Professional Documents
Culture Documents
Multiuser Detection in Heavy Tailed Noise
Multiuser Detection in Heavy Tailed Noise
doi:10.1006/dspr.2002.0457
1. INTRODUCTION
The problem of detection in multiple access channels has received much at-
tention. Multiuser detection (MUD) techniques are being intensively investi-
gated because they have the ability to combat the multiple access interference
(MAI) caused by the presence of more than one user in the channel. At best
MUD can achieve single user performance by removing MAI altogether. Most
MUD schemes are designed under the assumption of Gaussian noise and tend
1
This work was in part supported by the Australian Telecommunications Cooperative Research
Centre (AT-CRC).
262
1051-2004/02 $35.00
2002 Elsevier Science (USA)
All rights reserved.
Zoubir and Brcich: Multiuser Detection in Heavy Tailed Noise 263
to be linear in nature; the decorrelating detector which removes MAI for high
signal-to-noise ratios (SNRs) is simply a linear transformation [13].
Impulsive noise has been shown to be the more accurate description
of reality for many communications channels including urban and indoor
wireless channels [8, 9]. Impulsive noise generally arises from transients
such as electromagnetic discharge, automobile ignition, and the operation of
common appliances such as microwaves and fluorescent lighting. Unfortunately,
impulsive noise can severely degrade the performance of detectors based on the
Gaussian assumption, warranting the design of MUD schemes for impulsive
noise [1–3].
One way to proceed is to assume a parametric form for the noise distribution,
chosen from one of the many impulsive noise models, and then to design an
optimal or suboptimal detector. This assumes the chosen distribution is a good
model; of course it can only be an approximation to physical reality. How far
the model deviates from reality and what effect these deviations have on the
estimator become problematic and we can only hope that the estimator is robust
with respect to deviations in the assumed model.
To decrease the sensitivity of the estimator to the underlying distribution one
may turn to the theory of robust estimation and use M-estimators to implement
a suboptimal nonlinear detector which is robust to changes in distribution [14].
We propose a nonparametric procedure based on the M-estimation concept
but with the noise density estimated from the observations. This enables the
detector to adapt to various noise models with minimal a priori information; at
most we assume a unimodal symmetric density. To estimate the noise density
we use a nonparametric kernel estimator modified to yield good estimates for
impulsive noise.
This paper is organized as follows. In Section 2 we present the MUD problem.
The proposed nonparametric detector is presented in Section 3 after which
we discuss nonparametric density estimation in Section 4. Some results are
presented in Section 5 before conclusions are drawn in Section 6.
2. SIGNAL MODEL
K
xn = (S)nk Ak bk + vn n = 1, . . . , N. (1)
k=1
where ϕ(v) = −fV (v)/fV (v) is the location score function of fV (v). It is clear that
without a priori knowledge of fV (v) estimation of θ cannot be optimal.
In impulsive noise of unknown distribution both these methods have severe
disadvantages. First, the decorrelating detector is sensitive to impulsive noise,
and its performance rapidly degrades as the level of impulsive behavior
increases. Second, even for known fV (v), the computational cost of ML is often
too high.
To alleviate the general problem of estimation in the presence of outliers or
impulsive noise, Huber proposed the concept of M-estimators [5, 7]. In an M-
estimator the function − log fV (v) is replaced with a similarly behaved penalty
function, ρ(v). The penalty function is chosen to confer robustness on the
estimator under deviations from the assumed density. One then estimates θ by
solving
N K
(S)nk ψ xn − (S)nk θk = 0, k = 1, . . . , K, (6)
n=1 k =1
where ψ(v) = ρ (v). ψ(v) is commonly known as the influence function; when
ψ(v) = v the equations reduce to LS.
When fV (v) is unknown one cannot be sure whether ψ(v) is close to ϕ(v).
Huber considered estimation of location in a nominal Gaussian distribution
Zoubir and Brcich: Multiuser Detection in Heavy Tailed Noise 265
The principal of the soft limiter is intuitively obvious. When outliers are
present they will be clipped by the limiter, reducing their deleterious effect on
the estimator.
Solutions to (6) can be obtained by iteratively reweighted least squares or the
iterative modified residual method proposed by Huber, which is used here. To
achieve a consistent solution it is required that E[ψ(v)] = 0. Here we make the
assumption that the noise density is symmetric, so that this condition is met if
ψ(v) is antisymmetric.
M-estimates using Huber’s penalty function have been considered in the
context of MUD; see [14] for more details. Next we consider motivations for
and suggest a variation on M-estimation where the penalty function is not set
in advance but is estimated from the observations.
3. A NONPARAMETRIC DETECTOR
Although the soft limiter influence function is well motivated, like any
minimax solution it may be far from optimal for many distributions in the
class for which it was designed. Also, its asymptotic optimality property is not
indicative of its behavior for small samples.
If the penalty function could adapt itself to the observations the estimates
may exhibit robust behavior over a wider class of distributions. This is what
is proposed here; instead of fixing the penalty function we base it on the
observations. To achieve this a nonparametric density estimator is utilized
to estimate the score function, ψ(v), in (6). Score function estimation is
incorporated into the iterative parameter estimation procedure used to obtain
the M-estimates. At each step the parameters are updated and the residuals
found, these residuals are used to estimate the score function which in the
next step is utilized to update the parameters and so forth. The procedure is
summarized as Algorithm 1.
A LGORITHM 1 (Iterative procedure for the nonparametric detector).
1. Initialization
Set i = 0. Obtain an initial estimate of θ from the decorrelator, θ̂ 0 = θ̂LS .
266 Digital Signal Processing Vol. 12, Nos. 2,3, April and July 2002
v̂ = x − S θ̂ i
fˆV (v)
ψ(v) = − .
fˆV (v)
4. DENSITY ESTIMATION
where K(x) is the kernel function, h the global bandwidth, and λn the local
bandwidths. As kernel methods inherit the properties of K(x) it is recommended
to ensure K(x) is differentiable, nonnegative, and integrates to one; this
guarantees a valid density. We used a Gaussian kernel since it fulfills all these
requirements and gives good results over a wide range of distributions. The λn
are used to adjust for local structure in f (x) such as heavy tails, while h alters
fˆ(x) globally and may be used to control the smoothness.
Bandwidth selection. To determine h and λn we proceed as follows. First,
find a pilot estimate fˆ0 (x) by setting λn = 1 and for a suitable h, as will
be discussed. Second, evaluate λn = (fˆ0 (xn )/ N ˆ
n=1 f0 (xn )
1/N )−β where 0 ≤ β ≤ 1
ˆ
controls the sensitivity to f0 (x). A recommended choice is β = 1/2 since
asymptotically the resulting estimator will have a bias of lower order than when
β = 0. Finally, evaluate fˆ(x) at the required x.
For fixed λn , an optimal h can be obtained by minimizing a measure of distance
between fˆ(x) and f (x) with respect ∞to h. An oft-used measure is the mean
integrated squared error, MISE = E[ −∞ (fˆ(x) − f (x))2 dx]. For the special case
of a Gaussian distribution and kernel one obtains hopt = 0.79RN −1/5 , where R
is the interquartile range. In general hopt is a reasonable choice even for non-
Gaussian K(x) and f (x); however, should f (x) be multimodal this choice gives
large errors. In this case, or merely to refine h, one must turn to more intelligent
methods of bandwidth selection such as cross-validation or resampling schemes
such as the bootstrap. In cross-validation one estimates the MISE from the
observations and then minimizes this with respect to h. More details can be
found in [6, 10].
Bandwidth corrections for heavy tailed distributions. Even with local
bandwidth adjustments the AKE may undersmooth and produce spurious peaks
in the tails for heavy tailed distributions. Modes in the tails are undesirable
as they produce a score function which oscillates in the extremities. This can
lead to inaccurate estimates as the iterative scheme used in the parameter
estimation algorithm will only converge to a unique minimum if the step size
µ ≤ 1/|ψ (x)| [14]. To avoid this problem one can adaptively set the step size,
µ = 1/δ max |ψ (xn )|, where δ ≥ 1 allows for a margin of error when estimating
ψ (xn ); we used δ = 1.25. However, rapid oscillations result in small step sizes
which slow convergence.
Possibly the simplest way to remove these peaks is to selectively increase
the local bandwidths. We can then correct the tails while not distorting the
estimate near the mode. First we define the tail regions as the lower and upper
100ζ % of the observations; we used ζ = 0.25. For each tail we move from the
268 Digital Signal Processing Vol. 12, Nos. 2,3, April and July 2002
5. SIMULATIONS
We compare the decorrelating detector and a robust MUD [14] which uses M-
estimates with Huber’s penalty function to the proposed scheme for a variety
of noise models. We refer to the method of [14] as the robust detector and
the proposed scheme as the nonparametric detector. Unless otherwise stated
all noise models are symmetric and unimodal, this being incorporated into
the density estimator. Bit error rates (BER) versus SNR are shown for K = 6
users. The spreading codes are shifted maximal length sequences of length
N = 31, meaning density estimation is performed with a sample size of 31. The
amplitude of the first user is at −10 dB relative to the others; BERs are shown
for this user.
For Gaussian noise in Fig. 1 the decorrelator and robust detectors perform
similarly. This is expected as for Gaussian noise the influence function for
the robust detector clips only a minority of the incoming observations, but
Zoubir and Brcich: Multiuser Detection in Heavy Tailed Noise 269
FIG. 1. Bit error rate vs SNR in Gaussian noise (left) and generalized Gaussian noise with
p = 0.5 (right).
appears linear, the optimum shape in Gaussian noise, for the majority. Since the
nonparametric detector estimates the score function it will never be perfectly
linear for Gaussian noise, leading to a decrease in performance of approximately
1 dB.
The generalized Gaussian distribution has a density proportional to
exp(−|x|p ), 0 < p ≤ 2. A p less than 2 gives a heavy tailed distribution suitable as
an impulsive noise model. In Fig. 1 we have shown results for p = 0.5. The non-
parametric detector has the lowest BER up to an SNR of 9 dB. There appears to
be a range of p and SNR which favor use of the nonparametric detector, namely
for low p and SNR, a region where accurate detection is difficult. This suggests
that the nonparametric detector is more robust in highly impulsive noise.
Next we consider a special case of the ε contaminated mixture model where
the contaminant has a Gaussian distribution with variance κγ 2 . This is often
used as a model for impulsive noise as it approximates Middleton’s class A
model [9]. Results for ε = 0.01, κ = 100 and ε = 0.1, κ = 100 are shown in Fig. 2.
Both the robust and the nonparametric detectors outperform the decorrelator,
FIG. 2. Bit error rate vs SNR in ε mixture noise with ε = 0.01, κ = 100 (left) and ε = 0.1,
κ = 100 (right).
270 Digital Signal Processing Vol. 12, Nos. 2,3, April and July 2002
FIG. 3. Bit error rate vs SNR in SαS noise for α = 1.5 (left) and α = 1.0 (right).
while the robust detector outperforms the nonparametric detector in the first
case but not the second. Which detector is best obviously depends on the model
parameters, but we suggest a possible explanation. In the first case there is a
small proportion of very large noise values, making tail estimation difficult as we
have so few observations in the tail regions. In the second case large noise values
occur with a greater probability, aiding in tail estimation and hence estimation
of the score function, leading to smaller BERs. The set of (κ, ε) for which one
detector outperforms the other has yet to be determined.
Next we consider two cases where the noise is symmetric alpha stable (SαS).
SαS distributions are a popular model for impulsive noise for which there has
been shown to be a physical motivation in multiuser environments [8]. SαS
distributions are parameterized by their characteristic exponent, 0 < α ≤ 2,
and a scale parameter σ > 0. The level of impulsive behavior increases as α
decreases. For α = 2 the distribution is Gaussian, while for α = 1, we have the
highly impulsive Cauchy distribution. For α < 2 they have infinite variance and
for α ≤ 1 an infinite mean. Results for characteristic exponents of 1.5 and 1.0 are
shown in Fig. 3. In these cases the nonparametric detector outperforms both the
robust and the decorrelating detectors. Further experiments have shown this to
be true for smaller values of α. This gives more evidence to the suggestion that
the nonparametric detector is more robust to highly impulsive noise.
Finally we show an example where the noise distribution is symmetric but
bimodal. This is incorporated into the density estimation procedure as outlined
in Section 4. The bimodal density is composed of two Gaussian densities with
means of ±1 and a variance of σ 2 = 0.01. Figure 4 shows results for this
scenario. The nonparametric detector is able to adapt itself to the bimodal
nature of the noise and outperforms the other methods at low SNR.
Although asynchronous users and fading multipath channels were not
considered, experiments show a similar relative performance between the
detectors.
To briefly comment on computational complexity, we found the nonparamet-
ric detector took on average four iterations to converge while the robust detector
took two.
Zoubir and Brcich: Multiuser Detection in Heavy Tailed Noise 271
FIG. 4. Bit error rate vs SNR in symmetric bimodal noise generated from a two component
Gaussian mixture model where the components have a mean of ±1 and a common variance of 0.12 .
6. CONCLUSION
REFERENCES
ABDELHAK ZOUBIR received the Dipl.-Ing. degree (BSc/BEng) from Fachhochschule Nieder-
rhein, Germany, in 1983, the Dipl.-Ing, (MSc/MEng) and the Dr.-Ing. (PhD) degree from Ruhr Uni-
versity Bochum, Germany, in 1987 and 1992, all in Electrical Engineering. He is currently Professor
of Telecommunications at Curtin University of Technology and Head of the School of Electrical and
Zoubir and Brcich: Multiuser Detection in Heavy Tailed Noise 273
Computer Engineering. His general interest lies in statistical methods for signal processing with
applications to communications, sonar, radar, biomedical engineering and vibration analysis. He is
a Senior Member of the IEEE, serves on the IEEE SPS Technical Committee on Signal Processing
Theory and Methods and is an Associate Editor of IEEE Transactions on Signal Processing.
RAMON BRCICH received the B.Eng. degree in Aerospace Avionics (Hons 1) in 1998 from
Queensland University of Technology, Brisbane, Australia. At present he is pursuing the Ph.D.
degree in signal processing at Curtin University of Technology, Perth, Australia, where he is also
a Research Fellow in the School of Electrical and Computer Engineering. His current research
interests include irregular sampling and applications of heavy tailed modelling, particularly to
telecommunications.