Академический Документы
Профессиональный Документы
Культура Документы
Abstract cently, Moon and Soliman derived the ML rule for a time
invariant AWGN channel with intersymbol interference,
We derive the maximum likelihood (ML) decision rule together with two approximations [4].
for locating a frame synchronization pattern in data, dis- Here we present the derivation of the ML rule for a
turbed by flat fading and additive white Gaussian noise. flat fading channel with M-ary phase coherent signal-
We give a high signal to noise ratio approximation which ing. We assume that channel state information (CSI) is
is simpler to implement and almost reaches the perfor- available for each received symbol (which is the case, for
mance of the ML decision rule. The correlation rule is instance, when a subsequent ML Viterbi decoder uses
shown to perform catastrophically. Simulation results CSI to improve decoding performance; CSI perhaps gen-
for the Rayleigh channel show that the new ML decision erated using a kalman filter [5]). As will be explained
rule performs about 1 dB better than the Gaussian ML later, it is assumed that the sync word is spread over
rule. The rules presented here require channel state in- the data frame. We derive a high SNR rule, and show
formation for each received symbol; in many cases this that there is no useful equivalent of the correlation rule.
information is already available or can be obtained with The high SNR rule is easy to implement, and only mi-
relative ease. nor changes need to be made to a frame synchronizer
using the Gaussian high SNR rule from [3]. As Lui and
Tan pointed out, exact analytical analysis of the syn-
1 Introduction chronization rules does not generally seem to be feasible
for other than the correlation rule even for the Gaussian
In the past decades, there has been much research into channel. For this reason, we resort to computer simula-
optimum frame synchronization. Massey derived the tion to evaluate and compare performance.
maximum-likelihood (ML) rule for the Gaussian channel The use of channel state information to improve the
with coherent BPSK modulation, as well as suboptimum performance of a maximum likelihood (Viterbi) decoder
rules [l]. Nielsen analyzed these rules in more detail, in the case of fading was first proposed by Hagenauer
concluding that the well known correlation rule -which in [6]. The new metric was hardly more complex than
was believed to be optimum prior to Massey's work- per- the common soft decision metric, but gave significant
formed several dB worse than the true ML rule and its improvement. Due to the increased importance of mobile
high SNR approximation (high SNR rule) [2]. Further, communications, the fading cha.nne1 has attracted much
he gave an upper bound on the performance of a frame at tent ion.
synchronizer in a noiseless environment, where the ap-
pearance of the synchronization word in the random data
is the only disturbing factor. This upper bound is the
same for the ML, high SNR and correlation rules. 2 Derivation of the Maximum
Many years later, Lui and Tan [3] extended the anal- Likelihood Rule
ysis to M-ary phase (coherent and noncoherent) signal-
ing. Again it was shown that the correlation rule per- We assume a multiplicative flat fading channel followed
formed several dB worse than the ML and high SNR rule. by an additive white Gaussian noise channel (AWGN).
The high SNR rule offers the advantage of being almost Fig. 1shows the discrete equivalent channel model used,
as easy to implement as the correlation rule, and much together with the structure of a possible receiver employ-
easier than the true ML rule (which needs an estimate ing a Viterbi decoder. For simplicity, we use a similar
of the channel SNR and uses complex arithmetic). Re- notation and derivation to that in [3] (Gaussian chan-
348.4.1
1426 CH3132-810000-1426$3.00 0 1992 IEEE ICC '92
nel). The modulation scheme is M-aryl demodulation is 1 1 ~ 1 1is the Euclidean norm of v. Since we wish to maxi-
assumed to be coherent with perfect symbol timing. The mize this conditional probability over p , we can equiva-
set of transmitted symbols is { Wj ,0 5 j _< M - 1). The lently maximize the likelihood function:
data is transmitted in a stream of N-symbol frames, of
which L symbols in each frame are a known synchroniza-
tion sequence S = (SO, S I ,...,S L - ~ ) .B y assuming fast
fading and spreading the sync word across the frame we i=O
have independent channel states for each symbol, see
Fig. 2. The symbols in the data portion of the frame
d = (do, d l , ...,dN-,+l), are assumed to be chosen ran- (3)
all d i = O
domly and uniformly from the signal set. The fading
value (channel state) for each symbol i is represented by Equivalently (leaving out irrelevant factors):
ai and here we assume perfect channel estimation i.e.
ai = ai.
The demodulator output consists of a sequence of N
complex vectors. This sequence x = ( C O ,1 1 , ..., X N - 1 ) is i=O
vectors ni; see Fig. 1. The 2 components of the noise <,> denotes the inner product. Because of the follow-
vectors are iid with zero mean and variance N0/2. NO is ing:
the one-sided power spectral density.
The receiver observes any N consecutive channel out-
puts, and tries to determine the beginning of a new
frame, which must lie somewhere in this sequence x . all d i=O
Also:
i=O j=1
we can maximize:
348.4.2
1427
and, again assuming that all E, are equal:
L-1
j=1
except for the multiplication with the CSI factor corre-
sponding to each symbol.
Ej is the energy of the signal vector W,.
If we compare this likelihood function with that of [3]:
4 Correlation rule
L-1
LML-AWGN(P) = < zg(i)+p,Si > - Inspecting (7), we observe that the first term is equiva-
i=O lent to the normal correlation rule (except for the CSI):
L-1
LC(;L) = ag(i)+p < tg(i)+pr Si >. (14)
i=O
we notice that there is an extra term, and that the chan-
nel state information is used in several places in (7). Note We will compare the correlation rule with the ideal and
that even if all Ej are equal, the last term in (7) cannot high SNR rules.
be deleted.
5 Results
3 High SNR rule Monte Carlo computer simulation results are shown for
\.ire will now derive an approximation of (7) for the c a e BPSK, QPsK and 8PSK modulation schemes in Figs 3
of high signal to noise ratios. In order to do this, we to 5 . The three examples are taken from [3]. Each curve
approximate: shows the percentage of correctly synchronized frames
for the different rules depending on the average channel
SNR. Following the order of the legends, we start by
1
M
g(l)+P
ag(i)+, < zg(i)+,,, W j ' > - E j 7 ) M showing the performance of the ML and high SNR rules
j=1 for the AWGN channel, the difference between the curves
is low. Next we see the ML and high SNR rules with and
ag(i)+p
2
]
< z g ( i ) + p , W' > -Ej- a g ( i2) + P ), ('1
without CSI, on the Rayleigh distributed fading channel.
Finally the performance of the correlation rule without
and with CSI is given.
where 3 = j ( g ( i ) ) is that j which maximizes: We observe that the ML and high SNR rules with CSI
perform up to just over 1dB better than those without
a2 . CSI. We also notice that the high SNR rules approach the
ag(i)+p < z g ( i ) + p Wj > -Ej
3 - dl)tP
2 ' (lo) ML rules at high SNR. Interesting is that the difference
between the ML and high SNR rules at low SNR, is
or larger than for just the AWGN cha.nne1. The greatest
improvement that can be reached when starting with the
< z g ( i ) + p r wj >, (11) AWGN high SNR rule is almost 1.5 dB (ML rule with
348.4.3
1428
CSI). But the low additional complexity of the high SNR [2] P. Nielsen, “Some optimum and suboptimum frame
rule with CSI (assuming CSI is available) makes it seem synchronizers for binary data in Gaussian noise,”
very attractive in practice. An important result is that IEEE Trans. Commun., vol. 21, pp. 770-772, June
for BPSK the high SNR rule with CSI is always better 1973.
than the ML rule without CSI. This is true in the cases
of QPSK and 8PSK for SNRs greater than -1 dB and 2 [3] G. L. Lui and H. H . Tan, “Frame synchronization for
dB respectively. Gaussian channels,” IEEE Trans. Commun., vol. 35,
pp. 818-829, August 1987.
We notice that the correlation rule performs signifi-
cantly worse than the other rules. Looking at the per- [4] B. H . Moon and S. S. Soliman, “ML frame synchro-
formance at high SNR, it appears that the correlation nization for the Gaussian channel with ISI,” in Proc.
rules never give the same performance predicted by [3] ICC ’91, pp. 1698-1702, June 1991.
(random data limited). The explanation -in the case of
no noise- is simple: The correlation term (without CSI) [5] R. Haeb and H. Meyr, “A digital synchronizer
a t the correct position, does not necessarily equal the for linearly modulated signals transmitted over a
energy of the sync word, but will depend on the fading frequency-nonselective fading chanel,” in Proc. ICC
values. At other positions in the frame, which closely ’88, pp. 1012-1016, June 1988.
resemble the sync word, the correlation term might be
[6] J . Hagenauer, “Viterbi decoding of convolu-
higher than at the true position; this leads to a false
tional codes for fading- and burst-channels,” in
sync. The same also applies to the correlation rule us-
Proc. Zurich Seminar on Digital Communicatioiis,
ing CSI, but even to a greater extent: Using CSI in the
pp. G2.1-G2.7, 1980.
correlation rule degrades performance significantly.
6 Conclusions Transmitted
Symbols
348.4.4
1429
m
u
._
C
348.4.5
1430
'7