Dynamic Zero Finding for Algebraic Equations

T. Mylvaganam, R. Ortega, J. Machado and A. Astolfi

Abstract— In a variety of contexts, for example the H2 /H∞ control problems have been considered. Gen-
solution of differential games and the control of power eral algebraic equations have been considered in [17].
systems, the design of feedback control laws requires the In this paper we consider general nonlinear algebraic
solution of nonlinear algebraic equations: obtaining such
solutions is often not trivial. Motivated by such situa- equations for which we seek a solution. To this end
tions we consider systems of nonlinear algebraic equations a dynamical system, which is such that at any of its
and propose a method for obtaining their solutions. In equilibria components of the state coincide with a solu-
particular, a dynamical system is introduced and (locally) tion of the algebraic equations, is introduced. Feedback
stabilizing control laws which ensure that elements of the control laws which, subject to certain conditions, locally
state converge to a solution of the algebraic equations
are given. Illustrative numerical examples are provided. In asymptotically stabilize an equilibrium of the system
addition it is shown that the proposed method is applicable are provided. The solution of the underlying algebraic
to determine the equilibria of electrical networks with equation can then be determined from the steady-state
constant power loads. values of the trajectories of the closed-loop system.
The main results of this paper can be summarised as
I. I NTRODUCTION follows. The problem of solving a system of algebraic
Often the design of feedback control laws involve equations is posed as a feedback stabilization prob-
obtaining solutions of nonlinear algebraic equations. lem for a dynamical system. A systematic method of
This is, for example, the case when it comes to linear constructing (locally) asymptotically stabilizing control
quadratic differential games and in many aspects of laws is then proposed. The latter contribution draws
power systems analysis and design. In the former, inspiration from [18]. Therein the authors provide a
when considering linear feedback strategies, the so- method for constructing stabilizing control laws for
lution of a game requires solving a system of cou- systems with a particular structure with applications to
pled Algebraic Riccati Equations (AREs) as seen, for power systems (see also [6], [19]). More precisely, the
instance, in [1]–[4]. In the latter, equilibria of a power authors consider dynamical systems with a triangular
system typically satisfy algebraic equations, for which structure and provide stabilizing feedback control laws
the existence of solutions needs to be guaranteed (see, through the introduction of an augmented system. The
for example, [5]) and identified as part of the control results presented in this paper are applicable to a wider
design (see, for instance, [6]–[8]). With these motiva- class of problems.
tions in mind, the focus of this paper is on zero finding The remainder of this paper is organised as follows.
for algebraic equations. The problem considered is introduced in Section II. The
Similarities between classical problems in control the- main result, that is the design of a control law ren-
ory and numerical analysis (in particular iterative nu- dering the equilibria of a closed-loop system (locally)
merical methods) have been exploited both in terms of convergent to a solution of this problem, is provided
analysis and design of iterative algorithms for solving in Section III. Remarks and numerical examples con-
algebraic equations (see, for example, [9]–[14]). Other cerning some particular classes of algebraic equations
approaches for zero finding using a control theoretic are then provided in Sections IV and V. The results
approach appear in [15], [16], where coupled AREs are then illustrated on a power systems problem in
arising in the differential game formulation of mixed Section VI before some concluding remarks are given
in Section VII.
This work has been partially supported by the European Union’s Notation: The sets of natural numbers, real numbers
Horizon 2020 Research and Innovation Programme under grant and complex numbers are denoted by Z, R and C,
agreement No 739551 (KIOS CoE). respectively. The empty set is denoted by ∅. In denotes
T. Mylvaganam is with the Department of Aeronautics,
Imperial College London, London SW7 2AZ, UK, Email: the n×n identity matrix. Given x ∈ Rn we define |x|2 =
t.mylvaganam@imperial.ac.uk. xT x, and for a matrix A, A> denotes its transpose,
R. Ortega and J. Machado are with the Laboratoire des vec(A) denotes its vectorization and adj(A) denotes its
Signaux et Systèmes, Ècole Supèrieure d’Electricitè (SUPELEC),
Gif-sur-Yvette 91192, France, Email: ortega@lss.supelec.fr, adjoint matrix. Given a square matrix A, σ(A) denotes
juan.machado@l2s.centralesupelec.fr. its spectrum. The Kronecker product between two ma-
A. Astolfi is with the Department of Electrical and Electronic trices A and B is denoted by A ⊗ B. All functions are
Engineering, Imperial College London, London SW7 2AZ, UK and
the DICII, Università di Roma “Tor Vergata”, Via del Politecnico 1, assumed sufficiently smooth. The gradient of a function
00133 Roma, Italy, Email: a.astolfi@imperial.ac.uk. f is denoted by ∂f ∂x . The Euclidean basis vectors are
denoted by ei ∈ Rn , i = 1, . . . , n. The relations For instance one could select φ(x) =  tanh(x), for  > 0,
or φ(x) = x. Note that, provided Θ(t) is invertible for
vec(ABC) = (C > ⊗ A)vec(B) ,
all t ≥ 0,
tr(AC) = vec(A> )> vec(C) , zd = ẑd (x, Θ) := Θ−1 [q + φ(x)]. (7)
where A, B, and C denote matrices of appropriate The main contribution of this paper, provided in Sec-
dimensions, are used repeatedly. tion III-C, is the construction of a feedback control law
II. P ROBLEM F ORMULATION which locally asymptotically stabilizes an equilibrium
of the system (4)-(5). A Lyapunov-like design approach
Let f : Rn → Rn denote a function described by1 is adopted to construct the control law. To this end,
2
f (z) = q − P (z)z , (1) define the state vector X := col(x, z, vec(Θ)) ∈ R2n+n
and consider the function
where z ∈ Rn , q ∈ Rn and P : Rn → Rn×n . Assume
there exists z ∗ ∈ Rn such that V (X) = V1 (x) + V2 (X) + V3 (z, vec(Θ)) , (8)
2
q − P (z ∗ )z ∗ = 0 , (2) where
V1 (x) = 21 |x|2 , V2 (X) = 12 |z − zd |2 ,
and
V3 (z, vec(Θ)) = 21 tr (Θ − P )> (Θ − P ) ,


det[P (z ∗ )] 6= 0. (3)
Noting that tr (Θ − P )> (Θ − P ) = i,j (θij − pij )2 , it

P
Suppose, furthermore, that the image of f around z ∗ is
onto a neighborhood of 0. Note that (3) implies that z ∗ follows that V (X) > 0 for all X 6= (0, zd , vec(P )) and
is an isolated solution of (2). V (X) = 0 for X = (0, zd , vec(P )).
Define the dynamical system The time derivatives of each of the component func-
tions of V along the trajectories of the system (4)-(6)
ẋ = q − P (z)z , are given by
ż = u, (4)
Θ̇ = W, V̇1 = −x> φ − x> P (z − zd ) + (zd > ⊗ x> )vec(Θ − P ) ,

where x(t) ∈ Rn , z(t) ∈ Rn and Θ(t) ∈ Rn×n are the V̇2 = (z − zd )> u − (z − zd )> Θ−1 φ̇
states of the system and u(t) ∈ Rn and W (t) ∈ Rn×n +(z − zd )> R(x, Θ)vec(W ) ,
denote control inputs. For this system we consider the
problem of finding a static state feedback V̇3 = vec(Θ − P (z))> vec(W ) − vec(Θ − P )Pz (z)u ,
u = û(x, z, Θ) , W = Ŵ (x, z, Θ) , where we have defined the n × n2 matrix
R(x, Θ) := (q + φ)> Θ−> ⊗ Θ−1 ,


such that the closed-loop system has an asymptotically
stable equilibrium at (0, z ∗ , P (z ∗ )). Thus, the second and the n2 × n matrix
component of the steady-state of any trajectories con-  ∂P ∂P

verging to this equilibrium provide a solution to the Pz (z) := vec( ∂z 1
) ... vec( ∂z n
) , (9)
algebraic equations f (z) = 0. and used the fact that żd = Θ−1 φ̇−Θ−1 Θ̇Θ−1 (q +φ(x)).
Remark 1: The assumption that the image of f is onto It follows that
a neighbourhood of 0 around z ∗ expresses nothing else V̇ = − x> φ
than Brockett’s necessary condition for the feedback
+ (z − zd )> u + Rvec(W ) − P > x − Θ−1 φ̇
 
stabilization, via smooth feedback, of system (4), see
+ vec(Θ − P )> − Pz u + vec(W ) + (zd > ⊗ x> )> .
 
[20]. N
(10)
III. M AIN R ESULT 2 2

A. Design of the control law Now, define the matrix M ∈ R(n+n )×(n+n ) as
 
To provide a solution to the aforementioned problem In R(x, Θ)
M (X) := .
we introduce a variable zd ∈ Rn , which is implicitly −Pz (z) In2
defined by the relation It is clear from (10) that selecting the control u and
vec(W ) to satisfy
q − Θzd = −φ(x) , (5)  
u
where φ : Rn → Rn is selected such that M (X)
vec(W )
(11)
−kz (z − zd ) + P > x + Θ−1 φ̇
 
φ(0) = 0 ,
(6) = ,
x> φ(x) > 0, ∀x 6= 0 . −kΘ vec(Θ − P ) − (zd > ⊗ x> )>
1 We recall that any differentiable function f can be written in the 2 To simplify the notation, when clear from the context, we omit
form (1), but that the selection of P may be non-unique. the arguments of all mappings.
where kz > 0 and kΘ > 0 are tuning parameters, yields Thus, Vk̄ denotes the largest level set of V contained in
the set Ωµ .
V̇ = − x> φ − kz |z − zd |2 − kΘ tr (Θ − P )> (Θ − P ) .

(12) Proposition 1: Consider the dynamical system (4)-(5)

and a function φ(x) satisfying (6). Suppose the condi-
Note that if we select φ(x) such that x> φ(x) ≥ kx |x|2 , tion (3) and Assumption 1 hold and let X(0) ∈ Vk̄ .
with kx > 0, we have V̇ ≤ − min{kx , kz , kΘ }V. Then the dynamical system (4)-(5) in closed-loop with
It is clear that the derivations above rely on the the control law u and W satisfying (11) is such that the
assumptions that Θ and M (X) are invertible matrices— equilibrium Xe given in (13) is locally asymptotically
an issue that is addressed in the following two subsec- stable and, in particular, the following implication is
tions. true
Remark 2: The assumption that Θ is invertible for all [X(0) ∈ Ωµ ∩ N ⇒ lim z(t) = z ∗ ].
t→∞
t ≥ 0 can be interpreted as an observability condition Remark 3: In [18] a method for designing stabilizing
on the output zd implicitly defined in (5). N controllers for a class of triangular systems has been
proposed. The method therein considers systems simi-
B. A preliminary lemma
lar to (4) in which P (z) is diagonal. The result provided
In this subsection we give a verifiable condition for in Proposition 1 is applicable to such systems, as well
the invertibility of M (X). Towards this end, define the as a wider class of nonlinear systems. N
desired closed-loop equilibrium
Two special cases, namely systems of linear equa-
∗ ∗ 2n+n2 tions and systems of quadratic equations are consid-
Xe := col(0, z , vec(P (z ))) ∈ R , (13)
ered in the following sections. Whereas in the former
where z ∗ ∈ Rn is a solution of (2), the sets case P (z) is a constant matrix, in the latter P (z) is linear
2 in z. Numerical examples are provided for both cases,
Ωµ := {X ∈ R2n+n | |X − Xe | ≤ µ} (14)
which provide useful insights into the results expressed
2n+n2
N := {X ∈ R | det(Θ) 6= 0} (15) by Proposition 1.
and the mapping ∆δ : N → Rn×n IV. S YSTEMS OF L INEAR E QUATIONS
1 Consider the case in which f defined in (1) is linear.
∆δ (X) := In + [e> adj(Θ)(q + δ)adj(Θ)
(det(Θ))2 1 (16) In this case, P (z) = P̄ is a constant matrix that, in view
. . . e>
n adj(Θ)(q + δ)adj(Θ)]Pz (z) ,
of (3), is full rank. We seek to obtain a solution of the
system of linear equations
where δ ∈ Rn is a parameter.
The following assumption is made to ensure the q − P̄ z ∗ = 0 . (17)
(local) existence of u and W satisfying (11).
Recall the notation introduced previously, namely X =
Assumption 1: There exists  > 0 and µ > 0 such that col(x, z, vec(Θ)) and Xe = col(0, z ∗ , vec(P (z ∗ ))).
det[∆δ (X)] 6= 0, ∀ X ∈ Ωµ , ∀ |δ| ≤ . Corollary 1: Consider the system (4)-(5) with P (z) =
P̄ and a function φ(x) satisfying (6). Let Θ(0) be such
Assumption 1 and the following lemma enables the that
design of the locally asymptotically stabilizing control det(Θ(0)) det(P̄ ) > 0 . (18)
laws provided in Section III-C (and in Section IV and
Section V for linear and quadratic problems, respec- Then there exists u and W satisfying (11) and the
tively). system (4)-(5) in closed-loop with u and W is such
Lemma 1: Assumption 1 implies that det[M (X)] 6= 0 that Xe is an asymptotically stable equilibrium and
for all X ∈ Ωµ . lim z(t) = z ∗ .
t→∞
Remark 4: Corollary 1 establishes that for the case in
which P (z) is a constant matrix, the Assumption 1 in
C. A stabilizing controller Proposition 1 is trivially satisfied provided det(Θ(0))
To begin with we make the important observation and det(P (z ∗ )) have the same sign, see (18). N
that, in view of condition (3), the equilibrium Xe is Remark 5: In Proposition 1 it is assumed that P̄ is full
in the set N . This allows to design a controller which rank, such that (17) has a unique solution. If instead
locally stabilizes the equilibrium Xe as established in P̄ ∈ Rm×n , with m < n, (17) is an underdetermined sys-
the following proposition in which tem and may have no solutions or infinitely many solu-
2
Vk := {X ∈ R2n+n | V (X) ≤ k} , tions. In the latter case Corollary 1 may be generalised
to achieve local asymptotic stability for an equilibrium
denotes the level sets of the function V (X) defined in corresponding to a solution of the underdetermined
(8), with zd given in (7), and k̄ = max{k : Vk ⊂ Ωµ }. system. N
 TABLE I
1
T HE SELECTION OF INITIAL CONDITIONS CONSIDERED IN THE
0.5 SIMULATIONS .

x 0 Θ(0) zd (0)
-0.5 I4 Θ0,1 zd0,4
I3 Θ0,1 zd0,3
-1 I2 Θ0,1 zd0,2
-1.5
I−4 Θ0,2 zd0,−4
0 1 2 3 4 5 I−3 Θ0,2 zd0,−3
time (s)
I−2 Θ0,2 zd0,−3

Fig. 1. Time histories of the component of x.

1 Corollary 2: Consider the case in which (1) is a sys-

tem of quadratic equations. Let z ∗ denote a solution
z 0
of (1) such that the condition (3) is satisfied. Suppose
-1
¯δ ∩ 0 = ∅,


0 1 2 3 4 5 σ ∆ (19)
time (s)
1
for all |δ| < , and for some  > 0. It follows that
zd 0
Assumption 1 is satisfied.
-1
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 Remark 6: For the case in which (1) is quadratic, since
time (s)
Pz is constant, Assumption 1 is a condition on Θ(0)
(and P (z∗)). N
Fig. 2. Time histories of the components of z (top) and zd (bottom).
A. A numerical example
Consider the quadratic equation
A. A numerical example
−z1 z2
   
A system of linear equations, as studied in Section IV, q1
f (z) = − = 0, (20)
is considered and Corollary 1 is used to obtain the q2 z12 + z22
solution of (17). | {z } | {z }
q P (z)z
Considering the case in which n = 20, the parameters
P̄ and q are randomly generated in MATLAB3 such that q1 ∈ R
where  and q2 ∈ R. Consider the selection P (z) =
P̄ is full rank. The system (4)-(5) in closed-loop with −z2 0

the feedback control law satisfying (11) is simulated , and consider the case in which q1 = 8
z1 z2
with4 φ(x) = x, kz = 1 and kΘ = 1 and with the and q2 = 16. In this case the system of equations (20)
initial condition z(0) = 0, zd (0) = 0, Θ(0) = In has two solutions.
and x(0) = q. Note that the initial condition is such Following the main result given in Proposition 1, the
that det(Θ(0)) det(P̄ ) > 0 as required in (18) and system (4)-(5) in closed-loop with the feedback given by
Corollary 1. The time histories of the components of (11) is simulated for different selections of parameters.
x are shown in Figure 1 and the time histories of the Let Θ0,1 = diag{−1, 1} and Θ0,2 = diag{1, −1} and let
components of z (top) and zd (bottom) are shown in zd0,i = [0, i], for i ∈ Z. The initial condition Θ(0) is
Figure 2. varied between Θ0,1 and Θ0,2 , whereas zd (0) = zd0,i ,
for i = −4, . . . , 4. The considered initial conditions are
V. S YSTEMS OF Q UADRATIC E QUATIONS
referred to according to Table I. The remainder of the
Consider now the case in which (1) is quadratic, i.e. parameters are selected as follows5 : φ = x, z(0) = 0,
the case in which P (z) defined in (9) is linear in z. In x(0) = −q − Θ(0)zd (0) and kz = kΘ = 200.
this case Pz (z) = P̄z is a constant matrix. The trajectories of z are shown in Figures 3 for the
Let initial conditions I4 (solid, black line), I3 (solid, dark
¯ δ = In + 1 grey line), I2 (solid, light gray line), I−4 (dashed, black
∆ [e> adj(P (z ∗ ))(q + δ)adj(P (z ∗ ))
(det(P (z ∗ )))2 1 line), I−3 (dashed, dark grey line) and I−2 (dashed,
. . . e> ∗ ∗
n adj(P (z ))(q + δ)adj(P (z ))]P̄z ,
light grey line). The solid, square markers denote the
initial conditions, whereas the solutions of (2) to which
3 The system of equations was randomly generated in MATLAB
the trajectories converge are denoted by the diamond-
and can be replicated using the MATLAB code ”n=20; rng(’default’);
shaped markers.
SYS = rss(n); P = -(SYS.A - eye(n)); q = SYS.B;”
4 The selection of the function φ has been made in accordance 5 As in Section IV-A, the selection of the function φ has been made
with (6). in accordance with (6).
 3 where Pi ∈ R and Qi ∈ R are the active and reactive
2 power at port i, respectively, the CPLs constraint the
1
network via
z2 0 V̄i Iic = −Si , (23)
-1
where (·)c denotes the complex conjugate operator.
-2 Obviously, a necessary and sufficient condition for
-3 the existence of a sinusoidal steady–state (at a given
-4 -3 -2 -1 0 1 2 3 4
z1 frequency ω0 ) is that the complex equations (21), (22)
and (23) have a solution. It has been shown in [5] that
Fig. 3. Trajectories of z for the initial conditions I4 (solid, black line), this set of equations can be written in the form
I3 (solid, dark grey line), I2 (solid, light gray line), I−4 (dashed,
black line), I−3 (dashed, dark grey line) and I−2 (dashed, light grey q − P (z)z = 0 , (24)
line).
where q := col(P, Q), z := col(Re{I}, Im{I}) and
5
P (z) =
Θ 0
diag{e> diag{e>
 
− i Ḡz + K̄i } m+i Ḡz + K̄m+i } ,
-5
0 0.5 1 1.5 2 2.5 3 3.5 4
diag{e>m+i Ḡz + K̄m+i } −diag{e>
i Ḡz + K̄i }
5
with K̄ = col(Re{K}, Im{K}) and
Θ 0  
Re{G} −Im{G}
Ḡ = . (25)
-5 Im{G} Re{G}
0 0.5 1 1.5 2 2.5 3 3.5 4
time (s)
Hence, the problem of finding the equilibrium current I
can be solve using the method presented in this paper.
Fig. 4. Time histories of each component of Θ corresponding to the
intitial conditions I4 (top) and I−4 (bottom).

r L
  i
Θ11 Θ12
Let Θ = . The time histories of Θ11 icpl +
Θ21 Θ22 +
(solid, black lines), Θ12 (dashed, black lines), Θ21
vg C v rc CPL
(dashed, grey lines) and Θ22 (solid, grey lines) are
−
shown in Figure 4 for I4 (top) and I−4 (bottom). −

VI. A PPLICATION TO E LECTRICAL N ETWORKS WITH

C ONSTANT P OWER L OADS Fig. 5. AC Linear RLC circuit with a CPL.
In this section we apply the developed numerical
method to compute the steady state of an m-port linear Consider as an example the circuit shown in Fig. 5
time-invariant AC electrical network with Constant with the voltage source given by
Power Loads (CPLs) working in sinusoidal steady-state √
at a frequency ω0 ∈ R+ , see [5] and references therein vg (t) = 2Vg cos(ω0 t),
for additional details on this problem.
The description of this regime in the frequency do- 300
main is given by the relation
200

V̄ (jω0 ) = G(jω0 )I(jω0 ) + K(jω0 ), (21) 100

0
z
where6 V̄ ∈ Cm and I ∈ Cm are the vectors of -100
generalized Fourier transforms of the port voltages and -200
currents, respectively, and K ∈ Cm captures the effect -300
of the external (current or voltage) AC sources, all -400
evaluated at the frequency ω0 . 0 0.02 0.04 0.06 0.08 0.1
time (s)
Defining the apparent power at the CPLs as
Fig. 6. Time histories of the first (solid line) and second (dashed
Si := Pi + jQi , i ∈ {1, . . . m}, (22) line) components of z, which converge to the steady-state current
corresponding to the network of Fig. 5.
6 To simplify the notation the argument jω0 is omitted in the sequel.
and the CPL extracting power from the network ac- systematically ensure convergence to solutions with
cording to (22). In this case G and K are given by particular properties. It is also of interest to consider
r + jLw0 applications to problems arising in large-scale power
G=   systems control.
r L
(1 + rc − LCw02 ) + j rC + rc w0
R EFERENCES
Vg
K=   . [1] T. Basar and G. Olsder, Dynamic Noncooperative Game Theory.
r L
(1 + rc − LCw02 ) + j rC + rc w0 Academic Press, 1982.
[2] A. W. Starr and Y. C. Ho, “Nonzero-sum differential games,”
Journal of Optimization Theory and Applications, vol. 3, pp. 184–
Using the circuit parameters from Table VI, the above 206, 1969.
expressions result in G = 0.0412 + j0.0238 , K = [3] ——, “Further properties of nonzero-sum differential games,”
24.3592 − j0.6219. Journal of Optimization Theory and Applications, vol. 3, pp. 207–
219, 1969.
TABLE II [4] T. Mylvaganam, M. Sassano, and A. Astolfi, “Constructive
-Nash equilibria for nonzero-sum differential games,” IEEE
PARAMETERS FOR THE CIRCUIT IN F IG . 5 Transactions on Automatic Control, vol. 60, no. 4, pp. 950–965,
2015.
r = 0.04 Ω L = 78.0 µH C = 2.0 mF Vg = 24 V [5] J. E. Machado, R. Griñó, N. Barabanov, R. Ortega, and B. Polyak,
rc = 5 kΩ P = 100 W Q = −50 var ω0 = 2π50 rad/s “On existence of equilibria of multi-port linear ac networks with
constant-power loads,” IEEE Transactions on Circuits and Systems
I: Regular Papers, vol. 64, no. 10, pp. 2772–2782, 2017.
[6] D. Casagrande, A. Astolfi, D. Langarica, and R. Ortega, “Solu-
The existence of steady-states is guaranteed by [5, tion to the multi-machine transient stability problem and simu-
Proposition 3]. However, no indication is given on lated validation in realistic scenarios,” IET Generation, Transmis-
how to find these. In what follows, the main result of sion & Distribution, vol. 8, no. 8, pp. 1392–1405, 2014.
[7] J. Schiffer, E. Fridman, R. Ortega, and J. Raisch, “Stability of
this paper, i.e. Proposition 1, is used to determine the a class of delayed port-hamiltonian systems with application
steady-state current I of the circuit shown in Figure 5 to microgrids with distributed rotational and electronic genera-
with the circuit parameters given in Table VI. tion,” Automatica, vol. 74, pp. 71–79, 2016.
[8] S. Trip, M. Bürger, and C. D. Persis, “An internal model ap-
Consider the dynamical system (4)-(5) with P (z) proach to (optimal) frequency regulation in power grids with
given in (24). The system (4)-(5) in closed-loop with time-varying voltages,” Automatica, vol. 64, pp. 240 – 253, 2016.
the control law satisfying (11), with φ(x) = 100x, [9] A. Bhaya and E. Kaszkurewicz, “A control-theoretic approach
to the design of zero finding numerical methods,” IEEE Trans-
is simulated with the following selection of control actions on Automatic Control, vol. 52, no. 6, pp. 1014–1026, 2007.
parameters and initial conditions: kz = 10, kΘ = 600, [10] J. M. Ortega, “Stability of difference equations and convergence
> of iterative processes,” SIAM Journal on Numerical Analysis,
z(0) = [−1, 0] , Θ(0) = 10I, zd (0) = 0 and x(0) =
vol. 10, no. 2, pp. 268–282, 1973.
−q − Θ(0)zd (0) . The resulting time histories of the [11] K. Kashima and Y. Yamamoto, “System theory for numerical
first (solid line) and second (dashed line) component analysis,” Automatica, vol. 43, no. 7, pp. 1156 – 1164, 2007.
of the state z are shown in Figure 6. The trajectories [12] R. D. Harbor, “A control theory perspective of iterative nu-
merical methods,” in IEEE Proceedings of the 21st Southeastern
converge to the solution z ∗ = col(Re{I}, Im{I}) = Symposium on System Theory, 1989, pp. 71–73.
col(−436.3895, 271.827). [13] T. Han and Y. Han, “Solving large scale nonlinear equations by
a new ODE numerical integration method,” in Applied Mathe-
VII. C ONCLUSION matics, vol. 1, no. 3, 2010, pp. 222–229.
[14] B. Polyak and P. Shcherbakov, “Lyapunov functions: An op-
We consider the problem of obtaining solutions for timization theory perspective,” in Proceedings of the 20th IFAC
general, nonlinear, algebraic equations through the World Congress, vol. 50, no. 1, 2017, pp. 7456 – 7461.
[15] T. Mylvaganam and A. Astolfi, “A Nash game approach to
introduction of a dynamical system. The dynamical mixed H2 /H∞ control for input-affine nonlinear systems,” in
system is constructed such that it has an equilibrium Proceedings of the 10th IFAC Symposium on Nonlinear Control
when elements of its state coincides with the solution Systems, 2016.
[16] ——, “Dynamic algorithms for solving coupled algebraic riccati
of the algebraic equation. The problem is then solved equations arising in mixed H2 /H∞ control for scalar linear
through the design of a feedback control law which systems,” in Proceedings of the 55th IEEE Conference on Decision
(locally) asymptotically stabilizes an equilibrium of and Control (CDC), 2016, pp. 652–657.
[17] ——, “Zero finding via feedback stabilisation,” in Proceedings of
the system. The proposed method is demonstrated on the 20th IFAC World Congress, vol. 50, no. 1, 2017, pp. 8133 –
numerical examples and on the problem of determining 8138.
the steady-state current of an AC power system. [18] D. Casagrande, A. Astolfi, and R. Ortega, “Global stabilization
of non-globally linearizable triangular systems: Application to
Directions for future research include further analysis transient stability of power systems,” in Proceedings of the 50th
of the proposed method. In particular, the performance IEEE Conference on Decision and Control and European Control
of the method for specific classes of problems includ- Conference, 2011, pp. 331–336.
[19] D. Casagrande, A. Astolfi, R. Ortega, and D. Langarica, “A
ing matrix equations, such as coupled AREs arising solution to the problem of transient stability of multimachine
in the context of linear quadratic differential games, power systems,” in Proceedings of the 51st IEEE Conference on
will be studied. When considering coupled AREs one Decision and Control (CDC), 2012, pp. 1703–1708.
[20] W. Wong and R. W. Brockett, “Systems with finite communica-
often seeks a solution with certain properties (such tion bandwidth constraints. I. state estimation problems,” IEEE
as positive definiteness). It is therefore of interest to Transactions on Automatic Control, vol. 42, no. 9, pp. 1294–1299,
further extend the results presented in this paper to 1997.