Sunteți pe pagina 1din 7

2014 7th Advanced Satellite Multimedia Systems Conference and the 13th Signal Processing for Space Communications

Workshop (ASMS/SPSC)

A Novel Approach for Symbol Timing Estimation


Based on the Extended Zero-Crossing Property

Wilfried Gappmair1, Stefano Cioni2, Giovanni E. Corazza3, Otto Koudelka1


1
Institute of Communication Networks and Satellite Communications, Graz University of Technology, Austria
2
European Space Agency, ESA/ESTEC Noordwijk, The Netherlands
3
Department of Electronics, Computer Science and Systems, University of Bologna, Italy

AbstractCarrier- and data-blind recovery of the sym- be circumvented via a simple compensation coefficient
bol timing is of paramount importance in digital receivers, derived in [7]. Guided by the idea of a prefilter to avoid
i.e., detailed knowledge about carrier frequency and phase the annoying jitter floor, for the first time addressed by
or any pilot sequences is not necessary for proper operation.
In this context, feedforward algorithms are particularly Franks and Bubrouski [8] and in the sequel applied to
useful in packet-oriented systems, where rapid and stable Gardner synchronizers [9], [10], it could be shown in [11]
acquisition of the major transmission parameters is essen- that this approach works for FF concepts as well.
tial for subsequent processing stages. In the current paper,
we propose a novel approach for blind estimation of the With respect to the receiver matched filter (MF), pre-
symbol timing, which needs just one or two samples per filters are arranged in series, which might cause problems
symbol. The new method is characterized by introducing a in several scenarios due to the additional delay. In the
second filter in parallel to the receiver matched filter. Un- current contribution, however, we follow a completely
der Nyquist conditions, this filter exhibits an impulse re- different approach by introducing a second filter operated
sponse satisfying the extended zero-crossing property, i.e., it
in parallel to the MF so that the delay problem is avoided.
vanishes for all integer multiples of the symbol period,
including the origin! Using this idea for a suitably designed In case the noise power is known to the receiver, we end
timing estimator, it is shown that the annoying jitter floor up with an STR estimator, which needs just one sample
typical for most timing estimators and caused by pattern per symbol taken at the MF output and the output of the
noise can be avoided. parallel filter. On the other hand, if the noise power is not
KeywordsBlind feedforward recovery, symbol timing, known, which is perhaps more relevant from the practical
extended zero-crossing property, low-rate sampling. point of view, the algorithm requires twice the sampling
rate at the MF output.
I. INTRODUCTION Moreover, the parallel filter is to be designed such that
Symbol timing recovery (STR) is an indispensable the jitter floor disappears. For real-valued Nyquist pulses
task in digital receivers [1], [2]. One of the major reasons used for baseband shaping, we will show that this goal is
is the fact that many synchronization algorithms for carri- achieved, if the filter satisfies the extended zero-crossing
er frequency and phase require the symbol timing to be (XZC) property, i.e., the impulse response vanishes for all
established in advance. This means that STR methods integer multiples of the symbol period, including the
should be carrier-blind, i.e., knowledge of carrier infor- origin!
mation is not necessary to operate the related STR algo-
The remainder of the paper is organized as follows. In
rithm properly. Furthermore, it is to be noticed that, when
Section II, we introduce the signal model used for analyti-
carrier details are not known to the receiver, the underly-
cal and simulation work, including the design of the XZC
ing data are usually not available or their detection is not
filter. The novel STR approach is developed in Section
reliable enough. Therefore, it is most welcome, when the
III, with emphasis on the computation of the mean estima-
envisaged STR technique follows a carrier-blind and non-
tor value and the jitter variance. Numerical results are
data-aided (NDA) approach.
given in Section IV and, finally, conclusions are drawn in
Feedforward (FF) algorithms for STR are an attractive Section V.
option in packet-oriented systems, where rapid and stable
acquisition is of paramount importance [3]. The Oerder II. SIGNAL MODEL
and Meyr (O&M) estimator [4], normally implemented Let the independent and identically distributed (i.i.d.)
with four samples per symbol, represents perhaps the most symbols ci = ai + jbi be zero-mean and normalized to unit
prominent example in this respect. Mainly motivated by variance such that EC[|ci|2] = 1, where EC[] denotes expec-
the computational load of O&M schemes or modifications tation with respect to the M-ary symbol alphabet C. The
thereof [5], an FF solution with only two samples per unit-energy baseband pulse h(t) is assumed to satisfy the
symbol has been suggested by Lee in [6], whose bias can Nyquist condition; guided by practical reasons, we will

978-1-4799-5893-1/14/$31.00 2014 IEEE 59


2014 7th Advanced Satellite Multimedia Systems Conference and the 13th Signal Processing for Space Communications
Workshop (ASMS/SPSC)

concentrate on a root-raised cosine shape with roll-off Because H1(f) is proportional to the first-order deriva-
factor , where 0 1. Furthermore, it is assumed that tive of H(f), the additional filter needed for the new STR
the receiver signal r(t) is rotated by the carrier phase approach will be denoted as derivative matched filter
[, ) and shifted in time by [T/2, T/2). Therefore, (DMF); for = 0.25, the spectral evolution of H(f) and
with T as the symbol period, we have H1(f) is illustrated in Fig. 2. Observe also that the DMF
output is furnished by
r (t ) = e j ci h(t iT ) + w(t ) . (1)
i
x 1 (t ) = r (t ) h1 (t ) = e j s1 (t ) + n1 (t ) , (6)
In this context, w(t) represents a zero-mean white
with n1(t) = w(t) h1(t) and
Gaussian noise process with independent real and imagi-
nary parts, each of variance w2 = 1/(2s), where s = s1 (t ) = ci g1 (t iT ) . (7)
Es/N0 is the mean signal-to-noise ratio (SNR) per symbol. i

In the sequel, r(t) passes the receiver matched filter h*(t) In this context, it is to admit that the DMF concept has
as shown in Fig. 1. For root-raised cosines, it is to be been suggested in the open literature for a completely
observed that h*(t) = h(t) so that the corresponding out- different scenario, namely for NDA recovery of the carrier
put signal can be written as frequency via a closed-loop design [1].
x 0 (t ) = r (t ) h(t ) = e j s0 (t ) + n0 (t ) , (2) Nevertheless, by detailed inspection of Fig. 2, it is
clear that the sharp edges of H1(f) at f = 12+T involve a
where denotes convolution and n0(t) = w(t) h(t) is a
fairly slow roll-off in the time domain so that the related
zero-mean non-white Gaussian process with variance
FIR filter must be implemented with a correspondingly
2 w2 . Finally, by defining the raised cosine (RC) function
large number of taps. In order to reduce the computational
as g(t) = h(t) h(t), the signal part is given by
complexity, it is suggested that H1(f) is shaped by an addi-
s0 (t ) = ci g (t iT ) . (3) tional filter function P(f), i.e., H 1 ( f ) = H1(f) P(f); the
i overbar indicates that signals at the DMF output are
Matched
x0(t) shaped by P(f). Of course, P(f) has the same spectral sup-
filter
Symbol timing
port as H1(f) and, perhaps more important, it must not
r(t) yk
Derivative
recovery violate the XZC property introduced previously. This is
MF
x1(t) given, when it satisfies

Fig. 1. Dual-filter framework for symbol timing recovery. 1 1


P f + = P f , | f | < , (8)
As already mentioned in the introductory section, we 2T 2T 2T
will introduce a second filter operated in parallel to the as could be shown for a similar problem investigated in
receiver MF. We propose a filter design so that the convo- [13]. Among the shapes fulfilling the condition in (8), we
lution of the impulse response of the filter, henceforth will focus for pragmatic reasons on an RC-based profile,
denoted by h1(t), and the signal part of (1) is zero at all i.e., P ( f 21T ) = cos 2 |fT | , where | f | < 2T and 0 < .
sampling instants t = kT, i.e., h(t) h1(t)|t = kT = 0, k Z. Fig. 2 displays the spectrum of H 1 ( f ) , when forming the
Excluding k = 0, this condition would be satisfied by g(t). DMF with = 0.25.
However, if we consider g1(t) = tg(t) instead of g(t), then
the XZC property is established. This means that h1(t) is
determined by the integral equation h(t) h1(t) = g1(t), 1.5

which might be solved most elegantly by shifting the 1

problem to the frequency domain. Doing this, we have 0.5

that
0

1 G ( f )
H ( f ) H1 ( f ) = , (4) -0.5 __
H (fT )
j 2 f -.- H1 (fT )
-1 H1 (fT )

where H(f), H1(f), and G(f) are the Fourier transforms of -0.75 -0.5 -0.25 0 0.25 0.5 0.75 1

h(t), h1(t), and g(t), respectively [12]. Then, with G(f) = fT


H2(f), we simply obtain
Fig. 2. Spectral evolution for MF, DMF, and RC-shaped DMF ( = =
1 H ( f ) 0.25).
H1 ( f ) = 2t h(t ) = h1 (t ) . (5)
j f

978-1-4799-5893-1/14/$31.00 2014 IEEE 60


2014 7th Advanced Satellite Multimedia Systems Conference and the 13th Signal Processing for Space Communications
Workshop (ASMS/SPSC)

III. DEVELOPMENT AND ANALYSIS OF THE Putting all pieces together, we finally arrive at
NOVEL ESTIMATOR SCHEME
U 0 2 w2 0
Considering MF as well as DMF signals described in U0 = = cos(2 ) , (15)
20
(2) and (6), the corresponding T-spaced samples are im-
mediately obtained as which establishes a relationship between observables x0,k,
x0, k = x0 ( kT ) = e s0, k + n0, k ,
j
(9) processed by (11), and the (normalized) timing offset .
Unfortunately, the noise power 2 w2 has to be known in
x1, k = x1 ( kT ) = e j s1, k + n1, k . (10) advance to obtain (15). If this is not realistic, we could
resort to E[|x0,k|2] E[|x0, k1/2|2], which gives after some
In this context, the noise parts n0,k and n1,k are zero- algebra
mean non-white Gaussian variates, whereas the signal
parts s0,k and s1,k are simply provided by s0(kT) and s1(kT). U 0 = E[| x0,k |2 ] E[| x0,k 1 2 |2 ]
Both sorts of samples, x0,k and x1,k, will be processed ap- = E[| s0,k |2 ] E[ | s0,k 1 2 |2 ] (16)
propriately in the FF algorithm providing a timing esti- = 40 cos(2 ) .
mate , the interpolator is fed with to generate the cor-
rected samples yk, as it is sketched in Fig. 3. Hence, by means of an additional normalization step,
we have that
x0(t)
MF Interpolator yk U 0
U0 = = cos(2 ) . (17)
40
FF estimator
DMF
x1(t) Note that for processing U 0 solely one sample per
symbol is needed, whereas for U 0 two of them are re-
Fig. 3. Feedforward scheme for symbol timing recovery. quired. Note also that (15) and (17) are functions of the
(normalized) timing error, which could be used for a re-
A. Analysis of MF and DMF Output covery method by inverting the relationship as follows:
In order to develop the new STR estimator, we first = 21 arccos(U0). However, since this approach does not
compute E[|x0,k|2], which denotes expectation of |x0,k|2 with provide unambiguous estimates over the full range of
respect to data and noise. By taking into account that errors [1/2, 1/2), it is not really helpful in practice.
E[|n0,k|2] = 2 w2 and that n0,k is statistically independent As a consequence, we will take the DMF output into ac-
from s0,k, we obtain count as well.
U 0 = E[| x0,k |2 ] = E[| s0,k |2 ] + 2 w2 . (11) To this end, we analyze the expected value of
Re[ x0*, k x1, k ] . Recalling that the signal and noise contribu-
Evaluating E[|s0,k|2], it is to be recalled that the unit- tions in x0,k and x1,k are statistically independent, this
variance symbols ci C are i.i.d. so that, by applying the yields
auxiliary result (33) achieved in the Appendix,
U 1 = E[Re{x0*,k x1,k }]
E[| s0,k |2 ] = g 2 [( k )T ] (18)
k = Re{E[ s0*,k s1,k ]} + Re{E[n0*,k n1,k ]} .
(12)
1 m


=
T G( f ) G T
m
f e j 2m df ,

Then, by introducing A() = Re{E[ s0*,k s1,k ]} , the first
term in (18) is furnished as
where we used for convenient reasons the normalized
timing offset = T . Since G(f) is bandlimited by 12+T , A( ) = Re g [( k )T ] g1[( k )T ]
k
the relationship boils down to (19)
1
m j 2 m
E[| s0,k | ] = 0 + 20 cos(2 ) ,
2
(13) = Re
T
G1 ( f ) m G T f e df ,
with by tacitly assuming that the data symbols ci in s0,k and s1,k
(1+ ) 2 T are i.i.d. and that the DMF output is shaped by P(f).
1
0 = G ( f ) df = 1 , Again, the second line in (19) is due to (33) derived in the
2

T (1+ ) 2 T 4 Appendix. Regarding that G(f) is bandlimited and that


(14)
1
(1+ ) 2 T
1 G1 ( f ) = G1 ( f ) P( f ) = j 21 G(f f ) P( f ) , the equation
0 =
T
(1 ) 2 T
G ( f ) G f df = .
T 8 reduces to

978-1-4799-5893-1/14/$31.00 2014 IEEE 61


2014 7th Advanced Satellite Multimedia Systems Conference and the 13th Signal Processing for Space Communications
Workshop (ASMS/SPSC)

sin(2 )
(1+ ) 2 T
G( f ) 1 For comparison purposes, we furnish also the jitter
T 2 (1) 2T f
A( ) = P( f ) G f df . (20) variance for the maximum-likelihood (ML) NDA algo-
T
rithm [16], which is bounded by
In order to complete the analysis of (18), we must also
1
have a look at the contribution introduced by the noise ML
2
= . (27)
components of x0,k and x1,k, i.e., we have to compute 2 L s
Bn = E[n0*,k n1,k ] . Shifting the problem to the frequency Because of the mathematical intricacies, we skip the
domain, we directly obtain detailed analysis of the jitter variance in general form.
Instead, the focus is on the special, but important case
Bn = 2 w2 H ( f ) H 1 ( f ) df . (21) with negligible timing errors.

By examination of (15) and (17), it is obvious that


Because H(f) and H 1 ( f ) are counter-symmetric, as U0 1, when 0, so that we may assume that U 0 is
can be seen by inspection of Fig. 2, it is clear that the approximated by this value as well. In the next step, by
integral in (21) is zero so that Bn does not contribute to finding out the contribution of U 1 , we have to analyze the
(18). Hence, the normalized value of (18) develops as term Re[ x0*, k x1, k ] as can be seen from (25). With 0, it
2 2 is clear that x0, k e j ck + n0,k and x1,k n1,k , which
U 1 = U 1 = A( ) = sin(2 ) , (22) gives vk = Re[ x0*,k x1,k ]|0 Re[ e j ck* n1,k ] in the low-
1 1
noise range. In consequence, if s >> 1, we have that
where 1 = A( ) evaluated at = 0. With P ( f 21T ) = U 1 << U 0 such = 21 arg(U 0 + jU 1 ) 21 U 1 = 1 U 1 . 1

cos 2 |fT | , | f | < 2T and 0 < , the coefficient 1 is Putting all pieces together, this yields
given by
E[vk2 ] 12 E[| n1,k | ]
2

02 = 2 = = , (28)
A( ) 2 sin
2
2 0
12 L 12 L
1 = = 2
= . (23)
=0 4 2 2 3
=
where
Inspecting in detail the relationships (15), or (17), and (1+ ) 2 T
3
(22), we are now able to formulate the STR algorithm as E[| n1,k |2 ] = 4 w2 | H ( f ) P( f ) | df = . (29)
2

32 2 s
1
(1 ) 2 T
1
= arg(U 0 + jU 1 ) . (24) In order to obtain , which specifies the bandwidth of
2
P(f), we require that 02 is equal to ML
2
. To this end, the
Based on an observation length of L symbols at MF ratio (, ) = 0 ML is introduced and for {0.05,
2 2

and DMF output, U 0 and U 1 are obtained by approximat- 0.25} plotted in Fig. 4 as a function of , where 0 < .
ing U 0 or U 0 in (15) or (17), respectively, and U 1 in It is immediately verified that 02 approaches ML 2
, when
(22) with . This will be used in the next section, where nu-
1 L 1 merical results are presented in terms of the mean estima-
U 0 = | x0,k |2 , tor value and the jitter variance.
L k =0
1 L 1
U 0 = (| x0,k |2 | x0,k 1 2 |2 ) , (25)
5

L k =0 -.- = 0.25
4
1 L 1 - - = 0.05

U 1 = Re[ x0*,k x1,k ] .


L k =0 3
(, )

B. Jitter Variance 2

Powerful estimators exhibit no bias, i.e., = E[ ] , 1


and the jitter variance 2 = E[( ) 2 ] is close to the
Cramer-Rao lower bound (CRLB) as the theoretical limit
0.05 0.1 0.15 0.2 0.25
[14]. For computational simplicity, the latter is often em-
ployed in its modified form [15], which is for an observa-
tion length of L symbols and RC shapes with excess Fig. 4. Evolution of (, ) = 02 ML
2
.
bandwidth (roll-off factor) given by
IV. NUMERICAL RESULTS
1
MCRLB = . (26) Due to the related intricacy, we do not provide full
[ 3 (1 + 3 ) 8 2 ] 2 L s
1 2 2

analysis of the timing estimator in terms of bias and jitter

978-1-4799-5893-1/14/$31.00 2014 IEEE 62


2014 7th Advanced Satellite Multimedia Systems Conference and the 13th Signal Processing for Space Communications
Workshop (ASMS/SPSC)

variance. Instead, we will resort in the following to nu-


merical means verifying this way the major figures of XZC
O&M
merit.

0.01 mLee

Norm. jitter variance


We start with the mean error value (MEV), i.e., E[ ] , 0.001

which is approximated by averaging the estimates in (24)


over a number of simulation runs large enough to avoid
0.0001

any side effects. For 4-PSK, L = 100, and = = 0.25, 0.00001

Fig. 5 shows the results parameterized by different SNR


__
MCRLB
- - ML-NDA

values; the impulse response for MF as well as DMF 1. 106


5 10 15 20 25 30

spans a window of 21 symbol periods. For better visibil- Es/N0 [dB]


ity, simulation results are indicated by dots in different
Fig. 6. Evolution of the normalized jitter variance for 4-PSK (L = 100,
style, which have been interpolated by a cubic spline. As
= = 0.25).
can be seen, the estimates deviate from the ideal curve the
more the (normalized) timing offset 1/2; needless to
say that the phenomenon is more pronounced at lower XZC
O&M
SNRs and smaller values of L. The same behavior could

0.01 mLee

be observed with modulation schemes other than 4-PSK,

Norm. jitter variance


0.001
including multi-level signals like 16-QAM or 16-APSK,
and other values of and , although not shown because 0.0001
of limited space.
0.00001
With the same parameter setting used in Fig. 5, the jit-
__
MCRLB
- - ML-NDA
ter variance is illustrated in Fig. 6. As already mentioned 1. 106
5 10 15 20 25 30
previously, the jitter floor vanishes for the XZC estimator.
Es/N0 [dB]
Note that, for = , the XZC performance is close to the
ML-NDA bound, according to the result achieved previ- Fig. 7. Evolution of the normalized jitter variance for 16-QAM (L =
ously (see Fig. 4). For comparison purposes, the evolution 100, = = 0.25).
of the modified CRLB is plotted as well, indicating a gap Furthermore, the XZC estimator has been tested under
between ML-NDA bound and the theoretical limit, which very extreme SNR and roll-off conditions as they are
might be narrowed only by an increase of . Also for discussed by the channel model group for the new DVB-S
comparison purposes, the diagram includes the normal- standard [17], i.e., SNR values as low as -10 dB and =
ized jitter performance of O&M and modified Lee (mLee) 0.05. Due to the small roll-off factor, both MF and DMF
algorithms, both exhibiting a non-negligible jitter floor. are designed now for an impulse response spanning a
window of 81 symbol periods. Using 4-PSK signals, the
0.4
jitter performance is depicted in Fig. 8. An observation
window with L = 104 had to be chosen to achieve useful
Es /N0 = 3 dB
Es /N0 = 9 dB

0.2 results in the very low SNR range. Again, no jitter floor is
observed and the XZC performance is close to that of the
ML-NDA bound. Although not shown in the diagram to
MEV

avoid an overload of the plot, the same behavior has been


-0.2
verified for QAM and APSK signals as well.
-0.4

-0.4 -0.2 0 0.2 0.4 XZC


0.01 O&M
= /T mLee
Norm. jitter variance

0.001
Fig. 5. Evolution of the mean error value for 4-PSK (L = 100, = =
0.25). 0.0001

Using the same parameters as in Fig. 6, simulation re- 0.00001

sults for 16-QAM are visualized in Fig. 7. No significant


1. 106 __

discrepancy is detected with respect to the 4-PSK results;


MCRLB
- - ML-NDA

the only difference is that the jitter floor for O&M turns -5 0 5 10 15 20 25 30

out be somewhat increased. Es/N0 [dB]

Fig. 8. Evolution of the normalized jitter variance for 4-PSK (L = 104,


= = 0.05).

978-1-4799-5893-1/14/$31.00 2014 IEEE 63


2014 7th Advanced Satellite Multimedia Systems Conference and the 13th Signal Processing for Space Communications
Workshop (ASMS/SPSC)

V. CONCLUSIONS APPENDIX
A new approach for carrier-blind NDA recovery of the In the sequel, an alternative solution of the infinite
symbol timing has been introduced in the current paper. sum
The developed feedforward estimator needs just one or
two samples per symbol. However, it requires the opera- q( ) = x ( kT ) y ( kT ) (30)
k
tion of a second filter in parallel to the receiver matched
filter (MF). Assuming real-valued Nyquist pulses for is derived, which turns out to be most useful in case that
baseband shaping, the filter design must satisfy the ex- the spectra of x(t) and y(t) are bandlimited.
tended zero-crossing (XZC) property. It could be shown In the first step, by introducing the definition zk() =
that this corresponds to the first-order derivative of the x(kT ) y(kT ), the corresponding Fourier transform is
receiver MF in the frequency domain; sharp edges causing obtained as [12]
a slow roll-off in the time domain might be smoothed by a

further shaping function not violating the XZC character-
Z k ( f ) = F [ z k ( )] = e j 2kfT X ( f v ) Y ( v ) dv , (31)
istic.

Since the second filter is operated in parallel to the where X(f) and Y(f) are the Fourier transforms of x(t) and
MF, no additional delay needs to be taken into account. y(t), respectively. Then, by applying the Poisson identity
This is in contrast to conventional prefilter structures,
arranged in series to the MF and frequently suggested in 1
the open literature to avoid the annoying jitter floor.
e
k
j 2 kfT
( f m T )
T m
(32)

Furthermore, just to keep the analytical part tractable, and after some straightforward algebraic manipulations,
it has been decided to focus on a signal model, which does the relationship in (30) can be re-written as
not include phase noise effects and impairments caused by
residual frequency offsets. Preliminary simulation results q( ) Z k ( f ) e j 2f df
have shown that this would have an impact insofar as the k
(33)
jitter floor does not disappear, depending on the standard 1
m j 2m T
deviation of the phase noise and/or the percentage of the =
T Y (v ) m X T v e dv .
residual frequency error with regard to the symbol rate.
However, detailed investigations are out of scope for this
paper and will be deferred to future activities. The non- REFERENCES
vanishing floor is easily explained by the violation of [1] U. Mengali and A. N. DAndrea, Synchronization Techniques in
Digital Receivers, Plenum Press: New York, 1997.
Nyquist and XZC conditions; the same happens, if the
[2] H. Meyr, M. Moeneclaey, and S. A. Fechtel, Digital Communica-
root-raised cosine shape is distorted by a nonlinear chan- tion Receivers: Synchronization, Channel Estimation, and Signal
nel, e.g., a high-power amplifier operated close to the Processing, Wiley: New York, 1998.
saturation point. [3] E. Casini, R. De Gaudenzi, and A. Ginesi, DVB-S2 modem
algorithms design and performance over typical satellite chan-
Some words are also due in view of the computational nels, Int. J. Satellite Communications and Networking, vol. 22,
pp. 281318, May/June 2004.
complexity. To this end, we assume that the timing esti-
[4] M. Oerder and H. Meyr, Digital filter and square timing recov-
mate in (24) is evaluated by replacing U 0 and U 1 with ery, IEEE Trans. Commun., vol. 36, pp. 605612, May 1988.
U 0 and U 1 in (25). It is obvious that we need six addi- [5] M. Morelli, A. N. DAndrea, and U. Mengali, Feedforward ML-
tions and multiplications per symbol, all of them real- based timing estimation with PSK signals, IEEE Commun. Lett.,
valued. On the other hand, for O&M as well as mLee vol. 1, pp. 80 82, May 1997.
algorithms, eight additions and multiplications per symbol [6] S. J. Lee, A new non-data-aided feedforward symbol timing
estimator using two samples per symbol, IEEE Commun. Lett.,
are necessary. For all algorithms, the argument of a com- vol. 6, pp. 205207, May 2002.
plex number might be extracted via a simple look-up [7] Y. Wang, E. Serpedin, and P. Ciblat, An alternative blind feed-
table. Finally, comparing the complexity of the involved forward symbol timing estimator using two samples per symbol,
filter structures, no discrepancy is observed between XZC IEEE Trans. Commun., vol. 51, pp. 14511455, Sept. 2003.
and mLee algorithm, if we accept that parallel and prefil- [8] L. E. Franks and J. P. Bubrouski, Statistical properties of timing
jitter in a PAM timing recovery scheme, IEEE Trans. Commun.,
ter are both implemented with the same number of taps. vol. 22, pp. 913920, July 1974.
With O&M estimators using four samples per symbol, [9] A. N. DAndrea and M. Luise, Design and analysis of a jitter-
however, we would need twice the number of taps for MF free clock recovery scheme for QAM systems, IEEE Trans.
Commun., vol. 41, pp. 12961299, Sept. 1993.
and prefilter.
[10] A. N. DAndrea and M. Luise, Optimization of symbol timing
recovery for QAM data demodulations, IEEE Trans. Commun.,
vol. 44, pp. 399406, March 1996.

978-1-4799-5893-1/14/$31.00 2014 IEEE 64


2014 7th Advanced Satellite Multimedia Systems Conference and the 13th Signal Processing for Space Communications
Workshop (ASMS/SPSC)

[11] K. Shi, Y. Wang, and E. Serpedin, On the design of a digital [15] M. Moeneclaey, On the true and modified Cramer-Rao bounds
blind feedforward, nearly jitter-free timing-recovery scheme for for the estimation of a scalar parameter in the presence of nui-
linear modulations, IEEE Trans. Commun., vol. 52, pp. 1464 sance parameters, IEEE Trans. Commun., vol. 46, pp. 1536
1469, Sept. 2004. 1544, Nov. 1998.
[12] J. G. Proakis and D. G. Manolakis, Digital Signal Processing: [16] M. Moeneclaey and G. de Jonghe, Tracking performance com-
Principles, Algorithms, and Applications, Prentice Hall: Upper parison of two feedforward ML-oriented carrier-independent
Saddle River, NJ, 1996. NDA symbol synchronizers, IEEE Trans. Commun., vol. 40, pp.
[13] R. Lopez-Valcarce, J. Villares, J. Riba, W. Gappmair, and C. 14231425, Sept. 1992.
Mosquera, Cramer-Rao bounds for SNR estimation of over- [17] TM-S2 Channel Model Group, DVB-Sx Channel Models, Jan.
sampled linearly modulated signals, submitted to IEEE Trans. 2013.
Signal Processing, May 2014.
[14] S. M. Kay, Fundamentals of Statistical Signal Processing:
Estimation Theory, Prentice Hall: Upper Saddle River, NJ, 1993.

978-1-4799-5893-1/14/$31.00 2014 IEEE 65

S-ar putea să vă placă și