You are on page 1of 14

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 52, NO.

10, OCTOBER 2004 1733

Envelope-Aided Viterbi Receivers for GMSK Signals


With Limiter-Discriminator Detection
Ramón Sánchez-Pérez, Subbarayan Pasupathy, Fellow, IEEE, and Francisco Javier Casajús-Quirós, Member, IEEE

Abstract—Continuous phase modulation schemes, such as They are especially suited for fading or mobile channels, where
Gaussian minimum-shift keying (GMSK), are frequently used frequency and phase synchronism complicates coherent demod-
with limiter-discriminator (LD) detectors. This paper studies how ulation [7]–[9]. LD-based receivers have been widely studied
the side information derived from the signal envelope can enhance
the performance of a Viterbi algorithm (VA)-based receiver op- and employed for the last 50 years for analog frequency modu-
erating on the LD output of a GMSK scheme. By considering the lation (FM) demodulation purposes.
joint probability density function of envelope and frequency, dif- In discriminator-based systems, many postprocessing
ferent approximations yield different novel metrics for VA, using schemes have been studied, such as integrate and dump, linear
the three-variables envelope and its derivative, and frequency
error in different combinations. Simulation results confirm that equalizers, feedback equalizers, and others (see, for example,
such envelope-aided VA gives significant performance gains, and [8]), but only a few of them have coped with the problem
that envelope information complements the frequency information of postprocessing in the highly frequency-selective channel
output by the LD detector in frequency-selective fading channels. associated with high-data-rate time-division multiple-access
Index Terms—Envelope information, Gaussian minimum-shift (TDMA) systems [10]–[12]. The modulator, linear channel,
keying (GMSK), limiter discriminator (LD), maximum-likelihood and LD chain presents a strongly nonlinear behavior which, in
sequence detector (MLSD), multipath channel. the presence of a dispersive channel, can produce high signal
distortion at the output of the discriminator and yield large
I. INTRODUCTION error rates. Due to this nonlinear behavior [13], conventional
equalizers are not very effective. In contrast, nonlinear detector

C ONTINUOUS phase modulation (CPM) signaling


schemes. such as Gaussian minimum-shift keying
(GMSK), have gained great interest from a theoretical and
systems, such as the maximum-likelihood sequence detector
(MLSD) are suitable approaches [10]. In this paper, it is shown
that due to the particular distribution of the noise at the output
practical point of view in the last years, due to their attractive
of the LD, the performance of a MLSD scheme can be sig-
spectral properties (bandwidth efficiency and low out-of-band
nificantly improved with a minimum complexity increase, by
radiation) and their constant envelope [1]. The first property
means of the joint use of the LD output and the instantaneous
is important due to the intensive use of the limited spectral
envelope at the output of an envelope detector. Previous works
resource. The property of constant envelope allows the use of
by Asano and Pasupathy [14] and Rohani [15] have used
power-efficient nonlinear amplifiers.
envelope information to improve the performance of receivers
GMSK is being employed in global systems for mobile com-
based on LD detectors in fading channel environments. In a
munications (GSM) [2], digital enhanced cordless telecommu-
recent paper by Pawula [16], the squared envelope of the signal
nications (DECT) [3], and HIPERLAN/1 [4] standards at the
has been employed at the Viterbi decoder input of a coded
moment. Ideal coherent detection of these schemes has been
narrowband digital FM receiver.
thoroughly studied, and it has been shown that, for an addi-
tive white Gaussian noise (AWGN) channel, it provides high In order to describe and evaluate the proposed techniques, a
bandwidth and power efficiencies [5], [6]. However, schemes MLSD receiver will be considered. MLSD is highly descrip-
with good performance can also be implemented using non- tive and well known, and an efficient algorithm, the Viterbi al-
coherent detectors. Limiter-discriminator (LD)-based receivers gorithm (VA) [17], [18], is available for implementation. Some
have proved to be simple, cheap, and robust options for CPM. suboptimal simplifications of the VA have also been reported
[19]. In addition, an extensive use of MLSD has been done to
deal with both the inherent intersymbol interference (ISI) of par-
Paper approved by G. Cherubini, the Editor for CDMA Systems of the IEEE
Communications Society. Manuscript received March 27, 2000; revised August
tial-response modulation, and the ISI introduced by band limita-
8, 2000 and July 28, 2002. This work was supported in part by Spanish National tion at the intermediate frequency (IF) filter [20]. However, the
Projects TIC2003-09061-C03-01 and TIC2000-1395-C02-02, and in part by a same concepts can be applied to any other detection strategy
grant from the Researchers Training Program of the regional government of (e.g., symbol-by-symbol decision).
Madrid, Spain.
R. Sánchez-Pérez is with Agere Systems, 28033 Madrid, Spain (e-mail: To implement the VA, from a practical point of view, several
ramon.sanchez@agere.com; ramon@gaps.ssr.upm.es). parameters must be considered, such as the number of trellis
S. Pasupathy is with the Department of Electrical and Computer Engi-
neering, University of Toronto, Toronto, ON M5S 3G4, Canada (e-mail: states, decision depth, and others. In addition, the reference es-
pas@comm.utoronto.ca). timation procedure and the incremental metric definition must
F. J. Casajús-Quirós is with the Departamento de Señales Sistemas y Radio- be established. However, this paper deals with the problem of
comunicaciones, ETSI Telecomunicación, Universidad Politécnica de Madrid,
28040 Madrid, Spain (e-mail: javier@gaps.ssr.upm.es). the incremental metric definition and, in what follows, the refer-
Digital Object Identifier 10.1109/TCOMM.2004.836510 ences will be considered known and the number of trellis states
0090-6778/04$20.00 © 2004 IEEE
1734 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 52, NO. 10, OCTOBER 2004

Fig. 1. TX/RX based on LD.

and decision depth long enough to prevent the introduction of


inaccuracies.
Whenever the noise involved in the decision space in which
the VA is applied is Gaussian and identically distributed for any
sequence, the squared error is the optimal metric in the ML
sense. In our system, this assumption does not hold. It would be
a good approximation for Gaussian noise at the input of the re-
ceiver, constant envelope and high signal-to-noise ratio (SNR),
but it is not reasonable with a narrow IF filter and frequency-se-
lective channel. Simon [21] first pointed out the impact of the
mismatch between the Gaussian squared metric and the actual
Fig. 2. Phasor diagram representation of signal and noise components.
probability density function (pdf) on FM receivers with LD de- The (X; Y ) and (X ; Y ) frames are, respectively, related to the unmodulated
tection. It is intuitive that when multipath transmission produces and modulated signal phase.
a spread of the envelope range, the pdf of the phase and fre-
quency will spread as well. A careful analysis of the noise pdf The received baseband waveform , at the output of the IF
will provide an adequate metric for the VA. filter, is a signal distorted by the channel and the IF filter plus
This paper is organized as follows. In Section II, the signal Gaussian noise
model is introduced, following Rice’s [22] classical descrip-
tion. Section III discusses the effects of low envelope in the fre- (1)
quency noise statistics. Section IV presents the frequency error
pdf and its approximations. Section V proposes and describes where is the signal envelope and is the modulated
a quasi-Gaussian set of decision variables based on frequency, signal phase. The IF filter is assumed to be symmetrical and
envelope, and envelope derivative. A different approach, using centered at the carrier frequency. Therefore, noise can be repre-
a “pseudopolar” model of the frequency and envelope errors, sented as a function of the zero-mean Gaussian quadrature com-
is considered in Section VI. Section VII presents the details of ponents and in the unique form
the computer simulations that have been carried out, and Sec- (2)
tion VIII some numerical results. Finally, the conclusions and
a discussion about the importance of envelope information is The received waveform can also be written as
found in Section IX.

(3)
where the noise components and are defined relative
II. SIGNAL MODEL to the modulated phasor and related to and by

Fig. 1 shows a diagram of the typical baseband transmis- (4)


sion/reception (TX/RX) system based on a LD device. The filter (5)
with impulse response determines the modulation family.
The frequency modulation is easily performed using a voltage- A more useful representation for in terms of envelope
controlled oscillator (VCO). The signal travels from the TX and phase is
to the RX via the impulse response channel, and AWGN (6)
is added at the input of the RX. The baseband model of the
receiver includes a narrowband IF filter, whose effect will be where and are the envelope and phase error in the
often considered as part of the channel filter. The limiter re- presence of noise and are given by
moves the envelope information and the discriminator extracts
the instantaneous frequency. The post-processing block repre- (7)
sents any means of detection of the transmitted data, from a (8)
simple threshold to a more complex block, such as an equal-
izer or a Viterbi detector, as in our case.
Thus, the frequency error associated with the phase noise is
Rice’s model [22] of the signal and noise involved in LD de-
tection is well described through the phasor diagram at the input
(9)
of the LD of an FM receiver, as shown in Fig. 2.
SÁNCHEZ-PÉREZ et al.: ENVELOPE-AIDED VITERBI RECEIVERS FOR GMSK SIGNALS WITH LIMITER-DISCRIMINATOR DETECTION 1735

where the derivatives of and are given by If a low-pass filter is included at the output of the LD, the
energy of a spike will spread out into adjacent bit intervals, thus
(10) introducing ISI and memory effect. However, in our analysis, it
(11) will be considered that such a filter does not exist.
Just as the click detector was based on envelope detection, in
Properties of these processes are discussed in [23]. When the following sections, we will conclude that envelope detection
and are independent, their derivatives and are also can be the key concept needed to implement a much more effi-
independent. Also, and , as well as and , are cient and robust LD post-processor and detector.
independent when evaluated at the same instant of time. Inde-
pendence of and and of and follows from IV. FREQUENCY ERROR PDF
the symmetrical spectral density of the noise. Their joint pdf can
therefore be written as the product of four independent Gaussian It is worth indicating that in the context of MLSD, the con-
functions. sidered probabilities are conditioned to the transmitted signal
Due to the GMSK modulation, the narrow IF filter, and the candidates and are usually noted as . With the pur-
frequency-selective multipath channel effect (in what follows, it pose of alleviating the notation, we will not explicitly indicate
is useful to consider the effect of the narrow IF filter included in that dependency and will not include the function suscript, so
the frequency-selective channel), the instantaneous frequency, that the previous pdf is simply expressed as .
once sampled at the output of the LD, depends on a number In order to obtain the optimal distance definition to be used
of consecutive transmitted symbols. This is the basic as the differential metric in the VA for MLSD, the first step is
principle of MLSDs. Because of the channel, the received fre- to find the pdf associated with the frequency error at the output
quency , even under zero-noise conditions , is different of the discriminator, . The general probability distribution
from the frequency of the transmitted sequence . The noise- expression of the frequency at the output of the discriminator,
less instantaneous frequency at discrete time can be expressed as described in (6), is shown in [26] to be
as a function of the transmitted data
(14)
(12)
where is the well-known Marcum -function, with
where is the instantaneous phase.
Under multipath channel conditions, the envelope of the re- (15)
ceived signal is not constant, and its behavior will prove to be
critical. In fact, the non-constant value of the envelope will de-
pend on the transmitted sequence in a similar way. The equiva-
lent form of (12) for the envelope is

(13) (16)
and its reference values for the VA can be estimated with one In the case of a choice of signs, the upper sign refers to the upper
of the same techniques used to estimate for every possible quantity on the left, and the lower sign to the lower quantity.
transmitted symbol sequence in (12). Also

III. LOW-ENVELOPE EFFECTS IN FM RECEIVERS (17)


Noise in FM receivers has been widely studied, due to the
practical interest in analog and digital FM receivers in the past (18)
decades. However, most of these studies were done in a dif-
ferent context, in the absence of a frequency-selective multipath (19)
channel. Papers by Rice [22], Mazo and Salz [24], and more
recently [25], deal with the problem of reception in AWGN, with the power spectral density (PSD) of the complex
but without distortion. For full response signaling and with no noise, . Moreover
distortion, the most popular post-detection scheme was the in-
tegrate-and-dump device followed by a slicer. In these studies, (20)
together with Gaussian-modeled phase noise, rarer events
known as clicks were considered as major error-generating is the mean noise power. For a symmetrical noise-power spec-
mechanisms. These clicks are produced by signal segments that trum at the discriminator input, .
encircle the coordinates origin (zero envelope), thus producing The general expression (14) is intractable from a practical
a phase spike of area . When a path comes close to the origin, point of view. A simpler approximation is required. Let us sup-
even if it does not encircle it, the instantaneous frequency also pose now that and that the amplitude and phase vari-
displays a sharp spike, due to the rapid phase variation. This ations have small effect on the pdf, i.e., the frequency error is
led to the use of envelope detectors for detection and correction similar for our signals and for an unmodulated carrier. Evalua-
of clicks. tion of (14) has shown that these simplifications do not imply a
1736 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 52, NO. 10, OCTOBER 2004

that the noise-variance term results in a multiplicative constant


in the expression of the metric and can be removed. This is,
together with its simplicity, the most remarkable advantage of
the metrics based on a Gaussian pdf.

V. ORTHOGONAL NOISE COMPONENTS AT THE INPUT OF


LIMITER: MULTIVARIABLE METRIC
Equation (25) and Fig. 3 in the previous section suggest that
envelope information will be required to improve the perfor-
mance of the too-simple Euclidean distance of the frequency
variable. However, if envelope information is going to be
considered, it is mandatory to investigate the “optimum” use
of such information. That means using frequency and envelope
information together without a priori restrictions to find a
good metric: one that matches the noise pdf, is independent
of the noise power, and easy to calculate. In general, only a
distance derived from Gaussian pdf will have all these desirable
Fig. 3. PDF f (_) (21), evaluated for different signal envelope values A, and
0
 = 25 dB. A Gaussian pdf with  =A power has been included (dotted
properties.
line). The orthogonal noise components derived in (4), (5), (10),
and (11) are Gaussian and mutually independent, with joint pdf

critical degradation. Then, with the constraints and


, the pdf (26)

So, a convenient metric to feed the VA would be

(21) (27)

can be derived [16], [27] for the frequency error, with


or, equivalently
(22)
(28)
(23)
which is, of course, independent of the noise power.
(24) Unfortunately, in a noncoherent receiver, these noise com-
ponents are not directly observable. Let us try to approximate
and and are the modified Bessel functions of the first them. From (7) and (8), the alternative variables
kind of order 0 and 1, respectively.
The optimal MLSD metric associated with this pdf (i.e., (29)
, where and are arbitrary constant values),
would have, even with the simplifications assumed up to here, (30)
a complicated expression. In Fig. 3, the pdf in (21) is evaluated
for different signal-envelope values. A Gaussian pdf with
power has been included (dotted line), with results almost iden- (31)
tical for high values of but inaccurate for small ones. The
figure also shows that, as expected, the variance increases when (32)
the instantaneous envelope decreases. This suggests that the fre-
quency-error variance grows with the inverse square of the en- can be obtained, and after some simplifications, result in
velope. But even if this Gaussian approximation is accepted, the
resultant metric (or distance)
(33)
(25)

would have different values for different instantaneous en- (34)


velopes, and its calculation would require previous knowledge
of the instantaneous envelope for each different transmitted and together with the module of the Jacobian of the inverse
sequence. The well-known advantage of using a Gaussian pdf is transformation, , provides the joint pdf
SÁNCHEZ-PÉREZ et al.: ENVELOPE-AIDED VITERBI RECEIVERS FOR GMSK SIGNALS WITH LIMITER-DISCRIMINATOR DETECTION 1737

as shown in (35), at the bottom of the page. For a noncoherent As a consequence, the metric
receiver the phase is unknown. The marginal pdf is calculated
by integration of (35) with respect to , i.e., (44)

can be written as
(36)

(45)

(37)
In the MLSD criterion, the received value of will be used in
The exponent inside the integral can be expressed as (45) in order to evaluate the metric for several values of which
correspond to all the different signal candidates. The first term
in (45) will be constant for every , therefore it has no effect
when comparing the different values of the metric. In addition,
(38)
the distance can be multiplied by to obtain the equivalent
distance
where the value of is irrelevant, because

(46)

If is also approximated by

(39)

The argument of the Bessel function, can be approximated as (47)


follows:
can be written. If the last term is ignored (as it is reasonable
to suppose, since for signals with normalized bandwidth lower
(40) than 1, ) the subsequent metric is

(48)
(41)

which is very simple. It consists of three distinct terms. The first


and second terms contain the effect of the envelope and the
(42) envelope derivative , respectively. The last term corresponds
to the frequency error weighted by the signal envelope. It sug-
gests that, when the envelope is small, a higher frequency error
is more likely, and must be somehow weighted to show the right
(43) effect on the overall distance. Fig. 4 shows the building blocks
of a receiver that implements the metric in (48).
However, since the three terms are combined by addition,
For , , where is close to 1 ( they can be seen as three independent metrics that correspond to
). different dimensions of the same communications system, each

(35)
1738 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 52, NO. 10, OCTOBER 2004

Fig. 4. Tridimensional decision space receiver implementing the metric (48).

one having its own properties and contribution to the complete


equation. In the sequel to this section, further analysis of the
terms in (48) will be carried out. The performance provided by
the use of each contribution considered as a metric in itself and
combinations of some of them is available in Section VIII.

A. Envelope Weighted Frequency-Error Metric Component


The statistics of the envelope of the weighted frequency-error Fig. 5. f (R) and its Gaussian approximation f (R) for different values of
term in (48) can be easily derived. The frequency error and envelope A. SNR = 25.
envelope joint pdf can be derived by integration of [27, eq. (5.1)]
which is Gaussian! And, hence, the metric

(56)
(49)
is optimal for its variable , and so it is its contribution to (48).
integrated with respect to
B. Envelope Error-Metric Component
The envelope error term could be derived from (33)
after removing the effect of the phase error (i.e., ). If is
zero, as can be observed in Fig. 2, the envelope noise is exactly
(50) equal to the component of the orthogonal Gaussian noise.
where In Fig. 5, the actual pdf (53) is compared with the Gaussian
pdf assumed by the envelope error metric in (48). As expected,
(51) the approximation is good for high envelope (also high SNR),
and it is biased toward , which is always considered more
likely than it really is. It is worth remembering that a receiver
and results in
based only on this “envelope metric component” is not of prac-
tical interest, since for a nondispersive channel, the envelope of
the received signal is constant. Only when frequency-selective
(52) fading is present (also when a narrow IF filter is present), the
and together with [22] envelope values are spread and carry useful information. There-
fore, simulations based on this metric component alone have not
(53) been carried out. However, an example of how the envelope in-
formation is a useful complement in a frequency-detector device
results in is shown in Section IX.

C. Envelope Derivative Error-Metric Component


According to the simulations described in Section VII, the
(54) effect of the envelope derivative component is marginal. The
(where indicates a Gaussian pdf with mean and vari- difference between the performance results provided by the
ance ) that could be employed as a metric when is also known metric in (48) and a simplified version that does not include
in a more efficient way than through the unconditioned pdf . the term is about 1 dB of SNR (as shown later in
Going a step further, if the new decision variable Section VIII).
used in (48) is considered, and the associated pdf calculated, it
results in D. Joint Envelope-Frequency Error Metric
Up to now, the envelope and weighted frequency components
(55)
of (48) have been independently derived. According to (55),
SÁNCHEZ-PÉREZ et al.: ENVELOPE-AIDED VITERBI RECEIVERS FOR GMSK SIGNALS WITH LIMITER-DISCRIMINATOR DETECTION 1739

which is also close to Gaussian. In fact, a Gaussian pdf results if


approaches one (no amplitude error, again valid under
low noise conditions, see Fig. 5). The joint pdf (59) is, from (53)
and (55), equal to

(62)

and its approximation is

(63)

with the associated metric

(64)

(65)

Fig. 6. PDF f (;_ R) (52), for  = 0 20 dB and different instantaneous


(66)
envelope references. A 20-units offset has been added in order to show their
contour.
(67)
both components are independent, and the best way to combine (68)
them in a metric is simply by adding them. However, it is inter-
As a result, the same envelope and frequency metric compo-
esting to take a look at their joint distribution.
nents in (48) have been derived from their joint pdf. The prop-
To begin with, the straightforward metric for the VA making
erties of this metric (68) are remarkable: simple and noise-inde-
use of both the envelope and frequency information could have
pendent. This metric corresponds to the approximation of (62)
been derived directly from their joint pdf in (52), plotted in
with a bidimensional Gaussian pdf. Fig. 7 shows the contour
Fig. 6. Pdfs for three different instantaneous reference envelopes
plot of the logarithm of the actual pdf in (62) and its Gaussian
are shown in the same figure, to be compared. A 20-units offset
approximation, given by
has been added in order to show their contour. The differences
in shape and variance are clear, and they are far from being
(69)
Gaussian. Therefore, the classical Euclidean distance is not ad-
equate for our purposes.
as is implicit in (68).
Equation (52) can be approximated using
Let’s summarize. To obtain the metric (68), and the pdf from
for to
which is derived, the following simplifications have been made:
• envelope variation, ;
(57)
• central frequency is zero, (unmodulated signal);
• , so that
and an associated metric can be derived, for example
in (62);
• in (63), valid for high instantaneous SNR
(58) ;
• , equivalent baseband PSD of the noise after IF filter
that would depend, as expected, on the SNR is symmetrical.
However, if the decision variable is considered The first two items imply that the amplitude and frequency
instead of , (52) becomes error of a quite general modulated signal is approximated by
that of an unmodulated one. The two last items are good ap-
(59) proximations for high SNR. None of them is very realistic for
channels for which, although with a medium or high mean SNR,
or its approximation
the ISI produce very small instantaneous amplitude and strong
(60) variations, for some transmitted sequences. However, those sim-
plifications are necessary to obtain analytical expressions of
where the first term [from (55)] is Gaussian and independent practical interest. In addition, even for those cases where the
of , and the second one can be approximated from (53) using approximations are less realistic, the results obtained using the
for (low noise conditions), by proposed metrics are much better than those provided by the
conventional Euclidean distance. A detailed analysis of the ef-
fect of the inaccuracies should provide deeper insight, but is out
(61)
of the scope of this paper.
1740 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 52, NO. 10, OCTOBER 2004

Fig. 7. Contour plot of the logarithm of the actual pdf (62), for  = 025 dB. Contour plot of the Gaussian pdf, assumed in the approximated metric (68), is
plotted in dotted line for reference.

Fig. 8 depicts the sampled output of the discriminator are obtained. is again a shaping factor that balances the effect
and the envelope detector. Each instance corresponds to of amplitude and frequency. Finally, the new metric
an envelope/frequency pair for a sequence of symbols of
a GMSK signal with bandwidth-symbol duration product (74)
, modulation index , channel time response is derived.
, and SNR dB. From (52), the joint pdf of and is
Crosses correspond to the received points with multipath distor-
tion. Spots correspond to the actual signal with both distortion
and noise (SNR dB). The simulation considers that each
sample depends on four transmitted symbols, but the actual
channel could provide higher values for very narrow IF filters (75)
and larger multipath delays, . It clearly shows that clusters of A more intuitive expression can be obtained in terms of envelope
points with little envelope have a higher dispersion. In Fig. 8(b), and frequency
the new decision variables and (68) have been obtained
for the same signal. The clusters have a symmetric shape, as
expected, and look much more separable than in Fig. 8(a).
(76)
where and are those in (72) and (73). The same procedure
VI. PSEUDOPOLAR APPROACH can be applied to (57).
An alternative approach takes into account the similarity be- An example of this distance criterion is shown in Fig. 9 for the
tween the statistical behavior of the phase error and the fre- same transmission of Fig. 8. The distance between the received
quency error, and directly employs frequency error instead of samples of signal in the pseudopolar space (dots) and the refer-
phase error in a “pseudopolar” variable space. ences (crosses) corresponds to those expressed in (74).
That is to say, the orthogonal variables and involved in The distance criterion in this decision space (74) is very
the Euclidean metric (33) can be expressed in terms of the polar close to (68). In fact, for low frequency error, using the first and
coordinates and as second terms of the Taylor series expansion for , the
distance (74) can be approximated by
(70) (77)
(71)
For the particular case of , (77) is identical to (68). Its
Assuming that the phase and the frequency have a similar be- behavior (in terms of Gaussianity of ) is shown in
havior, two new variables, and , are derived from an anal- Fig. 10. The performance results provided by (74) are available
ogous change of coordinates (pseudopolar) applied on the fre- in Section VIII.
quency instead of the phase. Thus, the new variables
VII. SIMULATION
(72) A few simulations have been carried out to test the obtainable
(73) performance of the proposed reception postprocessing scheme.
SÁNCHEZ-PÉREZ et al.: ENVELOPE-AIDED VITERBI RECEIVERS FOR GMSK SIGNALS WITH LIMITER-DISCRIMINATOR DETECTION 1741

Fig. 9. Received symbols with pseudopolar metric for the same transmission
of Fig. 8.

where is the signal at the receiver input. The other terms


are defined as follows:
• , signal received by line-of-sight (LOS) path,
where has normalized amplitude;
• , signal received by a two-ray
Rayleigh fading path;
• , Gaussian white noise with variance ;
• and are complex, zero-mean, Gaussian random vari-
ables; their variances are and , respectively.
The channel quality ratio

(79)

is the ratio of power received through the LOS path to the power
received through the Rayleigh channel, where

(80)
Fig. 8. Typical behavior of received symbols in a high multipath distortion
transmission. Modulation is GMSK (BT = 0:5 and h = 0:5). Channel with
time response h(t) =  (t) + 0:95e  (t 0 0:75T ) and SNR = 20 dB. (81)
(a) Amplitude and frequency decision space. (b) Envelope and envelope
frequency decision space. Therefore, the mean power of the received signal turns out to
be
The classical three-ray Rician channel, with a variable
quality factor , has been employed. In this channel, the re-
ceived signal is modeled as the sum of a direct component, two (82)
Rayleigh fading components, and Gaussian noise. The ratio of The SNR can be expressed as
powers in the direct and delayed components characterizes
the nature of the channel. When , the channel is the (83)
land mobile channel (LMC) with a Rayleigh envelope. When
For simulation purposes, the power of the noise to be added
, the channel is the well-analyzed and well-known
at the input of the receiver depends on the channel quality
Gaussian channel; when , the channel has a
Rician envelope. Another important parameter is the time delay
(84)
between direct and delayed signal components, .
The Rician channel with two fading paths can be described The simulations have been done for static multipath channel
as averaged over Rayleigh variables. has been con-
sidered. The maximum length of the channel and the IF filter
combined response simulated is , with trellis
(78) states in the VA, and reference signals. The reference
1742 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 52, NO. 10, OCTOBER 2004

Fig. 10. Contour plot of the logarithm of the actual pdf (76), for  = 025 dB. Contour plot of the Gaussian pdf, assumed in the approximated metric (74), is
plotted in dotted line for reference.

signals are obtained through the reception of noiseless known The results of the simulations for different values of the
data. Practical methods for that task would include averaging or channel quality factor are shown in Fig. 11 (for )
channel identification (Chung and Rohani [10] present an anal- and 12 (for ). The curves are averaged for three dif-
ysis-synthesis procedure). Full decision depth has been used in ferent delays: , , and .
the VA. The IF filter is a second-order Butterworth filter with The bit-error rate (BER) improvement is considerable. If we
cutoff frequency at 0.651 times the data rate. consider as the maximum allowable error probability (for
example, this is the usage bound for the DECT standard), the
VIII. RESULTS following comments can be made.
Performance results have been obtained in the described sim-
ulation for the comprehensive metric in (48), as well as for some • Low channel quality, from Rayleigh channel (
simpler expressions based on one or two of its components, and dB) to : Only bi- or tridimensional decision spaces
for the pseudopolar metric. They have been described in detail (detection methods DM4–DM6) provide an acceptable
in Section V and its subsections, and in Section VI. Receivers BER. Pseudopolar approach DM6 results in an important
based on the simple threshold detector and on the VA with con- advantage over DM4 (about 6 dB) with a similar com-
ventional metric have been also simulated for reference. In what plexity. Use of envelope derivative DM5 does not make
follows, the seven resulting detection methods (DM) are re- much difference, with respect to DM4.
ferred as DM1–DM7, in order of growing complexity, and cor- • Moderate channel quality dB: Simple threshold
respond to the following: decision DM1 gives unacceptable BER. Classical VA
DM1) Simple threshold decision: if ; based on LD output alone (DM2) is almost enough,
, elsewhere; but only in very low noise conditions (about 35 dB).
DM2) VA with conventional squared frequency error Single-dimension VAs based on envelope-frequency
metric, ; product DM3 perform about 10 dB better, and could be
DM3) VA with squared weighted frequency error metric acceptable for medium noise environment, but not for
(56) described in Section V-A, which is just a systems that require lower BER (typically for data
weighted version of DM1: ; transmission). Bi- and tridimensional (DM4 and DM5)
DM4) VA with joint envelope-frequency error metric systems seem to have very good performances, and reach
(68) described in Section V-D, that includes in reasonable noise range. Pseudopolar ap-
also the contribution of the amplitude error: proach DM6 is much better, again.
; • High channel quality, dB: In low-distortion sys-
DM5) VA with joint envelope, envelope derivative, and fre- tems, the effect of the inclusion of MLSD is not very im-
quency error metric (48) derived in Section V, that portant when the induced ISI of the GMSK is not very
includes also the contribution of the envelope deriva- strong ( , Fig. 11). In this case, the maximum
tive: difference between threshold decision DM1 and the best
; MLSD, DM6, is approximately 4 dB. For higher induced
DM6) VA with metric derived from pseudopolar ap- ISI ( , Fig. 12), the difference is significant.
proach (74) in Section VI: However, in both cases, and , the
. different metric definitions give rise to similar results be-
SÁNCHEZ-PÉREZ et al.: ENVELOPE-AIDED VITERBI RECEIVERS FOR GMSK SIGNALS WITH LIMITER-DISCRIMINATOR DETECTION 1743

Fig. 11. Performance results for different quality factor K and detection methods. Modulation is GMSK (BT = 0:50 and h = 0:5). BERs have been averaged
for three different second ray delays: (7=8)T; (9=8)T , and (11=8)T . Detection methods are: 1) threshold decision; 2) VA with metric d(_ ) = _ ; 3) VA with
metric (56); 4) VA with metric (68); 5) VA with metric (48); 6) VA with metric (77).

cause they mainly differ in the use of envelope informa- complexity is not increased. (In a nonmultipath propagation
tion, and for high channel quality, the envelope is closer environment, the number of trellis states of the VA would be
to the case of no distortion and constant envelope. higher for .)
• Nondispersive channel, dB: The performance
enhancement due to the use of the VA is almost negligible On the Importance of Amplitude Information
(1 dB between DM1 and DM2 to DM6) for , We have emphasized the role of an adequate metric, but a
since the induced ISI is low, and therefore, the eye diagram good metric alone is not enough. Sampled signals that corre-
is wide open, but it is important for where spond to different transmitted data must be far enough from one
even a perfect synchronism can not avoid an strong ISI another. That is probably the key point of the proposed proce-
effect. dures that allows us to distinguish between different received
Threshold decision (DM1) is not suitable for low channel signals using both instantaneous frequency and envelope, when
quality, dB or lower. VA alone, DM2, is not enough their instantaneous frequency alone does not. In low-distortion
to get acceptable quality at medium or low SNR and bi- or channels, received signal envelope is almost constant whatever
tridimensional frequency-envelope postprocessing is needed the transmitted symbols are. In this case, the envelope is not
(DM4–DM6). very useful for discriminating the received sequence. On the
Results for show that modulations with small other hand, when multipath distortion is strong and the signal
, that have poorer performance for a simple threshold-based envelopes are very different for different sequences (near zero,
receiver, have a good behavior when MLSD is employed. sometimes), the frequency information can become insufficient,
Lower means lower bandwidth and lower out-of-band but the received signals can still be discriminated using envelope
radiation, and even though it has a poorer performance (than information.
) for threshold detection, the behavior for VA This effect is illustrated in Fig. 13, where the BER has been
detection and the proposed techniques is very similar and the calculated for a receiver that employs only the envelope error
1744 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 52, NO. 10, OCTOBER 2004

Fig. 12. Performance results for different quality factor K and detection methods. Modulation is GMSK (BT = 0:25 and h = 0:5). BERs have been averaged
for three different second ray delays: (7=8)T , (9=8)T , and (11=8)T . Detection methods are: 1) threshold decision; 2) VA with metric d(_) = _ ; 3) VA with
metric (56); 4) VA with metric (68); 5) VA with metric (48); 6) VA with metric (77).

IX. CONCLUSIONS

Our work shows that MLSD-based equalizers are suitable ap-


proaches for frequency-selective channels. However, simple fre-
quency-squared error metric is not adequate, since it does not
incorporate the important effect of envelope distortion in the
statistical distribution of noise. Our metric in (56) includes this
information and results in an appreciable performance improve-
ment. But as long as an envelope-based metric is to be used, an
optimal approach based on their joint statistics must be consid-
ered. Transformation of the decision space to get a Gaussian
type and noise-independent metric is the next important step.
Both the proposed transformations successfully deal with the
distortion issue with a minimum increase in complexity.
The metric criteria derived can be directly applied to other
CPM signaling schemes, such as Gaussian frequency-shift
keying (FSK) or continuous-phase FSK, since they have been
Fig. 13. Performance of envelope-based MLSD detector versus performance derived independently of the frequency impulse response .
of frequency-based MLSD for Rician channel snapshots. K = 0, SNR = 25,
 = (9=8)T . However, the assumption of no modulation in the calculation
of some pdfs will imply that it will represent a better ap-
( axis) and only the frequency error proximation for small modulation index values , where the
( axis) as distance criterion. unmodulated signal assumption is more realistic.
SÁNCHEZ-PÉREZ et al.: ENVELOPE-AIDED VITERBI RECEIVERS FOR GMSK SIGNALS WITH LIMITER-DISCRIMINATOR DETECTION 1745

It is also shown that the apparently small difference of the [7] M. K. Simon and C. C. Wang, “Differential versus limiter discriminator
pseudopolar version has a considerable effect over the final detection of narrowband FM,” IEEE Trans. Commun., vol. COM-31, pp.
1227–1234, Nov. 1983.
BER. The difference of decision methods DM4 and DM6, as [8] A. Svensson and C. E. Sundberg, “Performance evaluation of differen-
shown by (74), is restricted to the low-envelope case. Once tial and discriminator detection of continuous phase modulation,” IEEE
again, the fact that its effect is higher when it is more needed Trans. Veh. Technol., vol. VT-35, pp. 106–117, May 1986.
[9] K. Ohno and F. Adachi, “Performance evaluation of various decision
(i.e., close references with very small envelope) explains the schemes for frequency demodulation of narrowband digital FM signals
difference, and confirms that every improvement based on the in land mobile radio,” IEEE Trans. Veh. Technol., vol. 39, pp. 109–116,
small-envelope case is worthwhile. May 1990.
[10] K. S. Chung and B. Rohani, “Noncoherent detection scheme for GMSK
Envelope is not usually considered in systems based on con- system,” IEE Electron. Lett., vol. 34, pp. 728–730, Apr. 1998.
tinuous envelope modulation, since it is not supposed to con- [11] E. K. Wesel, J. Tellado-Mourelo, and J. M. Cioffi, “Adaptive DFE for
GMSK in indoor radio channels,” IEEE J. Select. Areas Commun., vol.
tain any information, but as our result shows, envelope provides 14, pp. 492–501, Apr. 1996.
an effective form of diversity (envelope information compen- [12] Y. Akaiwa and M. Maki, “An adaptive equalizer for digital FM with
sates the loss of frequency information) under multipath channel limiter-discriminator detection in frequency selective fading channel,”
in Proc. IEEE Vehicular Technology Conf., 1994, pp. 248–251.
conditions. [13] F. J. Casajs et al., “Predistortion techniques for DECT systems based
Multipath channel has an overall effect over the transmitted on limiter-discriminator receivers,” in Proc. IEEE Vehicular Technology
Conf., vol. 3, 1998, pp. 2469–2473.
signal. Even if only the phase of the transmitted signal depends
[14] D. K. Asano and S. Pasupathy, “Improved post-detection processing
on the transmitted data sequence, the waveform at the input of for limiter-discriminator detection of CPM in a Rayleigh, fast fading
the receiver contains important channel information that must channel,” IEEE Trans. Veh. Technol., vol. 44, pp. 729–734, Nov. 1995.
[15] B. Rohani, “Frequency discriminator detection in frequency-selective
be considered to implement a good receiver, and that appears in
fading environments,” Ph.D. dissertation, Curtin Univ. Technol., School
both envelope and phase components. It is widely accepted and a Elect. Comput. Eng., Bentley, WA, Australia, 1999.
common practice that coherent receivers working in a multipath [16] R. F. Pawula, “Improved performance of coded digital FM,” IEEE Trans.
Commun., vol. 47, pp. 1701–1708, Nov. 1999.
environment make use of both phase and envelope information, [17] G. D. Forney, “Maximum-likelihood sequence estimation of digital se-
but it is frequently forgotten that removing the envelope infor- quences in the presence of intersymbol interference,” IEEE Trans. In-
mation in the receiver for a frequency demodulator in a fading form. Theory, vol. IT-18, pp. 363–378, May 1972.
[18] J. F. Hayes, “The Viterbi algorithm applied to digital data transmission,”
channel would produce an important loss of information about IEEE Commun. Mag., vol. 13, pp. 15–20, Mar. 1975.
the channel. [19] K. J. Kerpez, “Viterbi receivers in the presence of severe intersymbol
The same can be said about the envelope variation informa- interference,” in Proc. IEEE GLOBECOM, 1990, pp. 907.5.1–907.5.5.
[20] Y. Iwanami, “Performance of sequence estimation scheme of narrow-
tion , which can be easily obtained in the receiver. However, band digital FM signals with limiter-discriminator detection,” IEEE J.
simulations show that it provides a low additional capability to Select. Areas Commun., vol. 13, pp. 310–315, Feb. 1995.
distinguish the transmitted symbols, and its contribution to the [21] M. K. Simon, “The impact of mismatch on the performance of coded
narrowband FM with limiter/discriminator detection,” IEEE Trans.
final performance is small. Commun., vol. COM-31, pp. 28–36, Jan. 1983.
Frequency and envelope are strongly complementary, when [22] S. O. Rice, “Noise in FM receivers,” in Time Series Analysis, M. Rosen-
frequency fails, envelope works, and vice-versa (see Fig. 13); blatt, Ed. New York: Wiley, 1963, ch. 25, pp. 395–422.
[23] , “Mathematical theory of random noise,” Bell Syst. Tech. J., vol.
only in a few cases are both of them useless. The proposed met- 23, pp. 282–332, July 1944.
rics are an effort to combine them in the best possible way. [24] J. E. Mazo and J. Salz, “Theory of error rates for digital FM,” Bell Syst.
Tech. J., pp. 1511–1535, Nov. 1966.
[25] I. Bar-David and S. Shamai, “On the Rice model of noise in FM re-
ceivers,” IEEE Trans. Inform. Theory, vol. 34, pp. 1406–1419, Nov.
1988.
[26] J. Salz and S. Stein, “Distribution of instantaneous frequency for signal
REFERENCES plus noise,” IEEE Trans. Inform. Theory, vol. IT-10, pp. 272–274, Oct.
1964.
[27] S. O. Rice, “Statistical properties of a sine wave plus random noise,” Bell
[1] J. B. Anderson et al., Digital Phase Modulation. New York: Plenum, Syst. Tech. J., vol. 27, pp. 109–153, 1948.
1986.
[2] “Recommendation GSM 05.04, Modulation,” Eur. Telecommun. Stan-
dards Inst. (ETSI), Sophia Antipolis, France, Tech. Rep. Ver. 3.1.2, 1992.
[3] “Digital Enhanced Cordless Telecommunications (DECT); Common In-
terface (CI); Part 2: Physical Layer (phl),” Eur. Telecommun. Standards
Inst. (ETSI), Sophia Antipolis, France, Tech. Rep. EN 300 175-2 V1.4.2,
1999.
[4] “Radio Equipment and Systems (RES); High Performance Radio Local
Area Network Type 1; Functional Specification,” Eur. Telecommun. Ramón Sánchez-Pérez received the M. Eng. and
Standards Inst. (ETSI), Sophia Antipolis, France, Tech. Rep. ETS 300 Ph.D. degrees in telecommunication engineering
652, 1996. from the Universidad Politécnica de Madrid, Madrid,
[5] K. Murota and K. Hirade, “GMSK modulation for digital mobile radio Spain, in 1996 and 2001, respectively.
telephony,” IEEE Trans. Commun., vol. COM-29, pp. 1044–1055, July He is currently with Agere Systems’ Ethernet Di-
1981. vision, Madrid, Spain. His main research interest is in
[6] O. Andrisano and N. Ladisa, “On the spectral efficiency of CPM systems signal theory and communications applied to broad-
over real channel in the presence of adjacent channel and cochannel in- band communications.
terference: A comparison between partial and full response systems,”
IEEE Trans. Veh. Technol., vol. 39, pp. 38–100, May 1990.
1746 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 52, NO. 10, OCTOBER 2004

Subbarayan Pasupathy (M’73–SM’81–F’91) was Francisco Javier Casajús-Quirós (S’81–M’89)


born in Chennai (Madras), India, on September 21, received the M.Eng. and Ph.D. degrees in telecom-
1940. He received the B.E. degree in telecommu- munication engineering from the Universidad
nications from the University of Madras, Madras, Politécnica de Madrid, Madrid, Spain, in 1982 and
India, in 1963, the M.Tech. degree in electrical 1988, respectively.
engineering from the Indian Institute of Technology, He has been an Associate Professor with the Uni-
Madras, India, in 1966, and the M.Phil. and Ph.D. versidad Politécnica de Madrid since 1989, where he
degrees in engineering and applied science from is Vice-Head of the Signals, Systems and Radiocom-
Yale University, New Haven, CT, in 1970 and 1972, munications Department. His main research interests
respectively. are in digital signal processing applied to wireless
He joined the faculty of the University of Toronto, communications and multimedia. In these fields, he
Toronto, ON, Canada, in 1973 and became a Professor of Electrical Engineering has authored and coauthored more than 100 publications in journals and con-
in 1983. He has served as Chairman of the Communications Group and as the ference proceedings.
Associate Chairman of the Department of Electrical Engineering at the Univer-
sity of Toronto. His research interests are in the areas of communications theory,
digital communications, and statistical signal processing.
Dr. Pasupathy is a Registered Professional Engineer in the Province of On-
tario. He was awarded the Canadian Award in Telecommunications in 2003 by
the Canadian Society of Information Theory, and was elected as a Fellow of
the Engineering Society of Canada in 2004. He has served as a Technical As-
sociate Editor for the IEEE Communications Magazine (1979–1982) and as an
Associate Editor for the Canadian Electrical Engineering Journal (1980–1983).
During 1982–1989, he was an Area Editor for Data Communications and Mod-
ulation for the IEEE TRANSACTIONS ON COMMUNICATIONS. From 1984 to 1998,
he wrote a regular column entitled “Light Traffic” for the IEEE Communications
Magazine.

You might also like