2016 American Control Conference (ACC)

Boston Marriott Copley Place

July 6-8, 2016. Boston, MA, USA

Structural Non-Minimum Phase Systems

Andreas Daasch, Matthias Schultalbers and Ferdinand Svaricek

Abstract Recently, there is increasing research interest in systems that are structurally equivalent. For this class the
investigation of structural properties, especially in the complex property under investigation holds numerically for almost
networks domain, i. e. [10]. For analysis of large sparse systems all admissible numerical realizations.
or systems with parameter uncertainty the structural approach
gives an early indication of system properties like controllability There have been a lot of publications considering this
and observability. In this paper the structural approach is topic. Structural systems were investigated probably start-
pursued to determine whether a system is structurally non-
minimum phase. A method for the representation of linear ing with the work [9] considering structural controllability
square MIMO systems as graphs is introduced. In this rep- and [7] considering structural solvability. The results for
resentation it is possible to determine the invariant zeros by structural controllability and observability, finite and infinite
the cycle families in the graph. A theorem for a structural zeros and poles, for linear systems where summarized in
non-minimum phase property is presented considering only the the book [14] and the survey [4]. In the book [13] the
structure of the system. The applicability is demonstrated.
topic of structural properties is discussed in a more math-
I. I NTRODUCTION ematical sense. Structural properties like differential rank,
infinite zeros and invertibility for nonlinear systems where
The structure of a system is described by the mutual described in the book [16]. Current results about stability
dependency of the state variables xi the inputs ui and the in structural systems were presented [1]. This result gave
outputs yi . That means, it is a yes or no criterion if e. g. a the idea to investigate the existence of the structural non-
state variable xi depends on another state variable xk or an minimum phase property and to give sufficient conditions
input uj . for that.
Definition 1. For a set of linear systems x = Al x + Bl u, This paper is organized as follows. In the next section
y = Cl x with the same number of state variables, inputs and we will give a short introduction to the representation of
outputs, their common structure can be defined by matrices, dynamical systems as a graph. In Section III we will show
A ,B and C . An element in these matrices is zero, denoted how to determine the zeros polynomial of the Rosenbrocks
by 0, if the element at the same position is identical zero system matrix using graph-theoretical methods. With that
respectively for all Al ,Bl and Cl and otherwise nonzero, the structural non-minimum phase property is stated in
denoted by , if the element at the same position is nonzero Section IV. An illustrative example of a structural non-
and generally independent respectively for all Al ,Bl and Cl . minimum phase system is given in Section V. In Section VI
Systems have an identical structure, if their states, inputs we draw a conclusion and give an outlook to further work.
and outputs have identical dependencies of each other. Such
systems are called structurally equivalent. II. G RAPH -T HEORETIC R EPRESENTATION

For structural as for numerical realizations, there are The linear system
methods to analyze system properties like controllability, x = Ax + Bu
observability, input-output decoupling, stability, etc. Obvi- (1)
ously, some information about a system is lost if only its y = Cx ,
structure is considered. Therefore we cannot always give is considered where the system state variables are denoted by
a precise answer whether the mentioned system properties x Rn , the input vector by u Rm and the output vector
hold numerically. However, the advantage of the structural by y Rp .
analysis is that if a property holds structurally it holds for First the basic structure of a graph is explained.
almost all systems of the same structure. We will call such
property structural and give the following definition. Definition 3. A graph G(V, E) consists of a set of vertices
v V and a set of edges e E. The graph exposes a structure
Definition 2. (Adopted from Definition 12.2 [14]) A struc- by connecting two vertices, vi and vj , with an edge ei,j .
tural property of a system is a property of a class of In a directed graph, edges establish a connection between
two vertices and assign its direction. This can be extended
Andreas Daasch and Matthias Schultalbers are with the Regelungs-
technisches Exzellenzzentrum, iav GmbH, Gifhorn 38158, Germany to a directed weighted graph G(V, E, W) which additionally
{andreas.daasch,matthias.schultalbers}@iav.de contains a set of weights w W. In the weighted graph
Ferdinand Svaricek is with the Department of Aerospace a value wi,j is assigned to every edge ei,j . Graphically,
Engineering, Chair of Control Engineering, University of the
Federal Armed Forces Munich, Neubiberg 85577, Germany vertices are represented by circles and edges by arrows
[email protected] connecting the cycles and revealing the direction of

978-1-4673-8682-1/$31.00 2016 AACC 3758

connection. If the graph is weighted the labels of the edges with the nonzero entries in the matrices A, B and C. From
reflect the value of their weights. this point of view we can conclude the following.
The dynamical system (1) can be represented as a graph Corollary 1. Systems that are structurally equivalent have
using the following rules. the same unweighted graph G(V, E).
Algorithm 1. The directed (weighted) graph G(V, E) Hence structural properties can by analyzed using the
(G(V, E, W)) of the system (1) consists of m input vertices unweighted graph G(V, E) and a structural property is valid
u1 , . . . , um , n states vertices x1 , . . . , xn and p output vertices for all systems that have the same unweighted graph.
y1 , . . . , yp . The vertices are connected by directed (and
III. G RAPH -T HEORETIC R EPRESENTATION OF THE
weighted) edges, generated by following rules:
I NVARIANT Z EROS P OLYNOMIAL
1) There exists a directed edge from input vertex uj to
state vertex xi if in B the element bi,j in the i-th row In this section a graph-theoretic method for calculating the
and j-th column is nonzero. (Then the weight wi,j of zeros of a linear system is introduced. A similar method is
the edge is given by bi,j .) also applicable for the calculation of the poles, see [14].
2) There exists a directed edge from state vertex xj to We consider the system (1) with the restriction that the
state vertex xi if in A the element ai,j in the i-th row number of outputs equals the number of inputs, i. e. m = p.
and j-th column is nonzero. (Then the weight wi,j of Such a system is called square.
the edge is given by ai,j .) The system (1) can be represented in the frequency domain
3) There exists a directed edge from state vertex xj to by     
X0 sI A -B X(s)
output vertex yi if in C the element ci,j in the i-th row = (4)
Y (s) C 0 U (s)
and j-th column is nonzero. (Then the weight wi,j of | {z }
the edge is given by ci,j .) P (s)

A simple example of a graph-theoretic representation of a where P (s) is called the Rosenbrocks system matrix. The
system is given by Fig. 1. invariant zeros s0i ([11]) of a non-degenerated square system
(1), (i. e. where rank(P (s)) = n + m holds for almost all
u b1,1 x1 a1,1 s), are the roots of the invariant zeros polynomial
det (P (s)) = pm snm + pm+1 snm1 + . . . + pn1 s + pn
Xn

a2,1 a1,2 c1,1

= pk snk .
b2,1
k=m
(5)
It is possible to map graph structures to the determinant
a2,2 x2 c1,2 y and the minors of the adjacency matrix. The structures used
for this are called cycle families.
Fig. 1. Example of a graph-theoretic realization of the generic linear SISO
system of order 2. Definition 5. ([14]) A (directed) path in the graph G(V, E) is
a sequence of edges {ei,j , ej,k , . . .} connecting the vertices
{vi , vj , vk , . . .} in forward direction, wherein every vertex
It is possible to transfer a graph to a matrix representation. is visited only once. If the first vertex and the last vertex
This will be useful for later steps. of a path are identical the sequence is called a cycle. The
Definition 4. ([15]) An adjacency matrix A of a graph length of a cycle is the number of vertices connected by it. A
G(V, E, W) represents the relation between the vertices vi cycle family is the set of disjoint cycles in the graph G(V, E),
V in matrix form. The matrix elements i,j A are i. e. cycles that do not share vertices. The width of a cycle
determined by the edges ei,j E in the following way. family is the number of state vertices it touches.
With that definition the determinant of the adjacency
(
0 if no directed edge between vi and vj exists
i,j = matrix A of a graph G(V, E, W) can be determined by the
wi,j if a directed edge from vi to vj exists.
following theorem.
(2)
Theorem 1. (See Theorem A2.1 and proof in [14]) The
The adjacency matrix A describes the system (1) by
determinant det(A) of an n n matrix A with the elements
y 0 C 0 y ai,ti is defined by the difference
x = 0 A B x . (3) n n
XY XY
u 0 0 0 u det(A) = ai,ti ai,ti (6)
| {z }
A () i=1 () i=1

The unweighted graph G(V, E) of system (1) reveals the where {t1 , t2 , . . . , tn } is a permutation of {1, 2, . . . , n} and
structure of the system, since the edges in the graph coincide () is the sum of even permutations and () the sum of

3759
odd permutations. If A is the adjacency matrix of a graph sum over all i1 under the condition that i1 has lower value
G(V, E, W), each summand a1,t1 a2,t2 . . . an,tn of (6) corre- than i2 . Now extend the determinant of P (s) to
sponds to a cycle family touching all vertices of G(V, E, W).  
The value of the summand is given by the product of the sI A -B
det =
weights of the edges in the cycle family. If the cycle family C 0
X X
consists of d disjoint cycles the sign factor of the summand = snm ...
is given by (-1)nd . im ...<im
X X -A -B i1 i2 ...im n+1n+2...n+m
The polynomial (5) can be constructed by the determinant ...
and minors of (4). The idea is to modify the open loop C 0 i i ...i n+1n+2...n+m
i2 <... i1 <i2 1 2 m
graph G(V, E) so that we can use it to get the graph-theoretic nm1
X X
+s ...
representation of the polynomial.
im+1 im <im+1
Definition 6. The feedback graph Gf (V, E, W) is con- X X -A -B i1 i2 ...im im+1 n+1n+2...n+m (10)
structed by inserting feedback edges in the graph G of system ...
C 0 i i ...i i n+1n+2...n+m
(1). The feedback edges connect the inputs ui to the outputs i2 <... i1 <i2 1 2 m m+1

yi by the feedback law + ...

X X
u = -Iy , (7) +s ...
in1 ...<in1
with I an m m identity matrix. A feedback cycle family is X X -A -B i1 i2 ...im ...in1 n+1n+2...n+m
a cycle family which contains exactly m feedback edges. ...
C 0 i i ...i ...i n+1n+2...n+m
i2 <... i1 <i2 1 2 m n1

With that it is possible to determine the coefficients of (5) 

-A -B

by the next theorem. + det
C 0
Theorem 2. In extension to Theorem 31.1 [14] the coeffi-
with ij [1, n] and j = 1, 2, . . . , m, . . . , n 1. Comparison
cients pk of (5) can be obtained by the following rules. The
of (10) with (5) and reordering yields
coefficient pk is determined by the feedback cycle families
of width k within the feedback graph Gf (V, E, W). X X -A -B i1 ...ik n+1...n+m
pk = ...
1) If no such cycle family of width k exists then the C 0 i ...i n+1...n+m
ik i1 <i2 1 k
coefficient pk equals zero. 1...mi1 ...ik n+m+1...n+2m
2) Otherwise, if only one cycle family of width k exists, X X I 0 0
then pk equals the product of the weights of their edges = ... 0 -A -B
multiplied by (-1)d , where d is the number of cycles ik i1 <i2 0 C 0 1...mi ...i n+m+1...n+2m
1 k
the family consists of. 1...mi1 ...ik n+m+1...n+2m
0 -C 0
3) If there is more than one cycle family of width k, pk
X X
= ... 0 -A -B
equals the sum of the corresponding weight products ik i1 <i2 I 0 0 1...mi ...i n+m+1...n+2m
of the cycle families. X 1 k

= ...
In order to prove this theorem we construct the adjacency ik
matrix of the feedback graph Gf (V, E, W) of system (1) with 1...mi1 ...ik n+m+1...n+2m
the feedback law (7) X 0 C 0
... (1)k+2m 0 A B .
0 C 0 i1 <i2 -I 0 0 1...mi ...i n+m+1...n+2m
1 k
Af = 0 A B . (8) (11)
-I 0 0
By Theorem 1 it is now possible to calculate the determinant
Since by Theorem 1 it is possible to determine the determi-
1...mi1 ...ik n+m+1...n+2m
nant and minors of Af , the proof tries to match them to the

0 C 0
determinant and minors needed to calculate the coefficients 0 A B (12)
of (5). -I 0 0 1...mi ...i n+m+1...n+2m
1 k

Proof. This proof follows the proof of Theorem 31.1 in [14]. from the weights of the feedback cycle families in
We introduce two notations. First Gf (V, E, W) of width k. Since the sign factor is (-1)k+2md
if the cycle families consist of d disjoint cycles, the overall
Mji11ji22 ...j
...ik
k
(9) sign factor becomes
is the determinant of the k k submatrix (minor) formed by (-1)k+2m (-1)k+2md = (-1)d . (13)
the common entries of the row vectors i1 , i2 , .P
. . , ik and the
column vectors j1 , j2 , . . . , jk of M . Second i1 <i2 is the

3760
 In the next section we will show how to use this result to holds.
deduct the structural non-minimum phase property.
With that we can state the proof.
IV. S TRUCTURAL N ON -M INIMUM P HASE P ROPERTY
Proof. (Proof of Theorem 3.) The property is structural since
The expression non-minimum phase is sometimes mis- for the general absence of a coefficient pk in (5) only the
leading. Originally it was introduced by Bode [2] describing absence of cycle families of width k in Gf (V, E) is relevant,
systems that have the identical gain over frequency but differ- regardless of the numerical values of the system.
ent (non-minimal) phase shift. The property was generalized The smallest cycle family maps to the coefficient which
(for nonlinear systems) to the stability of the zero dynamics precedes the highest order of (5), so it defines the number of
which are part of the internal dynamics of a system ([8]). invariant zeros n k. If there is a cycle family of the width
Definition 7. A system that has none or asymptotically k ]k, n] missing in Gf (V, E) the corresponding pk in (5) is
stable zero dynamics is called minimum phase. Otherwise also missing. Thus, by Theorem 4 not all invariant zeros have
the system is called non-minimum phase. strictly negative real parts. According to Lemma 1 this leads
to non-asymptotically stable zero dynamics. By Definition 7
As can be shown by an investigation of the correlation a system which exposes this property is to be called non-
between the zeros and the zero dynamics, this definition is minimum phase.
also applicable to linear systems and is in compliance with
the previous use of the term minimum phase. In some cases we can determine the position of some
It is directly possible to conclude from the position of the invariant zeros by structural investigation.
invariant zeros to the stability of the zero dynamics.
Remark 1. If the largest feedback cycle family has width
Lemma 1. The invariant zeros of a linear square non- ku < n than there are n ku invariant zeros in 0.
degenerated system (1) coincide with the eigenvalues of its
zero dynamics. The inversion of Theorem 3 does not hold due to the
following reason. If all cycle families of width k [k, n]
For the transmission zeros this was shown in [8] and can exist, it is not possible to conclude that the system of interest
be straightforwardly extended to all invariant zeros of square is minimum phase. Since we have not calculated the exact
non-degenerated systems ([3]). values of pk they may be positive or negative. Furthermore
With Definition 2 in mind, we introduce the following there may occur numerical cancellation of a pk because of
definition. the existence of more than one cycle family for a dedicated k.
Definition 8. Systems that are structurally equivalent are Both cases oppose Theorem 4 and hence we can not conclude
called structurally non-minimum phase, if almost all of their to a (structural) minimum phase property of a system.
numerical realizations are non-minimum phase. Remark 2. Theorem 3 is only a sufficient condition for the
Hence we propose the following rules for the graph non-minimum phase property of a numerical realization of
Gf (V, E) of a dynamical system (1) to determine if it is a system.
structurally non-minimum phase. An example of a system, that is structurally not non-
Theorem 3. A square system of order n with m inputs and minimum phase but its numerical realization is non-
outputs is structurally non-minimum phase if for its corre- minimum phase, is depicted in Fig. 2.
sponding feedback graph Gf (V, E) the following properties
hold: -1
1) Let k with m k < n be the width of the feedback
cycle family in Gf (V, E) with the smallest number of y
u -1 x3 1
state vertices, i. e. the smallest feedback cycle family.
2) There is at least one k with k < k n so that Gf (V, E) 8
-4
1
contains no feedback cycle family of width k.
x1 -3 x2
In order to prove this we recall the well-known Theorem
of Viete on polynomial roots ([5]). 1

Theorem 4. A necessary condition that the roots 0i of the Fig. 2. System with positive zeros that is not structurally non-minimum
polynomial equation phase.

c1 1 + c2 2 + . . . + c1 + c = 0 (14)
However there is an exception, where we can conclude to
of order 1 have strictly negative real parts is, that all a structural minimum phase property.
coefficients ci are nonzero and of equal sign, i. e.
Corollary 2. If the feedback graph Gf (V, E) of a system
ci > 0, i = 1, 2, . . . , (15) contains only one feedback cycle family and this feedback

3761
cycle family is of width ku = n than the system is thruster produce, we can define the inputs as
structurally minimum phase. T 2F
u1 = , u2 = cos (16)
This is obvious since in the mentioned case only the m m
coefficient pn in (5) may exist and hence the polynomial and describe system dynamics by
has no roots. This case is equivalent to the property of full
relative degree or flatness. x = u1 sin + u2 cos
Finally, we want to discuss under which circumstances a z =u1 cos + u2 sin g (17)
system that is structurally non-minimum phase, may not be =u2
non-minimum phase in the numerical sense.
with  = tan , = ml J , where J is the inertia respectively
Remark 3. Consider a system that is structurally non- to C and l the distance from C to the thrusters. A inertial
minimum phase, where in the corresponding feedback graph position sensor is placed somewhere on the body of the
Gf (V, E) there are more than one cycle families of the same aircraft. Its position y is given in the coordinates (a, b) w.r.t.
width k. Thus for distinct numerical realizations, numerical to the body fixed frame originated in the center of mass. If
cancellation by calculating the coefficient pk may occur so we use the sensor as output we get
that it vanishes. This will not be crucial for coefficients
that belong to cycle families of width k < k. But if the y1 = x + a cos b sin
(18)
coefficient pk vanishes, the numerical realisation may be y2 = z + a sin + b cos
minimum phase, since the polynomial (5) may fulfill the
necessary condition in Theorem 4 numerically. for its position in inertial coordinates.
So the system has n = 6 states x =
 T  T
x x z z and m = 2 inputs u = u1 u2
V. E XAMPLE  T
and p = 2 outputs y = y1 y2 . Linearization around the
 T
equilibrium (x = 0, u = g 0 ) and considering only the
nonzero entries yields the structural system matrices
F
T

0 0 0 0 0 0 0
0 0 0 0 0 0 0
y


0 0 0 0 0 , B = 0
0
a A = 0

l 0 0 0 0 0
b (19)
0 0 0 0 0 0 0
C 0 0 0 0 0 0 0
 
0 0 0 0
C = .
z 0 0 0 0
mg This structure can be drawn as graph and supplemented to
an unweighted feedback graph GfP V T OL shown in Fig. 4.
x
Applying Theorem 3 on GfP V T OL yields following:
F
The smallest width of a feedback cycle family is k = 4,
e. g. u1 x x y1 u1 , u2 y2
Fig. 3. Planar Vertical Take Off and Landing aircraft. u2 .
The largest width of a feedback cycle family is ku =
6 = n,
As a nonlinear example we will introduce a simple PV- e. g. u2 x x y1 u1 z z
TOL (Planar Vertical Take Off and Landing) aircraft, see y2 u2 .
Fig. 3, that was presented in [6] and whose non-minimum There is no cycle family of width k = 5.
phase behavior was well examined in [12]. The position of Thus we can conclude that the PVTOL is structurally non-
the plane is given by the coordinates x, z of its center of minimum phase. Indeed if we analyze the nonlinear system,
mass C and its angle with respect to an inertial frame. The it is possible to isolate the zero dynamics
motion of the system results from the forces acting on it. T
 
is the lift force generated by the wings due to the relative = g sin + a2 . (20)
forward speed, F represents the force produced in equal  b
amounts by the vertical thruster mounted rotated by fixed It can be shown that this nonlinear differential equation
w.r.t. the vertical axis of the plane and the gravity force is unstable i. e. not asymptotically stable ([12]). Thus the
mg. Considering that we can manipulate the forward speed example system is also non-minimum phase in the nonlinear
by the thrust in this direction, and the forces the vertical sense.

3762
 [4] Jean-Michel Dion, Christian Commault, and Jacob
van der Woude. Generic properties and control of
linear structured systems: a survey. In: Automatica
x x 39.7 (2003), pp. 11251144.
[5] A. Girard. Invention nouvelle En LAlgebre. Blauew,
u1 y1 1629.
[6] John Hauser, Shankar Sastry, and George Meyer.
Nonlinear Control Design for Slightly Non-minimum
Phase Systems: Application to V/STOL Aircraft. In:
Automatica 28.4 (July 1992), pp. 665679.
u2 y2 [7] Masao Iri, Junkichi Tsunekawa, and Keiji Yajima.
The Graphical Techniques Used for a Chemical Pro-
cess Simulator JUSE GIFS. In: Information Pro-
z z
cessing 71 (Proc. IFIP Congr. 71) 2 (1972), pp. 1150
1155.
[8] Alberto Isidori. Nonlinear Control Systems. 3rd.
Springer London, 1995.
[9] Ching-Tai Lin. Structural controllability. In: Au-
Fig. 4. Feedback graph GfP V T OL of the PVTOL. The dotted arrows are
tomatic Control, IEEE Transactions on 19.3 (June
the feedback edges.
1974), pp. 201208.
[10] Yang-Yu Liu, Jean-Jacques Slotine, and Albert-Laszlo
Barabasi. Controllability of complex networks. In:
VI. C ONCLUSION Nature 473.7346 (May 2011), pp. 167173.
We showed how to obtain the common graph-theoretic [11] A. G. J. MacFarlane and N. Karcanias. Poles and
representation of a linear system. For square linear systems zeros of linear multivariable systems : a survey of the
a method was introduced to calculate the systems invariant algebraic, geometric and complex-variable theory. In:
zeros from its graph-theoretic representation. This result was International Journal of Control 24.1 (1976), pp. 33
generalized to structural systems and was then used to state 74.
the structural non-minimum phase property, whose existence [12] P. Martin, S. Devasia, and B. Paden. A different look
was shown. A set of rules were given to determine this at output tracking: control of a VTOL aircraft. In:
property out of a graph-theoretic representation of a system. Decision and Control, 1994., Proceedings of the 33rd
The application of the method was demonstrated. IEEE Conference on. Vol. 3. Dec. 1994, 23762381
Further research concentrates on the structural non- vol.3.
minimum phase property for non-linear systems. Also an [13] K. Murota. Matrices and Matroids for Systems Anal-
extension of the structural non-minimum phase property to ysis. Algorithms and Combinatorics. Springer Berlin
sign non-minimum phase systems is considered. Heidelberg, 2009.
[14] Kurt J. Reinschke. Multivariable Control. A Graph-
R EFERENCES Theoretic Approach. Vol. 108. Springer Berlin Hei-
[1] M.-A. Belabbas. Sparse stable systems. In: Systems delberg, 1988.
& Control Letters 62.10 (Oct. 2013), pp. 981987. [15] K. Thulasiraman and M.N.S. Swamy. Graphs: Theory
[2] Hendrik W. Bode. Network analysis and feedback and Algorithms. Wiley, 2011.
amplifier design. D. van Nostrand New York, 1945. [16] Torsten Wey. Nichtlineare Regelungssysteme. Ein dif-
[3] Andreas Daasch. Zerodynamics of Linear Square Sys- ferentialalgebraischer Ansatz. Vieweg+Teubner Ver-
tems. Tech. rep. University of the Federal Armed lag, 2002.
Forces Munich, Department of Aerospace Engineer-
ing, Chair of Control Engineering, 2016.

3763