IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: ANALOG AND DIGITAL SIGNAL PROCESSING, VOL. 49, NO.

11, NOVEMBER 2002 721

Transactions Briefs__________________________________________________________________
PLL-Based Frequency Discriminator Using the Loop Filter
as an Estimator
Mohammad Hadi Izadi and Bosco Leung
Fig. 1. Digital FM discriminator.
Abstract—In this paper we present a new phase-locked-loop-based fre-
quency discriminator architecture that performs both A/D conversion and
frequency demodulation. In addition, the loop filter is designed as a linear
estimator. Two forms of linear estimators are investigated. Simulation re-
sults show that for a minimum shift key signal at 400 MHz IF, the proposed
frequency discriminator achieves an SNR improvement of up to 5 dB over
a large variety of noise/fading conditions.
Index Terms—Adaptive systems, Kalman filters, phase locked loops
(PLLs).

I. INTRODUCTION
Fig. 2. Block diagram of proposed frequency discriminator.
Traditionally, IF signal processing consists of an analog frequency
discriminator. Precise analog frequency discriminators, however, are
difficult to implement in CMOS technology. Alternatively, the IF signal noise and fading conditions, which are common in a wireless environ-
can be first digitized using either bandpass or passband analog-to-dig- ment. This adaptiveness results in a reduced requirement on the SNR
ital converters (ADCs) and then processed by a digital signal processor and, hence, dynamic range of the frequency discriminator. Two filter
(Fig. 1). Implementing these passband or bandpass ADCs poses great design approaches are explored. The first is a Weiner-based LMS adap-
design challenges [1], [2]. tive filter and the second is a first-order Kalman filter.
Recently, there have been other works on frequency discriminators Section II introduces our proposed architecture and discusses its use
[3]–[6]. Specifically, [5] works on trying to perform the A/D conversion with a Wiener filter as the loop filter. Section III investigates the use
function directly in the phase domain by using a 16 -based feedback of the Kalman filter in the proposed architecture. Section IV explains
loop. One of the disadvantages of 16 -based discriminators, however, the implementation of the architecture using a Wiener filter and for an
is that noise effects are not alleviated by threshold extension. This FM MSK input. Simulation results are then presented. Section V explains
threshold extension is defined as the point when the output SNR dete- the implementation using a Kalman filter and discusses simulation re-
riorates rapidly with decreasing carrier-to-noise (CNR) ratio. sults. Section VI contains the conclusions.
Our present approach also consists of doing the A/D conversion
in the phase domain. Essentially, we try to use a phase-locked-loop II. PROPOSED ARCHITECTURE USING THE WIENER FILTER AS AN
(PLL)-type structure to perform frequency demodulation and OPTIMUM LOOP FILTER
analog-to-digital conversion in a single step, which we call frequency
discrimination. A PLL structure has the benefit of threshold extension, The architecture that we propose is shown in Fig. 2. Referring to
which translates into low bit error rate (BER) detection in the case of Fig. 2, A/D conversion is done in the combined block of phase de-
digital FM signals, such as minimum shift key (MSK), at low SNR tector/phase-to-digital converter. Specifically, by using a digital filter
levels. A PLL extends this threshold level since it acts as a narrow as the loop filter, we are directly performing frequency-to-digital con-
tracking filter where the noise bandwidth it responds to is centered version, such that the output stream represents our demodulated wave-
about the instantaneous carrier frequency. form in the digital domain. Since it is a PLL-based discriminator, we
In addition, it has been shown that the optimum angle demodulator will be able to obtain threshold extension, which will translate into in-
for continuous-time signals, under certain approximations, is an analog creased BER performance.
phase locked loop [7]. We further assume, then, that a digital phase We are now in a position to discuss the loop filter design from an
locked loop, neglecting quantization effects, is also an optimum, or optimum discrete-time estimator viewpoint. Here, the loop filter is op-
near-optimum angle demodulator. In fact, it has been shown that under timized to make the PLL the best estimator. Since this use of loop filter
high SNR conditions, the optimum digital PLL (DPLL) loop filter is is quite different from its conventional use in a PLL (such as reducing
the Kalman filter [8]. phase noise, spurious tone, stability), it is worth elaboration.
A PLL with a fixed loop filter is not optimal when the input condi- We will first discuss using the nonadaptive Wiener filter as the loop
tions are changing. This paper explores the optimal design of a PLL- filter, as shown in Fig. 3(a). We will then show the limitation with
based frequency discriminator that will adapt its loop filter to changing this method for our application and show that making the Wiener filter
adaptive will overcome the limitation.
A discrete-time Wiener filter can be implemented using either a fi-
Manuscript received May 7, 2002; revised November 25, 2002. This paper nite-impulse response (FIR) or infinite-impulse response (IIR) filter.
was recommended by Associate Editor P. Carbone. We will only consider the FIR Wiener implementation. The input to
The authors are with the Department of Electrical and Computer Engi-
() ()
our filter, denoted as y k , consists of the desired message x k and
neering, University of Waterloo, Waterloo, ON N2L 3G1, Canada (e-mail:
[email protected]). ()
noise v k , as given by y k ( ) = ( )+ ( ) ( )
xk v k . x k is basically the
Digital Object Identifier 10.1109/TCSII.2002.807572 original message. It is first modulated and sent across a noisy channel

1057-7130/02$17.00 © 2002 IEEE

722 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: ANALOG AND DIGITAL SIGNAL PROCESSING, VOL. 49, NO. 11, NOVEMBER 2002

In practice, the channel characteristics are not known beforehand

and, hence, Ryy01 can only be estimated. Thus, Ryy 01 is replaced with
, which is known as the adaptation step size. In addition, the gradient
can be shown to be equal to the cross correlation of the error and the
input vector to the filter, r = 02Rey . Replacing Ryy01 with  and r
with 02Rey , we can rewrite (4) as
(a)

h(n; k + 1) = h(n; k) + Rey (k) (5)

where n is the tap number and k is the time index. The difficulty with
this equation is the expectation operator inherent in Rey (k). Since we
do not know the input statistics, we will ignore the expectation oper-
ator. We now have a random value whose mean is equal to the desired
gradient, thus giving us an unbiased estimate.
We are left with the least mean-square (LMS) coefficient update al-
gorithm, namely,
(b) h(n; k + 1) = h(n; k) + e(k)y(n 0 k) (6)
Fig. 3. Loop filter: nonadaptive and adaptive Wiener filter. where the subscript k again indicates the time index and e(k) is the
error term given by x(k) 0 x ^(k ). The adaptation step size  deter-
and becomes the analog IF input in Fig. 2. The PLL demodulates this mines the rate of convergence of the LMS algorithm, and also the av-
and produces y (k) at node 1 which contains this original message x(k) erage mean-square error. A fixed adaptation step size for all the filter
and the demodulated channel noise, which we have denoted as v (k). coefficients can lead to slow convergence if the eigenvalues of the filter
The estimate of our original message, assuming an FIR based filter, is input covariance matrix Ryy have a large spread. This is due to the fact
given by the output of the filter and denoted x
^(k ) that a single  cannot lead to fast convergence of all the coefficient
deviation components. Therefore we use the orthogonalized LMS al-
0
n 1 gorithm, where we replace  with an estimate of Ryy 01 . Fig. 3(b) depicts
x^(k) = h(i)y(k 0 i): (1) our final choice of the adaptive Wiener filter, which uses this update al-
i=0 gorithm. We have now finished coming up with the appropriate Wiener
filter as the optimum loop filter for our architecture. In Section IV, we
Here, h(i) is the impulse response of the filter, and n is the length
will discuss its specific implementation.
of the filter. Since this filter is used as an estimator, its design is guided
For completeness, it should be mentioned that there are several other
by trying to minimize the mean square error. Finding the minimum
types of adaptive filters based on discrete-time Wiener–Hopf equations.
mean-square error (MMSE) of the above filter will lead to the following
For example, the RLS algorithm is known for its self-orthogonalizing
Wiener–Hopf equation:
of filter coefficients, and its rapid tracking when neither the reference
0
n 1 signal nor the input (received) signal (or channel) characteristics can be
h0 (i)Ryy (i 0 j ) = Rxy (j ) j = 0; 1; . . . ; m 0 1 (2) controlled. The disadvantage of the filters based on the RLS algorithm
i=0 lies in their complexity, as even the most efficient type is still about
four times as complex as the LMS algorithm [9]. The LMS algorithm
where the filter input autocorrelation and the cross-correlation between is the simplest form available and we have chosen to use it for the loop
the filter input and the desired signal are referred to as Ryy (i; j ) and filter in order to implement it in reasonable chip area.
Rxy (j ), respectively. The solution of the discrete-time Wiener–Hopf
equation gives us the optimum filter coefficients h0 (i). Note that this
III. PROPOSED ARCHITECTURE USING THE KALMAN FILTER AS AN
design approach is quite different from the traditional loop filter design
OPTIMUM LOOP FILTER
approach, which is guided by trying to reduce phase noise and spurious
tones, and which is carried out by selecting proper transfer functions. We again start with the architecture shown in Fig. 2, but with the
The major limitation with the standard FIR (and IIR) Wiener filter Kalman filter as the loop filter.
presented is that it requires knowledge of Ryy , the autocorrelation se- The Kalman filter is the optimum filter for an autoregressive process
quence of the filter input, and Rxy , the cross-correlation sequence be- and is derived using the LMS error criterion. Strictly speaking, when
tween the filter input and some desired response. In our present applica- applied to the PLL structure, the optimality remains only when we treat
tion, our channel characteristics are unknown (since we have a wireless the whole PLL, and not just the loop filter, as the Kalman filter [7].
environment) and, hence, this information is unavailable to us. However, it has been shown [8] that simply using the Kalman filter for
In order to get around this problem, we will need an adaptive form the loop filter will significantly enhance the performance in an all dig-
of the Wiener filter which is shown in Fig. 3(b), such that the tap coef- ital PLL which is preceded by an explicit A/D converter. In our appli-
ficients are adjusted in a recursive manner toward the optimum values. cation, where we have combined the A/D and demodulation functions
By not solving for the MMSE, the Wiener–Hopf equation can be into the PLL block, we assume the advantage offered by the Kalman
rearranged in matrix form such that filter still remains and, thus, this is the approach we have adopted in the
paper. The following presents the Kalman filter for our application.
r~ = 2Ryy 01 R~ xy
~h 0 Ryy (3) The Kalman filter estimates ~ x(k) from the corrupted message ~y(k)
by a model based approach, as opposed to directly minimizing the mean
where r is the gradient of the mean square error with respect to the square error. The idea consists of the following.
filter tap coefficients h. Notice that the second term in the brackets is 1) We assume ~ x(k) is generatable by a model, called the signal
h0 from (2) in matrix form. We can once again rearrange this as model, as shown in Fig. 4. This signal model is described by a
first-order vector driven by a zero-mean white noise source
01 r:
~h0 = ~h 0 1 Ryy
2
(4) ~x(k) = A~x(k 0 1) + w~ (k 0 1): (7)
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: ANALOG AND DIGITAL SIGNAL PROCESSING, VOL. 49, NO. 11, NOVEMBER 2002 723

Fig. 6. Model for modulation and channel.

best estimate of the signal without any additional information.

Fig. 4. Kalman signal and observation model for loop filter. The second term is a correction term giving the difference be-
tween the new data sample and the observation estimate, shown
in Fig. 5. This term is weighted by the variable gain K (k), which
is referred to as the Kalman gain in the literature. The Kalman
gain determines the structure of the filter. When K (k) is small,
the estimator “believes” the model, and when K (k) is large, the
estimator “believes” the observation (or measurement). For in-
stance, if we refer to (12), we can see that if P (k) is small (the
system model is good) and cov (~v ; ~v ) is large (the observation is
Fig. 5. Loop filter: Kalman filter. noisy), then K (k) will be small and the estimator “believes” the
model. Conversely, if P (k) is large (the system model is bad)
Here, A is a signal model parameter used to describe the signal and cov (~v ; ~v ) is small (the observation is good), then K (k) will
x(k )
~ recursively and w ~ (k ) is the uncertainty in modeling. be large and the estimator “believes” the measurement.
2) We assume that when the message is modulated, sent through 4) We improve our knowledge of both models by tuning. The term
the channel, demodulated by the PLL, and sent to node 1 of y (k ) 0 AC ~
~ ^(k 0 1), is referred to as the innovation. In order for
x

Fig. 2, that the noise introduced can be incorporated in a model, the Kalman filter to work correctly and, thus, have optimal per-
called the observation model, as shown in Fig. 4. The observa- formance, the system and model parameters must be “tuned” ac-
tion model is described by a linear combination of the message cordingly. A starting point to check the optimality of the Kalman
x(k ) and the observation noise ~
~ v (k ), which is zero mean and filter are the innovation terms referred to as the innovation se-
white. ~v (k) is referred to as the observation noise because it af- quence. A necessary and sufficient condition for a Kalman filter
fects the observed loop filter input ~ y (k ) and it can come from, to be optimal is that the innovation sequence is zero mean and
for example, channel noise white [10].
y (k )
~ = x(k )
C~ +~
v (k ): (8) We have now finished coming up with the Kalman filter as the op-
timum loop filter for our architecture. In Section V we will discuss its
Here, C is an observation model parameter that relates the
implementation.
x(k ) to the observed signal ~
desired signal ~ y (k ).

3) We assume we have some initial knowledge of both models, and

IV. IMPLEMENTATION OF ARCHITECTURE USING WIENER FILTER FOR
we use them to guide our estimation ~x
^(k ) in the following sense.
MSK INPUT AND CORRESPONDING SIMULATIONS
If we assume that we have a prior estimate of the message, we
can form a linear combination of this prior estimate, weighted by In this section, we will apply the proposed architecture toward de-
the parameter A(k), along with the present noisy observation to signing a frequency discriminator for an MSK modulated input. Our
make a future estimate design is targeted for a 400-MHz IF input employing MSK modulation
~
^(k )
x = A(k )~
^(k
x 0 1) + K (k )y
~(k ): (9) and having a message bandwidth of 200 kHz. This can be used, for
example, for Global System for Mobile (GSM) Communication appli-
Here, ~x
^(k ) is our estimate of the message, ~
y (k ) is our observa-
cation.
tion, and A(k) and K (k) are used to weight our estimate and ob-
servation, respectively. In order to find the optimal estimator, we A. Model for Modulation and Channel
again use the MMSE criterion, giving A(k) = A(1 0 C K (k)),
Data bits are FM modulated using MSK and sent to the channel as
where A is the signal parameter in (7). Substituting this in (9),
shown in Fig. 6. The channel is modeled with additive white Gaussian
our estimation equation becomes
noise (AWGN) and Rayleigh fading. Since the spectrum of MSK is
~
^(k )
x = A~
^(k
x 0 1) + K (k )[~
y (k ) 0 AC ~
^(k
x 0 1)]
: (10) narrow, we assume that the fading introduced by the channel is fre-
This is shown in Fig. 5. quency nonselective. The input signal in Fig. 6 is assumed to be pro-
Here cessed by a front end, mixed down to IF and becomes the analog IF
input in Fig. 2.
K (k ) =C P (k )cov (~ v)
v; ~
The modulation index h for an MSK signal is equal to 0.5, and is
=C (AP (k 0 1) T + cov (w;
A ~ w~ )) given by h = (kf Am =fm ), where kf is the digital-controlled oscil-
2 cov (~ v ) + C cov (w;
v; ~ ~ w~ )C
T lator (DCO) sensitivity, Am is the amplitude of the binary source, and
fm is the modulation frequency. With unity amplitude, this means that
01
+ AC P (k 0 1) C
T AT (11) the DCO sensitivity must be 100 kHz to give us a modulation index of
0.5.
is determined recursively using
P (k ) = C
01 cov(~v; ~v)K (k): (12) B. Phase Detector and DCO
Equations (10)–(12) make up what is known as the Kalman al- The DCO takes the digital control voltage and generates a set of
gorithm. The first term of (10) is a prediction term that gives the square waves which are compared with the input. Each square wave has
724 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: ANALOG AND DIGITAL SIGNAL PROCESSING, VOL. 49, NO. 11, NOVEMBER 2002

an identical frequency, controlled by the digital input, whereas each of signal; therefore, the output will oscillate. The output of the loop filter
their phases are offset by the smallest quantization step. In our design is integrated over one symbol period at the VCO sampling frequency.
we choose the quantization step to be 11.25 , and so the DCO has 32 The output of the Integrate-and-Dump (I/D) filter is passed to a de-
taps. One convenient implementation for the DCO would be to use a cision device, which forms decisions based upon a predetermined av-
32 stage ring oscillator. These 32 square waves serve as references to erage output corresponding to logic 0 and logic 1 baseband values. Thus
the phase detector (PD). Conceptually, the PD can be thought of as a the difference between the decision device output (expected value) and
bank of 32 1-bit phase detectors. Each of these 1-bit phase detectors is the output of the I/D (estimated value) is used to generate an error signal
similar to the PD used in a conventional PLL. Each receives two inputs: which is fed back to the adaptive Wiener to adjust the tap coefficients
one from the analog IF input and the other from one of the 32 square in a manner that will reduce symbol error.
waves generated by the DCO. The 1-bit output of the PD then tells us Under zero noise conditions we can determine the value of x ^(k ),
which one of the two inputs are leading. In this way the PD acts like a which in our case is the output of the I/D, not the filter. We have chosen
flash ADC, except the digitization is performed in the phase domain. to do this because it allows us to average the output under heavy noise,
The phase to digital converter is analogous to the thermometer code which would otherwise lead to many errors. By taking a statistical av-
decoder in a flash ADC. The output y (k) is then a quantized version erage of the input waveform, we can find Ryy 01 , and thus apply this value
of the demodulated, corrupted MSK signal (corrupted by noise and to determine the updated value and then apply it to our filter tap coef-
fading). The loop filter then cleans the corruption and outputs the final ficients. Since we are using the output of the I/D, and not the output of
digital demodulated output. the filter, we will need to scale Ryy01 accordingly. Likewise, the value
of y (k) in (6) is the average value of y (k) over a symbol period.
In Section IV-D we present the simulation results of the frequency
C. Loop Filter Using Adaptive Wiener Filter discriminator based on this filter, and compare the results to a frequency
discriminator with a conventional FIR filter.
It is known that the general shape of the loop filter in a PLL should
be that of a low-pass filter [7]. In conventional frequency discriminator
PLL design, the bandwidth of the loop filter is designed such that it is
at least as large as the bandwidth of the message being demodulated D. Simulation Results
[11], which in our case is 200 KHz. To be precise, the message band-
width is larger than 200 kHz due to the spectral spreading caused by We performed simulations of the frequency discriminator using an
the FM modulation process and due to the fact that we are dealing with adaptive Wiener filter in a noisy environment with and without fading.
a square wave and not a sine wave; however, we can neglect these for The IF input is 400 MHz, message bandwidth is 200 kHz, and DCO
two reasons. First, since we are dealing with an MSK signal, which has gain is 390 kHz/V. The amplitude and phase gain variations due to
a modulation index of 0.5 (narrowband FM signal), spectral spreading fading are uniformly distributed between 0.2–1.8 and 6=4, respec-
is negligible. Second, since we will be implementing the FIR filter with tively. Referring to Node 1 of Fig. 2, since the signal y (k) contains
only a small number of taps (for reasonable chip area), the frequency the demodulated waveform, the essential information lies in a band of
response is not very sharp, so designing for a bandwidth of 200 kHz is 200 kHz; therefore, we pick a loop filter bandwidth that can pass this
adequate. information. Fig. 9 shows a plot of the BER of the discriminator em-
As mentioned in Section IV-C, we have decided to implement a form ploying the adaptive Wiener and conventional loop filters, as a func-
of adaptive Wiener filter using the orthogonalized LMS algorithm. In tion of the input SNR to the frequency discriminator in the presence
addition, for the present application, we have chosen to implement it of channel noise. The plots are extrapolated between two sets of data.
using an FIR filter because of its simpler structure. In implementing this Fig. 10 shows a similar plot, but with both channel noise and fading.
FIR filter, we have used a conventional nonpipelined FIR filter struc- The conventional filter used in the simulations is an eight-tap
ture in order to reduce the latency in the loop. We have implemented low-pass FIR filter designed using the Kaiser window approach.
the loop filter using only eight taps to reduce the complexity and in-
As can be seen from the plots, the adaptive Wiener filter performs
crease the speed of the circuit. The initial values of our filter taps are
better than the conventional filter in both cases. Due to the nature of the
set such that they correspond to a low-pass filter. This obviates the need
PLL, the input to the loop filter is time varying. The mixing operation
to send a training sequence to the adaptive filter. Adjustments are made
inherent in the sample-and-hold phase detector along with the nonuni-
to the filter tap coefficients every symbol, allowing the filter to change
form sampling brings about the time-varying nature of the filter input.
under different noise and fading conditions. A block diagram of the
The adaptive Wiener filter is able to adapt itself to the changing input
implementation of this adaptive Wiener filter is shown in Fig. 7, which
and, thus, produce better results than the conventional loop filter. We
maps closely with the high-level block diagram of Fig. 3(b). Here, clk
can see from Figs. 9 and 10 that the adaptive Wiener filter performs
is from the output of the DCO, and is used to clock in data from the
quite well at low SNR values, making it potentially better than [5].
output of the phase-to-digital converter. Coeff clk is used to update the
We assume the channel is frequency nonselective and therefore the
tap coefficients of the adaptive Wiener filter and averaged y (k) refers
to y (n 0 k) from (6).
fading is predominantly Rayleigh (envelope) fading [12]. Therefore,
the fading introduced into the channel does not cause intersymbol inter-
Fig. 8 shows a block diagram of the Matlab model used to simulate ference (ISI), however, it does increase the BER. The fading causes the
the frequency discriminator employing this adaptive Wiener filter. The amplitude and phase of the carrier signal to vary in a random manner.
sample-and-hold and A/D block correspond to the phase detector and When sampled by the PLL PD, this variation in the phase and ampli-
phase-to-digital converter block in Fig. 2. The adaptive Wiener filter tude are translated into noise-like behavior in the loop filter input. As in
block corresponds to the adaptive Wiener filter in Fig. 7. the case with noise alone, the adaptive Wiener filter is able to adapt it-
Under ideal conditions, without noise or fading, the output of our self to this changing input condition and produce better results than the
filter, x
^(k ), just consists of two values corresponding to the two fre- conventional loop filter, which is shown in Fig. 10. Finally, because of
quencies used to represent a 0 bit or a 1 bit. this increase in noise due to fading, when comparing the curves for the
The output of the filter is the recovered baseband data. Due to quan- adaptive Wiener filter in Figs. 9 and 10, the one in Fig. 10 has poorer
tization effects, however, our loop cannot precisely track the baseband results.
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: ANALOG AND DIGITAL SIGNAL PROCESSING, VOL. 49, NO. 11, NOVEMBER 2002 725

Fig. 7. Block diagram of adaptive Wiener filter implementation.

Fig. 8. Block diagram of frequency discriminator employing adaptive Wiener filter used for simulation.

V. IMPLEMENTATION OF ARCHITECTURE USING KALMAN FILTER FOR A. Loop Filter Using Kalman Filter
MSK INPUT AND CORRESPONDING SIMULATIONS
As discussed in Section III, in designing our Kalman filter, we need
In this section, the same frequency discriminator as in Section IV to determine our signal and observation models, which are given by (7)
will be used, except the loop filter will be implemented using a Kalman and (8).
filter. We start with the Kalman filter developed in Section III. This Since the signal model is responsible for generating the message, we
Kalman filter, in vector form, can be simplified for our present applica- need to know what the message is. We know that for MSK modulation,
tion because the frequency discriminator is designed for an MSK-mod- the message wavefrom comes from a binary source (see Fig. 6) and,
ulated signal. We will now discuss this simplification. therefore, the message is a square wave in shape, so we can expect
726 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: ANALOG AND DIGITAL SIGNAL PROCESSING, VOL. 49, NO. 11, NOVEMBER 2002

Fig. 11. Block diagram of Kalman filter.

Fig. 9. Comparison of discriminator performance using adaptive Wiener

versus conventional fixed loop filter. Only noise is present in the channel.
Testing conditions: 400-MHz IF MSK signal, 200-kHz message bandwidth,
DCO gain = 390 kHz/V.

Fig. 12. Kalman filter frequency response.

where we have normalized and made c = 1. The term v (k), referred to

as the observation noise, is used to represent the noise in the channel.
The two updated equations give us the following Kalman system
equation in scalar form with a = c = 1:
x^(k) = x^(k 0 1) + K (k)[y(k) 0 x^(k 0 1)]: (15)
From the system equation, we can draw the implementation of our
Kalman filter for an MSK-modulated input as shown in Fig. 11.
Our Kalman gain in scalar form is now
c a2 p(k 0 1) + w2
Fig. 10. Comparison of discriminator performance using adaptive Wiener
versus conventional fixed loop filter. Both noise and fading are present in cp(k)
K (k) = :
v2 + c2 w2 + c2 a2 p(k 0 1)
= (16)
the channel. Testing conditions: 400-MHz IF MSK signal, 200-kHz message v2
bandwidth, DCO gain = 390 kHz/V; fading conditions: phase gain variation
6
= =4 rad, amplitude gain variation = 0.2–1.8. Our filter transfer function is given by
X (z ) Kz
Hk (z ) =
z 0 (1 0 K )
= (17)
our estimate, ~x
^(k), to also be a square wave in shape. In this case, Y (z )
(7) can be simplified considerably. First of all, we can describe our where again a and c have been replaced by one.
message using a first-order recursive equation, so ~ x(k) becomes scalar, It can be seen that the filter characteristic is now solely determined
and (8)–(12) developed for a vector Kalman algorithm also become by the value of the Kalman gain K . Referring to (16), we can see that
2
scalar. Thus, the vectors ~ x, ~y, w
~ , ~v become scalars and the matrices with a and c fixed, that K is solely determined by the values w and
2
A; C; . . . become scalar a; c; . . . The covariance matrices cov(w;
~ w ~) v . More accurately, in fact, the value of K is determined by the ratio
2
and cov (~v ; ~v ) become scalar variances w and v2 . between w 2
and v2 , and not their specific values. Thus, our pole posi-
The message can be thought of as a square-wave continuous-time tion and, hence, the frequency response of our filter is determined by
2
(analog) signal, since the sampling rate of the PLL is much larger than the ratio between w and v2 .
the bit rate. In our case the sampling rate is 1000 times the baseband The challenge in designing this filter is choosing the appropriate
2
data rate. Thus, neglecting noise, during each bit period the desired value of w for a given input noise variance, v2 , and input frequency to
signal x(k) is constant for many consecutive samples such that x(k) = the filter. This translates to setting the appropriate filter bandwidth and,
x(k 0 1). Thus the signal model can be described by thus, the loop bandwidth of the PLL. This amounts to making some ini-
x(k) = x(k 0 1) + w(k 0 1):
tial estimate of the model (step 3 in Section III) and tuning (step 4 of
(13) Section III). Following these two steps for the present application re-
This recursive equation indicates that the present sample will be sults in a Kalman filter whose frequency response is shown in Fig. 12
equal to the past sample, with some correction term denoted by w(k 0 [13].
1) which we refer to as the model noise. Our observation model can be
described by B. Simulation Results
We performed simulations of the frequency discriminator with the
y(k) = x(k) + v(k) (14) Kalman filter in a noisy environment with and without fading. Again,
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: ANALOG AND DIGITAL SIGNAL PROCESSING, VOL. 49, NO. 11, NOVEMBER 2002 727

Fig. 15. Comparison of discriminator performance using Kalman versus

Fig. 13. Comparison of discriminator performance using Kalman versus adaptive Wiener versus conventional fixed loop filter. Only noise is present in
conventional fixed loop filter. Only noise is present in the channel. Testing the channel. Testing conditions: 400-MHz IF MSK signal, 200-kHz message
conditions: 400-MHz IF MSK signal, 200-kHz message bandwidth, DCO gain =
bandwidth, DCO gain 390 kHz/V.
= 390 kHz/V.

VI. CONCLUSION
We have presented a new PLL-based frequency discriminator archi-
tecture that performs both frequency demodulation and A/D conver-
sion. The loop filter is designed from a linear estimator viewpoint. Two
such forms of loop filters were investigated. Simulation results show
that for MSK signals at 400 MHz IF, an SNR improvement of up to 5
dB is achieved over a large variety of noise/fading conditions.

ACKNOWLEDGMENT
The authors would like to thank M. Sharifkhani for valuable discus-
sions.

REFERENCES
[1] A. Namdar and B. Leung, “A 400 MHz, 12-bit, 18 mW IF digitizer with
mixer inside a 61 modulator loop,” IEEE J. Solid-State Circuits, vol.
34, pp. 1765–1776, Dec. 1999.
Fig. 14. Comparison of discriminator performance using Kalman versus
conventional fixed loop filter. Both noise and fading are present in the channel.
[2] S. A. Jantzi, K. W. Marth, and A. S. Sedra, “Quadrature bandpass 16
modulation for digital radio,” IEEE J. Solid-State Circuits, vol. 32, pp.
Testing conditions: 400-MHz IF MSK signal, 200-kHz message bandwidth,
=
DCO gain 390 kHz/V; fading conditions: phase gain variation =6 4 = rad,
1935–1950, Dec. 1997.
16
=
amplitude gain variation 0.2 –1.8.
[3] I. Galton et al., “A PLL for 14-b, 50 kSample/s frequency-to-digital
conversion of a 10 MHz FM signal,” IEEE J. Solid-State Circuits, vol.
33, pp. 2042–2053, Dec. 1998.
[4] I. Galton, “Analog-input digital phase-locked loops for precise fre-
the IF input is 400 MHz, message bandwidth is 200 kHz, and DCO gain quency and phase demodulation,” IEEE Trans. Circuits Syst. II, vol.
is 390 kHz/V. The amplitude and phase gain variations due to fading are 42, pp. 621–630, Oct. 1995.
uniformly distributed between 0.2–1.8 and 6=4, respectively. Fig. 13 [5] R. D. Beards and M. A. Copeland, “An oversampling 16 frequency
discriminator,” IEEE Trans. Circuits Syst. II, vol. 41, pp. 26–32, Jan.
shows a plot of the bit error rate of the discriminator employing the 1994.
Kalman and conventional loop filters as a function of the input SNR [6] E. J. van der Zwan et al., “A 10.7 MHz IF-to-baseband 16 A/D conver-
to the frequency discriminator in the presence of channel noise. The sion system for AM/FM radio receivers,” IEEE J. Solid-State Circuits,
curves are extrapolated between simulated data points. Fig. 14 shows vol. 35, pp. 1810–1819, Dec. 2000.
[7] A. J. Viterbi, Principles of Coherent Communications. New York: Mc-
a similar plot, but with both channel noise and fading.
Graw-Hill, 1966.
As can be seen from the plots, the Kalman filter performed better [8] S. C. Gupta, “Phase-locked loops,” Proc. IEEE, vol. 63, pp. 291–306,
than the conventional loop filter in both cases. These better results are Feb. 1975.
in part due to our tuning of the Kalman filter bandwidth until the in- [9] S. Qureshi, “Adaptive equalization,” Proc. IEEE, vol. 73, pp.
1349–1386, Sept. 1985.
novation sequence is white. We can see from Figs. 13 and 14 that the
[10] J. V. Candy, Signal Processing: The Model-Based Approach. New
Kalman filter performs quite well at low SNR values. York: McGraw-Hill, 1986.
Fig. 15 shows the BER versus SNR trends in a noisy environment [11] D. H. Wolaver, Phase-Locked Loop Circuit Design. Englewood Cliffs,
of a discriminator with all the three different filters as loop filters. The NJ: Prentice-Hall, 1991.
adaptive Wiener-filter-based discriminator outperforms all other filters [12] B. Leung, VLSI for Wireless Communication. Englewood Cliffs, NJ:
Prentice-Hall, 2002.
under heavy noise conditions. When the input noise is not so severe, [13] M. H. Izadi, “PLL-based frequency discriminator using an estimation
it appears that the Kalman-filter-based discriminator performance is loop filter,” MASc. thesis, Dept. Elect. Comput. Eng., Univ. Waterloo,
better than the adaptive Wiener-filter-based discriminator. Waterloo, ON, Canada, 2002.