performance of the scheme of Example 1.

Fig. 5. Outage capacity, four transmit, four receive antennas.

interference suppression is used to suppress and decode corresponds to the transmission of 4 bits. Thus the horizontal
. Upon decoding , the contributions of the codewords axis denotes the receive signal-to-noise ratio per 4 bits.
transmitted from antennas one and two are subtracted from For comparison, we provide the graph of the outage capacity
the received signals. Finally, is decoded. In Fig. 4, we versus the signal-to-noise ratio (Fig. 5) for four transmit and
provide simulation results to demonstrate the performance of four receive antennas as computed by Foschini and Gans [4].
this multilayered space–time coded architecture. Each frame We observe that for a frame error probability of , we are
consists of 130 transmissions from each transmit antenna. about 6 dB away from the capacity.
It is assumed that the channel matrix is perfectly known at
the receiver. The horizontal axis shows the receive signal- Example 2: Here, eight transmit and eight receive antennas
to-noise ratio per transmission time. Each transmission time are used. The transmission rate is 8 bits/s/Hz. Let denote the
TAROKH et al: COMBINED ARRAY PROCESSING AND SPACE–TIME CODING 1127

Fig. 6. The performance of the scheme of Example 2.

Fig. 7. Outage capacity, eight transmit, eight receive antennas.

code given in Fig. 3. We use the product code antennas and is , the average power radiated from
where for transmission of 8 antennas and is , and the average power radiated
bits/s/Hz. At each time instance, upon the arrival of the 8 bits from antennas and is . Thus the total signal-to-noise
of the input data, the first, second, third, and fourth blocks of ratio at each receive antenna is . At the decoder,
length two of the input bits are respectively used as the input group interference suppression is used to decode . Upon
to encoders of , and . The output of encoders decoding , the contributions of the codewords transmitted
of are, respectively, transmitted by antennas from antennas one and two are subtracted from the received
and . We assume that the average power radiated signals. Using this, is decoded next and so forth. In Fig. 6,
from antennas and is , the average power radiated from we provide simulation results to demonstrate the performance
1128 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 45, NO. 4, MAY 1999

of this multilayered space–time coded architecture. Each frame [2] G. D. Forney, Concatenated Codes. Boston, MA: MIT Press, 1966.
consists of 130 transmissions from each transmit antenna. [3] G. J. Foschini, “Layered space-time architecture for wireless commu-
nication in a fading environment when using multielement antennas,”
It is assumed that the channel matrix is perfectly known at Bell Labs Tech. J., vol. 1, no. 2, Autumn 1996.
the receiver. The horizontal axis shows the receive signal- [4] G. J. Foschini, Jr. and M. J. Gans, “On limits of wireless communication
to-noise ratio per transmission time. Each transmission time in a fading environment when using multiple antennas,” Wireless
Personal Commun., vol. 6, no. 3, pp. 311–335, Mar. 1998.
corresponds to the transmission of 8 bits. Thus the horizontal [5] E. A. Gelblum and N. Seshadri, “High rate coded modulation schemes
axis denotes the receive signal-to-noise ratio per 8 bits. for 16 kbps speech in wireless systems,” in Proc. IEEE VTC’97, pp.
For comparison, we provide in Fig. 7 the graph of the outage 349–353.
[6] R. A. Horn and C. R. Johnson, Matrix Analysis. New York: Cambridge
capacity versus the signal-to-noise ratio for eight transmit and
Univ. Press, 1988.
eight receive antennas as computed by Foschini and Gans [4]. [7] W. C. Jakes, Microwave Mobile Communications. Piscataway, NJ:
We observe that for a frame error probability of , we are IEEE Press, 1993.
about 9 dB away from the capacity. [8] R. Lupas and S. Verdú, “Linear multiuser detectors for synchronous
code-division multiple-access channels,” IEEE Trans. Inform. Theory,
vol. 35, pp. 123–136, Jan. 1989.
VI. CONCLUSION [9] N. Sollenberger and S. Kustaria, “Evolution of TDMA (IS-54/IS-136)
to foster further growth of PCS,” in Proc. IEEE ICUPC’96.
We presented a new method of signal processing, namely [10] N. Seshadri and J. H. Winters, “Two signaling schemes for improving
group interference suppression method. This method was the error performance of frequency-division-duplex (FDD) transmission
systems using transmitter antenna diversity,” Int. J. Wireless Inform.
combined with space–time coding giving rise to combined Networks, vol. 1, no. 1, 1994.
array processing and space–time coding. Very high rates [11] V. Tarokh, N. Seshadri, and A. R. Calderbank, “Space-time codes for
at reasonable complexity and signal-to-noise ratios can be high data rate wireless communication: Performance analysis and code
achieved using this method. construction,” IEEE Trans. Inform. Theory, vol. 44, pp. 744–765, Mar.
1998.
[12] E. Telatar, “Capacity of multiantenna Gaussian channels,” AT&T-Bell
ACKNOWLEDGMENT Lab. Internal Tech. Memo., June 1995.
[13] M. K. Varanasi, “Group detection for synchronous Gaussian code-
The authors wish to thank the anonymous reviewers for division multiple-access channels,” IEEE Trans. Inform. Theory, vol.
thoughtful commentary that improved the presentation of the 41, pp. 1083–1096, July 1995.
[14] J. Winters, J. Salz, and R. D. Gitlin, “The impact of antenna diversity
results in this paper. on the capacity of wireless communication systems,” IEEE Trans.
Commun., vol. 42. pp. 1740–1751, Feb./Mar./Apr. 1994.
REFERENCES [15] A. Wittneben, “Base station modulation diversity for digital SIMUL-
CAST,” in Proc. IEEE’ VTC, May 1993, pp. 505–511.
[1] A. R. Calderbank, G. Pottie, and N. Seshadri, “Co-channel interference [16] , “A new bandwidth efficient transmit antenna modulation diver-
suppression through time/space diversity,” IEEE Trans. Inform. Theory, sity scheme for linear digital modulation,” in Proc. IEEE’ICC, 1993,
submitted for publication. pp. 1630–1634.