Proceedings of the American Control Conference

Chicago, Illinois June 2000

FREQUENCY TRACKING VIA
EXTENDED KALMAN FILTER: PARAMETER DESIGN

Sergio Bittanti, Sergio M. Savaresi

Politecnico cJi Miluno, Dipartiinento di Elettronica e Ii$orriicizione, Picizzci Leonardo dci Vinci,32, 20133,
Milcino, ITALY. E-iVIui1: :hiltun/i;sLi1:cii,esi~ ' ~ ~ l e ~ . ~ ~ o l i i i t i . i t .

Abstract. In this paper, the problem of estimating the
frequency of an harmonic signal embedded in broad-band
noise is considered. The paper focuses on the Extended
Kalman Filter Frequency Tracker, which is the application
of the Extended Kalman Filter (EKF) framework to the
frequency estimation problem. The EKF Frequency Tracker
recently proposed in the literature is characterised by a
vector of three design parameters, whose tuning is still a
controversial and unclear issue. In this paper it is shown
that the Extended Kalman Frequency Tracker can be
noises, having variances q and r, respectively:
w(t) W N ( 0 , q ) and v ( t )= W N ( 0 , r ) . The parameter E
(where & L O and, typically, & < < I ) has the goal of
enforcing stability into the third equation of ( I ) . The state
variable x 3 ( t ) represents the unknown frequency U .
Given the state-space model ( l ) , the EKF technique can be
applied straightforwardly (see e.g. [l], [13] for a general
description of EKF). The complete set of equations of the
EKF Frequency Tracker derived from (1) is here recalled
(see [7-91 for more details):

I
parametrized using just one design parameter, without any
loss of performance. This simplification allows an easier ,?(t//) = f ( $ t - I / / - I ) ) + K ( t ) ( y ( l ) - Hf(.i(t - I / / - I)))
and more transparent tuning of the EKF tracking behaviour. &(t) = C i ( r / t )
(2a)
Key words. Extended Kalman Filter; frequency tracking; K ( I ) = P ( ~ ) H( /H P ( ~ ) H+' r)-'
harmonic analysis; parameter estimation. P(t + 1) = F(t)[P(/)- P ( / ) H /( H P ( t ) H / + r ) - l H p ( t ) k /(/) + q/,
where
I . INTRODUCTION
This paper deals with the problem of estimating the
frequency o of an harmonic signal s ( t ) = Acos(ot + cp),
given its noisy measurement y ( t ) = s ( t ) + n ( t ) ,where n ( t )
is a broad-band stationary signal (e.g. a white noise).
This problem is frequently encountered in real-world
applications, especially in the fields of adaptive control and
signal processing, and numerous techniques have been
developed for its treatment (see e.g. [5-61, [lo-111). This
paper focuses on the use of the Extended Kalman Filter
(EKF), an approach which has been recently proposed by
Bitmead and co-authors in [7-91 for the design of a
Frequency Tracker. In particular, in [7] the EKF Frequency
Tracker is proposed, and some heuristic guidelines for the
tuning of its parameters are proposed. In [9] its stability Recursive eqs. (2) are suitable for on-line frequency
analysis is developed, whereas the case of high-noise tracking. In particular, the fourth equation in (2a), giving a
environments is considered in [SI. recursion for the auxiliary matrix P(t) of dimension 3x3, is
The basic idea around which papers [7-91 evolve around is the celebrated Riccafi equation ([3]).
to model the measured signal y ( t ) = s ( t ) + n ( t ) by means As any frequency tracker, from an external point of view
of a third-order state space model, as follows: the EKF Frequency Tracker (2) can be seen as a device

1
x,(t + 1) = cos(x3(t))xl(t) - sin(x,(t))x,(t) receiving the measured signal y ( t ) as input, and delivering
x2(t + 1) = sin(x,(t))x,(t) +cos(x,(t))x,(t) the estimated frequency h(t)= i , ( t / t ) as output. The
(1) internal structure is constituted by two main sub-parts: the
x;(t+1)=(1-E)X3(t)+W(f)
first one performs the updating of the state equations,
y ( t ) = XI ( t ) + 4 2 ) according to the first expression in (2a); the second one has
Here w(t) and v ( t ) are zero-mean incorrelated white the task of providing the gain vector at each time step, and
is based on the third and fourth equations of (2a).
0-7803-551 9-9/00 $10.00 0 2000 AACC 2225
Notice that the EKF is parameterised in { 4 , r , ~ }These .
three parameters are the ”knobs”, by means of which the
designer can obtain suitable tracking performance.
In [7] and [9] a preliminary analysis of the roles played by
the parameters { 4 . r , ~ }is given, and some guidelines for
their tuning are proposed. Such analysis, having the
objective of explaining the effects of each parameter on the
EKF performance. is mainly based upon Monte-Carlo
simulation trials. Some useful and interesting indications
have been obtained in this way, but the tuning of ( 4 , r , ~ )
still remains a difficult task due to the unclear cross-
relationships between such parameters.
Moving from the seminal works [7-91, the goal of this
paper is to go a step further in the analysis of the EKF
Frequency Tracker, in order to gain more insight in the
relationships between its behaviour and its tuneable
parameters (4,r, E } . Precisely, two simple but fundamental
results will be established:
0 Even in a noise-free setting ( n ( t ) = 0 V t ), if E > 0 the
EKF Frequency Tracker (2) provides an
asymptotically biased estimate of the unknown
frequency w . The bias is zero only for E = 0 .
Q The performance of the EKF Frequency Tracker
depends on the ratio h = r 14 only.
The main outcome of these statements is that a wise
parameterisation of (2) corresponds to set E = 0 from the
very beginning, and then using h = r14 as the only tuning
knob. This greatly simplifies the designer task.
The outline of the paper is as follows: in Section 2 a bias-
analysis of (2) is given, whereas Section 3 is devoted to the
analysis of the dependence of (2) upon r and q. These
results are both illustrated with a numerical example at the
end of Section 2 and Section 3, respectively.
2. BIAS ANALYSIS
A minimal requirement for a good frequency tracker is the
so-called unbiasedness in a noise-free setting. This means
that, if the measure y(t) of the harmonic signal
s ( t ) = Acos(wt) is not corrupted by noise ( n ( t ) = 0 V t ),
and the frequency w is constant, the frequency tracker is
expected to provide an estimate h(t) which, at steady-
state, coincides with w . Most of the frequency trackers
typically used in real-world applications (see e.g. [2], [4-61,
[IO-121) obviously have the above simple property. In the
subsequent result we will show that the EKF Frequency
Tracker (2) does not have this property if E > 0 .
Proposition 1. Consider the EKF Frequency Tracker (2).
Assume that n(t) = 0, V t and that the frequency w of
the harmonic signal s ( t ) = Acos(ot) is constant. At
steady-state, the frequency estimate & ( t ) = , f , ( t / t )
provided by (2) is unbiased if and only if EW = 0 .
2226
Proof. Consider the state equations of the EKF (2a), under
the assumption that the frequency estimate is at steady-
state, namely i 3 ( t / t )= i 3 ( t - 1 l t - 1) = X3 . Then, denoting
by { K , ( f ) ) i = , , 2the
, j three components of K(t), the original
equations of (2) can be written as:
1
*GI(///) = C O & ~ ) ? ~ ( / - I / / - 1) -sin(y?)-Q?(/- I / / - I ) +
+ K , (r)lv(/)-co+;)i-, (/ - I / i - i) +sin(i; )q2 ( I - I / / - I)]
(3)
.t2( I //) = siii(.Tl IQ,( I - I / / - I ) + cos(.^;).<-^ ( I - I / / - I ) +
I + K2(t)Lv(l)- c04.Y~PI(I - I / t - I)+sin(Y; ( I - I / / - I)]
X; = ( I -&)y3+ K ; ( t ) ~ ( / ) - C O ~ . Y ; ~ ~I ,/ (/ /-- I )t sin(Tl).Q2(/ - I / / - I)]
Note that the dynamics of (3) is confined into the first two
equations. Thus, i , ( f l t ) and i z ( t / t ) are the only state
variables of system (3). Y, can be seen as its output,
whereas y(t) plays the role of input signal. Interestingly
enough, while system (3) is non-linear, the system obtained
by setting &(t/t)=const. is linear. So, one can compute
the transfer function from y(t)to T3 . To this purpose, using
(3), by cumbersome but simple computations, one obtains:
0;; = K; (t)G(z;Ki ( t ) ,K2 ( t ) ,Z; >y(t> 7 (4)
where
G ( z ; K (, t ) , K 2 ( t ) . y 3=)
z ( 2 - 2ZCOS(Y3) + 1) .( 5 )
-
(K,(t)cos(jT,)-
Z* -z(K,(t)sin(jT3)+2cos(jT3))+ K , ( t ) + 1)
Eq. (4) expresses the relationship between the input y(t) and
the output Ex3, Under the assumption of noise-free setting,
the input signal of system (5) is a pure sinusoid:
y ( t ) = s ( t ) = Acos(wt) . Therefore we can resort to the
frequency-response theorem to determine the steady-state
value of the output. Equation (4) hence can be rewritten as:
6, K , ( / ) ,K 2 ( t ) ,Z3)lc~s(o>/+ L G ( e ” ; Kl(t), K ~ ( / ) , T , ) ) . ( ~ )
= K,(t)IG(e’“’;
Eq. (6) is peculiar in that it expresses how a sinusoidui
signal determines a constant signal. As a matter of fact,
what really matters is when the bias xj -a is zero. Now
from ( 5 ) it is apparent that the numerator evaluated in e’”
is zero if and only if 2, = w . Thus, from ( 6 ) it is possible
to conclude that the frequency estimation bias is zero if and
only if EO = 0 .
The above Proposition 1 states a very simple but interesting
twofold result:
0 At steady-state, the EKF frequency Tracker (2) is
unbiased (for every value of o ) in a noise-free setting
only if the parameter E is set to zero.
Since IG(e’”;K,(t),K,(t),2;)I=O V t if and only if
Z3= w , equation (4) points out that the bias must
depends on the value T, of the sinusoidal signal
frequency. Specifically, since the first member of
equation (4) grows as 2, increases, the bias is expected
 to grow with X3.Such a dependence of the bias upon has been set to ~=0.05;as a second trial, E is set to zero. The
the estimated frequency was already empirically results of the so-obtained frequency estimation are
pointed out in [7]. Proposition 1 now provides a simple displayed in Fig. 1. These diagrams clearly show the main
explanation to this fact. conclusions which can be drawn from Proposition 1 :
A straightforward consequence of the above Proposition 1 If E>O, the estimation is strongly biased even for small
is that the best choice for E is to set E=O; as a matter of fact values of E.
E>O may lead to a strong estimation bias which (unless o The bias increases as the value of the unknown
is very close to zero) dominates the mean square error of frequency increases;
the frequency estimate (this is confirmed by the heuristic The dynamic properties (stability and noise rejection)
analysis proposed in [7] - see also Example 1 below), of the EKF Frequency Tracker in the case ~=0.05and
without providing any clear benefit. In this regard, it is E=O are similar (the only clearly visible difference is
interesting to notice that the condition E>O was introduced just the presence of bias if E>O ).
in [7-91 with the goal of having the 3rd equation of model 25 frequency - _-
, 1
I
(1) asymptotically stable. As is well known, however, the ................................................
stability of the model is not required in order to come out
.*....'.
with a stable Kalman Filter: specifically, it is easy to check
by simulation that the stability properties of the EKF
Frequency Tracker using E=O or E>O (with ~ < < 1 )are
indistinguishable (see Example 1 below). This is confirmed J
by the fact that, in the stability analysis of the EKF 0 100 200 300 400 500
time
600 700 800 900 1000

proposed in [9], the assumption E>O does not play a crucial 2.5
I
frequency
I
-1
1

role.
On the basis of the above result, in the rest of the paper the
EKF Frequency Tracker (2) will be replaced by its
simplified version obtained by substituting E=O in (2),
namely: 05

(i(///) - I / / - I))+ K ( / ) ( Y ( / )- H f ( i ( / - I / / - 1)))

= f(i(/ 0
0 100 200 300 400 500 600 700 800 900 1000

I W ( / ) = Ci(///)

+ .)-'
K ( t ) = P(t)H'(HP(f)H'
( P ( t + 1) = F ( t ) [ P ( / ) - P ( t ) H 7(HP(/)H' + r)-'HP(l)JF' ( I ) + q/
dashed line indicates the true frequency).
tlme
Fig. 1. Frequency tracking of the EKF ~=0.05and E=O (the

where: 3. PARAMETERISATION ANALYSIS

[.=[I 0 01 c=[o 0 I]
In this way, the vector of "tuneable" parameters reduces to
two parameters only, q and r. We conclude this section with
a simulation example, having the goal of showing the effect
of setting E=O vs. E ~ Oin (2).
Example 1.
The signal considered in this numerical example is an
harmonic signal embedded in white noise, characterised by
a time-varying frequency. Specifically, y ( t ) = s ( t ) + n ( t ) ,
t=1,2, ... 1000, where n ( t ) = WN(O,O.Ol), s(t) = cos($t) for
t=1,2 ...300, and s ( t ) = c o s ( 4 t ) for ~ 7 0 1 , 7 0 ...
2 1000; for
~ 3 0 1 , 3 0 2..700
. s ( t ) is characterised by a time-varying
frequency which increases linearly from 4 6 to 2n/3.
In order to estimate the frequency of the harmonic signal
embedded in y ( t ) , the EKF Frequency Tracker (2), with
~0.01 and q=0.005, has been used. First the parameter E
2227
In the previous section it has been shown that, in order the
EKF Frequency Tracker to, be asymptotically unbiased in a
noise-free setting, the parameter E must be set to zero.
Another issue - even more important and interesting - is
the role played by q and r in determining the tracker
performance. Some indications and guidelines are given in
[7] and [9]. The issue, however, remains somewhat
controversial and unclear.
In this section it will be shown that the performances of the
EKF Frequency Tracker are not influenced independently
by q and r, but they only depend on the ratio h = r / q . This
result is particularly interesting since allows a further
simplification of the tracker (7). The result is formally
stated and proven in the following Proposition 2.
Proposition 2. Given a measured signal
y ( t ) = Acos(ot) + n(t) , the estimated frequency
6 ( f =) i , ( t / t ) obtained using the EKF Frequency
Tracker (8) depends upon the ratio h = r / q only.
Proof. Consider the time series ;(t / t ) , F(t), f ( t ) and K ( t )
obtained by means of the EKF Frequency Tracker (7)
when using the parameters q and r (and fed with y(t)). We
are now interested in understanding the relationships
between i ( f I t ) , F(t) P(t) and K(t) and the corresponding
time series, say F ( t l t ) , F ( t ) , F ( t ) and i ( t ) , obtained
when using = aq and 7 = ar (a>O)instead of q and r .
To this end, consider the Riccati Equation used in (7) given
by :
P ( / + l ) = ~ - ~ , ) [ P ~ / ) - f ( / ) H / ( H P (+/r). H / P ( / ) l F ’ ( / ) + q / :(8)
j’H
and plug in (8) aP(f+l)in place of f ( t + l ) , aP(t) in place of
P(t), aq in place of q, and ar in place of r :
@t -@t)H7 (hlxqt)H7 +w)-Ihlxqf)lp’(t)+CYq4 .
+ I) = qtkt)
After some simple manipulations of (9) we come to the
Riccati equation
P ( f+ 1) = F(f)[P(t)- P(t)H’ (HP(f)H7 + $1 HP(f)E’r(t)+ q I 3 , (10)
which actually coincides with (8). This means that if
9 = aq and F = ar (a>O) are used instead of q and r, the
resulting solution F ( t ) is equal to aP(t).
It is important to stress, however, that this is true under the
assumption that F(t) does not change if q and r are both
scaled by the same factor a, namely that F ( f )= F ( t ) . By
inspecting the structure of the EKF Frequency Tracker (7)
and that of the matrix F(t), it is easy to see that
F(f) = F ( t ) if z ( t / t ) = i ( t / t ) b’f . Moreover, K ( t ) = K ( t )
implies that Z ( t l t ) = i ( t / t ) . This can be easily proven
since
i(t/t)=f(i(t-~/t-~))+K(t)(y(t)-Hf(i(t-~/t-~))).
From this expression, it is apparent that, if g ( t ) = K ( t ) and
- -
;(t - 1 / t - I ) = ;(t - 1/ t - 1) , then i ( t l t ) =i ( t / t ) .
Since
the initial condition of the filter are chosen by the designer
at a value say X(0) , we have to take z ( 0 ) = i ( 0 ) = X(0) ,
and then, by induction, we can conclude that i ( t ) = K ( t )
implies % ( t / t ) = i ( t / t ) Vf . Thus, the following chain of
implications holds true:
-
i ( t ) =K(t)= i ( t l t ) = i ( f / t ) = F(t) = F ( t ) =3 F ( t ) =aP(t).
So far, we have to analyse the gain variation against the
scaling effect. To this end, recall that
K ( t ) = P(t)HT(HP(t)HT +,)-I . (1 1)
By plugging in (1 1) aP(t) in place of P(t), and ar in place
of r, after some manipulations we obtain:
K ( t ) = P ( t ) H T(HP(t)H + r)-’ .
The fact that (12) coincides with (1 1) means that
F ( t ) =aP(t)3 K ( t ) = K(t). Therefore, the following set of
circular implications holds:
~ ( t ) = K ( t ) ~ ~ ( t l t ) = ~ ( t / t ) == .F~( t() t- )p ( t ) = a P ( t )
=3 K(t)=K(t).
By inspecting the structure of the EKF (see (7)) we can
conclude that, if the initialisation of the filter is the same,
scaling the parameters q and r by the same positive real
number a has no effects on the state estimation. This means
,
that the estimated frequency & ( t ) obtained using the EKF
Frequency Tracker (7) depends on the ratio h = r / q only.
The above Proposition 2 is interesting since rigorously
.
allows a re-parameterisation of the EKF Frequency Tracker
(7). As a matter of fact, by conventionally setting q=l
(9) (hence h = r ), we obtain:
I i ( t / t )= f ( i ( t - l / r
&(f) = C i ( f / f )
- I ) ) + K(r)&r)
K ( t ) = P ( f ) H ‘ (HP(f)H7 +$I
- Hf(i(t- l / r - 1)))
P(r + 1) = F(i)[P(t)- P(r)H7 (HP(t)H7 + h)-lHP(r)b‘ ( r ) + 1;
where:
(13a)
1 ‘ I
cos(i,(t/t))i, ( t / t )- sin(i3(t/t))i2( t / t )
f ( i ( r / r ) )= sin(i3(t/t))il( t / t )+ cos(i,(t/t)P,(t/t)
.(13b)
i,(t/f)
H=[l 0 01 c=[o 0 I]
Tracker (1 3) is a significant simplification of (8); the vector
of the “tuneable” parameters of (13) is now constituted by a
single parameter il .
We conclude this section by presenting a numerical
example, which qualitatively shows the influence of the
parameter h on the tracker behaviour.
Example 2.
The signal considered in this example is an harmonic signal
embedded in white noise, characterised by a time-varying
frequency. Specifically, y ( t ) = s ( t ) + n(t) , t=1,2, ... 1000,
where n ( t ) = WN(O,O.Ol) and s ( t ) = cos(:t) for
t=1,2 ...500, s ( t ) = c o s ( g t ) for r=501,502 ... 1000.
In order to estimate the frequency of the harmonic signal
embedded in y ( t ) , the EKF Frequency Tracker (13), is
used. To assess the effect of h on the tracker behaviour,
three different values of h have been used: h=O, h=l, and
h=lO; the corresponding frequency estimation results are
displayed in Fig.2.
From Fig.2 some useful indications on the tracker
behaviour can be drawn. In particular, as for every on-line
frequency tracker (see e.g. [6], [12]), there is a trade-off
between the variance error of the estimated frequency and
the ability of the tracker to follow frequency variations (in
(12) the case considered herein we have considered the most
demanding situation, given by a large and sudden jump of
the frequency to be tracked). Specifically:
When h increases the variance error diminishes, as can
be seen by observing the decrement in the fluctuations
of the estimated frequency up to time 500; however,
2228
 the ability of the tracker to follow rapidly varying
frequency changes deteriorates.
When h decreases the variance error grows, but the
tracker is more robust towards frequency variations.
In particular notice from Fig.5 that if h=10 the frequency is
accurately estimated at steady-state, but the EKF gets
completely lost after the frequency jump. This problem can
be overcome using a lower value of h, at the price of a
worse noise rejection.
The correct choice of h clearly depends on the features of
the measured signal: if the frequency of the harmonic signal
is constant or slowly-varying a large value of h should be
used; on the other hand, if h is subject to large and sudden
variations, a small value of h should be used, in order to
guarantee robustness to the tracker.
25
frequency
2 m
05
of parameters tuning much easier and more transparent.
ACKNOWLEDGEMENTS
Paper supported by CARIPLO Foiindation for Scientific
Research. The authors are grateful to Professor R.R.
Bitmead, for his enlightening hints and suggestions.
REF E RENC E S
Anderson B.D.0, Moore J.B (1979). Optimal
Filtering. Prentice-Hall, Englewood Cliffs, NJ.
Bittanti S., Campi M., Savaresi S.M. (1997).
"Unbiased Estimation of a Sinusoid in Noise via
Adapted Notch Filters". Automatica, Vo1.33, n.2,
pp.209-2 15.
Bittanti S., Laub A.J., Willems J.C. (eds.), The Riccati
Equation. Springer-Verlag, 1991.
Handel P., Tichavsky P., Savaresi S.M. (1 998). "Large
error recovery for a class of frequency tracking
algorithms". International Journal on Adaptive Control

- 0 ' 3bO 4b0 Sbo

tl-
6k 7bo 6bo 9bo
I
1kO
and Signal Processing, Vol. 12, pp.417-436.
Hsu L., Ortega R., Damm G. (1999). "A Globally
frequency
Convergent Frequency Estimator". IEEE Trans. on
Automatic Control, Vo1.44, No.4, pp.698-713.
,.....
Kay S.M., Modern Spectral Estimation: Theory and
Applications. Prentice-Hall, 1988.
La Scala B., Bitmead R. (1996). "Design of an
05 Extended Kalman Filter Frequency Tracker". IEEE
-.L-. Trans. on Signal Processing, Vo1.44, n.3, pp.739-742.
treouenc"
La Scala B., Bitmead R., Quinn, B.G. (1996). "An
2 5 1 , I Extended Kalman Filter Frequency Tracker for High-
Noise Environments". IEEE Trans. on Signal
Processing, Vo1.44, n.2, pp.43 1-434.
1
La Scala B., Bitmead R., James M.R. (1995).
"Conditions for Stability of the Extended Kalman
05
Filter and Their Application to the Frequency Tracking
0
0 100 200 300 404
1"
500 6W 700 600 9W 1000 Problem 'I. Mathematics of Control, Signals and
Fig.2. Frequency tracking of the EKF using h=O, h=l, and Systems, Vo1.8, pp. 1-26.
h=10 (the dashed line indicates the true frequency). [IO] Quinn B.G, Femandes J.M. (1991). "A Fast Efficient
Technique for the Estimation of Frequency".
Biometrika, Vo1.78, pp.489-497.
4. CONCLUSIONS
[ 1 11 Pisarenko V.F. (1973). "The retrieval of harmonics
In this work the problem of estimating the frequency of an
from a covariance function". J.Roy.Astr., Vo1.33,
harmonic signal in noise, by means of an EKF Frequency
p.374-3 76.
Tracker, has been considered. The EKF Frequency Tracker
recently proposed in literature is characterised by a vector [12] Savaresi S.M. (1997). "Funnel Filters: a new class of
of three design parameters { 9 , r , ~ }In. this paper it is shown filters for frequency estimation of harmonic signals".
Automatica, Vo1.33, n.9, pp.1711-1718.
that, even in a noise-free setting, if E > O the EKF
Frequency Tracker (2) provides biased estimate of o , and [ 131 Song Y., Grizzle J.W. (1 995). "The Extended Kalman
that the performance of the EKF Frequency Tracker depend Filter as a Local Asymptotic Observer for Nonlinear
Discrete-Time Systems". Journal of Math. Systems,
on the ratio h = r / 9 only.
Estimation Control, Vo1.5, pp.59-78.
These results allow a simplification of the parameterisation
of the EKF Frequency Trackerand, and make the problem

2229