Introduction to Kalman filtering

Course key notes

Corrected exercises
Matlab training session

D.Alazard

December, 2011 - version 0.4

EMPTY PAGE
 3

Contents

Introduction 5

1 Random signal in linear systems 7

1.1 Problem statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2 Backgrounds and definitions . . . . . . . . . . . . . . . . . . . . . . . 9
1.2.1 Characterization of random variables . . . . . . . . . . . . . . 9
1.2.2 Caracterization of multivariate random variables . . . . . . . . 11
1.2.3 Random signal (or process) . . . . . . . . . . . . . . . . . . . 14
1.2.4 Moments for random signal . . . . . . . . . . . . . . . . . . . 15
1.2.5 Stationarity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.2.6 Spectrum in the s-plane . . . . . . . . . . . . . . . . . . . . . 17
1.2.7 White noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.3 Transmission of random signals in linear systems . . . . . . . . . . . . 21
1.3.1 Time-domain approach . . . . . . . . . . . . . . . . . . . . . . 21
1.3.2 Frequency-domain (spectral) approach . . . . . . . . . . . . . 22
1.4 Matlab illustration . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

2 Kalman filter 27
2.1 Principle of Kalman filtering . . . . . . . . . . . . . . . . . . . . . . 27
2.1.1 Kalman model . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.1.2 Assumptions: . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.1.3 Structure of an unbiased estimator . . . . . . . . . . . . . . . 30
2.2 Minimal error variance estimator . . . . . . . . . . . . . . . . . . . . 32
2.2.1 General solution . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.2.2 Kalman filter steady state . . . . . . . . . . . . . . . . . . . 34
2.2.3 Kalman filter tuning . . . . . . . . . . . . . . . . . . . . . . 35
2.3 Corrected exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

Introduction to Kalman filtering Page 3/80

4 Contents

2.3.1 First order system . . . . . . . . . . . . . . . . . . . . . . . . 36

2.3.2 Bias estimation . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.4 Discrete-time Kalman filter . . . . . . . . . . . . . . . . . . . . . . . 46
2.4.1 Discrete-time Kalman model . . . . . . . . . . . . . . . . . . 46
2.4.2 A particular case: continuous-time plant with discrete-time
measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
2.4.3 Recurrent equations of discrete Kalman filter . . . . . . . . . 49
2.4.4 Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
2.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
2.5.1 Second order system: . . . . . . . . . . . . . . . . . . . . . . . 54

3 About physical units 57

References 59

A State space equation integration 61

A.1 Continuous-time case . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
A.2 Discrete-time case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

B Transmission of random signals and noises in linear systems 65

B.1 Additional background: discrete random signals . . . . . . . . . . . . 65
B.2 Time-domain approach . . . . . . . . . . . . . . . . . . . . . . . . . . 66
B.2.1 Continuous-time case . . . . . . . . . . . . . . . . . . . . . . . 66
B.2.2 Discrete-time case . . . . . . . . . . . . . . . . . . . . . . . . . 68
B.3 Frequency-domain approach . . . . . . . . . . . . . . . . . . . . . . . 70
B.3.1 Continuous-time case . . . . . . . . . . . . . . . . . . . . . . . 70
B.3.2 Discrete-time case . . . . . . . . . . . . . . . . . . . . . . . . . 71

C Analytical solution of continuous-time differential Riccati equation 73

C.1 General analytical solution . . . . . . . . . . . . . . . . . . . . . . . . 73
C.2 Exemple . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74

D Matlab demo files 77

D.1 Function Kf t.m . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
D.2 Script file demoKalman.m . . . . . . . . . . . . . . . . . . . . . . . . . 77
D.3 Script file demoKalmand.m . . . . . . . . . . . . . . . . . . . . . . . . 79

Introduction to Kalman filtering Page 4/80

 5

Introduction

This document is an introduction to Kalman optimal filtering applied to linear

systems. It is assumed that the reader is already aware of linear control theory
(servo-loop systems), frequency-domain filtering (continuous and discrete-time) and
state-space approach to represent linear systems.
Generally, filtering consists in estimating a useful information (signal) from
a measurement (of this information) perturbed by a noise. Frequency-domain fil-
tering assumes that a frequency-domain separation exists between the frequency re-
sponse of the useful signal and the frequency response of the noise. Then, frequency-
domain filtering consists in seeking a transfer function fitting a template on its mag-
nitude response (and too much rarely, on its phase response). Kalman optimal
filtering aims to estimate the state vector of a linear system (thus, this state is
the useful information) and this estimate is optimal w.r.t. an index performance:
the sum of estimation error variances for all state vector components. First of all,
some backgrounds on random variables and signals are required then, assumptions,
structure and computation of Kalman filter could be introduced.
In the first chapter, we remind to the reader how a random signal can be
characterized from a mathematical point of view. The response of a linear system
to a random signal will be investigated in an additional way to the more well-
known response of a linear system to a deterministic signal (impulse, step, ramp,
... responses). In the second chapter, the assumptions, the structure, the main
parameters and properties of Kalman filter will be defined. The reader who wishes
to learn practical Kalman filter tuning methodology can directly start the reading
at chapter 2. But the reading of chapter 1, which is more cumbersome from a
theoritical point of view, is required if one wishes to learn basic principles in random
signal processing, on which is based Kalman filtering.
There are many applications of Kalman filtering in aeronautics and aerospace
engineering. As Kalman filter provides an estimate of plant states from an a priori
information of the plant behaviour (model) and from real measurement, Kalman
filter will be used to estimate initial conditions (ballistics), to predict vehicle po-
sition and trajectory (navigation) and also to implement control laws based on a
state feedback and a state estimator (LQG: Linear Quadratic Gaussian con-

Introduction to Kalman filtering Page 5/80

6 Introduction

trol). The signal processing principles on which is based Kalman filter will be
also very useful to study and perform test protocols, experimental data processing
and also parametric identification, that is the experimental determination of some
plant dynamic parameters.

Introduction to Kalman filtering Page 6/80

 7

Chapter 1

Random signal in linear systems

1.1 Problem statement

Let us consider a state-space model of a linear and stationary plant:

ẋ(t) = Ax(t) + Bu(t) + M w(t) (state equation)
(1.1)
y(t) = Cx(t) + Du(t) + v(t) (measurement equation)

where :

• x(t) ∈ Rn is the state vector of the plant,

• u(t) ∈ Rm is the deterministic input vector (known inputs: control signal,

...),

• w(t) ∈ Rq is the vector of unknown random signals (noises) that perturb the
state equation through the input matrix Mn×q (wx = M w denotes also the
state noise),

• y(t) ∈ Rp is the measurement vector,

• v(t) ∈ Rp is the vector of random signals (measurements noise) perturbing

the measurement (it is assumed there are as much noises as measurements).

Example 1.1 The ”quarter vehicle” model used to study a car active suspension
can be depicted by Figure 1.1. The control device (piston, jack) allows a force u
to be applied between the wheel (with a mass m and a vertical position z) and the
body of the car (with a mass M and a vertical position Z). K and f denote the
stiffness and the damping of the passive suspension. k is the tire stiffness between
wheel and ground. Finally, w denotes the vertical position of the point of contact
tire/ground excited by road unevenness (rolling noise). Body vertical acceleration

Introduction to Kalman filtering Page 7/80

 8 1. Random signal in linear systems

K f u
g
m
Z

k
z

ground w

Figure 1.1: Quarter vehicle model.

is measurement by the means of an inertial sensor (an accelerometer tacking into

account gravity g = −9.81 m/s2 ). The noise on this accelerometer is denoted v.
Let us denote δz et δZ small variations of z and Z around an equilibrium position
z0 et Z0 (equilibrium conditions are: w = 0, u = 0 and gravity effects are entirely
compensated by stiffness (k and K) compression). Then applying Newton principle
on both masses m and M , we get:
¨ = K(δz − δZ) + f (δz
M δZ ˙ − δZ)
˙ +u
mδz¨ = −K(δz − δZ) − f (δz˙ − δZ)
˙ − u − k(δz − w) .

˙ δz]
Let x = [δZ, δz, δZ, ˙ T be the state vector, then the following state space repre-
sentation can be derived:
     
0 0 1 0 0 0   0
 0 0 0 1   0   
ẋ =  x +  0  u +  0 w
 −K/M K/M −f /M f /M   1/M 0  g  0 
K/m −(K + k)/m f /m −f /m −1/m 0   1/k
  u
y = −K/M K/M −f /M f /M x + [1/M − 1] +v .
g
(1.2)
This model is in the form of (1.1) with n = 4, m = 2, q = 1 and p = 1.

We recognize in matrices (A, B, C, D) of model (1.1) the well known state

space representation of the transfer between deterministic input u and measurement

Introduction to Kalman filtering Page 8/80

1.2 Backgrounds and definitions 9

1
y: :
F (s) = D + C(sI − A)−1 B .
The response of this model to a deterministic input u(t) over a time range t ∈ [t0 , t]
and to initial conditions x(t0 ) reads :
Z t
A(t−t0 )
x(t) = e x(t0 ) + eA(t−τ ) Bu(τ ) dτ (1.3)
t0
y(t) = Cx(t) + Du(t) (1.4)

Proof: see appendix A where an illustrative exercise is also proposed.

What about the response of the model (1.1) to a random signal w(t) ?
To answer this question, it is required:

• firstly to remind how a random signal can be characterized from a mathemat-

ical (or stochastic) point of view,

• to state some assumptions on the stochastic properties of noises w(t) and v(t)
(gaussian white noise) to facilitate the determination of the response of model
1.1,

• to compute stochastic characteristics of the response (x(t) and y(t)).

The various definitions and background given below are extracted from refer-
ence [5] (chapter II and appendix A.I).

1.2 Backgrounds and definitions

1.2.1 Characterization of random variables
Let X be a random variable defined on the real space R. The probability dis-
tribution function F (x) associates, to each real value x, the probability of the
occurrence X < x. That is:

F : R → [0 1] / F (x) = P [X < x] .

Properties:

• ∀x1 < x2 , P [x1 ≤ X < x2 ] = F (x2 ) − F (x1 ),

• limx→+∞ F (x) = 1; limx→−∞ F (x) = 0.

1
s stands for Laplace variable.

Introduction to Kalman filtering Page 9/80

10 1. Random signal in linear systems

• F (x) is a monotonous, increasing function, and can be continuous or discon-

tinuous depending on X has continuous or discrete values, respectively.

If F (x) is differentiable, then its derivative w.r.t. x is called probability density

function and is denoted p(x):

dF (x)
p(x) = soit : p(x)dx = P [x ≤ X < x + dx] .
dx
To characterize a random variable X , one can also used the moments of this
variable. The first moment is called mean value or expected value. The second
central moment is called variance and is denoted varx = σx2 where σx is the
standard deviation. That is:

• expected value:
Z +∞ Z +∞
E[X ] = x p(x) dx = x dF (x) , (1.5)
−∞ −∞

• k-th moment: Z +∞
k
E[X ] = xk p(x) dx , (1.6)
−∞

2
• k-th central moment:
Z +∞
k
E[(X − E[X ]) ] = (x − E[X ])k p(x) dx .
−∞

Full description of a random variable requires the characterization of all its

moments. But from a practical point of view 3-th and higher moments are not used
because they cannot be computed or derived easily. The (mathematical) interest of
normal (gaussian) random variable lies is the fact they are entirely defined by the
first and second moments. Indeed, let X a gaussian random variable with a mean
value m and a standard deviation σ, then:
1 (x−m)2
p(x) = √ e− 2σ2 , E[x] = m, E[(x − m)2 ] = σ 2 .
2πσ

Example 1.2 [Uniform distribution] The probability density function of the random
variable X is constant between two values a and b with b > a.
Z b  b
x 1 1 2 1 b 2 − a2 a+b
E[X ] = dx = x = =
a b−a b−a 2 a 2 b−a 2
2
Recall: varx = E[X 2 ] − E[X ]2 .

Introduction to Kalman filtering Page 10/80

1.2 Backgrounds and definitions 11

F(x) p(x)

1 1/(b−a)

a 0 b x a 0 b x

Figure 1.2: Probability distribution and density functions for a uniform distribution.

Z b 2 "  3 #b
1 a+b 1 1 a+b
varx = E[(X − E[X ])2 ] = x− dx = x−
b−a a 2 b−a 3 2
a
 3 2
1 1 b−a (b − a)
⇒ varx = = .
3b−a 2 12

Discrete random variable: If a random variable is defined on a set of N discrete

values xi , i = 1, · · · , N then the density function is no more used and one can directly
define the ”probability that X = xi ” noted P (X = xi ). The definition of moments
involves a discrete sum :
X
N
k
E[X ] = xki P (X = xi ) .
i=1

Example 1.3 The random variable X corresponding to a dice tossing:

1 21 35
P (X = i) = , ∀ i = 1, · · · , 6 ; E[X ] = ; varx = .
6 6 12
2

1.2.2 Caracterization of multivariate random variables

Let X = [X1 , · · · , Xq ]T a vectorial random variable with q components taking on the
values in Rq .

Probability distribution function

(i.e. the joint distribution of the q random variables Xi , i = 1, · · · q)

F (x1 , · · · , xq ) = P (X1 < x1 and X2 < x2 and · · · and Xq < xq ) .

Introduction to Kalman filtering Page 11/80

12 1. Random signal in linear systems

Density function

∂ q F (x1 , · · · , xq )
p(x1 , · · · , xq ) = .
∂x1 · · · ∂xq

Moments

Define x = [x1 , · · · , xq ]T . Only the vector or first moments (that is the mean
vector) and the matrix of second central moments (that is the covariance matrix)
are considered:

• mean: E[X ] = [E[X1 ], · · · , E[Xq ]]T .

• covariance: Covx = E[(X − E[X ])(X − E[X ])T ]. The component Covx (i, j) at
row i and column j of this covariance matrix verifies:
Z
Covx (i, j) = (xi − E[Xi ])(xj − E[Xj ]) dF (xi , xj ) .
R2

The covariance matrix is definite, positive and symmetric.

Gaussian random vector with mean m and covariance ∆

1 1 T −1
p(x) = √ e− 2 (x−m) ∆ (x−m) .
(2π)q/2 det∆

The gaussian random vector X with mean m and covariance ∆ can be build from
the normal gaussian vector N (that is the vector with a null mean and an unitary
covariance) in the following way:

X = m + GN

where G is such that: GGT = ∆.

Independence

Two random variables X1 and X2 are independent if and only if:

F (x1 , x2 ) = F (x1 )F (x2 ) .

An independence necessary condition is:

E[X1 X2 ] = E[X1 ]E[X2 ] . (1.7)

Introduction to Kalman filtering Page 12/80

1.2 Backgrounds and definitions 13

Exercise 1.1 Lest us consider 2 independent random variables X1 and X2 with a

uniform density between −1 and 1 for each of this 2 variables. Define a new random
variable as Y = X1 +X
2
2
.
Compute E[Y], vary and covx1 ,y the covariance matrix of the vector: (X1 , Y)T .

Correction: From exercise 1.2, one can derive:

1
E[X1 ] = E[X2 ] = 0; varx1 = varx2 = .
3
 
X1
By definition: Y = (1/2 1/2) . Then :
X2
 
X1
E[Y] = (1/2 1/2) E =0,
X2
     
X1 1/2 1/2
vary = (1/2 1/2)E (X1 X2 ) = (1/2 1/2)covx1 ,x2
X2 1/2 1/2
 
1/3 0
X1 and X2 are independent, thus: covx1 ,x2 =
0 1/3

1
⇒ vary = .
6
In the same manner:
 
T 1 0
covx1 ,y = Gcovx1 ,x2 G , with : G =
  1/2 1/2
.
1/3 1/6
=
1/6 1/6

Remark : At the cost of more tedious calculus, one can also give a full caracterization
of the random variable Y by its distribution function F (y) or density function p(y)
and then compute first and second moments:
  Z
X1 + X2
F (y) = P < y = P (X1 < 2y − X2 ) = P (X1 < 2y − x2 )p(x2 )dx2 .
2 Dx2

For a given value y, we can write (see Figure 1.2):

P (X1 < 2y − x2 ) = 0, ∀ x2 / 2y − x2 < −1 ⇒ ∀ x2 > 2y + 1,

2y − x2 + 1
P (X1 < 2y − x2 ) = , ∀ x2 / − 1 ≤ 2y − x2 < 1 ⇒ ∀ x2 / 2y − 1 < x2 ≤ 2y + 1,
2
P (X1 < 2y − x2 ) = 1, ∀ x2 / 2y − x2 ≥ 1 ⇒ ∀ x2 ≤ 2y − 1.

Introduction to Kalman filtering Page 13/80

14 1. Random signal in linear systems

And p(x2 ) = 1/2 if −1 ≤ x2 < 1, p(x2 ) = 0 else. Then 3 :

Z 1 Z 2y+1
1 2y − x2 + 1 (y + 1)2
• if y < 0 F (y) = 0 dx2 + dx2 + 0 = ,
2y+1 2 −1 4 2
Z 1 Z 2y−1
2y − x2 + 1 1 −y 2 + 2y + 1
• if y ≥ 0 F (y) = 0 + dx2 + dx2 = .
2y−1 4 −1 2 2
That yields to:

p(y) = y + 1 ∀y ∈ [−1, 0],

p(y) = −y + 1 ∀y ∈ [0, 1],
R1 R0 R1
E[Y] = −1 y p(y)dy = −1 y(y + 1)dy + 0 y(−y + 1)dy = 0,
R1 R0 R1
vary = −1 y 2 p(y)dy = −1 y 2 (y + 1)dy + 0 y 2 (−y + 1)dy = 61 .

F(y) p(y)
1 1

−1 0 1 y −1 0 1 y

Figure 1.3: Distribution and density functions of Y.

In the following, random variable (resp. signal) stands for both single (q = 1)
or multivariate (q components) variable (resp. signal).

1.2.3 Random signal (or process)

Given a random variable X , the random signal x(t) is a function of time t such
that for each given t, x(t) corresponds to a value (a sample) of X .
3
Using conditional probabilities, one can also write:
Z Z
2y − x2 + 1
F (y) = 0 p(x2 |x2 > 2y + 1)dx2 + p(x2 |2y − 1 < x2 ≤ 2y + 1)dx2
Dx2 Dx2 2
Z
+ 1 p(x2 |x2 ≤ 2y − 1)dx2 .
Dx2

Introduction to Kalman filtering Page 14/80

1.2 Backgrounds and definitions 15

1.2.4 Moments for random signal

The second moment of a random signal is called the auto-correlation function.
Let w(t) be a random signal, then:

first moment: m(t) = E[w(t)] (1.8)

T
second moment: φww (t, τ ) = E[w(t)w(t + τ ) ] . (1.9)

Remark 1.1 if w(t) is a vectorial (multivariate) signal with q components then

φww (t, τ ) is a q × q definite positive matrix for each value of t and τ . The diagonal
terms are the scalar auto-correlation functions of each components and the cross
terms are the inter-correlation functions between components.

A centered gaussian random signal, that is a signal generated at each

instant from a sample of a gaussian random variable with a null mean value, is
therefore entirely defined by its auto-correlation function.

1.2.5 Stationarity
A random signal is defined to be wide sense stationary 4 if its mean is constant
(m(t) = m) and if its auto-correlation function depends only on τ (φww (t, τ ) =
φww (τ )).
The quadratic mean or variance 5 of a stationary random signal is the auto-
correlation function value at the origin:

σw2 = φww (τ )|τ =0

Exercise 1.2 A random signal b(t) is generated in the following way: from the
initial time t0 = 0, the signal b(t) is hold every dt seconds on the value (sample) of
a gaussian centered random variable X with a standard deviation σ; all samples xi
of the variable X are independent to each others. So: b(t) = xi ∀t ∈ [idt, (i + 1)dt[
(see Figure 1.4).

• Compute the mean m(t) and the auto-correlation function φbb (t, τ ) of signal
b(t). Is b(t) wide sense stationary ?.

• Answer again previous questions assuming now that the initial instant is a
random variable with a uniform distribution between 0 and dt.

Solution keys:
4
One can also defined the strict sense stationary (see reference[3]).
5
or covariance matrix in the case of a vectorial random signal.

Introduction to Kalman filtering Page 15/80

16 1. Random signal in linear systems

b(t)

2 σ

dt
σ
Probability
(−2σ < b(t) < 2σ)
=0.95
0

− σ

−2 σ

0 10 dt 20 dt .... t

Figure 1.4: Time domain response of a realization of random signal b(t).

• m(t) = E[b(t)] = E[X ] = 0, ∀t,

• φbb (t, τ ) = E[b(t)b(t + τ )] = σ 2 if t and t + τ are in the same time interval dt;
0 else (because the samplings are independent and centered).

• the auto-correlation function depends on t and τ ; so b(t) is not a stationary

signal. For instance, for t = i dt + ε (∀i, ∀ ∈ [0, dt[) , the response of φbb (t, τ )
is plotted in Figure 1.5.

• if initial time t0 is now a random variable with a uniform distribution taking

value between 0 and dt (let b0 (t) be this new random signal) then one can
consider in the previous calculus that ε has a uniform density between 0 and
dt: Z
1 dt
φb0 b0 (t, τ ) = φbb (i dt + ε, τ )dε
dt 0
– for dt > τ ≥ 0, φbb (i dt + ε, τ ) = σ 2 iff 0 < ε < dt − τ , 0 else. That is:
Z dt−τ
σ2 σ2
φb0 b0 (t, τ ) = dε = (dt − τ ) ∀τ / 0 ≤ τ < dt .
dt 0 dt

– for −dt < τ < 0, φbb (i dt + ε, τ ) = σ 2 iff −τ < ε < dt, 0 else. That is:
Z dt
σ2 σ2
φb0 b0 (t, τ ) = dε = (dt + τ ) ∀τ / − dt < τ < 0 .
dt −τ dt

Introduction to Kalman filtering Page 16/80

1.2 Backgrounds and definitions 17

Figure 1.5: Autocorrelation function of signal b(t).

−dt 0 dt

Figure 1.6: Autocorrelation function of signal b0 (t).

So φb0 b0 (t, τ ) = σ 2 (1− |τdt| ) ∀τ ∈ [−dt dt], 0 else (see Figure 1.6) and depends
only on τ , b0 (t) is now wide sense stationary.

1.2.6 Spectrum in the s-plane

Wide sense stationary random signals can also be characterized by their frequency-
domain representation called Power Spectral Density (PSD) or also their spec-
trum in the s-plane (the PSD is given by the spectrum where s is replaced by
jω). The spectrum in the s-plane of a Wide sense stationary random signal is the
bilateral Laplace transform of the auto-correlation function 6 .
Z ∞
Φww (s) = LII [φww (τ )] = φww (τ )e−τ s dτ . (1.10)
−∞

Power Spectral Density (PSD): Φww (ω) = Φww (s)|s=jω .

6
The inverse bilateral Laplace transform inverse is defined by:
Z +j∞
1
φww (τ ) = L−1
II Φww (s) = Φww (s)esτ ds
2π j −j∞

Introduction to Kalman filtering Page 17/80

18 1. Random signal in linear systems

Remark 1.2 :
Z ∞ Z ∞
−τ s
Φww (s) = φww (τ )e dτ + φww (−u)eus du
0 0
−
= Φ+
ww (s) + Φww (−s)

with :

Φ+ +
ww (s) = L[φww (τ )] and φ+
ww (τ ) = φww (τ ) if τ ≥ 0, φ+
ww (τ ) = 0 else ,

Φ− −
ww (s) = L[φww (τ )] and φ−
ww (τ ) = φww (−τ ) if τ ≥ 0, φ−
ww (τ ) = 0 else .
−
If the function φww (τ ) is pair then Φ+ ww = Φww and the spectrum in the s-plane
Φww (s) is a function of s2 (that is, the PSD is real).
The initial value theorem of the mono-lateral Laplace transform allows the
variance to be computed directly from the spectrum in the s-plane:

σw2 = φww (τ )|τ =0 = lim sΦ+

ww (s) .
s→∞

Remark 1.3 : From the PSD, one can write:

Z +∞ Z +∞
1 1
φww (τ ) = Φww (ω)ejωτ dω and σw2 = Φww (ω)dω .
2π −∞ 2π −∞

The variance of the noise w is (with a factor 1/2π) the integral of the PSD in the
frequency domain.

Example 1.4 Let us consider again the signal b0 (t) of section 1.2 whose autocor-
relation function is: φb0 b0 (t, τ ) = σ 2 (1 − |τdt| ) (pair function). The spectrum in the
s-plane of this signal is:
Z ∞
|τ | −τ s
Φb0 b0 (s) = σ 2 (1 − )e dτ = Φ+ +
b0 b0 (s) + Φb0 b0 (−s) with:
−∞ dt

Z dt
τ
Φ+
b0 b0 (s) = σ 2 (1 − )e−τ s dτ
0 dt
( dt Z dt )
σ2 dt − τ −τ s 1
= e − e−τ s dτ (integration by parts)
dt −s 0 s 0
(  −τ s dt )
σ2 dt e σ2 −s dt


= + = s dt + e − 1 .
dt s s2 0 dt s2

Introduction to Kalman filtering Page 18/80

1.2 Backgrounds and definitions 19

σ2


Φb0 b0 (s) = dt s2
e−s dt + es dt − 2 .
Thus the PSD is:
σ2 −jω dt jω dt

2σ 2 4σ 2 ω dt
Φb0 b0 (ω) = − 2
e + e − 2 = 2
(1 − cos(ω dt)) = 2
sin2 ( )
dt ω dt ω dt ω 2

sin2 ( ω2dt )
Φb0 b0 (ω) = σ 2 dt ( ω2dt )2
.

Φb0 b0 (ω) is plotted in Figure 1.7 7 . From a practical point of view, the logarithmic
scale representation given in Figure 1.8 is preferred and highlights that the PSD
can be considered as constant for pulsations very lower than the sampling pulsation
2π/dt. This signal can be used as a white noise signal with a PSD R in a
given pulsation range if 2 π/dt is chosen very large w.r.t. this pulsation
range and if σ 2 = R/dt (see next section).

Φ (ω)
b’b’

2
σ dt

−4π/dt −2π/dt 0 2π/dt 4π/dt ω

Figure 1.7: PSD for signal b0 (t).

This frequency-domain representation of random signal is particularly useful to

study and characterized the transmission of random signals in stationary linear
systems (see section 1.3.2).
7
It is recalled that limx→0 ( sin x
x )=1

Introduction to Kalman filtering Page 19/80

20 1. Random signal in linear systems

Φb’b’(ω) |dB

20 log10(σ2 dt)

0 2π/dt ω
−∞ 4π/dt

Figure 1.8: PSD for signal b0 (t) (logarithmic scales).

1.2.7 White noise

Lastly, a white noise is a random signal with an infinite variance and whose auto-
correlation function is proportional to a Dirac function (that is a constant PSD for
any pulsation ω). This profile expresses that two values of the signal, even sampled
at two very closed instants, are not at all correlated.
Central gaussian white noise w(t) and v(t) we are going to used in Kalman
filtering are then entirely defined by their respective PSD W (t) and V (t):

E[w(t)w(t + τ )T ] = W (t)δ(τ ), E[v(t)v(t + τ )T ] = V (t)δ(τ ) (1.11)

Matrices W (t) and V (t) become constant in the case of stationary gaussian white
noises. The normalized gaussian white noise is such that W (t) = Iq×q (q components
in the noise).

Remark 1.4 From a practical point of view, it is not possible to simulate, on a

numerical computer, a perfect continuous-time white noise characterized by a finite
PSD R (but with an infinite variance). The approximation proposed in example 1.2
(see Figure 1.8), which consists in holding, over a sample period dt (which must be
chosen very low w.r.t. the settling time of the system), a gaussian random variable
with a variance σ 2 = R/dt, will be used. This approximation corresponds to the
”Band-limited white-noise” proposed in Matlab/Simulink).

Introduction to Kalman filtering Page 20/80

1.3 Transmission of random signals in linear systems 21

1.3 Transmission of random signals in linear sys-

tems
1.3.1 Time-domain approach
The assumptions:

• the model is linear (equation 1.1),

• and w is a centered gaussian white noise

allows to assert that the state x and the output y are also gaussian multivariate
signals and are therefore entirely characterized by first and second moments. The
following theorem will allow these characteristics to be computed. It is assumed
that the deterministic input u is null (u(t) = 0).

Theorem 1.1 let us consider the linear system:

˙ = Ax(t) + M w(t) .
x(t) (1.12)
w(t) is a centered stationary Gaussian white noise with a PSD W . m(t0 ) et P (t0 ) are
the mean value and the covariance matrix of the initial state x(t0 ) (also a gaussian
random variable). Then x(t) is a gaussian random signal with:

• mean vector:
m(t) = E[x(t)] = eA(t−t0 ) m(t0 )

• covariance matrix P (t) = E[(x(t) − m(t))(x(t) − m(t))T ] solution of the dif-

ferential equation:
˙ = AP (t) + P (t)AT + M W M T .
P (t) (1.13)
If the system is stable (all the eigenvalues of A have a negative real part), then
a steady state (stationary) response can be reached : Ṗ = 0 and P (t) = P
solves the continuous-time Lyapunov equation :
AP + P AT + M W M T = 0 . (1.14)

Proof : (see annex B.2).

Remark 1.5 If the output equation is considered without noise i.e. y(t) = Cx(t)
then the covariance matrix Py (t) of y(t) is:
Py (t) = CP (t)C T
(if the white noise on the measurement is considered i.e. y(t) = Cx(t) + v(t), then
the covariance of y is infinite).

Introduction to Kalman filtering Page 21/80

22 1. Random signal in linear systems

1.3.2 Frequency-domain (spectral) approach

Only the steady state is considered here.

Theorem 1.2 (Transmission of a white noise in the linear system) Let us con-
sider a stable linear continuous-time system defined by the transfer matrix G(s)p×q
between the input w and the output y. The steady state response y to a station-
ary random signal w, characterized by a spectrum Φww (s)q×q in the s-plane, is a
stationary random signal characterized by a spectrum Φyy (s) in the s-plane such
that:
Φyy (s)p×p = G(−s)Φww (s)GT (s) .

Proof : (voir annexe B.3).

In the case of a SISO (Single Input Single Output) transfer G, this result can
be depicted by Figure 1.9.

G(s)

Figure 1.9: SISO case.

Reciprocally : Let us consider a random signal w(t) with a given real 8

spectrum in the s-plane Φww (s), then the decomposition (such a decomposition in
not unique as it is shown in example 1.3):
Φww (s) = G(−s)GT (s)
where G(s) is the transfer gathering all the stable poles of Φww (s), allows to provide
a Marlov model of the signal w(t), i.e. a state space realization of the ”generator”
filter G(s). This state space realization is denoted (G(s) is assumed to be strictly
proper): 
xG˙(t) = AG xG (t) + BG b(t)
w(t) = CG xG (t)
where b(t) is a normalized gaussian white noise such that Φbb (s) = Iq×q .

Exercise 1.3 Let us consider again example 1.1 on the 1/4 vehicle model. Some
experiments in a real environment have shown the PSD of the rolling noise for a
vehicle at 90 km/h on a secondary road could be approximated by:
ω 2 + 104
Φww (ω) =
ω 4 − 1.75 104 ω 2 + 108
8
That is: a spectrum with only exponents of s2 ; that is: the correspondent auto-correlation
function φww (τ ) is pair.

Introduction to Kalman filtering Page 22/80

1.3 Transmission of random signals in linear systems 23

a) Give the spectrum in the s-plane for w(t).

b) Give a Markov representation of w(t).

c) Compute the variance of w(t) (the signal is assumed to be centered).

Solution :
−s2 +104
a) By definition: Φww (s) = s4 +1.75 104 s2 +108
.

b) A decomposition Φww (s) = G(−s)G(s) is:

  
102 + s 102 − s
Φww (s) = .
s2 − 50s + 104 s2 + 50s + 104

All the stable poles of Φww (s) must be gathered in G(s), that allows the de-
nominator of G(s) to be determined but there is no condition on the stability
of G(s) zeros. So there are 2 different filters that provide an output signal with
the spectrum Φww (s) when the input is a normalized white noise:

102 − s 102 + s
G(s) = and G0 (s) =
s2 + 50s + 104 s2 + 50s + 104
If G(s) is chosen then a Markov representation of w(t) is (an horizontal
companion form is proposed, for instance):
    
 ˙ 0 1 0
xG (t) = 4 xG (t) + b(t)
−10 −50 1 . (1.15)
 2
w(t) = [10 − 1]xG (t)

c) The variance can be computed in two different ways:

– from the spectrum in the s-plane and the initial value theorem (see remark
1.2):

1/50 s + 1/2 −1/50 s + 1/2 −

Φww (s) = 2 4
+ 2 = Φ+
ww (s) + Φww (−s)
s + 50s + 10 s − 50s + 104

σw2 = φww (τ )|τ =0 = lim sΦ+

ww (s) = 1/50.
s→∞

– from the Markov model and the theorem 1.1: during steady-state, the
covariance matrix P of the state vector xG is the solution of the Lya-
punov equation:
       
0 1 0 −104 0 0 0 0
P +P + =
−104 −50 1 −50 0 1 0 0

Introduction to Kalman filtering Page 23/80

24 1. Random signal in linear systems

 
10−6 0
⇒ P =
0 10−2
 −6  2 
2 2 10 0 10
⇒ σw = [10 − 1] −2 = 1/50 .
0 10 −1

1.4 Matlab illustration

The file bruit.m, given below, allows to illustrate the way to use Matlab macro-
functions to solve equation 1.3. This file also shows how to simulate, on a numerical
computer, such a colored noise from a pseudo-white noise generated with a sample
period dt. The Simulink file simule_bruit.mdl depicted in Figure 1.10 9 can be
simulated. Responses of pseudo-white noise b(t) and rolling noise w(t) are plotted in
Figures 1.11 and 1.12, respectively. These responses highlight that the variance of
w(t) est independent of dt (since the sampling pulsation 2π/dt is fast w.r.t the filter
G(s) dynamics) while the variance of b(t) is equal to 1/dt: one can see that this
variance tends toward infinity if dt tend to zero as the variance of a continuous-time
white noise is infinite (but the PSD of b(t) is independent of dt and is equal to 1).
The example B.1 proposed in appendix completes this illustration by a discrete-
time analysis (with a sampling period dt).

%=============================================================
clear all close all
% Filter definition:
G=tf([-1 100],[1 50 10000])
% State space realization:
[A,B,C,D]=ssdata(G);
% Variance determination using Lyapunov equation:
P=lyap(A,B*B’); var_w=C*P*C’
% ==> on can find the wanted result: variance=1/50.
% (standard deviation: sigma=0.14)

% Validation using a simulation: see SIMULINK file: simule_bruit.mdl

% Choice of a sampling period fast enough w.r.t. the filter dynamics (100 rd/s):
dt=0.0001;
% Simulation:
9
Feel free to send an e-mail to [email protected] with ”Introduction Kalman” for the subject
if you want a copy of the 2 files bruit.m and simule bruit.mdl.

Introduction to Kalman filtering Page 24/80

1.5 Conclusion 25

sim(’simule_bruit’);
% Results plotting:
plot(b.time,b.signals.values,’k’);
% Numerical computation of the variance of b(t):
var(b.signals.values) % One can find again: var_b=1/dt=10000.
figure
plot(w.time,w.signals.values,’k’)
% Numerical computation of variance of w(t):
var(w.signals.values) % One can find: var_w=1/50 (approximately).

% The sampling period is now multiplied by a factor 10:

dt=0.001;
sim(’simule_bruit’);
figure(1) hold on
plot(b.time,b.signals.values,’g-’);
var(b.signals.values) % One can find again: var_b=1/dt=1000.
figure(2) hold on
plot(w.time,w.signals.values,’g-’)
var(w.signals.values) % One can find: var_w=1/50 (approximately).
%=============================================================

b(t) −s+100 w(t)

w
s2+50s+10000
Band−Limited G(s) To Workspace
White Noise:
Noise power =1
Sample time=dt b

To Workspace1

Figure 1.10: SIMULINK file simule bruit.mdl for the rolling noise simulation.

1.5 Conclusion
Mathematical tools used to analyse continuous-time random signals and their trans-
mission in linear systems were presented in this chapter. The notion of gaussian
white noise was also introduced (this assumption is required in the Kalman filter
developed in next chapter). The interest of such gaussian (centered) noises lies in
the fact that they are entirely characterized by their autocorrelation function (or
their spectrums in the s-plane if they are stationary).
The reader will find in appendix B some additional developments for the anal-
ysis of discrete-time random signals.

Introduction to Kalman filtering Page 25/80

26 1. Random signal in linear systems

400

300

200

100
b(t)

−100

−200

−300

−400
0 0.5 1 1.5 2
Temps (s)

Figure 1.11: Responses of normalized white-noise b(t) (with dt = 0.0001 s: black;

with dt = 0.001 s: grey).

0.4
2σ =0.28
w
0.3

0.2

0.1

0
w(t)

−0.1

−0.2

−0.3

−0.4

−0.5
0 0.5 1 1.5 2
Temps (s)

Figure 1.12: Responses of rolling noise w(t) (with dt = 0.0001 s: black; with dt =
0.001 s: grey).

Introduction to Kalman filtering Page 26/80

 27

Chapter 2

Kalman filter

2.1 Principle of Kalman filtering

2.1.1 Kalman model
We consider again the model presented at the beginning of chapter 1. This model
involves deterministic inputs u(t) and random inputs w(t) and v(t). So we will as-
sume that the model of the perturbed plant can be always described by the following
state space realization, also called Kalman model:

˙ = Ax(t) + Bu(t) + M w(t) state equation, x ∈ Rn , u ∈ Rm , w ∈ Rq
x(t)
.
y(t) = Cx(t) + Du(t) + v(t) measurement equation, y ∈ Rp , v ∈ Rp
(2.1)
The objective of Kalman filtering is to estimate the plant state x(t). This estimate,
commonly noted x b(t), is the output of the Kalman filter.
The Kalman model (2.1) suggests an input-output oriented system between
deteministic input u and output y. In the field of signal processing, this input-output
relationship can be quite different than the one used in servo-loop theory in that
sense that u is not necessary the system control signal. Indeed, and more particularly
in the design of navigation filter for aerospace vehicles, the Kalman model (2.1) can
be used to describe the relationship between two kinds of measurements: primary
measurements (gathered in vector u) and secondary mesurements (gathered in vector
y). Then, the objective of the Kalman filter is to perform the “best” measurement
fusion between primary and secondary measurements.

Example 2.1 Let us consider the motion of an aerospace vehicle (with a mass m)
along a single translation degree-of-freedom. Its position along this degree-of-freedom
is named p(t). To control the postion p(t), the spacecraft is fitted with an actuator
(thruster). This actuator can be considered itself as a dynamic sub-system with
a transfer function A(s) between its input uc (t) (in fact the output of the control

Introduction to Kalman filtering Page 27/80

28 2. Kalman filter

algorithm) and the output F (t): the real thrust applied to the spacecraft (N ). The
spacecraft is instrumented with:
d2 p(t)
• a linear accelorometer which measures the acceleration p̈(t) = dt2
with a
noise w(t), that is: p̈m (t) = p̈(t) + w(t),

• a position sensor (star tracker) which measures the position p(t) with a noise
v(t), that is: pm (t) = p(t) + v(t).

The control algorithm (out of the scope of this document) requires estimates of the
position pb(t) and the velocity vb(t) = ḃp(t) of the spacecraft in order to compute the
control signal uc (t) to be sent to the actuator. Note that the velocity is not measured.
To solve such a problem, one can use a Kalman filter involving the accelerometer
as primary measurement u(t) = p̈m (t) (seen as the input of the Kalman model)
and the position sensor as secondary mesurement y(t) = pm (t) (seen as the output
of the Kalman model). Obviously, the Kalman model between u (acceleration)
and y (position) is a double integration and reads:

ṗ(t) = v(t)
v̇(t) = p̈(t) = u(t) − v(t)
y(t) = p(t) + v(t)

or using the general state-space representation (2.1) with x(t) = [p(t) v(t)]T :
      
 ˙ 0 1 0 0
x(t) = x(t) + u(t) + w(t)
 0 0 1 −1 . (2.2)

y(t) = 1 0 x(t) + 0u(t) + v(t)

The block-diagram sketch of the whole system is represented in Figure 2.1. Of course
such a Kalman filter cannot be used to estimate the internal state of the actua-
tor sub-system A(s) as the model of the actuator is not taken into account in the
Kalman model.

2.1.2 Assumptions:
The Kalman model (2.1) must meet the following assumptions:

H1: the (A, C) pair is detectable, i.e. there is no unstable and unobservable eigen-
value in the model,

H2: signals w(t) and v(t) are centered gaussian white noise with Power Spec-
tral Densities (PSD) W and V respectively, that is:

Introduction to Kalman filtering Page 28/80

2.1 Principle of Kalman filtering 29

uc F p̈ ṗ p
Control Actuator: A(s) 1 1
algoritm 1/m s s

w + p̈m = u
pb
+ (primary)
Kalman
v + pm = y filter vb
+ (secondary)
Plant model

Figure 2.1: Block-diagram sketch for the control of an aerospace vehicle using a
Kalman filter to hybridize primary (acceleration) and secondary (posi-
tion) measurements.

– E[w(t) w(t + τ )T ] = W δ(τ ),

– E[v(t) v(t + τ )T ] = V δ(τ )
– E[w(t) v(t + τ )T ] = 0 (this last equality expresses the stochastic indepen-
dence of noises w(t) and v(t) : this assumption is introduced for reason
of clearness but is not required in the general formulation of Kalman
filter: the reader will find in reference [5] this general formulation taking
into account a correlation between state noise and measurement noise).

H3: V is invertible (i.e. there is as much independent noise sources as measure-

ments or each measurement is perturbed by its own noise 1 ).

Remarks:

• Although the whole theory of Kalman filter can be applied to the non-
stationary case (for the plant and the noises), we will assume in the following
that the plant and noises are stationary, i.e: matrices A, B, M , C, D, W are
V independent of time t.

• The mean of a random signal, which is also called a bias, is considered to be

deterministic and must be, if it is, extracted from noises w(t) or v(t) to meet
the assumption H2 (w and v are centered noises). For instance, if random
signal w(t) in state equation (2.1) is biased and if this bias E[w(t)] is known,
then the Kalman filter will be designed from the following model:
  
 ˙ u(t)
x(t) = Ax(t) + [B M ] + M (w(t) − E[w(t)])
E[w(t)] .

y(t) = Cx(t) + Du(t) + v(t)
1
V is a positive definite matrix and W is a semi-definite positive matrix.

Introduction to Kalman filtering Page 29/80

30 2. Kalman filter

The new state noise w̄(t) = w(t) − E[w(t)] is now centered. If the bias E[w(t)]
is unknown then it can be modelled as an initial condition on a new integral
state and the Kalman filter will allow this bias to be estimated (see example
in section 2.3).

• If noises are colored and characterized by spectrums in the s-plane then re-
sults of section 1.3.2 will allow the ”color” (or frequency response) of the noises
to be taken into account with a Kalman model augmented by the Markov
representation of these noises. For instance: if the spectrum in the s-plane
Φww (s) of the centered random signal w(t) in (2.1) is known, then the decom-
position Φww (s) = G(−s)GT (s) will allow a Markov representation of w(t)
to be determined, that is: a state space realization of G(s) (see example 1.3):

xG˙(t) = AG xG (t) + BG b(t)
w(t) = CG xG (t)

where b(t) is a centered random signal with a unitary de spectrum in the s-

plane
Φbb (s) = Iq×q (i.e. a normalised white noise).
The augmented model:
 " #       

 ˙
x(t) A M CG x(t) B 0

 = + u(t) + b(t)
xG˙(t) 0 AG xG (t) 0 BG
 

 x(t)

 y(t) = [C 0]
xG (t)

meets now all the assumptions of the Kalman model.

In fact all the deterministic information it is possible to know in the system must
be gather in the model (i.e. ẋ = Ax + Bu, y = Cx + Du and matrix M ). All
the random information must be gather in noises w(t) and v(t). The state noise
wx = M w represents all the external perturbations (wind in the case of an aircraft,
road unevenness in the case of a car, ...) and/or also modelling errors or uncertain-
ties (difference, due to linearization, between the tangent model and the non-linear
model, neglected dynamics, ...): wx is an upper bound of all what makes that the
state does not evolve exactly as the deterministic model predicts (ẋ = Ax + Bu).

2.1.3 Structure of an unbiased estimator

A Kalman filter is a dynamic system with 2 (vector) inputs: the deterministic
control signal u and the measurement y, that is: all known signals of the plant. The
b (or the output) of this filter is an estimate of the state x of the plant.
state x

Introduction to Kalman filtering Page 30/80

2.1 Principle of Kalman filtering 31

Let:
 
u(t)
ḃ
x(t) b(t) + [Bf Kf ]
= Af x (2.3)
y(t)
b(t) + Bf u(t) + Kf y(t)
= Af x (2.4)

be the state space realization of this filter. Of course this filter must be initialized
with x b(t0 ): the estimate of the state of the plant at the initial time t0 .
Let us denote ε(t) = x(t) − x b(t) the state estimation error and ε(t0 ) = x(t0 ) −
b(t0 ) the initialization error.
x
Substituting the equation (2.4) to the state equation of (2.1) and using the
measurement equation, we can write:
b − Bf u − Kf (Cx + Du + v)
ε̇ = Ax + Bu + M w − Af x
b + (B − Bf − Kf D)u + M w − Kf v
= (A − Kf C)x − Af x
= (A − Kf C)ε + (A − Kf C − Af )b
x + (B − Kf D − Bf )u + M w − Kf v . (2.5)

As noises w and v are gaussian and as the system is linear, on can state that
ε(t) is a gaussian random signal. We are going to characterize the expected value
of ε(t).
Unbiased estimator: first of all, we wish the estimator to be unbiased, that is:

• whatever the input profile u(τ ) applied in the time interval τ ∈ [t0 , t],

b(t0 ),
• whatever the initialization x

we wish the estimation error expected value tends towards 0 when t tends towards
infinity.
As noises w and v are centered, we can write:
˙
E[ε(t)] = E[ε̇(t)] = (A−Kf C)E[ε(t)]+(A−Kf C −Af )E[b
x(t)]+(B −Kf D−Bf )u(t)

Then limt→∞ E[ε(t)] = 0, ∀ u(t), ∀ E[b

x(t)], if and only if:

Af = A − Kf C, B f = B − Kf D (2.6)

and A − Kf C is stable. (2.7)

Indeed, from theorem 1.1 (page 21):

E[ε(t)] = e(A−Kf C)(t−t0 ) ε(t0 ) and lim E[ε(t)] = 0 .

t→∞

If (2.6) is used in (2.4), the Kalman filter realization becomes:

ḃ = (Ab
x x + Bu) + Kf (y − Cb
x − Du) . (2.8)

Introduction to Kalman filtering Page 31/80

32 2. Kalman filter

We recognize, in the first term of the right-hand member of equation (2.8), the deter-
ministic model of the plant (Ab x + Bu). This model is used to predict the evolution
of the plant state from the current estimation xb. This prediction is in fact an in-line
simulation of the plant model. This model being wrong, the prediction is corrected
(updated) with the difference between the measurement y and the prediction of the
measurement yb = Cb x + Du through the filter gain Kf . This difference y − yb is
also called the innovation. The block-diagram of such a filter (in the case D = 0)
is depicted in Figure 2.2. This structure ensures that the estimator is unbiased
whatever the matrices A, B, C, D of the plant and the gain Kf such that A − Kf C
is stable (that justifies the assumption H1: the existence of an unstable and unob-
servable mode in the plant does not allow a stabilizing gain Kf to be determined
and so an unbiased estimator to be built).

u y + −
Plant

Kf
B x^
Z
C
+
+
+

A
Kalman filter
^
x

Figure 2.2: Kalman filter block diagram (case D = 0).

2.2 Minimal error variance estimator

The gain Kf is computed according to the confidence we have in the model (this
confidence is expressed by the PSD W : the lower this confidence is, the greater W
must be) with respect to the confidence we have in the measurement (expressed
by the PSD V ). If the model is very precise (W is very low) and if the measurement
is very noisy (V is very great) then the gain Kf must be very low (the prediction
is good). In fact, among all the stabilizing gains Kf , we are going to choose the
one that minimizes the variance of the state estimation error ε(t) ∀t (in fact
the sum of estimation error variances for all components of the state). We recall
(see previous section) that ε(t) = x(t) − xb(t) is a multivariate (with n components),
centered (unbiased), gaussian random signal. The gaussian feature of this centered
signal allowes to state that if the variance of the estimation error is really minimized
b(t) is the best estimate of x(t).
then, x

Introduction to Kalman filtering Page 32/80

2.2 Minimal error variance estimator 33

2.2.1 General solution

So we seek Kf minimizing:
X
n
J(t) = E[εi (t)2 ] = E[εT (t)ε(t)] (2.9)
i=1
= trace E[ε(t)εT (t)] (2.10)
= trace P (t) . (2.11)

b(t))(x(t) − x
P (t) = E[(x(t) − x b(t))T ] is the estimation error covariance matrix.
Returning (2.6) in (2.5), the evolution of ε(t) is governed by the following state
space equation:
 
w(t)
ε̇(t) = (A − Kf C)ε(t) + [M − K] , (2.12)
v(t)

with:     
w(t) T T Wq×q 0q×p
E [w (t + τ ) v (t + τ )] = δ(τ ) .
v(t) 0p×q Vp×p
The theorem 1.1 (page 21) can be applied and allows to conclude that the estimation
error covariance P (t) is governed by the differential equation:
  
T W 0 MT
Ṗ (t) = (A − Kf C)P (t) + P (t)(A − Kf C) + [M − Kf ]
0 V −KfT
= (A − Kf C)P (t) + P (t)(A − Kf C)T + M W M T + Kf V KfT . (2.13)

To minimize trace P (t) (w.r.t. Kf and ∀ t > t0 ), it is sufficient to minimize

˙ :2
trace P (t)
˙
∂(trace P (t))
= −P (t)C T − P (t)C T + 2Kf V
∂Kf
⇒ Kf (t) = P (t)C T V −1 . (2.14)
From(2.14) and (2.13), we get:

Ṗ (t) = AP (t) + P (t)AT − P (t)C T V −1 CP (t) + M W M T . (2.15)

This Riccati differential equation must be integrated from t0 to t and must be

initialized with P (t0 ) that expresses the confidence we have in the filter ini-
tialization xb(t0 ):

b(t0 ))T ] .
b(t0 ))(x(t0 ) − x
P (t0 ) = E[(x(t0 ) − x
2
Indead: starting from the prescribed initial condition P (t0 ), to minimize trace P (t) ∀ t > t0 ,
it is sufficient to minimize its time-domain derivative at each time t > t0 .

Introduction to Kalman filtering Page 33/80

34 2. Kalman filter

In appendix C, an analytical solution for the general matricial differential Riccati

equation (2.15) is proposed and illustrated on an academic example.
The gain Kf is then computed from P (t) and equation (2.14). The Kalman
filter is thus an un-stationary system.
Equations (2.8), (2.14) and (2.15) constitute the continuous-time Kalman fil-
ter equations and must be integrated from initialization x b(t0 ) and P (t0 ). The in-
tegration of (2.15) and the computation of Kf (t) (2.14) can be performed in-line
or off-line. In the last case, the time-law Kf (t) must be uploaded in the on-board
computer. From a practical point of view, the implementation of the Kalman filter
will be performed in discrete-time on a numerical computer. The Kalman filter
state equation (2.8) can be discretized (integration using rectangular (Euler or
trapezoidal (Tustin approximations, ..., if only the steady state behavior is consid-
ered). One can also prefer to directly design the Kalman filter in discrete-time (see
section 2.4.2). Lastly, all the filter equations are defined by the data of the problem,
that is, the set of matrices A, B, M , C, D, W and V .

2.2.2 Kalman filter steady state

During the steady state (that is: once the transient response due to initialization
error is finished), the estimation error becomes a stationary random signal (this is
the case for all signals in the block diagram of Figure 2.2). So we can write:

Ṗ (t) = 0 .

P is a constant positive definite matrix (the covariance of the state estimation er-
ror in steady state) and is the only positive solution of the algebraic Riccati
equation:
AP + P AT − P C T V −1 CP + M W M T = 0 (2.16)
The filter gain Kf = P C T V −1 becomes also stationary.
On can easily verify (from the steady state of Lyapunov equation (2.13)) that
the positiveness of P implies the stability of the filter, that is all the eigenvalues of
A − Kf C have a negative real part.
The reader will find in [1], a general method to compute the positive solution
of an algebraic Riccati equation. From a practical point of view, keep in mind
that such solvers are available in most softwares for computer-based control design:
macro-function lqe of Matlab or Scilab control packages. This function provides
directly P and Kf from matrices A, B, M , C, D, W and V (see also macro-function
care and kalman under Matlab).

Introduction to Kalman filtering Page 34/80

2.2 Minimal error variance estimator 35

2.2.3 Kalman filter tuning

For a given model (matrices A, B, M , C, D), the filter gain Kf and its evolution as
a function of time depend only on:

• W : the ”confidence” we have in the state equation,

• V : the ”confidence” we have in the measurement,

• P (t0 ): the ”confidence” we have in the initialization.

In steady state, if the system and the noises are stationary, the Kalman gain Kf
is constant and its value depends only on W and V .
Furthermore, the gain Kf and the estimation error response ε(t) depend only on
the relative weight of P (t0 ) w.r.t. V (during the transient response) and the relative
weight of W w.r.t. V (during the steady state). Indeed, it can be easy to check that
the gain Kf (and so the filter) does not change if these 3 tuning matrices W , V and
P (t0 ) are multiplied by a scaling factor α. But keep in mind that the estimation
error covariance P will be multiplied by α. So the designer must be sure that these
3 matrices are realistic if he wants to use P (t) to analyse the estimation quality and
conclude for instance:
p the probability forpthe i-th component of the state xi (t) to be
between x bi (t) − 2 P(i,i) (t) and x
bi (t) + 2 P(i,i) (t) is equal to 0.95 (gaussian variable
property). From a practical point of point, the Kalman filter must be validate on
a validation model which takes into account realistic measurement noises and a
modelling, as accurate as possible, of all the disturbances (non-linearities, external
perturbations,...) that have been overvalued by a state noise w(t) in the Kalman
model also called synthesis model. Such a validation model cannot be used as
synthesis model for 2 reasons: fisrtly, it does not generally fulfill the assumptions
H1, H2, H3 and secondly it will lead to a too much complex Kalman filter from
the implementation point of view.
Whatever the tuning, one can check that the Kalman estimation error covari-
ance P (t) is always lower than the covariance propagated in the state equation (that
is without using measurement and governed by Ṗ = AP + P AT + M W M T ) and the
estimation error covariance of the output yp (without noise: yp = Cx + Du), that is:
CP (t)C T , is always lower that the measurement noise covariance (which is infinite
in the continuous-time case !! but this property is also true in the discrete-time case
where the measurement noise covariance is finite, see section 2.4). So it is always
better to use ybp = Cb x + Du rather than the direct measurement y to estimate the
real output yp of the plant.
The designer could use the following trade-off to tune qualitatively the estima-
tion error time-response.
Influence of P (t0 ) ./. V : during the transient response, the initial estimation
error ε0 will be corrected as faster (and the Kalman gain will be as greater) as

Introduction to Kalman filtering Page 35/80

36 2. Kalman filter

P (t0 ) is greater w.r.t. V . But the estimate will be affected by the measurement
noise because we have a good confidence in this measurement.
Influence of W ./. V : during the steady state, the gain Kf will be as lower,
and the estimation response will be as smoother, as W is weaker w.r.t. V (we have
confidence in the model). But, if the plant is subject to a perturbation which has
been under-estimated by this too weak choice of W , the estimate x b will not track
the real state x or will have a too long updating time. Furthermore the estimation
error covariance P will not be representative of this phenomena.
These behaviors will be illustrated in the following example. Although this is a
first order example, it allows to highlight general trade-offs that can be encountered
on more complex applications (continuous or discrete-time).

2.3 Corrected exercises

2.3.1 First order system
Statement: let us consider a stable first order system
1
G(s) = with a < 0 .
s−a
The measurement y of the system output x is perturbed by a white noise v(t) with
a unitary PSD (V = 1). The input of this system is subject to an unknown and
variable transmission delay (the maximal value of this delay is denoted T ). For the
design of the Kalman filter we want to perform on this system, this perturbation
(unknown delay) will be roughly modelled by a white noise w(t), with a PSD W ,
independent of v, acting directly on the input signal according to Figure 2.3.

w v
u + 1 x + y
+ s−a +

Figure 2.3: Kalman synthesis model.

Questions:

a) Set up the Kalman model for this problem and compute the Kalman filter,
to estimate the state (or output) x of the plant. Give results as a function
of a, W and P0 (the initial estimation error variance). Study the asymptotic
behavior of the filter gain Kf when W tends towards 0.

b) Build a simulator using M atlab/Simulink (involving a validation model with

a real transmission delay T ) in order to analyse the influence of parameters W

Introduction to Kalman filtering Page 36/80

2.3 Corrected exercises 37

b(t) and on its variance

and P0 on the estimation error response ε(t) = x(t) − x
P (t).
Numerical application:

b0 = 0,
• a = −1, x0 = 20, x
• T = 0, 0.1, 0.5, 1(s) (these various values will be tested),
• u(t): square signal with a magnitude of 10 and a period of 5 s,
• dt = 0.01 s (integration step for the simulator).

c) What do you propose to improve the estimation quality in the case of an

important and known delay T (for instance: T = 1 s) ?

Correction: Question a): the Kalman model is:


ẋ(t) = ax(t) + u(t) + w(t)
(n = 1; m = 1; p = 1),
y(t) = x(t) + v(t)

with E[w(t)w(t + τ )] = W δ(τ ), E[v(t)v(t + τ )] = δ(τ ) and E[v(t)v(t + τ )] = 0.

The first step in the determination of the Kalman filter consists in solving the
differential Riccati equation (2.15) which is a scalar equation in this case:

Ṗ (t) = 2aP (t) − P (t)2 + W . (2.17)

In appendix C, an analytical solution for the general matricial differential Riccati

equation is proposed and illustrated (see section C.2) on this example. The reader
will also find in ([8]) a general method to solve such a differential equation in the
scalar case. This method consists in:

• looking for a constant particular solution (because the coefficients of the dif-
ferential equation do not depend on time t). This solution corresponds in fact
to steady-state solution p = p∞ which must solve:
2
P∞ − 2aP∞ − W = 0 .
√
The only positive solution is P∞ = a + a2 + W ,

• looking for a general solution by quadrature, that is a solution of the form:

1
P (t) = P∞ + .
z(t)

The expansion of Ṗ (t) leads to:

√
ż(t) = 2 a2 + W z(t) + 1 .

Introduction to Kalman filtering Page 37/80

38 2. Kalman filter

The integration of this differential equation using the varying constant method
gives: h √ i
1 √
2 2
z(t) = √ e2 a +W (t−t0 ) − 1 + e2 a +W (t−t0 ) z0
2
2 a +W
where t0 is the initial time and z0 = z(t0 ) is the initial condition on z(t) :
1
z0 =
P0 − P ∞
√
where P0 is the initial condition on P (t). Let us denote k = 2 a2 + W , then:
k(P0 − P∞ )
P (t) = P∞ +
(P0 − P∞ )(ek(t−t0 ) − 1) + kek(t−t0 )
k(P0 − P∞ )
= P∞ + k(t−t0 ) . (2.18)
e (P0 − P∞ + k) + P∞ − P0
Lastly: Kf (t) = P (t).

If W = 0 then P∞ = a + |a| = 0 if a < 0 (stable system)

⇒ lim Kf = 0 .
t→∞

Once the initial error is corrected, only the model is used to estimate x. This is
logic because the model is assumed to be perfect when W is set to 0. That property
holds only if the system is stable. Indeed, if the system is unstable (a > 0) then:

⇒ lim Kf = 2a .
t→∞

This is the value of the gain which allows the filter dynamics to be stable while
minimizing the effect of the measurement noise on the estimation error variance.
Question b): the Matlab/Simulink simulator is depicted in Figure 2.4. The
various parameters used in this Simulink file (simuKalman.mdl) are written in this
Figure. The Matlab function Kf_t.m used to implement the non-stationary gain
Kf is given in appendix D with the script-file demoKalman.m to declare the various
parameters and to plot results 3 . Different simulation results are presented in Figures
2.5 to 2.9 :

• Figure 2.5: The confidence P0 in the initialization is not at all representative

b0 =
of the initial estimation error which is in fact very important (ε0 = x0 − x
20). The transient response to correct
p this initial error is thus long and the
estimation of x ±2σ (with σ(t) = P (t)) does not allow to frame the real
value of x during this transient response.
3
Thanks to send an email to [email protected] with ”Introduction Kalman”” for the subject
if you wish a copy of these files.

Introduction to Kalman filtering Page 38/80

2.3 Corrected exercises 39

• Figure 2.6: If this confidence in the initialisation is degraded (P0 = 100),

this transient response becomes more faster. The measurement is now more
exploited by the filter. The estimate is therefore a little bit noisy during this
transient response. In steady state, this estimate becomes smoother because
we are confident in the model (W is low). One can note that the estimate
is unbiased. Lastly the estimation is good because in this simulation, the
validation model (with T = 0) coincides exactly with the Kalman synthesis
model.

• Figure 2.7: If now a 1 s delay is taken into account in the validation model,
the (good) confidence we have in the synthesis model (W = 1) does not allow
to be representative of real model errors. The filter is confident in a wrong
model and does not use the measurement enough: this filter has a very long
reaction time.

• Figure 2.8: If now it is specified that the model is not so good (W = 100), the
filter is more confident with measurements: the estimate is more sensitive to
the measurement noise but have a good reaction time.

• Figure 2.9 (answer to question c)): if now the delay is known (T = 1 s), it
is possible to take this delay into account in the Kalman synthesis model in
two different ways: firstly in propagating this delay in the prediction equation
or secondly (what it is proposed here: see file demoKalman.m) in using a Pade
filter in series with the model.

Introduction to Kalman filtering Page 39/80

40 2. Kalman filter

v
SIMULATION PARAMETERS
Band−Limited Start time: 0s Stop time: 10s
White Noise: Solver: ode4 (Runge Kutta)
Noise power=V; Fixed step=dt
Sample Time=dt; y

u x
x’ = Ax+Bu
y = Cx+Du
Signal Transport Delay 1/(s+1)
Generator Time Delay=T A=−1;B=1;C=1;
D=0;X0

Scope
Clock
output

MATLAB To Workspace
Kf_t Function StructureWithTime

1
b c
s
Integrator
Initial Condition=0

Figure 2.4: SIMULINK file simuKalman.mdl for the Kalman filter simulation.

50
y(t)
x(t)
u(t)
40
xhat(t)
xhat(t)+2 σ
xhat(t)−2 σ
30

20

10

−10

−20

−30

−40
0 1 2 3 4 5 6 7 8 9 10

Figure 2.5: Simulation with P0 = 1, W = 1, T = 0.

Introduction to Kalman filtering Page 40/80

2.3 Corrected exercises 41

50
y(t)
x(t)
u(t)
40
xhat(t)
xhat(t)+2 σ
xhat(t)−2 σ
30

20

10

−10

−20

−30

−40
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Figure 2.6: Simulation with P0 = 100, W = 1, T = 0 (zoom around the transient

response) .

50
y(t)
x(t)
40 u(t)
xhat(t)
xhat(t)+2 σ
30 xhat(t)−2 σ

20

10

−10

−20

−30

−40

−50
0 1 2 3 4 5 6 7 8 9 10

Figure 2.7: Simulation with P0 = 100, W = 1, T = 1.

Introduction to Kalman filtering Page 41/80

42 2. Kalman filter

50
y(t)
x(t)
40 u(t)
xhat(t)
xhat(t)+2 σ
30 xhat(t)−2 σ

20

10

−10

−20

−30

−40

−50
0 1 2 3 4 5 6 7 8 9 10

Figure 2.8: Simulation with P0 = 100, W = 100, T = 1.

40
y(t)
x(t)
u(t)
30
xhat(t)
xhat(t)+2 σ
xhat(t)−2 σ
20

10

−10

−20

−30

−40

−50
0 1 2 3 4 5 6 7 8 9 10

Figure 2.9: Simulation of the stationary Kalman filter designed on a model taking
into account a second order Pade approximation of the delay (W = 1,
T = 1 s).

Introduction to Kalman filtering Page 42/80

2.3 Corrected exercises 43

2.3.2 Bias estimation

Statement: a vehicle is moving along an Ox axis. The velocity ẋ and the position
x of this vehicle are measured. These measures are denoted vm and xm , respectively.
The measurement xm is perturbed by a centered gaussian white noise v(t) with a
unitary PSD V = 1 (m2 /Hz) : xm (t) = x(t) + v(t).
The measurement vm is biased by a signal b(t) which can be modelled as a step
˙ + b(t).
function with a unknown magnitude b0 : vm (t) = x(t)
From these 2 measurements vm et xm , we want to built a Kalman filter to estimate
the position x(t) and the bias b(t).

1) Find the state space equations of the Kalman model with x and b as state
variables, vm as input variable, xm as output variable.

2) Draw a block diagram representation of this model.

In fact the bias b(t) can derive with time. To take into account these variations, it
is assumed that the time-derivative of the bias is perturbed by a white noise w(t)
with a PSD W = q 2 (independent of v(t)):
˙ = w(t) .
b(t)

3) Give the new Kalman model equation, compute the steady state Kalman
gain (as a function of q) and give the state space representation of the filter
b and bb to be computed from xm and vm .
allowing estimates x
ḃ
4) How would you proceed to estimate the velocity of the vehicle x?

5) Compute the transfer matrix F (s) of the filter:

" #  
b
X(s) Xm (s)
ḃ = F (s) .
X(s) Vm (s)

ḃ
6) Comment this transfer (particularly X(s)/V m (s)) as a function of q. Demon-
strate that this filter F (s) provides perfect estimates if the measurements are
perfect.

Correction: Question 1. On can directly state:

 
vm (t) = ẋ(t) + b(t) ẋ = −b + vm
so: .
xm (t) = x(t) + v(t) xm = x + v

Introduction to Kalman filtering Page 43/80

44 2. Kalman filter

The bias is modelled as an integral effect (ḃ = 0) with an unknown initial condition
b0 . Thus we have:
       

 ẋ 0 −1 x 1
 = + vm
ḃ 0 0  b 0
(2.19)

 x
 xm = [ 1 0 ] +v
b

with E[v(t)v T (t + τ )] = δ(τ ).

Question 2. The block diagram of the Kalman model is depicted in Figure 2.10

b0
s v

− x
. x +
vm xm
+ +

Figure 2.10: Synthesis model block-diagram.

Question 3. If the Kalman filter is designed on model 2.19, we will find a null gain
in steady state because the state equation is not perturbed by a noise: once the bias
is estimated, the filter will not be able to detect an eventual drift of this parameter.
To overcome this problem and to take into account the bias is not constant, a noise
w is introduced on ḃ(t):
         

 ẋ 0 −1 x 1 0
 = + vm + w
ḃ 0 0 b
  0 1
(2.20)

 x
 xm = [ 1 0 ] +v
b

with E[w(t)wT (t + τ )] = q 2 δ(τ ). This model is in the form of (2.1) and matrices A,
B, C, M can be identified.  
T −1 p1 p12
The steady state solution reads: Kf = P C V with P = positive
p12 p2
solution of the algebraic Riccati equation:
             
0 −1 p1 p12 p1 p12 0 0 p1 p12 1 0 p1 p12 0 0 0 0
+ − + =
0 0 p12 p2 p12 p2 −1 0 p12 p2 0 0 p12 p2 0 q2 0 0

So: 
 2p12 + p21 = 0
p + p1 p12 = 0 .
 2
p212 = q 2

Introduction to Kalman filtering Page 44/80

2.3 Corrected exercises 45

The positive solution is:

 √   √ 
2q −q 2q
P = √ ⇒ Kf = .
−q q 2q −q

The filter state space equation is (from (2.8)):

" #         
ḃ
x b
x b
x b
x xm
= A b +Bvm +Kf xm − C b = (A−Kf C) b +[Kf B] ,
ḃb b b b vm

" #  √    √  
ḃ
x − 2q −1 b
x 2q 1 xm
or: = bb + −q 0 .
ḃb q 0 vm

Question 4. An unbiased estimate of the velocity can be directly obtained by

ḃ = vm − bb.
removing the estimate of the bias from the velocity measurement vm : x
Question 5. A state space realization of F (s) is:
 " #  √    √  

 ḃ
x − 2q −1 b
x 2q 1 x m
 =
 ḃb q 0 bb + −q 0 vm
       

 b
x 1 0 b
x 0 0 xm

 ḃ = +
x 0 −1 bb 0 1 vm

The associated transfer matrix F (s) is:

       √ −1  √ 
0 0 1 0 s 0 − 2q −1 2q 1
F (s) = + − .
0 1 0 −1 0 s q 0 −q 0

That is:  √ 
" # 2qs + q s
√  
b
X(s) qs s2 + 2qs Xm (s)
ḃ = √ . (2.21)
X(s) s2 + 2qs + q Vm (s)
Question 6.
ḃ √
X s2 + 2qs
(s) = 2 √ .
Vm s + 2qs + q
√
This is a second√ order high pass filter with a cutting pulsation q (rd/s) and a
damping ratio 2/2 (∀q). The steady state (D. C.) gain of this filter is null (the
null frequency component, that is the bias, is filtered). The cutting frequency is as
greater as q is great, that is, as the bias may derive a lot.
If measures are perfect then xm = x and vm = ẋ. So we have:

Introduction to Kalman filtering Page 45/80

46 2. Kalman filter

• Xm (s) = X(s) and Vm (s) = sX(s). Reporting this result in the first row of
(2.21), we get:
√
b ( 2qs + q)X(s) + s2 X(s) b
X(s) = √ ⇒ X(s) = X(s) .
s2 + 2qs + q

• Vm (s) = Ẋ(s) and Xm (s) = 1/s Ẋ(s). Reporting that in second row of (2.21),
we get:
√
ḃ Ẋ(s) + (s2 + 2qs)Ẋ(s) ḃ
X(s) = √ ⇒ X(s) = Ẋ(s) .
s2 + 2qs + q

This filter does not induce any degradation of the quality of the measurement. Such
a filter is called a complementary filter.
2

2.4 Discrete-time Kalman filter

2.4.1 Discrete-time Kalman model
By analogy with the continuous-time case the discrete-time Kalman model is:

x(k + 1) = Ad x(k) + Bd u(k) + Md wd (k) state equation, x ∈ Rn , u ∈ Rm , wd ∈ Rq
y(k) = Cd x(k) + Dd u(k) + vd (k) measurement equation, y ∈ Rp , vd ∈ Rp
(2.22)
Assumptions : we will assume that:
H1: The pair (Ad , Cd ) is detectable,
H2: Signals wd (k) and vd (k) are centered gaussian pseudo-white noises with
covariance matrices Wd and Vd respectively, that is:
– E[wd (k) wd (k + l)T ] = Wd δ(l),
– E[vd (k) vd (k + l)T ] = Vd δ(l)
– E[wd (k) vd (k + l)T ] = 0 (with δ(l) = 1 if l = 0; 0 else).
H3: Vd is invertible.

Remark 2.1 : While in the continuous-time case, white noises in the Kalman
model are defined by their PSD W and V (variances are infinite), discrete-time
Kalman model noises are defined by their covariance matrices Wd and Vd (variances
are finite). PSD of these discrete-time noises are constant (and equal to Wd and Vd )
but on limited range of the reduced pulsation θ ∈ [−π π] (see appendix B.1); this
is why this noise is said to be pseudo-white.

Introduction to Kalman filtering Page 46/80

2.4 Discrete-time Kalman filter 47

w(t)
v(t)
dt dt
u(k) u(t) y(t) y(k)
Bo(s) PLANT

Figure 2.11: Continuous-time plant with zero-order holds on the inputs and sampled
outputs.

u(k) u(t)

u(k) dt Bloqueur u(t)

d’ordre 0
t
2 k
O 1 3 4 5 6 0 dt

Figure 2.12: The zero-order hold: Bo (s).

Remark 2.2 : like in the continuous-time case, if the noise wd is a colored noise
and characterized by a spectrum in the z-plane Φww (z) the factorization Φww (z) =
G(z −1 )GT (z) (see section B.3.2) allows a Markov representation of wd to be de-
rived and to be taken into account in an augmented Kalman model.

2.4.2 A particular case: continuous-time plant with discrete-

time measurements
As it was previously mentioned, practical implementation of Kalman filter, even
for continuous-time plant, is performed on a numerical computer and in discrete-
time. So we will consider that the measurement of the continuous-time plant (1.1)
is sampled with a sampling period dt. We will also assume that zero-order holds
are placed on deterministic inputs u (see Figures 2.11 and 2.12) and we are going
to determine a discrete-time representation of this model.
We denote x(k dt) = x(k). From the general solution (1.3), the state equation
integration between t0 = k dt and t = (k + 1) dt reads:
Z ! Z
(k+1)dt (k+1)dt
x(k+1) = eAdt x(k)+ eA((k+1)dt−τ ) Bdτ u(k)+ eA((k+1)dt−τ ) M w(τ )dτ .
k dt k dt

The change of variable: (k + 1)dt − τ = v leads to the following results:

Z dt  Z dt
Adt Av
x(k + 1) = e x(k) + e Bdv u(k) + eAv M w((k + 1)dt − v)dv .
0 0

The measurement equation is a static equation and then, at each sampling time, we

Introduction to Kalman filtering Page 47/80

48 2. Kalman filter

Φvv(ω)

−π/dt 0 π/dt ω

Figure 2.13: Frequency-domain limitation of the continuous-time measurement

noise.

can write:
y(k) = Cx(k) + Du(k) + v(k)
The discrete time state equation is expressed under the form of (2.22) with:
R dt
Ad = eAdt , Bd = 0
eAv Bdv, Md = In , Cd = C, Dd = D (2.23)

The discrete-time state and measurement noises read:

Z dt
wd (k) = eAv M w((k + 1)dt − v)dv, vd (k) = v(kdt) ,
0

It is required to characterized them by their respective covariance matrices Wd and

Vd .
Due to sampling, the covariance matrix of discrete-time measurement noise
becomes:
Vd = V /dt . (2.24)
Justification: The continuous-time measurement equation y(t) = Cx(t) + Du(t) +
v(t) involves a white noise v(t) with an infinite variance and a finite PSD V . The
sampling of the measurement with a sampling step dt provides a numerical series
y(k) with an infinite variance and makes the Kalman filter design to be singular.
The frequency domain response (or the PSD) of noise v(t) must be limited between
−π/dt and π/dt to sample correctly the measurement y(t). One can recognize an
analogy with the Shannon condition for deterministic signals; that is: the sampling
cannot allow to take into account the signal components with a pulsation higher
than π/dt. The PSD of the continuous-time noise, limited in frequency domain, is
depicted in Figure 2.4.2. The variance of such a random signal is now finite and is

Introduction to Kalman filtering Page 48/80

2.4 Discrete-time Kalman filter 49

equal to (see remark 1.3):

Z +∞ Z π
1 V dt V
σv2 = Φvv (ω)dω = dω = .
2π −∞ 2π π
− dt dt

As the sampling of a signal does not change its variance, the discrete-time measure-
V
ment equation is: y(k) = Cx(k) + Du(k) + vd (k) with E[vd (k) vd (k + l)T ] = dt δ(l).

Now we have to compute the covariance matrix Wd of the state noise wd (k):
Z dt Z dt 
T Av T T AT τ
Wd = E[wd (k)wd (k)] = E e M w((k + 1)dt − v)dv w ((k + 1)dt − τ )M e dτ
0 0
Z Z dt
T
= eAv M E[w((k + 1)dt − v)wT ((k + 1)dt − τ )]M T eA τ dvdτ
0
Z Z dt
T
= eAv M W δ(τ − v)M T eA τ dvdτ
0

R dt T
Wd = 0
eAv M W M T eA v dv . (2.25)

Remark 2.3 : We mention in appendix B (see remark B.2, equation (B.13)) how
Wd can be computed; but we can also use the approximation Wd ≈ dtM W M T if dt is
small with respect to the settling time of the plant. Lastly, we have to keep in mind
that all these formulaes are embedded in Matlab macro-functions lqed or kalmd.
These functions allow a discrete-time Kalman filter to be designed directly from
the continuous-time data (A, B, C, D, M , W and V ) and a sampling period dt (see
the Matlab illustration in appendix D.3). Such an approach (and such functions)
cannot bear correlation between measurement and state noises.

2.4.3 Recurrent equations of discrete Kalman filter

The discrete Kalman filter principle is the same than in the continuous-time case.
It involves a prediction based on the deterministic model and a correction (updating)
based on the innovation (difference between the measurement and its prediction) but
in the discrete-time case, we will distinguish:

• the predicted state at time k + 1, denoted x b(k + 1|k), knowing all the mea-
sures until time k; which is associated with the prediction error covariance
matrix denoted:
h  T i
P (k + 1|k) = E x(k + 1) − x b(k + 1|k) x(k + 1) − x b(k + 1|k) .

Introduction to Kalman filtering Page 49/80

 50 2. Kalman filter

• the estimated (or updated) state, denoted x b(k + 1|k + 1), knowing all the
measures until time k + 1 (just after the correction or the updating); which is
associated with the estimation error covariance matrix, denoted:
h  T i
P (k + 1|k + 1) = E b(k + 1|k + 1) x(k + 1) − x
x(k + 1) − x b(k + 1|k + 1) .

b(k|k) and we predict the state at time k + 1

Prediction: at time k, we know x
using the deterministic model

b(k + 1|k) = Ad x
x b(k|k) + Bd u(k) . (2.26)

At time k, the estimation error was characterized by P (k|k). The prediction model
being wrong, the error can only increase and this error at time k + 1 will be char-
acterized by (see theorem B.1 in appendix) :

P (k + 1|k) = Ad P (k|k)ATd + Md Wd MdT (2.27)

Correction: at time k + 1, the prediction is updated with the innovation

through the filter gain Kf (k + 1):
 
b(k + 1|k + 1) = x
x b(k + 1|k) + Kf (k + 1) y(k + 1) − Cd x
b(k + 1|k) − Dd u(k + 1) .
(2.28)
Using the measurement equation of model (2.22), we can write:
 
x(k+1)−bx(k+1|k+1) = In −Kf (k+1)Cd (x(k+1)−b x(k+1|k))−Kf (k+1)vd (k+1) and :

   T
P (k + 1|k + 1) = In − Kf (k + 1)Cd P (k + 1|k) In − Kf (k + 1)Cd + Kf (k + 1)Vd KfT (k + 1)
= P (k + 1|k) − Kf (k + 1)Cd P (k + 1|k) − P (k + 1|k)CdT KfT (k + 1) · · ·
 
· · · + Kf (k + 1) Cd P (k + 1|k)CdT + Vd KfT (k + 1) . (2.29)

Like in the continuous-time case, we are looking for Kf (k + 1) which minimises

trace(P (k + 1|k + 1)) :

∂trace(P (k + 1|k + 1))  

= −2P (k + 1|k)CdT + 2Kf (k + 1) Cd P (k + 1|k)CdT + Vd .
∂Kf (k + 1)

One can deduce:

 −1
Kf (k + 1) = P (k + 1|k)CdT Cd P (k + 1|k)CdT + Vd . (2.30)

Introduction to Kalman filtering Page 50/80

 2.4 Discrete-time Kalman filter 51

Returning this expression in 2.29, we get:

 
P (k + 1|k + 1) = In − Kf (k + 1)Cd P (k + 1|k) . (2.31)

The set of equations (2.26), (2.27), (2.28), (2.30) and (2.31) constitutes the discrete-
time Kalman filter. Equations (2.26) and (2.28) are initialized with x b(0|0), the
initial estimate and equations (2.27), (2.30) and (2.31) are initialized with P (0|0),
the confidence we have in the initialization.
If the model and noises are stationary, equations (2.27), (2.30) and (2.31) can
be integrated off-line. Removing (2.30) and (2.31) in (2.27), one can find a recurrent
Riccati equation in the prediction error covariance:
 −1
T T T
P (k+1|k) = Ad P (k|k−1)Ad −Ad P (k|k−1)Cd Cd P (k|k−1)Cd +Vd Cd P (k|k−1)ATd +Md Wd MdT .
(2.32)
Lastly, in steady state, Kf (k + 1) = Kf (k) = Kf , but on can distinguish:

• Pp = P (k + 1|k) = P (k|k − 1) = · · ·: the prediction error covariance matrix in

steady state which is the positive solution of the algebraic discrete Riccati
equation:

Pp = Ad Pp ATd − Ad Pp CdT (Cd Pp CdT + Vd )−1 Cd Pp ATd + Md Wd MdT . (2.33)

• Pe = P (k + 1|k + 1) = P (k|k) = · · ·: the estimation error covariance matrix

in steady state:
Pe = (I − Kf Cd )Pp .

Then, the state space realization of the stationary (steady state) Kalman filter is
(with (2.28) in (2.26)):
  

 y(k)
b(k + 1|k) = Ad (I − Kf Cd )b
 x x(k|k − 1) +[Ad Kf Bd − Ad Kf Dd ]
  u(k)

 y(k)
 b(k|k) = (I − Kf Cd )b
x x(k|k − 1) +[Kf − Kf Dd ]
u(k)
(2.34)
The state of this realization is the predicted state, the output is the estimated state.

Remark 2.4 :

• 0 < Pe < Pp . Indeed, from (2.30) and (2.31), we get:

Pe = Pp − Pp CdT (Cd Pp CdT + Vd )−1 Cd Pp .

Introduction to Kalman filtering Page 51/80

52 2. Kalman filter

The second term of the right-hand member is always positive, thus: Pe < Pp ,
that is the estimation error variance is always lower than the prediction error
variance (or the state noise variance propagated in the state equation).

• Lastly the (un-noised) output yp (k) = Cd x(k) + Dd u(k) can be estimated by

ybp (k) = Cd x
b(k|k) + Dd u(k). The estimation error covariance for this output
reads:

Cd Pe CdT = Cd Pp CdT −Cd Pp CdT (Cd Pp CdT +Vd )−1 Cd Pp CdT = Cd Pp CdT (Cd Pp CdT +Vd )−1 Vd .

That is:
 
Cd Pe CdT −Vd = Cd Pp CdT (Cd Pp CdT +Vd )−1 −I Vd = −Vd (Cd Pp CdT +Vd )−1 Vd < 0 .

So: Cd Pe CdT < Vd i.e. ybp (k) is a better estimate of yp that the direct measure-
ment.

• The way to solve Discrete-time Algebraic Riccati Equations (DARE) are not
detailed here. Macro-functions (lqe in Scilab or dlqe, dare, kalman in Mat-
lab) allow such equations to be solved, that is: to compute Kf and to provide
the state space realization (2.34) of the filter in steady state. The way to use
these function is illustrate in the Matlab session presented in appendix D.3.

Exercise 2.1 Demonstrate that in the case of a continuous-time plant with discrete-
time sampled output (with a sampling period dt),

• the continuous-time Kalman filter designed from continuous-time data (A,

B, M , C, W , V ) and then discretized by theEuler method

• and discrete-time Kalman filter designed from equations (2.23), (2.24) and
(2.25)

tend towards the same solution as dt tends towards 0 (first order calculus in dt).
Solution : The continuous-time Kalman filter is defined by equations (2.8), (2.14)
and (2.15). The Euler method consists in removing x(t) ˙ and x(t) by x(k+1)−x(k)
dt
and x(k), respectively. Removing (2.14) in (2.8), this first discrete-time filter is
defined by:
  
 P (k + 1) = P (k) + dt AP (k) + P (k)AT − P (k)C T V −1 C(k)P + M W M T
  
 x b(k + 1) = x b(k) + dt Abx(k) + Bu(k) + P (k)C T V −1 y(k) − Cb x(k) − Du(k)
(2.35)

Introduction to Kalman filtering Page 52/80

2.4 Discrete-time Kalman filter 53

The prediction error covariance of the discrete-time Kalman filter is given by equa-
tion (2.32), let us denote P (k|k − 1) = P (k), then using equations (2.23), (2.24)
and (2.25) and with the first order approximation Wd ≈ dtM W M T , we get:

T dt
 −1 T
P (k+1) ≈ eAdt P (k)eA −eAdt P (k)C T CP (k)C T +V /dt CP (k)eA dt +dtM W M T .

Or with a first order expansion:

P (k + 1) ≈ (I + Adt)P (k)(I + AT dt) + dtM W M T

 
−dt(I + Adt)P (k)C T I − dtV −1 CP (k)C T V −1 CP (k)(I + AT dt)
 
≈ P (k) + dt AP (k) + P (k)AT − P (k)C T V −1 CP (k) + M W M T + dt2 (· · ·) .

So one can recognize (at the first order) the first equation of (2.35). The gain Kf
becomes:
 −1
Kf (k) = P (k)C T CP (k)C T V −1 dt + I V −1 dt ≈ P (k)C T V −1 dt,

and the state equation of the filter (equation (2.34)) becomes (we denote x b(k|k −1) =
b(k) and we use also the first order approximation: Bd ≈ dtB):
x
 
b(k + 1) = eAdt I − dtP (k)C T V −1 C x
x b(k) + dteAdt P (k)C T V −1 y(k)
 
+ Bd − dt eAdt P (k)C T V −1 D u(k)
 
T −1
≈ (I − Adt) I − dtP (k)C V C x b(k) + dt(I − Adt)P (k)C T V −1 y(k)
 
+ Bd − dt(I − Adt)P (k)C T V −1 D u(k)
  
T −1
≈ x b(k) + dt Ab x(k) + Bu(k) + P (k)C V y(k) − Cbx(k) − Du(k) + dt2 (· · ·) .

In the first order approximation, we recognize the second equation (2.35). So both
discrete-time filters are equivalent when dt tends toward 0.

2.4.4 Example
Let us consider again exercise 2.3.2 on the bias estimation. We wish to implement
the filter on a numerical computer, both measurements vm and xm being sampled
with the sampling period dt.

• Provide state-space equations of the discrete-time Kalman model with [x(k), b(k)]T
as state vector.

Introduction to Kalman filtering Page 53/80

54 2. Kalman filter

• Compute using Matlab the gain Kf and the estimation error covariance in
steady state.
Numerical application: dt = 0.01 s, V = 1, q = 1.

Correction: The computation of discrete-time model consist in applying formulaes

(2.23), (2.24) and (2.25) to the continuous-time model defined by equation (2.20).
We assume the velocity measurement vm (k) constant on a sampling period (that
corresponds to introduce a zero-order hold). So we can write:
 
0 −dt        
0 −dt 1 −dt
 
1 0 0 0
Ad = e 0 0 = + + + ··· = ,
0 1 0 0 0 0 0 1
Z dt     
1 −v 1 dt
Bd = dv = ,
0 0 1 0 0
1
Cd = [1 0] , Dd = 0 , Vd = ,
dt
Z dt     Z dt  2 2 
1 −v 0 0 1 0 q v −q 2 v
Wd = dv = dv
0 0 1 0 q2 −v 1 0 −q 2 v q2
 1 2 3   
q dt − 21 q 2 dt2 0 0
⇒ Wd = 3 ≈ .
− 21 q 2 dt2 q 2 dt 0 q 2 dt
The discrete-time Kalman model is thus defined by:
       

 x(k + 1) 1 −dt x(k) dt
 = + vm (k) + wd (k)
b(k + 1) 0 1  b(k) 0
(2.36)

 x(k)
 xm (k) = [ 1 0 ] + vd (k)
b(k)

The reader will find in appendix D.3 the Matlab file demoKalmand allowing the
various model parameters to be computed and also the gain Kf and the steady state
estimation error covariance (from function dlqe). We also show how the function
lqed can be used directly from the data of the continuous-time model.

2.5 Exercises
2.5.1 Second order system:
A mobile with a mass m is moving along Ox axis due to a force (command) u(t).
A perturbation force w(t) acts also on this mass. w(t) is modelled as a gaussian

Introduction to Kalman filtering Page 54/80

2.5 Exercises 55

centered white noise with a PSD W . The position x(t) of this mass is measured.
We denote xm (t) this measurement which is perturbed by a gaussian centered white
noise v(t) with a PSD V .
A.N.: m = 1 (Kg); W = 1 (N 2 /Hz); V = ρ2 (m2 /Hz) .
From the measurement xm and the command u, we wish to design a Kalman
˙ of the mobile to be estimated.
filter allowing the position x(t) and the velocity x(t)
These estimates are denoted x ḃ
b(t) and x(t).

1) Give the state space equations of the Kalman model.

2) Compute the Kalman gain in steady state (as a function of ρ).

3) Compute the transfer matrix F (s) of the filter:

" #  
b
X(s) Xm (s)
ḃ = F (s)
X(s) U (s)

4) Comment this transfer (frequency-domain responses of the various compo-

nents, influence of ρ).

5) This filter will be implemented in a numerical computer with a sampling period

dt. Give the state space equations of the discrete-time model and the recurrent
equations of the discrete-time Kalman filter. We denote x ḃ0 the initial
b0 and x
position and velocity, and P0 the initial estimation error covariance.

Introduction to Kalman filtering Page 55/80

EMPTY PAGE
 57

Chapter 3

About physical units

It is important to provide physical units for the characteristics of random signals we

are going to use or simulate. The aim of this brief chapter is to clarify this point.
First of all, we recall that: if u stands for the physical unit of a random variable
X , then its distribution function has no unit and the probability density function
f (x) is in u−1 .
If u stands for the physical unit of a continuous-time random signal w(t) than
physical units of various stochastic characteristics are given in Table 3.1 (equations,
allowing the dimension homogeneity to be checked, are also referenced).

Variable Notation Physical unit Equations

signal w(t) u
mean E[w(t)] u (1.8) and (1.5)
auto-correlation (variance) φww (t, τ ) u2 (1.9) and (1.6)
spectrum in the s-plane (and PSD) Φww (s) u2 s or (u2 /Hz) (1.10)

Table 3.1: Physical units for continuous-time random signal characteristics.

In the case of a discrete-time random signal, the reader will consult Table 3.2
(see also appendix B.1).

Variable Notation Unit physique

signal w(n) u
mean E[w(n)] u
auto-correlation (variance) φww (n, k) u2
spectrum in the z-plane (and PSD) Φww (z) u2

Table 3.2: Physical units for discrete-time random signal characteristics.

Introduction to Kalman filtering Page 57/80

58 3. About physical units

One can remark a factor time (s) between the continuous-time PSD (or spec-
trum in the s-plane) and the discrete-time PSD (or spectrum in the z-plane). Equa-
tion (2.24), used for the discretization of the measurement equation of a continuous-
time system is homogeneous with that. To characterize a continuous-time noise on a
signal expressed√in u, some authors specify the square root of the PSD which is thus
expressed in u/ Hz (and which is wrongly considered as a standard deviation).
If we consider now the continuous-time Kalman model (2.1) and if u stands
for the physical unit of the state variable x(t), then the physical unit of the state
noise wx (t) = M w(t) is u/s and its DSP Wx = M W M T is expressed in u2 /s.
In the case of discrete-time Kalman filter (2.22), if u stands for the physical
unit of x(k) (and also of x(k + 1)) then Md wd (k) is also expressed in u and the PSD
of the state noise Md Wd MdT is expressed in u2 . Equation (2.25) used to discretize
a continuous-time state noise or the approximation Wd = W dt respect also this
homogeneity.

Introduction to Kalman filtering Page 58/80

 59

References

[1] D. Alazard, C. Cumer, P. Apkarian, M. Gauvrit, and G. Ferreres.

Robustesse et Commande Optimale.
CEPADUES dition, 1999.
[2] B.D.O. Anderson and J.B. Moore.
Linear optimal control.
Prentice-Hall, 1971.
[3] E.R. Dougherty.
Random processes for image and signal processing.
SPIE Optical Engineering Press, 1999.
[4] H. Kwakernaak and R. Sivan.
Linear optimal control system.
John Wiley and Sons, 1972.
[5] M. Labarrre, J. P. Krief, and B. Gimonet.
Le filtrage et ses applications.
Cpadus Editions, 1982.
[6] J.L. Pac.
Les processus stochastiques.
polycopi SUPAERO.
[7] Wikipedia:.
http://en.wikipedia.org/wiki/Matrix exponential.
[8] Wikipedia:.
http://en.wikipedia.org/wiki/Riccati differential equation.

Introduction to Kalman filtering Page 59/80

EMPTY PAGE
 61

Appendix A

State space equation integration

A.1 Continuous-time case

Let us consider a continuous state space model:

ẋ(t) = Ax(t) + Bu(t)
(A.1)
y(t) = Cx(t) + Du(t)
The response of this model to deterministic input on the time range t ∈ [t0 , t] and
to initial conditions x(t0 ) is:
Z t
A(t−t0 )
x(t) = e x(t0 ) + eA(t−τ ) Bu(τ ) dτ (A.2)
t0
y(t) = Cx(t) + Du(t) (A.3)

Proof: the general solution of the state equation ”without second member”
ẋ(t) − Ax(t) = 0 is:
x(t) = eAt K .
A particular solution can be derived using the ”varying constant method” (K →
K(t)) :
AeAt K(t) + eAt K̇(t) = AeAt K(t) + Bu(t) (A.4)
K̇(t) = e−At Bu(t) (A.5)
Z t
K(t) = e−Aτ Bu(τ ) dτ + K0 (A.6)
t0
Z t
At
⇒ x(t) = e K0 + eA(t−τ ) Bu(τ ) dτ . (A.7)
t0

Taking into account the initial condition at t = t0 yields to K0 = e−At0 x(t0 ) and
allows to find a unique solution (A.2). Lastly, the output equation is static : y(t) =
Cx(t) + Du(t).

Introduction to Kalman filtering Page 61/80

62 A. State space equation integration

2
Such a solution involves the exponential matrix eA(t−t0 ) ([7]) which can be computed
using a eigen-value/eigen-vector decomposition of matrix A if A is diagonalizable
(see example in section C.2 or using a power series expansion (see example below).

Remark A.1 Using (A.2) with t0 =0, x0 = 0 and u(t) = δ(t) ( Dirac function),
the impulse response of the system defined by (A, B, C, D) is:
f (t) = CeAt B + Dδ(t) ∀t ≥ 0 (f (t) = 0 si t < 0) .
R +∞
The Laplace transform: F (s) = 0 f (τ )e−τ s d τ allows to find the well-known
result:
Z +∞
F (s) = (CeAτ B+Dδ(τ ))e−τ s d τ = C(sI−A)−1 B+D (∀s ∈ convergence domain).
0
(A.8)
Y (s) 1
Example A.1 Let us consider the second order model: U (s)
= (s+1)2
.
• a) compute the system impulse response,
• b) compute the response of the system to initial conditions y(0) = y0 and ẏ(0) =
y˙0 (with t0 = 0).
Correction: If the use of Laplace transform pair table allows the question a) to be
solved directly: y(t) = te−t , it is strongly recommended to use a state space approach
and equation (A.2) to solve question b). Let us consider a Jordan realization of
the plant:     
 −1 1 0
ẋ = x+ u
0 −1 1 . (A.9)

y= [ 1 0 ]x
As the dynamic matrix A is not diagonalizable (one multiple eigenvalue of order 2),
the computation of the matrix exponential eAt can be performed using a power series
expansion:
A2 t2 A3 t3 A4 t4
eAt = I + At + + + + ··· .
2! 3! 4!
Then we get:
 
−t t    
t2 t3 3 4
1−t+ − + · · · t − t2 + t2! − t3! + · · · e−t te−t
 

e 0 −t = 2! 3!
2 3 = .
0 1 − t + t2! − t3! + · · · 0 e−t
Impulse response: with t0 = 0, x(t0 ) = 0 and u(t) = δ(t) ( Dirac function), equation
(A.2) becomes :
Z t   −t 
(t − τ )eτ −t te
x(t) = τ −t δ(τ )dτ = −t et y(t) = [1 0]x(t) = te−t .
0 e e

Introduction to Kalman filtering Page 62/80

A.2 Discrete-time case 63

Response to initial conditions (u(t) = 0): a relationship between the state vector x =
[x1 , x2 ]T and the vector composed of the output y and its derivative ẏ, on which the
initial conditions are formulated, must be established. The measurement equation
and its derivation allow to write:
      
y 1 0 1 0 y0
= x ⇒ x(t0 ) = .
ẏ −1 1 1 1 y˙0
So:
   
e−t te−t 1 0 y0
At
y(t) = Ce x(t0 ) = [1 0] = e−t (1 + t)y0 + te−t y˙0 .
0 e−t 1 1 y˙0
Remark: an horizontal companion realization can be also used: then the state vector
is composed of the output and its derivative:
      
ẏ 0 1 y 0
= + u. (A.10)
ÿ −1 −2 ẏ 1
The computation of the matrix exponential is now:
 
0 t  
e−t (1 + t) te−t
 

e −t −2t = .
−te−t e−t (1 − t)
2

A.2 Discrete-time case

Let us consider a discrete-time state space model:

x(k + 1) = Ad x(k) + Bd u(k)
(A.11)
y(k) = Cd x(k) + Dd u(k)
The response to deterministic inputs u(.) over a time range [k0 , k] and to initial
conditions x(k0 ) = x0 is:
X
k−1
x(k) = Ak−k
d
0
x0 + Ai−k
d
0
Bd u(k − 1 − i + k0 ) (A.12)
i=k0
y(k) = Cd x(k) + Du(k) . (A.13)
The proof is directly given solving the state-space recurrent equation:
x(k) = Ad x(k − 1) + Bd u(k − 1)
= A2d x(k − 2) + Ad Bd u(k − 2) + Bd u(k − 1)
= ···

Introduction to Kalman filtering Page 63/80

64 A. State space equation integration

Remark A.2 Using (A.12) with k0 =0, x0 = 0 and u(i) = δ(i) (discrete Dirac
function: δ(0) = 1; δ(i) = 0 if i 6= 0), the impulse response of the system defined by
(Ad , Bd , Cd , Dd ) is therefore:

f (0) = Dd , f (i) = Cd Ai−1

d Bd ∀i ≥ 1 (f (i) = 0 si i < 0) .
P∞
The Z-transform: F (z) = i=0 f (i)z −i allows to find the well-known result:
∞
X
−i −1
F (z) = Cd Ai−1
d Bd z +Dd = Cd (zI−Ad ) Bd +Dd (∀z ∈ domaine de convergence).
i=1
(A.14)

Introduction to Kalman filtering Page 64/80

 65

Appendix B

Transmission of random signals

and noises in linear systems

The following proofs are extracted from reference [5] and adapted to the notation
of this document.

B.1 Additional background: discrete random sig-

nals
The notions introduced in chapter 1 are extended here to the discrete-time case .
Let w(n) be a sequence of random variables.
• Expected value (mean) : m(n) = E[w(n)].
• Autocorrelation function : φww (n, k) = E[w(n)wT (n + k)].
• Wide-sense stationarity: m(n) = m; φww (n, k) = φww (k) ∀n .
• Variance : σw2 = φww (k)|k=0 .
• Spectrum in the z-plane :
∞
X
Φww (z) = φww (k)z −k , and reciprocally :
k=−∞
Z Z π
1 k−1 1
φww (k) = Φww (z)z dz = Φww (ejθ )ejθk dθ .
2πj unit circle 2π −π
• Power Spectral Density (PSD): this function of the pulsation ω assumes the
introduction of a time scale. So, the sampling period dt between each sample
of the sequence w(n) is introduced.

Φww (ω) = Φww (z)|z=ejωdt

Introduction to Kalman filtering Page 65/80

66 B. Transmission of random signals and noises in linear systems

Then, we can state:

Z π/dt Z π/dt
dt jω dt k dt
φww (k) = Φww (ω)e dω and σw2 = Φww (ω)dω .
2π −π/dt 2π −π/dt

The variance is equal (with a factor dt/2/π) to the integral of the PSD between
−π/dt et π/dt.

B.2 Time-domain approach

B.2.1 Continuous-time case
Theorem 1.1 (recall of page 21). Let us consider the continuous-time linear
system:
˙ = Ax(t) + M w(t) .
x(t) (B.1)
w(t) is a centered gaussian white noise with a PSD W . Let us denote m(t0 ) and
P (t0 ) the mean vector and the covariance matrix of the initial state x(t0 ) (also a
gaussian random variable independent of w(t)). It is shown that x(t) is a gaussian
random signal:
• with mean vector:
m(t) = E[x(t)] = eA(t−t0 ) m(t0 )

• and a covariance matrix P (t) = E[(x(t) − m(t))(x(t) − m(t))T ] solution of the

differential Lyapunov equation:
˙ = AP (t) + P (t)AT + M W M T .
P (t) (B.2)

If the system is stable (all the eigenvalues of A have a negative real part) a
steady state is reached: Ṗ = 0 and P (t) = P is then solution of the continuous-
time Lyapunov equation:

AP + P AT + M W M T = 0 . (B.3)

Proof: The integration of equation (B.1) from initial time t0 and running time t
yields to: Z t
A(t−t0 )
x(t) = e x(t0 ) + eA(t−τ ) M w(τ )dτ
t0

x(t) is thus a combination of random gaussian signals (x(t0 ) and w(τ )), so x(t) is
also a gaussian random signal. Let us compute its mean m(t) = E(x(t)] and its
covariance matrix P (t) = E[(x(t) − m(t))(x(t) − m(t))T ].

Introduction to Kalman filtering Page 66/80

B.2 Time-domain approach 67

Mean m(t):

Z t
A(t−t0 )
m(t) = e E[x(t0 )] + eA(t−τ ) M E[w(τ )]dτ ,
t0

now, E[w(τ )] = 0 (centered noise) and E[x(t0 )] = m(t0 ) so:

m(t) = eA(t−t0 ) m(t0 ) .

Covariance P (t):

  Z t
A(t−t0 )
x(t) − m(t) = e x(t0 ) − m(t0 ) + eA(t−τ ) M w(τ )dτ
t
 Z t0 
= eA(t−t0 ) x(t0 ) − m(t0 ) + eA(t0 −τ ) M w(τ )dτ . (B.4)
t0

  T   T
A(t−t0 )
x(t) − m(t) x(t) − m(t) = e x(t0 ) − m(t0 ) x(t0 ) − m(t0 ) +
Z t  T
+ eA(t0 −τ ) M w(τ ) x(t0 ) − m(t0 ) dτ +
t
Z 0t   T
+ x(t0 ) − m(t0 ) wT (τ )M T eA (t0 −τ ) dτ +
t
Z Z0 t 
A(t0 −τ ) T T AT (t0 −u) T
+ e M w(τ )w (u)M e dτ du eA (t−t0 ) . (B.5)
t0

 Z t   T 
A(t−t0 ) A(t0 −τ )
P (t) = e P (t0 ) + e M E w(τ ) x(t0 ) − m(t0 ) dτ +
t0
Z t   
T
+ E x(t0 ) − m(t0 ) w (τ ) M T eA (t0 −τ ) dτ +
T
t
Z Z0 t   
A(t0 −τ ) T T AT (t0 −u) T
+ e M E w(τ )w (u) M e dτ du eA (t−t0 ) . (B.6)
t0

From assumption (x(t0 ) and w(τ ) are independent ∀ τ > t0 and w(t) is a centered
white noise), one can state:
  T 
• E w(τ ) x(t0 ) − m(t0 ) =0
 
T
• E w(τ )w (u) = W δ(τ − u).

Introduction to Kalman filtering Page 67/80

68 B. Transmission of random signals and noises in linear systems

So:
 Z t 
A(t−t0 ) A(t0 −τ ) T AT (t0 −τ ) T
P (t) = e P (t0 ) + e MWM e dτ eA (t−t0 ) . (B.7)
t0

and
   
˙ dP (t) A(t−t0 ) AT (t−t0 ) A(t−t0 ) T
P (t) = = Ae P (t0 ) + · · · e +e P (t0 ) + · · · eA (t−t0 ) AT
dt
 
A(t−t0 ) A(t0 −t) T AT (t0 −t) T
+e e MWM e eA (t−t0 ) . (B.8)

That is:
˙ = AP (t) + P (t)AT + M W M T .
P (t) (B.9)

Remark B.1 In steady state, the general solution of equation

AP + P AT + M W M T = 0 is:
Z ∞
T
P = eAt M W M T eA t dt .
0

Indeed:
Z ∞
T Tt T
AP + P A = AeAt M W M T eA + eAt M W M T eA t AT dt
0
Z ∞ T
d[eAt M W M T eA t ]
= dt
0 dt
T
= [e M W M T eA t ]∞
At
0 = 0 − MW M
T
(iff A is stable) .

Let us denote h(t) = eAt B, ∀t ≥ 0, the impulse response of the state x, one can
write:
Z ∞ Z ∞
T 1
P = h(t)W h (t)dt = H(−jω)W H T (jω)dω ( Parseval equality)
0 2π −∞
= [L−1 Φ
II xx (s)]s=0 with: Φ T
xx (s) = H(−s)W H (s) .

These last equalities allow to do the link with the following frequency-domain ap-
proach (theorem 1.2) and to show that the variance is (with a factor 2π) the integral
of the noise frequency response square.

B.2.2 Discrete-time case

Theorem B.1 Let us consider the discrete-time linear system:

x(k + 1) = Ad x(k) + Md wd (k) . (B.10)

Introduction to Kalman filtering Page 68/80

B.2 Time-domain approach 69

wd (k) is a pseudo-white gaussian centered noise with a PSD Wd (that is

E[wd (k)wd (k + j)T ] = Wd δ(j)) . m(0) and P (0) denote the mean vector and covari-
ance matrix of initial state x(k0 ) = x(0) (also a random gaussian variable indepen-
dent of wd (k)). Then x(k) is also a gaussian random signal:
• with a mean vector:
m(k) = E[x(k)] = Adk−k0 m(0)
• and a covariance matrix P (k) = E[(x(k) − m(k))(x(k) − m(k))T ] solution of
the recurrent Lyapunov equation:
P (k + 1) = Ad P (k)ATd + Md Wd MdT . (B.11)
If the system is stable (that is all the eigenvalues of Ad have a modulus lower
than 1) a steady state is reached: P (k + 1) = P (k) = P is solution of the
discrete-time Lyapunov equation:
P = Ad P ATd + Md Wd MdT . (B.12)

Proof: (this proof is simpler than in the continuous-time case). In equation (A.12),
x(k) is a linear combination of gaussian random variable. Then one can conclude
that x(k) is a gaussian random variable. wd (k) being centered (∀k), one can state
that E[x(k)] = Ak−k
d
0
m(0).
Lastly, one can note that x(k + 1) − m(k + 1) = Ad (x(k) − m(k)) + Md wd (k).
The independence, at each instant k, of centered variables x(k) − m(k) et wd (k)
allow us to conclude that P (k + 1) = Ad P (k)ATd + Md Wd MdT .

Remark B.2 From equation B.7, one can derive the discrete-time Lyapunov equa-
tion for a continuous system sampled with a sampling period dt. It is sufficient to
choose t0 = n dt et t = (n + 1) dt where dt is the sampling period. Then we denote
P (t) = P (n + 1) and P (t0 ) = P (n), we have:
Z dt
A dt AT dt T
P (n + 1) = e P (n)e + eAu M W M T eA u du
0

Let us remark :
• eAdt = Ad : discrete-time dynamic matrix,
R dt T
• 0 eAu M W M T eA u du = Wd is the covariance matrix integrated on a sampling
period,
Then we can write:
Pn+1 = Ad Pn ATd + Wd
(whose steady state is: Pn = Ad Pn ATd + Wd ). One find again equation (B.11) with
Md = I.
One can verify that :

Introduction to Kalman filtering Page 69/80

70 B. Transmission of random signals and noises in linear systems

• if A is invertible, Wd satisfy the Lyapunov equation:

T dt
AWd + Wd AT + M W M T − eAdt M W M T eA = 0, (B.13)

• Wd ≈ dtM W M T if dt est small w.r.t. the plant dynamics.

Example B.1 (Example with Matlab) Let us consider again exercise 1.3 and
complete it by a discrete-time analysis of the variance of the noise w:

% Filter definition:
G=tf([-1 100],[1 50 10000])
% State space realization:
[A,B,C,D]=ssdata(G);
% variance determination using continuous-time Lyapunov equation:
P=lyap(A,B*B’); var_w=C*P*C’ % ==> var_w=1/50.

% Discrete-time Analysis:
dt=0.001; A_d=expm(A*dt);
Wd=lyap(A,B*B’-A_d*B*B’*A_d’); %In this example: W=I; M=B.
Pd=dlyap(A_d,Wd);var_wd=C*Pd*C’
% ==> We find excatly the variance of w(t): var_w=1/50.

% Approximative calculus using: Wd=dt*B*B’

Pdp=dlyap(A_d,dt*B*B’);var_wdp=C*Pdp*C’
% ==> That does not work too bad !!.

B.3 Frequency-domain approach

B.3.1 Continuous-time case
Theorem 1.2 (rappel). Let us consider a stable linear continuous-time system
defined by the transfer matrix G(s)p×q between the input w and the output y. The
steady state response y to a stationary random signal w, characterized by a spectrum
Φww (s)q×q in the s-plane, is a stationary random signal characterized by a spectrum
Φyy (s) in the s-plane such that:

Φyy (s)p×p = G(−s)Φww (s)GT (s) .

Proof : without loss of generality the proof will consider a strictly proper system
(without direct feed-through). Let (A, M , C) be a state space realization of the
transfer G(s), i.e.: G(s) = C(sI − A)−1 M .

Introduction to Kalman filtering Page 70/80

B.3 Frequency-domain approach 71

By definition:
Z ∞ Z ∞
−τ s
 
Φyy (s) = φyy (τ )e dτ = E y(t)y T (t + τ ) e−τ s dτ . (B.14)
−∞ −∞

To compute y(t) during the steady state, the formulae (A.2) will be used with
x(t0 ) = 0 and t0 = −∞:
Z t Z ∞
A(t−u)
y(t) = Ce M w(u)du = CeAv M w(t−v)dv (change of variable: t − u = v) .
−∞ 0

Z +∞ Z ∞ Z ∞ 
T AT u
Φyy (s) = E Av
Ce M w(t − v)dv w (t + τ − u)M e C du e−τ s dτ
T T
−∞ 0 0
Z +∞ Z Z ∞ 
  T AT u T
= Ce M E w(t − v)w (t + τ − u) M e C du dv e−τ s dτ
Av T
−∞ 0
Z +∞ Z Z ∞ 
Av vs −(τ +v−u)s T AT u T −us
= Ce M e φww (τ + v − u)e M e C e du dv dτ
−∞ 0
Z ∞ Z +∞ Z ∞
−(τ +v−u)s T
= Av
Ce M e dvvs
φww (τ + v − u)e dτ M T eA u C T e−us du
0 −∞ 0
T
= G(−s)Φww (s)G (s) from (A.8) and (B.14) .

B.3.2 Discrete-time case

Theorem B.2 (Transmission of a noise in a discrete-time linear system) Let
us consider a stable linear discrete-time system defined by the transfer matrix
G(z)p×q between the input w and the output y. The steady state response y to a
stationary random signal w, characterized by a spectrum Φww (z)q×q in the z-plane,
is a stationary random signal characterized by a spectrum Φyy (z) in the z-plane such
that:
Φyy (z)p×p = G(z −1 )Φww (z)GT (z) .
Proof : without loss of generality the proof will consider a strictly proposer system
(without direct feed-through). Let (A, M , C) be a state space realization of the
transfer G(s), i.e.: G(s) = Cd (zI − Ad )−1 Md .
By definition:
X
+∞ X
+∞
 
Φyy (z) = φyy (i)z −i = E y(k)y T (k + i) z −i (B.15)
i=−∞ i=−∞

To compute y(k) during the steady state, the formulae (A.12) will be used with
x0 = 0 and k0 = −∞:
X
k−1 X
+∞
y(k) = Cd Aj−k
d
0
Md w(k − 1 − j + k0 ) = Cd Ajd Md w(k − 1 − j) (j ← j − k0 ) .
j=−∞ j=0

Introduction to Kalman filtering Page 71/80

72 B. Transmission of random signals and noises in linear systems

" +∞ #
X
+∞ X X
+∞
Tl
Φyy (z) = E Cd Ajd Md w(k − 1 − j) wT (k + i − 1 − l)MdT Ad CdT z −i
i=−∞ j=0 l=0
( +∞ +∞ )
X
+∞ XX   l
= Cd Ajd Md E w(k − 1 − j)wT (k + i − 1 − l) MdT ATd CdT z −i
i=−∞ j=0 l=0

X
+∞ X
+∞ X
+∞
l
= Cd Ajd Md z j+1 φww (i − l + j)z −(i−l+j) MdT ATd CdT z −l−1
i=−∞ j=0 l=0

X
+∞ X
+∞ X
+∞
l−1
= Cd Aj−1
d Md z
j
φww (k)z −k MdT ATd CdT z −l (k = i − l + j)
j=1 k=−∞ l=1
−1 T
= G(z )Φww (z)G (z) from (A.14) and (B.15) .

Introduction to Kalman filtering Page 72/80

 73

Appendix C

Analytical solution of
continuous-time differential
Riccati equation

C.1 General analytical solution

Let us consider the general differential Riccati equation (2.15) presented in section
2.2.1 where P (t) is the unknown n × n symmetric matrix:

Ṗ (t) = AP (t) + P (t)AT − P (t)C T V −1 CP (t) + M W M T , (C.1)

with the initial condition: P (t0 ) = P0 .

The following procedure to compute P (t) is derived from chapter 15 of reference
[2].
Let us consider the augmented differential equation:
    
Ẋ(t) −AT C T V −1 C X(t)
= (C.2)
Ẏ (t) MW MT A Y (t)
| {z }
H

with initial conditions 1 : X(t0 ) = In , Y (t0 ) = P0 .

H is called the Hamiltonian matrix associated with Riccati equation.
Then: P (t) = Y (t)X −1 (t).
Indeed:
Ṗ (t) = Ẏ (t)X −1 (t) − Y (t)X −1 (t)Ẋ(t)X −1 (t)
= M W M T X(t)X −1 (t) + AY (t)X −1 (t) + Y (t)X −1 (t)AT X(t)X −1 (t) − Y (t)X −1 C T V −1 CY (t)X −1
= AP (t) + P (t)AT − P (t)C T V −1 CP (t) + M W M T .
1
In is the n × n identity matrix.

Introduction to Kalman filtering Page 73/80

 C. Analytical solution of continuous-time differential Riccati
74 equation

Then:
      
X(t) I Φ11 (t0 , t) Φ12 (t0 , t) I
= Φ(t0 , t) = (C.3)
Y (t) P0 Φ21 (t0 , t) Φ22 (t0 , t) P0

where Φ(t0 , t) is the state-transmition matrix of (C.2) (see also: http://en.wikipedia.org

/wiki/State-transition matrix), partionned in four n × n submatrices Φij (t0 , t).
Finally the general solution reads:

P (t) = [Φ21 (t0 , t) + Φ22 (t0 , t) P0 ][Φ11 (t0 , t) + Φ12 (t0 , t) P0 ]−1 . (C.4)

This analytical solution of P (t) is still valid in the non-stationnary case, that
is when equation matrix data A, C, M , W and V are functions of time t. In the
stationnary case:
Φ(t0 , t) = eH(t−t0 ) .
This matrix exponiential can be computed very efficiently using linear algebra library
(function expm in Matlab).
In the next section, this general solution is used to find the solution presented
in the exercice of section 2.3.1.

C.2 Exemple
Let us consider the Riccati differential equation (2.17):

Ṗ (t) = 2aP (t) − P (t)2 + W .

The associated hamiltonian matrix reads:

 
−a 1
H= .
W a

The matrix exponential eHt is computed using an eigen-value/eigen-vector decom-

position of matrix H:
 λt 
e 1 0 ··· 0
 0 eλ2 t 0 · · · 
−1 Ht Λt −1   −1
H = M ΛM ⇒ e = M e M = M  .. . . M ,
 . . 0 
0 ··· 0 e λn t

where:

Introduction to Kalman filtering Page 74/80

C.2 Exemple 75

 
λ1 0 · · · 0
 0 λ2 0 · · · 
 
• Λ is the eigenvalue diagonal matrix Λ =  .. ... ,
 . 0 
0 · · · 0 λn
• M is the eigenvector matrix M = [v1 , v2 · · · vn ].
Eigen-values:

det(sI−H) = s2 −a2 −W = (s−λ)(s+λ) with: λ2 = a2 +W (always > 0 since: W > 0) .

Eigen-vectors:
   
−a − λ 1 1
• λ1 = λ then v1 / v1 = 0 ⇒ v1 = ,
W a−λ a+λ
   
−a + λ 1 1
• λ2 = −λ then v1 / v2 = 0 ⇒ v2 = ,
W a+λ a−λ
   λ−a 1

1 1 −1 2λ 2λ
Then: M = and M = λ+a 1 ,
a+λ a−λ 2λ
− 2λ
and:
   λ(t−t )   λ−a 1

H(t−t0 ) 1 1 e 0
0 2λ 2λ
e =
a+λ a−λ 0 e−λ(t−t0 ) λ+a
2λ
− 2λ1
 
1 eλ(t−t0 ) (λ − a) + e−λ(t−t0 ) (λ + a) eλ(t−t0 ) − e−λ(t−t0 )
= ,
2λ eλ(t−t0 ) (λ2 − a2 ) + e−λ(t−t0 ) (a2 − λ2 ) eλ(t−t0 ) (λ + a) + e−λ(t−t0 ) (λ − a)
and finally:
eλ(t−t0 ) (λ2 − a2 ) + e−λ(t−t0 ) (a2 − λ2 ) + eλ(t−t0 ) (λ + a)P0 + e−λ(t−t0 ) (λ − a)P0
P (t) =
eλ(t−t0 ) (λ − a) + e−λ(t−t0 ) (λ + a) + eλ(t−t0 ) P0 − e−λ(t−t0 ) P0
λ2 − a2 + (a + λ)P0 + e−2λ(t−t0 ) [a2 − λ2 + (λ − a)P0 ]
=
λ − a + P0 + e−2λ(t−t0 ) (λ + a − P0 )
λ2 −a2 +(a+λ)P0
Steady state: limt→∞ P (t) = P∞ = λ−a+P0
=a+λ
√
P∞ = a + λ = a + a2 + W (P∞ is independant of initial condition P0 )

P (t) can be expressed as a function of P0 , P∞ and λ (with: λ − a = 2λ − P∞ ):

P∞ (2λ − P∞ ) + P∞ P0 + e−2λ(t−t0 ) (2λ − P∞ )(P0 − P∞ )
P (t) =
2λ − P∞ + P0 + e−2λ(t−t0 ) (P∞ − P0 )
2λ − P∞ + P0 + e−2λ(t−t0 ) (P∞ − P0 ) P∞P−2λ
∞
P (t) = P∞
2λ − P∞ + P0 + e−2λ(t−t0 ) (P∞ − P0 )
−2λe−2λ(t−t0 ) (P∞ − P0 )
P (t) = P∞ +
2λ − P∞ + P0 + e−2λ(t−t0 ) (P∞ − P0 )

Introduction to Kalman filtering Page 75/80

 C. Analytical solution of continuous-time differential Riccati
76 equation

2λ(P0 −P∞ )
P (t) = P∞ + e2λ(t−t0 ) (P0 −P∞ +2λ)+P∞ −P0
.

One can recognize the solution proposed in equation (2.18) with k = 2λ.

Introduction to Kalman filtering Page 76/80

 77

Appendix D

Matlab demo files

D.1 Function Kf t.m

function y=Kf_t(u)
% y=Kf_t(u)
% input:
% * u(1): time (t),
% * u(2:length(u)): innovation (y(t)-yhat(t))
% output:
% * y = Kf(t)*(y(t)-yhat(t))
global a c W P0

% Compute P(t):
Pinf=a+sqrt(a^2+W);
k=2*sqrt(a^2+W);
P=Pinf+k*(P0-Pinf)./(exp(k*u(1))*(P0-Pinf+k)+Pinf-P0);

% Compute Kf(t):
Kf=P*c’;

% Output
y=Kf*u(2:length(u));

D.2 Script file demoKalman.m

% Global Variable declaration
global a c W P0
% Kalman model data:

Introduction to Kalman filtering Page 77/80

78 D. Matlab demo files

a=-1; % state space date

b=1;
c=1;
W=1; % Process noise spectral density
V=1;
P0=1; % Initial estimation error variance
X0=20; % Initial condition for the process output
% Simulation data
dt=0.01; % Integration sampling period
T=0; % Time delay in the validation model

% Start simulation:
sim(’simuKalman’);

% Output plots:
figure plot(output.time,output.signals.values(:,1),’g-’) hold on
plot(output.time,output.signals.values(:,2),’k-’,’LineWidth’,2)
plot(output.time,output.signals.values(:,3),’k-’)
plot(output.time,output.signals.values(:,4),’r-’,’LineWidth’,2)
% Compute state estimation error variance as a function of time:
t=output.time;
Pinf=a+sqrt(a^2+W);
k=2*sqrt(a^2+W);
Pdet=Pinf+k*(P0-Pinf)./(exp(k*t)*(P0-Pinf+k)+Pinf-P0);
% plot estimation+2*sigma:
plot(output.time,output.signals.values(:,4)+2*sqrt(Pdet),’r-’)
% plot estimation-2*sigma:
plot(output.time,output.signals.values(:,4)-2*sqrt(Pdet),’r-’)
legend(’y(t)’,’x(t)’,’u(t)’,’xhat(t)’,...
’xhat(t)+2\sigma’,’xhat(t)-2 \sigma’)

return

% Solution with a 2nd order Pade approximation of a 1 second delay

% taken into account in the Kalman filter:
T=1; sys=ss(a,b,c,0);
[num,den]=pade(T,2);
ret=tf(num,den);
sysret=sys*ret;
[a,b,c,d]=ssdata(sysret);
W=1;

Introduction to Kalman filtering Page 78/80

D.3 Script file demoKalmand.m 79

[Kf,P]=lqe(a,b,c,W,1);
% Output estimation error variance:
var=c*P*c’;

% Start simulation only in steady state behavior:

X0=0;
sim(’simuKalman_p’); % same as simuKalman.mdl but
% the matlab Function Kf_t is
% replaced by a static Gain

% Output plots:
figure plot(output.time,output.signals.values(:,1),’g-’) hold on
plot(output.time,output.signals.values(:,2),’k-’,’LineWidth’,2)
plot(output.time,output.signals.values(:,3),’k-’)
plot(output.time,output.signals.values(:,4),’r-’,’LineWidth’,2)
% plot estimation+2*sigma:
plot(output.time,output.signals.values(:,4)+2*sqrt(var),’r-’)
% plot estimation-2*sigma:
plot(output.time,output.signals.values(:,4)-2*sqrt(var),’r-’)
legend(’y(t)’,’x(t)’,’u(t)’,’xhat(t)’,...
’xhat(t)+2\sigma’,’xhat(t)-2 \sigma’)

D.3 Script file demoKalmand.m

% Continuous-time model data:
A=[0 -1;0 0];
B=[1;0];
M=[0;1];
C=[1 0];
D=0;
sysc=ss(A,B,C,D);
V=1;
q=1; W=q^2;
% Continuous-time Kalman filter:
[Kf,P]=lqe(A,M,C,W,V)

dt=0.01;
% Compute Discrete-time model:
sysd=c2d(sysc,dt,’zoh’);
[A_d,B_d,C_d,D]=ssdata(sysd)

Introduction to Kalman filtering Page 79/80

80 D. Matlab demo files

Vd=V/dt
Wd=[1/3*q^2*dt^3 -1/2*q^2*dt^2;-1/2*q^2*dt^2 q^2*dt] % Wd can not
% be computed by a Lyapunov equation (equation B.13) since A
% is not invertible !! (see how Wd can be computed in LQED)
% Discrete-time Kalman filter using dlqe:
[Kfd1,Pd1]=dlqe(A_d,eye(2),C_d,Wd,Vd)
% Discrete-time Kalman filter using lqed (from continuous-time data):
[Kfd2,Pd2]=lqed(A,M,C,W,V,dt) %==> same solution !

Introduction to Kalman filtering Page 80/80