IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 42, NO.

21314, FEBRUARYIMARCHIAPRIL 1994

1391
The Modified Cramer-Rao Bound and Its
Application to Synchronization Problems
Aldo N. D' Andrea, Senior Member, IEEE, Umberto Mengali, Fellow, IEEE, and Ruggero Reggiannini
Abstract-We introduce the modified Cramer-Rao bound
(CRB) which, like the hue CRB, is a lower bound to the error
variance of any parameter estimator. The modified CRB proves
useful when, in addition to the parameter to be estimated, the
observed data also depend on other unwanted parameters.
The relationship between the modified and true CRB is
established and applications are discussed regarding the
estimation of carrier-frequency offset, carrier phase, and timing
epoch in linearly modulated signals.
Modified CRBs for phase and timing estimation have been
already discussed in previous works where it is shown that several
practical carrier-phase and clock recovery circuits do attain such
bounds. Frequency discrimination, instead, is not so well-
represented in the literature and a significant contribution of this
paper is the calculation of the modified CRB for frequency
estimation. This bound is compared with the performance of
some frequency detectors and it is concluded that further work is
needed in search of more efficient frequency discrimination
methods.
I. INTRODUCTION
Synchronization is a fundamental function in modern
digital communication systems. Its task is to estimate certain
signal parameters, such as carrier frequency, carrier phase and
timing epoch, which are necessary in the demodulation and
data detection processes. Several textbooks on the design and
performance of synchronization systems [ 11-[6] are available
and a host of papers have been published to compare
alternative synchronization schemes suited for various
applications. References [7]-[8] provide valuable guidelines
for understanding the rationale of synchronization algorithms;
they also give a perspective which helps compare these
algorithms according to certain criteria of optimality.
In this context the question arises of the ultimate accuracy
that can be achieved in synchronization operations.
Establishing bounds to such an accuracy is an important goal
since it provides benchmarks for evaluating the performance of
actual synchronizers. Tools to approach this problem are
available from the parameter estimation theory [9]-[lo] in the
form of Cramer-Rao bounds (CRBs), which give fundamental
lower limits to the variance of any parameter estimator. Other
bounds in signal parameter estimation are discussed in [ 111-
CRBs have been derived in [ 151-[ 161 for clock and carrier-
phase recovery in linear modulations and, in [ 17] , in
~141.
(nonlinear) continuous-phase modulation. Unfortunately, an
analogous treatment for carrier frequency estimation is not
available in the literature and a first objective of this paper is to
fill this gap.
Another objective is to make clear the real meaning of
some CRBs which have been found in the synchronization
area. As we shall see, in many cases of practical interest the
computation of true CRBs is a demanding task; in fact the
bounds derived in [15]-[17] (as well as the one we give here
for carrier-frequency estimation) are not exactly CRBs.
Nevertheless, as they have the same structure as the true
CRBs, they will be referred to as modzjied Cramer-Rao bounds
(MCRBs). The relationship between true CRBs and MCRBs is
investigated.
The third topic is to complete the comparison between the
performance of practical synchronizers and the limits predicted
by the theory. This subject has been thoroughly discussed in
[18] for carrier phase and clock recovery. By contrast, results
concerning carrier frequency detectors are comparatively
limited and our aim is to supplement prior literature with some
new findings.
The rest of the paper is organized as follows. In Section I1
we introduce the concept of MCRB and we point out its
relationship to the CRB. In Section I11the MCRBs for carrier
frequency, carrier phase and symbol timing are derived. The
last two bounds coincide with those obtained in [15]-[16]
where they are simply denoted CRBs. In Section IV we show
that the above MCRBs coincide with the true CRBs calculated
under the assumption that some further knowledge is given on
the signal characteristics. Comparisons between MCRBs and
the performance of a few carrier-frequency detectors are made
in Section V. Finally, in Section VI we draw some
conclusions.
11. MODIFIED CRAMER-RAO BOUND
We assume that the received waveform has a complex
(1)
envelope
r(t) =s ( t ) +w(t ) ,
Paper approved by Costas N. Georghiades, the Editor for Synchronization
and Optical Detection of the IEEE Communications Society. Manuscript
received December 27, 1991; revised July 7, 1992.
The authors are with the Dipartimento di Ingegneria della Informazione,
Universita di Pisa, Via Diotisalvi 2, 56126 Pisa, Italy.
0090-6??8/94$04.00 0 1994 IEEE
IEEE Log Number 9401037.
which is observed over an interval TO. In (l), s(t) is the
information-bearing signal and w(t) represents complex-valued
additive white Gaussian noise with two-sided power spectral
density 2N0. The signal is known in most of its basic
characteristics (nominal carrier frequency, modulation format,
signalling interval and so on) but a few of its parameters are
unknown. Typically this happens with the carrier phase 8, the
symbol epoch z, the carrier frequency error v and the
1392
IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 42, NO 21314, FEBRUARYMARCWAPRIL 1994
transmitted data. In general one is interested in estimating e, z
and v but, for practical reasons, overlooks the more
challenging goal of jointly estimating 0, z, v and the data [ 191-
[20]. In other words, one concentrates on a subset of {e,z,v}
and looks at the other parameters as unwanted parameters.
To further simplify the discussion in this paper we limit
ourselves to the estimation of a single element of {Q,sv},
generically denoted by A, which we assume to be deterministic
(non-random). All the other parameters, including the data, are
collected in a random vector u having a known probability
density function p( u) which does not depend on A. An exact
representation of the observed waveform r(t) would require
infinite-dimensional vector spaces but, for the time being, we
assume that a finite-dimensional vector r can be found to
represent r(t) with aequate accuracy.
As is known, if A@) is any unbiasedAestimator of A, a lower
bound to the variance of the error A.(r)--A is given by the
Cramer-Rao formula [9, Ch. II]
where E, denotes statistical expectation with respect to the
subscripted variable, and p(rlA) is the probability density
function of r for a given A.
To compute CRB(A) we need p(rlA) which, in principle, can
be obtained from the integral
-
p(r I a) =1 p( r I u, AM^) du, ( 3)
4
where p(rlu,A), the conditional probability density function of
r given u and 1, is easily available, at least for additive
Gaussian channels. Unfortunately, in most cases of practical
interest, the computation of (2) is impossible because either the
integration in (3) cannot be carried out analytically or the
expectation in (2) poses insuperable obstacles.
A way out of this impasse is to resort toAa different bound.
Another lower bound to the variance of $+A (henceforth
referred to as the modified CRB) is the following
1
MCRB(il) =
Er, , {[ a 1 n y AT}
This bound is found observing that
(4)
where the first iiequality derives from application of the CRB
to the estimator A@) for a fixed u , while the second is true in
view of J ensen's inequality [21, p.3891 and the convexity of the
function l l x for DO.
Although (4) has the same structure as (2), ibis much easier
to use. I n fact, for the Gaussian channel
density in (4) is a well-known exponential function whose
argument is a quadratic form in the difference between r and
the signal s. Thus, the logarithm of p(rlu,A) equ
quadratic form and the expectation in (4) is readily de
through a simple example. Assume that r has L components
The difference between CRB and MCRB is stressed
rk =j l +u+wk, k=1, 2; - . , L, (6)
where u is a (scalar) unwanted parameter and wk represents
additive noise. The {wk} are independent and eq
distributed Gaussian random variables with zero mea
standard deviation ow *he parameter u is Gaussian, with zero
mean and standard deviation o,, and is independent of {wk}.
Thus, the probability density functions p(rlu,A) and p ( u )
are:
while p(rlA), as computed through (3), results in
where C is a constant independent of A, Hence, substituting (7)
and (9) into (4) and (2), we obtain
(10)
2
CRB(1) =0," +5
MCRB(A) =s. (1 1)
L
2
L
From (10)-(11) it is seen that MCRB(A) is less than CRB(A)
and one wonders if this is a general rule or just an accidental
case. In Appendix A we show that the inequality MCRB(A) 5
CRB(A) holds true, which means that MCRBs are generally
looser than CRBs. The question then arises whether MCRBs
are tight enough for use in practical applications. This
1393
D'ANDREA et al.: MODIFIED CRAMER-RAO BOUND AND ITS APPLICATION TO SYNCHRONIZATION
important issue will be addressed in Sections IV and V.
111. MCRBS IN THE ESTIMATION OF CARRIER FREQUENCY,
CARRIER PHASE AND TIMING EPOCH
(15)
1
MCRB(2) =
Ew,u[[aln;;a~~)]2i
and substituting (14) into (15) one gets, after some
manipulations,
A. Signal Model and Basic Assumptions
..
The signal s( t ) is modelled as follows
s ( t ) =e x p [ j [ 2 m( t - t o ) +8 ] } ~c i g ( t - i T- z ) , (12)
where v is the offset of the carrier frequency from its nominal
value, 8 is the carrier phase at some reference time +to, z
represents the symbol epoch, T is the symbol spacing, c {ci}
are complex-valued data, and g(t) is the (real-valued)
signalling pulse.
Our goal is to compute the MCRBs for the separate
estimation of v, 8 and z. In particular, in computing MCRB(v),
we consider v as a fixed parameter, while u , 4 ( 8, s~) is a
random vector. Similar assumptions are made in dealing with 8
and z, and we denote by u o L (z,v,c) and u , 4 ( 8, v, c) the
associated unwanted parameter vectors.
The following assumptions are made on the statistics of
8,z,v and c . The timing epoch z in u , and u is uniformly
distributed between 0 and T. The probability densiiy function
of 8 in u, and u r , and of v in u 6 and u , , are assigned but need
not be specified here because they do not affect the final
results. The data symbols {c l } are zero-mean independent
random variables with
i
This is a general result which is now be specialized to the
case il=v and u=u, Using (12) it is found that
with
m( t ) 4 x c i g ( t - i T - z) .
1
As indicated in (16), we need the statistical expectation of
(17) over c , z and 8. However, since (17) is independent of 0,
we may limit ourselves to c and z. To this purpose we note
Ec(lm(t)r} =M2 x g 2 ( t - i T - z ) (19)
i
and that, as a consequence of the Poisson formula, one has
x g 2 ( t - iT - z) =T . exp[j2ni(t - z)/T] (20)
1
for i = k
otherwise
where G2@ is the Fourier transform of g2(t). Thus, substituting
(20) into (19) and averaging with respect to zyields E{ckcT} ={M2 0
(13)
M2 being some positive constant. Finally, the synchronization
parameters appearing in any of the vectors u,, u e and u are
independent of each other and of the data.
It is worth noting that the signal representation (12) fits
conventional QAM and PSK modulations. Extensions to linear
staggered modulations (such as OQPSK and MSK) are
possible but, for space limitations, will not be pursued here.
B. Calculation of MCRB(v)
In computing MCRB( v) we face the conceptual difficulty
that an exact representation of r(f) would require an infinite-
dimensional space, while in the previous section we assumed r
to be finite dimensional. This problem is thoroughly discussed
in [9, ch. IV] where it is shown that in the limit, as the number
of dimensions of r tends to infinity, a formula like (4) does still
apply provided that p(rlu,A) is replaced by the likelihood
function
To denoting the observation interval, and the expectation over r
is replaced by the expectation over the noise process w(t).
and from (17) one gets
The expectation in (22) depends on the position of to within
To. Therefore, since MCRB( v) is inversely proportional to this
expectation, we have different bound values as to varies. Now,
since MCRB( v ) is a lower bound, we are interested in its
maximum value which is achieved by choosing to as the mid-
point of To. With this choice the integral in (22) amounts to
(Ln3/12, where LT is the length of To.
Also, it is easily checked that the quantity M2G2(0) in (21)
is related to the average signal energy per symbol by
Thus (22) becomes
With these changes (4) becomes and from (16) we eventually obtain
1394 EEE "SACTIONS ON COMMUNICATIONS, VOL. 42, NO. U3/4, FEBRUARY/MARCWAPRIL 1994
3T 1
MCRB(V) =
2 7 c 2 ( ~ ~ ) 3 E, I N,
which is the desired result.
To compare MCRB( v) with the performance of some
feedback frequency detectors discussed in the literature, it is
useful to express (25) in terms of the (one-sided) equivalent
noise bandwidth BL=1/(2LT) of a moving-window estimator
operating on LT seconds of observed data (see [15]-[16]).
Letting LT=1/(2Bd in (25) one has
12(BLT)3 1
MCRB(V) =
n2T2 E,lNO'
C. Calculation of MCRB(0) and MCRB(7)
Letting .kQ and u=uoin (16) we have
On the other hand, using (12), we find
and bearing in mind (21) and (23)
Finally, substituting into (16), we obtain
(30)
1 1
MCRB(0) =- -
2L E, /No
or, in terms of the equivalent noise bandwidth BL=ll(2LZ),
B T
MCRB(0) =-.-L-- ,
NQ
which coincides with the result in [16].
final result is [ 15]-[ 161
The MCRB for z is found with similar reasoning and the
BLT T2
4n25 E, I No
MCRB(z) =-- ,
where 4 is an adimensional coefficient depending on the shape
of GO
D. Remark
In Section I1 we pointed out that MCRB(A) 5 CRB(A). A
problem of interest here is to see whether, for any of the
specific cases considered above, the equality MCRB(A)=
CRB(A) may occur. The answer is negative: from the material
in Appendix A it appears that equality takes place only if the
logarithm of the likelihood function A(A,u) in (14) is a linear
function of A or, in other terms, if the integral in (14) is
proportional to A. This however is not the case, as can be seen
by inspection.
IV. FURTHER PROPERTIES OF MCRBS
Since MCRBs are looser than CRBs and no estimator can
provide a lower variance than that established by CRB, one
may argue that the bounds indicated in (26) and (31)-(32)
cannot be reached by any practical synchronizer. This
conclusion however is valid only if the available information
on the signal characteristics is limited just as we have assumed
so far. We shall see shortly that MCRBs can actually be
attained if more information is provided.
To elaborate this point we consider first carrier-phase
estimation. In the preceding di ssion we assumed that data,
timing and carrier-frequency offset were all unknown. Under
these conditions, the bound (3 1) cannot be achieved. Suppose
instead that v, z and, perhaps, the data are available: is the
bound (31) still far from the performance of a practical phase
recoveIy circuit? Note that this is a concrete problem because,
in many modem implementations, frequency offset correction
and timing extraction occur prior to (or simultaneously with)
phase recovery. In addition, phase-recovery is data-aided or
decision-directed so that the transmitted symbols are either
known exactly or are provided by the decision circuit with
good accuracy.
A perfectly analogous point may be raised for the
estimation of the timing epoch: can the bound (32) be attained
by a clock synchronizer if frequency offset, carrier phase and,
perhaps, data symbols are known?
The case of frequency estimation has some peculiarities
which are worth mentioning. If the carrier frequency is
unknown, a reasonable assumption on 1'3 is a uniform
distribution over (0,2n). For the data values and symbol
timing, we may envision two extreme situations :
* Frequency acquisition: v is greater than a significant frac-
Frequency tracking:
tion of the symbol rate UT.
vi s very small compared to UT.
In the first case, data and timing are likely to be unknown
and we are right in the assumptions of Section III. I n the
second case, instead, timing extraction may be carried out in
parallel with frequency tracking so that the frequency tracker
can use timing estimates [22]-[23]. As for the data, th
usually unknown if coherent demodulation is perfo
However, if the signal is PSK modulated and undergoes
differential (rather than coherent) detection, then differential
decisions may be used to help frequency tracking.
In summary, for frequency estimation a practical question
is: can the MCRB be attained if carrier phase is totally
unknown but symbol timing and, perhaps, differential
decisions are available?
We address the foregoing questions in two manners. One
1395
DANDREA et al.: MODIFIED CRAMER-RAO BOUND AND ITS APPLICATION TO SYNCHRONIZATION
consists of looking at the existing literature to see how the best
synchronizers compare with MCRBs. Such an approach is
straightforward and is pursued in the next Section. Its only
limitation is that it is just a picture of the present state of things
and does not provide guidelines for unexplored cases.
As for the other approach, let us denote by standard
assumptions the assumptions on the unwanted parameters we
made in the previous section. We wonder whether a CRB, as
computed under some non-standard assumptions, may equal
the corresponding MCRB as expressed in (26) or (31)-(32)
(and, thus, computed under standard assumptions). The
importance of this issue is that, as long as the CR33is greater
than MCRB, there is no possibility whatsoever for a practical
synchronizer to achieve MCRB. Vice-versa, if they coincide,
there is no logical obstacle to speculation about the existence
of efficient algorithms reaching the MCRB under the specified
non-standard assumptions.
Here we simply summarize the conclusions of our
investigation (see Appendix B for the details). Approximate
equality between CRB and MCRB is found to occur for:
(i ) Estimation of 8 when V, 5 and data are known
(ii) Estimation of z when v, 8, and data are known
(iii) Estimation of v with M-PSK modulation, when z and
differential data are available but 0 is unknown.
Further cases in which CRB equals MCRB do possibly exist
but we could not deal with them because of insurmountable
analytical difficulties. Anyway, if synchronizers approaching
MCRB exist, then CRB and MCRB must be close to each
other.
The following has been taken into account in Appendix B.
The performance of a generic data-aided estimator depends in
general on the particular data pattern being transmitted.
However, if the pattern is random and the observation interval
To is so long that all the data sequences are evenly
represented in To, such a dependence fails. In discussing cases
(i)-(iii) we have assumed that this asymptotic condition is met.
As a final remark we note that, at first appearance, our
being able to compute CRBs for (i)-(iii) seems in contradiction
with the statement made in Section I1that CRBs are usually
difficult to obtain. The explanation is that, under the
assumptions ( i ) or (ii), the signal has no unwanted parameters
whatsoever, so that CRB is readily computed by standard
methods [9]. This is not true with (iii) and in fact, in this case,
we have made approximations which are valid only at high
SNRs.
v. PERFORMANCE OF PRACTICAL SYNCHRONIZERS
A. Carrier Phase and Clock Recovery
As mentioned earlier, a thorough performance evaluation of
the most important algorithms suitable for carrier phase and
clock recovery in digital satellite transmissions is given in
[ 181. From that material it appears that MCRBs for 8 and z are
attained or closely approached by several synchronization
schemes.
FREQUENCY
ERROR
I I
. . .
I
I ! kT+? ek
LOOP
FILTER
vco
--
Fig. 1. Functional block diagram of the NDA loop.
In particular, if timing and carrier frequency offset are
known, MCRB( 0) is reached both by maximum likelihood
(ML) decision-directed (DD) methods, like that proposed in
[19], and by ad hoc non data-aided (NDA) methods discussed
in [22].
MCRB( z) is attained by ML-DD algorithms which,
however, are seldom used in practice as they involve
expensive Derivative Matched Filters (DMFs). An exception is
represented by the DD early-late scheme with sample spacing
Tl2, proposed in [18], which operates closely to the MCRB(z)
for all values of the excess-bandwidth factor 01 without
requiring DMFs. Simple alternative solutions are the DD
scheme by Mueller and Muller [24] and the NDA scheme by
Gardner [25]. They may be seen as approximations to the ML
method but the former requires that 8 be known and has better
performance when a is small, while the latter has carrier-phase
independent behaviour and has superior performance for a
close to unity. For example, at Es/No=10 dB, the Mueller and
Muller algorithm is about 2 dB from MCRB( z) for -0, while
the Gardner algorithm is a fraction of dB from the same limit
for -1.
B. Carrier-Frequency Estimation
In the rest of this section we overview three carrier
frequency estimation methods and we provide a noise
performance analysis for the first two.
I ) Algorithm No. I : The first method is an NDA algorithm
that has been brought out in [26] as an approximation to the
ML carrier frequency estimator when the data are unknown
and the S N R is low. Figure 1 shows a frequency control loop
that includes such a detector.
Input to the system is r(t) (the complex envelope of the
received waveform), whose signal component is affected by a
frequency offset v (see (12)). Compensation for this offset is
accomplished by acting on the VCO frequency. To this end the
error signal
ek =Re{xkY;} (34)
is formed, where x k and Y k are the strobes at t=kT+? of the
outputs of the matched filter G*( f ) and of the frequency-
matched filter dG*(f)/df If theAloop filter is a simple digital
integrator, the VCO frequency vk is governed by a first-order
equation of the type
c k + l = c k - Y e k 7 (35)
1396 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 42, NO 21314, FEBRUARYMARCHIAPRL 1994
ERROR
DETECTOR
I I
LOOP ek
FILTER
vco
Fig. 2. Functional block diagramof theDDD loop.
where yis the step-size.
The noise performance of this scheme is now evaluated
under the assumption that GV) is a root-Nyquist cosine-rolloff
function with excess bandwidth factor a (the overall channel
transfer function, after matched filtering, is thus G(f)G*#) and
has a full-Nyquist cosine rolloff shape). As discussed in [26],
the loop performance depends on the value of the sampling
phase T. Minimum frequency jitter is achieved when z" equals
z, the timing epoch in (12). In practice such a condition can be
met only if v-Gk is small. In the sequel we assume that this is
the case and we take "zz.
The calculation of the variance of v-0, is made by standard
methods [27] and one eventually arrives at
where BL is the loop noise bandwidth. The presence of the
factor a in (36) might lead to the erroneous conclusion that the
frequency jitter vanishes as the signal excess bandwidth
approaches zero. Actually (36) is derived from a linearized
model of the frequency recovery loop which loses its validity
when a approaches zero [28]. Before comparing (36) with
MCRB( V) we consider a second frequency detection scheme.
2) Algorithm No. 2: This is a differential decision-directed
(DDD) algorithm of rotational type [29] for use with PSK
signals. To understand its operation let us consider the tracking
loop in Fig. 2, which coincides with that in Fig. 1 except for
the frequency error generator. We assume that: ( i ) the
frequency errors are small compared to the symbol rate; (ii)
Gcf, is as defined before; (iii) timing is accurately established
and z"=z. Also, we denote by ck=exp(jvk) the transmitted
symbols, with vktaking discrete values between 0 and 2n.
Under these conditions the phase difference between
successive strobes xk and xk-1 is found as the sum of three
terms
A4k =avk +(v- ?k)T+ 6k, (37)
the first of which, Avk4vk-vk-l, is due to modulation, the
second to the residual frequency error VGk, the third to
thermal noise and intersymbol interference. If A vk were
available, we could remove it from (37) and this would yield
(V-$k)T+&, whose average is an estimate of w$k.
In practice this idea is implemented by forming the error
signal
ek =I*{xkX;-, exp(-jA@k)} , (38)
V
= I
I
io-sw T' MCRB( v )
1 o - ~ I I I I 1
10.0 12.5 15.0 17.5 20 0
E,/%, dB
Fig. 3. Normalized variance of estimated frequency jitter for NDA and DDD
loops. The MCRB(v) is also shown for comparison (circles show simulation
results).
where A$k is the estimate of Avk provided by the detection
circuit. In the ideal case of correct decisions A@k equals Ayk,
and the average of ek turns out to be proportional to
ce of vak is very
cumbersome but, assuming correct decisions, the final result is
simply:
As before, the calculation of the v
gv
=--
BLT [2BLT+
n2T2 E, INo
Equations (36) and (39) are drawn in Fig. 3 along with
simulation results, assuming BLT=5.10-3 and QPSK signals
with excess bandwidth factor a =0.5. The discrepancy
between theory and simulations for DDD is due to decision
errors which are not accounted for in deriving (39).
Examination of the figure indicates that the performance of the
considered schemes is quite far from MCRB(z).
3) Algorithm No. 3: This is a feedforward NDA algorithm
suitable for M-PSK modulation. It has been proposed in [23]
and further discussed in [30]. Its basic idea is as follows. Call
Zk the strobe of the matched filter output at t=kT+z and denote
by the argument (taken modulo 274 of the M-th power of zk.
As explained in [23], a k consists of three parts: (i) a piece-wise
increasing quantity 2nMvkT due to the frequency error v; (ii) a
constant term corresponding to the initial carrier phase; (iii) a
zero-mean random term contributed by thermal noise and
intersymbol interference. Estimating v amounts to evaluating
the average slope of ak, and this can be done by fitting (in the
least mean-square sense) the measurements of a k with a linear
1391
DANDREA et al. : MODIFIED CRAMER-RAO BOUND AND ITS APPLICATION TO SYNCHRONIZATION
function of k.
A theoretical analysis of this algorithm appears formidable,
and reference [23] concentrates on a simplified case with
negligible intersymbol interference. In particular, a formula is
given for the estimation error variance, which is valid at high
SNRs. This formula does coincide with MCRB(v) in (25).
Simulation results reported in [30] indicate, however, that the
values of EbINo at which that bound is reached may be quite
large ( Eb is the average energy per information bit). For
example, for QPSK with rectangular pulses and an observation
interval of 15 symbols, the estimation error variance at Eb/No
=10 dB is still 5 dB greater than MCRB(v) for vT=O, and
orders of magnitude greater for vT=5.1W2.
VI. CONCLUSIONS
The concept of MCRB has been introduced and its
relationship to the true CRB has been discussed in the general
context of parameter estimation theory. Applications to
estimation of synchronization parameters have been illustrated
and, in particular, the MCRB for carrier-frequency estimation
has been derived.
In computing MCRBs the assumption has been made that
essentially no information is available on the unwanted
parameters. In these hypotheses the MCRB is looser than the
true CRB. However, in several practical situations the
information is not so meagre; the possibility even exists that
MCRB and CRB are either very close or coincident. Some
cases are found in which this occurs.
Finally, the question has been raised of how far the
performance of practical synchronizers is from the MCRB. In
particular we have concentrated on three frequency detectors
which are believed to represent a good sample of the existing
methods. Two of them are found quite far from MCRB but the
third does attain that limit, even if under some restrictive
conditions. Further research is needed to lessen such
conditions by either improving the estimation method or by
resorting to entirely new techniques.
ACKNOWLEDGEMENTS
The authors wish to thank Dr. C. N. Georghiades and the
anonimous reviewers for their valuable advice, which helped
improve significantly the quality of the present paper.
APPENDIX A
We maintain that CRB(A) 2 MCRB(A). To prove our claim
we show that the expectation in (2) of the text is less than or
equal to the expectation in (4), i.e.,
We start from
where, in writing the second line, we have used the formula
Next, we observe that
p( r I a) =j.r I u, n)p(u) du . (44.4)
Assuming that p( u) is independent of il and differentiating
(A.4) with respect to A we get
-Qo
or, bearing in mind (A.3),
Application of the Schwarz inequality to the second integral in
(A.6) yields
(A.7)
and substituting this result into the last line of (A.2) we
eventually obtain (A.1).
Notice that the equality holds in (A.7) (and therefore in
(A. 1)) if and only if
a Inp( r I IC, A)
an
=constant.
It should be stressed that the above conclusion is valid only as
long as u is a true random vector, with a regular probability
density function. For example, if u were a known deterministic
quantity instead, p( u) would be a delta. Then, from (A.4) one
would have p(rlA)=p(rlu,A) and the equality would apply in
(A. 1) anyway.
APPENDIX B
A. CRB for Case (i) of Section IV
Denote by CRB(8 I ue) the Cramer-Rao bound for 8 when
the vector ue&(sv,c) is given. This bound is readily computed
[9, p. 2751 since, if ue is specified, the signal is perfectly
known:
Using (12) in the text and assuming that To comprises many
signalling intervals (L >>l), it is found that
1398 IEEE TRANSACTf4 3NS ON COMMUNICATIONS, VOL. 42, NO. 21314, FEBRUARYMARCHIAPRIL 1994
where
oa
(B.3)
A
Y1- k =j&t)g[t f (i - k)T1 dt 9
and no is the integer such that noT is nearest to the beginning of
the observation interval.
Next we re-write (B.2) in the form
For L sufficiently large, the quantity in square brackets
approximates the symbol autoconelation function which is M2
if m=O and is zero otherwise (see(l3)). Thus, bearing in mind
that
2
Yo =j- IG(f >I df =G2(0), (B.5)
.-cc
from (B.l), (B.4) we get
which, in view of (23), coincides with MCRB(6) in (31).
B. CRB for Case (ii) of Section IV
The CRB for z when u, A (6,v, c ) is given can be obtained
by similar arguments and coincides with the MCRB(z) in (32).
C. CRB for Case (iii) of Section IV
In this case the analysis is much longer and we illustrate its
major steps only. We start with the mathematical model of the
signal under the assumptions (iii). The modulation is M-PSK
and the symbols in (12) have the form ci =p exp(jaJ , where p
is a constant and a,belongs to the alphabet S A {0, 2dM, ...,
27G(M-l)lM}. The differences {ai -a,l} are known and,
defining no as in (B.2), it is easily seen that
a, =a,, +4, for i 2 no 03-71
with
k=nO +1
Therefore, letting Z, A pexp(j4J and using (B.7) in (12), we get
for te To
8 +a,, +2m(t - tO)]}c Ci g( t - iT - z) . (B.9)
1
Notice that the only unknown quantities here are V, 8 and am.
The first is a constant, while 8 and a,, are random variables,
uniformly distributed over (0,27c) and S , respectively.
Next we consider the likelihood A(v,O,a,,) of s(t) for fixed
values of v, Band 4,. Bearing in mind (14) and (B.9) we find,
after some passages,
A(v,8,ano)=Cexp KO -c0s(6+an0 -y)], (B.lO)
where C is a constant independent of (v,6,%,) and A and yare
such that
Aexp(jy)= j r(t)o*(t)dt (B.ll)
TO
with
o(t) A exp[j2m(t - tO)]cC, g(t - iT - z). (B.12)
We need the likelihood A(v), which is obtained by
I
averaging (B. 10) with respect to 0 and a,,. This yields
(B.13)
where I&) is the modified Bessel function of order zero. For
x>>l one has [31, p. 3771
en
I, z-
G'
(B.14)
so that, assuming high SNRs (hence, p/No>>l), from (B.13)-
(B. 14) we get
alnA(v) I 1 aA2
-
J v 2ANo J v .
(B.15)
This equation is useful to compute CRB(v Iz,Z)), which is
the CRB for v once timing and differential data are given:
-. . (B.16)
where Ew{.} means expectation over the noise. Such an
expectation, however, is lengthy to compute and we only give
the final result
CRB( v I 7, c") =- No ro (B.17)
4n2 r0r2 -r;
with
(B.18)
rk =A jTo fk]l%.(t)r dt
k(t) A zElg( t - iT - z) . (B.19)
Looking at (B.17)-(B.19) it is seen that CRB(vlz, c")
depends on the actual symbol pattern. However it can be
shown that, when the length of TO grows large, the bound tends
to a limit which is obtained by replacing the quantities rk in
(B.17) by their averages with respect to Fl and z, the latter
taken uniformly distributed over a symbol interval. Such
averages are computed bearing in mind that, since the c, in
(12) are independent and equiprobable, so are the;, in (B.9)
and therefore (see (21))
I
DANDREA ef al.: MODIFIED CRAMER-RAO BOUBD AND ITS APPLICATION TO SYNCHRONIZATION
1399
[23] S. Bellini, C. Molinari and G. Tartara, Digital Frequency Estimation in
Burst ModeQPSK Transmission, IEEE Trans. Commun , vol. COM-38,
[24] K. H. Mueller and M. Muller, Timing Recovery in Digital Synchronous
Data Receivers, IEEE Trans. Commun., vol. COM-24, pp. 516-530,
May 1976.
[25] F. M. Garduer, A BPSWQPSK Timing &or Detector for Sampled
Receivers, IEEE Trans. Commun., vol. COM-34, pp. 423-429, May
(B.20) pp. 959-961, July 1990.
M2G2 (0)
Ec , r ( l wl l j = .
Carrying out the indicated calculations one finds
(B.21)
3T 1
CRB(V I T,t) =
w ~ ( L T ) ~ EJN, 1986.
which coincides with ( 25) in the text. In summary, for theCase
(iii) of Section IV, the CRB tends to MCRB at high SNRs
when the observation interval grows large.
REFERENCES
[I] A. J . Viterbi, Principles of coherent Communication. New York:
[2] J. J. Stiffler, Theory of Synchronous Communicarions. Englewood Cliffs,
[3] W. C. Lindsey, Synchronization Systems in Communication and Control.
[4] W. C. Lindsey and M. K . Simon, Telecommunicution Systems
[5] F. M. Gardner, Phaselock Techniques, 2nd ed. New York Wiley, 1979.
[6] H. Meyr and G. Ascheid, Synchronization in Digital Communications.
[7] L. E. Franks. Carrier and Bit Svnchronization in Data Communication -
McGraw-Hill, 1966.
NJ: Prentice Hall, 1971.
Englewood Cliffs, NJ: Prentice Hall, 1972.
Engineering. Englewood Cliffs, NJ: Prentice Hall, 1973.
New York Wiley, 1990.
A Tutorial Review, IEEE Trans. Commun., vol. COM-28, pp. 1107-
1121, Aug. 1980.
F. M. Gardner, Demodulator Reference Recovery Techniques Suited for
Digital Implementation, ESA Final Report, ESTEC Contract No 6847-
H. L. Van Trees, Detection, Estimation and Modulation Theory. New
York Wiley, 1968.
H. W. Sorenson, Parameter Estimation. New York: Marcel Dekker,
1980.
L. P. Seidman, Performance Limitations and Error Calculations for
Parameter Estimation, Proc. IEEE, vol 58, pp 644652, May 1970.
J . Ziv and M. Zakai, Some Lower Bounds on Signal Parameter
Estimation, IEEE Trans Inform. Theory, vol. IT-15, pp.386-391, May
1969.
S Bellini and G. Tartara, Bounds on Error in Signal Parameter
Estimation, IEEE Trans. Commun., vol. COM-22, pp. 340-342, March
1974.
S. C. White and N. C. Beaulieu, On the Application of the Cramer-Rao
and Detection Theory Bounds to Mean Square Error of Symbol Timing
Recovery, IEEE Trans. Commm., vol. COM-40, p.p. 16351643, Oct.
1992.
M. Moeneclaey, A Simple Lower Bound on the Linearized Performance
of Practical Symbol Synchronizers, IEEE Trans. Commun., vol. COM-
31, pp.1029-1032, Sept. 1983.
M. Moeneclaey, A Fundamental Lower Bound to the Performance of
Practical J oint Carrier and Bit Synchronizers, IEEE Trans. Commun.,
vol. COM-32, pp. 1007-1012, Sept. 1984.
M. Moeneclaey and I. Bruyland, The J oint Carrier and Symbol
Synchronizability of Continuous Phase Modulated Waveforms, Con$
Rec. ICC86, vol. 2, paper 31.5.
T. J esupret, M. Moeneclaey and G. Ascheid, Digital Demodulator
Synchronization, ESA Draft Final Report, ESTEC Contract No 8437-
89-NL-RE, Feb. 1991.
H. Kobayashi, Simultaneous Adaptive Estimation and Decision
Algorithmfor Carrier Modulated Data Transmission Systems, IEEE
Trans. Commun., vol. COM-19, pp. 268-280, June 1971.
D. D. Falconer and J . Salz, Optimal Reception of Digital Data Over the
Gaussian Channel with Unknown Delay and Phase J itter, IEEE Tram.
Inform. Theory, vol. IT-23, pp. 117-126, Jan. 1977.
C. W. Burrill, Measure, Integration and Probabiliq. New York:
McGraw-Hill, 1972.
A. J. Viterbi and A. M. Viterbi, Nonlinear Estimation of PSK-
Modulated Carrier Phase with Application to Burst Digital
Transmission, IEEE Trans. Inform. Theory, vol. IT-29, pp. 543-55 I ,
July 1983.
86-NL-DG, AUg. 1988.
1261 F. M. Gardner, Frequency Detectors for Digital Demodulators Via
Maximum-Likelihood Derivation, ESA Final Report: Part 11, ESTEC
Contract No 8022-88-NL-DG, March 1990.
[27] U. Mengali, Joint Phase and Timing Acquisition in Data-Transmission,
IEEE Trans. Commun., vol. COM-25, pp. 1174-1 185, Oct. 1977.
[28] N. A. DAndreaand U. Mengali, Noise Performance of Two Frequency-
Error Detectors Derived from Maximum Likelihood Estimation
Methods, to appear in IEEE Trans. Commun
[29] D. G. Messerschmitt, Frequency Detectors for PLL Acquisition Timing
and Canier Recovery, IEEE Trans. Commun., vol. COM-27,pp. 1288-
1295, Sept. 1979.
[30] S. Bellini, C. Molinari and G. Tartara, Digital Carrier Recovery with
Frequency Offset in TDMA Transmission, Con$ Record ICC91, vol.
11, paper 25.6.
[31] M. Abramowitz and I. A. Stegun, Handbook of Muthemtical Functions.
New York: Dover, 1970.
Aldo N. DAndrea (M82-SM91) received the Dr. Ing. degree in Electronic
Engineering from the University of Pisa, Italy, in 1977.
From 1977 to 1981 he was a Research fellow engaged in research on
digital phase-locked loops at the Centro Studi per i Metodi e i Dispositivi di
Radiotrasmissione of the Consiglio Nazionale delle Ricerche (CNR). Since
1978 he has been involved in the development of the Italian Air Traffic
Control Program (ATC). Currently, he is an Associate Professor of
Communication Networks at the Dipartimento di Ingegneria della
Informazione, Universith di Pisa. His interests include the design and analysis
of digital communication systems, signal processing and synchronization.
Umberto Mengali (M69-SM85-F90) received the Dr. Ing. Degree in
Electrical Engineering from the University of Pisa in 1961 and the Libera
Docenza in Telecommunications fromthe Italian Education Ministry in 1971.
Since 1963 he has been with the Department of Information Engineering
of the University of Pisa where heis a Professor of Telecommunications. His
research interests are in digital communication theory, with emphasis on
synchronization methods and modulation techniques.
Professor Mengali is a member of the IEEE Communication Theory
Committee and a former Editor of the IEEE Transactions on Communications
(1985-1991). He is a Fellow of IEEE and is listed in American Men and
Women in Science.
Ruggero Reggiannini received the Dr. Ing. degree in Electronic Engineering
fromthe University of Pisa, Italy, in 1978.
From1978 to 1983 he was with USEA S.p.A., where he was engaged in
the design and development of underwater acoustic systems. Since 1984 he
has been with the Department of Information Engineering of the University of
Pisa, where heis currently Associate Professor of Radio Communications. His
research interests are in the field of digital satellite and mobile communication
systems.