Professional Documents
Culture Documents
EQUALIZERS
Timothy F. Settle, Michael D. Zoltowski, and Venkataramanan Balakrishnan
Purdue University
Electrical and Computer Engineering Department
West Lafayette, Indiana
settle@ecn.purdue.edu, mikedz@ecn.purdue.edu, ragu@ecn.purdue.edu
past samples (not symbols) from each channel, encompassing a fraction of the mainlobe of the symbol waveform [3].
Equalization with sampling L times per symbol with L > 1
will be referred to as sample-spaced equalization.
There are several advantages that accrue from the fact
that we have access to multiple channel outputs. For example, it is well-known that in the case of a single channel, if
the channel frequency response contains a null in the baseband, a linear equalizer will amplify noise power in the
vicinity of the null. This problem is commonly referred to
as noise enhancement. However, this is not true with multiple channels; here, noise enhancement arises only when
the channels have a spectral null at or near the same frequency. Another advantage with multiple channels is that
equalization can be achieved with FIR filters whose length
roughly equals that of the channel impulse responses; in
contrast, exact equalization of a single channel almost always requires an IIR equalizer.
Referring to Figure 1, designing a linear equalizer for
linear time-invariant channels requires the equalizing filter
coefficients to effect a frequency response for the overall
system from s(n) to y (n) which has unit magnitude and
linear phase. We however, propose that by deliberately permitting equalization errors, i.e., allowing deviations from
unit magnitude and linear phase of the overall frequency response, yields performance gains under certain conditions.
In particular we seek to improve the signal-to-noise ratio
(SNR) in the symbol decision statistics beyond that achieved
with exact equalization design methods.
We will make use of the bit-error-rate (BER) of the symbol decisions as a performance parameter to substantiate our
claims. We demonstrate that as we increase the allowable
equalization error, an attendant improvement in the BER
of the symbol decisions is achieved. In other words, by
permitting an increase in ISI we are able to achieve a significant decrease in the power of the additive noise in the
symbol decision statistic, while having negligible impact on
the signal power in the symbol decision statistic.
ABSTRACT
Multi-channel linear equalizers are commonly employed in
digital communication systems to mitigate channel distortions and thereby reduce inter-symbol interference (ISI).
Traditional multi-channel linear equalizer design for linear time-invariant channels requires synthesis of equalizing
filters such that the frequency response of the overall system has unit magnitude and linear phase. In this work, we
investigate the efficacy of deliberately allowing equalization errors, i.e., allowing deviations from unit magnitude
and linear phase of the overall frequency response, with
the goal of increasing the signal-to-noise ratio (SNR) in the
symbol decision statistics. Our approach relies on numerical convex optimization based on linear matrix inequalities.
Our investigations reveal that as the allowable equalization
error is increased, the mean-square error (MSE) of the postequalizer signal constellation decreases despite the attendant
increase in ISI. This implies that the trade-off between SNR
and ISI in the symbol decision statistic weighs more heavily
in favor of SNR improvement for a surprisingly large range
of allowable equalization error.
1. INTRODUCTION
To our knowledge, in all previous equalizers proposed for
narrowband digital communication systems that use Nyquist
pulses, the decision statistic for a given symbol value is a
linear combination of past and future symbol-spaced samples. The total number of samples required depends on
the effective duration of the tails of the symbol waveform.
These samples may be taken across multiple channels which
may be either different antenna outputs or virtual channels
as synthesized in fractionally-spaced equalization. We will
refer to this type of equalization as symbol-spaced equalization.
In many scenarios the multi-path delay spread, , is a
fraction of the symbol time interval. In some cases, oversampling permits one to compute the equalized decision
statistic for a given symbol value from a small number of
This research work is supported by the National Science Foundation
under grant no. MIPS-9708309.
2. PROBLEM FORMULATION
2.1. SIGNAL MODEL
Transmitter
x(n)
Modulator
hRF1(n)
N
X
hRF2(n)
i=1
s(n)
psr(n)
u1(n)
r2(n)
p*sr(-n)
g1(n)
p*sr(-n)
g2(n)
p*sr(-n)
gNc(n)
y(n)
Receiver
uNc(n)
i
i
(n)
a
h (n)
Consider the baseband multi-channel model for a digital communication system shown in Figure 1, where each
of the Nc channels is realized through a separate antenna
element at the receiver. For the model shown in Figure 1,
the sampled version of the transmitted signal is
=
1
X
k=1
b(k )psr (n
kL):
(1)
D );
(2)
psr (n)
(nL):
i (n)
hRF
psr (
n);
psr (n)
psr ( n), which has a raised cosine spectrum with
excess bandwidth parameter . The symbol-spaced channels, hai (n), are the down-sampled versions of the corresponding hi (n)s. Exact equalization for a symbol-spaced
equalizer would then require the gi (n)s to satisfy,
N
X
c
(n) = K (n
s(n)
u2(n)
rNc(n)
where K ( ) is the Kronecker delta function. This relationship implies that only a fixed time delay is imposed on the
transmitted symbol waveform sequence and therefore ISI
is eliminated given that we are using square-root Nyquist
pulses psr ( ) for the symbol waveform.
Due to the fact that we are sampling with L
2(1 +
) , we must down-sample the outputs of the matched filters by L in order to permit symbol-spaced processing for
symbol-spaced equalizer design. Assuming sample-spaced
estimates of the channel impulse responses are available,
symbol-spaced (aliased) channels may be constructed as follows,
Channels
hRFNc(n)
r1(n)
i (n)
hRF
i=1
i (n)
i
g
(n) = K (n
D ):
(3)
fjPN
b(k )
jg
(4)
jg
(5)
+ (m) 2
fj PN
b(k )
+ (m) 2
i=1 giH gi
c
In (4) and (5), scales the information symbol and the term
(m) is the residual ISI. Since b(k ) is constant, if we can
constrain 1 and (m) 0 in (4) or (5), then it follows
that minimizing the denominator of (4) or (5) improves the
SNR of the symbol decision statistic.
PNc H
i=1 gi Rvv gi
(6)
C = [h1 ; : : : ; hNg 1 ];
D = h0 ;
(7)
where fh0 ; h1 ; : : : ; hN 1 g are the impulse response coefficients. Given state-space realizations for Hi (z ), Gi (z ), and
zD of the form shown in (7), it is straightforward to write a
state-space realization for E (z ) with (AE ; BE ; CE ; DE ),
with CE and DE being affine functions of the filter coefficients of g1 (n); : : : ; gNc (n).
From standard results in system theory [1, x2.7.3] the
constraint in Problem (6), i.e., kE (z )k1 ", can be shown
to be equivalent to the LMI, in PE = PH
E,
2 H
3
AE PE AE PE
AH
PE BE
CH
E
E
4 BHE PE AE
BH
E PE BE "2 I DH
E 5 0:
CE
DE
I
(8)
We denote the matrix in (8) as Q(PE ; CE ; DE ).
The objective function, f (g1 ; : : : ; gNc ), in (6) may be
rewritten in a form suitable for LMI optimization as follows. Let gi = [DGi j CGi ]T . By introducing a slack
variable , we can minimize f (g1 ; : : : ; gNc ) by requiring
f (g1 ; : : : ; gNc ) and minimizing . From standard
results in system theory [1, x10.1.1], the constraint
f (g1 ; : : : ; gNc ) can be shown to be equivalent to the LMI,
2
[DG1 j CG1 ]
66 DHG1
R
vv1
66 CHG1
..
..
66
.
64 H.
DGNc
0
CH
G
Nc
..
3
GNc j CGNc ]
7
7
0
7
7
0:
..
7
7
.
7
5
R
vv1
[D
(9)
We will denote the matrix in (9) as F(; CGi ; DGi ). Note
that this same matrix can be used for symbol-spaced design
by replacing Rvv with Ravv and using the equalization error
function defined for symbol-spaced equalizer design.
The matrices F() and Q() are affine functions of PE
and the equalizer filter coefficients. We may now write our
original multi-channel equalizer optimization problem as a
LMI optimization problem.
Minimize:
Subject to: PE
= PH
E
F(; CGi ; DGi ) 0
Q(PE ; CE ; DE ) 0:
(10)
This problem can be very efficiently solved using standard numerical techniques [1]. In this section we developed the design equation for a sample-spaced zero-forcing
equalizer based on LMI optimization. The simulation results in the next section will compare five versions of the
zero-forcing equalizer. These versions are: Sample-Spaced
Minimum-Norm (Sas-ZF-MN), Sample-Spaced LMI (SaSZF-LMI, developed in this section), Sample-Spaced Windowed LMI (SaS-ZF-WLMI), Symbol-Spaced MinimumNorm (SyS-ZF-MN), Symbol-Spaced LMI (SyS-ZF-LMI).
Each of the design methods based on the minimum-norm
solution yield zero-forcing equalizers which achieve exact
equalization, i.e., the equalization error is zero. Whereas
the LMI based designs deliberately allow a non-zero equalization error. The reader should consult [2] for details on
the other design methods not covered in this paper.
RF1
and H
RF2
Equalized Spectrum
Magnitude
1.5
SaSZFMN
SaSZFLMI
SaSZFWLMI
0.5
0
2
Frequency in Hz
10
4
x 10
Equalized Spectrum
SySZFMN
SySZFLMI
1.2
1.1
Magnitude
3. SIMULATION RESULTS
We now compare the five different design methodologies
listed in Section 2.5, using 16-QAM modulation. For each
simulation the symbol rate is 40 KHz, L = 5, i.e., Ts =
5s, the symbol waveform has the square-root raised cosine
spectrum with = 0:35, the sample-spaced channels are
modeled as a three tap FIR filters with unity energy gain and
the symbol-spaced channels are modeled as seven tap FIR
filters. Two channels are used in all simulations (Nc = 2).
For the sample-spaced designs the length of each equalizer
is three and for the symbol-spaced designs each equalizer
has a length of seven. The design parameter D is chosen
such that D = ((Nh +Ng 1)=2), where () is a function
which rounds its argument to the nearest integer. Our results
are summarized in Figures 3, 4, and 5 shown below.
1
0.9
0.8
1.5
0.5
0
Frequency in Hz
0.5
1.5
2
4
x 10
Frequency Responses
Magnitude in dB
10
2
4
HRF1
H
10
RF2
8
3
0
Frequency in Hz
a
1
10
3
4
x 10
a
2
10
BER
Magnitude in dB
0
2
4
H1
a
H2
10
10
8
2
1.5
0.5
0
Frequency in Hz
0.5
1.5
SaSZFMN
SaSZFLMI
SySZFMN
SaSZFWLMI
SySZFLMI
10
x 10
10
10
MSE versus
SaSZFMN
SaSZFLMI
SySZFMN
SaSZFWLMI
SySZFLMI
9.5
10
MSE (dB)
15
20
25
30
35
10.5
11
11.5
12
12.5
13
13.5
10
SNR (dB)
0.05
0.1
0.15
0.2
0.25
0.3
0.35