Download as pdf or txt
Download as pdf or txt
You are on page 1of 21

INTERNATIONAL JOURNAL OF ADAPTIVE CONTROL AND SIGNAL PROCESSING

Int. J. Adapt. Control Signal Process. 2003; 17: 763–783 (DOI: 10.1002/acs.774)

Exponential convergence of the Kalman filter based parameter


estimation algorithm

Liyu Caoa,n,y and Howard M. Schwartzb


a
Micro Optics Design Corp., 40 Rooney Crescent, Moncton, NB.E1E 4M3, Canada
b
Department of Systems and Computer Engineering, Carleton University, 1125 Colonel By Drive, Ottawa, On. K1S 5B6,
Canada

SUMMARY
In this paper we shall present a new method to analyse the convergence property of the Kalman filter based
parameter estimation algorithms. This method for convergence analysis is mainly based on some matrix
inequalities and is more simple than some of the existing approaches in the literature. This method can
simultaneously provide both lower and upper bounds on the exponential convergence rate as the functions
of bounds of the related matrices, such as the covariance matrices. A simulation example is provided to
illustrate the convergence property of the Kalman filter based algorithms. Copyright # 2003 John Wiley &
Sons, Ltd.

KEY WORDS: Kalman filter; recursive parameter estimation; exponential convergence; convergence rate

1. INTRODUCTION

The exponentially weighted least squares (EWLS) parameter estimator and the Kalman filter
based (KFB) estimator are widely used in adaptive control systems to track the time-varying
parameters. As has been shown in Reference [1], exponential convergence of these kinds of
algorithms is of fundamental importance for their tracking ability. Further, one is often
concerned not only with the property of exponential convergence but also with the exponential
convergence rate, which is a measure of the transient tracking performance.
The convergence analysis is usually carried out for the ideal case where the system model is
assumed to be known, the data is free from noise and the parameters are time invariant, as has
been done in References [2–4]. For the EWLS algorithm with a forgetting factor, this kind of
analysis has been carried out in an elegant way in References [4, 5] and has shown that the
EWLS algorithm is exponentially convergent and the convergence rate is mainly determined by
the forgetting factor. For the KFB parameter estimator, although it has been known for a very
long time that it is exponentially convergent or stable, but it seems that an explicit
representation of the convergence rate in terms of the estimator’s parameters has not been
established. This could probably be explained by the key difference of the two kinds of

n
Correspondence to: Dr. Liyu Cao, Micro Optics Design Corp., 40 Rooney Crescent, Moncton, NB. E1E 4M3, Canada
y
E-mail: lcao@micro-optics.com

Received 23 April 2002


Copyright # 2003 John Wiley & Sons, Ltd. Accepted 1 November 2002
764 L. CAO AND H. M. SCHWARTZ

algorithms. In the EWLS algorithm the tracking ability is defined or controlled by a scalar
forgetting factor; while in a Kalman filter, the scalar forgetting factor is replaced by a positive
definite matrix, which considerably complicates the mathematical analysis as to its convergence
rate.
The KFB algorithms considered in this paper are represented by the following equations:
y# t ¼ y# t1 þ Kt ðyt  jTt y# t1 Þ ð1Þ

Pt1 jt
Kt ¼ ð2Þ
rt þ jTt Pt1 jt

Pt1 jt jTt Pt1


Pt ¼ Pt1  þ Qt ð3Þ
rt þ jTt Pt1 jt

where y# t represents the estimated n-dimensional system’s parameter vector, and jt is the
n-dimensional regression vector and is assumed to be deterministic. The designer-adjustable
parameters in the algorithm are the scalar rt and the positive definite matrix Qt :
As has been shown in Reference [6], stability of the above algorithms can be analysed by using
the general Kalman filter theory for the case where the regression vector jt is deterministic in
nature [7]. Further, it is shown in Reference [6] that the above algorithm is uniformly
exponentially stable if the sequence jt is persistently exciting and the positive definite matrix
sequence Qt is bounded. This is a very important result in the stability theory of parameter
estimation algorithms, because, as indicated in Reference [6], it is applicable to a very general
glass of algorithms including the Kalman filter. In addition, it can be used to prove the
exponential stability of an algorithm even in the case where the unknown parameters are time-
varying. In References [7, 8] a more challenging problem is addressed, where the stability
property of the KFB algorithm is intensively analysed based on a stochastic excitation
condition. Although the general Kalman filter theory is very powerful, it is difficult to use it to
provide a quantitative description on the exponential convergence rate, which is very useful in
algorithm analysis and design. Another kind of approach to the exponential stability(conver-
gence) is algorithm-specific and attempts to establish the exponential convergence in a relatively
simple and direct way. This is achieved by using a Lyapunov-type function to derive a recursive
equation for the parameter estimation error, and then to show that the parameter estimation
error is exponentially decreasing. While proving exponential stability of an algorithm, this kind
of approach also provides estimates on the convergence rate from the Lyapunov function based
analysis, although sometimes the estimates may be very conservative. The approach that will be
used in this paper belongs to this category. In Reference [2] a Lyapunov function based analysis
is carried out for the EWLS algorithm, which led to an apparent description on the convergence
rate. In Reference [3] the same kind of analysis is applied to the so-called modified directional
forgetting algorithm. From Reference [3] it can be seen that the modified directional forgetting
algorithm has the same form as the Kalman filter based algorithms but with a special choice on
the parameter rt and on the matrix Qt : Reference [3] is the only article found that contributes to
the convergence analysis of a Kalman filter like algorithm using a Lyapunov function. Although
the major objective of Reference [3] is to provide a proof of exponential convergence of the
modified directional forgetting algorithm, a bound on the convergence rate can be derived from
[3] as will be shown in Section 5.

Copyright # 2003 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2003; 17:763–783
EXPONENTIAL CONVERGENCE OF THE KALMAN FILTER 765

In this paper, we will develop a new analytical method using the Lyapunov function that can
be used to establish the exponential convergence of the KFB algorithms based on some matrix
inequalities. The basic assumptions for this method are the boundedness of some matrices.
Basically we address the same problem as that addressed in Reference [3], but our approach is
completely different and more apparent. The fundamentals of our approach are several simple
matrix inequalities. Based on these matrix inequalities we are able to establish the exponential
convergence of the KFB algorithms by simple matrix operations. The key step of the analysis is
to express the Lyapunov function in terms of the sum of the covariance matrix Pt and a newly
defined positive definite matrix. As will be shown in Section 3, the proof of exponential
convergence is made easy by simply applying these matrix inequalities to the above Lyapunov
function. Our method can simultaneously provide both upper and lower bounds on the
exponential convergence rate in terms of the algorithm’s matrices. Compared with the available
results in the literature, our method provides a new kind of description of the convergence rate.
In addition, as will be shown in Section 5, the newly developed bounds on the convergence rate
are more concise and easy to use than the similar results in the literature.
The paper is organized as follows. Basic assumptions in the analysis and the preliminary
lemmas on some matrix inequalities are given in Section 2. In Section 3, based on the matrix
inequalities, the exponential convergence of the KFB algorithms is established under the
condition that two matrices (one is Pt ) in the Lyapunov function are bounded. It is shown in
Section 4 that these two matrices are bounded under the condition of persistent excitation. In
Section 5 the bounds developed in Section 3 are compared with the other bounds available in the
literature. In Section 6 a simulation example is provided to verify the theoretical analysis and to
show the convergence property of a KFB algorithm. Finally concluding remarks are given in
Section 7.

2. BASIC ASSUMPTIONS AND PRELIMINARY LEMMAS

In this section the basic assumptions and some preliminary lemmas used in the analysis are
summarized.
The data model is assumed to be ideal and is given by
yt ¼ jTt y0 ð4Þ
where y0 is a constant vector.
The matrix sequence Qt is assumed to satisfy
q1 I  Qt  q2 I ð5Þ
where q1 and q2 are some scalars and satisfy q2  q1 > 0: The inequality (5) shows that Qt is
positive definite and bounded.
The regression vector sequence jt is assumed to be persistently exciting satisfying
X
tþs
aI  ji jTi  bI ð6Þ
i¼tþ1

where 05a5b51 and s > 0 is an integer. The above inequality also means that jt is bounded
from above. That is, there is a scalar c > 0 such that jjt j  c for all t:

Copyright # 2003 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2003; 17:763–783
766 L. CAO AND H. M. SCHWARTZ

It is assumed that the scalar rt in (3) is equal to a constant r in the following analysis. For the
case where this is not true, the analysis is still effective as long as rt is bounded.
Some facts on the positive definite matrices are frequently used in this paper and are
summarized in the following lemmas.

Lemma 2.1
Let A be a positive definite matrix, and let lm ðAÞ and lM ðAÞ denote its minimum and maximum
eigenvalue respectively. Then
lm ðAÞI  A  lM ðAÞI

Proof
Refer to Theorem 7.7.3 in Reference [9]. &

Lemma 2.2
Assume that n  n matrices A and B are positive definite, and A  B: Let the eigenvalues of A
and B be arranged in the same order. Then
lk ðAÞ  lk ðBÞ
for k ¼ 1; 2; . . . ; n:

Proof
Refer to Corollary 7.7.4 in Reference [9]. &

Lemma 2.3
Let n  n matrices A and B be real symmetric: If A  B; then
C T AC  C T BC
for any n  m real matrix C:

Proof
Refer to Observation 7.7.2 in Reference [9]. &

The matrix inversion lemma is also used in this paper and is given in the following.

Matrix Inversion Lemma


Assume that the n  n matrix A and m  m matrix C are non-singular, and B is an n  m matrix,
D is an m  n matrix. Then
ðA þ BCDÞ1 ¼ A1  A1 BðC 1 þ DA1 BÞ1 DA1 ð7Þ

The key idea in the analysis is to use some matrix inequalities for the sum of two positive
definite matrices. The matrix inequalities are presented in the following lemma.

Copyright # 2003 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2003; 17:763–783
EXPONENTIAL CONVERGENCE OF THE KALMAN FILTER 767

Lemma 2.4
Let the matrices A and B be positive definite and satisfy
a1 I  A  a2 I ð8Þ

b1 I  B  b2 I ð9Þ
for some scalars 05a1  a2 and 05b1 5b2 : Then
1 1
A1  ðA þ BÞ1  A1 ð10Þ
1 þ b2 =a1 1 þ b1 =a2

Proof
We have
b1
B  b1 I ¼ a2 I
a2
b1
 A ð11Þ
a2
Then from the above inequality we get
 
b1
AþB 1þ A ð12Þ
a2
From the fact that if two positive definite matrices satisfy A1  A2 > 0; then A1 1
2  A1 ; we have
1
ðA þ BÞ1  A1 ð13Þ
1 þ b1 =a2
In the same way it can be proven that
1
ðA þ BÞ1  A1 ð14Þ
1 þ b2 =a1
&

If a1 is equal to the minimum eigenvalue of A; then a1 is called the tight lower bound of A in
the sense that the matrix A  a1 I is positive semidefinite rather than positive definite. Similarly,
if a2 is equal to the maximum eigenvalue of A; then a2 is the tight upper bound of A: If ai ; i ¼ 1; 2
and bi ; i ¼ 1; 2 are tight bounds of A and B; then (10) could be a rather tight inequality in the
sense that there exist some matrices A and B such that the matrices ðA þ BÞ1  ð1 þ b2 =a1 Þ1 A1
and ð1 þ b1 =a2 Þ1 A1  ðA þ BÞ1 are positive semidefinite rather than positive definite.

3. EXPONENTIAL CONVERGENCE OF THE KFB ALGORITHM

In this section we propose a simple method based on Lemma 2.4 to prove the exponential
convergence of the KFB algorithms. For this purpose, we need the boundedness of the related
matrices.

Copyright # 2003 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2003; 17:763–783
768 L. CAO AND H. M. SCHWARTZ

First, we assume that the covariance matrix Pt is bounded both from below and above, that is,
there exist 05a5b51 such that
aI  Pt  bI ð15Þ

One can show that Pt has a lower bound as follows. Define


Pt1 jt jTt Pt1
P% t1 ¼ Pt1  ð16Þ
r þ jTt Pt1 jt
By using the matrix inversion lemma we can obtain the inverse of P% t1 as follows (setting
1
A ¼ Pt1 ; B ¼ jt ; C ¼ 1=r; D ¼ jTt in (7))
P% 1 1 1 T
t1 ¼ Pt1 þ r jt jt ð17Þ

Thus (3) can be written as


Pt ¼ P% t1 þ Qt ð18Þ

From (17) we can see that if Pt1 > 0 then P% t1 > 0: From (18) it can also be seen that if P% t1 > 0
then Pt > 0: Therefore, if P0 > 0 then P% t > 0 for all t  0: Thus, an obvious lower bound for Pt is
given by
P t > Q t  q1 I ð19Þ

However, q1 is not a tight bound for Pt because (19) is a strict inequality.


We will prove that Pt is also bounded from above in the next section. In the following, we
establish the exponential convergence of the KFB algorithm based on Lemma 2.4.
Define the parameter estimation error as
y* t ¼ y# t  y0 ð20Þ

Then from (1) and (4) we can get


y* t ¼ y# t1  y0 þ Kt jTt ðy0  y# t1 Þ
¼ ðI  Kt jTt Þy* t1
 
Pt1 jt jTt
¼ I y* t1
r þ jTt Pt1 jt
 
Pt1 jt jTt Pt1 1 *
¼ Pt1  P yt1
r þ jTt Pt1 jt t1
1 *
¼ P% t1 Pt1 yt1 ð21Þ

Define the following Lyapunov function


Vt ¼ y* Tt Pt1 y* t ð22Þ

Substituting (21) into (22) we can get


Vt ¼ y* Tt1 Pt1
1 % 1 *
Pt1 Pt1 P% t1 Pt1 yt1 ð23Þ

Copyright # 2003 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2003; 17:763–783
EXPONENTIAL CONVERGENCE OF THE KALMAN FILTER 769

First we take a look at the product P% t1 Pt1 P% t1 : From (18) we have

P% t1 Pt1 P% t1 ¼ P% t1 ðP% t1 þ Qt Þ1 P% t1


¼ ½P% 1 % % 1 1
t1 ðPt1 þ Qt ÞPt1 

¼ ðP% 1 % 1 % 1 1
t1 þ Pt1 Qt Pt1 Þ

Furthermore we get
1 %
Pt1 Pt1 Pt1 P% t1 Pt1
1 1 % 1
¼ Pt1 ðPt1 þ P% 1 % 1 1 1
t1 Qt Pt1 Þ Pt1

¼ ðPt1 P% 1 % 1 % 1
t1 Pt1 þ Pt1 Pt1 Qt Pt1 Pt1 Þ
1

¼ ðPt1 þ r1 Pt1 jt jTt Pt1 þ MtT Qt Mt Þ1 ð24Þ


where Mt is given by
Mt ¼ P% 1 1 T
t1 Pt1 ¼ I þ r jt jt Pt1 ð25Þ
Define
Ot ¼ r1 Pt1 jt jTt Pt1 þ MtT Qt Mt ð26Þ
Noting that Mt is non-singular, based on Lemma 2.3 one can see that if Qt > 0 then Ot > 0: Thus
(24) can be written as
1 %
Pt1 Pt1 Pt1 P% t1 Pt1
1
¼ ðPt1 þ Ot Þ1 ð27Þ
It can be proven that (refer to the next section) the matrix Ot given in (26) is bounded in the
following way
gI  Ot  dI ð28Þ
for some scalars d > g > 0: Let lM ðÞ and lm ðÞ denote the maximum and minimum of the
eigenvalues of a matrix respectively. Define
lm ðOt Þ
m1 ¼ min ð29Þ
t lM ðPt1 Þ

lM ðOt Þ
m2 ¼ max ð30Þ
t lm ðPt1 Þ

ð31Þ
Then we have the following lemma.

Lemma 3.1
Assume that Pt satisfies (15) and Ot satisfies (28). Then we have
g d
 m1 5m2  ð32Þ
b a

Copyright # 2003 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2003; 17:763–783
770 L. CAO AND H. M. SCHWARTZ

Proof
Assume that at t ¼ t1 the ratio lm ðOt Þ=lM ðPt1 Þ reaches its minimum. Then we have
lm ðOt1 Þ
m1 ¼
lM ðPt1 1 Þ
g
 ð33Þ
b

Similarly, we can get the following inequality


d
m2  ð34Þ
a

It is obvious that m1 5m2 : Combining (33) and(34) we get (32). &

Based on (23) and (27), the Lyapunov function Vt can be expressed as


Vt ¼ y* T ðPt1 þ Ot Þ1 y* t1
t1 ð35Þ
Since Ot > 0; from the above equation we get
1 *
Vt 5y* Tt1 Pt1 yt1 ¼ Vt1 ð36Þ
which shows that the positive definite function Vt is monotonically decreasing and hence tends to
zero as t ! 1: The Lyapunov function expressed in the form of (35) is the main contribution of
this paper and is of fundamental importance in the analysis of the KFB algorithms. First, based
on this expression it is direct to prove that the KFB algorithm is exponentially convergent if
both Pt and Ot are bounded. Furthermore, based on Lemma 2.4 we can obtain an expression on
the convergence rate at which the tracking error y* t tends to zero.

Theorem 3.1
Assume that Qt satisfies (5), jt satisfies (6) and the matrix Ot satisfies (28). Then the Kalman
filter based algorithms are exponentially convergent, and the convergence rate is bounded by the
following inequality
a b
V0  jy* t j2  V0 ð37Þ
ð1 þ m2 Þt ð1 þ m1 Þt
where V0 ¼ y* T P 1 y* 0 :
0 0

Proof
From Lemma 2.1 we have lm ðPt ÞI  Pt  lM ðPt ÞI and lm ðOt ÞI  Ot  lM ðOt ÞI: Then applying
Lemma 2.4 we have
1
ðPt1 þ Ot Þ1  P 1
1 þ lm ðOt Þ=lM ðPt1 Þ t1
1
 P 1
1 þ m1 t1

Copyright # 2003 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2003; 17:763–783
EXPONENTIAL CONVERGENCE OF THE KALMAN FILTER 771

Therefore,
Vt ¼ y* Tt1 ðPt1 þ Ot Þ1 y* t1
1 * T 1 * 1
 y P yt1 ¼ Vt1
1 þ m1 t1 t1 1 þ m1

Repeating the above procedure, we have


1
Vt  V0
ð1 þ m1 Þt

From Pt  bI; we have Vt  b1 y* Tt y* t : Thus we get


b
jy* t j2  V0 ð38Þ
ð1 þ m1 Þt

From Lemma 3.1 we know that m1 > 0: Thus (38) shows that the estimate error y* t tends to zero
exponentially fast.
Similarly, we can get
a
jy* t j2  V0 ð39Þ
ð1 þ m2 Þt
Inequality (37) follows from (38) and (39). &

Remark 3.1
From the proof of Theorem 3.1, we see that based on (35) it is straight-forward to get the
inequality Vt  ð1=ð1 þ m1 ÞÞVt1 ; which directly leads to the conclusion of exponential
convergence. While in Reference [3] the convergence analysis is basically based on the
inequality Vt  Vt1 ; and much more complex derivations are needed to reach the conclusion of
exponential convergence. This apparent difference shows the analytical advantage of the method
developed in this paper.
Theorem 3.1 not only shows the algorithm is exponentially convergent, but also
represents a quantitative description on the convergence rate by giving both its lower
and upper bounds in terms of the algorithm matrices. Theorem 3.1 indicates that the
convergence rate is bounded by the ratios of the eigenvalues of Pt and Ot : However,
since Ot is a complex function of Pt ; Qt and jt ; it is difficult to evaluate the ratios m1 and m2 : To
overcome this difficulty, one may use the following relaxed bounds obtained directly from
Lemma 3.1,
a * 2 b
t V0  jyt j  V0 ð40Þ
ð1 þ d=aÞ ð1 þ g=bÞt
However, except the case where lm ðOt1 Þ ¼ g; lM ðPt1 Þ ¼ b for some t1 ; the number m1 may be
much larger than g=b: Similarly, except some specific case the number m2 may be much smaller
than d=a: Therefore, the bounds given in (40) may be very conservative.
Since Ot is a complex function of Pt and Qt ; (37) does not tell much about how the
convergence rate is affected by Pt and Qt : In the following, we will derive an upper bound for the
convergence rate which is directly given in terms of Pt and Qt :

Copyright # 2003 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2003; 17:763–783
772 L. CAO AND H. M. SCHWARTZ

Theorem 3.2
Under the same assumptions as in Theorem 3.1, the convergence rate of the KFB algorithm is
bounded by
a
jy* t j2  V0 ð41Þ
ð1 þ m3 Þt
where m3 is defined by
l2m ðPt1 Þlm ðQt Þ
m3 ¼ min ð42Þ
t l3M ðPt1 Þ

Proof
From (26) and Lemma 2.1 as well as Lemma 2.3 we have
Ot  MtT Qt Mt
 lm ðQt ÞMtT Mt
¼ lm ðQt ÞPt1 ðP% 1 2
t1 Þ Pt1 ð43Þ
From Lemma 2.1 we also have
ðP% 1
t1 Þ
2
 lm ½ðP% 1 2 2 % 1
t1 Þ I ¼ lm ðPt1 ÞI

From (17) we see that P% 1 1 % 1 1


t1  Pt1 ; and by Lemma 2.2 we have lm ðPt1 Þ  lm ðPt1 Þ: Therefore,
we have
ðP% 1 2 2 1
t1 Þ  lm ðPt1 ÞI

and therefore by Lemmas 2.3 and 2.1,


MtT Qt Mt  lm ðQt Þl2m ðPt1
1 2
ÞPt1
 lm ðQt Þl2m ðPt1
1 2
Þlm ðPt1 ÞI
lm ðQt Þl2m ðPt1 Þ
¼ I ð44Þ
l2M ðPt1 Þ
Then based on Lemma 2.4, we have
ðPt1 þ Ot Þ1  ðPt1 þ MtT Qt Mt Þ1
1 1
 Pt1 ð45Þ
lm ðQt Þl2m ðPt1 Þ
1þ l3M ðPt1 Þ

1
 P 1 ð46Þ
1 þ m3 t1
The rest of the proof is exactly the same as that for Theorem 3.1. &

Copyright # 2003 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2003; 17:763–783
EXPONENTIAL CONVERGENCE OF THE KALMAN FILTER 773

Compared with (37), the upper bound given in (41) bounds the convergence rate directly in
terms of the eigenvalues of Qt and Pt1 ; thus provides the direct insights into how the upper
bound of the convergence rate is affected by the design parameter Qt as well as the matrix Pt :

Remark 3.2
Theorem 3.2 can directly be used to give an upper bound on the convergence rate of the
directional forgetting algorithm discussed in Reference [3] by setting Qt ¼ dI: Such a bound is
much simpler in form than the corresponding bound in Reference [3], as will be shown in
Section 5.

4. THE BOUNDEDNESS OF Pt AND Ot

In this section, we will give the bounds of Pt and Ot to prove that the assumptions in Theorem
3.1 are true. Since as shown in (40) these bounds may also be used to provide an approximate
range for the rate of exponential convergence, we shall give the tightest bounds as much as
possible.
Recall that Pt can be written as (see (18))
Pt ¼ P% t1 þ Qt
From the above equation, it is expected that the bounds of Pt have the form of
ðp1 þ q1 ÞI  Pt  ðp2 þ q2 ÞI ð47Þ
where qi > 0; i ¼ 1; 2 are the bounds on the eigenvalues of Qt as defined in (5), and pi > 0; i ¼ 1; 2
are the bounds on the eigenvalues of P% t : The matrix P% t is determined by Pt (refer to (16)), as such
it is generally difficult to get tight bounds on P% t
As stated in Section 3, q1 is a lower bound for Pt but not a tight one, because of the effect of p1
in (47) is completely ignored. Here we provide another lower bound on Pt having the form of
p1 þ q1 and being tighter than q1 :
From (18) and (17) we have
 1 1
Pt ¼ Pt1 þ r1 jt jTt þQt ð48Þ
From the above equation we have Pt > Qt  q1 I for t ¼ 1; 2; . . . : Assume P0  q1 I: Then we
have
1 1 c2
Pt1 þ r1 jt jTt 5 Iþ I
q1 r
for t ¼ 1; 2; . . . : From (48) we get
Pt > a1 I
 1
1 c2
a1 ¼ q 1 þ þ ð49Þ
q1 r
We see that a1 > q1 : Therefore, a1 is tighter than q1 :
Next we show Pt is bounded from above and give an upper bound on Pt :

Copyright # 2003 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2003; 17:763–783
774 L. CAO AND H. M. SCHWARTZ

Lemma 4.1
Assume that the sequences Qt and jt satisfy (5) and (6) respectively. Furthermore assume that
Pt  aI for t  0: Then there exists a positive number b such that the covariance matrix Pt of
the Kalman filter based algorithm satisfies
Pt  bI ð50Þ

Lemma 4.1 can be proved by using the approach of Reference [10] (refer to the proof of
Theorem 1 in Reference [10]). Here, we use a different approach that can lead to an upper bound
on Pt represented explicitly as a function of the algorithm’s parameters.

Proof
Define the matrix Rt as
R1 %
t ¼ Pt1 ¼ Pt  Qt ð51Þ
From (17) and (18) we have
1
Rt ¼ Pt1 þ r1 jt jTt
1
¼ ðR1
t1 þ Qt1 Þ þ r1 jt jTt ð52Þ
On the other hand, we also have
1
Rt1 ¼ Pt2 þ r1 jt1 jTt1 ð53Þ
1
Note that Pt2  a1 I and jjt1 j  c for t  2: Thus, from the above equation we get
Rt1  a1 I þ r1 c2 I ð54Þ
or
R1
t1  ða
1
þ r1 c2 Þ1 I ð55Þ
for t  2: Applying Lemma 2.4 to (52) we can get
Rt  mRt1 þ r1 jt jTt ð56Þ
where m is defined by
1
m¼ ð57Þ
1 þ q3
and q3 is given by
q2 q2 2
q3 ¼ þ c ð58Þ
a r
Obviously we have q3 > 0 and m51:
From (56) we have
Rtþs  mRtþs1 þ r1 jtþs jTtþs ð59Þ
Recursively applying the above inequality we can get
X
s
Rtþs  ms1 Rtþ1 þ r1 msi jtþi jTtþi ð60Þ
i¼2

Copyright # 2003 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2003; 17:763–783
EXPONENTIAL CONVERGENCE OF THE KALMAN FILTER 775

From (56) we have Rtþ1 > r1 jtþ1 jTtþ1 ; then from the above inequality we can get
X
s
Rtþs > r1 msi jtþi jTtþi
i¼1

X
s
¼ r1 ms1 m1i jtþi jTtþi
i¼1

X
s
 r1 ms1 jtþi jTtþi ð61Þ
i¼1

Using the condition of persistent excitation given in (6) we have the following
inequalityz
a
Rtþs > ms1 I ð62Þ
r
Assume s  2: Then from (51) and the above inequality we can get for t > s  2
Pt ¼ R1
t þ Qt
r
5 ðq2 þ mðs1Þ ÞI ð63Þ
a

ð64Þ
It is obvious that Pt is bounded from above for t  2: Thus we conclude that for all t  0 there
exist a positive number b such that
Pt  bI &

From the proof of Lemma 4.1, we see that for t > s we have b5b1 ; where b1 is
defined by
r
b1 ¼ q2 þ mðs1Þ
a
r
¼ q2 þ ð1 þ q3 Þs1 ð65Þ
a
The scalar b1 can be used as an approximate upper bound for Pt : From (58) and (65) we see that
b1 is proportional to the upper bound q2 of Qt ; but inversely proportional to the lower bound a
of Pt : Since b1 is affected by a; to make a good estimation on b we need a good estimation on a:
It should be noted that although b1 has the form of p2 þ q2 ; it is not a tight bound of Pt since
(63) is a strict inequality.
Now we turn to the boundedness of Ot to prove the inequality (28).

z
The approach used here is quite similar to the approach used in Reference [4] to derive a lower bound for Pt1 in the
EWLS algorithm. However, the lower bound given by (62) is larger than the corresponding one obtained by the
approach of Reference [4], which is proportional to the ratio of ðms1  ms Þ=ð1  ms Þ ¼ ms1 ð1  mÞ=ð1  ms Þ: In the other
words, the approach in this paper can be used to derive a tighter upper bound for the matrix Pt in the EWLS algorithm.

Copyright # 2003 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2003; 17:763–783
776 L. CAO AND H. M. SCHWARTZ

Lemma 4.2
Assume that Qt and jt satisfy (5) and (6) respectively. Then there exist positive scalars g and d
such that
gI  Ot  dI

Proof
From the proof of Theorem 3.2 we have
Ot  MtT Qt Mt
lm ðQt Þl2m ðPt1 Þ
 I
l2M ðPt1 Þ
q1 a2
 I
b2
From the above inequality we see that there exist a positive number g such that Ot  gI and
q1 a2
g ð66Þ
b2
On the other hand, from (26) and Lemma 2.1 as well as Lemma 2.3 we have
Ot  r1 jPt1 jt j2 I þ q2 MtT Mt
 r1 jPt1 jt j2 I þ q2 lM ðMtT Mt Þ ð67Þ
Note that lM ðMtT Mt Þ
¼ s2M ðMt Þ;
where sM ðÞ denotes the maximum singular value of a matrix.
Also note that the the maximum singular value of a matrix is equal to its spectral norm.
Applying the triangle inequality on a matrix norm to Mt we get
sM ðMt Þ ¼ sM ðI þ r1 jt jTt Pt1 Þ
 1 þ r1 sM ðjt jTt Pt1 Þ
From the definition of matrix singular value, we have
s2M ðjt jTt Pt1 Þ ¼ jjt j2 lM ðPt1 jt jTt Pt1 Þ
¼ jjt j2 jTt Pt1
2
jt
 b2 c4
Therefore, we get the following bound for sM ðMt Þ
sM ðMt Þ  1 þ r1 bc2
Then we have
Ot  r1 jPt1 jt j2 I þ q2 s2M ðMt Þ
"  2 #
b 2 c2 bc2
 þ q2 1 þ I ð68Þ
r r

Copyright # 2003 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2003; 17:763–783
EXPONENTIAL CONVERGENCE OF THE KALMAN FILTER 777

The above inequality indicates that there exists a positive number d such that Ot  dI and
 2
b2 c2 bc2
d þ q2 1 þ ð69Þ
r r
Thus the lemma is proven. &

5. COMPARISONS WITH THE OTHER BOUNDS IN THE LITERATURE

In this section, the bounds of the convergence rate provided in Section 3 will be compared with
the other bounds that can be obtained in the literature.
Theorem 3.1 is the main result of the paper, which shows that the KFB algorithm is
exponentially convergent and also provides bounds on the convergence rate. Compared with the
other results available in the literature, Theorem 3.1 has some unique properties. First, it
simultaneously gives both the upper and lower bounds in a simple way, while in the literature
usually only upper bounds are available. Second, the bounds provided by Theorem 3.1 are given
in terms of the maximum or minimum ratio of the eigenvalues of the matrices Pt and Ot as
shown in (37),(29) and (30), and the matrix Ot originally defined in (26) is firstly introduced into
the analysis of exponential convergence. This is a new feature of the bounds and no similar
results have been found in the literature. The bounds given in (37) are simple in form, the upper
or lower bound is presented in terms of only one parameter: m1 or m2 : However, since m1 (or m2 )
is defined as the minimum (or maximum) ratio of the eigenvalues of the matrices Pt and Ot (refer
to (29) and (30)), and since Ot appears as a complex function of Pt ; Qt and jt ; it is very difficult to
evaluate m1 (or m2 ). Therefore, the bounds given in Theorem 3.1 are simple in form, but their
implications are complex. As will be shown by the simulation example in Section 6, determining
the convergence rate of the Kalman filter is a complex problem, this is why m1 and m2 are difficult
to analyse. When one attempts to replace the bounds given in (37) with more insightful bounds
in terms of the algorithm’s parameters, one should keep in mind that these bounds may become
very conservative.
In the following, more detailed comparisons will be made to the bounds based on the results
of References [2, 3]. Since there are no comparable bounds available in these references to the
bounds given by Theorem 3.1, comparisons will be made based on the bound of Theorem 3.2
and the corresponding bounds in References [2, 3].
To simplify the problem we will compare the bounds for the KFB algorithm with the specific
choice Qt ¼ dI: This choice is consistent with the algorithms considered in References [2, 3].

5.1. The bound based on Theorem 3.2


When Qt ¼ dI; (42) becomes
l2m ðPt1 Þ
m3 ¼ d min ð70Þ
t l3M ðPt1 Þ
Then from Theorem 3.2 the convergence rate of the KFB algorithm, r; is bounded by
1
r ð71Þ
1 þ m3

Copyright # 2003 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2003; 17:763–783
778 L. CAO AND H. M. SCHWARTZ

Since a  lm ðPt1 Þ and b  lM ðPt1 Þ; from (70) we have

da2
m3  ð72Þ
b3

Therefore, the convergence rate is bounded by

1
r ð73Þ
1 þ da2 =b3

5.2. The bound based on Reference [3]


Based on the proof of Lemma in Reference [3], the convergence rate is bounded by
r  ð1  adÞ1=s ð74Þ
where s is defined in (6), a is defined in (15), and d is defined by
k2
d¼ ð75Þ
1 þ r%
In (75) r% is defined as
jTt Pt1 jt  r% ð76Þ
and k is given by
rffiffiffi
a 1
k¼ ð77Þ
s 1 þ 2bsb
where a and b are defined in (6), and b is defined in (15).
Substituting (75) and (77) into (74) leads to
 1=s
aa
r 1 ð78Þ
sð1 þ r%Þð1 þ 2bsbÞ2
Equation (78) is the upper bound of the convergence rate based on the results of
Reference [3].

5.3. The bound based on Reference [2]


In Reference [2] (Theorem 3.2), exponential convergence of a general forgetting algorithm is
established under the conditions of persistent excitation and the bounded covariance matrix Pt :
It is shown in [2] that the Kalman filter can be viewed as a member of the family of the general
forgetting algorithm. Therefore, we can derive a bound for the convergence rate of the Kalman
filter based on the result of Reference [2]. From the proof of Theorem 3.2 of [2], we can get the
following bound for the Kalman filter
a
r1 2 2 ð79Þ
kcmax s ð1 þ c2 bÞ

Copyright # 2003 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2003; 17:763–783
EXPONENTIAL CONVERGENCE OF THE KALMAN FILTER 779

where k and cmax are defined by


1
k¼ þ c2 ð80Þ
a
cmax ¼ maxð1; c2 bÞ ð81Þ
and c is the upper bound of the regression vector jt as defined in Section 2.
Now we can make comparisons between (73) and (78) as well as (79). First we should note these
bounds share some common property, that is, all of them are proportional to b; the upper bound
of Pt ; and are inversely proportional to a; the lower bound of Pt : Therefore, all of the bounds
indicate that the Kalman filter may converge faster when a becomes larger or b becomes smaller.
The differences between the bound of (73) and the bound of (78) as well as (79) are also
apparent. The bound of (73) has a simpler form than (78) and (79) in the sense that it depends only
on three parameters. In particular (73) depends only on the matrices Pt and Qt ; while the bounds
given in (78) and (79) depend on Pt and the excitation condition as well. The simplicity of (73)
comes from the simplicity of the method developed in the paper. Note that both a and b depend
on the excitation condition, and particularly b is a complex function of the excitation parameters,
as has been shown in Section 4. Therefore, if the bounds given in (78) and (79) are represented in
terms of the excitation parameters, (78) and (79) will become more complex functions of the
excitation parameters. Simplicity is one of the properties of the bounds developed in this paper.
Another major difference between (73) and (78) as well as (79) is that (73) depends on d; while
(78) and(79) do not depend on d: As has been shown in Section 4, both a and b depend on d:
Therefore, (78) and(79) actually depend on d: We know that d represents the ‘size’ of the matrix
Qt ¼ dI; and the function of Qt is exactly the same as that of the forgetting factor in the
exponentially weighted least squares(EWLS) algorithm. Therefore, just as the forgetting factor
the number d plays a major role in determining the convergence rate. The bound in (73) reflects
this fact in an apparent and direct way, while the bounds in (78) and (79) reflect this fact in an
indirect and implicit way.
Finally, in order to make the bound of (79) be useful, one should ensure that the following
inequality
a
51 ð82Þ
kc2max s2 ð1 þ c2 bÞ
is true. Otherwise, (79) will be meaningless. The same arguments are applicable to (78). On the
other hand, the bound in (73) is meaningful as long as da > 0 and b51; which are true by
assumption. The inequality (73) itself is enough to show exponential convergence. However, this
is not true for (78) and (79). To make the inequality (78) (or (79)) meaningful, one must provide
it with good enough estimates on the algorithm’s parameters so that (82) is true.

6. SIMULATION EXAMPLE

Theorem 3.1 provides the lower and the upper bounds of the convergence rate of a KFB
algorithm. In this section we will show, through a simulation example, that the convergence rate
of a KFB algorithm is not only time-varying, but also change as the initial conditions change.
This fact suggests that generally we cannot describe the convergence rate of a KFB algorithm by
a single number; given its lower and upper bound as shown in Theorem 3.1 is a natural way to
describe the speed of the algorithm.

Copyright # 2003 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2003; 17:763–783
780 L. CAO AND H. M. SCHWARTZ

Figure 1. Input (up) and output (down) of the system.

We consider the following second-order system,


yðtÞ þ a1 yðt  1Þ þ a2 yðt  2Þ ¼ b1 uðt  1Þ þ b2 uðt  2Þ
The parameter vector y ¼ ½a1 a2 b1 b2 T is needed to be estimated. In the simulation, y is a
constant vector and given by
y ¼ ½1 0:8 1 0:2T
The input and output sequences used in the simulation are shown in Figure 1. The algorithm’s
parameters are chosen as r ¼ 1; Q ¼ 0:01I and P0 ¼ 0:1I:
In the simulation, we take the following two different initial parameter estimation vectors y# 0
to show the difference in the resulting convergence rates.
(i) y# 0 ¼ ½0 0 0 0T :
(ii) y# 0 ¼ ½0 0 2 0:4T :

From the simulation results, it is found that


a ¼ 0:0288; b ¼ 0:2647
m1 ¼ 0:0368; m2 ¼ 3:024
The above results are independent of y# 0 :
Based on Theorem 3.1 we know the convergence rate is a number between r2 ¼ 1=ð1 þ m1 Þ ¼
0:2485 and r1 ¼ 1=ð1 þ m1 Þ ¼ 0:9646; and the tracking error y* t is bounded in the following way:
!2
t jy* t j
10ar2   10brt1
jy* 0 j
The above inequality is also independent of the initial estimation error.

Copyright # 2003 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2003; 17:763–783
EXPONENTIAL CONVERGENCE OF THE KALMAN FILTER 781

Figure 2. Normalized tracking errors ð1 : y# 0 ¼ ½0 0 0 0T ; 2 : y# 0 ¼ ½0 0 2 0:4T Þ; the upper bound (3) and
the lower bound (4).

Figure 3. Normalized tracking error(solid line) with y# 0 ¼ ½0 0 0 0T and the graph 0:82t (dashed line).

The normalized tracking error jy* t j2 =jy* 0 j2 and its bounds are shown in Figure 2. Noting that
the only difference between the graph 1 and 2 in Figure 2 is the initial estimation vector y# 0 ; we
see that the convergence speed is highly dependent on the initial condition.
From Figure 2 it can also be seen that during the tracking process the tracking errors decrease
with different rates at different instants, which means that the convergence rate is time-varying.

Copyright # 2003 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2003; 17:763–783
782 L. CAO AND H. M. SCHWARTZ

Figure 4. Normalized tracking error (solid line) with y# 0 ¼ ½0 0 2 0:4T and the graph 0:95t (dashed line).

However, we can approximately describe the tracking process by the average of the time-varying
convergence rate. It can be determined that the average convergence rate for the graph 1 shown
in Figure 2 is close to 0.82, while the average convergence rate for the graph 2 is close to 0.95.
The graph 0:82t is shown in Figure 3 together with the actual tracking error with the initial
condition y# 0 ¼ ½0 0 0 0T ; the graph 0:95t is shown in Figure 4 together with the actual tracking
error with the initial condition y# 0 ¼ ½0 0 2 0:4T : From Figures 3 and 4 we see that the
normalized tracking error graphs fit well with some function of the form rt ; r51; which
indicates that the tracking process is basically dominated by a constant convergence rate. We see
that r2 50:8250:955r1 ; and therefore the averaged convergence rate is also bounded by
Theorem 3.1.
From Figure 2 we see that the differences between the actual tracking errors (graph 1 and
graph 2) and their upper bound (graph 3) is large. However, the difference between the upper
bound r1 ¼ 0:9646 of the convergence rate and the averaged convergence rate 0.95 in Figure 4 is
small. Thus, in this example, the upper bound of the convergence rate given in Theorem 3.1 is
not very conservative.

7. CONCLUDING REMARKS

A new analytical method has been developed to prove the exponential convergence of the
Kalman filter based parameter estimator. This method is simpler and much easier to perform
than some existing methods in the literature [3]. In addition, both lower and an upper bounds on
the convergence rate are also given in terms of the ratios of the eigenvalues of the related
matrices. These bounds provide a quantitative description on the convergence rate, but are
difficult to evaluate. Comparisons between the results of this paper and the others in the
literature have shown the new features of the approach developed in this paper.

Copyright # 2003 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2003; 17:763–783
EXPONENTIAL CONVERGENCE OF THE KALMAN FILTER 783

A simulation example has shown that the convergence rate of a KFB algorithm is
time-varying and the tracking process is basically dominated by the average of the time-varying
convergence rate. The average convergence rate is dependent on the initial conditions and is
bounded by the lower and upper bounds given in Theorem 3.1.
Finally, an important fact about the convergence property of a KFB algorithm is that unlike
the exponentially weighted least squares algorithm whose the convergence rate is mainly
determined by the forgetting factor, for the Kalman filter its convergence rate appears as a
complex function of the related matrices, which makes it difficult to get insights into how the
convergence rate is affected by the adjustable parameters.
REFERENCES
1. Anderson BDO, Johnstone RM. Adaptive systems and time varying plants. International Journal of Control 1983;
37(2):367–377.
2. Parkum JE, Poulsen NK, Holst J. Recursive forgetting algorithms. International Journal of Control 1992; 55(1):
109–128.
3. Bittanti S, Bolzern P, Campi M. Exponential convergence of a modified directional forgetting identification
algorithm. Systems & Control Letters 1990; 14:131–137.
4. Johnstone RM, Johnson Jr. CR, Bitmead RR, Anderson BDO. Exponential convergence of recursive least squares
with exponential forgetting factor. Systems & Control Letters 1982; 2(2):77–82.
5. Johnson Jr. CR. Lectures on Adaptive Parameters Estimation. Prentice-Hall: Englewood Cliffs, 1988.
6. Ljung L, Gunnarsson S. Adaptation and tracking in system identification-a survey. Automatica 1990; 26(1):7–21.
7. Guo L. Estimating time-varying parameters by the Kalman filter based algorithm: stability and convergence. IEEE
Transactions on Automatic Control 1990; AC-35(2):141–147.
8. Guo L, Ljung L. Exponential stability of general tracking algorithms. IEEE Transactions on Automatic Control
1995; AC-40(8):1376–1387.
9. Horn RA, Johnson CR. Matrix Analysis. Cambridge University Press: Cambridge, 1985.
10. Bittanti S, Bolzern P, Campi M. Convergence and exponential convergence of identification algorithms with
directional forgetting factor. Automatica 1990; 26(5):929–932.

Copyright # 2003 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2003; 17:763–783

You might also like