Topic 07: Controllability and observability

Advanced System Theory

Winter semester 2021/22

Dr.-Ing. Oliver Wallscheid

Automatic Control
Faculty of Electrical Engineering, Computer Science, and Mathematics

Oliver Wallscheid AST Topic 07 1

Table of contents

1 Basic concept explanation

2 Reachability and controllability in discrete time
3 Reachability and controllability in continuous time
4 Observability and constructibility in discrete time
5 Observability and constructibility in continuous time
6 Dual systems

Learning objectives
Students will be able to explain the qualitative system theory concepts of controllability,
reachability, observability and constructibility for LTI systems. Moreover, they can utilize
different mathematical tools to investigate if a state / system has any of these properties.

Oliver Wallscheid AST Topic 07 2

Course outline

No. Topic
1 Introduction (admin, system classes, motivating example)
2 Linear algebra basics
3 Response of linear systems incl. discrete-time representations
4 Laplace and z-transforms (complex frequency domain)
5 Frequency response
6 Stability
7 Controllability and observability
8 State transformation and realizations
9 State feedback and state observers

Oliver Wallscheid AST Topic 07 3

Basic concepts (1)
Definition 7.1: Reachability
A state x1 is called reachable if there exists an input that transfers x(t) from the origin to x1
in some finite time T . If this holds true for arbitrary x1 , the system is called reachable.

Definition 7.2: Controllability

A state x0 is called controllable if there exists an input that transfers x(t) from x0 to the
origin in some finite time T . If this holds true for arbitrary x0 , the system is called controllable.

Fig. 7.1: Graphical representation of reachability (left) and controllability (right)

Oliver Wallscheid AST Topic 07 4
Basic concepts (2)
Definition 7.3: Constructibility
A state x0 is called constructable if it can be uniquely determined from the past and present
{y(t), u(t)} over a finite time T . If this holds for any x0 , then we call the system constructable.

Definition 7.4: Observability

A state x0 is called observable if it can be uniquely determined from the present and future
{y(t), u(t)} over a finite time T . If this holds for any x0 , then we call the system observable.

=output ? System

=input Xo

Fig. 7.2: Can the state be determined based on limited sections of the input and output trajectories?
Oliver Wallscheid AST Topic 07 5
Delimitation of the considered system classes
The previously introduced qualitative system theory concepts will be handled only for LTI
system in this course! Nevertheless, we will consider both systems in continuous time

ẋ(t) = Ax(t) + Bu(t),

y(t) = Cx(t) + Du(t),

and discrete time

x[k + 1] = Ax[k] + Bu[k],
y[k] = Cx[k] + Du[k].

While illustrative and straightforward procedures exist to verify the existence of the addressed
properties for LTI system, corresponding extensions for LTV and (certain) nonlinear systems
exist but are mathematically more complex.

Oliver Wallscheid AST Topic 07 6

Table of contents

1 Basic concept explanation

2 Reachability and controllability in discrete time

3 Reachability and controllability in continuous time

4 Observability and constructibility in discrete time

5 Observability and constructibility in continuous time

6 Dual systems

Oliver Wallscheid AST Topic 07 7

Derivation reachability (1)
Lets start with the discrete-time case, where it is sufficient to consider
we don't need output
x[k + 1] = Ax[k] + Bu[k], k ≥ k0 . equation here! (7.1)
From Sec. 4 we know that the state at time step k is
k−1
X
x[k] = Φ[k, k0 ]x0 + Φ[k, i + 1]Bu[i] (7.2)
i=k0

with the state transition matrix

Φ[k, k0 ] = Ak−k0 , k ≥ k0 . (7.3)
For an arbitrary state x1 = x[k] and k0 = 0 we obtain
k−1
X
k
x1 = A x0 + Ak−(i+1) Bu[i]. (7.4)
i=0

Oliver Wallscheid AST Topic 07 8

Derivation reachability (2)
Rewriting the previous equation yields
x1 = Ak x0 + C k U k (7.5)
with h i h iT
C k = B, AB, . . . , Ak−1 B , U k = uT [k − 1], uT [k − 2], . . . , uT [0] . (7.6)
Here, C ∈ Rn×k·m is the controllability matrix and U stores the input sequence {u[k]}.

Theorem 7.1: Discrete-time response in n steps

Consider the LTI system (7.1) with given x0 ∈ Rn . If there exists an input sequence U k =
 T T
u [k − 1], uT [k − 2], . . . , uT [0] transferring x0 to x1 , this can be achieved with not more
than n steps. For k = n the input sequence can be found by solving

C n U n = x1 − An x0 . (7.7)

The proof can be found applying the Cayley-Hamilton theorem.

Oliver Wallscheid AST Topic 07 9
Reachability criterion
Due to Theo. 7.1, we generally denote

C = C n = B, AB, . . . , An−1 B ∈ Rn×n·m

 
(7.8)
we can say that, if controllability matrix has full rank
the pair AB is completely state reachable
as the controllability matrix.

Theorem 7.2: Reachability in discrete time

The LTI discrete-time system (7.1), i.e., the pair (A, B), is (completely state) reachable, if
and only if,
rank(C) = n. (7.9)
If C is not full rank: Then there might exists some
states which are NOT reachable
I Binary, system-wide all or nothing criterion.
I Only if (7.9) holds true, one can find an input sequence solution from (7.7).
due to calaye hamilton property
I If a solution U n exists, it is generally not unique. n is number of states store in our x vector

I If a state is not reachable in n steps, it will not be reachable in n + 1, n + 2, . . . steps.

Oliver Wallscheid AST Topic 07 10
 Controllability matrix would have been non
Example (1) square if we would have more inputs here

Square matrix

Consider the system * Normally we don't

input trajectories bu
   
0 1 0
x[k + 1] = x[k] + u[k]. we have A=square
=X1 1 1 =Xo 1 have found specific
trajectory u(1) and u
This system is reachable since
 
0 1 Summary
rank(C) = rank([B, AB]) = rank = 2. 1) Write cont. Matrix
1 1 2) Check its Rank
 T  T
Assuming x1 = a b and x0 = a0 b0 we can derive
           
0 1 u[1] a 1 1 a0 u[1] b − 1 − b0
= − ⇔ =
1 1 u[0] b 1 2 b0 u[0] a − a0 − b0
=A(sq) (because its a Two state system)
from (7.7), i.e., the goal state can be reached with two steps.

Oliver Wallscheid AST Topic 07 11

 Check States are in column or in
Example (2) rows. Prof Explanation is
ambigious

Input state
Now, consider another system given by
State space    
1 1 1
x[k + 1] = x[k] + u[k].
0 1 0
Two Linearly Dependent Col
This system is not (completely) reachable since
ows, or
 
1 1
rank(C) = rank([B, AB]) = rank = 1 < n = 2.
0 0

Obviously, only states of the form

 T
x=α 1 0 with α ∈ R
which one is this State?
are reachable since they are a basis of the reachability subspace.

Oliver Wallscheid AST Topic 07 12

Controllability
x1 is the origin and Xo (where we want to reach
should not be at origin
For evaluating controllability we can again utilize (7.7) for x1 = 0 and x0 6= 0:

C n U n = −An x0 . (7.10)

Hence, it becomes clear that again rank(C n ) = n must hold to find a solution of the above
equation system, leading to the following theorem:

Theorem 7.3: Reachability implies controllability

The LTI discrete-time system (7.1), i.e., the pair (A, B), is (completely state) controllable if
it is reachable.

Obviously, the question arises if the opposite also holds, i.e., if controllability implies
reachability?

Oliver Wallscheid AST Topic 07 13

Reachability, controllability and the state transition matrix
If a system is controllable, we require the time reversibility property of the state transition
matrix to prove that reachability also holds true. Recall from Sec. 4
(Φ[k, k0 ])−1 = Φ[k0 , k] = A−(k−k0 ) . (7.11)
Inverse of state transition matrix

Theorem 7.4: Controllability and reachability in discrete time

If A is nonsingular controllability also implies reachability for discrete-time LTI systems.

controllability

reachability

Oliver Wallscheid AST Topic 07 14

Table of contents

1 Basic concept explanation

2 Reachability and controllability in discrete time

3 Reachability and controllability in continuous time

4 Observability and constructibility in discrete time

5 Observability and constructibility in continuous time

6 Dual systems

Oliver Wallscheid AST Topic 07 15

Derivation reachability (1)
In continuous time we start recalling that the solution of

ẋ(t) = Ax(t) + Bu(t), x0 = x(t0 ), (7.12)

is Z t
x(t, t0 , x0 ) = Φ(t, t0 )x0 + Φ(t, τ )Bu(τ ) dτ (7.13)
t0

with the state transition matrix

Φ(t, τ ) = Φ(t − τ, 0) = eA(t−τ ) . (7.14)

For an arbitrary state x1 = x(t = T ) and x0 = x(t0 = 0) we obtain

Z T
x1 − eAT x0 = eA(t−τ ) Bu(τ ) dτ. (7.15)
0

Oliver Wallscheid AST Topic 07 16

Derivation reachability (2)
Using the Cayley-Hamilton theorem, that is
n−1
X
e−Aτ = αi (τ )Ai (7.16)
i=0

where αi (τ ), i = 1, . . . , n − 1, are scalar time functions, we have

n−1
X Z T
AT i
x1 − e x0 = AB αi (τ )u(τ ) dτ
i=1 0
 RT
(7.17)

0 α0 (τ )u(τ )
= B, AB, . . . , An−1 B 
  ..
 dτ.

| {z } RT .
C∈Rn×n·m 0 αn−1 (τ )u(τ )

Note hat the controllability matrix is defined the same in continuous time as in discrete time.
Oliver Wallscheid AST Topic 07 17
Derivation reachability (3)

Investigating (7.17) it should become clear that

I The left-hand side of the equation is fully known, i.e., it is a given vector in Rn .
I To find a solution we require n linear independent equations on the right-hand side as well.
I Consequently, the controllability matrix C must have full rank.

Theorem 7.5: Reachability in continuous time

The LTI continuous-time system (7.12), i.e., the pair (A, B), is (completely state) reachable,
if and only if,
rank(C) = n. (7.18)

Oliver Wallscheid AST Topic 07 18

Controllability

In contrast to discrete time, the transition matrix in the continuous time is always nonsingular
since  −1
(Φ(t, τ ))−1 = eA(t−τ ) = eA(τ −t) = Φ(τ, t)

applies. Hence, one can find:

Theorem 7.6: Controllability and reachability in continuous time

If a LTI continuous-time system is (completely state) reachable it is also (completely state)
controllable and vice versa.

If a system is completely reachable / controllable, the question remains how to calculate a

corresponding control input trajectory in order to transfer the system states between arbitrary
x0 and x1 . One possible solution will be discussed next.

Oliver Wallscheid AST Topic 07 19

 Reachability Gramian Is used to calculate corresponding control input trajectory
Definition 7.5: Reachability Gramian for continuous-time LTI systems
The n × n reachability Gramian of the continuous-time LTI system (7.12) is
is is Singular, its inverse
s not exists
Z T
T
Wr (0, T ) = e(T −τ )A BBT e(T −τ )A dτ. (7.19)
0

Note that Wr is symmetric and positive semi-definite for any time interval T > 0.

Theorem 7.7: System input solution for reachable continuous-time LTI systems
Let the system (7.12) be (completely state) reachable. Then, there exists an input that will
transfer any state x0 to any other state x1 in some finite time T . Such input is given by
T (T −t)
Wr−1 (0, T ) x1 − eAT x0 .
u(t) = BT eA
 
(7.20)
RT RT
It can be shown that the above solution minimizes 0 ||u(τ )||2 dτ = 0 uT (τ )u(τ )dτ , i.e.,
investing the minimum amount of control effort (least squares solution of (7.17)).
Oliver Wallscheid AST Topic 07 20
 NOT important for EXAM, Prof. SKipped the slide
Example (1)

Consider the scalar system

ẋ = −ax + bu
which is obviously completely reachable. The task is to find u(t) transferring x0 = x(0) to the
origin x(T ) = 0 in finite time T . First, we require the reachability Gramian:
Z T Z T
−(T −τ )a −(T −τ )a −2aT 2
Wr (0, T ) = e bbe dτ = e b e2aτ dτ
0 0
1  2aT b2 
= e−2aT b2 1 − e−2aT .
 
e −1 =
2a 2a
Then, we can apply Theo. 7.7:
2a 1
−2a 1
u(t) = be−(T −t)a 2 −2aT
−e−aT x0 = 2aT
eat x0 .
b 1−e b e −1

Oliver Wallscheid AST Topic 07 21

 NOT important for EXAM, Prof. SKipped the slide
Example (2)
We can verify the result by inserting it into the system response solution
Z t
−at
x(t, t0 , x0 ) = e x0 + e−a(t−τ ) bu(τ ) dτ
0

yielding
 t  
−2a
Z
−at aτ 1
x(t, t0 , x0 ) = e x0 + e b 2aT
eaτ x0 dτ
0 b e −1
2at
 
−at e −1
=e 1 − 2aT x0 .
e −1

Thus, for t = T we receive x(T ) = 0. Further observations:

I For t > T the state does not remain at the origin.
I For decreasing T → 0 the control input magnitude increases |u| → ∞.
Oliver Wallscheid AST Topic 07 22
 NOT important for EXAM, Prof. SKipped the slide
Interpretation of the reachability Gramian
Looking at the input solution (7.20), two major observations an be made when investigating
the eigenvalues
Wr (0, T )vi = λi vi (7.21)
of the reachability Gramian Wr .

Theorem 7.8: Completely state reachable system’s Gramian

If a given system is completely state reachable, all eigenvalues of Wr are greater than zero
(λi > 0). Otherwise, the inverse Wr−1 does not exist.

Theorem 7.9: Partially state reachable system’s Gramian

All zero eigenvalues of Wr correspond to non-reachable states.

In fact, the larger λi the better the controllability of the corresponding state xi , i.e., one
requires less input ’power’ to transfer the specific state.
Oliver Wallscheid AST Topic 07 23
Stabilizability

Based on previous discussions we can introduce a slightly relaxed definition compared to

stability:

Definition 7.6: Stabilizability

A state is called stabilizable if it either stable or controllable (if it is unstable). If all states of
a system are stabilizable, the system is completely state stabilizable.

In simple words:

One can introduce a state feedback for all unstable states such that the controlled system has
only stable states.

Oliver Wallscheid AST Topic 07 24

 is a tool to find out the states which are not reachable, as in case of
Popov-Belevitch-Hautus (PBH) test controllability matrix, by checking rank we can only find it the states are
reachable / controllable or not, but if the answer is No, we cannot find
information about NON reachable spaces
The system ẋ(t) = Ax(t) + Bu(t) is

controllable / reachable
if for all eigenvalues λi of A it holds
i = 1, 2, 3...... (we have to evaluate for each state)
rank ([ λi I − A, B ]) = n for those values of i, we get rank defeciencyy (7.22)
say that these are non reachable states
we can

for all eigen values which are UNSTABLE values, These

and eigen movements are Controllable
stabilizable we can also extend the concet of PBH test to conept of stabilizability

if for all eigenvalues λi ≥ 0 of A it holds Those Eigen values which are stable (WIth negative real part) does not interest us
because they will decay over time, so they don't need to be stabilize

rank ([ λi I − A, B ]) = n (7.23)

Note that with the PBH test we can evaluate which specific states are non-(reachable /
controllable / stabilizable) compared to the binary, system-wide controllability matrix test from
Theo. 7.5.
Oliver Wallscheid AST Topic 07 25
Table of contents

1 Basic concept explanation

2 Reachability and controllability in discrete time

3 Reachability and controllability in continuous time

4 Observability and constructibility in discrete time

5 Observability and constructibility in continuous time

6 Dual systems

Oliver Wallscheid AST Topic 07 26

 Observability derivation (1)
Recall that observability refers to the ability to uniquely determine the state from knowledge of
current and future outputs and inputs. This translates to if we can determine x0 with
{u[k], y[k]} samples for k ≥ 0. Hence, a good starting point is to write down LTI
discrete-time output equation
k−1
Input / output X we don't need input
trajectories (Both are y[k] = CAk x0 + CA(k−1)−i Bu[i] + Du[k]. equation here! (7.24)
known quantities)
Initial state response i=0 Forced state response

For k > 0 and y[0] = Cx0 + Du[0] we can rewrite

UnKnown quantity
Transformed output at step K = ỹ[k] = CAk x0 (7.25)

with Initial state response

"k−1 #
X
Known quantity ỹ[k] = y[k] − CA(k−1)−i Bu[i] + Du[k] . (7.26)
i=0

Oliver Wallscheid AST Topic 07 27

 Observability derivation (2)
if x is in null space of CA^k, The system is unobsereavble or
if x=0, the system is unobservable. so we cannot distinguish if
Reflecting upon X (the state we are observing) is in null space of CA^k or, the
k state its self is a null state
ỹ[k] = CA x0 --Eq no 1
one can ask if an arbitrary initial state x0 leads to an (observable) zero-input response.

Definition 7.7: Unobservable states in discrete-time LTI systems

A state x is unobservable if the zero-input response of (7.24) is zero for all k ≥ 0, i.e., if
our state is
e of CA^k CAk x = 0 ∀ k ≥ 0. (7.27)

I Hence, a state is unobservable when it cannot be distinguished to a system response

based on the initial condition x0 = 0.
I As discussed previously, it is sufficient to consider k = 0, . . . , n − 1 time steps for a
system with n states, i.e., if the zero-input response is not observable within this time
interval, it will not be observable for k = n, n + 1, n + 2, . . . time steps.
Oliver Wallscheid AST Topic 07 28
Observability derivation (3)
Hence, for k = 0, . . . , n − 1 time steps we can rewrite (7.25)
Ỹ0,n−1 = O n x0 Same as eq no 1, but evaluated at n steps (7.28)
UnKnown quantity
with
h iT h iT
Ỹ0,n−1 = ỹT [0], ỹT [1], . . . , ỹT [n − 1] , O n = CT , (CA)T , . . . , (CAn−1 )T (7.29)
Evaluated at n steps, i.e. n=0,1,2.. n-1 | {z }
observability matrix, Rp·n×n
for a system with y ∈ Rp outputs. Hence, we can derive:

Theorem 7.10: Null space of the observability matrix

If an arbitrary state x is in the null space of O, i.e, If O has a full rank, then there cannot be any null space as by
defination of vectors which leads us that all states must be fully
obserevable
Ox = 0, (7.30)

it is unobservable.
Oliver Wallscheid AST Topic 07 29
Observability derivation (4)

Extending the investigation to a system wide view angle:

Theorem 7.11: Observability in discrete time

The LTI discrete-time system (7.24), i.e., the pair (A, C), is (completely state) observable, if
and only if,
rank(O) = n. (7.31)

Why is the observability property important?

I Certain states might not be directly measurable (e.g., internal charge of a battery).
I Even if a state could be measured, one might want to reduce the number of required
sensors to a minimum (e.g., to save costs or installation space).
I With the investigation tools from this section, one can find suitable sensor arrangements.

Oliver Wallscheid AST Topic 07 30

Example (1)

Consider the system

 
0 1  
x[k + 1] = x[k], y[k] = 0 1 x[k].
1 1
has linearly
Independent columns
Here, the observability matrix 
  
C 0 1
O= =
CA 1 1
has full rank and, therefore, the system is completely state observable. In view of (7.28) we
require two output measurements to determine x0 :
        
y[0] 0 1 x1 [0] x1 [0] y[1] − y[0] we can see that we don't need
= ⇔ = . input to determine unkown
y[1] 1 1 x2 [0] x2 [0] y[0]

Oliver Wallscheid AST Topic 07 31

 Example (2)

Additionally, the slightly different system

 
1 0  
x[k + 1] = x[k], y[k] = 1 0 x[k]
1 1
has linearly dependent columns
is investigated. In contrast, the observability matrix
   
ows, or C 1 0
O= = ERROR
CA 1 0

has two linearly-dependent rows and, therefore, rank(O) = 1 results. Hence, all states of the
form  T
x=α 0 1 with α ∈ R
are unobservable since they are in the null space of O. Consequently, one can measure
identical output series {y[k]} although the system is initialized with different x0 .
Oliver Wallscheid AST Topic 07 32
Constructibility (1)

Recall that constructibility refers to the ability to uniquely determine the state from knowledge
of current and past outputs and inputs. This translates to if we can determine x[k = K] with
{u[k], y[k]} samples for k = 0, . . . , K. Formally we introduce:

Definition 7.8: Unconstructable states in discrete time

A state x is called unconstructable if for every k ≥ 0, there exists a x̃ ∈ Rn such that

x = Ak x̃, C x̃ = 0. (7.32)

I In simple words, this expresses that the dynamics of the system states are not evident in
the output measurement.
I The definition might be a little unintuitive in the first place, but allows us to transfer
much insights from the observability discussion to constructibility (see next slide).

Oliver Wallscheid AST Topic 07 33

Constructibility (2)
If A is nonsingular we can simply rearrange (7.32) to
CA−k x = 0, (7.33)
i.e., all states x satisfying this equation are unconstructable. Compare this with Def. 7.7
CAk x = 0,
defining unobservable states.
Since the constructibility and observability concepts have opposite temporal perspectives
(backward vs. forward in time), the same mathematical tests used for observability can be
applied to constructibility if A is nonsingular.

Theorem 7.12: Observability and constructibility in discrete time

The LTI discrete-time system (7.24), i.e., the pair (A, C), is (completely state) constructable
if it is observable. In contrast, constructibility implies observability only if A is nonsingular.

Oliver Wallscheid AST Topic 07 34

Table of contents

1 Basic concept explanation

2 Reachability and controllability in discrete time

3 Reachability and controllability in continuous time

4 Observability and constructibility in discrete time

5 Observability and constructibility in continuous time

6 Dual systems

Oliver Wallscheid AST Topic 07 35

 Derivation observability (1)
In continuous time we start recalling that the solution of
ẋ(t) = Ax(t) + Bu(t), y(t) = Cx(t) + Du(t), x0 = x(t0 ), (7.34)
assuming an initial time t0 = 0 is
Z t
y(t) = CΦ(t, 0)x0 + CΦ(t, τ )Bu(τ ) dτ + Du(t)
0
Z t
= CeAt x0 + [CeA(t−τ ) B + Dδ(t − τ )]u(τ ) dτ.
0
Initial state response Forced state response
Again, it will be convenient to rewrite
ỹ(t) = CeAt x0 UnKnown quantity (7.35)
Transformed output at step K =

with Z t 
y ỹ(t) = y(t) − [CeA(t−τ ) B + Dδ(t − τ )]u(τ ) dτ . (7.36)
0
Oliver Wallscheid AST Topic 07 36
Derivation observability (2)

Definition 7.9: Unobservable states in continuous-time LTI systems

A state x is unobservable if the zero-input response of (7.34) is zero for all t ≥ 0, i.e., if

CeAt x = 0 ∀ t ≥ 0. x should not be in the null space of Ce^At(7.37)

Likewise in the reachability analysis, we can state that the left-hand side in
ỹ(t) = CeAt x0 (7.38)
is known and the question is under which conditions one can find a unique solution for x0 . We
can investigate this by applying again the Cayley-Hamilton theorem, that is
n−1
X
At
e = αi (t)Ai (7.39)
i=0

where αi (τ ), i = 1, . . . , n − 1, are scalar time functions.

Oliver Wallscheid AST Topic 07 37
Derivation observability (3)
Inserting (7.39) into (7.38) results in
 
C
n−1
X  CA 
CAi αi (t)x0 = α0 (t)x0 · · · αn−1 (t)x0 

ỹ = . (7.40)
 
..
i=1
 . 
CAn−1
| {z }
O∈Rn·p×n

Note hat the observability matrix is defined the same in continuous time as in discrete time.

Theorem 7.13: Observability in continuous time

The LTI continuous-time system (7.34), i.e., the pair (A, C), is (completely state) observable,
if and only if,
rank(O) = n. (7.41)

Oliver Wallscheid AST Topic 07 38

Constructibility
Analogous to observability property, we formally introduce:

Definition 7.10: Unconstructable states in continuous time

A state x is called unconstructable if for every t ≤ 0 the zero-input response of the system
(7.34) is zero, i.e.,
CeAt x = 0, ∀t ≤ 0. (7.42)

Taking into account that the time is reversible in the continuous-time state transition matrix

−1
(Φ(t, 0))−1 = eAt = e−At = Φ(−t, 0)
we can directly find the following.

Theorem 7.14: Constructibility and observability in continuous time

If a LTI continuous-time system is (completely state) observable it is also (completely state)
constructable and vice versa.

Oliver Wallscheid AST Topic 07 39

Observability Gramian
Definition 7.11: Observability Gramian for continuous-time LTI systems
The n × n observability Gramian of the continuous-time LTI system (7.34) is
Z T
T
Wo (0, T ) = eτ A CT Ceτ A dτ. (7.43)
0

Note that Wo is symmetric and positive semi-definite for any time interval T > 0.

Theorem 7.15: Initial state solution for observable continuous-time LTI systems
Let the system (7.34) be (completely state) observable. Then, the initial state x0 (t0 = 0) can
be determined by evaluating {y(t), u(t)} over some finite time interval T resulting in
Z T 
τ AT
x0 = Wo−1 (0, T ) e T
C ỹ(τ )dτ . (7.44)
0

Note: for small T also ||Wo || is small leading to numerical issues when calculating Wo−1 .
Oliver Wallscheid AST Topic 07 40
Example (1)
Consider the system  
0 1  
ẋ(t) = x(t), y(t) = 1 0 x(t).
0 0
From that we can derive
 
At 1 t
and CeAt = 1 t .
 
e =
0 1
The observability Gramian results in
Z T Z T  Z T 
Aτ T

Aτ

1   1 τ
Wo (0, T ) = Ce Ce dτ = 1 τ dτ = dτ
0 0 τ 0 τ τ2
1 2
 
T 2 T
= 1 2 1 3 .
2T 3T
1 4
Since det(Wo (0, T )) = 12 T 6= 0 for T > 0 the inverse Wo−1 (0, T ) exists and one can find x0
when monitoring {y(t), u(t)}.
Oliver Wallscheid AST Topic 07 41
Example (2)
Lets further assume that the system has an initial state
 T
x0 = 0 1 .

By the system’s definition we have u = 0, resulting in the response

 T
x(t) = t 1 , y(t) = Cx(t) = t = ỹ(t).

Inserting into (7.44) we can find that the initial state is successfully observed using the
Gramian approach:
" 1 −1
− 12 T −2 Z T 1
Z T # " #
3T

−1 τ AT T
x0 = Wo (0, T ) e C ỹ(τ )dτ = 12 τ dτ
0 − 21 T −2 T −3 0 τ
" 1 −1
− 21 T −2 12 T 2
#" # " #
3T 0
= 12 = .
− 21 T −2 T −3 1 3
3 T 1

Oliver Wallscheid AST Topic 07 42

Interpretation of the observability Gramian
Looking at the initial state solution (7.44), two major observations an be made when
investigating the eigenvalues
Wo (0, T )vi = λi vi (7.45)
of the observability Gramian Wo .

Theorem 7.16: Completely state observable system’s Gramian

If a given system is completely state observable, all eigenvalues of Wo are greater than zero
(λi > 0). Otherwise, the inverse Wo−1 does not exist.

Theorem 7.17: Partially state observable system’s Gramian

All zero eigenvalues of Wo correspond to non-observable states.

In fact, the larger λi the better the observability of the corresponding state xi , i.e., the
’signal-to-noise ratio’ improves when applying some form of estimator.
Oliver Wallscheid AST Topic 07 43
Detectability

Based on previous discussions we can introduce a slightly relaxed definition compared to

observability and constructibility:

Definition 7.12: Detectability

A state is called detectable if it either observable or it decays to zero asymptotically (if it is
unobservable). If all states of a system are detectable, the system is completely state detectable.

In simple words:

The non-observable states of a given system are to some extent ’irrelevant’ since they
asymptotically tend towards zero.

Oliver Wallscheid AST Topic 07 44

Popov-Belevitch-Hautus (PBH) test
The system ẋ(t) = Ax(t) + Bu(t), y(t) = Cx(t) + Du(t) is
observable / constructable
if for all eigenvalues λi of A it holds
 
λi I − A
rank =n (7.46)
C

and
detectable
if for all eigenvalues λi of A holds
       
λi I − A λi I − A
rank =n ∨ rank <n ∧ λi < 0 . (7.47)
C C

With the PBH test we can evaluate which specific states are non-(observable / constructable /
detectable) compared to the binary, system-wide observability matrix test from Theo. 7.13.
Oliver Wallscheid AST Topic 07 45
Table of contents

1 Basic concept explanation

2 Reachability and controllability in discrete time

3 Reachability and controllability in continuous time

4 Observability and constructibility in discrete time

5 Observability and constructibility in continuous time

6 Dual systems

Oliver Wallscheid AST Topic 07 46

Dual systems (1)

Controllability and observability are dual aspects of the same problem.

Definition 7.13: Dual system

Consider the system described by

ẋ(t) = Ax(t) + Bu(t), y(t) = Cx(t) + Du(t),

x[k + 1] = Ax[k] + Bu[k], y[k] = Cx[k] + Du[k].

The dual system is denoted by

ẋD (t) = AD xD (t) + BD uD (t), yD (t) = CD xD (t) + DD uD (t),

(7.48)
xD [k + 1] = AD xD [k] + BD uD [k], yD [k] = CD xD [k] + DD uD [k],

where AD = AT , BD = CT , CD = BT and DD = DT .

Oliver Wallscheid AST Topic 07 47

Dual systems (2)

Theorem 7.18: Dual system

A LTI system denoted by {A, B, C, D} is reachable if and only if its dual {AD , BD , CD , DD }
is observable, and vice versa.

The proof is quite simple by evaluating the dual observability and controllability matrices:
BT
 
 BT AT 
OD =   = CT ,
 
..
 .  (7.49)
BT (AT )n−1
C D = CT AT CT · · · (AT )n−1 CT = OT .
 

The duality property is useful to transfer proofs from observability to controllability and vice
versa. It also facilitates, if necessary, the examination of a given system for these properties.
Oliver Wallscheid AST Topic 07 48
Important Matlab /GNU Octave commands

1 [bool, ncon] = isctrb(A, B); % binary controllability test with ncon controllable states
2 bool = isstabilizable(A, B); % binary stabilizability test
3 Co = ctrb(A,B); % returns the controllability matrix
4
5 [bool, nobs] = isobsv(A, C); % binary observability test with nobs observable states
6 bool = isdetectable(A, C); % binary detectability test
7 Ob = obsv(A,C); % returns the observability matrix
8
9 W = gram(sys, mode) % computes the controllability / observability Gramian for T => infinity

Note:
I The commands require to install the Control (System) Toolbox in Matlab or GNU Octave.
I The binary ’is...’ commands are only available in GNU Octave.
I Only applicable to LTI systems (in continuous or discrete time).

Oliver Wallscheid AST Topic 07 49

Summary: what you’ve learned in this section

I Reachability (’controllable-from-the-origin’) and controllability

(’controllable-to-the-origin’) are state and/or system properties, referring to the ability if
and how states can be transferred through the state space.
I Reachability always implies controllability. The opposite also holds for continuous-time
systems, while in the discrete-time case A must be nonsingular.
I Observability (’forwards in time’) and constructibility (’backwards in time’) are state
and/or system properties, referring to the ability if and how states can be uniquely
determined by monitoring the system inputs and outputs.
I Observability always implies constructibility. The opposite also holds for continuous-time
systems, while in the discrete-time case A must be nonsingular.
I Checking the rank of the controllability / observability matrix determines if an entire
system is reachable / observable.
I The PBH test can evaluate if and which specific states are uncontrollable / unobservable.
I The reachability / observability Gramian offers insights how well a state is controllable /
observable and one possible solution to the reachability / observability problem.
Oliver Wallscheid AST Topic 07 50
End of this section

Thanks for your attention!

Oliver Wallscheid AST Topic 07 51