A Comparative Study of MIMO Detection

Algorithms for Wideband Spatial Multiplexing

Systems
Jingming Wang

Babak Daneshrad

Electrical Engineering Department

University of California, Los Angeles, CA 90095
Email: {wangj, babak}@ee.ucla.edu
Telephone: (310) 825-7792

Abstract The implementation of wideband MIMO system

posts a major challenge to hardware designers due to the huge
processing power required for MIMO detection. To achieve this
goal with complete VLSI solution, channel coding and MIMO
detection are preferrably separated so that each of them can
be fit into a single chip. In this paper, a comparative study is
presented regarding various uncoded adaptive and non-adaptive
MIMO detection algorithms. Intended to serve as a reference
for system designers, this comparison is performed from several
different perspectives including theoretical formulation, simulated BER/PER performance, and hardware complexity. All
the simulations are conducted within MIMO-OFDM framework
and with a packet structure similar to that of IEEE 802.11a/g
standard. As the comparison results show, the RLS algorithm
appears to be an affordable solution for wideband MIMO system
targeting at Giga-bit wireless transmission. As a direct result of
this work, an ASIC for 25MHz wideband 8 8 MIMO-OFDM
system using RLS has been designed and fabricated.
Index Terms MIMO systems, signal detection, packet radio,
wireless LAN

I. I NTRODUCTION
In recent years, multiple input - multiple output (MIMO)
based wireless communications has received widespread attention in the communication community. To date, a majority
of the work in this area has been of a theoretical nature [1],
[2], [3] and little attention has been paid to the implementation
requirements of MIMO systems. Recently the UCLA Wireless
Integrated Research (WISR) group embarked on a project to
develop a wideband (25MHz) real-time MIMO-OFDM testbed
at 5.2GHz RF. The ultimate objective is to develop both system
solution and novel VLSI architecture to enable real-time Gigabps indoor wireless communications.
One of the challenges in building a wideband MIMO system
is the trememdous processing power required at the receiver
side. While coded MIMO schemes offer better performance
than separate channel coding and modulation scheme by fully
exploring the tradeoff between multiplexing and diversity
[4], its hardware complexity can be practically formidable,
especially for wideband system with more than 4 antennas
on both transmitter and receiver sides. On the other hand,
its much easier to find a VLSI solution using traditional
channel coding schemes such as convolutional code and Turbo

IEEE Communications Society / WCNC 2005

Rich-Scattering Fading Channel

Input
bits Channel
Coding

MIMO
Tx

Fig. 1.

MIMO
Rx

MIMO
Detect

Channel
Decoding

Output
bits

Uncoded spatial multiplexing system

code for data rate of hundreds of Mbps. For this reason, we

start off by considering the uncoded MIMO schemes, also
called spatial multiplexing as shown in Fig. 1, and carry out
a side-by-side comparative study to evaluate a number of
uncoded MIMO detection algorithms from both performance
and implementation point of view.
This paper is organized as follows. Section II describes the
channel model and review a list of well-established adaptive and non-adaptive algorithms. The adaptive algorithms
were originally targeted for equalization and beamforming
applications. Here they have been extended specifically for
the detection of MIMO signals. The performance of these
algorithms are simulated and compared in Section III. To make
the results comparable, the simulation is conducted with a
packet structure similar to that of the IEEE 802.11a/g standard.
Section IV compares the hardware implementation complexity
of these algorithms and Section V concludes the paper by
identifying RLS as the preferred solution for Giga-bps wireless
MIMO systems.
II. MIMO D ETECTION FOR F LAT-FADING C HANNEL
A. MIMO Channel Model
Consider a MIMO system where Nt different signals are
transmitted and arrive at an array of Nr (Nt Nr ) receivers
via a rich-scattering flat-fading environment. Grouping all the
transmitted and received signals into vectors, the system can
be viewed as transmitting an Nt 1 vector signal x through
an Nr Nt matrix channel H, with Nr 1 Gaussian noise
vector v added at the input of the receiver
y = Hx + v,

(1)

where y is the received Nr 1 vector. The (nr , nt )-th element

of H, hnr nt , is the complex channel response from the nt -th

408

0-7803-8966-2/05/$20.00 2005 IEEE

transmit antenna to the nr -th receive antenna. x is zero-mean

and has a covariance matrix of Rx = E{xx } = x2 I. v is
also zero-mean and Rv = E{vv } = v2 I.
In frequency-selective fading channels, the channel frequency response hnr nt is no longer characterized by a constant, but rather a function of the frequency
y(f ) = H(f )x(f ) + v(f ).

(2)

When OFDM modulation is used, the entire channel is divided

into a number of subchannels. These subchannels are spaced
orthogonally to each other such that no inter-carrier interference (ICI) is present at the subcarrier frequency under perfect
sampling and carrier synchronization. When sampled at the
subcarrier frequency of fnc , the channel model becomes
(nc )

(nc ) (nc )

(nc )

2) Maximum Likelihood (ML): ML detection has the best

performance among all the MIMO detection algorithms. It
finds the x that minimizes
||y Hx|| = ey ey = (y Hx) (y Hx),

(6)

i.e. the most likely transmitted signal that causes the smallest
difference (squared error) from the received signal. The problem can be solved by enumerating over all possible x and
finding the one that causes the smallest ey ey . As the signal
constellation M and Nt increase, the computational complexity increases exponentially and could become prohibitively
high for practical applications.
D. Linear Adaptive MIMO Detection

, nc = Nc /2, , Nc /2 1.
(3)
With Nc sufficiently large, the subchannel at each of the
subcarriers can be regarded as flat-fading. Therefore, when
using OFDM, the MIMO detection over frequency-selective
channels is transformed into MIMO detection over Nc narrowband flat-fading channels. For this reason, we only focus
on the MIMO detection algorithms in flat-fading channels in
the rest of the paper.

Instead of assuming known channel matrix H, which usually requires channel probing before each transmission and
then calculating W in a bursty manner, adaptive algorithms
estimate W directly through iteration via the use of a known
training sequence at the beginning of each transmission.
1) Least Mean-Square (LMS): LMS is an estimate of the
steepest descent algorithm [5] and updates W according to

B. Linear MIMO Detection

where is the update step size. For LMS to converge in the

mean-squared sense, i.e. Ji J , needs to satisfy 0 <
< 2/max , where max is the largest eigenvalue of Ry =
x2 HH + v2 I. Therefore, the convergence of LMS depends
on both channel condition and signal to noise ratio at the input
of the receiver. The final residual error also depends on the
value of .
2) Recursive Least-Squares (RLS): RLS is the recursive
solution to the exponentially weighted least-squares (LS) problem [5]. The recursive optimal solution at time instant i is

=H

+v

A straightforward approach to recover x from y is to use an

Nt Nr weight matrix W to linearly combine the elements
= Wy.
of y to estimate x, i.e. x
1) Zero-Forcing (ZF): The ZF algorithm attempts to null
out the interference introduced from the matrix channel by
directly inverting the channel with the weight matrix
WZF = H = (H H)1 H .

(4)

2) Minimum Mean-Squared Error (MMSE): A drawback of

the ZF is that nulling out the interference without considering
the noise could boost up the noise power significantly, which in
turn results in performance degradation. To solve this, MMSE
minimizes the mean squared-error, i.e. J(W) = E{(x
)}, with respect to W, and the optimum solution is
) (x x
x
[5], [6]



2
2 1
Wo = Rxy R1
H .
(5)
y = H H + v /x I
C. Nonlinear MIMO Detection
1) VBLAST: A popular nonlinear combining approach is
the vertical Bell Labs layered space time algorithm (VBLAST)
[3]. It uses the detect-and-cancel strategy similar to that of
decision-feedback equalizer. Either ZF or MMSE can be
used for detecting the strongest signal component used for
interference cancellation . The performance of this procedure
is generally better than ZF and MMSE, as will be shown in
Section III.

IEEE Communications Society / WCNC 2005

Wi = Wi1 + [xi Wi1 yi ] yi ,

Wi = Wi1 + (xi Wi1 yi ) yi Pi ,

Pi =


1 Pi1 yi yi Pi1
Pi1
,
1 + 1 yi Pi1 yi

(7)

(8)
(9)

where 
0  < 1 is the
exponential forgetting factor, and
1
i
ik

yk yk
is the inverse of the weighted
Pi =
k=0
correlation matrix of yi with initial condition P1 = 0 I.
The scalar 0 is usually a large positive number and
is very close to 1. Compared to the stochastic estimation
problems given previously which require the signal statistics
such as correlation matrix, LS problem is deterministic [5],
[6]. Therefore, RLS can be used to find the LS solution to a
non-stationary process, or simply said, RLS can track nonstationary process in the LS sense. When xi , H, and vi
are all stationary, it is the weighted time-average estimate to
MMSE as i if Rxy and Ry in (5) are replaced by
i
i
ik
xk yk and k=0 ik yk yk , respectively.
k=0

409

0-7803-8966-2/05/$20.00 2005 IEEE

10

10

ZF
ZFVBLAST

=1

MMSE
MMSEVBLAST

=1
1

10

10

=2

=2

1x1

Bit Error Rate (BER)

Bit Error Rate (BER)

=4
2

10

2x4

1x1

2x2

4x4

8x8

2x4
3

10

8x8

10

=4
1x4

10

2x2
4x4

2x2

1x4
2x8

4x4
1x2

8x8

10

4x8

2x8

10

2x4 1x2

8x8

2x2

4x4

16x16

4x8
2x4

4x8
4x8
5

10

10
0

Fig. 2.

15

10

20
SNR (dB)

25

30

35

40

BER performance of ZF and ZF-VBLAST

Simulations of performance are conducted within MIMOOFDM framework using 25MHz bandwidth. The packet structure used in simulation can be found in [7], which is similar
to the IEEE 802.11a/g standard in frequency domain. In time
domain, each packet has 400 OFDM blocks and a duration of
1.28ms excluding the training blocks in adaptive algorithms.
4-QAM is assumed unless otherwise noted. The simulation
is concluded by calculating uncoded bit/packet error rate
(BER/PER) when 400 packets with errors are collected. But
the total number of packets simulated is no less than 1, 000 and
no more than 40, 000. Perfect sampling and carrier frequency
offset synchronization are assumed throughout the simulation.
The channel is assumed to be quasi-static constant for an
entire packet, but independent among different packets. Each
channel path is generated independently using the exponential
decaying Rayleigh fading channel model [8]. The impulse
response of the channel is composed of equally spaced i.i.d.
complex Gaussian taps with a power-delay profile of
(10)

where Ts is the sampling period. The last tap is determined by

30dB power degradation from the first tap. P0 = 1eTs /rms
ensures that the channel has unity average energy such that
SN R is uniquely determined by
(11)

rms is the RMS delay spread of the channel and is 50ns

in our simulations. Perfect channel knowledge is assumed
at the receiver for non-adaptive algorithms. For LMS/RLS
simulations, a different pseudo-random training sequence is
transmitted before each packet.

IEEE Communications Society / WCNC 2005

10

15

20
SNR (dB)

25

30

35

40

BER performance of MMSE and MMSE-VBLAST

B. BER Performance

A. Simulation Setup

SN R = Nt x2 /v2 .

Fig. 3.

III. C OMPARISON OF P ERFORMANCE

P (k) = P0 ekTs /rms ,

The BER results are shown in Fig. 2 and Fig. 3. In

general, the BER vs. SNR curves are clustered according to
= Nr /Nt . Different algorithms with the same tend to
yield close BER performance, and the higher is, the better
performance because of the diversity gain. Among all the
detection algorithms, the general order from best to worst is:
MMSE-VBLAST, ZF-VBLAST, MMSE, ZF, as expected. At
low SNR, VBLAST, especially ZF-VBLAST, may underperform ZF/MMSE since the gain from successive interference
cancellation has been largely offset by the error propagation
from the substantial amount of decision errors. As the SNR
increases, the number of decision errors decreases, and the performance gain from interference cancellation when compared
to ZF/MMSE becomes evident. However, in ZF-VBLAST
case, this gain is not significant. For instance, for 2 2 at
BER = 103 , ZF-VBLAST is only 2dB better than MMSE,
while for 4 4, the gain is only 2.5dB. In contrast, MMSEVBLAST offers much higher gain. Therefore, if VBLAST is
to be implemented, MMSE-VBLAST is preferred over ZFVBLAST. Note that when Nt = 1, the simulated BER of
MMSE matches the theoretical performance of maximum ratio
combining (MRC) in Rayleigh flat-fading channel. Moreover,
for Nt = 1 with 4-QAM, ZF/MMSE/VBLAST are effectively
the same since the corresponding Ws differ only by a positive
multiple.
The difference between ZF and MMSE diminishes as the
SNR increases, due to the similarity between (4) and (5), and
the relationship between x2 /v2 and SNR in (11). Such effect
is more obvious when > 1 or Nt is large. When > 1,
it requires smaller SNR for ZF and MMSE to become close
to each other due to the diversity gain. For the same SN R
range, x2 /v2 is smaller when Nt is large and therefore ZF
and MMSE are closer. For the same reason, at = 1, MMSEVBLAST performs significantly better than ZF-VBLAST. But
for > 1, at relatively low SNR (when > 1 is really

410

0-7803-8966-2/05/$20.00 2005 IEEE

10

4QAM
16QAM

=1
16QAM
1

1x1

2x2

4x4

1x2
0.7

10

Packet Error Rate (PER)

=2
16QAM

=1
4QAM
8x8
2x4

1x4
8x8

0.6

2x8
0.5

4x8

0.4

2x4

0.3

4x4
2x4

4x8

2x2

4x4

=2
4QAM

10

1x1

=2

0.8

10

Bit Error Rate (BER)

=1

=4

0.9

2x2

1x2

10

0.2

0.1

MMSE
MMSEVBLAST

10

Fig. 4.

10

15

20
SNR (dB)

25

30

35

40

BER performance of MMSE 4-QAM vs. 16-QAM

10

Fig. 5.

15

20
SNR (dB)

25

30

35

40

PER performance of MMSE and MMSE-VBLAST

10

C. PER Performance
The PER performance curves are shown in Fig. 5. ZF and
MMSE yield very close PER performance, and similarly, ZFVBLAST and MMSE-VBLAST (except for Nt = Nr > 1).
For this reason, the curves in Fig. 5 only illustrate MMSE
and MMSE-VBLAST. The performance of VBLAST is consistently better than ZF/MMSE since for PER to reach below
100%, the SNR is already sufficiently high for infrequent error
propagation. At = 1, the PER for MMSE increases with the
number of antennas as compared to the roughly overlapped
curves previously observed in the BER plot.
For rms = 50ns, the channel selectivity leads to a degradation in PER compared to flat-fading channel. This is because
for each packet, its more likely to see bit errors caused by
a deep null in the channel frequency response (corresponding
to lower SNR), than in channels with smaller rms . This is
readily mitigated via interleaved channel coding techniques.
On the other hand, the BER stays the same for different rms

IEEE Communications Society / WCNC 2005

SNR=30dB
SNR=20dB

LMS, SNR=30dB
1

10

Bit Error Rate (BER)

needed), this difference becomes much smaller.

More antennas can generate more diversity gain, even
though is kept constant. Take MMSE for example, when
= 2, BER from best to worst is 4 8, 2 4, 1 2. Similar
argument holds for = 4. For = 1, this trend is not as
pronounced. For VBLAST-based solutions, this is valid in all
cases. The only exception to this trend is ZF with = 1
due to its noise enhancement effect. But as the SNR increases
beyond the range that is shown, ZF yields similar results as
MMSE as they converge to each other.
The benefits of using multiple antennas is further shown in
Fig. 4. For the same throughput rate, using multiple antennas
and smaller QAM constellation is more efficient compared to
larger constellation with a smaller number of antennas. From
Fig. 4, it is clearly seen that increasing power is not the most
efficient way to increase throughput. At BER = 103 , it takes
5 10dB more SNR for Nt Nr 16-QAM system to reach
the same BER performance as 2Nt 2Nr 4-QAM.

RLS, SNR=20dB

2x2

4x4

10

8x8

1x1
3

10

RLS, SNR=30dB

1x1, LMS/RLS, SNR=30dB

4

10

Fig. 6.

50

100

150

250
300
200
Number of iterations

350

400

450

500

The learning curves of LMS ( = 0.02) and RLS ( = 0.99)

as long as the cyclic prefix is sufficiently long compared to

rms . In our simulations, we have used a cyclic prefix length
of 16 40ns= 640ns, long enough for rms = 50ns.
D. Convergence of LMS/RLS
LMS and RLS are resursive alternatives to the matrix
inversion-based solutions. Under the simulation environments,
the BER/PER performance of RLS will ultimately converge
to the MMSE results shown before with sufficient training
while that of LMS should get very close when is very
small. Whats key here is the convergence speed, as shown in
Fig. 6 in the form of ensemble average BER learning curves.
As expected, at the same SNR, LMS converges much slower
than RLS except for 1 1 case where the learning curves
actually overlap with each other. The required training length
depends on various factors including number of antennas,
SNR, and updating factor or . For RLS, the training
length is roughly on the order of Nt Nr . For LMS, its
about 10 times longer. At higher SNR, it takes RLS longer
to converge because the learning curve at higher SNR has

411

0-7803-8966-2/05/$20.00 2005 IEEE

TABLE I
C OMPARISON OF C OMPUTATIONAL C OMPLEXITY

a deeper BER floor to reach while the initial (also fastest)

learning slope for different SNR is similar. Larger or
smaller can increase the speed of convergence for LMS
and RLS, respectively. But on the other hand, they tend to
adversely affect the final convergence performance at the same
time. The convergence of LMS also depends on the channel
condition number (eigenvalue spread). RLS guarantees the
convergence at the price of higher hardware complexity and
less robustness to quantization effects [5], [6]. Similar to the
BER performance, the effect of channel delay spread on the
convergence of LMS/RLS is negligible when the cyclic prefix
is long enough.

LMS
RLS

IV. C OMPUTATIONAL C OMPLEXITY

The comparison of different MIMO detection algorithms
is not complete without the implementation complexity also
factored in. Since the hardware cost of each algorithm is highly
implementation-specific, we try to provide a rough estimation
of the required multiplications for each algorithm based on the
following assumptions:
Complex matrix multiplication ALM BM N requires
L M N complex scalar multiplications;
The matrix inversion operation of matrix AN N takes
O(N 3 ) multiplications [9];
Each complex multiplication consists of 4 real ones, and
each real multiplication takes one operation to finish;
Only the highest and second highest order components,
e.g. N 3 + 2N 2 in N 3 + 2N 2 + N , are counted for the
required multiplications;
3
3
For simplicity, O(N ) is counted as N in calculating
the number of Giga-operation per second (GOPS).
The corresponding results based on these assumptions are
given in Table I. The GOPS figure is calculated for 25MHz
bandwidth and 4-QAM modulation. Due to the over-simplified
nature of these assumptions, the estimation in Table I is only
meaningful in the order of magnitude sense. The estimations
for LMS and RLS are per iteration and those for the nonadaptive algorithms are for estimating W and do not take
into account the cost of estimating the channel matrix H.
For all algorithms except ML, there is an additional cost for
= Wy, as shown in the bottom row of the table.
estimating x
This overhead is only significant for LMS and RLS.
From the comparison, ML is easily ruled out for practical purpose. All the non-adaptive algorithms demand higher
processing power than adaptive algorithms. As a reference,
the fastest commercially available DSP from TI runs at 5.7
GIPS, or about 1.2 GOPS. For single ASIC solution, LMS and
RLS appear to be the only realistic candidates for wideband
application with 4 to 8 antennas. A closer observation reveals
that, RLS is no longer much more expensive compared to
LMS, as conventional wisdom might suggest. This is mainly
due to the fact that in MIMO scenario, the cost of LMS is
multiplied by Nt since there are Nt signals to estimate, while
in RLS, updating Nr Nr inverse correlation matrix Pi brings

IEEE Communications Society / WCNC 2005

Giga Op. per Second

Real
Algorithm

Multiplications

44

88

16 16

8Nt Nr + 2Nt

13

52

14Nr2

+ 8Nt Nr + 6Nr

36

143

19

154

1,229

+2Nt3 + 4Nt2 Nr

33

452

6,622

ML

4Nt Nr M Nt

410

4 105

1011

Wy

4Nt Nr

26

ZF/MMSE

4Nt3 + 8Nt2 Nr

ZF/MMSE

Nt4 + 83 Nt3 Nr

-VBLAST

in no extra cost for 1 < Nt Nr than Nt = 1. Furthermore,

RLS offers the advantage of superior convergence speed and
insensitivity to the eigenvalue spread of the channel [5]. The
above implementation complexity is estimated on subcarrierby-subcarrier basis for MIMO-OFDM system. By utilizing the
frequency domain correlation between adjacent subcarriers,
the hardware cost can be further reduced by linearly interpolating W of non-pilot subcarriers from the estimates on pilot
subcarriers [7].
V. C ONCLUSIONS
In this paper, we have provided an overview of various
MIMO detection algorithms for spatial multiplexing systems.
The simulated performance of these algorithms is compared,
and this comparison is further extended to a first order estimation of their hardware costs. From these comparisons, it
is observed that VBLAST generally outperforms ZF/MMSE,
at the cost of significantly higher implementation complexity.
In fact, ZF-VBLAST is only slightly better than MMSE
(2 3dB at uncoded BER of 103 ). The advantages of
VBLAST over ZF/MMSE become much less significant when
> 1 because of the antenna diversity gain. MMSE-VBLAST
performs much better than ZF-VBLAST. However, in practice,
inaccurate estimates of x2 /v2 in (5) as welll as the channel
matrix itself tend to reduce this gain.
The study shows that RLS has a much lower computation
intensity than ZF/MMSE/VBLAST and achieves the performance of MMSE with sufficient training. This is done by
spreading out the computation through multiple iterations.
RLS is superior to LMS in terms of convergence speed, and
its hardware cost is on the same order as LMS. Compared
to ZF/MMSE/VBLAST, RLS doesnt require explicit channel
information and subsequent matrix inversion, and can be
implemented using the QR-decomposition based systolic array
architecture [5] [6]. Therefore, it can be concluded that RLS
presents the best performance to complexity metric among the
surveyed algorithms for Giga-bps MIMO wireless systems.
Based on these findings, RLS has been chosen for the MIMO

412

0-7803-8966-2/05/$20.00 2005 IEEE

detection in the UCLA MIMO-OFDM testbed [10], [11].

In parallel, a MIMO detection chip using QR-decomposition
based RLS systolic array structure and frequency domain linear interpolation [7] has been developed with TSMC 0.18m
technology to accomodate up to 8 8 antenna configuration,
512 OFDM subcarriers in 12.5 MHz bandwidth using a single
chip. This architecture is further scalable in the frequency
domain to support 88 transmissions with n512 subcarriers
in n 12.5 MHz bandwidth through using n identical chips.
R EFERENCES
[1] G. J. Foschini and M. J. Gans, On limits of wireless cmmunications in
a fading environment when using multiple antennas, Wireless Personal
Communications, vol. 6, pp. 311335, March 1998.
[2] G. J. Foschini, Layered space-time architecture for wireless communications in a fading environment using multi-element antennas, Bell
Labs Tech. J., vol. 1, no. 2, pp. 4159, Autumn 1996.
[3] P. W. Wolniansky, G. J. Foschini, G. D. Golden, and R. A. Valenzuela,
V-BLAST: An architecture for realizing very high data rates over the
rich-scattering wireless channel, in Proc. ISSSE-98, Italy, Sep. 1998.
[4] L. Zheng and D. Tse, Diversity and multiplexing: a fundamental
tradeoff in multiple antenna channels, IEEE Tran. Inform. Th., vol. 49,
pp. 107396, May 2003.
[5] S. Haykin, Adaptive Filter Theory, 3rd ed. Prentice-Hall, 1996.
[6] A. Sayed, Fundamentals of Adaptive Filtering. Wiley-IEEE Press, 2003.
[7] J. Wang and B. Daneshrad, Performance of linear interpolation-based
MIMO detection for MIMO-OFDM systems, in Proc. IEEE WCNC04,
vol. 2, Atlanta, GA, USA, 2004, pp. 981986.
[8] N. Chayat, Tentative criteria for comparison of modulation methods,
IEEE P802.11-97-96, Sep. 1997.
[9] G. H. Golub and C. F. V. Loan, Matrix Computations, 3rd ed. John
Hopkins University Press, 1996.
[10] R. Rao and et.Al., Multi-antenna testbeds for research and education
in wireless communications, IEEE Commun. Mag., vol. 42, no. 12, pp.
S4150, December 2004.
[11] S. Lang, R. Rao, and B. Daneshrad, Design and development of a 5.25
GHz software defined wireless OFDM communication platform, IEEE
Commun. Mag., vol. 42, no. 6, pp. S612, June 2004.

IEEE Communications Society / WCNC 2005

413

0-7803-8966-2/05/$20.00 2005 IEEE