IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 45, NO.

6, JUNE 1997 647

Unbiased Timing-Error Estimation in the Presence of Nonideal Interpolation

Daeyoung Kim, Madihally J. Narasimha, and Donald C. Cox

Abstract— We propose a method to eliminate the bias term timing-error detector to yield an unbiased estimate, which
present in the timing-error estimator employed in digital receivers eliminates the undesirable low-frequency spectral lines from
where the input signal is sampled by a fixed clock which is not the loop noise spectrum. In practice, the bias function can be
synchronized to the transmitter clock. This bias error results
from the nonideal interpolation that precedes the timing-error precalculated and stored in a look-up table of the receiver
estimator. We show that it can be derived as a function of the for some quantized values of the previous symbol-timing
previously estimated symbol timings. An unbiased timing-error estimates.
estimate can then be obtained by subtracting this bias term from
the output of the timing-error detector. Simulation results are
II. SYSTEM DESCRIPTION
included to show the performance improvement realizable by
employing this method. Fig. 1 shows a block diagram of the channel and the data
Index Terms—Interpolation, symbol timing recovery, synchro- receiver, which incorporates the all-digital-timing recovery
nization, timing error detector. scheme considered in this paper. The input signal to
the receiver, after matched filtering, is sampled by a fixed
clock with period . Since the symbol period is derived
I. INTRODUCTION from a source clock that is independent of the sampling

I NTERPOLATION filters are needed to obtain symbol syn-

chronized samples in digital receivers where the input
analog-to-digital (A–D) conversion clock is not synchronized
clock, the ratio is typically irrational. The function of
the interpolator is to generate symbol synchronized samples
suitable for detection.
to the transmitter clock [1]. The interpolator computes the For a sequence of symbols transmitted, the received
synchronized samples based on the timing-error offset supplied signal prior to the sampler is
by a timing-error estimator. The timing-error estimator, in turn,
operates on the output of the interpolator, thereby forming a (1)
closed-loop system.
The interpolated samples, however, are different from the
where is the channel delay, is the overall baseband
actual synchronized samples due to the nonideal characteristics
impulse response, and is the noise
of the interpolation filter—even if it is assumed that the
with known power-spectral density . To recover ,
timing-error estimator supplies the correct timing offset. This
we need the sample of at , whereas only
distortion, in turn, affects the performance of the symbol-
the samples , are available after
timing recovery process. Bucket and Moeneclaey [2] show
sampling by a fixed clock. We first estimate the symbol-
that the nonideal interpolation leads to a bias term in the
timing as , where is an integer and
output of the timing-error detector. This biasing gives rise to a
. Then, we approximate by interpolation
loop noise spectrum consisting of low-frequency spectral lines.
They propose a narrow-bandwidth loop filter to suppress the
spectral lines. However, this method is difficult to implement (2)
when the spectral lines converge toward , which occurs
when the ratio of the symbol period to the A–D sampling where is the interpolation filter for fractional timing .
period is close to an integer. The interpolated value differs from
In this paper, we discuss a method to eliminate these low- because of the nonideal characteristics of the interpolation
frequency spectral lines without employing narrow-band loop filter and errors in the symbol-timing estimate. The symbol-
filters. We first derive the bias term in the timing-error detector timing estimation error is defined as
as a function of the previous symbol-timing estimates. This
bias term can then be subtracted from the output of the (3)
The desired characteristic of the -curve, which is defined
Paper approved by M. Luise, the Editor for Synchronization of the IEEE
Communications Society. Manuscript received August 26, 1995; revised as the average timing-error detector output for the timing-
February 19, 1996 and September 9, 1996. This paper was presented in part estimation error , is [3]. Therefore, we scale
at the IEEE International Conference on Communications, Dallas, TX, June by to obtain , i.e.,
23–27, 1996.
D. Kim was with the STAR Laboratory, Electrical Engineering Department, (4)
Stanford University, Stanford, CA 94305 USA. He is now with Motorola
Information Systems Group, Mansfield, MA 02048 USA. whose -curve has the same slope as the desired -curve at
M. Narasimha and D. Cox are with the STAR Laboratory, Electrical
Engineering Department, Stanford University, Stanford, CA 94305 USA. [6]. In this paper, and stand for the -
Publisher Item Identifier S 0090-6778(97)04166-4. curve of and , respectively. Then is passed through
0090–6778/97$10.00  1997 IEEE
648 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 45, NO. 6, JUNE 1997

Fig. 1. Channel modeling and data receiver incorporating all-digital-timing recovery (without dashed line) and the proposed unbiased timing-error estimator
(with dashed line).

the loop filter, whose output is used to update the symbol-

timing estimate, i.e., and , employing typical control
algorithms [1].
It is known that the -curve is biased (i.e., )
if the interpolator is nonideal [2]. This causes performance
degradations as explained in [2]. In the following section, we
propose a method to eliminate the bias term. First, we derive
the bias term as a function of the two previous symbol-timing
estimations, i.e., and . Then, an unbiased
estimate can be obtained by subtracting this bias value from
the output of the timing-error detector.

III. THE UNBIASED TIMING-ERROR DETECTOR

In this section, we assume that is a sequence of inde-
1
pendent equiprobable symbols belonging to the set
and the decoded output symbol has no error, i.e., Fig. 2. The bias function versus estimated fractional timing
. We also assume that is a white Gaussian noise (T = 2Ts ; = 0:5). (a) Jitter (b) Jitter energy.
independent of .
As derived in the Appendix, for Mueller–Müller’s (M&M)
TABLE I
timing-error detector [4], the expectation2 of with respect INTERPOLATION FILTER COEFFICIENTS
to and , i.e., the -curve, is
(5)
where

(6)

Note that is the interpolated value of .

By introducing the new function , we can express the
equation for the -curve in a form similar to that obtained then is unbiased for all values of . The
when no interpolation (i.e., synchronized sampling clock) is data receiver with this modification is shown in Fig. 1 as the
employed [3]. dashed lines with bias function. Note that the bias function is
The bias factor in (note that ) can be found independent of the input SNR.
from (5). It is given by The performance of the symbol-timing estimator can be
characterized by the mean and the variance of the (steady-
bias state) timing-estimation error defined in (3). From now
(7) on, we will refer to the symbol-timing estimation error as
jitter.
If we modify the timing-error detector output as
Fig. 2 shows the bias function versus when and
(8) is a raised cosine pulse with roll-off factor [7] ,
1 We assume BPSK to make the derivation simple. It is easy to extend the
for two kinds of interpolation filters: a linear filter [5] and
result of this paper for other signal constellations. a second-order polynomial filter with the design parameter
2 Note that we do not take the expectation with respect to  ; n ; and e . [2], [5]. Their coefficients are given in Table I.
k k k
KIM et al.: UNBIASED TIMING-ERROR ESTIMATION 649

IV. EXAMPLES
In this section, to make the analysis simple, we consider
only the case of steady-state performance. We also assume
that is an exact integer so that the jitter power spectrum
due to nonideal interpolation contains a spectral line at .
This cannot be removed by a loop filter. In practice, will
be slightly different and the line spectrum will be located near
. If this spectral line is not significantly attenuated by
the loop filter, the actual performance obtained in practice will
be close to the results shown here.
At steady state, the mean of the jitter will converge to
such that . We can easily show that , i.e.,
the bias value, since has a slope of 1. Therefore, the
energy of the jitter is given by

(a)
(9)

where the variance of the jitter for a loop bandwidth

can be derived as explained in the appendix. We set the loop
bandwidth in the following examples.3
Example 1: We assume that the overall impulse response
is a raised cosine pulse with roll-off factor [7], which
is set to 0.5 in this example. The input SNR is defined as
, where is the energy of the transmitted
pulse. We assume dB and .
Fig. 3(a) shows the jitter versus for the second-order
polynomial filter [5] .4 The solid lines show the
mean of the jitter, and the dashed lines show the value of
mean plus or minus standard deviation . Significant jitter
performance gains are realizable in the case of the second-
order polynomial filter as the magnitude of and the mean
are comparable.
When is exactly two, the nominal value of will
(b) be fixed independent of time, and we can obtain the jitter
Fig. 3. The symbol-timing estimation error versus  for the biased estimator energy from Fig. 3(b). The jitter energy of the biased estimator
6)
with second-order filter. (a) Jitter (solid line: mean, dashed line: mean  . is about 0.85 dB worse than that of the unbiased one in
(b) Jitter energy. (a) Gain versus input SNR. (b) Gain versus . the worst case. We will denote this gain as maximum jitter
energy gain. This value is important since the receiver works
As we are sampling at a point away from the maximum with the same for a long time when is close to an
eye opening [7], the value that determines the symbol-timing integer.
estimation performance is not the variance of the jitter but the The mean and variance of timing-estimation error, when
total energy of the jitter (i.e., mean variance). Removing Gardner’s timing-error detection algorithm [8] is employed
the bias of the timing-error estimation reduces the energy of instead of the M&M algorithm, can be derived using the
the jitter by making the mean zero. method described in the Appendix. The bias term of Gardner’s
When the designed value of is an exact integer (e.g., algorithm is about twice that of the M&M algorithm when
) the actual value in practice will be slightly different . The maximum jitter energy gain for this case is
(say ) due to clock tolerances. In this case, about 5.5 dB.
varies slowly as a function of time. On the other hand, if the Fig. 4(a) shows the jitter energy gain versus SNR for these
design value is not an integer (e.g., ), varies two timing-error detection algorithms. Fig. 4(b) shows the
more rapidly with time. The bias term present in can be jitter energy gain versus .
viewed as the sampling of the bias versus function shown
in Fig. 2. Therefore, when slowly changes as a function of V. CONCLUSION
time, will have low-frequency terms due to the bias. These
cannot be suppressed by the loop filter. However, when We have proposed a method to remove the bias of the
changes rapidly with time, the bias term present in does timing-error estimation caused by nonideal interpolation. The
not have appreciable low-frequency components. Therefore, 3 This value is chosen to make e  3% when the input SNR 10 > dB.
the scheme proposed in this paper will be useful when the 4 We will not look at the linear filter since its jitter variance is far bigger
designed value of is close to an integer. than that of the second-order filter.
650 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 45, NO. 6, JUNE 1997

APPENDIX
BIAS AND AUTOCOVARIANCE OF TIMING-ERROR
DETECTOR OUTPUT IN THE PRESENCE OF INTERPOLATOR
In this Appendix, we explain how to derive the bias and
the autocovariance of the timing-error detector output for
two kinds of timing-error detectors: M&M algorithm [4] and
Gardner’s algorithm [8].
From (1) to (3)

(10)

where is defined in (6) and is

(11)
(a)

We can easily see that (10) has the same form as (4)
in Cowley and Sabel’s paper (C&S) [3]. We can get (10)
from C&S by substituting the interpolated versions of the
transmitted pulse and noise instead of the original. Therefore,
all the results in C&S can readily be used by modifying
the transmitted pulse and the noise to their interpolated
versions, i.e., and .

REFERENCES
[1] F. M. Gardner, “Interpolation in digital modems—Part I: Fundamentals,”
IEEE Trans. Commun., vol. 41, pp. 501–507, Mar. 1993.
[2] K. Bucket and M. Moeneclaey, “Symbol synchronizer performance
affected by non-ideal interpolation in digital modems,” in Int. Conf.
Commun. New Orleans, LA, May 1994, pp. 929–933.
[3] W. G. Cowley and L. P. Sabel, “The performance of two symbol timing
recovery algorithms for PSK demodulators,” IEEE Trans. Commun., vol.
42, pp. 2345–2355, June 1994.
(b) [4] K. H. Mueller and M. Müller, “Timing recovery in digital synchronous
data receivers,” IEEE Trans. Commun., vol. COM-24, pp. 516–531, May
Fig. 4. Jitter energy gain for unbiased estimator over biased estimator. (a) 1976.
Maximum jitter energy gain versus input SNR. (b) Maximum jitter energy [5] L. Erup, F. M. Gardner, and R. A. Harris, “Interpolation in digi-
gain versus . tal modems—Part II: Implementation and performance,” IEEE Trans.
Commun., vol. 41, pp. 998–1008, June 1993.
[6] M. Moeneclaey, “A simple lower bound on the linearized performance
of practical symbol synchronizers,” IEEE Trans. Commun., vol. COM-
bias is derived as a function of the two previous symbol- 31, pp. 1029–1032, Sept. 1983.
timing estimates and is then subtracted from the timing-error [7] E. A. Lee and D. G. Messerschmitt, Digital Communication, 2nd ed.
detector output to yield unbiased estimates. Our simulation Norwell, MA: Kluwer Academic, 1994.
[8] F. M. Gardner, “A BPSK/QPSK timing-error detector for sampled
results indicate that the unbiased estimator can greatly reduce receivers,” IEEE Trans. Commun., vol. COM-36, pp. 423–429, May
the jitter energy of the symbol-timing estimation error. 1986.