910 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 45, NO.

5, MAY 2000

Gaussian Filters for Nonlinear Filtering Problems

Kazufumi Ito and Kaiqi Xiong

Abstract—In this paper we develop and analyze real-time and use Gaussian sum filters for the development of nearly optimal
accurate filters for nonlinear filtering problems based on the filters. The Gaussian sum filter has been studied in [1] and
Gaussian distributions. We present the systematic formulation [20]. However, we adapt our Gaussian filter for the update of
of Gaussian filters and develop efficient and accurate numerical
integration of the optimal filter. We also discuss the mixed Gaussian distributions. We also suggest some new update rules
Gaussian filters in which the conditional probability density of weights for Gaussian sum filters. Through our experimental
is approximated by the sum of Gaussian distributions. A new study we found that the filters developed in the paper perform
update rule of weights for Gaussian sum filters is proposed. better than or as good as the filter of Julier–Uhlmann [12]. They
Our numerical testings demonstrate that new filters significantly have a significant improvement over the extended Kalman filter
improve the extended Kalman filter with no additional cost and
the new Gaussian sum filter has a nearly optimal performance. with no additional cost.
An outline of the paper is as follows. In Section II we de-
Index Terms—Author: please supply index terms. E-mail: key-
[email protected] for information. velop the Gaussian filter based on a single Gaussian distribu-
tion. In Section III we discuss the efficient numerical integration
of the Gaussian filter based on quadrature rules, and introduce
I. INTRODUCTION the Gauss–Hermite filter (GHF) and the central difference filter
(CDF). In Section IV we formulate the Gauss–Hermite filter and
T HE NONLINEAR filtering problem consists of estimating
the state of a nonlinear stochastic system from noisy ob-
servation data. The problem has been the subject of consider-
the central difference filter as the filter algorithms. In Section V
we introduce the mixed Gaussian filter and the new update rules
able research interest during the past several decades because of weights. In Section VI we discuss the relation of the Gaussian
it has many significant applications in science and engineering filter for the discrete time system to the continuous-time optimal
such as navigational and guidance systems, radar tracking, sonar filter governed by the Zakai equation. In Section VII we ana-
ranging, and satellite and airplane orbit determination [11], [14], lyze the stability and performance bound of the Gaussian fil-
[15]. As is well-known, the most widely used filter is the ex- ters and the mixed Gaussian filters developed in the paper. In
tended Kalman filter for nonlinear filtering problems. It is de- Section VIII we report our numerical findings and comparison
rived from the Kalman filter based on the successive lineariza- studies. Also, we demonstrate a nearly optimal performance of
tion of the signal process and the observation map. The extended the mixed Gaussian filter. Finally, we conclude our results in
Kalman filter has been successfully applied to numerous non- Section IX.
linear filtering problems. If nonlinearities are significant, how-
ever, its performance can be substantially improved. Such ef- II. GAUSSIAN FILTERS
forts have also been reported in [1], [5], [7], [12], [13], and [20]. We discuss the nonlinear filtering problem for the dis-
In this paper our objective is to develop and analyze real-time crete-time signal system for -valued process
and accurate filters for nonlinear filtering algorithms based on (2.1)
Gaussian distributions. We present the systematic formulation
of Gaussian filters and mixed Gaussian filters. and the observation process is given by
We first develop the Gaussian filter. The proposed filter is (2.2)
based on the steps: 1) we assume the conditional probability where and are white noises with covariances and
density to be a Gaussian distribution (i.e., assumed density) respectively. We assume that the initial condition and
and 2) we obtain the Gaussian filter by equating the Bayesian are independent random variables. The optimal non-
formula with respect to the first moment (mean) and the second linear filtering problem is to find the conditional expectation
moment (covariance). Our approach is based on the efficient of the process given the observation data
numerical integration of the Bayesian formula for optimal The probability density function of the
recursive filtering. The direct evaluation of the Jacobian matrix conditional expectation is given by Bayes’ formula
associated with the extended Kalman filter is avoided, which
is similar to the one recently reported in [12]. Secondly, we

Manuscript received November 7, 1997; revised November 30, 1998 and June
16, 1999. Recommended by Associate Editor, T. E. Duncan. This work was sup-
ported in part by the Office of Naval Research under Grant N00014-96-1-0265 (2.3)
and MURI-AFOSR under Grant F49620-95-1-0447.
The authors are with the Center for Research in Scientific Computation, and
North Carolina State University, Raleigh, NC 27695-8205 USA (e-mail:
[email protected]; [email protected]).
Publisher Item Identifier S 0018-9286(00)04155-6. (2.4)
0018–9286/00$10.00 © 2000 IEEE
ITO AND XIONG: GAUSSIAN FILTERS FOR NONLINEAR FILTERING PROBLEMS 911

where is the one-step prediction and is the probability and

density function of conditioned on That is, the re-
cursive filter (2.3), (2.4) consists of the prediction step (2.3)
and the correction step (2.4). We consider a Gaussian approx-
imation of this recursive formula and develop Gaussian filters.
We assume that is a single Gaussian distribution with
mean and covariance Then we construct
the Gaussian approximation of in the following two steps.
First, we consider the predictor step. We approximate
by the Gaussian distribution that has the same mean and covari- (2.6)
ance as By Fubini’s theorem the mean and covariance
of are given by Next, we discuss the corrector step

where is the normalization constant and we assume that

is given by the Gaussian approximation defined by (2.5) and
(2.6). Define the innovation process

Here we again approximate for the conditional distribution

of given the observations up to time by
a Gaussian. That is, we approximate by its
Gaussian approximation This means that the probability
and density function of is given by the Gaussian distribution with
mean and covariance defined by

(2.7)
and

Remark 1.1: To derive the mean and covariance of

(2.8)
we also can use (2.1) and the independence of and
For example Finally, we construct the Gaussian approximation of with
mean and covariance defined by

(2.9)
and
which is precisely the same as the above, though differently ex- (2.10)
pressed.
Thus, if is a Gaussian with mean and co- where the Kalman-filter gain is defined by
variance then the Gaussian approximation of
has mean and covariance defined by (2.11)

and the covariance is defined by

(2.5)
(2.12)
912 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 45, NO. 5, MAY 2000

In order to implement the Gaussian filter we must develop If we assume and change the coordinate of integration
the approximation methods to evaluate integrals (2.5)–(2.12). by then
In Section III we discuss the approximation methods for the
integration of the form (3.2)

with We apply the Gauss–Hermite quadra-

ture rule. The Gauss–Hermite quadrature rule is given by
where is a given function. Our discussions include the
Gauss–Hermite quadrature rule and finite difference approxi-
mation. Note that a different approximation to this integration
results in a variant of the Gaussian filter. For example, if we ap-
where the equality holds for all polynomials of degree up to
proximate the integration by linearizing at then we obtain
and the quadrature points and the weights are de-
the extended Kalman filter.
termined (e.g., see [9]) as follows. Let be the symmetric tridi-
In order to improve the performance of Gaussian filter we
agonal matrix with zero diagonals and
also discuss the mixed Gaussian filter in Section V. We approx-
Then are the eigenvalues of and equal
imate by the linear combination of multiple Gaussian
to where is the first element of the th normalized
distributions, i.e.,
eigenvector of Thus, is approximated by

(3.3)

where , and is exact for all polyno-

mials of the form with In order
to evaluate we need -point function evaluations. For ex-
ample we have

Then, each Gaussian distribution is updated separately by the

and
Gaussian filter (2.5)–(2.12). In the corrector step we update the
weights For example, we can determine the weights
by the -projection, i.e., minimize and requires nine-point function evaluations for
In Julier–Uhlmann [12] the standard normal distribution is
approximated by the discrete distribution as

and

over where is defined as in the corrector step (2.4),

assuming

where is a constant and is th unit vector in

Here, this discrete distribution has the same first, second,
and higher odd moments as the standard normal distribution.
Julier–Uhlmann developed a Gaussian filter based on the
following quadrature rule for (3.1):

(3.4)
III. QUADRATURE RULES
The JU-rule requires -point function evaluation and is
In this section we discuss the approximation methods for the exact for all quadratic polynomials. If we set and
integral of the form then the JU-rule coincides with the Gauss–Hermite rule
Finally we consider the polynomial interpolation methods.
We approximate by the quadratic function that satisfies

(3.1)
ITO AND XIONG: GAUSSIAN FILTERS FOR NONLINEAR FILTERING PROBLEMS 913

at the points in consisting of and

and

where is the stepsize. That is, is given by

(3.10)
(3.5)
If we remove the second-order correction term in (3.9) and
where is the th coordinate of a point and the vector (3.10), then this coincides with the extended Kalman filter with
in and the symmetric matrix are the central difference approximation of the Jacobian of
defined by
IV. FILTER ALGORITHMS
In this section we describe the filter algorithms that are based
on the Gaussian filter and quadrature methods described in Sec-
tion III. We first consider the Gauss–Hermite filter based on the
Gaussian–Hermite quadrature rule. Let be the quadra-
ture rule for

(4.1)
Thus, we approximate (3.1) by
Let and be the starting values for the mean and co-
variance of the random variable and set and
In Gaussian–Hermite filter we apply the following
(3.6) predictor and corrector steps recursively.

and the integral A. Gauss–Hermite Filter

Predictor Step: Compute the factorization
and set Update
by
by

(3.7)
Corrector Step: Compute the factorization
and set Update by
where is the quadratic approximation of for
Next, we consider the approximation of defined by

where

(3.8)

which only bases on the values

and and uses the diagonal second order correc-
tion of the central difference approximation of
in (3.5). Then the integrals and are approximated by

(3.9) We use the Choresky decomposition for finding and [6]

in our calculations. One of the advantages of the quadrature
914 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 45, NO. 5, MAY 2000

based filter is that we are not required to have the derivative of V. MIXED GAUSSIAN FILTER
and In this section we discuss the mixed Gaussian filter. We ap-
Secondly, we present the filter algorithm based on the cen- proximate the conditional probability density by the
tral difference approximation with second order diagonal cor- linear combination of multiple Gaussian distributions, i.e.,
rection. We summarize the filter algorithm of the central differ-
ence approximation with the second-order diagonal correction
as follows.

B. Central Difference Filter

Predictor Step: Compute the factorization
and set approx-
imated by Update
by (5.1)

Here we apply the Gaussian filter (2.5)–(2.12) to each Gaussian

distribution and obtain the update
Each update is independent from the others and
can be performed in a parallel manner.
Next, we update the weights for the new update
where are the central difference approximations of at the end of corrector step. We discuss three update formulas
in what follows.
Corrector Step: Compute the factorization First, by equating the zero moment of each Gaussian distri-
and set approximated by bution we obtain
Update by

where

Here we approximate the right-hand side by the Gaussian dis-

tribution as in (2.5) and (2.6) and obtain
where are the central difference approximations of

In the algorithm we avoid calculating the derivatives of and

Instead, we use the central difference. (5.2)
The following are the operation counts in terms of function
evaluations. where are defined by (2.7) and (2.8), which is the update
• If we use the quadrature rule based on the -point formula discussed in [1].
Gauss–Hermite rule then the algorithm requires Next, we apply the collocation condition at
function evaluations.
• If we use the Julier–Uhlmann discrete Gaussian rule then
the algorithm requires function evalua-
tions.
• If we use the central difference algorithm, the
function evaluations are required.
As we will see, our numerical testings indicate that the algo-
rithm based on the Gauss–Hermite rule performs better than the
others. If we use the three-point rule, then we require 729
function evaluations for the case which can be done in
real-time.
ITO AND XIONG: GAUSSIAN FILTERS FOR NONLINEAR FILTERING PROBLEMS 915

to obtain the update where is chosen so that the singularity of the matrix
is avoided and the matrices are defined by

(5.3)
Finally we discuss the simultaneous update of the weights.
We determine the weights by the -projection, i.e.,
minimize

Thus, we solve (5.6) to obtain the weights at each corrector

step by using the existing numerical optimization method (e.g.,
see [6]).
The theoretical foundation of the Gaussian sum approxima-
tion as above is that any probability density function can be ap-
proximated as closely as desired by a Gaussian sum. More pre-
(5.4) cisely, we state the following error estimate, the proof of which
is found in [17].
over where is defined as in the corrector step (2.4) Theorem 5.1: Let be a nonnegative integer and
with For any there must exist and a mask

such that for any density function

and all the estimate

In order to perform the minimization (5.4) we need to evaluate (5.7)

the integral of the form
holds, where is a linear combination of Gaussian distri-
butions given by

(5.8)

and it is relatively expensive. Hence we propose the minimiza- with

tion of the sum of collocation distances

and is a constant independent of Here and

are defined by

(5.5) and
over satisfying A positive constant
is chosen so that the likelihood of each Gaussian distribution is
nonzero [e.g., ]. Problem (5.5) is
formulated as the quadratic programming respectively.
Roughly speaking, the estimate (5.7) shows that any proba-
subject to (5.6) bility density function can be
916 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 45, NO. 5, MAY 2000

approximated by a sum of Gaussian distributions each of whose (6.3). We approximate the process by the discrete-time
components is given by process by

(5.9) (6.6)
with
with order for

VI. RELATION TO CONTINUOUS-TIME FILTER Here the interval is subdivided into the
In this section let us discuss the nonlinear filtering problem subintervals
for the continuous-time signal process in generated by
(6.1)
where and
where is the standard Brownian motion, i.e., is the So, denotes the approximation of the process at
diffusion process [3], [11], [19]. So, (6.1) holds in the sense of and denotes the discretization of the ordi-
Ito and satisfies nary differential equation

for all We employ a family of Runge–Kutta methods in our calculation,

Now we consider the continuous observation process including the Euler approximation

(6.2)
The predictor step (2.5), (2.6) is applied successively times
to obtain i.e.,
where is the standard Brownian process that is independent
of and
Or, we consider the discrete observation process
(6.3)
where is the stepsize and is white noise with
covariance We assume that the initial condition and
are independent. (6.7)
Then the conditional probability density and
satisfies the Zakai equation
(6.4)
where is the Fokker–Planck operator [3], [19].
As shown in [8], [10], and [15], the discrete-time filter (2.3),
(2.4) applied to the time-discretized signal system of (6.1), (6.2)

(6.5) (6.8)
provides an approximation method for the continuous-time op-
timal filter to (6.1), (6.2). Note that denotes the finite dif- with
ference approximation

for Hence, the mean and covariance of are

for the ordinary differential equation given by
(6.9)
Then the corrector step (2.7)–(2.12) is applied as it is to obtain
The proposed Gaussian filter does not converge to the optimal
filter (6.4) as unless is Gaussian. However, it
converges to a Gaussian filter, which approximates (6.4) and
will be discussed elsewhere. VII. STABILITY ANALYSIS OF FILTER ALGORITHMS
Next, we discuss the discrete-time filter for the contin- In this section we analyze the stability and performance
uous-time process (6.1) with the discrete-time observation bound of the Gaussian filters (2.5)–(2.12) for the signal process
ITO AND XIONG: GAUSSIAN FILTERS FOR NONLINEAR FILTERING PROBLEMS 917

(6.5) with i.e., a discrete approximation by assumption (7.1). That is

of the continuous signal system (6.1) can be given by
(7.5)

where is the white noise with covariance Note that for any the covariance is nonegative and satis-
We assume that fies the following property:
for (7.1)
Then we have the following theorem. The theorem shows sta- Thus by (7.5) we obtain the estimate (7.2).
bility of the proposed Gaussian filter independent of the fineness Remark 7.1: Using the similar arguments that lead to (7.2),
of on a fixed bounded time interval. we can prove that the same result holds for the approximation
Theorem 7.1: The solution to (2.5)–(2.12) has the esti- method discussed in Sections III and IV. That is, the estimate
mate (7.2) is still valid for the Gaussian–Hermite filter, provided that

Next, we discuss the stability of the mixed Gaussian filter.

The mixed Gaussian filter assumes the form

(7.6)

where the th Gaussian distribution is com-

(7.2) puted by the Gaussian filters described in Sections III and IV
Proof: From (2.5), (2.6) the Gaussian approximation of in a parallel manner. The weights are deter-
has the mean and covariance defined by mined at the end of the corrector step based on the likelihood
of each Gaussian distribution. Under the assumption (7.1) each
(7.3) Gaussian filter in the Gaussian sum (7.6) is stable in the sense
of (7.2).
and Next we discuss the performance bound. In general, let the
operators and denote the one-step update
of conditional probability density functions that are based on the
(7.4) exact filter (2.3), (2.4) and the mixed Gaussian filter discussed
in Section V, respectively. We estimate the difference between
Note that the exact filter and the mixed Gaussian filter

(7.7)
where is the probability density function of random vari-
Thus by the Cauchy and Schwarz inequality we have able and

where
is the mixed Gaussian approximation of Let
Then we have the following result.
Theorem 7.2: Assume that for any probability density func-
tion and
Hence it follows from (7.4) that
(7.8)
Then the error estimate

(7.9)

holds, where is the initial error defined by

918 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 45, NO. 5, MAY 2000

and the error is given by ters against that of the extended Kalman filter (EKF) and the
filter of Julier–Uhlmann (JUF) [12].
(7.10)
We use the average root mean square error for our comparison
Proof: Note that of the methods. The average root mean square error is defined
by

(8.1)

which is based on different simulation runs. Here subscript

denotes the th component of the vector and its corre-
sponding estimate and the superscript denotes the th
simulation run.
Note that for a one-dimensional system the filter of
Julier–Uhlmann [12] coincides with the three-point Gauss–Her-
mite filter when Also when the stepsize is the
Thus we have central difference filter has the same estimate algorithm as the
filter of Julier–Uhlmann except for the estimation Hence,
in this section the stepsize is always chosen as for our study.
We also note that our proposed filter algorithms maintain the
which leads to (7.9).
nonnegative definiteness of the covariance update.
Remark 7.2: Suppose is a mixed Gaussian distribution
We carried out our comparison study using various examples
and we set Also assume that and
and different starting states and different problem data, but
are affine functions. Then we can show that i.e., the
we only present the selected results in what follows.
mixed Gaussian filter is exact, since the Gaussian filter is exact
Example 8.1: We consider the one-dimensional signal
for the Gaussian distribution in this case.
process
To understand the validity and usage of error formula (7.9)
we adopt the following filter (which is different from the one we (8.2)
implemented), i.e., the mixed Gaussian filter with fixed means
and the observation process
and covariances. For simplicity of our discussion we assume
Let us consider the fixed Gaussian elements (8.3)
where

with nodal points and covariance

Then we define the subspace of the probability distribu- and
tions on by
It is motivated from the continuous time systems (6.1) and (6.2)
of the form
Define is the projection of onto
We consider the mixed Gaussian filter with fixed
means and covariances, i.e., (8.4)
as described in Section VI.
This example is used as a benchmark because of the following
reasons. The deterministic equation
where
(8.5)
has two stable equilibria, 1 and 1, and one unstable equi-
librium, 0. So, we observed that the signal process is dis-
Thus is the fixed operator in this case. Hence
tributed around one of stable equilibria. The observation func-
can be estimated based on the error estimate formula (7.9),
tion is the square of shifted distance from
assuming the smoothness of
the origin and distinguishes the two stable equilibria marginally.
The performance of the filter is possibly tested because of the
VIII. NUMERICAL RESULTS AND COMPARISONS
minor margin of detecting measurement function of the
In this section we demonstrate the feasibility of our proposed signal process.
filter algorithms, Gauss–Hermite filter (GHF), and central dif- We use the following problem data: the time elapse
ference filter (CDF), using the three test examples as in Exam- the initial condition i.e., the
ples 8.1–8.3. We compare the performance of our proposed fil- starting point and the system
ITO AND XIONG: GAUSSIAN FILTERS FOR NONLINEAR FILTERING PROBLEMS 919

Fig. 1. The average of root mean square error.

parameters and We consider the time from the system. Here, we choose and
0 to 4. is chosen as the best value, i.e., (see [12]). which is a mathematically very interesting case, because in the
Fig. 1 shows the average root mean square errors committed case there are three unstable equilibria and a strange attractor
by each filter across a simulation consisting of 50 runs. From for the equations.
Fig. 1 we see that the five-point Gauss–Hermite filter performs We use the following problem data: the system parameters are
better than others. The central difference filter has similar per- chosen as and the initial condition is
formance like the filter of Julier–Uhlmann [12] in this example. i.e.,
All the three filters perform superior to the extended Kalman and We consider the time from 0 to 4. The initial
filter. estimate is with covariance
Example 8.2: In this example, we consider the Lorenz In Figs. 2–4 we show the average of root mean square errors
system for each component of system state committed by the algorithms
of the Gauss–Hermite quadrature rule, the central difference ap-
proximation, and the JU-rule, respectively, across a simulation
consisting of 50 runs. As shown in Figs. 2–4, the GHF and the
where is the -valued process. The CDF have smaller errors than the JUF and the EKF. From these
-valued processes and are white noises with the figures we conclude that in this example the GHF has a substan-
same covariance Here is a constant vector and tial improved performance, and the GHF performs a little better
is a constant scalar. Also than the CDF.
Example 8.3: We further discuss the three-dimensional con-
tinuous signal process
(8.6)
(8.8)
The observation function is chosen as the shifted distance from
and the discrete observation process
the origin given by
(8.9)

It is motivated from the Lorenz stochastic differential system where and are zero mean and uncorrelated
of the form (8.4) as described in Section VI where noises with covariances given by and
The three-dimensional deterministic equations respectively. The function is given by

(8.7)
is the Lorenz equation in which are positive The above system states and represent altitude
parameters. The three parameters have a great deal of impact on (in feet), velocity (in feet per second), and constant ballistic
920 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 45, NO. 5, MAY 2000

Fig. 2. The average of root mean square error.

Fig. 3. The average of root mean square error.

coefficient (in per second), respectively. The detailed physical linearity of signal process [2], [12]. That is, in the predicator
meaning of the system and its parameters can be found in [2]. steps of EKF and JUF, the means are calculated using the nu-
We chose this example as a benchmark because it contains merical scheme. Then their covariance are propagated from the
significant nonlinearities in the signal and observation processes th to th step using
and had been discussed widely in the literature.
In the previous literature, a fourth-order Runge–Kutta scheme
with 64 steps between each observation is employed for numer-
ical integration of (8.8) in order to deal with the significant non-
ITO AND XIONG: GAUSSIAN FILTERS FOR NONLINEAR FILTERING PROBLEMS 921

Fig. 4. The average of root mean square error.

Fig. 5. The absolute value of average altitude error.

with (8.9). In the predictor step of GHF, we directly use the Euler
approximation of signal process (8.8) based on 32 or 64 steps
(i.e., 32 or 64) between each observation (they are shortly
denoted by GHF32 or GHF64, respectively).
where was evaluated at and Thus, However, in NGHF the signal process system (8.8) is
rewritten as the discrete form
In the comparison, we implement the proposed approxima-
tion schemes (6.7)–(6.9) to solve the filtering problem of (8.8),
922 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 45, NO. 5, MAY 2000

Fig. 6. The absolute value of average velocity error.

Fig. 7. The absolute value of average x -error.

where is the approximation of the process at We use the following data: the system parameters are chosen
and function with as and
the initial condition is
and We consider the time from 0 to 30. The
initial estimate is with
covariance

which is the fourth-order Runge–Kutta scheme. Specifically,

in NGHF64.
ITO AND XIONG: GAUSSIAN FILTERS FOR NONLINEAR FILTERING PROBLEMS 923

Fig. 8. The comparison of signal; estimates of two mixed Gaussian filter and optimal filters.

Fig. 9. The weights of two mixed Gaussian filters.

We also choose the optimization number in JUF for the A. Mixed Gaussian Example
three-dimensional system.
In Figs. 5–7 we show the absolute value of average error In this section we first demonstrate a nearly optimal perfor-
for each component committed by the algorithms of the mance of the mixed Gaussian method described in Section V.
Gauss–Hermite quadrature rule, the JU-rule, and the EKF, We consider the same example as in Example 8.1 with
respectively, across a simulation consisting of 50 runs. These We employ the two Gaussian distributions starting from
figures show that in this example the GHF has superior perfor-
mance than the JUF and the EKF.
924 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 45, NO. 5, MAY 2000

Fig. 10. The comparison of the probability density functions of mixed Gaussian filter and optimal filter at t = 1:

Fig. 11. The comparison of the probability density functions of mixed Gaussian filter and optimal filter at t = 2:75:

and against the one of the optimal Zakai filter (6.4) with
in Figs. 8–11. Here GHF1 (GHF2) represents the three-point
Gauss–Hermite filter which is the first (second) component of
respectively, and the initial weights are In the two mixed Gaussian filter with Weight 1 (Weight 2). We ap-
order to demonstrate the effectiveness of the proposed update proximate the Zakai equation by the operator-splitting method
(5.5), (5.6) we jumped the process at [which as described in [10]. We observed that the mixed Gaussian
corresponds to in the continuous time process (8.4)] filter performs nearly optimally up to We also note
to It may represent an impulse force at that the Zakai filter is no longer optimal after because
We compare the performance of the mixed Gaussian filter of the jump. As seen as Figs. 8 and 9, the update formula (5.5),
ITO AND XIONG: GAUSSIAN FILTERS FOR NONLINEAR FILTERING PROBLEMS 925

Fig. 12. The comparison of altitude errors of each Gaussian filter.

Fig. 13. The comparison of velocity errors of each Gaussian filter.

(5.6) quickly captures the phase change from the one steady and
point to the other The observations are also
supported by the comparison in Figs. 10 and 11.
Example 8.3 (Revisited): Next, we also apply our mixed with the same covariance
Gaussian method to Example 8.3. We employ the two Gaussian
distribution starting from
926 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 45, NO. 5, MAY 2000

Fig. 14. The comparison of x errors of each Gaussian filter.

Fig. 15. The weights of two mixed Gaussian filters.

respectively, and the initial weights are IX. CONCLUSIONS

We demonstrate the performance of the mixed Gaussian filter
with in Figs. 12–15. We see that the mixed Gaussian The paper presents the systematic formulation of Gaussian
filter based on the update formula (5.5), (5.6) captures the filters and mixed Gaussian filters based on the efficient numer-
signal process very well. That is, the weights of the two mixed ical integration of the Bayesian formula for optimal recursive
Gaussian filters shown in Fig. 15 effectively pick up the com- filter. Based on our formulation we develop the two filter algo-
bined performance of the Gaussian filters: GHF1 and GHF2. rithms, namely, the Gauss–Hermite filter (GHF) and the central
More detailed discussions and numerical testings of the mixed difference filter (CDF). We demonstrated the feasibility of our
Gaussian filter will be presented in the forthcoming paper. proposed filter algorithms for testing nonlinear filtering prob-
ITO AND XIONG: GAUSSIAN FILTERS FOR NONLINEAR FILTERING PROBLEMS 927

lems. Our numerical results indicate that both the Gauss–Her- [14] R. E. Kalman, “A new approach to linear filtering and prediction prob-
mite filter and the central difference filter have superior perfor- lems,” Trans. ASME, J. Basic Eng., vol. 82D, pp. 35–45, Mar. 1960.
[15] H. J. Kushner, “Approximation to optimal nonlinear filters,” IEEE
mance to the filter of Julier–Uhlmann and the extended Kalman Trans. Automat. Contr., vol. 12, pp. 546–556, 1967.
filter. We also proposed the new update rules for the Gauss sum [16] , private communication.
filters and show that they can perform near optimally. [17] V. Mazya and G. Schmidt, “On approximate approximations using
gaussian kernels,” IMA J. Numer. Anal., vol. 16, pp. 13–29, 1996.
[18] M. Piccioni, “Convergence of implicit discretization schemes for linear
ACKNOWLEDGMENT differential equations eith application to filtering,” in Stochastic Partial
Differential Equations and Application, G. Da Prato and L. Tabbaro,
The authors thank the referees for their careful reviews and Eds. Trento, 1985, vol. 1236. (Lecture Note in Mathematics).
helpful suggestions. [19] B. L. Rozovskii, Stochastic Evolution Systems, Linear Theory and Ap-
plication to Nonlinear Filtering. Norwell, MA: Kluwer, 1991.
[20] H. W. Sorenson and D. L. Alspace, “Recursive Bayesian estimation
REFERENCES using Gaussian sums,” Automatica, vol. 7, pp. 465–479, 1967.
[1] D. L. Alspace and H. W. Sorenson, “Nonlinear Bayesian estimation
using Gaussian sum approximation,” IEEE Trans. Automat. Contr., vol.
17, pp. 439–448, 1972.
[2] M. Athans, R. P. Wishner, and A. Bertolini, “Suboptimal state estima-
tion for continuous-time nonlinear systems from discrete noise measure-
ments,” IEEE Trans. Automat. Contr., vol. 13, pp. 504–514, 1968. Kazufumi Ito received the Ph.D. degree in system
[3] A. Bensoussan, “Nonlinear filtering theory,” in Progress in Automation science and mathematics from Washington Univer-
and Information Systems, Recent Advances in Stochastic Calculus, J. S. sity in 1981. Subsequently, he held the positions
Baras and V. Mirelli, Eds. New York: Springer-Verlag, 1990. in ICASE, Brown University and University of
[4] A. Bensoussan, R. Glowinski, and A. Rascanu, “Approximation of the Southern California before joining the Department
Zakai equation by the splitting up method,” SIAM J. Contr. Optim., vol. of Mathematics, North Carolina State University,
28, pp. 1420–1431, 1990. where he is currently a Professor. His research inter-
[5] Y. Bar-Shalom and X. R. Liu, Estimation and Tracking: Principles, ests include nonlinear filtering, stochastic process,
Techniques, and Software. New York: Artech, 1993. control and analysis of partial differential equations.
[6] J. E. Dennis and R. B. Schnabel, Numerical Methods for Unconstrained He co-organized an AMS-IMS-SIAM Joint Summer
Optimization and Nonlinear Equations. Englewood Cliffs, NJ: Pren- Research Conference on Identification and Control
tice-Hall, 1983. in Systems governed by PDE’s, 1992 and the workshop on Stochastic Control
[7] C. P. Fang, “New algorithms of Gaussian assumed density filter and a and Nonlinear filtering, at NC State University, 1996.
convergence result,” IEEE Trans. Automat. Contr..
[8] P. Florchinger and F. Le Gland, “Time-discretization of the Zakai
equation for diffusion process observed in correlated noise,” in Proc.
9th Conf. Analysis and Optimization of Systems, A. Bensoussan and J.
L. Lions, Eds. New York: Springer-Verlag, 1990, vol. 144. (Lecture
Notes on in Control and Inform. Sci.). Kaigqi Xiong received the M.S. and Ph.D. de-
[9] G. H. Golub, “Some modified matrix eigenvalue problems,” SIAM Rev., grees in applied mathematics from the Claremont
vol. 15, pp. 318–334, 1973. Graduate School in 1995 and 1996, respectively. In
[10] K. Ito, “Approximation of the Zakai equation for nonlinear filtering,” 1995–1996, he was also a Researcher of Electrical
SIAM J. Contr. Optim., vol. 34, pp. 620–634, 1996. Engineering at the University of California, River-
[11] A. H. Jazwinski, Stochastic Process and Filtering Theory. New York: side. From 1996–1999, he was a Visiting Assistant
Academic, 1970. Professor in the Center for Research in Scientific
[12] S. J. Julier and J. K. Uhlmann. (1994) A general method for approxi- Computation at the North Carolina State University.
mating nonlinear transformations of probability distributions. [Online]. As a Researcher, he is visiting the Department of
Available: www http://phoebe.robots.ox.ac.uk/default.htm Mechanical and Aerospace Engineering at the Uni-
[13] S. J. Julier, J. K. Uhlmann, and H. F. Durrant-Whyte, “A new approach versity of California, Irvine. His research interests
for filtering nonlinear systems,” in Proc. Amer. Contr. Conf., Seattle, include nonlinear systems, mathematical control theory, filtering algorithms,
WA, June 1995, pp. 1628–1632. parameter estimation, flight control, neural networks and applications.