ku leuven - the behavioral approach to ...sistawww/smc/jwillems/...jan c. willems (k.u. leuven) –...
TRANSCRIPT
European Embedded Control Institute
HYCON-EECI Graduate School on Control
Spring 2009 - Module M4
THE BEHAVIORAL APPROACH
TO
MODELING AND CONTROL
LecturersPaolo Rapisarda (Un. of Southampton)
andJan C. Willems (K.U. Leuven)
– p. 1/82
Lecture 1
Monday 02-02-2009 14.00-17.30
Models and Behaviors
Lecturer: Jan C. Willems
– p. 2/82
Outline
1. Part I:
◮ Mathematical models◮ The universum◮ The behavior◮ A variety of examples
2. Part II:
◮ Dynamical systems◮ Systems described by ODEs◮ Linearity, time-invariance◮ Properties of dynamical systems◮ Other sets of independent variables
– p. 3/82
Mathematical models
A bit of mathematics & a bit of philosophy...
– p. 4/82
Mathematical models
Assume that we have a ‘real’ phenomenon.
(like a circuit with wires, or a moving object, or a chemicalsubstance with its attributes – we do not formalize the term‘phenomenon’ further, but we will give many examples).
The phenomenon produces‘events’, synonymously called‘outcomes’.
– p. 5/82
Mathematical models
Assume that we have a ‘real’ phenomenon. The phenomenonproduces‘events’, synonymously called‘outcomes’.
Phenomenon
event, outcome
– p. 5/82
Mathematical models
Assume that we have a ‘real’ phenomenon. The phenomenonproduces‘events’, synonymously called‘outcomes’.
Phenomenon
event, outcome
We view a deterministic model for a phenomenon as aprescription of which eventscanoccur, and whichcannot.
– p. 5/82
Mathematical models
Assume that we have a ‘real’ phenomenon. The phenomenonproduces‘events’, synonymously called‘outcomes’.
Phenomenon
event, outcome
We view a deterministic model for a phenomenon as aprescription of which eventscanoccur, and whichcannot.
‘Can’ means: ‘in principle possible’. Which event willactually occur in a particular ‘experiment’ depends on theenvironment to which the system is exposed.
– p. 5/82
Aim of this lecture
◮ In the first part of this lecture, we develop this point ofview into a mathematical formalism.
◮ In the second part, we apply this formalism to dynamicalsystems.
– p. 6/82
The universum
– p. 7/82
Mathematization
The events can be described in the language of mathematics,in terms of mathematical concepts, by answering:
To which universum do the (unmodelled) events belong?
– p. 8/82
Mathematization
The events can be described in the language of mathematics,in terms of mathematical concepts, by answering:
To which universum do the (unmodelled) events belong?
◮ Do the events belong to a discrete set?; discrete event phenomena.
◮ Are the events real numbers, or vectors of real numbers?; continuous phenomena.
◮ Are the events functions of time?; dynamical phenomena.
◮ Are the events functions of space, or time & space?; distributed phenomena.
– p. 8/82
Mathematization
The events can be described in the language of mathematics,in terms of mathematical concepts, by answering:
To which universum do the (unmodelled) events belong?
◮ Do the events belong to a discrete set?; discrete event phenomena.
◮ Are the events real numbers, or vectors of real numbers?; continuous phenomena.
◮ Are the events functions of time?; dynamical phenomena.
◮ Are the events functions of space, or time & space?; distributed phenomena.
The set where the events belong to is called theuniversumdenoted byU .
– p. 8/82
Discrete event phenomena
Examples:
◮ Words in a natural language(a word is thought of as an ‘event’)U = A
∗
= all (finite) strings of letters fromthe alphabetA = {a,b,c, . . . ,x,y,z}
– p. 9/82
Discrete event phenomena
Examples:
◮ Words in a natural languageU = A
∗
= all (finite) strings of letters fromthe alphabetA = {a,b,c, . . . ,x,y,z}
Sentences in a natural language
◮ DNA sequences
◮ Fortran codeLATEX code
◮ Error detecting and correcting codesISBN numbers
– p. 9/82
Continuous phenomena
Examples:
◮ The pressure, volume, quantity, and temperature of a gasin a vessel
Gas
(pressure, volume, quantity, temperature)
; U = (0,∞)× (0,∞)× (0,∞)× (0,∞)
– p. 10/82
Continuous phenomena
◮ A spring
F1 F2
L
Event = (forceF1, force F2, length L)
; U = R×R× (0,∞)
– p. 11/82
Continuous phenomena
◮ The gravitational attraction of two bodies
+
force
position1
position2
mass1
mass2
Event = mass1, mass2, position1, position2, force
; U = (0,∞)× (0,∞)×R3×R
3×R3
– p. 12/82
Continuous phenomena
◮ The voltage across and the current through a resistor
V
I
R+
−
Event = (voltage, current) ; U = R2
– p. 13/82
Dynamical phenomena
Examples:
◮ The voltage across and the current through a capacitoror an inductor
V
I
+
−C
I
V+
−l
Event = (voltage, current): time→ R2
– p. 14/82
Dynamical phenomena
Examples:
◮ The voltage across and the current through a capacitoror an inductor
V
I
+
−C
I
V+
−l
Event = (voltage, current): time→ R2 U =
(R
2)R
Recall the standard notation:AB := the set of maps from B to A = { f : B → A}
Dynamical phenomena; this course.– p. 14/82
Dynamical phenomena
◮ The motion of a point mass
+
mass
positionforce
Event = (force, position): time→ R3×R
3 =(R
3)R
– p. 15/82
Dynamical phenomena
◮ Planetary motion
SUN
PLANET
The events are maps fromR to R3
; U = {w : R → R3} =
(R
3)R
– p. 16/82
Dynamical phenomena
◮ The voltage across and the current into an electrical portwith ‘dynamics’
V
I
−
+����
RL
C
C
LR ��
��
The events are maps fromR to R2
; U = {(V, I) : R → R2} =
(R
3)R
– p. 17/82
Dynamical phenomena
◮ The input and the output of a signal processor
SYSTEM OUTPUTINPUT
Events: maps from Z to Rm×R
p
; U = {(u,y) : Z → Rm×R
p} = (Rm×Rp)Z
– p. 18/82
Dynamical phenomena
◮ The input and the output of a signal processor
SYSTEM OUTPUTINPUT
Events: maps from Z to Rm×R
p
; U = {(u,y) : Z → Rm×R
p} = (Rm×Rp)Z
◮ Trajectories of variables associated with mechanicaldevices, electrical circuits, chemical reactions,multi-domain transducers, economic processes, ...phenomena with ‘memory’.
– p. 18/82
Distributed phenomena
◮ Temperature profile of, and heat absorbed by, a rod
�������������������������������������������������������������
���
q(x,t)
T(x,t)x
Events: maps from R×R to [0,∞)×R
; U = {(T,q) : R2 → [0,∞)×R} = ([0,∞)×R)R2
– p. 19/82
Distributed phenomena
◮ EM fields.
In each point of space & at each time, there is an
electric field ~E(t,x,y,z)magnetic field ~B(t,x,y,z)current density ~j(t,x,y,z)charge density ρ(t,x,y,z)
Events: maps from R×R3 to R3×R3×R3×R
; U = {(~E,~B,~j,ρ) : R4 → R
10} =(R
10)R4
– p. 20/82
Distributed phenomena
◮ EM fields.
In each point of space & at each time, there is an
electric field ~E(t,x,y,z)magnetic field ~B(t,x,y,z)current density ~j(t,x,y,z)charge density ρ(t,x,y,z)
Events: maps from R×R3 to R3×R3×R3×R
; U = {(~E,~B,~j,ρ) : R4 → R
10} =(R
10)R4
◮ Images
◮ Phenomena in which related things happen at differentpoints in space
– p. 20/82
A model is a subset: the ‘behavior’
– p. 21/82
The behavior
Given is a phenomenon with event universumU .Without further scrutiny, any event in U is possible.
After studying the situation, the conclusion is reached that theevents are constrained, that some laws are in force.
– p. 22/82
The behavior
Given is a phenomenon with event universumU .Without further scrutiny, any event in U is possible.
After studying the situation, the conclusion is reached that theevents are constrained, that some laws are in force.
Modeling means that certain events are declared to beimpossible, that they cannot occur.
The possibilities that remain constitute what we call the‘behavior’ of the model.
– p. 22/82
The behavior
Given is a phenomenon with event universumU .Without further scrutiny, any event in U is possible.
After studying the situation, the conclusion is reached that theevents are constrained, that some laws are in force.
A mathematical model:⇔ a pair (U ,B), with
U the universum of events
B a subset of U , called the behavior of the model
Notation: Mathematical model (U ,B)
– p. 22/82
The behavior
B
U
possible
allowed
forbidden
– p. 23/82
The behavior
Every “good” scientific theory is prohibition:it forbids certain things to happen...The more a theory forbids, the better it is.
Karl PopperConjectures and Refutations:The Growth of Scientific KnowledgeRouthledge, 1963
Karl Popper(1902-1994)
– p. 24/82
Examples
– p. 25/82
Discrete event phenomena
Examples:
◮ Words in a natural languageU = A
∗
= all (finite) strings of letters fromthe alphabetA = {a,b,c, . . . ,x,y,z}
B = all words recognized by the spelling checker.For example, SPQR/∈ B.
B is basically defined by enumeration, by listing itselements.
– p. 26/82
Discrete event phenomena
Examples:
◮ Words in a natural languageU = A
∗
= all (finite) strings of letters fromthe alphabetA = {a,b,c, . . . ,x,y,z}
B = all words recognized by the spelling checker.For example, SPQR/∈ B.
B is basically defined by enumeration, by listing itselements.
Sentences in a natural language.B = all ‘legal’ sentences.SpecifyingB is a complicated issue.Usually determined using grammars.
– p. 26/82
Discrete event phenomena
Examples:
◮ Words in a natural languageU = A
∗
= all (finite) strings of letters fromthe alphabetA = {a,b,c, . . . ,x,y,z}
B = all words recognized by the spelling checker.For example, SPQR/∈ B.
B is basically defined by enumeration, by listing itselements.
Sentences in a natural language.B = all ‘legal’ sentences.
◮ DNA sequences.B =???
◮ LATEX code.B = all LATEX files that ‘run’.
– p. 26/82
Discrete event phenomena
◮ 32-bit binary strings with a parity check.
U = {0,1}32
B =
{
a1a2 · · ·a31a32 | ak ∈ {0,1} and a32(mod 2)
=31
∑k=1
ak
}
– p. 27/82
Discrete event phenomena
◮ 32-bit binary strings with a parity check.
U = {0,1}32
B =
{
a1a2 · · ·a31a32 | ak ∈ {0,1} and a32(mod 2)
=31
∑k=1
ak
}
B can be expressed in other ways. For example,
B = {a1a2 · · ·a31a32 | ak ∈{0,1} and32
∑k=1
ak(mod 2)
= 0}
B =
a1a2...
a31a32
| ∃
b1b2...
b30b31
s.t.
a1a2...
a31a32
(mod 2)=
1 0 0 ··· 01 1 0 ··· 0... ... ... ...... ... ... ...0 0 ··· 1 10 0 ··· 0 1
b1b2...
b30b31
– p. 27/82
Discrete event phenomena
◮ 32-bit binary strings with a parity check.
U = {0,1}32
B =
{
a1a2 · · ·a31a32 | ak ∈ {0,1} and a32(mod 2)
=31
∑k=1
ak
}
B can be expressed in other ways. For example,
B = {a1a2 · · ·a31a32 | ak ∈{0,1} and32
∑k=1
ak(mod 2)
= 0}
B =
a1a2...
a31a32
| ∃
b1b2...
b30b31
s.t.
a1a2...
a31a32
(mod 2)=
1 0 0 ··· 01 1 0 ··· 0... ... ... ...... ... ... ...0 0 ··· 1 10 0 ··· 0 1
b1b2...
b30b31
input/output representation
kernel representation
image representation– p. 27/82
Continuous phenomena
Examples:
◮ The pressure, volume, quantity, and temperature of a gasin a vessel
Gas
(pressure, volume, quantity, temperature)
U = (0,∞)× (0,∞)× (0,∞)× (0,∞)
Gas law: B = {(P,V,N,T ) ∈ U | PV = NT }
– p. 28/82
Continuous phenomena
◮ A massless spring
F1 F2
L
Event = (forceF1, force F2, length L)
U = R×R× (0,∞)
B ={
F1 = F2 L = ρ(F1)}
– p. 29/82
Continuous phenomena
◮ The gravitational attraction of two bodies
1
+
position q mass M 1 2force F
position q 2
mass M
; U = (0,∞)× (0,∞)×R3×R
3×R3
B =
{
~F =M1M2~1M2→M1||~q1−~q2||2
}
‘inverse square law’Isaac Newton, 1642-1727
– p. 30/82
Continuous phenomena
◮ The voltage across and the current through a resistor
V
I
R+
−
Event = (voltage, current) ; U = R2
‘Ohm’s law’ B = {(V, I) | V = RI }
Georg Ohm, 1789 – 1854
– p. 31/82
Dynamical phenomena
◮ The voltage across and the current through a capacitor
V
I
+
−C
Event = (voltage, current) as a function of time
; U =(R2
)R
B = {(V, I) : R → R2 | C d
dtV = I }
– p. 32/82
Dynamical phenomena
◮ The voltage across and the current through a capacitoror an inductor
V
I
+
−C
I
V+
−l
Event = (voltage, current) as a function of time
; U =(R2
)R
B = {(V, I) : R → R2 | C d
dtV = I }
B = {(V, I) : R → R2 | V = L d
dt I }– p. 32/82
Dynamical phenomena
◮ Planetary motion U =(R3
)R
PLANET
SUN
DC
B
A 1 year
34 months
Kepler’s laws ; B
– p. 33/82
Dynamical phenomena
◮ Planetary motion U =(R3
)R
PLANET
SUN
DC
B
A 1 year
34 months
Kepler’s laws ; B = the orbits R → R3 with:
K.1 periodic, ellipses, with the sun in one of the foci;K.2 the vector from sun to planet sweeps out equal areas
in equal times;K.3 the square of the period
divided by the third powerof the major axis is thesame for all the planets
– p. 33/82
Dynamical phenomena
◮ Newton’s second law
Isaac Newtonby William Blake
+
massM
position~qforce ~F
U =(R
3×R3)R
B =
{
(~F,~q) : R → R3×R
3 | ~F = M d2
dt2~q
}
– p. 34/82
Static versus dynamic
For the pointmass, we could have chosen
event = (force, acceleration)
; U = R3×R
3B = { ~F = M~a }.
– p. 35/82
Static versus dynamic
For the pointmass, we could have chosen
event = (force, acceleration)
; U = R3×R
3B = { ~F = M~a }.
For the capacitor and for the inductor, we could have chosen
event = (voltage, charge); U = R×R B = { Q = CV }
event = (current, flux) ; U = R×R B = { NΦ = LI }with N = number of turns
Michael Faraday1791 – 1867
– p. 35/82
Static versus dynamic
For the pointmass, we could have chosen
event = (force, acceleration)
; U = R3×R
3B = { ~F = M~a }.
For the capacitor and for the inductor, we could have chosen
event = (voltage, charge); U = R×R B = { Q = CV }
event = (current, flux) ; U = R×R B = { NΦ = LI }with N = number of turns
Michael Faraday1791 – 1867
Combining with
~a = d2
dt2~q, I = ddt Q, V = N d
dt Φ,
recovers the dynamical equations.
– p. 35/82
Static versus dynamic
Which of these descriptions is most appropriate depends onthe application.
But if we think in terms of ‘observable’ variables, or, better,in terms of the variables by which a device interacts with itsenvironment, it is most logical to choose
(~F,~q)
and(V, I)
as the phenomenological variables.
– p. 36/82
Dynamical phenomena
◮ The input and the output of a signal processor
System outputs inputs
Events: maps fromZ toRm×R
p; U = {(u,y) : Z → R
m+p}
– p. 37/82
Dynamical phenomena
◮ The input and the output of a signal processor
System outputs inputs
Events: maps fromZ toRm×R
p; U = {(u,y) : Z → R
m+p}
For an MA system
B ={
(u,y) : Z → R2 | y(t) = 12T+1 ∑t+T
t′=t−Tu(t ′)
}
many variations.– p. 37/82
Distributed phenomena
◮ The temperature profile of, and heat absorbed by, a rod
�������������������������������������������������������������
���
q(x,t)
T(x,t)x
Events: maps fromR×R to [0,∞)×R
U = {(T,q) : R2 → [0,∞)×R}
B ={
(T,q) : R2 → [0,∞)×R | ∂
∂ t T = ∂ 2
∂ x2 T +q}
‘the diffusion equation’– p. 38/82
Distributed phenomena
◮ Maxwell’s equations for EM fields in free space
∇ ·~E =1ε0
ρ ,
∇×~E = −∂∂ t
~B,
∇ ·~B = 0 ,
c2∇×~B =1ε0
~j +∂∂ t
~E.
B = maps fromthe independent variables:(t,x,y,z) time and spaceto the dependent variables:(~E,~B,~j,ρ)
electric & magnetic field, current & charge densitythat satisfy Maxwell’s equations.
James Clerk Maxwell1831 – 1879
– p. 39/82
Behavioral models
Behavioral models fit the tradition of modeling, but have notbeen approached or formalized as such in a deterministicsetting.The behavior captures the essence of what a model is.
– p. 40/82
Behavioral models
Behavioral models fit the tradition of modeling, but have notbeen approached or formalized as such in a deterministicsetting.The behavior captures the essence of what a model is.
The behavior is all there is.Equivalence of models, properties of models,
symmetries, system identification, etc.,must all refer to the behavior.
– p. 40/82
Behavioral models
Behavioral models fit the tradition of modeling, but have notbeen approached or formalized as such in a deterministicsetting.The behavior captures the essence of what a model is.
The behavior is all there is.Equivalence of models, properties of models,
symmetries, system identification, etc.,must all refer to the behavior.
Every ‘good’ scientific theory is prohibition: it forbidscertain things to happen...The more a theory forbids, the better it is.
Replace ‘scientific theory’ by ‘mathematical model’ !– p. 40/82
Recapitulation
◮ A phenomenon produces ‘events’, ‘outcomes’.; the event universumU
◮ A mathematical model specifies a subsetB of U .B is the behavior of the model.
◮ A mathematical model is a pair(U ,B).
– p. 41/82
End of Part I
– p. 42/82
Part II
– p. 43/82
Outline
1. Part I:
◮ Mathematical models◮ The universum◮ The behavior◮ A variety of examples
2. Part II:
◮ Dynamical systems
◮ Systems described by ODEs◮ Linearity, time-invariance◮ Properties of dynamical systems◮ Other sets of independent variables
– p. 44/82
Dynamical systems
– p. 45/82
The dynamic behavior
In dynamical systems, the ‘events’ are maps, with thetime-axis as domain, hence events are functions of time.
Phenomenon
eventsignal space
time
– p. 46/82
The dynamic behavior
In dynamical systems, the ‘events’ are maps, with thetime-axis as domain, hence events are functions of time.
Phenomenon
eventsignal space
time
It is convenient to distinguish in the very notation
the domain of the maps, thetime setand the codomain, thesignal space
the set where the functions take on their values.
– p. 46/82
The dynamic behavior
In dynamical systems, the ‘events’ are maps, with thetime-axis as domain, hence events are functions of time.
It is convenient to distinguish in the very notation
the domain of the maps, thetime setand the codomain, thesignal space
the set where the functions take on their values.
The behavior of a dynamical system, hence a family oftime-trajectories, is often described by a system of ordinarydifferential equations (ODEs) or difference equations.
– p. 46/82
The dynamic behavior
Formal definition: A dynamical system:⇔ (T,W,B)
T ⊆ R time setW signal space
B ⊆ WT the behavior
a family of trajectories T → W
w : T → Rw ∈ B ⇔ w is compatible with the modelw : T → Rw /∈ B ⇔ the model forbids w
– p. 47/82
The dynamic behavior
Formal definition: A dynamical system:⇔ (T,W,B)
T ⊆ R time setW signal space
B ⊆ WT the behavior
a family of trajectories T → W
w : T → Rw ∈ B ⇔ w is compatible with the modelw : T → Rw /∈ B ⇔ the model forbids w
mostly, T = R,R+,Z, or N (= {0,1,2, . . .}),and, in this course,W = Rw, for somew ∈ N
B is a family of time trajectoriestaking values in a (finite-dimensional) vector space.
T = R or R+ ; ‘continuous-time’ systems and ODEsT = Z or N ; ‘discrete-time’ systems and difference eqn’sWe mainly deal with the caseT = R.
– p. 47/82
ODEs as dynamical systems
– p. 48/82
ODEs as models for dynamical systems
Consider the ODE
f
(
w,ddt
w,d2
dt2w, . . . ,dn
dtnw
)
= 0
withf : W×R
w×·· ·×Rw
︸ ︷︷ ︸
n times
→ R•
W ⊆ Rw
Notational purists may prefer to write
f ◦
(
w,ddt
w,d2
dt2w, . . . ,dn
dtnw
)
= 0
But we leave this to puritans.
– p. 49/82
ODEs as models for dynamical systems
Consider the ODE
f
(
w,ddt
w,d2
dt2w, . . . ,dn
dtnw
)
= 0
withf : W×R
w×·· ·×Rw
︸ ︷︷ ︸
n times
→ R•
W ⊆ Rw
Defines the dynamical system(R,W,B) with
B = {w : R → W, sufficiently smooth|
f
(
w(t),dwdt
(t),d2wdt2 (t), . . . ,
dnwdtn
(t)
)
= 0 ∀ t ∈ R}
other solution concepts may be appropriate as well ...
– p. 49/82
Examples
◮ Newton’s second law
Isaac Newtonby William Blake
+
massM
position~qforce ~F
~F = M d2
dt2~q w =
[
~F~q
]
W = R3×R
3
f : (~F,ddt
~F,d2
dt2~F ,~q,
ddt
~q,d2
dt2~q) 7→ ~F −Md2
dt2~q
– p. 50/82
Examples
◮ A pointmass in a gravitational field
~q
SUN
PLANET
inverse square law
d2~qdt2 + k
~1~q
||~q||2= 0 (∗)
f : (~q,ddt
~q,d2
dt2~q) 7→d2
dt2~q+ k~1~q
||~q||2
– p. 51/82
Examples
◮ A pointmass in a gravitational field
~q
SUN
PLANET
inverse square law
d2~qdt2 + k
~1~q
||~q||2= 0 (∗)
All orbits satisfying K1, K2, and K3 are solutions.
PLANET
SUN
DC
B
A 1 year
34 months
Is (∗) a ‘representation’ of K1, K2, and K3?Do K1, K2, and K3 give all the solutions of(∗)?
– p. 51/82
Examples
The convolution
y(t) =
∫ t
0e(t−t ′)u(t ′)dt ′ w(t) =
[
u(t)y(t)
]
can be represented by
ddt
y = y+u
f : (u,ddt
u,y,ddt
y) 7→ddt
y− y−u
Note advantages/disdavantages of each of theserepresentations.
– p. 52/82
Linear time-invariant differential systems
LTIDSs
– p. 53/82
LTIDSs
The dynamical system(R,Rw,B) ; B is said to be
[[ linear ]] :⇔ [[ [[w1,w2 ∈ B,α ∈ R]] ⇒ [[αw1 +w2 ∈ B]] ]]
– p. 54/82
LTIDSs
The dynamical system(R,Rw,B) ; B is said to be
[[ linear ]] :⇔ [[ [[w1,w2 ∈ B,α ∈ R]] ⇒ [[αw1 +w2 ∈ B]] ]]
[[ time-invariant ]] :⇔ [[ [[w ∈ B, σ t the t-shift]] ⇒ [[σ t(w) ∈ B]] ]]
σ t( f )(t ′) := f (t ′ + t)
σ
t−shift
f
map
t f
t
– p. 54/82
LTIDSs
The dynamical system(R,Rw,B) ; B is said to be
[[ linear ]] :⇔ [[ [[w1,w2 ∈ B,α ∈ R]] ⇒ [[αw1 +w2 ∈ B]] ]]
[[ time-invariant ]] :⇔ [[ [[w ∈ B, σ t the t-shift]] ⇒ [[σ t(w) ∈ B]] ]]
σ t( f )(t ′) := f (t ′ + t)
σ
t−shift
f
map
t f
t
[[ differential ]] :⇔ [[B can be ‘described’ by an ODE]].
– p. 54/82
Linearity
In many examples,B is the graph of a map
u 7→ y = F(u) w =
[
uy
]
with F a map: u ∈ a space of inputs7→ y ∈ a space of outputs.
Then B = {w =
[
uy
]
| y = F(u)} = the ‘graph’ of F
When F is a linear map, the resulting system is linear.Def. of linear system has the classical one as a special case.
– p. 55/82
Linearity
In many examples,B is the graph of a map
u 7→ y = F(u) w =
[
uy
]
with F a map: u ∈ a space of inputs7→ y ∈ a space of outputs.
Then B = {w =
[
uy
]
| y = F(u)} = the ‘graph’ of F
When F is a linear map, the resulting system is linear.Def. of linear system has the classical one as a special case.
But a dynamical system, even in input/output form(say p( d
dt )y = q( ddt )u) is seldom a map !
y depends onu and on the initial conditions.– p. 55/82
LTIDSs
The dynamical system(R,Rw,B) is
a linear time-invariant differential system (LTIDS) :⇔the behavior consists of the set of solutions of a system oflinear constant coefficient ODEs
R0w+R1ddt
w+ · · ·+Rn
dn
dtnw = 0.
R0,R1, · · · ,Rn ∈ R•×w real matrices that parametrize thesystem, andw : R → Rw.
– p. 56/82
LTIDSs
The dynamical system(R,Rw,B) is
a linear time-invariant differential system (LTIDS) :⇔the behavior consists of the set of solutions of a system oflinear constant coefficient ODEs
R0w+R1ddt
w+ · · ·+Rn
dn
dtnw = 0.
R0,R1, · · · ,Rn ∈ R•×w real matrices that parametrize thesystem, andw : R → Rw. In polynomial matrix notation
; R(
ddt
)w = 0
with R(ξ ) = R0 +R1ξ + · · ·+Rnξ n ∈ R [ξ ]•×w
a polynomial matrix , usually ‘wide’ or square.
– p. 56/82
Examples
Examples:
◮d2
dt2~q = ~F ,w =
[
~F~q
]
; R(ξ ) =[
I3×3... −I3×3ξ 2
]
– p. 57/82
Examples
Examples:
◮d2
dt2~q = ~F ,w =
[
~F~q
]
; R(ξ ) =[
I3×3... −I3×3ξ 2
]
◮ddt x = Ax+Bu, y = Cx+Du, w =
xuy
;
R(ξ ) =
[
−ξ In×n +A B 0C D −Ip×p
]
◮ p0w+ p1ddt w+ · · ·+ pn dn
dtn w = 0; R(ξ ) = p(ξ ) with p(ξ ) = p0 + p1ξ + · · ·+ pnξ n.
– p. 57/82
LTIDSs
We should define what we mean by a solution of
R(
ddt
)w = 0
For ease of exposition, we takeC ∞ (R,Rw) solutions.Hence the behavior defined is
B =
{
w ∈ C∞ (R,Rw) | R
(ddt
)
w = 0
}
– p. 58/82
LTIDSs
We should define what we mean by a solution of
R(
ddt
)w = 0
For ease of exposition, we takeC ∞ (R,Rw) solutions.Hence the behavior defined is
B =
{
w ∈ C∞ (R,Rw) | R
(ddt
)
w = 0
}
B = kernel(R
(ddt
))‘kernel representation’ of this B.
LTIDS B ↔ kernel(R
(ddt
))
The theory of LTIDSs is basically the study of real polynomialmatrices, matrices over the ringR [ξ ].
– p. 58/82
LTIDSs
B = kernel(R
(ddt
))‘kernel representation’ of this B.
LTIDS B ↔ kernel(R
(ddt
))
The theory of LTIDSs is basically the study of real polynomialmatrices, matrices over the ringR [ξ ].
The theory of LTIDSs is the topic of lectures 2 and 3.
– p. 58/82
Properties of dynamical systems
– p. 59/82
Controllability
Assume thatΣ = (R,W,B) is time-invariant(to avoid irrelevant complications)
and T = R (for the sake of concreteness)
Σ is said to be controllable :⇔for all w1,w2 ∈ B, there existsT ≥ 0 and w ∈ B such that
w(t) =
{
w1(t) for t < 0w2(t −T ) for t ≥ T
– p. 60/82
A picture is worth a thousand words
1
w2
w
time
W
2w
1
w
w
time
W W
controllability ⇔ concatenability of trajectories after a delay
– p. 61/82
Stabilizability
Assume thatΣ = (R,Rw,B)T = R,W = Rw (for the sake of concreteness)
Σ is said to be stabilizable :⇔for all w ∈ B, there existsw′ ∈ B such that
w′(t) = w(t) for t < 0
andw′(t) → ∞ for t → ∞
– p. 62/82
In pictures
w’
w
0
time
Rw
stabilizability ⇔ trajectories can be steered to0
– p. 63/82
Autonomous systems
Assume thatΣ = (R,W,B)T = R (for the sake of concreteness)
Σ is said to be autonomous :⇔
w1,w2 ∈ B and w1(t) = w2(t) for t < 0 ⇒ w1 = w2.
– p. 64/82
In pictures
time
time
FUTURE
PAST
W
W
In an autonomous system, the past determines the future.
– p. 65/82
Autonomous systems
Assume thatΣ = (R,W,B)T = R (for the sake of concreteness)
Σ is said to be autonomous :⇔
w1,w2 ∈ B and w1(t) = w2(t) for t < 0 ⇒ w1 = w2.
Equivalently, there is a map
F : W(−∞,0) → W
[0,+∞)
such that
w ∈ B ⇒ w|[0,+∞)= F
(
w|(−∞,0)
)
– p. 66/82
Autonomous systems
Assume thatΣ = (R,Rw,B)T = R,W = Rw (for the sake of concreteness)
Σ is said to be autonomous :⇔
w1,w2 ∈ B and w1(t) = w2(t) for t < 0 ⇒ w1 = w2.
Examples:
◮ Kepler’s laws.
◮dndtnw = f
(
w, ddt w, . . . , dn−1
dtn−1w)
◮ In particular, ddt x = f (x)
– p. 67/82
Autonomous systems
Assume thatΣ = (R,Rw,B)T = R,W = Rw (for the sake of concreteness)
Σ is said to be autonomous :⇔
w1,w2 ∈ B and w1(t) = w2(t) for t < 0 ⇒ w1 = w2.
Examples:
◮ Kepler’s laws.
◮dndtnw = f
(
w, ddt w, . . . , dn−1
dtn−1w)
◮ In particular, ddt x = f (x)
Autonomous models aim at‘closed’ systems.Our aim is principally ‘open’ systems.
– p. 67/82
Observability
Consider Σ = (T,W1×W2,B).
w2 is said to be observable fromw1 in Σ :⇔
(w1,w′2),(w1,w
′′2) ∈ B ⇒ w′
2 = w′′2
– p. 68/82
In pictures
observed w2to−be−deduced SYSTEMw1variables variables
– p. 69/82
In pictures
observed w2to−be−deduced SYSTEMw1variables variables
observability ⇔ unobserved can be deduced from observed
Note: knowing the laws that govern the system !!
– p. 69/82
Observability
Consider Σ = (T,W1×W2,B).
w2 is said to be observable fromw1 in Σ :⇔
(w1,w′2),(w1,w
′′2) ∈ B ⇒ w′
2 = w′′2
Equivalently, iff there exists a mapF : WT1 → W
T2 such that
(w1,w2) ∈ B ⇒ w2 = F (w1)
– p. 70/82
Observability
Consider Σ = (T,W1×W2,B).
w2 is said to be observable fromw1 in Σ :⇔
(w1,w′2),(w1,w
′′2) ∈ B ⇒ w′
2 = w′′2
Equivalently, iff there exists a mapF : WT1 → W
T2 such that
(w1,w2) ∈ B ⇒ w2 = F (w1)
In applications, often‘observed’ = manifest, ‘to-be-deduced’ = latent (see lecture 2).In this case we can speak about‘an observable model’But in general observability depends on the partitionW = W1×W2.
– p. 70/82
Detectability
Consider Σ = (T,Rw1 ×Rw2,B).
w2 is said to be detectable fromw1 in Σ :⇔
(w1,w′2),(w1,w
′′2) ∈ B ⇒ w′
2(t)−w′′2(t) →t→∞ 0
w2 can be asymptotically deduced fromw1 (knowing the lawsof the system).
– p. 71/82
Controllability versus observability
In our setting, controllability is an intrinsic property of asystem.
Observability depends on the variable partition.
There is no ‘duality’ - we need more structure for that.
– p. 72/82
LTIDSs
Controllability, observability, stabilizability, auton omy, etc.will be studied in great detail for LTIDSs in lecture 2.
– p. 73/82
State controllability and observability
Consider classical Kalman definitions for
x = f (x,u) y = f (x,u)state spaceX, input spaceU, output spaceY
– p. 74/82
State controllability and observability
Consider classical Kalman definitions for
x = f (x,u) y = f (x,u)state spaceX, input spaceU, output spaceY
state controllability :⇔∀ x1,x2 ∈ X,∃ T ≥ 0 and u : [0,T ] → U
such that x(0) = x1 yields x(T ) = x2.
1
x2X2
x
This is aspecial caseof our controllability:variables: w = x = stateor w = (x,u) = (input, state)
– p. 74/82
State controllability and observability
Consider classical Kalman definitions for
x = f (x,u) y = f (x,u)state spaceX, input spaceU, output spaceY
Similar for state observabilityKalman definition: observed (u,y); to-be-deducedx.
– p. 74/82
State controllability and observability
Consider classical Kalman definitions for
x = f (x,u) y = f (x,u)state spaceX, input spaceU, output spaceY
Why should we be so concerned with the state?
If a system is not (state) controllable, what is it caused by?insufficient influence of the control?bad choice of the state?not properly editing of the equations?
Kalman’s definitions address a rather special situation.
– p. 74/82
Other sets of independent variables
– p. 75/82
Other time sets
The setting is identical for LTIDSs with time set
[0,∞),(−∞,0] or [t1, t2] .
– p. 76/82
Other time sets
The setting is identical for LTIDSs with time set
[0,∞),(−∞,0] or [t1, t2] .
For discrete-time systems with time-axisN := {0,1,2, . . .},the differential operator should be replaced by the shiftσ ,σ( f )(t) := f (t +1).Only non-negative powers ofσ make sense.
– p. 76/82
Other time sets
The setting is identical for LTIDSs with time set
[0,∞),(−∞,0] or [t1, t2] .
For discrete-time systems with time-axisN := {0,1,2, . . .},the differential operator should be replaced by the shiftσ ,σ( f )(t) := f (t +1).Only non-negative powers ofσ make sense.
For discrete-time systems with time-axisZ, the differentialoperator should be replaced by the shiftσσ( f )(t) := f (t +1).Both positive and negative powers ofσ make sense.σ−1( f )(t) := f (t −1).
– p. 76/82
Distributed phenomena
In distributed systems, ‘events’ are maps from a set ofindependent variables to a set ofdependent variables .Usually the independent variables are time and space.
���������������������������������������������������������������������������������������������������������������������������������������������������
���������������������������������������������������������������������������������������������������������������������������������������������������
������������������������������������������������������������������������������������������������������������������������������������������������������������������������
������������������������������������������������������������������������������������������������������������������������������������������������������������������������
������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������
������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������
event
��������������������������������������������������������������������������������������������������������������������������������������������������������������������������������
��������������������������������������������������������������������������������������������������������������������������������������������������������������������������������
Phenomenon
t
spacex
w
time
– p. 77/82
Distributed phenomena
In distributed systems, ‘events’ are maps from a set ofindependent variables to a set ofdependent variables .Usually the independent variables are time and space.
The behavior of a distributed system is usually described bya system of partial differential equations (PDEs).
– p. 77/82
Example: heat diffusion
�������������������������������������������������������������
���
q(x,t)
T(x,t)x
Events: maps fromR×R to R×R
B ={
(T,q) : R2 → [0,∞)×R | ∂∂ t T = ∂ 2
∂ x2 T +q}
set of independent variables:R2 or R+×R, time and space.set of dependent variables:R2 or R+×R,
temperature and heat flow.
– p. 78/82
Example: Maxwell’s equations
∇ ·~E =1ε0
ρ ,
∇×~E = −∂∂ t
~B,
∇ ·~B = 0 ,
c2∇×~B =1ε0
~j +∂∂ t
~E.
independent variables:(t,x,y,z) time and space ; R4
dependent variables:(~E,~B,~j,ρ)
electric & magnetic field, current & charge density ; R10.
James Clerk Maxwell1831 – 1879
– p. 79/82
Summary of Lecture 1
– p. 80/82
The main points
◮ A model is a subsetB of a universumU .B is the behavior of the model.
– p. 81/82
The main points
◮ A model is a subsetB of a universumU .B is the behavior of the model.
◮ Dynamical systems: events are time trajectories.
– p. 81/82
The main points
◮ A model is a subsetB of a universumU .B is the behavior of the model.
◮ Dynamical systems: events are time trajectories.
◮ Important system properties: controllability,stabilizability, observability, detectability, autonomy.
– p. 81/82
The main points
◮ A model is a subsetB of a universumU .B is the behavior of the model.
◮ Dynamical systems: events are time trajectories.
◮ Important system properties: controllability,stabilizability, observability, detectability, autonomy.
◮ Controllability and observability generalize theirclassical counterparts in a meaningful way.
– p. 81/82
The main points
◮ A model is a subsetB of a universumU .B is the behavior of the model.
◮ Dynamical systems: events are time trajectories.
◮ Important system properties: controllability,stabilizability, observability, detectability, autonomy.
◮ Controllability and observability generalize theirclassical counterparts in a meaningful way.
◮ Distributed phenomena fit well in this setting by taking amore general set of independent variables than just time.
– p. 81/82
End of Lecture 1
– p. 82/82