Vander Schaft

The port-Hamiltonian approach to physical system modeling and control, Namur, November, 2008 1
The port-Hamiltonian approach to

physical system modeling and control
Arjan van der Schaft, University of Groningen
Hans Zwart, University of Twente
In collaboration with Bernhard Maschke, Romeo Ortega,

Jacquelien Scherpen, Stefano Stramigioli, Alessandro Macchelli,
Peter Breedveld, Dimitri Jeltsema, Morten Dalsmo, Guido
Blankenstein, Damien Eberard, Goran Golo, Ram Pasumarthy,
Javier Villegas, Gerardo Escobar, Guido Blankenstein, Aneesh
Venkatraman, Rostyslav Polyuga ..
Part I : Network Modeling and Analysis

Part II : Control of Port-Hamiltonian Systems
Part I : Network Modeling and Analysis

1. Motivation and context
2. From passive systems to port-Hamiltonian systems
3. Some properties of port-Hamiltonian systems
4. Intro to distributed-parameter port-Hamiltonian systems
5. Mixed lumped- and distributed-parameter port-Hamiltonian

systems
6. Lumping of distributed-parameter physical systems, and

port-Hamiltonian dynamics on graphs and k -complexes.
7. Model reduction
8. Conclusions of Part I
From passive systems to port-Hamiltonian systems

A square nonlinear system
ẋ = f (x) + g(x)u, u ∈ Rm
Σ:
y = h(x), y ∈ Rm
where x ∈ Rn are coordinates for an n-dimensional state space X , is

passive if there exists a storage function H : X → R with
H(x) ≥ 0 for every x, such that
Z t2
H(x(t2 )) − H(x(t1 )) ≤ uT (t)y(t)dt
t1
for all solutions (u(·), x(·), y(·)) and times t1 ≤ t2 .

The system is lossless if ≤ is replaced by =.
If H is differentiable then ’passive’ is equivalent to

d
H ≤ uT y
dt
which reduces to (Willems, Hill-Moylan)
∂T H
∂x (x)f (x) ≤ 0
h(x) = g T (x) ∂H
∂x (x)
while in the lossless case ≤ is replaced by =.

In the linear case
ẋ = Ax + Bu
y = Cx
is passive if there exists a quadratic storage function H(x) = 12 xT Qx,
with Q = QT ≥ 0 satisfying the LMIs
AT Q + QA ≤ 0, C = BT Q
Every linear passive system with storage function H(x) = 21 xT Qx,

satisfying
ker Q ⊂ ker A
can be rewritten as a linear port-Hamiltonian system
ẋ = (J − R)Qx + Bu, J = −J T , R = RT ≥ 0
y = B T Qx,
in which case the storage function H(x) = 21 xT Qx is called the

Hamiltonian.
• Passive linear systems are thus port-Hamiltonian with

non-negative Hamiltonian.
• Conversely every port-Hamiltonian system with

non-negative Hamiltonian is passive.
Mutatis mutandis ’most’ nonlinear lossless systems can be written

as a port-Hamiltonian system
ẋ = J(x) ∂H
∂x (x) + g(x)u
y = g T (x) ∂H
∂x (x)
∂H
with J(x) = −J T (x) and ∂x (x) the column vector of partial
derivatives. Note that
∂H
ẋ = J(x) (x)
∂x
is the internal Hamiltonian dynamics known from physics, which in
classical mechanics can be written as
∂H
q̇ = ∂p (q, p)
ṗ = − ∂H
∂q (q, p)
with the Hamiltonian H the total (kinetic + potential) energy.

Similarly, most nonlinear passive systems can be written as a

port-Hamiltonian system (with dissipation)
ẋ = [J(x) − R(x)] ∂H
∂x (x) + g(x)u
y = g T (x) ∂H
∂x (x)
with R(x) = RT (x) ≥ 0 specifying the energy dissipation
d ∂T H ∂H
H=− (x)R(x) (x) + uT y ≤ uT y
dt ∂x ∂x
However, in network modeling it is the other way around: one

derives the system in port-Hamiltonian form (and if the Hamiltonian
H ≥ 0 then it is the storage function of a passive system).
The matrix J(x) corresponds to the internal power-conserving
interconnection structure of physical systems due to:
• Basic conservation laws such as Kirchhoff’s laws.

• Powerless constraints; kinematic constraints.
• Transformers, gyrators, exchange between different types of
energy.
The matrix R(x) corresponds to the internal energy dissipation

in the system (due to resistors, damping, viscosity, etc.)
Main message: start with port-Hamiltonian models instead of
passive models. Closer to physical modeling, and capturing more
information than just the energy-balance of passivity.
A bit of port-based network modeling

The passivity framework considers a system component, and its
power-exchange with other system components:
d
H ≤ uT y
dt
The feedback interconnection of two passive systems
d d
dt H1 ≤ uT1 y1 , dt H2 ≤ uT2 y2
u1 = −y2 + v1 , u2 = y1 + v2
leads to an interconnected system that is again passive, since
d
(H1 + H2 ) ≤ uT1 y1 + uT2 y2 = v1T y1 + v2T y2
dt
The feedback interconnection is a typical example of a

power-conserving interconnection (total power is zero):
uT1 y1 + uT2 y2 + v1T y1 + v2T y2 = 0

In port-based modeling, e.g. bond graphs, one looks at the
system as the power-conserving interconnection of ideal basic
system components: (energy-) storage elements, resistive
elements, transformers, gyrators, constraints, etc.
• The Hamiltonian of the resulting port-Hamiltonian system is

the sum of the energies of the storage elements.
• The J - and B -matrix is determined by the transformers,

gyrators, constraints, and the power-conserving
interconnection.
• The R-matrix is determined by the resistive elements, and the

way they are connected.
An k dimensional storage element is determined by a

k -dimensional state vector x = (x1 , · · · , xk ) and a Hamiltonian
H(x1 , · · · , xk ) (energy storage), defining the lossless system
ẋi = fi , i = 1, · · · , k
∂H
ei = ∂xi (x1 , · · · , xk )
d
Pk
dt H = i=1 fi ei
Such a k - dimensional storage component is written in vector

notation as a port-Hamiltonian system with J = 0, R = 0, and B = I :
ẋ = f
∂H
e = ∂x (x)
The elements of x are called energy variables, those of ∂H

∂x (x)
co-energy variables. Furthermore the elements of f are flow
variables, and of e effort variables.
Note that eT f is power.
Example: The ubiquitous mass-spring-damper system:

Two storage elements:
• Spring Hamiltonian Hs (q) = 21 kq 2 (potential energy)
q̇ = fs = velocity
dHs
es = dq (q) = kq = force
1 2
• Mass Hamiltonian Hm (p) = 2m p (kinetic energy)
ṗ = fm = force
dHm p
em = dp (p) = m = velocity
interconnected by
fs = em = y, fm = −es + u
(power-conserving since fs es + fm em = uy ) yields the

port-Hamiltonian system
      
∂H
q̇ 0 1 (q, p) 0
  =    ∂q  (q, p) +   u
∂H
ṗ −1 0 ∂p (q, p) 1

h i ∂H (q, p)
y = 0 1  ∂q 
∂H
∂p (q, p)
with
H(q, p) = Hs (q) + Hm (p)
Addition of the damper
dR 1 2
ed = = cfd , R(fd ) = cfd (Rayleigh function)
dfd 2
via the extended interconnection
fs = em = fd = y, fm = es − ed + u
leads to the mass-damper-spring system in port-Hamiltonian

form
         
∂H
q̇ 0 1 0 0 (q, p) 0
  = ( − )  ∂q  +  u
∂H
ṗ −1 0 0 c ∂p (q, p) 1
 
h i ∂H (q, p)
y = 0 1  ∂q 
∂H
∂p (q, p)
Example: Electro-mechanical systems
      
∂H
q̇ 0 1 0 ∂q (q, p, φ) 0
  
 ṗ  = −1 0

∂H
   ∂H
0   ∂p (q, p, φ)
 
0 V,
+  I= (q, p, φ)
    ∂ϕ
ϕ̇ 0 0 − R1 ∂H
∂ϕ (q, p, φ) 1
Coupling electrical/mechanical domain via Hamiltonian H(q, p, φ).
Example: LC circuits. Two inductors with magnetic energies

H1 (ϕ1 ), H2 (ϕ2 ) (ϕ1 and ϕ2 magnetic flux linkages), and capacitor
with electric energy H3 (Q) (Q charge). V denotes the voltage of
the source.
L1 L2
C
ϕ1 Q ϕ2
V
Question: How to write this circuit as a port-Hamiltonian system

in a modular way?
Hamiltonian equations for the components of the LC-circuit:
Inductor 1 ϕ̇1 = f1 (voltage)

∂H1
(current) e1 = ∂ϕ1
Inductor 2 ϕ̇2 = f2 (voltage)

∂H2
(current) e2 = ∂ϕ2
Capacitor Q̇ = f3 (current)
∂H3
(voltage) e3 = ∂Q
All are port-Hamiltonian systems with J = 0 and g = 1.
If the elements are linear then the Hamiltonians are quadratic, e.g.
ϕ1
H1 (ϕ1 ) = 2L1 1 ϕ21 , and ∂H
∂ϕ1
1
= L1 = current , etc.
Kirchhoff’s interconnection laws in f1 , f2 , f3 , e1 , e2 , e3 , f = V, e = I are

    
−f1 0 0 1 −1 e1
    
−f   0 0 −1 0  e 
 2    2
 =  
−f3  −1 1 0 0  e3 
 
  
e 1 0 0 0 f
Substitution of eqns. of components yields port-Hamiltonian
system  
    ∂H  
ϕ˙1 0 0 −1  ∂ϕ1  1
      
ϕ˙2  = 0 0 ∂H  +   f
   1   ∂ϕ2  0
 
 
Q̇ 1 −1 0 ∂H 0
∂Q
∂H
e = ∂ϕ1
with H(ϕ1 , ϕ2 , Q) := H1 (ϕ1 ) + H2 (ϕ2 ) + H3 (Q) total energy.

However, this class of port-Hamiltonian systems is not closed

under interconnection:
Figure 1: Capacitors and inductors swapped
Interconnection leads to algebraic constraints between the state

variables Q1 and Q2 .
How to model DAEs as port-Hamiltonian systems ?

Intermezzo: what is the appropriate generalization of the
skew-symmetric mapping J ? Answer: Dirac structures.
(’From skew-symmetric mappings to skew-symmetric relations’)
Power is defined by
P = e(f ) =: < e | f >= eT f, (f, e) ∈ V × V ∗ .
where the linear space V is called the space of flows f (e.g.
currents), and V ∗ the space of efforts e (e.g. voltages).
Symmetrized form of power is the indefinite bilinear form , on
V × V ∗:
(f a , ea ), (f b , eb ) := < ea | f b > + < eb | f a >,
(f a , ea ), (f b , eb ) ∈ V × V ∗ .
Definition 1 (Weinstein, Courant, Dorfman) A (constant)

Dirac structure is a subspace
D ⊂ V × V∗
such that
D = D⊥ ,
where ⊥ denotes orthogonal complement with respect to the
bilinear form , .
For a finite-dimensional space V this is equivalent to
(i) < e | f > eT f = 0 for all (f, e) ∈ D ,
(ii) dim D = dim V .
For any skew-symmetric map J : V ∗ → V its graph
{(f, e) ∈ V × V ∗ | f = Je}
is a Dirac structure !
For many systems, especially those with 3-D mechanical

components, the interconnection structure will be modulated by
the energy or geometric variables.
This leads to the notion of non-constant Dirac structures on
manifolds.
Definition 2 Consider a smooth manifold M . A Dirac structure on

M is a vector subbundle D ⊂ T M ⊕ T ∗ M such that for every x ∈ M
the vector space
D(x) ⊂ Tx M × Tx∗ M
is a Dirac structure as before.

Geometric definition of a port-Hamiltonian system
ẋ fx f
H D(x)
∂H ex e
∂x (x)
Figure 2: Port-Hamiltonian system
The dynamical system defined by the DAEs

∂H
(−ẋ(t) = fx (t), (x(t)) = ex (t), f (t), e(t)) ∈ D(x(t)), t∈R
∂x
is called a port-Hamiltonian system.
Particular case is a Dirac structure D(x) ⊂ Tx X × Tx∗ X × F × F ∗

given as the graph of the skew-symmetric map
    
f −J(x) −g(x) e
 x =  x ,
e g T (x) 0 f
leading (fx = −ẋ, ex = ∂H

∂x (x)) to a port-Hamiltonian system as
before
ẋ = J(x) ∂H ∂x (x) + g(x)f, x ∈ X , f ∈ Rm
e = g T (x) ∂H
∂x (x), e ∈ Rm
Energy-dissipation is included by terminating some of the ports

by static resistive elements
fR = −F (eR ), where eTR F (eR ) ≥ 0, for all eR .
If H ≥ 0 then the system stays passive with respect to the

remaining flows and efforts f and e:
d
H ≤ eT f
dt
This leads, e.g. for linear damping, to input-state-output
port-Hamiltonian systems in the form

 ẋ = [J(x) − R(x)] ∂H (x) + g(x)f
∂x
 e = g T (x) ∂H (x)
∂x
where J(x) = −J T (x), R(x) = RT (x) ≥ 0 are the interconnection and

damping matrices, respectively.
Example: Mechanical systems with kinematic constraints
Constraints on the generalized velocities q̇ :
AT (q)q̇ = 0.
This leads to constrained Hamiltonian equations
∂H
q̇ = ∂p (q, p)
ṗ = − ∂H
∂q (q, p) + A(q)λ + B(q)u
0 = AT (q) ∂H
∂p (q, p)
y = B T (q) ∂H
∂p (q, p)
with H(q, p) total energy, and λ the constraint forces.

By elimination of the constraints and constraint forces one derives
a port-Hamiltonian model without constraints.
Can be extended to general multi-body systems.

Intermezzo: Relation with classical Hamiltonian equations
∂H
ẋ = J (x)
∂x
with constant or ’ integrable’J - matrix admits coordinates
x = (q, p, r) in which
 
0 I 0 q̇ = ∂H∂p (q, p, r)
 
J = −I 0 0 , ∂H

 ṗ = − ∂q (q, p, r)
0 0 0 ṙ = 0
For constant or integrable Dirac structure one gets Hamiltonian
DAEs
q̇ = ∂H∂p (q, p, r, s)
ṗ = − ∂H
∂q (q, p, r, s)
ṙ = 0
∂H
0 = ∂s (q, p, r, s)
Some properties of port-Hamiltonian systems

Port-Hamiltonian systems modeling encodes more
information than energy-balance.
The Dirac structure determines all the Casimir functions

(conserved quantities which are independent of H ).
Example: In the first LC circuit the total flux φ1 + φ2 is a
conserved quantity that is solely determined by the circuit topology.
(In Part II this will be used for set-point control.)
Furthermore, the Dirac structure directly determines the algebraic
constraints.
Example: In the second LC-circuit the state variables Q1 and Q2
are related by
Q1 Q2
=
C1 C2
Any power-conserving interconnection of

port-Hamiltonian systems is again port-Hamiltonian
• The resulting Hamiltonian is the sum of the Hamiltonians of
the individual systems.
• The Dirac structure is determined by the Dirac structures of

the individual systems, and the way they are interconnected.
• The resistive structure is determined by the resistive structures

of the individual systems.
Conclusion: port-Hamiltonian systems theory provides a

modular framework for modeling and analysis of complex
multi-physics lumped-parameter systems.
Network modeling is prevailing in modeling and simulation of

lumped-parameter physical systems (multi-body systems, electrical
circuits, electro-mechanical systems, hydraulic systems, robotic
systems, etc.), with many advantages:
• Modularity and flexibility. Re-usability (‘libraries’).
• Multi-physics approach.
• Suited to design/control.
Disadvantage of network modeling: it generally leads to a large set

of DAEs, seemingly without any structure.
Port-based modeling and port-Hamiltonian system theory
identifies the underlying structure of network models of
physical systems, to be used for analysis, simulation and
control.
Distributed-parameter port-Hamiltonian systems
fa fb
ea eb
a b
Figure 3: Simplest example: Transmission line
Telegrapher’s equations define the boundary control system

∂Q ∂ ∂ φ(z,t)
∂t (z, t) = − ∂z I(z, t) = − ∂z L(z)
∂φ ∂ ∂ Q(z,t)
∂t (z, t) = − ∂z V (z, t) = − ∂z C(z)
fa (t) = V (a, t), e1 (t) = I(a, t)
fb (t) = V (b, t), e2 (t) = I(b, t)
Transmission line as port-Hamiltonian system

Define internal flows fx = (fE , fM ) and efforts ex = (eE , eM ):
electric flow fE : [a, b] → R

magnetic flow fM : [a, b] → R
electric effort eE : [a, b] → R
magnetic effort eM : [a, b] → R
together with external boundary flows f = (fa , fb ) and boundary
efforts e = (ea , eb ). Define the infinite-dimensional Dirac structure
    
∂
fE 0 ∂z e
  =   E 
∂
fM ∂z 0 eM
   
fa,b eE|a,b
  =  
ea,b eM |a,b
This defines a Dirac structure on the space of internal flows and

efforts and boundary flows and efforts.
Substituting (as in the lumped-parameter case)

fE = − ∂Q∂t

∂ϕ  x
f = −ẋ
fM = − ∂t

Q ∂H
eE = C = ∂Q
 ∂H
e =
eM = ϕ
= ∂H  x ∂x
L ∂ϕ
with, for example, quadratic energy density
1 Q2 1 ϕ2
H(Q, ϕ) = +
2 C 2 L
we recover the telegrapher’s equations.
Of course, the telegrapher’s equations can be rewritten as the

linear wave equation
∂2Q ∂ ∂I ∂ ∂ φ
∂t2 = − ∂z ∂t = − ∂z ∂t L =
2
∂ 1 ∂φ ∂ 1 ∂ Q 1 ∂ Q
− ∂z L ∂t = ∂z L ∂z C = LC ∂z 2
(if L(z), C(z) do not depend on z ), or similar expressions in φ, I or V .

The same equations hold for a vibrating string, or for a
compressible gas/fluid in a one-dimensional pipe.
Basic question:
Which of the boundary variables fa , fb , ea , eb can be considered to

be inputs, and which outputs ? See Lecture of Hans Zwart.
Example 2: Shallow water equations; distributed-parameter

port-Hamiltonian system with non-quadratic Hamiltonian
The dynamics of the water in an open-channel canal can be

described by      
h v h h
∂t   +   ∂z  = 0

v g v v
with h(z, t) the height of the water at position z , and v(z, t) the
velocity (and g gravitational constant).
This can be written as a port-Hamiltonian system by recognizing
the total energy
b
1
Z
H(h, v) = [hv 2 + gh2 ]dz
2 a
yielding the co-energy functionsa
eh = ∂H
∂h = 12 v 2 + gh Bernoulli function
∂H
ev = ∂v = hv mass flow
It follows that the shallow water equations can be written, similarly

to the telegraphers equations, as
∂h ∂ ∂H
∂t (z, t) = − ∂z ∂v
∂v ∂ ∂H
∂t (z, t) = − ∂z ∂h
with boundary variables −hv|a,b and ( 21 v 2 + gh)|a,b .

a DanielBernoulli, born in 1700 in Groningen as son of Johann Bernoulli, profes-
sor in mathematics at the University of Groningen and forerunner of the Calculus
of Variations (the Brachistochrone problem).
Paying tribute to history:
Figure 4: Johann Bernoulli, professor in Groningen 1695-1705.
Figure 5: Daniel Bernoulli, born in Groningen in 1700.

We obtain the energy balance

b
d 1
Z
[hv 2 + gh2 ]dz = −(hv)( v 2 + gh)|ba
dt a 2
which can be rewritten as
−v( 21 gh2 )|ba − v( 12 hv 2 + 12 gh2 ))|ba =
velocity × pressure + energy flux through the boundary

Conservation laws
All examples sofar have the same structure
∂α1 ∂ ∂H ∂
∂t (z, t) = − ∂z ∂α2 = − ∂z β2
∂α2 ∂ ∂H ∂
∂t (z, t) = − ∂z ∂α1 = − ∂z β1
with boundary variables β1 |{a,b} , β2 |{a,b} , corresponding to two

coupled conservation laws:
d
Rb Rb ∂
dt a
α1 = − β
a ∂z 2
= β2 (a) − β2 (b)
d
Rb Rb ∂
dt a
α2 = − β
a ∂z 1
= β1 (a) − β1 (b)
(In the transmission line, α1 and α2 is charge- and flux-density, and

β1 , β2 voltage V and current I , respectively.)
For some purposes it is illuminating to rewrite the equations in

terms of the co-energy variables β1 , β2 :
   2 2
   2 2
 
∂β1 ∂ H ∂ H ∂α1 ∂ H ∂ H ∂β2
2 ∂α1 α2   ∂t  ∂α21 ∂α1 α2   ∂z 
 ∂t  =  ∂α
2
1 = − 
∂β2 ∂ H ∂2H ∂α2 ∂2H ∂2H ∂β1
∂t ∂α2 α1 ∂α22 ∂t ∂α2 α1 ∂α22 ∂z
For the transmission line this yields

    
∂V 1 ∂V
0
 ∂t  = −  C   ∂z 
∂I 1 ∂I
∂t L 0 ∂z
The matrix is called the characteristic matrix, whose eigenvalues

1 1
are the characteristic velocities √LC and − √LC corresponding to
the characteristic eigenvectors (and curves).
For the shallow water equations this yields

    
∂β1 ∂β1
v g
 ∂t  = −    ∂z 
∂β2 ∂β2
∂t h v ∂z
with
1 2
β1 = v + gh, β2 = hv
2
being the Bernoulli function and mass flow, respectively.
√
This corresponds to two characteristic velocities v ± gh, which
are, like in the transmission line case, of opposite sign (subcritical
or fluvial flow) if
v 2 ≤ gh
Because the Hamiltonian is non-quadratic, and thus the pde’s are
nonlinear, the characteristic curves may intersect, corresponding
to shock waves.
Mixed lumped- and distributed-parameter

port-Hamiltonian systems
Typical example: power-converter connected via a transmission
line to a resistive load or an induction motor:
• The power-converter is a port-Hamiltonian system (with

switching Dirac structure).
• Transmission line is distributed-parameter port-Hamiltonian

system.
• Induction motor is a port-Hamiltonian system, with

Hamiltonian being the electro-mechanical energy.
Power converter connected to the load via transmission line
L D Il
Vl
E S C VC = VL line R
Figure 6: The Boost converter with a transmission line

Boost power converter
+ L − D
+
+ +
E S C R
− −
−
Figure 7: Boost circuit with clamping diode
The circuit consists of a capacitor C with electric charge qC , an

inductor L with magnetic flux linkage φL , and a resistive load R,
together with an ideal diode and an ideal switch S , with switch
positions s = 1 (switch closed) and s = 0 (switch open).
The voltage-current characteristics of the ideal diode and switch

are depicted in Figure 8.
iS
6iD 6
vS
-
−vD
-
Figure 8: Voltage-current characteristic of an ideal diode and ideal

switch
The ideal diode thus satisfies the complementarity conditions:
vD iD = 0, vD ≤ 0, iD ≥ 0.
This yields the port-Hamiltonian model

1 2 1 2
(with H(qC , φL ) = 2C qC + 2L φL ):
        
qC
q̇C − R1 1−s ∂H
∂qC = C 0 siD
  =   + E +  
∂H φL
φ̇L s−1 0 ∂φL = L 1 (s − 1)vD
φL
I = L
Assume that the switch and the diode are coupled in the following
sense: if the switch is closed (s = 1) then the diode is open
(iD = 0), while if the switch is open (s = 0), then the diode is
closed (vD = 0). (This means that we disregard the so-called
discontinuous modes.)
In this case we obtain the switching port-Hamiltonian system

      
qC
q̇C − R1 1−s ∂H
= 0
  =    ∂qC C 
+  E
∂H φ
φ̇L s−1 0 ∂φL = L
L
1
 
∂H qC
h i
∂qC = C φL
I = 0 1  φL
=
L
∂H
∂φL = L
A port-Hamiltonian system where the J -matrix depends on the

switch position.
In general, a switching port-Hamiltonian system (without algebraic

equality and inequality constraints) is defined as
∂H
ẋ = F (ρ)z + g(ρ)u, z= ∂x (x)
y = g T (ρ)z
with ρ ∈ {0, 1}p , and
F (ρ) = J(ρ) − R(ρ), J(ρ) = −J T (ρ), R(ρ) = RT (ρ) ≥ 0
Note that the system is passive for every switching sequence:
d ∂T H ∂H
H=− (x)R(ρ) (x) + uT y ≤ uT y
dt ∂x ∂x
Oriented graphs and Kirchhoff behavior
Figure 9: Kirchhoff
An oriented graph G consists of a finite set N of nodes (vertices)

and a finite set B of branches (edges). To every branch b ∈ B
there corresponds an ordered pair (n, m) ∈ N 2 representing the
starting node n and the final node m of this branch.
An oriented graph is specified by its incidence matrix D , which is

an n̄ × b̄ matrix, n̄ being the number of nodes and b̄ being the
number of branches, with (i, j)−th element dij equal to 1 if the j -th
branch is a branch towards node i, equal to −1 if the j -th branch is
a branch originating from node i, and equal to 0 otherwise.
The node space Λ0 is the vector space of all functions from N to

R. Clearly, Λ0 can be identified with Rn̄ .
The branch space Λ1 is the vector space of all functions from B to
R. Again, Λ1 can be identified with Rb̄ .
For an electrical circuit Λ1 will be the vector space of currents
over the branches in the circuit.
The dual space of the vector space Λ1 will be denoted by Λ1 ; the
vector space of voltages over the branches.
The product < V |I > of a vector of currents I ∈ Λ1 with a vector of
voltages V ∈ Λ1 is the total power over the circuit.
Λ0 is the vector space of potentials over the nodes, with dual
space Λ0 .
(Note: since Λ0 and Λ1 have a canonical basis we can identify them
with their dual spaces Λ0 and Λ1 .)
The incidence matrix D is the matrix representation of a map
∂ : Λ1 → Λ0
called the incidence operatora . The adjoint map of the incidence

operator ∂ is the linear map
d : Λ0 → Λ1
which is called the co-incidence operator. The matrix

representation of the map d is given by D T .
a In the literature this operator is usually called the boundary operator.
Kirchhoff’s current laws (KCL) for a circuit are given as
∂I = 0
while Kirchhoff’s voltage laws (KVL) take the form
V ∈ im d
(The kernel of the incidence operator ∂ is the cycle space of the

graph, the image of the co-incidence operator d is its cut space.)
Kirchhoff’s voltage laws are thus
V = dφ
for some φ ∈ Λ0 (the vector of potentials at every node). In matrix

notation Kirchhoff’s laws are
DI = 0, V = DT φ
Tellegen’s theorem automatically follows from Kirchhoff’s laws.

Take any current distribution I satisfying KCL, and any voltage
distribution V satisfying KVL. Then
X X X
Vb Ib = (dφ)b Ib = φn (∂I)n = 0
b b n
since I satisfies KCL ∂I = 0.

In particular, Tellegen’s theorem implies that for any actual current
and voltage distribution over the circuit the total power is equal to
zero.
The Kirchhoff behavior BK (G) of a graph G with incidence

operator ∂ is
BK (G) := {(I, V ) ∈ Λ1 × Λ1 | I ∈ ker ∂, V ∈ im d}
It immediately follows that the Kirchhoff behavior defines a Dirac

structure.
D ⊂ V × V ∗ is a Dirac structure if < v ∗ |v >= 0 for every (v, v ∗ ) ∈ D
while dim D = dim V .
Open graphs
An open graph G is obtained from an ordinary graph with set of
nodes N by identifying a subset Ne ⊂ N of external nodes. The
remaining subset Ni := N − Ne are the internal nodes of the open
graph.
Kirchhoff’s current laws now take the form
 
Ie
∂I +  =0
0
with Ie ∈ Λe the vector of external currents entering the graph at

the external nodes (terminals).
In this case we obtain

P P P
b Vb Ib = b (dφ)b Ib =
φn (∂I)n =
n
P P P (1)
ni φni (∂I)ni + ne φne (∂I)ne = − ne φne Ine
since I satisfies the Kirchhoff’s current laws (∂I)ni = 0 and

(∂I)ne = −Ine .
Thus, for open graphs the total power over the graph is not zero
P
but equal to − ne φne Ine (the incoming power at the external
nodes).
Even though the vector of potentials φe at the external nodes is

not uniquely determined by the vector V of voltages, the
P
expression ne φne Ine is uniquely determined.
The freedom in the choice of φ corresponding to the same vector

of voltages V is given by all vectors ψ such that D T ψ = 0. Hence
the freedom in φe is given by all vectors ψe such that for some ψi it
holds that DeT ψe + DiT ψi = 0, where De is the submatrix of D
consisting of the first n̄e rows and Di is the submatrix consisting of
the last n̄ − n̄e rows. For any such ψe we have
ψeT Ie = −ψeT De I = ψiT Di I = 0

The image of D , or equivalently the kernel of D T , can be

characterized as follows. First
X
T
0 = 11 DI = Ine (2)
ne
Hence the external part of the Kirchhoff behavior of an open graph

is constrained by the obvious fact that all external currents sum up
to zero.
In general, the rank of D is equal to n̄ − kG , where kG denotes the
number of connected components.
  if Gconsists
For example,  of two connected components, then the
11 0
vectors   and   span the kernel of D T . This implies that both
0 11
the sum of the external currents belonging to the first component
as well as the sum of the external currents of the second
component are equal to zero.
It turns out that an open graph G can be closed (in a unique

fashion) to an ordinary graph Ḡ while keeping the same number of
connected components.
Let G be connected. Then add one virtual node n̄, and virtual
edges from this virtual node n̄ to every external node ne ∈ Ne . The
Kirchhoff behavior of this graph Ḡ extends the Kirchhoff behavior
of the open graph G . In fact, to the virtual node n̄ we may
associate an arbitrary potential φn̄ (a ground-potential), and we
P
may rewrite the righthand-side of (1) as (since e Ine = 0)
X X
− (φne − φn̄ )Ine = − Vne Ine
ne ne
where Vne := φne − φn̄ and Ine denote the voltage, respectively
current, over the virtual edge towards the external node ne .
Thus the incoming power is rewritten as a product of external
currents and voltages.
Constitutive relations for every branch.

Simplest case is a resistive relation between Ib and Vb such that
Vb Ib ≥ 0.
In the case of a capacitive relation one defines the charge Qb
together with the electric energy Hb (Qb ). The constitutive relations
are then given by
dHb
Q̇b = −Ib , Vb = (Qb )
dQb
Alternatively, in the case of an inductor one specifies the magnetic
energy Hb (ϕb ), where ϕb denotes the flux, together with the
relations
dHb
ϕ̇b = −Vb , Ib = (ϕb )
dϕb
Substituting these constitutive relations into the Kirchhoff behavior
(and corresponding Dirac structure) defined by the graph results in
a port-Hamiltonian system description.
Interconnection of open graphs
Consider two open graphs G A and G B with respective sets of

external nodes NeA and NeB . Suppose we want to interconnect
these open graphs over a set of shared external nodes
N̄ ⊂ NeA ∩ NeB
This may be done by identifying the shared nodes corresponding to

the two graphs, leading to an interconnected open graph G A k G B
with resulting set of internal nodes
NiA ∪ NiB ∪ N̄
and external nodes
(NeA − N̄ ) ∪ (NeB − N̄ )
Extension to higher-order networks
An oriented graph G with incidence operator ∂ defines a

1-dimensional complex. Indeed, the sequence
∂ 11
Λ1 → Λ0 → R
satisfies 11 ◦ ∂ = 0.
The ’algebraic-topological invariants’ of this 1-complex (the
so-called Betti numbers) are nothing else than the number of
connected components.
An arbitrary k -complex Λ is specified by a sequence of linear spaces

Λ0 , Λ1 , · · · , Λk , together with a sequence of incidence operators
∂k ∂k−1 ∂ ∂
Λk → Λk−1 → · · · Λ1 →1 Λ0 →0 0
with the property that
∂j−1 ◦ ∂j = 0, j = 1, · · · , k
The vector spaces Λj , j = 0, 1 · · · , k, are called the spaces of

j -chains.
Each Λj is generated by a finite set of j -cells in the sense that Λj is
the set of functions from the set of j -cells to R.
A typical example of a 3-complex is the triangularization of a

3-dimensional manifold, with the j -cells, j = 0, 1, 2, 3, being the sets
of vertices, edges, faces, and tetrahedra, and ∂1 , ∂2 , ∂3 capturing
the incidence structure.
Denoting the dual linear spaces by Λj , j = 0, 1 · · · , k, we have the
following dual sequence
d d d
Λ0 →1 Λ1 →2 Λ2 · · · Λk−1 →
k
Λk
having the analogous property
dj ◦ dj−1 = 0, j = 2, · · · , k
The elements of Λj are called j -cochains.

Consider any k -complex Λ, with k -chains α ∈ Λk and k -cochains

β ∈ Λk . We define, similarly as in the case of a graph (1-complex)
its Kirchhoff behavior as
BK (Λ) := {(α, β) ∈ Λk × Λk |
∂k α = 0, ∃φ ∈ Λk−1 s.t. β = dk φ}
As before, it is immediately seen that BK (Λ) ⊂ Λk × Λk is a Dirac

structure. In particular, it follows that
< β | α >k = 0
for every (α, β) ∈ BK (Λ), where < · | · >k denotes the duality product
between the dual linear spaces Λk and Λk .
Open k -complexes
An open k -complex is obtained by identifying a subset N(k−1)e of

the set of all (k − 1)-cells, called the external (k − 1)-cells.
Define the linear space of functions from this subset of (k − 1)-cells
to R as Λe ⊂ Λk−1 (the space of ’external currents’).
As before, Kirchhoff’s voltage laws remain unchanged
β = dk φ,
while Kirchhoff’s current laws are modified into

 
αe
∂k α +  =0
0
where αe is the vector of external currents associated to the

external (k − 1)-cells. We obtain
< β | α >k =< dk φ | α >k =< φ | ∂k α >k−1 =

 
−αe
<φ|   >k−1 = − < φe | αe >k−1
0
where φe denotes the vector of potentials at the external

(k − 1)-cells.
Similar to graphs it follows that the Kirchhoff current laws for open
k -complexes Dke α = −αe imply certain constraints for the external
’currents’ αe . Indeed, by the fact that
∂k−1 ◦ ∂k = 0
it follows that
D(k−1)e αe = 0
Furthermore, we can uniquely extend the open k -complex to an

ordinary k -complex while keeping the same Betti numbers.
This means again that the external power can be rewritten as a
product of ’external currents’ and ’external voltages’.
Hamiltonian dynamics on a k -complex
There are a number of canonical ways to define ’physical’ dynamics

on k -complexes.
First option: define the Dirac structure
fx = −dk f, fx ∈ Λk , f ∈ Λk−1
e = ∂k ex , ex ∈ Λk , e ∈ Λk−1
Associate to every k -cell an energy storage, leading to a total

energy storage H(x), where x ∈ Λk denotes the vector of energy
variables, with
∂H
ẋ = −fx , ex = (x)
∂x
Furthermore, associate to every (k − 1)-cell the resistive relation
f = −Re, R = RT ≥ 0
This yields the relaxation dynamics
∂H
ẋ = −dk e = dk R f = −dk R ∂k (x), x ∈ Λk
∂x
with the property that
dH ∂H T ∂H
= −(∂k (x)) R ∂k (x) ≤ 0
dt ∂x ∂x
For an open complex with external (k − 1)-cells and external

’currents’ Λ̄ ⊂ Λk−1 the definition is modified as follows.
Consider instead
   
f k  
f
fx = −dk  , fx ∈ Λ , ∈ Λk−1 , fb ∈ Λ̄k−1
fb f
   b (3)
e e
  = ∂k ex , ex ∈ Λk ,   ∈ Λk−1 , fb ∈ Λ̄k−1
eb eb
with fb , eb corresponding to the external (k − 1)-cells,

and f, e corresponding to the internal (k − 1)-cells.
∂H
Imposing the same storage relations fx = −ẋ, ex = ∂x (x) and
resistive relations f = −Re we arrive at
ẋ = −drk R ∂kr ∂H b
∂x (x) + dk fb
eb = ∂kb ∂H
∂x (x)
 
h i ∂kr
where we have split dk as dk = drk dbk and ∂k =  .
∂kb
This defines a port-Hamiltonian system with inputs fb and
outputs eb .
Heat transfer on a 2-complex

We will write the heat transfer as a conservation law in terms of
the conservation of internal energy:
The internal energy u of the 2-complex is a 2-cochain, u ∈ Λ2 (with
every component of u denoting the energy of the corresponding
2-cell).
The thermodynamic properties are defined by Gibbs’ relation, and
generated by the entropy function s = s(u) as thermodynamic
potential. Since we consider transformations with constant volume
and without mass transfer, Gibbs’ relation reduces to the definition
of the vector of intensive variables eu , conjugated to the extensive
variables u by
∂s
eu = (u)
∂u
The components of the vector eu ∈ Λ2 are equal to the reciprocal
of the temperature in each 2-cell.
The heat conduction is given by the heat flux
f ∈ Λ1
describing the heat flux through every 1-cell (edge). This flux
arises from thermal non-equilibrium, defined by the fact that the
temperature is varying over the 2-cells.
Its conjugate vector of variables is the thermodynamical driving
force vector
e ∈ Λ1
given as the vector of the differences of the reciprocals of the

temperatures of the 2-cells with common edges (1-cells)
e = ∂2 eu
By Fourier’s law the heat flux due to thermal non-equilibrium is

expressed as
f = R(eu ) e,
with R(eu ) = RT (eu ) ≥ 0 depending on the heat conduction

coefficients. (Note the sign-difference !). Finally
du
= d2 f
dt
Hence the resulting system is a port-Hamiltonian system (of
relaxation type) defined on the 2-complex, with vector of state
variables x given by the internal energy vector u, and Hamiltonian
s(u).
By the different sign the entropy s(u) satisfies
ds ∂s ∂s
= (∂2 (u))T R(eu )∂2 (u) ≥ 0
dt ∂u ∂u
expressing the fact that the entropy s(u) is monotonously

increasing.
The exchange of heat through the boundary of the system can be
incorporated by splitting the edges (1-cells) into internal edges with
the resistive relation and external (boundary) edges. This would
lead to
ds ∂s T ∂s
= (∂2 (u)) R(eu ) ∂2 (u) + eb fb
dt ∂u ∂u
with fb , eb denoting the heat flux, respectively, thermodynamical
driving force, through the boundary 1-cells.
Another type of Hamiltonian dynamics is obtained by considering

the relations
fx1 = dk ex2 , fx1 ∈ Λk , ex2 ∈ Λk−1

fx2 = −∂k ex1 , ex1 ∈ Λk , fx2 ∈ Λk−1
defining again a Dirac structure, together with two energy-storage

relations
1
ẋ1 = −fx1 , ex1 = ∂H∂x1 (x1
)
2 ∂H 2 2
ẋ = −fx2 , ex2 = ∂x2 (x )
leading to the oscillatory dynamics
∂H
ẋ1 = −dk ∂x 1 2
2 (x , x )
∂H
ẋ2 = ∂k ∂x 1 2
1 (x , x )
Example 1: (k = 2) Discretized Maxwell’s equations on a

2-dimensional domain.
Example 2: Formation controla
Consider n point masses
ṗi = ui , i = 1, · · · , n
pi
with pi denoting their momenta, and vi = m i
their velocities.
Suppose v̄ is a desired joint velocity vector, and moreover, we want
their position vectors qi converge to a certain desired formation,
e.g. (for n = 3)
|q1 − q2 | = |q2 − q3 | = |q3 − q1 | = 1
This determines a graph structure with incidence matrix D .

a M. Arcak, IEEE TAC, 52, August 2007
Apply feedback to obtain
q̇i = ξi + v̄
ξ˙i = fi
and use
(D T ⊗ I3 )(q̇) = (D T ⊗ I3 )(ξ)
as an input to a Hamiltonian integrator dynamics defined on the

edges, leading to an input
f := −(D ⊗ I3 )(ψ)
for the dynamics at the nodes.

By designing carefully the Hamiltonian corresponding to the
dynamics on the edges, and by adding damping, one obtains the
desired formation control.
Model reduction of port-Hamiltonian systems

• Network modeling of complex lumped-parameter systems
(circuits, multi-body systems) often leads to high-dimensional
models.
• Structure-preserving spatial discretization of

distributed-parameter port-Hamiltonian systems yields
high-dimensional port-Hamiltonian models.
• Lumped-parameter modeling of systems like MEMS gives

high-dimensional port-Hamiltonian systems.
• Controller systems may be in first instance

distributed-parameter, and need to be discretized to low-order
controllers.
In many cases we want the reduced-order system to be again

port-Hamiltonian:
• Port-Hamiltonian model reduction preserves passivity.
• Port-Hamiltonian model reduction may (approximately)

preserve other balance laws /conservation laws.
• Physical interpretation of reduced-order model.
• Reduced-order system can replace the high-order

port-Hamiltonian system in a larger context.
Thus there is a need for structure-preserving model reduction of

high-dimensional port-Hamiltonian systems.
Controllability analysis
Consider a linear port-Hamiltonian system, written as
ẋ = F Qx + Bu, F := J − R, J = −J T , R = RT ≥ 0
y = B T Qx, Q = QT ≥ 0
Take linear coordinates x = (x1 , x2 ) such that the upper part of

         
ẋ1 F11 F12 Q11 Q12 x1 B1
  =       +  u
ẋ2 F21 F22 Q21 Q22 x2 B2
   
h i Q11 Q12 x1
y = B1T B2T    
Q21 Q22 x2
is the reachability subspace R.

By invariance of R this implies
F21 Q11 + F22 Q21 = 0

B2 = 0
If follows that the dynamics restricted to R is given as
ẋ1 = (F11 Q11 + F12 Q21 )x1 + B1 u

y = B1T Q11 x1
−1
Now solve for Q21 as Q21 = −F22 F21 Q11 . This yields
−1
ẋ1 = (F11 − F12 F22 F21 )Q11 x1 + B1 u
y = B1T Q11 x1
which is again a port-Hamiltonian system.

Observability analysis
Suppose the system is not observable. Then there exist

coordinates x = (x1 , x2 ) such that the lower part is the
unobservability subspace N . By invariance of N it follows that
F11 Q12 + F12 Q22 = 0

B1T Q12 + B2T Q22 = 0
Then the dynamics on the quotient space X /N is
ẋ1 = (F11 Q11 + F12 Q21 )x1 + B1 u

y = B1T Q11 x1 + B2T Q21 x1
It follow from that F12 = −F11 Q12 Q−1

22 and B T
2 = −B T
1 Q12 Q−1
22 .
Substitution yields
ẋ1 = F11 (Q11 − Q12 Q−1

22 Q21 )x1 + B1 u
y = B1T (Q11 − Q12 Q−1

22 Q21 )x1
which is again a port-Hamiltonian system with Hamiltonian

H̄ = 21 xT1 (Q11 − Q12 Q−1
22 Q21 )x1 .
Remark Note that the Schur complement (Q11 − Q12 Q−1

22 Q21 ) ≥ 0 if
Q ≥ 0.
This suggests two canonical ways for structure-preserving
model reduction.
Conclusions of Part I
• Port-Hamiltonian systems provide a unified framework for
modeling, analysis, and simulation of complex
lumped-parameter multi-physics systems.
• Inclusion of distributed-parameter components.
• Direct lumping of distributed-parameter systems to

finite-dimensional port-Hamiltonian systems.
• Structure-preserving model reduction.
◦ Extensions to thermodynamic systems and chemical reaction

networks.
◦ Further exploration of the network (graph) information.
See www.math.rug.nl/˜arjan, for further info.

The port-Hamiltonian approach to

physical system modeling and control
Part II: Control of Port-Hamiltonian systems
Contents
• Use of passivity for control
• Control by interconnection: set-point stabilization
• The dissipation obstacle
• A state feedback perspective; shaping the Hamiltonian
• New control paradigms
• Model reduction of port-Hamiltonian systems

Use of passivity for control and beyond

• The storage function can be used as Lyapunov function,
implying some sort of stability for the uncontrolled system.
• The standard feedback interconnection of two passive systems
is again passive, with storage function being the sum of the
individual storage functions.
• Passive systems can be asymptotically stabilized by adding
artificial damping. In fact,
d
H ≤ uT y
dt
together with the additional damping u = −y yields
d
H ≤ − k y k2
dt
proving asymptotic stability provided an observability
condition is met.
Example The Euler equations for the motion of a rigid body

revolving about its center of gravity with one input are
I1 ω̇1 = [I2 − I3 ]ω2 ω3 + g1 u

I2 ω̇2 = [I3 − I1 ]ω1 ω3 + g2 u
I3 ω̇3 = [I1 − I2 ]ω1 ω2 + g3 u,
T
Here ω := (ω1 , ω2 , ω3 ) are the angular velocities around the principal
axes of the rigid body, and I1 , I2 , I3 > 0 are the principal moments
of inertia. The system for u = 0 has the origin as an equilibrium
point. Linearization yields the linear system
   
−1
0 0 0 I1 g1
   
A= 0 0 0  −1
B =  I2 g2  .
  
0 0 0 I3−1 g3
Hence the linearization does not say anything about stabilizability.

Stability and asymptotic stabilization by damping injection

Rewrite the system in port-Hamiltonian form by defining the
angular momenta
p1 = I1 ω̇1 , p2 = I2 ω̇2 , p3 = I3 ω̇3
and defining the Hamiltonian H(p) as the total kinetic energy
1 p21 p22 p23

H(p) = ( + + )
2 I1 I2 I3
Then the system can be rewritten as

        
∂H ∂H
ṗ1 0 −p3 p2 g1
     ∂p1    h i  ∂p1 
ṗ2  =  p3  ∂H  + ∂H 
0 −p1  g2  u, y = g1 g2 g3 

     ∂p2   ∂p2 
∂H ∂H
ṗ3 −p2 p1 0 ∂p3 g3 ∂p3
Since Ḣ = 0 and H has a minimum at p = 0 the origin is stable.

Damping injection amounts to the negative output feedback
p1 p2 p3
u = −y = −g1 − g2 − g3 = −g1 ω1 − g2 ω2 − g3 ω3 ,
I1 I2 I3
yielding convergence to the largest invariant set contained in

p1 p2 p3
S := {p ∈ R3 | Ḣ(p) = 0} = {p ∈ R3 | g1 + g2 + g3 = 0}
I1 I2 I3
It can be shown that the largest invariant set contained in S is the

origin p = 0 if and only if
g1 6= 0, g2 6= 0, g3 6= 0,
in which case the origin is rendered asymptotically stable (even,

globally).
Beyond control via passivity: What can we do if the

desired set-point is not a minimum of the storage function ??
Recall the proof of stability of an equilibrium (ω1∗ , 0, 0) 6= (0, 0, 0) of

the Euler equations.
The total energy H = 2Ip 2
1
+ 2I2
p 2 + 2I3
p 2 = 1
2 I 1 ω1
2
+ 1
2 I 2 ω2
2
+ 1 2
2 I3 ω3 has a
1 2 3
minimum at (0, 0, 0). Stability of (ω1∗ , 0, 0) is shown by taking as
Lyapunov function a combination of the total energy K and
another conserved quantity, namely the total angular momentum
C = p21 + p22 + p23 = I12 ω12 + I22 ω22 + I32 ω32
This follows from

 
h i 0 −p3 p2

p1 p2 p3 
 p3 0 =0
−p1 
−p2 p1 0
In general, for any Hamiltonian dynamics
∂H
ẋ = J(x) (x)
∂x
one may search for conserved quantities C , called Casimirs, as
being solutions of
∂T C
(x)J(x) = 0
∂x
d
Then dt C = 0 for every H , and thus also H + C is a candidate
Lyapunov function.
Note that the minimum of H + C may now be different from the
minimum of H .
Control by interconnection: set-point stabilization:

Consider first a lossless Hamiltonian plant system P
ẋ = J(x) ∂H
∂x (x) + g(x)u
y = g T (x) ∂H
∂x (x)
where the desired set-point x∗ is not a minimum of the

Hamiltonian H , while the Hamiltonian dynamics ẋ = J(x) ∂H
∂x (x) does
not possess useful Casimirs.
How to (asymptotically) stabilize x∗ ?
Control by interconnection:
Consider a controller port-Hamiltonian system
ξ˙ = Jc (ξ) ∂H
∂ξ (ξ) + gc (ξ)uc ,
c
ξ ∈ Xc
C:
yc = g T (ξ) ∂H
∂ξ (ξ)
c
via the standard feedback interconnection
u = −yc , uc = y
u y
P
C
yc uc
Then the closed-loop system is the port-Hamiltonian system

    
∂H
ẋ J(x) −g(x)gcT (ξ) (x)
 =   ∂x 
ξ˙ gc (ξ)g T (x) Jc (ξ) ∂Hc
∂ξ (ξ)
with state space X × X c , and total Hamiltonian H(x) + Hc (ξ).
Main idea: design the controller system in such a manner

that the closed-loop system has useful Casimirs C(x, ξ) !
This may lead to a suitable candidate Lyapunov function
V (x, ξ) := H(x) + Hc (ξ) + C(x, ξ)
with Hc to-be-determined.
Thus we look for functions C(x, ξ) satisfying

 
h
T
∂ C T
∂ C
i J(x) −g(x)gcT (ξ)
=0
∂x (x, ξ) ∂ξ (x, ξ)

gc (ξ)g T (x) Jc (ξ)
such that the candidate Lya[unov function
V (x, ξ) := H(x) + Hc (ξ) + C(x, ξ)
has a minimum at (x∗ , ξ ∗ ) for some (or a set of) ξ ∗ ⇒ stability.
Remark: The set of such achievable closed-loop Casimirs C(x, ξ)

can be fully characterized.
Subsequently, one may add extra damping (directly or in the

dynamics of the controller) to achieve asymptotic stability.
Example: the ubiquitous pendulum
Consider the mathematical pendulum with Hamiltonian
1 2
H(q, p) = p + (1 − cos q)
2
actuated by a torque u, with output y = p (angular velocity).
Suppose we wish to stabilize the pendulum at a non-zero angle q ∗
and p∗ = 0.
Apply the nonlinear integral control
ξ˙ = uc = y
∂Hc
−u = yc = ∂ξ (ξ)
which is a port-Hamiltonian controller system with Jc = 0.

Casimirs C(q, p, ξ) are found by solving

 
h i 0 1 0 
∂C ∂C ∂C  =0
∂q ∂p ∂ξ  −1 0 −1
0 1 0
leading to Casimirs C(q, p, ξ) = K(q − ξ), and candidate Lyapunov

functions
1 2
V (q, p, ξ) = p + (1 − cos q) + Hc (ξ) + K(q − ξ)
2
with the functions Hc and K to be determined.
For a local minimum, determine K and Hc such that

Equilibrium assignment
∂K ∗
sin q ∗ + ∂z (q − ξ∗) = 0
∂Hc ∗
− ∂K ∗
∂z (q − ξ ) +
∗
∂ξ (ξ ) =0
Minimum condition
 
∂2K ∗ ∂2K ∗
cos q + ∂z 2 (q − ξ ∗ )
∗
0 − ∂z 2 (q − ξ∗)
 
 0 1 0 >0
 
∂2K ∗ ∂2K ∗ ∂ 2 Hc ∗
− ∂z 2 (q − ξ ∗ ) 0 ∂z 2 (q −ξ )+∗
∂ξ 2 (ξ )
Many possible solutions.

Example: stabilization of the shallow water equations
The dynamics of the water in a canal can be described by

     
h v h h
∂t   +   ∂z  = 0

v g v v
with h(z, t) the height of the water at position z , and v(z, t) its
velocity (and g the gravitational constant).
Recall that by recognizing the total energy
b b
1 2
Z Z
H(h, v) = Hdz = [hv + gh2 ]dz
a a 2
this can be written (similarly to the telegrapher’s equations) as the

port-Hamiltonian system
∂h ∂ ∂H
∂t (z, t) = − ∂z ∂v (h, v)
∂v ∂ ∂H
∂t (z, t) = − ∂z ∂h (h, v)
with the 4 boundary variables
hv|a,b
−( 21 v 2 + gh)|a,b
denoting respectively the mass flow and the Bernoulli function

at the boundary points a, b.
(Note that the product hv · ( 21 v 2 + gh) equals power.)
Suppose we want to control the water level h to a desired

height h∗ .
An obvious ’physical’ controller is to add to one side of the canal,
say the right-end b, an infinite water reservoir of height h∗ ,
corresponding to the port-Hamiltonian ’source’ system
ξ˙ = uc
∂Hc
yc = ∂ξ ( = gh∗ )
with Hamiltonian Hc (ξ) = gh∗ ξ , by the feedback interconnection

1 2
uc = y = h(b)v(b), yc = −u = v (b) + gh(b)
2
This yields a closed-loop port-Hamiltonian system with total
Hamiltonian Z l
1 2
[hv + gh2 ]dz + gh∗ ξ
0 2
By mass balance,
Z b
h(z, t)dz + ξ + c
a
is a Casimir for the closed-loop system. Thus we may take as

Lyapunov function
1
Rb Rb
V (h, v, ξ) := 2 a
2 2
[hv + gh ]dz + gh ξ − gh [ ∗ ∗
a
h(z, t)dz + ξ] + 12 g(b − a)h∗2
1
Rb 2 ∗ 2
= 2 a
[hv + g(h − h ) ]dz
which has a minimum at the desired set-point (h∗ , v ∗ = 0, ξ ∗ )

(with ξ ∗ arbitrary).
Remark Note that the source port-Hamiltonian system is not
passive, since the Hamiltonian Hc (ξ) = gh∗ ξ is not bounded from
below.
An alternative, passive, choice of the Hamiltonian controller system

is to take e.g.
1
Hc (ξ) = gh∗ ξ 2
2
leading to the Lyapunov function
1 b 2 1 ∗
Z
V (h, v, ξ) = ∗ 2
[hv + g(h − h ) ]dz + gh (ξ − 1)2
2 a 2
Asymptotic stability of the equilibrium (h∗ , v ∗ = 0, ξ ∗ = 1) can be
obtained by adding ’damping’, that is, replacing uc = y = h(b)v(b) by
∂V
uc := y − (ξ) = h(b)v(b) − gh∗ (ξ − 1)
∂ξ
leading to (if there is no power flow through the left-end a)
d
V = −gh∗ (ξ − 1)2
dt
(See also the work of Bastin & co-workers for related and more
refined results.)
The dissipation obstacle

Surprisingly, the presence of dissipation R 6= 0 may pose a problem !
C(x) is a Casimir for the Hamiltonian dynamics with dissipation
∂H
ẋ = [J(x) − R(x)] (x), J = J T , R = RT ≥ 0
∂x
iff
∂T C ∂T C ∂C ∂ T C ∂C ∂T C
[J − R] = 0 ⇒ [J − R] =0⇒ R =0⇒ R=0
∂x ∂x ∂x ∂x ∂x ∂x
and thus C is a Casimir iff
∂T C ∂T C
(x)J(x) = 0, (x)R(x) = 0
∂x ∂x
The physical reason for the dissipation obstacle is that by using a
passive controller only equilibria where no energy-dissipation takes
place may be stabilized.
Similarly, if C(x, ξ) is a Casimir for the closed-loop port-Hamiltonian

system then it must satisfy
 
h T T
i R(x) 0
∂ C ∂ C =0
∂x (x, ξ) ∂ξ (x, ξ)

0 Rc (ξ)
implying by semi-positivity of R(x) and Rc (x)

∂T C
∂x (x, ξ)R(x) = 0
∂T C
∂ξ (x, ξ)Rc (ξ) = 0
This is the dissipation obstacle, which implies that one cannot

shape the Lyapunov function in the coordinates that are directly
affected by energy dissipation.
Remark: For shaping the potential energy in mechanical systems
this is not a problem since dissipation enters in the differential
equations for the momenta.
To overcome the dissipation obstacle
Suppose one can find a mapping C : X → Rm , with its (transposed)

Jacobian matrix K T (x) := ∂C
∂x (x) satisfying
[J(x) − R(x)]K(x) + g(x) = 0
Construct now the interconnection and dissipation matrix of an

augmented system as
   
J JK R RK
Jaug :=   , Raug :=  
K T J K T JK K T R K T RK
By construction
[K T (x) | −I]Jaug = [K T (x) | −I]Raug = 0
implying that the components of C are Casimirs for the

Hamiltonian dynamics
   
∂H
ẋ (x)
  = [Jaug − Raug ]  ∂x 
∂Hc
ξ̇ ∂ξ (ξ)
Furthermore, since [J(x) − R(x)]K(x) + g(x) = 0

 
J −R [J − R]K
Jaug − Raug =  
K T [J − R] K T JK − K T RK
 
J −R −g
=  
[g − 2RK]T K T JK − K T RK
Thus the augmented system is a closed-loop system for a different

output !
Port-Hamiltonian systems with feedthrough term take the form
ẋ = [J(x) − R(x)] ∂H
∂x (x) + g(x)u
y = (g(x) + 2P (x))T ∂H
∂x (x) + [M (x) + S(x)]u,
with M skew-symmetric and S symmetric, while

 
R(x) P (x)
 ≥0
P T (x) S(x)
The augmented system is thus the feedback interconnection of the

nonlinear integral controller
ξ˙ = uc
∂Hc
yc = ∂ξ (ξ)
with the plant port-Hamiltonian system with modified output with

feedthrough term
ẋ = [J(x) − R(x)] ∂H
∂x (x) + g(x)u
ymod = [g(x) − 2R(x)K(x)]T ∂H

∂x (x) + [−K T
(x)J(x)K(x) + K T
(x)R(x)K(x)]u
Remark: See Jeltsema, Ortega and Scherpen for further

explorations.
Generalization to feedback interconnection with

state-modulation.
Recall that K T (x) := ∂C
∂x (x) is a solution to
[J(x) − R(x)]K(x) + g(x) = 0. This can be generalized to
[J(x) − R(x)]K(x) + g(x)β(x) = 0
with β(x) an m × m design matrix.

The same scheme as above works if we extend the standard
feedback interconnection u = −yc , uc = y to the state-modulated
feedback
u = −β(x)yc , uc = β T (x)y
Note that K(x) is a solution for some β(x) iff
g ⊥ (x)[J(x) − R(x)]K(x) = 0
(In fact, β(x) := −(g T (x)g(x))−1 g T (x)[J(x) − R(x)]K(x) does the job.)
A state feedback perspective: shaping the

Hamiltonian
Restrict (without much loss of generality) to Casimirs of the form
C(x, ξ) = ξj − Gj (x)
It follows that for all time instants
ξj = Gj (x) + cj , cj ∈ R
Suppose that in this way all control state components ξi can be
expressed as function
ξ = G(x)
of the plant state x. Then the dynamic feedback reduces to a
state feedback, and the Lyapunov function H(x) + Hc (ξ) + C(x, ξ)
reduces to the shaped Hamiltonian
H(x) + Hc (G(x))
A direct state feedback perspective:

Interconnection-Damping Assignment (IDA)-PBC control
A direct way to generate candidate Lyapunov functions Hd is to

look for state feedbacks u = ûIDA (x) such that
∂H ∂Hd
[J(x) − R(x)] (x) + g(x)ûIDA (x) = [Jd (x) − Rd (x)] (x)
∂x ∂x
where Jd and Rd are newly assigned interconnection and damping
structures.
Remark: For mechanical systems IDA-PBC control is equivalent to
the theory of Controlled Lagrangians (Bloch, Leonard, Marsden, .).
For Jd = J and Rd = R (Basic IDA-PBC) this reduces to
∂(Hd − H)
[J(x) − R(x)] (x) = g(x)ûBIDA (x)
∂x
and thus in this case, there exists an ûBIDA (x) if and only if
∂(Hd − H)
g ⊥ (x)[J(x) − R(x)] (x) = 0
∂x
which is the same equation as obtained for stabilization by Casimir
generation with a state-modulated nonlinear integral controller !
Conclusion: Basic IDA-PBC ⇔ State-modulated Control by
Interconnection.
Shifted passivity w.r.t. a controlled equilibrium

(see Jayawardhana, Ortega). Consider a port-Hamiltonian system
∂H
ẋ = F z + gu, z= ∂x (x)
y = gT z
where F = J − R, g are constant, and a controlled equilibrium x0 :
∂H
F z0 + gu0 = 0, z0 = (x0 )
∂x
Define the shifted storage function
∂Hp T
V (x) := Hp (x) − (x − x0 ) (x0 ) − Hp (x0 )
∂x
∂V
Note that ∂x = z − z0 . It follows that
d
dt V = (z − z0 )T ẋ = (z − z0 )T (F z + gu) =
(z − z0 )T F (z − z0 ) + (z − z0 )T g(u − u0 ) + (z − z0 )T (F z0 + gu0 ) ≤ (y − y0 )T (u − u0 )
implying passivity w.r.t. the shifted inputs u − u0 and outputs y − y0 .
Application to switching control

Consider the port-Hamiltonian model of a power-converter
∂Hp
ẋ = F (ρ)z + g(ρ)E + gl u, z= (x), F (ρ) := J(ρ) − R(ρ)
∂x
with vector of Boolean variables ρ ∈ {0, 1}k , Hp (x) the total stored
electromagnetic energy, and output vector y = glT z .
Let x0 be an equilibrium of the averaged model, that is
∂H
F (ρ0 )z0 + g(ρ0 )E + gl u0 = 0, z0 = (x0 )
∂x
for some ρ0 ∈ [0, 1]k and u0 . Then
ẋ = F (ρ)(z − z0 ) + F (ρ)z0 + g(ρ)E + gl u

= F (ρ)(z − z0 ) + [F (ρ) − F (ρ0 )]z0 + [g(ρ) − g(ρ0 )]E + gl (u − u0 )
+F (ρ0 )z0 + g(ρ0 )E + gl u0
= F (ρ)(z − z0 ) + [F (ρ) − F (ρ0 )]z0 + [g(ρ) − g(ρ0 )]E + gl (u − u0 )
For many power converters we know that

Pp
F (ρ) − F (ρ0 ) = i=1 Fi (ρi − ρ0i )
Pp
g(ρ) − g(ρ0 ) = i=1 gi (ρi − ρ0i )
and thus
p
X
ẋ = F (ρ)(z − z0 ) + [Fi z0 + gi E](ρi − ρ0i ) + gl (u − u0 )
i=1
Take as Lyapunov/storage function

∂Hp T
V (x) := Hp (x) − (x − x0 ) (x0 ) − Hp (x0 )
∂x
Then
d ∂Hp ∂Hp T T
dt V (x) = [ ∂x (x) − ∂x (x0 )] ẋ = (z − z0 ) ẋ =
Pp
(z − z0 )T F (ρ)(z − z0 ) + i=1 (z − z0 )T [Fi z0 + gi E](ρi − ρ0i ) + (z − z0 )T gl (u − u0 )
with (z − z0 )T F (ρ)(z − z0 ) ≤ 0.
Thus at any time we can choose ρi ∈ {0, 1} such that

d
V (x) ≤ (z − z0 )T gl (u − u0 )
dt
implying passivity of the switched system with respect to the input
vector u − u0 and output vector y − y0 = glT (z − z0 ). As a
consequence, if the converter is terminated on a static resistive
load then the switched converter is (asymptotically) stable around
x0 . Thus the voltage over the resistive load can be stabilized
around any set-point.
This can be immediately generalized to converters connected to a
load via a transmission line (see Part I).
Note that for linear capacitors and inductors we have
1 T 1
Hp (x) = x Qx, V (x) = (x − x0 )T Q(x − x0 )
2 2
(cf. Buisson & co-workers)
New control paradigms

Example: Energy transfer control
Consider two port-Hamiltonian systems Σi
ẋi = Ji (xi ) ∂H
∂xi (xi ) + gi (xi )ui
i
yi = giT (xi ) ∂H
∂xi (xi ),
i
i = 1, 2
Suppose we want to transfer the energy from the port-Hamiltonian

system Σ1 to the port-Hamiltonian system Σ2 , while keeping the
total energy H1 + H2 constant.
This can be done by using the output feedback

    
u1 0 −y1 y2T y
 =   1
u2 y2 y1T 0 y2
It follows that the closed-loop system is energy-preserving.

However, for the individual energies
d
H1 = −y1T y1 y2T y2 = −||y1 ||2 ||y2 ||2 ≤ 0
dt
implying that H1 is decreasing as long as ||y1 || and ||y2 || are
different from 0. On the other hand,
d
H2 = y2T y2 y1T y1 = ||y2 ||2 ||y1 ||2 ≥ 0
dt
implying that H2 is increasing at the same rate. Has been
successfully applied to energy-efficient path-following control of
mechanical systems (cf. Duindam & Stramigioli).
Impedance control
Consider a system with two (not necessarily distinct) ports
ẋ = [J(x) − R(x)] ∂H
∂x (x) + g(x)u + k(x)f, x ∈ X , u ∈ Rm
y = g T (x) ∂H
∂x (x) u, y ∈ Rm (4)
e = k T (x) ∂H
∂x (x) f, e ∈ Rm
The relation between the f and e variables is called the
’impedance’ of the (f, e)-port. In Impedance Control (Hogan) one
tries to shape this impedance by using the control port
corresponding to u, y .
Typical application: the (f, e)-port corresponds to the end-point of
a robotic manipulator, while the (u, y)-port corresponds to
actuation.
Basic question: what are achievable impedances of the
(f, e)-port ?
Conclusions of Part II
• Beyond passivity by port-Hamiltonian systems theory.
• Control by interconnection and Casimir generation, IDA-PBC

control.
• Allows for ’physical’ interpretation of control strategies.

Suggests new control paradigms for nonlinear systems.
◦ Use of passivity generally yields good robustness, but

performance theory is yet lacking.
THANK YOU !

Vander Schaft

Загружено:

Сведения о документе

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

Vander Schaft

Загружено:

Авторское право:

Доступные форматы

The port-Hamiltonian approach to physical system modeling and control, Namur, November, 2008 1

The port-Hamiltonian approach to

In collaboration with Bernhard Maschke, Romeo Ortega,

Part I : Network Modeling and Analysis

Part I : Network Modeling and Analysis

2. From passive systems to port-Hamiltonian systems

3. Some properties of port-Hamiltonian systems

4. Intro to distributed-parameter port-Hamiltonian systems

5. Mixed lumped- and distributed-parameter port-Hamiltonian

6. Lumping of distributed-parameter physical systems, and

From passive systems to port-Hamiltonian systems

where x ∈ Rn are coordinates for an n-dimensional state space X , is

for all solutions (u(·), x(·), y(·)) and times t1 ≤ t2 .

If H is differentiable then ’passive’ is equivalent to

while in the lossless case ≤ is replaced by =.

Every linear passive system with storage function H(x) = 21 xT Qx,

can be rewritten as a linear port-Hamiltonian system

in which case the storage function H(x) = 21 xT Qx is called the

• Passive linear systems are thus port-Hamiltonian with

• Conversely every port-Hamiltonian system with

Mutatis mutandis ’most’ nonlinear lossless systems can be written

with the Hamiltonian H the total (kinetic + potential) energy.

Similarly, most nonlinear passive systems can be written as a

with R(x) = RT (x) ≥ 0 specifying the energy dissipation

However, in network modeling it is the other way around: one

• Basic conservation laws such as Kirchhoff’s laws.

The matrix R(x) corresponds to the internal energy dissipation

A bit of port-based network modeling

leads to an interconnected system that is again passive, since

The feedback interconnection is a typical example of a

uT1 y1 + uT2 y2 + v1T y1 + v2T y2 = 0

• The Hamiltonian of the resulting port-Hamiltonian system is

• The J - and B -matrix is determined by the transformers,

• The R-matrix is determined by the resistive elements, and the

An k dimensional storage element is determined by a

Such a k - dimensional storage component is written in vector

The elements of x are called energy variables, those of ∂H

Example: The ubiquitous mass-spring-damper system:

• Spring Hamiltonian Hs (q) = 21 kq 2 (potential energy)

(power-conserving since fs es + fm em = uy ) yields the

Addition of the damper

via the extended interconnection

leads to the mass-damper-spring system in port-Hamiltonian

Example: Electro-mechanical systems

Example: LC circuits. Two inductors with magnetic energies

Question: How to write this circuit as a port-Hamiltonian system

Hamiltonian equations for the components of the LC-circuit:

Inductor 1 ϕ̇1 = f1 (voltage)

Inductor 2 ϕ̇2 = f2 (voltage)

All are port-Hamiltonian systems with J = 0 and g = 1.

Kirchhoff’s interconnection laws in f1 , f2 , f3 , e1 , e2 , e3 , f = V, e = I are

with H(ϕ1 , ϕ2 , Q) := H1 (ϕ1 ) + H2 (ϕ2 ) + H3 (Q) total energy.

However, this class of port-Hamiltonian systems is not closed

Figure 1: Capacitors and inductors swapped

Interconnection leads to algebraic constraints between the state

How to model DAEs as port-Hamiltonian systems ?

Definition 1 (Weinstein, Courant, Dorfman) A (constant)

For many systems, especially those with 3-D mechanical

Definition 2 Consider a smooth manifold M . A Dirac structure on

is a Dirac structure as before.

Geometric definition of a port-Hamiltonian system

The dynamical system defined by the DAEs

Particular case is a Dirac structure D(x) ⊂ Tx X × Tx∗ X × F × F ∗

leading (fx = −ẋ, ex = ∂H