International Journal of Pure and Applied Mathematics

Volume 91 No. 1 2014, 33-41

ISSN: 1311-8080 (printed version); ISSN: 1314-3395 (on-line version)
url: http://www.ijpam.eu
doi: http://dx.doi.org/10.12732/ijpam.v91i1.4
AP
ijpam.eu

CONVERGENCE ANALYSIS OF EXTENDED KALMAN

FILTER IN A NOISY ENVIRONMENT THROUGH
DIFFERENCE EQUATIONS

S. Elizabeth1 , R. Jothilakshmi2 §
1,2 Departmentof Mathematics
Auxilium College
Vellore, 632 006, Tamilnadu, INDIA

Abstract: In this paper, the convergence aspects of the Extended Kalman

Filter, when used as a deterministic observer for a nonlinear discrete-time sys-
tems, are addressed and analyzed. The conditions needed to ensure the bound-
edness of the error covariances which are related to the observability properties
of the nonlinear systems are identified through difference equations. Further-
more, boundedness and stability conditions are provided in a noisy environment
systems.

AMS Subject Classification: 39A10, 39A11, 39A99, 34C99

Key Words: boundedness, stability, extended Kalman filter, positive definite
matrices

1. Introduction

Difference equations have many applications in several applied sciences (such as

Biology, Ecology, Economics, Population dynamics, Genetics, etc), Engineer-
ing ( Control Systems, Digital Signal Processing, Image Processing, etc.) and
Medicine [1]. For this reason, there exist an increasing interest in studying
difference equations and systems of difference equations.
c 2014 Academic Publications, Ltd.

Received: August 4, 2013 url: www.acadpubl.eu
§ Correspondence author
34 S. Elizabeth, R. Jothilakshmi

Designing an observer for a nonlinear system is quite challenge. Thus, as

a first step, it is interesting to see how classical linearization techniques work
with nonlinear systems and what their limitations are. All biological systems
are non-linear, since growth or change cannot continue in a linear manner for
very long time, without causing reduction in available resources. Baras [3] de-
scribe a method for constructing observers for dynamic systems as asymptotic
limits of filters. They discuss the method as applied to the linear case, and a
class of nonlinear systems with linear observations in the continuous-time do-
main. Essentially the extended Kalman filter (EKF) is used as their observer
of both linear and nonlinear system of equations. The extended Kalman filter
(EKF) is a well-known standard linearization method for approximate nonlin-
ear filtering [9]. The Kalman filter is one of the most widely used methods
for tracking and estimation due to its simplicity, optimality, traceability and
robustness. However the application of the Kalman filter to nonlinear systems
can be difficult. The most common approach is to use the Extended Kalman fil-
ter which simply linearises all nonlinear models, so that the traditional Kalman
filter is applied.
The paper is organized as follows. Section 2 describes the basic concepts.
Section 3 provides the convergence aspects of the Extended Kalman Filter,
when used as a deterministic observer for a nonlinear discrete-time system.
Systems with nonlinear output maps are also considered. The conditions needed
to ensure the boundedness of the error covariances which are related to the
observability properties of the nonlinear systems are identified. Furthermore,
the asymptotic convergence of the observation errors are analyzed in a noisy
environment. Section 4 concludes the paper.

2. Preliminaries

We denote by |.|, the Euclidean norm of a vector, and by ||.|| and |||.|||, the
induced norms on matrices and tensors. We also adopt the following nota-
tions. Consider two symmetric matrices P and Q, of the same dimension, the
inequality P ≥ Q means that the difference P − Q is non-negative definite
matrix. Similarly, P > Q means that P − Q is positive definite matrix. A
symmetric matrix Q is said to be bounded from above (below, respectively) if
there is a number q > 0 such that Q ≤ qI (qI ≤ Q). It is well known that, under
stochastic controllability and observability assumptions, the Kalman filter for
a time-varying system with artificial noises can be used as a global asymptotic
observer for the deterministic system [2].
CONVERGENCE ANALYSIS OF EXTENDED KALMAN FILTER... 35

Consider a system of difference equations [12],

xk+1 = Ak xk + Bk uk , x0 : unknown,
(1)
y k = Ck xk ,

where Ak is assumed as invertible matrix, xk ∈ Rn is the input of the system

state, yk ∈ Rm is the measured output at a time instant k. Also consider the
associated noisy system,

zk+1 = Ak zk + Bk uk + N wk ,
(2)
ξk = Ck zk + Rvk ,

uk , vk , wk are the Gaussian random vectors with zero mean covariances and
mutually independent. The design parameters N and R are to be chosen as
positive definite matrices. Filtering is a procedure of estimating hidden states
based on observable data [4]. The Kalman filter equations for (2) are given as
follows [5].
Now Consider the measurement and time update of a noisy environmented
system of equations,

Measurement Update

x̂k = x̄k + Kk (ξk − Ck x̄k ) ,

−1 (3)
Pk−1 = P̄k−1 + CkT RRT Ck ,

Time Update

x̄k+1 = Ak x̂k + Bk uk ,
(4)
P̄k+1 = Ak Pk ATk + N N T ,

−1
−1
Kk = Pk CkT RRT = P̄k CkT Ck P̄k CkT + RRT
where P̄k and Pk are the a priori and a posteriori covariances, and x̄k and x̂k
be the priori and a posteriori estimates of the state at time k, respectively [10].
The filter is initiated with x̄0 and P̄0 ; P̄0 is used as a design parameter, assumed
also positive definite.
36 S. Elizabeth, R. Jothilakshmi

To obtain an error dynamics, let’s rewrite the Kalman filter in terms of the
priori variables [5]. From (3) and (4) we use yk instead of ξk ,

x̄k+1 = Ak (I − Kk Ck ) x̄k + Bk uk + Ak Kk yk , (5)

P̄k+1 = Ak (I − Kk Ck ) P̄k ATk + N N T . (6)

If we define the error as ek = xk − x̄k , then the error dynamics is given as [8]

ek+1 = Ak (I − Kk Ck )ek . (7)

The associated Riccati difference equations for the error covariances are [6]
h
−1 i−1 T
P̄k+1 = Ak P̄k−1 + CkT RRT Ck Ak + N N T , (8)

−1
−1
−1
= Ak Pk ATk + N N T + CkT RRT

Pk+1 Ck . (9)

Note that P̄0 > 0 and rank N = n implies P̄k > 0 and Pk > 0 for all 0 ≤ k < ∞
[7].

3. Main Results

Economic agents often possess limited information. In addition, economic in-

formation is often noisy, possibly because of measurement errors [11]. In this
section, we study control problems under partial (or imperfect) information.
Since we are interested in the asymptotic behaviour of the error ek , it is neces-
sary to obtain bounds for ||P̄k || and ||Pk−1 || [13, 14].
Lemma 3.1. Consider the following noisy system:

xk+1 = Ak xk + N wk ,
(10)
yk = Ck xk + Rvk .

Suppose that there are positive real numbers α1 , α2 , β1 , β2 such that the
following conditions hold for some finite M ≥ 0 and for all k ≥ M :
k−1
X
α1 I ≥ φ(k, i + 1)N N T φT (k, i + 1) ≥ α2 I, (11)
i=k−M
CONVERGENCE ANALYSIS OF EXTENDED KALMAN FILTER... 37

k
X
β1 I ≤ φT (i, k)CiT (RRT )−1 Ci φ(i, k) ≤ β2 I; (12)
i=k−M

then
1
I ≤ Pk ≤ (α1 + 1/β1 )I,
β2 + 1/α2
where
φ(k, i) = Ak−1 Ak−2 ...Ai .

Remark 3.2. Under the above conditions, it can be shown that P̄k is
bounded from above and below. Indeed, from (4),

P̄k ≤ (α1 + 1/β1 ) kAk2 + kN k2 .

Also, from (3),

1
P̄k ≥ Pk ≥ I.
β2 + 1/α2
Therefore,
1
I ≤ P̄k ≤ (α1 + 1/β1 ) kAk2 + kN k2 I.
β2 + 1/α2
It is obvious that Pk−1 and P̄k−1 are both bounded from above and below, which
ensures the boundedness.
Now we state a theorem on the convergence of the error, on which an
extention is made to the nonlinear systems.
Lemma 3.3. Assume for some a > 0 there exists a function V such that:

1. V : n+ × Ba → RR+ ; V (k, 0) = 0 is positive definite and continuous with

respect to the second argument;
2. ∆V (k, ek ) = V (k + 1, ek+1 ) − V (k, ek ) ≤ −µ(|ek |), where µ is of class K.
Then the origin of (7) is asymptotically stable.
Theorem 3.4. Consider the system (1) and the Kalman filter equations (3)
and (4) for the associated system (2). Suppose that Ak is invertible for all k
and that assumption (1) holds. Suppose that ||A|| = sup {||Ak || : k = 0, 1, ..., }
and ||C|| = sup {||Ck || : k = 0, 1, ..., } are bounded. Then the Kalman filter
for the noisy system (2) is a global asymptotic observer for the deterministic
system (1), as long as N has rank n, R and Q̄0 are positive definite matrices.
38 S. Elizabeth, R. Jothilakshmi

−
Proof. Let PK = (Q− −1
k ) . From (4) we have,

A−1 − −T −1 T −T
k Qk+1 Ak = Qk + Ak N N Ak .

Inverting the above equation

−
ATk Pk+1 Ak = Q−1 −1 −1 T T −1 −1 −1
k − Qk (Qk + Ak (N N ) Ak ) Qk .

If we note
Qk = (I − Kk Ck )Q−
k or
Q−1
k = (Q− −1
K ) (I
−1
− Kk Ck )

then

−
ATk Pk+1 Ak =
 −
Pk − Pk− (I − Kk Ck )−1 (Q−1 T T −1 −1 −
(I − Kk Ck )−1 . (13)

k + Ak (N N ) Ak ) Pk

Thus, from (7) and (13) we have,

− −
eTk+1 Pk+1 ek+1 = eTk (I − Kk Ck )T ATk Pk+1 Ak (I − Kk Ck )ek
= ek (I − Kk Ck ) Pk − Pk− (I − Kk Ck )−1 (Q−1
T T −

k
T T −1 −1 −

+Ak (N N ) Ak ) Pk ek .

Since Qk = (I −Kk Ck )Q− T − −

k is symmetric, then (I −Kk Ck ) = Pk (I −Kk Ck )Qk .
Therefore,
−
eTk+1 Pk+1 ek+1 = eTk Pk− (I − Kk Ck ) − Pk− (Q−1 T T −1 −1 −

k + Ak (N N ) Ak ) Pk ek
= eTk Pk− ek − eTk Pk− Kk Ck + Pk− (Q−1

k
+ATk (N N T )−1 Ak )−1 Pk− ek .

Now if we let V (k, ek ) = eTk Pk− ek then V satisfies the conditions given in
lemma (3.3). Moreover, noting that

Pk− Kk Ck = CkT (Ck Q− T T −1

k Ck + RR ) Ck .

−
∆V (k, ek ) = eTk+1 Pk+1 ek+1 − eTk Pk− ek
= −eTk CkT (Ck Q− T T −1 − −1

k Ck + RR ) Ck + Pk (Qk
+ATk (N N T )−1 Ak )−1 Pk− ek

≤ −eTk Pk− (Q−1 T T −1 −1 −

k + Ak (N N ) Ak ) Pk ek .
CONVERGENCE ANALYSIS OF EXTENDED KALMAN FILTER... 39

Since

||Q−1 T T −1 −1
k + A + k (N N ) Ak || ≤ ||Qk || + ||N
−1
Ak ||2
≤ p + ||N −1 ||2 ||A||2 = r
1 − 2
eTk Pk− (Q−1 T T −1 −1 −
k + Ak (N N ) Ak ) Pk ek ≥ r |Pk ek | .

If we use |Pk− ek | ≥ 1q |ek |, the above equation will be of the form,

−1 −1
∆V (k, ek ) ≤ 2
|ek |2 ≤ 2 V (k, ek ),
rq rq p1

where we used the bounds given in assumption (1) and p1 = p + ||R−1 ||2 ||C||2 .
Therefore by Lemma (3.3), ek converges to zero asymptotically.

4. Conclusion

In this paper we have argued that the principle difficulty for identifying the
error in the noisy environment. We applied the Riccati difference equation
techniques to the Extended Kalman filter in a noisy environment which con-
sistently predict the new state and observation of the system. Also we extend
the arrival of convergence and stability concepts of the Extended Kalman filter
through difference equations.

Acknowledgments

This research is funded by the University Grants Commission (F. No. 42 - 20/
2013 (SR) UGC), New Delhi, India.
40 S. Elizabeth, R. Jothilakshmi

References

[1] R.P. Agarwal, Difference Equations and Inequalities, Marcel Dekker, New
York, 2000.

[2] Alfredo, Costanzo, The polynomial extended Kalman filter as an exponen-

tial observer for nonlinear discrete-time systems, In: Proceedings of the
47-th IEEE Conference on Decision and Control Cancun, Mexico (2008),
9-11.

[3] J.S. Baras, A. Bensoussan, N.R. James, Dynamic observers as asymptotic

limits of recursive filters: Special cases, SIAM J. Appl. Math., 48, No. 5
(2008), 1147-1158.

[4] D.E. Catlin, Estimation, control and the discrete Kalman filter, Applied
Mathematical Sciences, Springer-Verlag (1989), 71-84.

[5] S.T. Chung, J.W. Grizzle, Sampled-data observer error linearization, Au-
tomatica, 26, No. 6 (2001), 997-1007.

[6] S. Elaydi, An Introduction to Difference Equations, Springer-Verlag, New

York, 1996.

[7] E.M. Elsayed, On the solution of some difference equations, European Jour-
nal of Pure and Applied Mathematics, 4, No. 3 (2011), 287-303.

[8] Eric J. Kostelich, James A. Yorke, Noise reduction: Finding the simplest
dynamical system consistent with the data, Physica D: Nonlinear Phenom-
ena (1990), 183-196.

[9] S.J. Julier, J.K. Uhlmann, H.F. Durrant-Whyte, A new approach for fil-
tering nonlinear systems, In: The Proceedings of the American Control
Conference, Seattie, Washington (1995), 1628-1632.

[10] K.H. Kim, G.I. Jee, C.G. Park, J.G. Lee, The stability analysis of the
adaptive fading extended Kalman filter using the innovation covariance,
International Journal of Control, Automation and Systems, 7 (2009), 49-
56.

[11] R. Lopez-Valcarce, S. Dasgupta, Stable estimates in equation error iden-

tification, In: Proceedings of the 36-th IEEE Conference on DEcision and
Control, 2 (1997), 1678-1679.
CONVERGENCE ANALYSIS OF EXTENDED KALMAN FILTER... 41

[12] J.C. Peyton Jones, S.A. Billings, Recursive algorithm for computing the
frequency response of a class of non-linear difference equation models, In-
ternational Journal of Control (2007), 1925-1940.

[13] Michel Verhaegen, Patrick Dewilde, Subspace model identification Part

1. The output-error state-space model identification class of algorithms,
International Journal of Control (1991), 1187-1210.

[14] Y. Song, J.W. Grizzle, The extended Kalman filter as a local asymptotic
observer for discrete-time nonlinear systems, J. Math. Systems Estimate
and Control, 5 (1995), 59-78.
42