530

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 52, NO. 3, MAY 2003

T-BLAST for Wireless Communications:

First Experimental Results
Mathini Sellathurai, Member, IEEE, and Simon Haykin, Fellow, IEEE

AbstractIn earlier papers [1][5], we described a novel

multitransmit, multireceive (MTMR) antenna system for wireless
communications. This new system, turbo Bell-Labs layered
spacetime (T-BLAST) architecture, combines the benefits of
layered spacetime coding concepts and turbo principles in the
multitransmit, multireceive antenna system design. In particular,
the random layered spacetime codes designed by using a set of
block convolutional codes and random spacetime interleavers
and the spacetime turbo-like decoding operation allow T-BLAST
to realize the benefits of MTMR systems in a computationally
feasible manner. The goal of this paper is to present experimental
results of T-BLAST based on real-life data collected using the
Bell-Labs experimental multiple antenna system with eight
transmit and five and six receive antennas. The experimental
results show the practical virtues of T-BLAST.
Index TermsBell-Labs layered space-time (BLAST) architecture, interference cancellation, spacetime codes, turbo processing.

I. INTRODUCTION

NE of the major concerns of wireless communications

research is to provide techniques that use a scarce
resource, namely, the available transmission spectrum, efficiently. A groundbreaking technique for wireless research,
offering tremendous potential to increase systems capacity and
spectrum efficiency, is the multitransmit, multireceive (MTMR)
antenna schemes popularized as Bell-Labs layered space-time
(BLAST) architecture. The first version of BLAST, originated
by Foschini [6], used a novel diagonal layered spacetime
architecture, hence the terminology D-BLAST. However, from
a practical perspective, D-BLAST is inefficient for short packet
transmissions due to its boundary spacetime wastage. The
challenge is to design an MTMR wireless communications
system that is capable of achieving a high spectral efficiency
and yet maintains a manageable system complexity.
Along the series of BLAST innovations, vertical BLAST
(V-BLAST), described in [7], was the first practical system
demonstrated in real time. In V-BLAST, every antenna
transmits its own independent substream of data using a
simple vector- and linear-decoding structure. By using this
Manuscript received May 23, 2002; revised August 28, 2002, November 22,
2002, and January 21, 2003. The authors are grateful to the Natural Sciences and
Engineering Research Council of Canada and Dr. R. A. Valenzuela, Department
of Wireless Communications Research, Lucent Technologies, Bell Laboratories, Holmdel, NJ, USA, for financial support.
M. Sellathurai was with McMaster University, Hamilton, ON L8S 4K1,
Canada. She is currently with the Communications Research Center (CRC) of
Canada, Ottawa, ON K2H 8S2, Canada.
S. Haykin is with McMaster University, Hamilton, ON L8S 4K1, Canada
(e-mail: haykin@soma.crl.mcmaster.ca).
Digital Object Identifier 10.1109/TVT.2003.810986

architecture, it has been demonstrated that up to 50% of the

channel capacity can be achieved even with no channel coding.
However, V-BLAST suffers from two limitations.
It demands more receive than transmit antennas. The
ability to work with fewer receivers than transmitters is
necessary in most cellular communications systems since
the base station is typically designed with more antennas
than mobile hand-communication devices.
No built-in spacetime codes are used to overcome deep
fades from any of the transmit antennas.
The next major innovation in this series is turbo BLAST,
or T-BLAST for short, which uses a relatively simple layered
spacetime encoder and an iterative joint detectordecoder that
is turbo-like in operation. Due to the structure of the spacetime
encoder and the turbo-like operation of the receiver, T-BLAST
offers the following combination of features:
a spectral efficiency attained in the course of two to four
iterations of the receiver, which is significantly superior to
that attainable with V-BLAST;
a built-in capability of accommodating the multiple-antenna configuration, including the case of fewer receive
than transmit antennas, which is achieved with manageable computational complexity.
These novel features of T-BLAST are confirmed by the
real-life experiments presented here. The data set was acquired
with the cooperation of the Department of Wireless Communications Research, Lucent Technologies, Bell Laboratories
at Holmdel, New Jersey, USA, and was collected using the
indoor narrow-band BLAST test bed located on the second
floor of the Crawford-Hills Laboratory. Basic to the operation
of T-BLAST is the idea of iterative processing, which has
also found applications in channel equalization and multiuser
detection [10][13].
The rest of the paper describes the experimental setup and
the real-life experimental results. For a detailed exposition
of T-BLAST architecture, readers are referred to references
[1][5].

II. EVALUATION TESTS

This section discusses the performance of T-BLAST using
the Bell-Labs test bed, described in [14], with eight transmit and
five and six receive antennas. It begins by first describing the indoor narrow-band test bed and receiver digital signal-processing
(DSP) operations. Next, the bit- and frame-error performances
of T-BLAST configuration are presented.

0018-9545/03$17.00 2003 IEEE

SELLATHURAI AND HAYKIN: T-BLAST FOR WIRELESS COMMUNICATIONS

Fig. 1. Bit-error performance of T- BLAST n = 8 and n = 6 with

convolutional code of rate R = 1=2 constraint length 3 and of 4-QAM
modulated SNR = 9 dB.

A. Indoor Test Bed

The antenna arrays consist of wire dipoles mounted in various (horizontal- and vertical-polarization) arrangements, with
about half-wave length separation between adjacent elements.
The system operates at a carrier frequency of 1.95 GHz with
30-kHz signal bandwidth. Detailed descriptions of the indoor
BLAST test-bed hardware and measurement approach are given
in [7] and [14].
The measurements were taken in the Bell-Labs building in
Crawford Hill, NJ, in a laboratory of 12 24 , located on the
second floor of the building. For all measurements, the transmitter was located in the southwest corner of the lab and was
facing north. The receiver was located in the northeast corner
of the lab and was facing south. The measurements were taken
while people were walking around the area and 100 continuously measured channel conditions were considered in this
paper. At each channel condition, a packet of 132 4-QAM modulated symbols per antenna is transmitted at a rate of 25 k-symbols/s. Among those 132 symbols, the first 32 are dedicated to
synchronization and training.
1) Synchronization: The first 16 symbols are used for frame
and symbol-timing recovery.
2) Training sequence: The next 16 symbols (symbols 17 to
32) are used for matrix channel-response estimation. For
each substream, mutually orthogonal and equal powertraining sequences are generated by using 16-dimensional
Hadamard sequences.
3) Information symbols: The last 100 symbols are used for
information transmission. In the transmit end, each substream of 100 information bits is independently encoded
, constraint
using a convolutional code with rate
length 3, and then interleaved using spacetime interleavers. The space interleavers are designed according to
the D-BLAST architecture but with no edge wastage [5].
Eight independent time interleavers are chosen randomly

531

Fig. 2. Number of errors versus packets in T-BLAST n = 8 and n = 6

with convolutional code of rate R = 1=2 and constraint length 3 and 4-QAM
modulation SNR = 9 dB.

and no attempt is made to optimize their design. The dimensions of the time interleavers are 200 b.
B. Receiver
At the receiving end, the signal-detection process involves the
following sequence of operations:
1) Frame initialization: The receiver waits until it finds a sufficiently strong signal to indicate the start of data transmission.
2) Symbol synchronization: The sampled received signal
is cross-correlated with a predefined synchronization
sequence and the condition that results in highest cross
correlation is used to establish symbol synchronization.
The received signal is oversampled with four samples
per symbol period. Binary Barker sequences are used
for synchronization due to their good autocorrelation
properties.
3) Hardware-induced intersymbol interference (ISI) mitigation: The spectrum shaped with an analog low-pass filter
is usually distorted by the radio-frequency front-end of
the transmitter during the transmission process. The ISI
caused by the spectrum shaping and its distortions is mitigated by a precalculated fixed-coefficient FIR filter.
4) Channel estimation: The matrix channel response is estimated by using a mutually orthogonal 16-dimensional
Hadamard sequence transmitted between the parallel antennas for training purpose.
5) Information recovery: The iterative detection and decoding receiver described in Section II is used to recover
the transmitted signals. In this scheme, we separate the
receiver into two stages: soft interference-cancellation
detector and a set of parallel single-inputsingle-output
(SISO) channel decoders. Extrinsic information learned
from one stage is applied to the other stage iteratively
until the receiver converges. These two stages of pro-

532

Fig. 3.

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 52, NO. 3, MAY 2003

Signalspace diagram at the receiver output in T-BLAST for packet 1

n = 8 and n = 6 with convolutional code of rate R = 1=2 and constraint

length 3 and 4-QAM modulation SNR = 9 dB.

cessing are separated by the corresponding spacetime

interleavers and de-interleavers.
C. Experiments With (8,6)-BLAST
This section discusses the performance of T-BLAST using
the Bell-Labs test bed with eight transmit and six receive antennas. In particular, we describe the bit- and frame-error rate
performance obtained using real-life received signals. The estimated signal-to-noise ratio (SNR) of the real-life system is
roughly 9 dB.
Bit Error Rate (BER): Fig. 1 displays the BER performance
of T-BLAST at each iteration of the receiver. Here the instantaneous (broken trace) and average (solid trace) BER are demonstrated at each iteration.
Subplots 1 and 2, respectively, show the BER performance
before decoding the detected coded signals (200 b/packet) and
the decoded information signals (98 information bits + 2 flushed
b/packet) at the first iteration. As expected, the second scheme
performs better due to the coding gain. However, the performance at the first iteration is poor since each substream (transmit
antenna) sees all the other (seven) parallel substreams as interference. The subsequent subplots 36 show the bit-error performance at iterations 25. These figures clearly illustrate a significant performance improvement in the course of a few iterations.
The first three iterations are sufficient to achieve a significant
performance gain. Evidently, the performance gain due to the
subsequent iterations (4 and 5) is minimal.
Frame Error: In Fig. 2, we show the corresponding bit-error
traces in each packet and the packets in error. As in Fig. 1, subplot 1 shows the receiver bit-error trace (out of 200 b/packet) at
each packet with no channel encoding. The subsequent subplots
26 show the bit-error traces (out of 98 information b/packet) at
iterations 15. The iterative action of the receiver significantly
reduces the packet error rate.

MSE at the receiver output of T-BLAST for packet 1 n = 8 and

n = 6 with convolutional code of rate R = 1=2 and constraint length 3 and
4-QAM modulation SNR = 9 dB.
Fig. 4.

With no channel encoding, 100150 bits were detected incorrectly in each packet; indeed a 100% packet-error rate was occasionally observed. With channel encoding, the number of errors
in each packet is reduced to below 15 and the packet-error rate is
reduced by over 25%. In fact, the packet-error rate is further reduced to 17% at iteration 2. Thereafter, only about 4 packets are
corrupted and among the corrupted packets, only 12 bits are in
error per packet. Even though the frame-error performance has
converged at iteration 3, the appearance and disappearance of
errors are observed between packets, from iteration 3 to 6. However, a packet-error rate of 45% is maintained. For example, at
iteration 3, packets 7, 35, 50, and 75 are in error whereas at iteration 5, packets 15, 35, 50, and 75 are in error.
SignalSpace Diagram: Another measure of convergence of
the iterative receiver is the mean square error (MSE) between the
detected signals and the transmitted constellation points in the
signalspace diagram. Here we consider two examples. First,
we show a packet of data that has converged to zero bit error in
three iterations. The second example shows the appearance and
disappearance of errors from one iteration to the next.
Example 1: Perfect Convergence: In this example, we illustrate a perfect convergence behavior. Fig. 3 displays the softdecoded signals in the signalspace diagram for packet 1. Here
the and axes, respectively, represent the real and imaginary
parts of the 4-QAM signal. Subplot 1 shows the positions of the
200 8 coded bits, whereas subplots 29 show the positions of
the 98 8 decoded bits at iterations 18 of the T-BLAST receiver.
The corresponding MSEs of the detected soft bits are shown
in Fig. 4. A bit error occurs if the MSE exceeds 1. MSE between 0 and 1 means that a bit is classified appropriately but
has a residual error that can propagate through the soft interference-cancellation receiver. Subplot 1 shows the MSE before
decoding. Almost all the bits have residual error and about 30%
are detected incorrectly. The figure illustrates how the test error

SELLATHURAI AND HAYKIN: T-BLAST FOR WIRELESS COMMUNICATIONS

Fig. 5.

Signalspace diagram at the receiver output for packet 2 in T-BLAST

n = 8 and n = 6 with convolutional code of rate R = 1=2 and constraint

length 3 and 4-QAM modulation SNR = 9 dB.

533

MSE at the receiver output of T-BLAST for packet 2 n = 8 and

n = 6 with convolutional code of rate R = 1=2 and constraint length 3 and
4-QAM modulation SNR = 9 dB.
Fig. 6.

decreases as new iterations are added to the receiver. In particular, all bits have been detected correctly in three iterations.
Example 2: Appearance and Disappearance of Errors: In
this example, we illustrate the appearance and disappearance of
errors from one iteration to the next. Fig. 5 displays the soft decoded signals in the signalspace diagram for packet 2. The corresponding MSEs of the detected soft bits are shown in Fig. 6.
From the figures, we observe the following:
the residual errors do not converge within eight iterations;
all the bit errors are corrected at iteration 3. However, a
single bit error appears in the fourth, sixth, and eighth iterations and disappears during the fifth and seventh iterations, which demonstrates the appearance and disappearance of errors in the course of convergence.
In both examples, little benefit results from increasing the
number of iterations beyond three; thus, it is reasonable to
accept this number as the practical number of iterations in
terms of both error reduction and receiver complexity.
D. Experiments With (8,5)-BLAST
The above tests were repeated with eight transmit and five
receive antennas. Even though the performance decreased from
the previous experiment, good convergence behavior is still observed. Figs. 7 and 8 display the soft-decoded signals in the
signalspace diagram and the corresponding MSEs of the detected soft bits, respectively. In contrast to the previous experiment, five iterations were needed to achieve a perfect convergence. Moreover, this point is also borne out by the bit-error rate
performance, Fig. 9, over 100 packets versus the number of itand 6 plotted using solid traces. Note
erations for both
that the broken traces show the corresponding performances assuming no channel coding.

Fig. 7. Signalspace diagram at the receiver output for packet 1 in T-BLAST

n = 8 and n = 5 with convolutional code of rate R = 1=2 and constraint
length 3 and 4-QAM modulation SNR = 9 dB.

E. Simulated Results: BER versus SNR

This section compares the performance of 4-QAM modulated T-BLAST and horizontal-coded V-BLAST using indoor
real-channel measurements. We refer to horizontal-coded
V-BLAST when each of the substreams is provided with an
amount of channel coding equal to that used in T-BLAST.
Note that V-BLAST does not use any spacetime or iterative
decoding. We synthesize the received signal using the measured channel characteristics and evaluate the performance of
T-BLAST over a wide range of SNRs using various BLAST

534

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 52, NO. 3, MAY 2003

Fig. 8. MSE at the receiver output of T-BLAST for packet 1 n = 8 and

n = 5 with convolutional code of rate R = 1=2 and constraint length 3 and
of 4-QAM modulation SNR = 9 dB.

combinations. In all the experiments presented here, it is

assumed that the exact channel matrix is known. The antenna
and
, 6, 7, and 8.
configurations considered are
Fig. 10 displays the BER performance of T-BLAST (solid
trace) and coded V-BLAST (broken trace) for measured
real-life channels (using thin curves) and simulated i.i.d
complex Gaussian matrix channels (using thick curves). The
antenna configurations of 8 transmit and 58 receive antennas
and the T-BLAST gives us the best performance within the
first 10 iterations. The figure reveals a major limitation of
the V-BLAST system: the inability to work with fewer receive antennas than transmit antennas. In terms of T-BLAST
performance, the following observations can be made from
Fig. 10. The bit-error performance of T-BLAST improves with
increasing number of receivers (due to the added diversity),
with T-BLAST outperforming V-BLAST in all four cases.
Moreover, the performances of indoor measured channels are
very close to that of i.i.d. Gaussian matrix channels. In indoor
rich-scattering environments, it is typical to get well-conditioned channels [7] and [8].
III. DISCUSSION
Previously, the BLAST test bed has been used to demonstrate uncoded V-BLAST architecture for 8 transmit and 12
receive antennas [7]. However, a major limitation of V-BLAST
is its inability to work with fewer receive antennas than
transmit antennas. The ability to work with fewer receivers
than transmitters is necessary in most cellular systems since
the base station is typically designed with more antennas than
mobile transceivers. In contrast, T-BLAST accommodates any
multiple-antenna configuration, including the case of fewer
receive antennas than transmit antennas with manageable computational complexity. The experimental results using real-life
indoor wireless communication data with eight transmit and

Fig. 9. BER performances for n = 8 and n = 6, 5 with convolutional code

of rate R = 1=2 and constraint length 3, and 4-QAM modulation SNR = 9 dB.

Fig. 10. Bit-error performance for n = 8 and n = 5, 6, 7 and 8 using

convolutional code with rate R = 1=2 and constraint length 3 and 4-QAM
modulation using measured indoor channels (thin curves) and i.i.d. Guassian
channels (thick curves).

5 and 6 receive antennas presented here confirm the practical

virtues of T-BLAST.
ACKNOWLEDGMENT
The experiments were carried out while the first author of this
paper was at Bell Labs, Lucent Technologies, Crawford Hill, NJ
as a visiting researcher in 2000. The cooperation and help provided by Dr. D. Samardzija, Lucent Technologies, in the experiments are gratefully acknowledged.
REFERENCES
[1] M. Sellathurai and S. Haykin, Turbo-BLAST for high speed wireless
communications, in Proc. Wireless Commun. Network Conf., vol. 1,
Chicago, IL, Sept. 2000, pp. 315320.

SELLATHURAI AND HAYKIN: T-BLAST FOR WIRELESS COMMUNICATIONS

[2]
[3]

[4]
[5]
[6]
[7]

[8]
[9]
[10]
[11]
[12]
[13]
[14]

, A nonlinear iterative beamforming technique for wireless communications, in 33rd ASILOMAR Conf. Signals, Systems, and Computers, vol. 2, Pacific Grove, CA, Oct. 1999, pp. 957961.
, Turbo-BLAST: A novel technique for multi-transmit multi-receive wireless communications, in Multiaccess, Mobility, and
Teletraffic for Wireless Communications: Kluwer, 2000, vol. 5, pp.
1324.
, A simplified diagonal BLAST architecture with iterative parallelinterference cancellation, in Proc. IEEE Int. Conf. Commun. (ICC), vol.
10, Helsinki, Finland, June 2001, pp. 30673071.
, Turbo-BLAST for wireless communications: Theory and experiments, IEEE Trans. Signal Proc., vol. 50, pp. 25382546, Oct. 2002.
J. G. Foschini, Layered space-time architecture for wireless communication in a fading environment when using multi element antennas,
Bell Labs Tech. J., 1996.
G. D. Golden, J. G. Foschini, R. A. Valenzuela, and P. W. Wolniansky,
Detection algorithm and initial laboratory results using V-BLAST
space-time communication architecture, Electron. Lett., vol. 35, pp.
1415, Jan. 1999.
D. Gesbert, H. Bolcskei, D. Gore, and A. Paulraj, MIMO wireless channels: Capacity and performance prediction, Proc. IEEE Globecom02,
vol. 1, pp. 10831087, Nov. 2000.
C. Berrou, A. Glavieux, and P. Thitmajshima, Near Shannon limit
error-correcting coding and decoding: Turbo codes, in Proc. Int. Conf.
Commun., Geneva, Switzerland, May 1993, pp. 10641070.
J. Hagenauer, The turbo principle: Tutorial introduction and state of the
art, in Int. Symp. Turbo Codes, Best, France, Sept. 1997.
X. Wang and V. Poor, Iterative (Turbo) soft interference cancellation
and decoding for coded CDMA, IEEE Trans. Commun., vol. 47, pp.
10461061, July 1999.
K. M. Chugg, A. Anaastasopoulos, and X. Chen, Iterative Detection:
Adaptivity, Complexity Reduction and Application: Kluwer, 2001.
M. Tuchler, R. Koetter, and A. C. Singer, Turbo equalization: Principles
and new results, IEEE Trans. Commun., vol. 50, pp. 754767, May
2002.
H. Zheng and D. Samardzija, Performance evaluation of indoor wireless video system using BLAST test-bed, in The 35th Annu. Conf. Inform. Sciences Syst., Baltimore, MD, Mar. 2001.

535

Mathini Sellathurai (S95M01) received the

Ph.D. degree in electrical engineering from McMaster University, Hamilton, ON, Canada, in 2001
and the Technical Licentiate degree in electrical
engineering from the Royal Institute of Technology,
Stockholm, Sweden, in 1997.
She is currently with Communications Research
Centre of Canada, Ottawa, ON, Canada, as a Senior
Research Scientist. Her research interests include
the applications of adaptive signal processing to
spacetime wireless communications, satellite
communications, and broadband multimedia systems.
Dr. Sellathurai was awarded the doctoral price in engineering and computer
sciences from the Natural Sciences and Engineering Research Council of
Canada for her Ph.D. dissertation.

Simon Haykin (F86) received the B.Sc. degree

(First Class Honors) in 1953, the Ph.D. degree in
1956, and the D.Sc. degree in 1967, all in electrical
engineering from the University of Birmingham,
Birmingham, U.K.
His research interests include signal processing,
neural networks and adaptive filters and their
applications in radar and communication systems,
adaptive hearing systems, and computational
neuroanatomy.
In 1980, he was elected Fellow of the Royal Society of Canada. He was
awarded the McNaughton Gold Medal, IEEE (Region 7), in 1986 and was the
recipient of the Booker Gold Medal from URSI. He is the founding Director of
the Communications Research Laboratory at McMaster University, Hamilton,
ON, Canada. In 1996, he was awarded the title "University Professor," the first
faculty member in the Faculty of Engineering at McMaster University to be
awarded this prestigious title.