IEEE TRANSACTIONS ON COMMUNICATIONS, VOL 36. NO.

5, MAY 1988
605
_
I_Ilu IlCI uHd 5quuIC JImIH_ HCCCVCIy
MARTIN OERDER, STUDENT MEMBER, IEEE, AND HEINRICH MEYR, FELLOW, IEEE
A bstract-Digital realizations of timing recovery circuits for digital
data transmission are of growing interest. In this paper, we present a new
digital algorithm which can be implemented very efficiently also at high
data rates. The resulting timing jiller has been computed and verified by
simulations. In contrast to otber known algorithms, tbe one presented
here allows free running sampling oscillators and a new planar filtering
metbod whicb prevents synchronization hangups.
1. INTRODUCTION
_
IGITAL realizations of reccivcrs for synchronous data
signals-baseband as well as QPSK or QAM signals-are
of growing interest as the capabilities of signal processors (for
low data rates) and application spccifc intcgratcd circuits (for
high data rates) increase. These receivers have to include
algorithms for timing recovery. Several such discrete-time
algorithms have been proposed during the last few years [1]
[3]. The majority of these solutions, howcvcr, include only the
integration of one part of the timing synchronization, namely,
the generation of some kind of timing error signal, into the
digital part of the receiver. This error signal is then typically
used to control an analog veo which generates thc sampling
strobes.
Due to the advantages of an integrated realization, however,
as much of the receiver as possible should b digital. This
means that the input signal should be sampled at a fxed rate by
a free running oscillator and all frther processing should then
be done digitally using these samples. For symbol detection,
this means that thc optimum decision metrics must be
generated from the given samples by some sort of interpolation
which is controlled by an estimate of the current timing offset
[4]. Therefore, wc need an algorithm which determines this
absolute timing offset (not only a timing error signal) from thc
given samples of the signal.
Such an algorithm is proposed in this paper. It is the digital
counterpart of the well-known continuous-time filter and
square timing recover [5], [6], but it extracts the timing
information from the squared signal in a new way. The
analysis of the timing jitter presented in this paper leads to
results that are similar to the continuous-time case, although
the metho of analysis is different.
Another main contribution of thc paper is a new method of
hangup-free filtering of the timing signal . With all other
known timing recovery methods a major problem is that the
synchronization loop can get stuck at an unstable equilibrium
point. In this paper, we show how this can be avoided through
planar filtering of two-dimensional timing esti mates.
The final section of the paper presents a digital realization of
the timing detector which is suitable for VLSI integration also
at high data rates.
Paper approved by the Editor f(lf Signal Design, Modulation, and Detection
of the IEEE Communications Society. Manuscript received April 18. 1987;
revised September 23. 1987. This paper was presented at GLOBECOM '87,
Tokyo, Japan. November 1987.
The authors are with Lehrstuhl fur Elektrische Regelungstechnik, Aachen
University of Technology, West Germany.
IEEE Log Number 8820399.
II. TIMING ESTIMATION
Here we consider the timing recovery for digital data
transmission by linear modulation schemes (PAM, QAM,
PSK). The rcccived signal (PAM) or the equivalent low-pass
signal (QAM, PSK) can be written as
O
r(l)= angr(t-nT-E(t)T)+n(t)
u(t) + n(t).
(I)
Where an are the complex valucd transmitted symbols with
mean power I (e.g., I, i with QPSK), gr(t) is the
transmission signal pulse, Tis the symbol duration, n(t) is the
channel noise which is assumed to be white and Gaussian with
power density No, and (t) is an unknown, slowly varying
time delay.
Now timing recovery means the estimation of the delay (t)
to enable the optimal dctcction of the data. Because varies
very slowly, in a digital realization, we can process the
rcccived signal section by section. For each section Lm' we
can assume E to be constant and obtain an estimate Em. This
estimate must then be combined with the previous estimates
(i.e., it must be fltered) such that the optimal estimate Em is
obtained. The latter can b used to control an analog or digital
sampler for the detection.
Below we consider a special type of timing estimator which
is particularly suited for digitl realization. It is similar to the
continuous-timc flter and square synchronizer in that the input
signal is squared and the resulting spectral component at the
symbol rate is extracted by a filtering operation. In Fig. I the
algorithm is shown. Afer a receiving filter [impulse response
gR(t)l the signal f(t) = r(t) * gR(t) is sampled at rate NIT
("*" denotes a convolution). We thus have samples
h=f(kTIN).
(2)
The sequence
with
represents the samples of the filtered and squared input signal
and contains a spectral component at liT. This spectral
component, which in a conventional synchronizer is cxtracted
by a PLL or a narrow-band flter is here determined for every
section of length LT (i.e., from LN samples) by computing
the complex Fourier coefficient at the symbol rate
(m+l)U\'-1
Xm=
xke-
j2dIN. (4)
k-
mLN
As is shown in the next section, the normalized phase Em =
-1/27 arg (Xm) of this coeflcient is an unbiased estimate for
E.
090-6778/88/0500-0605$01.00 1988 IEEE
606
TIN

Fig. I.
Ilm.11 LN 1

-jlnk/N
xke
x,
k: mLN
Discrete-time filter and square estimator.
The sampling rate must be such that the spectral component
at IIT can still be represented, i.e., NIT> 21T We use N "
4 for practical reasons. In the case of bandwidth efficient
modulation with a singlesided bandwidth of less than liT, the
receiving flter gR(t) also has a singlesided bandwidth of less
than liT and thus the squared signal has a singlesided
bandwidth of less than 21T. Therefore, with N 4. the
sequence Xk completely describes the underlying continuous
time signal.
III. STATISTICS OF THE ESTIMATE
In this section. we compute the statistics of the estimate m
as a function of the pulse form g(t) and the noise power
density No of the additive noise. We assume m = and omit
IEEE TRANSACTIONS ON COMMUNICATIONS. VOL. 36, NO.5. MAY 19
Using the identity (A6) fom Appendix A, we obtain for tl
expectation of X
LN-l
E[X] = E[xk]e-iM1N (1
kO
with
5lx(t)]= _x(t)e-jh/t d
t
.
At this point, we introduce the following functions to simpli
the notation:
pAt)
= g(t)g*(t -nT)
p"
(
f
)
= 5 [Pn(t)].
(1
(1
the index m for the sake of simpler notation. We then have
A. Mean
The mean of the estimate is
E[]=E

arg (X)| (5)

LN
E[Xl =
"
5[Po(t-ET)J/ liT
LN
= Po(l/T)e-i2n
T
(1
For small variance of the estimates wc can linearize the arg
and thus
operation.
-I
E[fl= arg (E[X])
27
=--arg
.
E[XkJe-J2dIN
27
kO
(6)
The linearization is valid, of course, only for I arg(X)1 7.
However, due to the subsequent filter operations, which are
discussed in Section IV, this is the only case of interest.
We first have to compute the expectation of the squared
signal
E[xd=E
..
a"g(kTIN-nT-ET)+f(kTIN)
(7)
The expectation must be taken with respect to the Jomt
distribution of the symbols a
n
and the noise n(t). Noise and
symbols are independent of each other. Therefore, and with
E[n(t)] " 0, the cross term of the binomial in (7) vanishes.
The remaining terms are
..

..
ang(kTIN-nT-ET)
+ E[I f(kTI N)i2]
O O
= E[ana!lg(kTIN-nT-ET)
. g*(kTIN-mT-ET)+E[lf(kTIN)i21. (8)
With noise power u2 and independently distributed symbols of
E[] =-- arg
LN
p0(1IT)e-J2,,

2 T
1
=E--arg Po(lIT).
21
Therefore, under the assumption
arg Po(1IT)=O,
(I
(I
is an unbiased estimate of the clock phase E. But even if (I
is not valid, the mean of exactly equals the required sampli
offset as we show below. We assume
gR(t) = gj( - t + 0 T) (generalized matched filter).
(J
We then have
g(t)=go(/-oT) with gO(t)=gT(t) * gj(-t)
(J
1(t)= ango(/-nT-ET-OT)+fi(t).
C
Since go(t) is symmetrical, the optimal sampling instant is
goU
0), i.e., for the symbol an at
topt, n = nT +ET +0 T,
i.e., the required sampling offset is ( + O)T. Evaluating (
then yields
Since go(t) is symmetrical, thc Fourier transform in (21)
rcal.
Therefore, we have
ag Po (1 IT) = -20
mean power 1, we have and thus
(9)
E[E] = E+ 0 (
:
which is exactly what is required for symbol detection.
OERDER AND MEYR: DIGITAL FILTER AND SQCARE TIME RECOVERY
B. Variance
Here we determine the variance of the random variable E,
i.e., the mean square error of the estimation. We assume
arg Po(lIT)=O (24)
to simplif the notation. (It can easily be shown that the results
are valid for arbitrary and arg Po.) We ten have
var l;J
=
E[;2]
1
=
(2)2
E [(arg (X2]
607
The three terms [30a)-c)] represent the parts of the timing
jitter that are generated by (signal x signal), (signal x noise),
and (noise x noise) interaction.
C. Conditions for Asymptoticaly Jitter-Free Timing
Recovery
In this section, we study the conditions to be fulfilled by the
transmit and receive flters necessary for the timing estimate to
have zero variance in the noiseless case (No = 0), i.e., the
conditions for the s x s-portion of the variance
(34)
1 E [(1m X)2]
-
(21)2 (E [Re X])2 .
to be zero.
(25) We have
The latter approximation is valid since the imaginary part of X
has zero mean and the variances of both imaginary and real
part are small compared to the squared real mean.
From (14) and (24) follows
LN
E [Re Xl =E[Xl =-Po(lIT). (26)
T
The variance of the imaginary part is
LN-\ LN-\
= E[XkXk'] sin (hkIN) sin (hk'IN) (27)
k'"O k"O
with
(28)
By using some approximations which are valid for large L, the
expectation can be computed (Appendix B). If the results are
used in (25), we obtain
with
(29)
(30a)
1 1 2h
u;xn
=
(2
1
)
2
L
N
o
(
Po
(l IT2
(30b)
T
1 1
'
Re c(l/T)
u2
= N2 (30e)

(21)2 L
0
(Po(1/T2
h= g(t)g*{t')r(t-t')
sin (21tlT) sin (21t' IT) dt dt' (31 )
'(f) = 5["2(t)l
(32)
"(7)= gR(t)g(t+7) dt.
(33)
Pm(t) = g(t)g*(t -mT) (35)
Pm (f) = G(f)
*
(G*( -f)e-j2'fmT
)
G(f- v)G*( -v)CJ2wnT
d
v
(36)
and use of the abbreviation
H(/)= G(l/T-f)G*( -f) (37)
yields
For real valued g(t). i.e., symmetrical joint transfer function
of the transmit and the receive filter, we have
G*( -/) = G(f)
H(f)=G(l/T-f)G(f).
That means that H(f) is symmetrical around 1I2T and
(39)
(40)
1m Pm(l/T) = |1m H(v) eos (21vmT) dv. (41)
Therefore, a sufficient condition for zero jitter is
1m H(f)=O
which can be obtained, for example, with
gel) = g( t) (symmetrical pulse shape)
and also of course with all linear-phase pulses
g(7+t)=g(7-t)
(42)
(43)
(44)
as they act like the corresponding symmetrical pulse g(t - T)
with an additional timing delay fO T. The conditions (43)
and (4), however, show that the optimal receive filter in the
synchronization path is a matched filter
(45)
These results are valid, of course, only with the approxima
tions made in Section III-B, in particular, only for large
estimation intervals LT. In the case of short intervals, the
estimate exhibits jitter, but the spectrum of the jitter has a zero
at the origin and can thus be suppressed by low-pass filtering.
True absence of jitter can be obtained in general only by using
nonoverlapping pulses.
This is in contrast to the conventional continuous-time filter
and square timing recovery. In the continuous-time case, the
timing is determined by detecting the zeros of the timing wave,
608
Therefore. true jitter-free timing recovery is possible if the
timing wave exhibits only amplitude jitter. but no phase jitter.
The latter can be achieved. for example. with locally
symmetric pulses [7 J. In our case. however. the estimation is
based on samples which have an arbitrary offset from the zeros
ami thus exhibit random amplitude fluctuations. Therefore.
only asymptotically jitter-free recovery can be obtained.
D. Simulation Results
Fig. 2 shows the variance of the estimates f (29) for several
estimation intervals L where both transmit and receive filter
are fourth-order Butterwortb filters with corner frequency 0.71
T and the modulation format is 8PSK (solid lines). The
markers show the results of Monte Carlo simulations (5000
estimates for each point). In addition, for L 64, the three
parts of (29) are shown by the dotted lines. The simulations are
very close to the theoretical results. Only for L < 4 are there
errors due to the approximations in the computation of the
variance which are valid for large L only. For L = I and EI
No " 0 dB the simulation result is smaller than the theoretical
result. This is due to the finite range of E. Thc variance tcnds to
1112 when f is uniformly distributed in the estimation range.
Fig. 3 shows the corresponding curves for linear phase
filters with a transfer function amplitude similar to thc above
Butterworth filter. For L 2 16 and with moderate EINo the
simulation results match the theory very well. In particular,
the predicted missing of an s x s-portion of the variance can
be seen. Because the absolute variances are much smaller than
in Fig. 2, the effects of the finite observation intervals LT are
much more visible here, especially for large EI No
E. Frequency Offset
Since we usc a free running sampling oscillator, a frequency
offset between transmit and receive timing may be present
rcsulting in a continuously rising or falling ,. In contrast to
carrier recovery, however, the frequency offset m tlmmg
recovery is vcry small (10-5 10
-
2 of thc symbol rate). We
can thcrefore always find an observation length L L' for
which we can consider E to be approximately constant. Thcn
all considerations of the previous sections apply. For L > L'
inspection of the estimation algorithm reveals that the estimate
X is nothing but the average over estimates from shorter
intervals. Therefore, also the mean of the estimate f is just the
average of the timing delay E over the observation interval, as
long as the variation of E is smailer than T12. The latter
condition, of course, limits the possihle observation length L.
Similarly, for small frequency offsets, the variance of the
estimates can be expected to be nearly independent of the
frequency offset. For larger frequency ofsets, one would have
to examine whether the algorithm behaves like its continuous
time counterpart that exhibits a signifcant increase in timing
jitter in the presence of frequency offset.
IV. PLANAR FILTERING OF THE ESTIMATES
Due to frequency offset and random variations of the delay
E, the observation length L is limited. The estimation,
however. can be significantly improved if the knowledge of
the statistical properties of E is used to postfilter the estimates.
For example, a simple "random walk" model for the time
delay E leads to a first-order Kalman filter. The variance of the
fltered estimates can be computed from the variance of the
unfiltered estimates and the random walk parameters [8J.
Since the range of the estimates fm is finite, the filter
innovation must be reduced to the range [-0.5,0.5] as shown
in Fig. 4.
With this kind of filtering, however, the following situation
can occur. If the true value E is at a value 0.5 distant from the
filtered value E, the estimates E also vary around this value and
thus the innovation is at 0.5 and vanishes in the mean. Then
tEEE TRANSACTIONS ON COMMUNICATIO'S. VOL :In. NO. ', MAY 198
1t-02
t
L
: lE-94
IF-B5
1
M
iE-lb
^ L=l
5xn
IE B7----------------
15 2B 25 30 35
4
B
E/No
Fig. 2. Computed and simulated variance of the timing estimate for sever<
estimation intervals L. Transmit and receive filter are fourth-orde
Butterworth, I, 0.7/ T. Dotted lines: 5 x 5, 5 X n, and n X n-part fa
L 64.
IE-01
1E-02
L
IE-B]

<
; 1E-04
IE 05-
C

IE -0 7

--

o Ii 15 20 25
I
3B
Fig 3, As fig, 2. but hnear-phase filters,
Fig. 4. Filtering of the estimates.
x L=I I
L=
4
I
L=1b I
L=b4 I x
35
I
40
E/No
the filter is in an unstable equilihrium amI it can remain ther
some time in spite of the large error.
This is a problcm which arises also for other models an
filters and with almost all known synchronization methods
even if they-like the one presented above-could be calle,
.
'open loop" at first sight. Because of the periodic behavior c
the filtered delay "hangups" can occur in the filter loop.
Below we present a realization which avoids these prot
lems. The central idea is to filter a complex phasor instead c
the corresponding (periodic) angle. The term Xm from Fig.
is such a phasor. Instead of first determining the angle of thi
phasor and filtering the angle, we can apply a Kalman filter t
the phasor itself and use the angle of the filtered phasor t
control the sampling (Fig. 5).
In Figs. 6 and 7, a situation for filtering of E and fltering c
X is shown. In the first case a hangup can occur if the error i
E is approximately 0.5 (corresponding to an angle of 7). Wit
planar filtering, however. the fltered value X movcs eorreetl
(with a step width which depends on the filter coefficient
present) towards its place. Thus, hangup problems cann<
occur any more.
OER]ER AN] MEYR DIGITAL FILTER AND SQUARE TIME RECOVERY
FIg. 5. Planar filtering at the e'timale,.
Fig. 6. Example for the trajectory of estimate and filtered value with
filtering of the delay values
Re
Fig. 7. fig. 6, but planar filtering.
The filtered estimate E-m of Fig. 5 has of course the fnite
range L 0.5, 0.5J again and with only small variation if X",
jumps can occur bctwccn 0.5 in Em. The interpolation unit.
however, which is controlled by (m can easily discriminate
these "wrap-around" jumps from true variations of the time
delay and therefore determine the underlying infinite range
estimate and correctly compute the decision metric [4].
As a final remark. let us note that with the digital filter and
square timing estimation, the planar fltering is nothing but a
(weighted) summation of succcssivc values X",; and this is
merely an extension of the estimation interval L T in the
algorithm for computing Xm (Fig. 1) with an additional
weighting of the terms.
V. REALlZATlON OF THE DETECTOR
Fig. 8 show, a possible realization of the computation ofXm
which allows high data rates through the use of parallel
processing and pipelining.
With a double set of latches. the quadruples of samples
belonging to an estimation interval of lenglh L " I are
collected. Since the sin and cos functions take on only values 0
and I, no multiplications arc necessary. The samples can
then be processed at the symbol rate II T rather than at the 41 T
sampling rate. Squaring and addition can be divided by latches
into further pipeline stages. Thus, the fact that the estimation
algorithm needs 4 samples per symbol (instead of one or two
as other algorithms do) is relevant virtually only 10 the AID
converter and therefore the estimator can be used even at high
data rates. We are currently incorporating the detector into a
CMOS standard cell chip which will run at about \0 Mbits/s.
In cases where a low number of samples per symbol is
important (e
.
g . . when adaptive ceho cancellers are used), the
actual sampling rate can be reduced to 2 samples per symbol
by using a simple all-pass filter to generate the missing
samples [2].
609
[: Lat,h
T
s

T" Jc- == -nL.
T. -- nL. _
T,

T, ...nL ..
Fig. 8. Fast digital realization of the detector.
VI. CONCLUSIONS
The proposed timing recovery enahles a VLSI realization of
digital receivers which can operate on a sampled input signal
without any feedback to the sampling device. The latter can
operate at a fixed rate with a free running oscillator. The
planar fltering algorithm results in very fast and hangup-free
timing recovery.
ApPENDIX A
Equivalence of discrete-time and continuous-time computa
tion of the Fourier coeffcients of periodic band-limited
functions. Assuming x(t) to be a Tperiodic and NI2T-band
limited signal, we show that
kO LN 1
LN
'

x(kTIN)e )hIN
=

x(t)
e-J2fII
T
d
t,
T 0
(AI)
To do so, we stali with the integral form. Due to the low-pass
limitation, we can rewrite the signal by using sine-interpola
tion (sine x = sin x/x)
\
T
x(t)e-
j2,(/T dt
"0
=

;
x(n
TIN) s
in
e
_
t
_n_ T_IN

'0
_
TIN
kO I /-1
x(kTIN)
k=kO
T
kO+I"-1

x(k
T
IN)e-j2dIN
.
LN
*-'0
e-)2.1IT dt
(A2)
(A3)
(A4)
(AS)
610
In paricularq with x(t) " 2:_ _ y(t nT), we havc
k
O
+L.-I
LN
x(kTIN) e-j2.
k
IN=-Y(IIT) ,
T
ApPENDIX B
We frst compute the expectation
E
lxkX
k'
]
=E anjg(kTIN-nT) +i(kTIN
an2g(kTIN -nT) + ii(kTI N
*
"
2
an3g(k'TIN-nT) +i(k'TIN
"3

an4
g(k'TIN-nT) +ii(k'TIN
*
|
(A6)
IEEE TRANSACTIONS ON COMMUNICATIONS. VOL. 36. NO.5. MAY II
"
g*(kTIN-n
2
T) E[n(k' TIN) n*(k' TIN) ]
O
=No Po(kTIN-iT) <(O) (BII
i= T
E
3
= E[a
nj
a:
2] g(kTIN-nj T)
n, n2
g*(k' TIN-n
2
T) E[n(k' TIN) n*(kTIN) ]
O
=No g(kTIN-iT) g*(k'TIN-iT)
i= -0
"
ck-k') TIN) (81
E
4
=E[n(kTIN) n*(kTIN) n(k' TIN) n*(k' TIN) ]
=N&<2(0) +N&<2k-k') TIN) . (B1
The corresponding terms in (27) are now computed. TI
approximations are valid for large estimation intervals LT.
LN-I LN-I ! !
(BI)
Sll= Po(kTIN-iT)
With
E[a;i(t) ]=0
E[a;] =0
E[a;an = O'j
E[a,aj] =0
E[a,oj",aI1 = ( i
for i =j*k /
for i=/*j=k
for i=j=k=/
otherwise
(B2)
(B3)
(B4)
(B5)
(B6)
(B7)
10 of thc 16 terms which result from (B 1) vanish. The
following terms remain:
with
E, = E[an
j
a:
2
an3a:
4
]
T
l
n2 '3 n4
. g(kTI N -n1 T) g*(kT/N-n2 T)
.
g(k'TIN-n3T) g*(k'TIN-n4T)
!
= Po(kTiN-iT) Po(k'TIN-jT)
1= ! J=-o
O 0
+
pj(kTIN-iT) p/k'TIN-iT)
i= -0)= !
O
+(--2) Po(kTIN-iT) Po(k'TIN-iT)
J= -Q
(B8)
(89)
k=O k'=O i_-o j=-O
"
PoCk' TIN-jT) sin (21kIN) sin (27k'IN)

LN
|
2
=
T
Im Po(lIT) =0
LN-\ LN-l 0 0
(B1:
S12= pJ(kTIN-iT)
k=O k'=O i=-o j=-o

p
iCk'TIN-iT) sin (27kIN) sin (27k'IN)
O O O
"L pj(kTIN) pj(k'TIN)
k=-o k'=-Oj=-O
sin (21kIN) sin (27 k' IN)
!
=L(NIT) 2 (1m PiC1IT2 (81
J
= -0
L'-l UV-l
SI3=(y-2) Po(kTIN) Po(k
'
TIN)
k=O k' =0
sin (27 kIN) sin (21fk'IN)
O O !
"(y-2)L Po(kTIN-iT)
"
Po(k'TIN-iT) sin (27kIN) sin (21k'IN)
=L(NITj2(y-2) (1m Po(1IT))
2=0
.
(B1
Therefore,
and
O
SI "L(NIT) 2

(1m Pj(lIT2
LN-1 LN-J Q
j= ~
S
2
=No Po(kTIN-iT) <(O)
k=O k' ",0 i_-o
. sin (27kIN) sin (21k'IN) =0
(B1
(Bl
OERDER AD MEYR: DIGITAL FILTER AND SQUARE TlME RECOVERY
LN-! LN-I
S3=No g(kTIN-iT)g*(k'TIN-iT)
k=O k'-O i--o
lk-k')TIN) sin (27kIN) sin (27k' IN)
O
g(kTIN)g*(k'TIN)
k--ck'=-o
Ik - k' ) TIN) sin (27kIN) sin (27k' IN)
"L{NIT)2No |g(t)g*(t')I(t-t')
. sin (27tIT) sin (27t'IT) dt dt'
: " L(NIT)2NoIJ
LN-l LN-l
S4= (12(O)+12k-k')TIN
kO k' O
sin (27kIN) sin (27k'IN)
LN-l LN-I
=N6 12k-k')TIN)
kO k' O
sin (27kIN) sin (27k'IN)
2LI-2
"
N


12{nTIN)
m=O n=
(m n) LVLT
Q

r2(nTIN)
m=O n=-o
(m n) CYV
I
- (cos (27mIN)-cos (27nIN
2
O
=N (LNI2)r2(nTIN) cos (27nIN)
LN2
=-N Re \(1IT)
2 T
Finally, we can write
(BIg)
(BI9)
(B20)
The three terms represent the parts that are generated by
(signal x signal), (signal x noise) and (noise x noise).
Correspondingly, the variance of is
(B2 1)
with

L
Po(1IT)
(1m Pm{lIT2
1
m
I I 213
=
(27)2
L
N
o
Po(1IT2
N' 1
-
LTN6 - Re \(1/T)
1 T 2
(27)
2 LN 2
-Po(IIT)
T
T
1
1
- Re '{lIT)
=
___
N
2 _2
(27)2 L 0 (Po(1/T2
ACKOWLEDGMFNT
611
(B2 2)
(B23)
(B24)
We would like to thank the reviewers for their valuable
comments and suggestions,
REFERENCES
[I] K. H. Mueller and M, Muller, "Timing recovery in digital synchro
nous data receivers," IEEE Trans. Corman., vol. COM-14, pp,
516-530, May 1976.
[2] O. Agazzi, C. -P. J. T7eng, D. G. Messerschmitt. and D. A. Hodges,
"Timing recovery in digital subscriber loops ," IEEt' Trans. Com
mun., vol. COM-33, pp. 558-569, June 1985.
13] F. M. Gardner, "A BPSKiQPSK tlilling-error detector for sampled
receivers," IEEE Trans. Commun., vol. COM-34, pp. 423-429.
May 1986.
[4] M. Oerder. G, Aschid, R. Hib, and H. Meyr. "An all digital
implementation of a receiver for bandwidth effcient communication,"
in Signal Processing 1I1: Theories and Applications. I. T. Young el
al .. Eds. New York: Elsevier, 1986, pp. 1091 -1094.
[5J L E. Franks and J. P. Bubrouski, "Statistical properties of timing jitter
in a PAM timing recovery scheme." IfEE Trans. Corman . . vol.
COM-22, pp. 913-920, July 1974.
r6] M. Moeneclaey, "Preiliter optimization for the filter and 'quare
synchronizer," Arch. Elektr. Ubertragung, pp, 257-261, 1984
[7] A. N, D'Andrea, U. Mcngali, and M. Moro, "Nearly optimum
preljltering in clock recovery," IEEE Trans, Corman . . vol. COM-
34, pp. 1081-1088. Nov, 1986.
[X] R. D. O. Anderson and J. B, Moore, Optimal Filterin". Englewood
Cliffs. NJ: Prentice-Hall, 1979.
*
Marlin Oerder (S '88) was born in Aachen, West
Germany, on August 26, 1957. He received his
Dipl.-[ng. degree in electrical engineering from
Aachen University of Technology in 1984, Cur
rently, he is pursuing the Ph,D. degree at Aachen
University of Technology.
His interests are in the field of digital communi
cations with emphasis on digital synchronizatIOn
techniques.
612
Heinrich Meyr (M'75-SM'83-F'86) receive the
Dipl.-Ing. and Ph.D. degrees from te Swiss
Federal Institute of Technology (ETH) , Zurich, in
1967 and 1973, respctively.
From 1968 to 1970 he held research psitions at
Brown Boveri Corpration, Zurich, and the Swiss
Federal Institute fo Reactor Research, From 1970
to the summer of 1977 he was with Hasler Research
Lboratory, Ber, Switzerland, His last position at
Hasler was Manager of the Research Department.
During 1974 he was a Visiting Assistant Professor
with the Department of Electrical Engineering University of Souther
Califoria, Los Angeles. Since the sumer of 1977 he has been Professor of
IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 36, NO.5, MAY 198:
Electrical Engineering at the Aachen Technical University (RWTH), Aachen
West Germany, His research fouses on synchronization digital signa
proessing aDd in particular, on agoritms and architectures suitble fc
VLSI implementtion, In tis area he is frequently in demand as a consultar
to industrial concers, He has published work in various fields and joural
and holds over a dozen patents.
Dr. Meyr served as a Vice Chairman for the 1978 IEEE Zurich Sm
and as a Interational Chairman for the 1980 National Telecomunication
Conference, Houston, TX, He served as Assoiate Editor for the lEE.
TRANSACTIONS ON ACOUSTICS, SPEECH, AND SIGNAL PROESSING fror
1982 to 1985, and as Vice President for Interational Affairs of the lEE
Comunications Soiety,