Positively Invariant Sets for ODEs and Numerical Integration
Peter Giesl
1 a
, Sigurdur Hafstein
2 b
and Iman Mehrabinezhad
2 c
1
Department of Mathematics, University of Sussex, Falmer, BN1 9QH, U.K.
2
Science Institute, University of Iceland, Dunhagi 3, 107 Reykjav
´
ık, Iceland
Keywords:
Positively Invariant Sets, Ordinary Differential Equations, Numerical Integration.
Abstract:
We show that for an ordinary differential equation (ODE) with an exponentially stable equilibrium and any
compact subset of its basin of attraction, we can find a larger compact set that is positively invariant for both
the dynamics of the system and a numerical method to approximate its solution trajectories. We establish this
for both one-step numerical integrators and multi-step integrators using sufficiently small time-steps. Further,
we show how to localize such sets using continuously differentiable Lyapunov-like functions and numerically
computed continuous, piecewise affine (CPA) Lyapunov-like functions.
1 INTRODUCTION
In this paper we study the ordinary differential equa-
tion (ODE)
˙
x = f(x), f C
s
(R
n
;R
n
), s 1, (1)
and numerical methods to approximate its solution for
given initial values. The solution x(t) to the initial
value problem (1) with x(0) = ξ
ξ
ξ is denoted by φ
φ
φ(t,ξ
ξ
ξ).
Our ultimate goal is to characterize the long term
behavior of system (1) by computing Lyapunov func-
tions and contraction metrics. As many real-world
systems are modelled by (1), such a characterization
has numerous practical uses in engineering and sci-
ence. In this paper we assume that (1) possesses an
exponentially stable equilibrium at x
0
R
n
, i.e. the
Jacobian Df(x
0
) R
n×n
of f at x
0
is Hurwitz (the real
part of all its eigenvalues is negative). We denote the
equilibrium’s basin of attraction by
A(x
0
) := {x R
n
: lim
t
φ
φ
φ(t,x) = x
0
}.
The Lyapunov stability theory is a generalization of
the concept of dissipative energy in physics. It is
the centerpiece of practical and theoretical stabil-
ity analysis and is treated in various detail in virtu-
ally all textbooks and monographs on ODE systems,
cf. e.g. (Hahn, 1967; Yoshizawa, 1966; Zubov, 1964;
Khalil, 2002; Sastry, 1999; Vidyasagar, 2002).
a
https://orcid.org/0000-0003-1421-6980
b
https://orcid.org/0000-0003-0073-2765
c
https://orcid.org/0000-0002-6346-9901
For a general ODE there is no obvious candi-
date for a Lyapunov function and there are no an-
alytical methods to compute a Lyapunov function.
Hence, various methods for the numerical compu-
tation of Lyapunov functions have been developed.
To name a few, in (Valmorbida and Anderson, 2017;
Vannelli and Vidyasagar, 1985) the computation of
rational Lyapunov functions was studied, in (An-
derson and Papachristodoulou, 2015; Parrilo, 2000)
sum-of-squared (SOS) polynomial Lyapunov func-
tions were computed using semi-definite optimization
(SOS method), see also (Ratschan and She, 2010;
Kamyar and Peet, 2015) for other approaches using
polynomials, and in (Giesl, 2007) a Zubov type PDE
was numerically solved using radial basis functions
(RBF method). For an overview of more methods see,
e.g., the review paper (Giesl and Hafstein, 2015b).
In (Julian et al., 1999; Marin
´
osson, 2002) linear pro-
gramming was used to compute continuous and piece-
wise affine (CPA) Lyapunov functions; this approach
is referred to as the CPA method. In the CPA method
the domain, where the Lyapunov function is to be
computed, is triangulated, i.e. subdivided into sim-
plices, and a feasibility problem is derived, such that
its solution can be used to define a CPA Lyapunov
function for the system. In (Giesl and Hafstein, 2014;
Hafstein, 2004; Hafstein, 2005) it was shown that the
CPA method always succeeds in computing a Lya-
punov function for an ODE with an asymptotically
stable equilibrium, if the simplices in the triangulation
are sufficiently small and non-degenerate in a certain
sense.
44
Giesl, P., Hafstein, S. and Mehrabinezhad, I.
Positively Invariant Sets for ODEs and Numerical Integration.
DOI: 10.5220/0012189700003543
In Proceedings of the 20th International Conference on Informatics in Control, Automation and Robotics (ICINCO 2023) - Volume 1, pages 44-53
ISBN: 978-989-758-670-5; ISSN: 2184-2809
Copyright © 2023 by SCITEPRESS Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)
In (Giesl and Hafstein, 2015a), the CPA and the
RBF method were combined to deliver a method that
is as fast as the RBF method and delivers a verified
Lyapunov function as the CPA method. This was
achieved by solving a system of linear equations in the
RBF method rather than a linear optimization prob-
lem as in the CPA method. At the same time, the
obtained function is verified to be a Lyapunov func-
tion by checking that it satisfies the constraints of the
feasibility problem. Further, the authors proved that
this approach is constructive and one is always able
to compute a verified Lyapunov function in any com-
pact subset of an exponentially stable equilibrium’s
basin of attraction. A similar approach uses numer-
ical integration of solution trajectories of the ODE
to generate values for the variables of the feasibility
problem of the CPA method and then verifies the con-
straints, see (Bj
¨
ornsson et al., 2014; Bj
¨
ornsson et al.,
2015; Bj
¨
ornsson and Hafstein, 2017; Doban, 2016;
Doban and Lazar, 2016; Hafstein et al., 2014a; Haf-
stein et al., 2014b; Hafstein et al., 2015; Hafstein and
Valfells, 2017; Hafstein and Valfells, 2019; Li et al.,
2015) and also (Gudmundsson and Hafstein, 2015;
Hafstein, 2019a) for more implementation oriented
papers. This technique works well in practice and
in (Giesl and Hafstein, 2023) it is proved that it al-
ways works, assuming that one can find a compact set
that is positively invariant for both the ODE and for a
numerical scheme to approximate its solution trajec-
tories. The main contribution of this paper is to prove
the existence of such a positively invariant set.
The results in this paper also make an important
contribution in the context of numerical methods for
contraction metrics, see (Giesl et al., 2023c), where
the results of this paper are used to derive a uniform
error estimate on compact sets. Contraction metrics
are Riemannian metrics such that the corresponding
distance of adjacent solution trajectories decreases ex-
ponentially over time. They have received consider-
able attention in the literature (Lewis, 1949; Lewis,
1951; Demidovi
ˇ
c, 1961; Krasovski
˘
i, 1963; Borg,
1960; Hartman, 1961; Hartman, 1964; Lohmiller and
Slotine, 1998; Aminzare and Sontag, 2014; Simpson-
Porco and Bullo, 2014; Forni and Sepulchre, 2014;
Giesl, 2015). The analytical computation of a con-
traction metric for an ODE is notoriously difficult,
even more difficult than the computation of a Lya-
punov function, as it requires the computation of a
matrix-valued function. Numerical methods for the
construction of contraction metrics include (Aylward
et al., 2008; Giesl and Hafstein, 2013; Giesl, 2019;
Giesl et al., 2023a), see also the recent review (Giesl
et al., 2023b).
Let us give an overview of the paper: In Section
2 we recall some facts about numerical integration
methods of ODEs and prove a theorem about approxi-
mations of solutions with multi-step methods. In Sec-
tion 3 we establish the existence of positively invari-
ant sets, both for the dynamics of the system (1) and a
numerical integration scheme to approximate the so-
lution trajectories in the basin of attraction of an expo-
nentially stable equilibrium; the main result is Theo-
rem 3.5. Such positively invariant sets are very useful,
in fact necessary, to prove that Lyapunov functions
and contraction metrics can be approximated arbitrar-
ily close on compact subsets of basins of attraction,
using numerical integration with subsequent numeri-
cal quadrature. We prove our results using the fourth-
order Adams-Bashforth (AB4) multi-step scheme ini-
tialized with fourth-order Runge-Kutta (RK4), but we
discuss how the results can be extended to AB-RK
numerical schemes of arbitrary order. Finally we con-
clude the paper in Section 4.
Notation: We define N
0
:= {0, 1, 2,...} as the set
of the natural numbers and N
+
:= N
0
\{0} as the
set of the positive natural numbers. We denote the
usual p-norms on R
n
and the corresponding induced
matrix norms by ·
p
, 1 p < . For both vec-
tors in R
n
and matrices in R
n×n
we write ·
max
for the maximum absolute value norm, i.e. x
max
:=
max
i=1,2,...,n
|x
i
| for a vector x R
n×n
and A
max
:=
max
i, j=1,2,...,n
|a
i j
| for a matrix A =
a
i j
R
n×n
.
Apart from the usual equivalence estimates for the p-
norms on R
n
, recall the norm equivalence A
max
A
2
nA
max
for a matrix A R
n×n
and that
·
max
is not sub-multiplicative, but Ab
max
nA
max
b
max
for b := B R
n×n
or b := b R
n
.
We denote the closure of a set U R
n
by U and its
boundary by U.
2 NUMERICAL INTEGRATION
METHODS
For computing Lyapunov functions or contraction
metrics for the ODE (1), initial value problems can
be solved numerically for all vertices of a triangu-
lation of a relevant compact subset of R
n
, as dis-
cussed in the last section. As the number of vertices
can be very high, it is advantageous to use multi-step
methods rather than single-step methods, as these are
considerably faster for the same degree of precision.
Additionally, the examples in (Hafstein, 2019b) indi-
cate that the corrector step in the Adams-Bashforth-
Moulton predictor-corrector methods does not deliver
better results than the Adams-Bashforth method with-
out the corrector step. Therefore we will concentrate
Positively Invariant Sets for ODEs and Numerical Integration
45
in this paper on the Adams-Bashforth method, and
in order to be concrete, we will concentrate on the
Adams-Bashforth method of order four (AB4) initial-
ized with the usual Runge-Kutta method of the same
order (RK4). However, it is straight-forward to adapt
the proofs to the Adams-Bashforth method of any or-
der initialized with the Runge-Kutta of the same order
and we will elaborate at the appropriate places.
The true solution to the ODE (1) with initial-value
ξ
ξ
ξ is denoted by t 7→ φ
φ
φ(t,ξ
ξ
ξ) and its numerical approx-
imations at t
i
= hi, i N
0
, by
e
φ
φ
φ
i
(ξ
ξ
ξ) or
e
φ
φ
φ
i
, i.e.
e
φ
φ
φ
i
:=
e
φ
φ
φ
i
(ξ
ξ
ξ),
dependent on whether the initial-value ξ
ξ
ξ is clear from
the context or not. The time-step h > 0 is a parameter
of the numerical method. Thus, we set
e
φ
φ
φ
0
= ξ
ξ
ξ and for
i = 0,1,2 we use the RK4 formulas
k
1
= hf(
e
φ
φ
φ
i
) (2)
k
2
= hf(
e
φ
φ
φ
i
+ k
1
/2)
k
3
= hf(
e
φ
φ
φ
i
+ k
2
/2)
k
4
= hf(
e
φ
φ
φ
i
+ k
3
)
e
φ
φ
φ
i+1
=
e
φ
φ
φ
i
+
1
6
(k
1
+ 2k
2
+ 2k
3
+ k
4
).
For i 3 we use the AB4 formula and set
e
φ
φ
φ
i+1
=
e
φ
φ
φ
i
(3)
+
h
24
55f(
e
φ
φ
φ
i
) 59f(
e
φ
φ
φ
i1
) + 37f(
e
φ
φ
φ
i2
) 9f(
e
φ
φ
φ
i3
)
.
Let S R
n
be a compact set and f C
4
(R
n
;R
n
). It
is well known that one can find a constant C
RK4
such
that for every ξ
ξ
ξ S we have
e
φ
φ
φ
1
(ξ
ξ
ξ) φ
φ
φ(h,ξ
ξ
ξ)
max
C
RK4
h
5
, (4)
where the constant C
RK4
depends on the derivatives of
f, up to and including the fourth order, in a compact,
convex set
e
S S, fulfilling
e
φ
φ
φ
1
(ξ
ξ
ξ),φ
φ
φ(h,ξ
ξ
ξ)
e
S for all
ξ
ξ
ξ S.
The existence of such a set
e
S is clear from the for-
mulas (2). The reason for this is simply that the er-
ror estimate (2) is derived by using Taylor polynomi-
als for t 7→ φ
φ
φ(t,ξ
ξ
ξ) and using the fact that
˙
φ
φ
φ(t,ξ
ξ
ξ) =
f(φ
φ
φ(t,ξ
ξ
ξ)). From this it is obvious, that for τ
> 0
there exists a constant C such that (4) holds true with
C
RK4
= C and all 0 < h τ
and all ξ
ξ
ξ S. This is
the property we need for the assumptions in Theorem
3.5.
For multi-step methods like AB4 the error esti-
mate is usually formulated differently. Namely that
there exists a constant C
AB4
> 0 such that
e
φ
φ
φ
i+1
(ξ
ξ
ξ) φ
φ
φ((i + 1)h,ξ
ξ
ξ)
max
C
AB4
h
5
if
e
φ
φ
φ
j
= φ
φ
φ( jh, ξ
ξ
ξ) for j = i,i 1, i 2, i 3 in formula
(3), i.e. if the previous approximations
e
φ
φ
φ
j
are exact.
Again, the constant C
AB4
depends on the derivatives
of f, up to and including the fourth order, in a com-
pact, convex set
e
S S, fulfilling
e
φ
φ
φ
i+1
(ξ
ξ
ξ),φ
φ
φ( jh, ξ
ξ
ξ)
e
S
for j = i,i 1, i 2,i 3 for all ξ
ξ
ξ S.
In this form the error estimate is not useful for
our application of computing Lyapunov functions and
contraction metrics. Therefore we prove the follow-
ing theorem.
Theorem 2.1. (Error estimate for AB4-RK4 method)
Consider the system (1) and assume that f
C
4
(R
n
;R
n
). Then AB4 initialized with RK4 fulfills the
property:
For any compact set S R
n
there exist constants
C, h
> 0 such that for all step-sizes 0 < h h
we
have for any i N
0
that
e
φ
φ
φ
i+1
(ξ
ξ
ξ) φ
φ
φ(h,
e
φ
φ
φ
i
(ξ
ξ
ξ))
2
Ch
5
whenever
e
φ
φ
φ
0
(ξ
ξ
ξ),
e
φ
φ
φ
1
(ξ
ξ
ξ),...,
e
φ
φ
φ
i
(ξ
ξ
ξ) S.
Proof. Fix a constant 0 < h
1
1 and a compact, con-
vex set S
S such that φ
φ
φ(h,ξ
ξ
ξ) S
for all ξ
ξ
ξ S and
e
φ
φ
φ
i+1
(ξ
ξ
ξ) S
, whenever
e
φ
φ
φ
j
(ξ
ξ
ξ) S, for j = 0,1,2,...,i,
i N
0
, and when using step-size 0 < h h
1
. Further-
more, let 0 < h
2
h
1
be a constant and
e
S be a com-
pact, convex set such that φ([3h
2
,0],S
)
e
S. Fix an
arbitrary, but constant ξ
ξ
ξ S for the rest of the proof.
For a fixed step-size 0 < h h
2
denote φ
φ
φ
i
(y) :=
φ
φ
φ(ih,y), i Z, and by
e
φ
φ
φ
i
, i N
0
, the approximation
to φ
φ
φ
i
(ξ
ξ
ξ) generated by the numerical method, i.e. AB4
initialized with RK4. There exists a constant C
RK4
>
0, independent of ξ
ξ
ξ S, such that for 0 < h h
2
we
have
e
φ
φ
φ
i+1
φ
φ
φ
1
(
e
φ
φ
φ
i
)
max
C
RK4
h
5
, (5)
as long as 0 < h h
2
and
e
φ
φ
φ
j
S, j = 0,1,...,i, for i =
0,1,2; this is just the classical local truncation error
estimate for RK4.
For the steps with the AB4 method it is advanta-
geous to define
AB4
j=i
(x
j
) :=
x
i
+
h
24
[55f(x
i
) 59f(x
i1
) + 37f(x
i2
) 9f(x
i3
)]
where either x
j
=
e
φ
φ
φ
j
or x
j
= φ
φ
φ
j
(y). Note that x
j
refers
to the sequence (x
j
)
j
, and j in j = i refers to the index
j of the sequence (x
j
)
j
. i is the value of the last index
used in the sequence when AB4
j=i
(x
j
) is computed
from x
i
,x
i1
,x
i2
,x
i3
. In the case x
j
=
e
φ
φ
φ
j
we have
AB4
j=i
(
e
φ
φ
φ
j
) =
e
φ
φ
φ
i+1
if i 3
ICINCO 2023 - 20th International Conference on Informatics in Control, Automation and Robotics
46
and in the case x
j
= φ
φ
φ
j
(y) there exist a constant
C
AB4
> 0, such that
φ
φ
φ
i+1
(y) AB4
j=i
(φ
φ
φ
j
(y))
max
C
AB4
h
5
(6)
as long as 0 < h h
2
and φ
φ
φ
j
(y)
e
S for
j = i, i 1,i 2,i 3, in particular for all y S
. We
now prove the theorem by induction.
Denote by (I) the proposition:
There exist constants C,C
,h
> 0, such that for
every time-step 0 < h h
we have for any i N
0
that
e
φ
φ
φ
i
φ
φ
φ
1
(
e
φ
φ
φ
i1
)
max
Ch
5
, (7)
whenever
e
φ
φ
φ
k
S for k = 0,1,...,i 1, and addition-
ally we have for i 3 and with j = 0,1,2,3, that
e
φ
φ
φ
ij
φ
φ
φ
j
(
e
φ
φ
φ
i
)
max
C
h
5
. (8)
The assertions of the theorem clearly follow from (I).
To prove (I) let us first fix the constants. Let L > 0
be a Lipschitz constant for f on
e
S and set
C := 2 max{C
RK4
,C
AB4
}, (9)
C
:= e
L
(1 + e
L
+ e
2L
)C, (10)
and
h
:= min
h
2
,
C C
AB4
5C
L
. (11)
We now show that (I) holds true for i = 0,1,2,3.
Indeed, (7) follows immediately from (5), and (8) fol-
lows immediately from the standard error estimate for
an explicit numerical integrator with local truncation
error Ch
5
since 0 < h h
, see e.g. (Sauer, 2012).
Now assume that (I) holds true for all natural num-
bers up to and including some i 3. We assume that
e
φ
φ
φ
i
S and show that (I) also holds true for i + 1.
Let us first consider (7) with i replaced by i + 1.
Observe that
e
φ
φ
φ
i+1
φ
φ
φ
1
(
e
φ
φ
φ
i
)
max
= AB4
j=i
(
e
φ
φ
φ
j
) φ
φ
φ
1
(
e
φ
φ
φ
i
)
max
AB4
j=i
(
e
φ
φ
φ
j
) AB4
j=0
(φ
φ
φ
j
(
e
φ
φ
φ
i
))
max
+ AB4
j=0
(φ
φ
φ
j
(
e
φ
φ
φ
i
)) φ
φ
φ
1
(
e
φ
φ
φ
i
)
max
(12)
and for the second term on the right-hand-side we
have the bound
AB4
j=0
(φ
φ
φ
j
(
e
φ
φ
φ
i
)) φ
φ
φ
1
(
e
φ
φ
φ
i
)
max
C
AB4
h
5
(13)
by (6).
To bound the first term on the right-hand-side of
(12) we use the formula for AB4, that φ
φ
φ
0
(
e
φ
φ
φ
i
) =
e
φ
φ
φ
i
,
the Lipschitz condition on f on
e
S, and induction hy-
pothesis (8), and we get
AB4
j=i
(
e
φ
φ
φ
j
) AB4
j=0
(φ
φ
φ
j
(
e
φ
φ
φ
i
))
max
(14)
=
h
24
59f(
e
φ
φ
φ
i1
) + 59f(φ
φ
φ
1
(
e
φ
φ
φ
i
))
+ 37f(
e
φ
φ
φ
i2
) 37f(φ
φ
φ
2
(
e
φ
φ
φ
i
))
9f(
e
φ
φ
φ
i3
) + 9f(φ
φ
φ
3
(
e
φ
φ
φ
i
))
max
h
24
59f(
e
φ
φ
φ
i1
) f(φ
φ
φ
1
(
e
φ
φ
φ
i
))
max
+ 37f(
e
φ
φ
φ
i2
) f(φ
φ
φ
2
(
e
φ
φ
φ
i
))
max
+ 9f(
e
φ
φ
φ
i3
) f(φ
φ
φ
3
(
e
φ
φ
φ
i
))
max
hL
24
59
e
φ
φ
φ
i1
φ
φ
φ
1
(
e
φ
φ
φ
i
)
max
+ 37
e
φ
φ
φ
i2
φ
φ
φ
2
(
e
φ
φ
φ
i
)
max
+ 9
e
φ
φ
φ
i3
φ
φ
φ
3
(
e
φ
φ
φ
i
)
max
105hL
24
C
h
5
< 5C
Lh
6
.
Hence, (12), (13), and (14) deliver
e
φ
φ
φ
i+1
φ
φ
φ
1
(
e
φ
φ
φ
i
)
max
5C
Lh
6
+C
AB4
h
5
Ch
5
, (15)
because
5C
Lh +C
AB4
5C
Lh
+C
AB4
C
by (11).
Hence, the bound (7) in (I) holds true for i re-
placed by i + 1.
Let us now consider the bound (8) in (I) for i re-
placed by i + 1.
The case j = 0 is obvious and from
e
φ
φ
φ
i+1j
φ
φ
φ
j
(
e
φ
φ
φ
i+1
)
max
= φ
φ
φ
j
(φ
φ
φ
j
(
e
φ
φ
φ
i+1j
)) φ
φ
φ
j
(
e
φ
φ
φ
i+1
)
max
e
jLh
φ
φ
φ
j
(
e
φ
φ
φ
i+1j
)
e
φ
φ
φ
i+1
max
the case j = 1, i.e.
e
φ
φ
φ
i
φ
φ
φ
1
(
e
φ
φ
φ
i+1
)
max
e
Lh
φ
φ
φ
1
(
e
φ
φ
φ
i
)
e
φ
φ
φ
i+1
max
C
h
5
follows from (15) and e
Lh
C C
. Here we used the
well known
φ
φ
φ(t,a) φ
φ
φ(t,b)
max
e
L|t|
a b
max
.
The cases j = 2 and j = 3 now follow similarly from
(15) and the induction hypothesis (8). For j = 2 we
have
e
φ
φ
φ
i1
φ
φ
φ
2
(
e
φ
φ
φ
i+1
)
max
e
φ
φ
φ
i1
φ
φ
φ
1
(
e
φ
φ
φ
i
)
max
+ φ
φ
φ
1
(
e
φ
φ
φ
i
) φ
φ
φ
2
(
e
φ
φ
φ
i+1
)
max
e
Lh
φ
φ
φ
1
(
e
φ
φ
φ
i1
)
e
φ
φ
φ
i
max
+ e
2Lh
φ
φ
φ
1
(
e
φ
φ
φ
i
)
e
φ
φ
φ
i+1
max
e
Lh
(1 + e
Lh
)Ch
5
Positively Invariant Sets for ODEs and Numerical Integration
47
and
e
Lh
(1 + e
Lh
)Ch
5
C
.
For j = 3 we have
e
φ
φ
φ
i2
φ
φ
φ
3
(
e
φ
φ
φ
i+1
)
max
e
φ
φ
φ
i2
φ
φ
φ
1
(
e
φ
φ
φ
i1
)
max
+ φ
φ
φ
1
(
e
φ
φ
φ
i1
) φ
φ
φ
2
(
e
φ
φ
φ
i
)
max
+ φ
φ
φ
2
(
e
φ
φ
φ
i
) φ
φ
φ
3
(
e
φ
φ
φ
i+1
)
max
e
Lh
φ
φ
φ
1
(
e
φ
φ
φ
i2
)
e
φ
φ
φ
i1
max
+ e
2Lh
φ
φ
φ
1
(
e
φ
φ
φ
i1
)
e
φ
φ
φ
i
max
+ e
3Lh
φ
φ
φ
1
(
e
φ
φ
φ
i
)
e
φ
φ
φ
i+1
max
e
Lh
(1 + e
Lh
+ e
2Lh
)Ch
5
,
and
e
Lh
(1 + e
Lh
+ e
2Lh
)C = C
C
.
Thus, we have proved the bound (8) of (I) for i
replaced by i + 1, which concludes the proof.
Definition 2.2. (Order of numerical methods)
A numerical method to solve
˙
x = f(x), f C
p
(R
n
;R
n
), (16)
is said to be of order p N, if for any compact set
S R
n
there exist constants C
φ
φ
φ
,h
> 0 such that for
all step-sizes 0 < h h
we have for any i N
0
that
e
φ
φ
φ
i+1
(ξ
ξ
ξ) φ
φ
φ(h,
e
φ
φ
φ
i
(ξ
ξ
ξ))
2
C
φ
φ
φ
h
p+1
whenever
e
φ
φ
φ
0
(ξ
ξ
ξ),
e
φ
φ
φ
1
(ξ
ξ
ξ),...,
e
φ
φ
φ
i
(ξ
ξ
ξ) S.
With Definition 2.2, Theorem 2.1 can be formu-
lated as: Assume f C
4
(R
n
;R
n
) in (1). Then AB4
initialized with RK4 is of order 4 in the sense of Def-
inition 2.2.
Remark 2.3. It is straight forward to adapt the proof
of Theorem 2.1, under the assumption that f in (1)
is in C
p
(R
n
;R
n
), to the Adams-Bashforth method of
order p initialized with Runge-Kutta of the same or-
der. Hence, an AB-RK pair of order p is a numerical
method of order p in the sense of Definition 2.2.
We are now ready to study positively invariant sets
for the ODE (1), that are also positively invariant for
the numerical method.
3 POSITIVELY INVARIANT SETS
A positively invariant set for system (1), i.e. a set
P R
n
such that φ
φ
φ(t,x) P for all t 0 whenever
x P, is not necessarily positively invariant for a nu-
merical procedure to approximate its solution trajec-
tories. The following example is quite revealing for
the general situation; we show for a simple system
and Euler’s Method that
no matter how small the time-step h > 0 is, the dis-
crete semi-dynamical system defined by Euler’s
Method does not necessarily have the same basins
of attraction as the original system, and
if we restrict Euler’s Method to certain compact
and positively invariant subsets of the basins of at-
traction of the system, then these sets are also pos-
itively invariant for the discrete semi-dynamical
system defined by Euler’s Method for sufficiently
small step sizes.
Recall that semi-dynamical systems are dynamical
systems, with the exception that solution trajectories
are not defined for negative times.
Example 3.1. Consider the system
˙
θ = 1 and ˙r =
r(1 r
2
) in polar coordinates; the origin is an
asymptotically stable equilibrium and the circular
disc B
1
:= {x R
2
: x
2
< 1} is its basin of attrac-
tion. In Cartesian coordinates the equations of mo-
tion are
˙x
˙y
=
x(1 x
2
y
2
) y
y(1 x
2
y
2
) + x
=: f(x,y). (17)
Using Euler’s method at z
0
:= (0, y) and with step-
size h > 0 delivers the approximation
z :=
0
y
+ h
y
y(1 y
2
)
at time h. Now, for y
2
< 1, we have
z
2
2
= h
2
y
2
+ (y hy(1 y
2
))
2
> 1
if
0 < g(y,h)
:= y
2
(1 + (1 y
2
)
2
)
| {z }
=:a>0
h
2
+ 2y
2
(y
2
1)
| {z }
=:b<0
h + y
2
1
|{z}
=:c<0
,
i.e. for
h >
b +
b
2
4ac
2a
> 0.
Because g is continuous and lim
y→±1
g(y,h) = h
2
, for
a fixed h > 0 we can always find y close enough to 1
(or 1), such that z
2
> 1. Hence, B
1
is not posi-
tively invariant for this system when Euler’s method
is used, no matter how small h > 0 is.
Now fix 0 < r < 1 and consider the compact set
B
r
:= {x R
2
: x
2
r}. Note that B
r
is positively
ICINCO 2023 - 20th International Conference on Informatics in Control, Automation and Robotics
48
invariant for the dynamical system defined by (17). If
h > 0 satisfies
(1 + (1 r
2
)
2
)h + 2(r
2
1) 0 (18)
i.e. h
2(1 r
2
)
1 + (1 r
2
)
2
,
then for all z
0
= (x,y) with z
0
2
r one has
z
2
2
= z
0
+ hf(z
0
)
2
2
z
0
2
2
,
i.e. B
r
is positively invariant for Euler’s method with h
fufilling (18). Note that condition (18) can be derived
from considering the special case z
0
= (0,y).
We now show in Theorem 3.2 and Corollary 3.5
that for general systems, sublevel-sets S R
n
of cer-
tain Lyapunov-like functions V are positively invari-
ant for both the system (1) and numerical methods of
order p in the sense of Definition 2.2, p N
+
and
1 p s, to approximate its solution trajectories in S
and sufficiently small step size h > 0.
Theorem 3.2. (Positively invariant sets) Consider the
system (1), let V C
1
(R
n
;R) and assume S is a com-
pact connected component of {x R
n
: V (x) m},
m R. Further assume that V (x) ·f(x) < 0 and that
V (x) points out of S for every x S. Then S is
positively invariant for (1).
Further, assume that f C
p
(R
n
;R
n
) and that we
have a numerical method of order p in the sense of
Definition 2.2. Then there is an h
> 0 such that if
the time-step h of the numerical method fulfills 0 <
h h
, then
e
φ
φ
φ
i+1
(ξ
ξ
ξ) S, whenever
e
φ
φ
φ
k
(ξ
ξ
ξ) S for k =
0,1,2,...,i, i N
0
.
Proof. Define
δ :=
1
2
max
xS
V (x) ·f(x) > 0
and let ε > 0 be such that V (x) ·f(x) δ < 0 for
all
x W := {x R
n
: d(x, S) 2ε},
where
d(x, K) := min
yK
x y
2
for a compact set K.
To see that S is positively invariant for (1) consider
that if it is not, then some solution trajectory starting
in S must intersect S at some point x and then leave
S, that is, there exists an x S and an τ
> 0 such
that φ
φ
φ(τ,x) W \S for all 0 < τ τ
. Then
m < V (φ
φ
φ(τ
,x)) = V (x) +
Z
τ
0
d
dτ
V (φ
φ
φ(τ,x))dτ
= m +
Z
τ
0
V (φ
φ
φ(τ,x)) ·
d
dτ
φ
φ
φ(τ,x)dτ m δτ
,
a contradiction.
In order to prove the desired property for the nu-
merical method, set
V := {x R
n
: d(x, S) ε} W ,
F := max{max
xS
f(x)
2
,1}, and
h
:= min{ε/(2max{F,C
φ
φ
φ
}),1,τ
,h
}.
Here and later in the proof C
φ
φ
φ
,h
> 0 are the constants
for the numerical method from Definition 2.2.
Then, for x S \V and 0 h h
we have
φ
φ
φ(h,x) x
2
Z
h
0
f(φ
φ
φ(s,x))
2
ds hF
and it follows that
d(φ
φ
φ(h,x),S \V ) ε/2, x S \V . (19)
Note that from (19), Definition 2.2 and for
e
φ
φ
φ
k
(ξ
ξ
ξ)
S for k = 0, 1, 2, . . . , i, x :=
e
φ
φ
φ
i
(ξ
ξ
ξ), we have
d(
e
φ
φ
φ
i+1
(ξ
ξ
ξ),S \V ) d(φ
φ
φ(h,x),S \V )
+ φ
φ
φ(h,x)
e
φ
φ
φ
i+1
(ξ
ξ
ξ)
2
ε/2 + ε/2 = ε.
Hence,
e
φ
φ
φ
i+1
(x) S if the time-step of the numerical
method fulfills 0 h h
, whenever
e
φ
φ
φ
k
(ξ
ξ
ξ) S for
k = 0, 1, 2,...,i and
e
φ
φ
φ
i
(ξ
ξ
ξ) = x S \V .
To finish the proof we need to show the statement
in the case
e
φ
φ
φ
k
(ξ
ξ
ξ) S for k = 0, 1, 2, . . . , i and
x =
e
φ
φ
φ
i
(ξ
ξ
ξ) S V .
We assume on the contrary that there are sequences
ξ
ξ
ξ
j
S and 0 < h
j
h
, h
j
0 as j , such that
e
φ
φ
φ
j
i
j
+1
(ξ
ξ
ξ
j
) / S
for all j, although
e
φ
φ
φ
j
0
(ξ
ξ
ξ
j
),
e
φ
φ
φ
j
1
(ξ
ξ
ξ
j
),...,
e
φ
φ
φ
j
i
j
(ξ
ξ
ξ
j
) S, x
j
:=
e
φ
φ
φ
j
i
j
(ξ
ξ
ξ
j
) SV .
Here
e
φ
φ
φ
j
0
(ξ
ξ
ξ
j
),
e
φ
φ
φ
j
1
(ξ
ξ
ξ
j
),...,
e
φ
φ
φ
j
i
j
(ξ
ξ
ξ
j
),
e
φ
φ
φ
j
i
j
+1
(ξ
ξ
ξ
j
)
is the sequence generated by the numerical method
with initial value ξ
ξ
ξ
j
S and step-size h
j
.
Note that since S is positively invariant for (1)
we have φ
φ
φ(h
j
,x
j
) S and therefore V (x
j
) m and
V (φ
φ
φ(h
j
,x
j
)) m for all j. Further, there exists I N
+
such that for all j I we have φ
φ
φ(θh
j
,x
j
) W S
for all θ [0,1] and V (
e
φ
φ
φ
j
i
j
+1
(ξ
ξ
ξ
j
)) > m. Moreover,
there is a convex and compact set
e
S S such that
e
φ
φ
φ
j
i
j
+1
(ξ
ξ
ξ
j
)
e
S for all j. Let L
V
be a Lipschitz con-
stant for V on
e
S and recall that by Definition 2.2 we
have
e
φ
φ
φ
j
i
j
+1
(ξ
ξ
ξ
j
) φ
φ
φ(h
j
,x
j
)
2
C
φ
φ
φ
h
p+1
j
.
Positively Invariant Sets for ODEs and Numerical Integration
49
Now
V (
e
φ
φ
φ
j
i
j
+1
(ξ
ξ
ξ
j
)) V (φ
φ
φ(h
j
,x
j
))
h
j
L
V
e
φ
φ
φ
j
i
j
+1
(ξ
ξ
ξ
j
) φ
φ
φ(h
j
,x
j
)
2
h
j
L
V
C
φ
φ
φ
h
p+1
j
h
j
= L
V
C
φ
φ
φ
h
p
j
,
V (
e
φ
φ
φ
j
i
j
+1
(ξ
ξ
ξ
j
)) V (x
j
)
h
j
>
m m
h
j
= 0,
and
V (
e
φ
φ
φ
j
i
j
+1
(ξ
ξ
ξ
j
)) V (x
j
)
h
j
=
V (
e
φ
φ
φ
j
i
j
+1
(ξ
ξ
ξ
j
)) V (φ
φ
φ(h
j
,x
j
))
h
j
+
V (φ
φ
φ(h
j
,x
j
)) V (x
j
)
h
j
(20)
for all j I.
Define g
j
(t) := V (φ
φ
φ(t,x
j
)). By the Mean-Value
theorem there exist θ
j
(0,1) such that
g
j
(h
j
) g
j
(0) = g
(θ
j
h
j
)h
j
δh
j
holds for all j I. From (20) it follows that
0 limsup
j
V (
e
φ
φ
φ
j
i
j
+1
(ξ
ξ
ξ
j
)) V (x
j
)
h
j
limsup
j
V (
e
φ
φ
φ
j
i
j
+1
(ξ
ξ
ξ
j
)) V (φ
φ
φ(h
j
,x
j
))
h
j
+ limsup
j
V (φ
φ
φ(h
j
,x
j
)) V (x
j
)
h
j
0 δ < 0,
which is a contradiction and thus the theorem is
proved.
Remark 3.3. Note that V (x,y) = x
2
+ y
2
is a Lya-
punov function for system (17) on B
1
and for 0 < r < 1
the set V
1
([0,r
2
]) = {x R
2
: x
2
r} B
1
is a
compact sublevel set that is positively invariant.
Further,
V (x,y) ·f(x,y) = 2r
2
(1 r
2
) < 0
for x
2
+ y
2
= r
2
, i.e. V (x, y) ·f(x,y) is less than a
negative constant at the boundary of V
1
([0,r
2
]). The
last theorem then tells us that there must be an h
> 0
so small that V
1
([0,r
2
]) is positively invariant for
Euler’s Method with step size 0 < h h
too.
In applications it can be more convenient to have
the following version of Theorem 3.2, which can be
used for Lyapunov functions computed by first ap-
proximately solving the Zubov’s PDE
V (x) ·f(x) =
q
δ
2
+ f(x)
2
2
using generalized interpolation in reproducing kernel
Hilbert spaces and then interpolating the values over
the simplices of a triangulation T , see (Giesl and Haf-
stein, 2015a; Giesl et al., 2021) for more details. Note
that in the following theorem, CPA[T ] denotes the
set of these interpolating functions, called continuous
piecewise affine (CPA) functions, which are affine on
each simplex of the triangulation T and continuous
overall. Further, V
ν
R
n
denotes the constant gra-
dient of V in the interior of a simplex S
ν
T .
Theorem 3.4 (CPA version of Thm. 3.2). Let V
CPA[T ] and assume there is a compact connected
component S of {x R
n
: V (x) m}, m R. Assume
that V
ν
points out of S at x for every x SS
ν
and
every S
ν
T , and that there is a constant c > 0 such
that V
ν
·f(x) c for every x in a neighbourhood of
S and every ν such that x S
ν
. Then S is positively
invariant for (1).
Further, assume that f C
p
(R
n
;R
n
) and that we
have a numerical method of order p in the sense of
Definition 2.2. Then there is an h
> 0 such that if
the time-step h of the numerical method fulfills 0 <
h h
, then
e
φ
φ
φ
i+1
(ξ
ξ
ξ) S, whenever
e
φ
φ
φ
k
(ξ
ξ
ξ) S for k =
0,1,2,...,i.
Proof. Essentially, the proof is the same as the proof
of Theorem 3.2; the existence of δ = c/2 and ε > 0
now follow directly from the assumptions. The only
reasoning that needs modification is why
g
j
(h
j
) g
j
(0) δh
j
.
For V CPA[T ] it follows because for
D
+
V (x) := lim sup
h0+
V (φ
φ
φ(h,x)) V (x)
h
we have
D
+
V (x) max
ν: xS
ν
V
ν
·f(x) δ,
see e.g. (Hafstein, 2020, Lem. 2.2), and by a gener-
alized Mean Value Theorem, see (Scheeffer, 1884) or
(Walter, 2004, Thm. 12.24).
Finally, we can state and prove the main result of
this paper.
Theorem 3.5. (Positively invariant sets for the system
and the numerical method) Let x
0
be an exponentially
stable equilibrium of (1), where f C
p
(R
n
;R
n
) with
ICINCO 2023 - 20th International Conference on Informatics in Control, Automation and Robotics
50
p N, and let K A(x
0
) be compact. Then there
exists a compact and connected set S, K S A(x
0
),
with the following property:
Assume we have a numerical method of order p
in the sense of Definition 2.2. Then there exists a con-
stant h
> 0, such that S is positively invariant both for
the original flow φ
φ
φ(0,ξ
ξ
ξ) = ξ
ξ
ξ S, t 7→ φ
φ
φ(t,ξ
ξ
ξ), induced
by (1), and for the sequences
e
φ
φ
φ
i
(ξ
ξ
ξ), i N
0
, generated
by the numerical method with step-size h, 0 < h h
,
for the initial-values ξ
ξ
ξ S. In other words,
e
φ
φ
φ
i
(ξ
ξ
ξ) S
for all ξ
ξ
ξ S and all i N
0
.
Proof. By (Giesl, 2007, Thm. 2.46) there exists a
Lyapunov function V for the system (1) fulfilling
V (x) ·f(x) = −∥x x
0
2
q
1 + f(x)
2
2
for all x A(x
0
). We set r := max
xK
V (x) and S :=
V
1
([0,r]). Since V is also a Lyapunov function for
the system
˙
x = f(x)(1 + f(x)
2
2
)
1/2
,
with a bounded right-hand-side, the set S A(x
0
) is
compact. Since V (φ
φ
φ(t,ξ
ξ
ξ)) r for all t 0 and
x
0
φ
φ
φ([0,),ξ
ξ
ξ) V
1
([0,r]) = S
for all ξ
ξ
ξ S, the set S is also connected. Using this
Lyapunov function and Theorem 3.2 for the numeri-
cal method, the existence of h
> 0 with the claimed
properties follows.
Remark 3.6. By Theorem 2.1 the AB4 method initial-
ized by RK4 is a numerical method of order 4 in the
sense of Definition 2.2, and thus, fulfills the assump-
tions in Theorem 3.5. By Remark 2.3 the same applies
to the Adams-Bashforth method of any order, initial-
ized with the Runge-Kutta method of the same order.
These results are used in (Giesl and Hafstein, 2023)
and (Giesl et al., 2023c) to prove that Lyapunov func-
tions and contraction metrics can be approximated
arbitrarily close on compact sets using numerical in-
tegration and quadrature.
4 CONCLUSIONS
We have shown that for an ODE with an exponen-
tially stable equilibrium x
0
and any compact subset K
of its basin of attraction A(x
0
), we can find a compact
and connected set S, K S A(x
0
), that is positively
invariant, both for the ODE and its numerical approx-
imation. We considered the concrete case of using the
Adams-Bashforth method of fourth order, initialized
with the usual Runge-Kutta method of fourth order,
but we also discussed obvious extensions to an arbi-
trary order. Finally, we demonstrated how such posi-
tively invariant sets can be computed in practice.
ACKNOWLEDGEMENT
The research done for this paper was partially sup-
ported by the Icelandic Research Fund in the grant
228725-051, Linear Programming as an Alternative
to LMIs for Stability Analysis of Switched Systems,
which is gratefully acknowledged.
REFERENCES
Aminzare, Z. and Sontag, E. (2014). Contraction methods
for nonlinear systems: A brief introduction and some
open problems. In Proceedings of the 53rd IEEE Con-
ference on Decision and Control, pages 3835–3847.
Anderson, J. and Papachristodoulou, A. (2015). Advances
in computational Lyapunov analysis using sum-of-
squares programming. Discrete Contin. Dyn. Syst. Ser.
B, 20(8):2361–2381.
Aylward, E., Parrilo, P., and Slotine, J.-J. (2008). Stability
and robustness analysis of nonlinear systems via con-
traction metrics and SOS programming. Automatica,
44(8):2163–2170.
Bj
¨
ornsson, J., Giesl, P., Hafstein, S., Kellett, C., and Li,
H. (2014). Computation of continuous and piecewise
affine Lyapunov functions by numerical approxima-
tions of the Massera construction. In Proceedings
of the CDC, 53rd IEEE Conference on Decision and
Control, pages 5506–5511, Los Angeles (CA), USA.
Bj
¨
ornsson, J., Giesl, P., Hafstein, S., Kellett, C., and Li, H.
(2015). Computation of Lyapunov functions for sys-
tems with multiple attractors. Discrete Contin. Dyn.
Syst. Ser. A, 35(9):4019–4039.
Bj
¨
ornsson, J. and Hafstein, S. (2017). Efficient Lya-
punov function computation for systems with multiple
exponentially stable equilibria. Procedia Computer
Science, 108:655–664. Proceedings of the Interna-
tional Conference on Computational Science (ICCS),
Zurich, Switzerland, 2017.
Borg, G. (1960). A condition for the existence of orbitally
stable solutions of dynamical systems. Kungl. Tekn.
H
¨
ogsk. Handl. 153.
Demidovi
ˇ
c, B. (1961). On the dissipativity of a certain
non-linear system of differential equations. I. Vestnik
Moskov. Univ. Ser. I Mat. Meh., 1961(6):19–27.
Doban, A. (2016). Stability domains computation and sta-
bilization of nonlinear systems: implications for bio-
logical systems. PhD thesis: Eindhoven University of
Technology.
Doban, A. and Lazar, M. (2016). Computation of Lyapunov
functions for nonlinear differential equations via a
Yoshizawa-type construction. IFAC-PapersOnLine,
49(18):29 – 34.
Positively Invariant Sets for ODEs and Numerical Integration
51
Forni, F. and Sepulchre, R. (2014). A differential Lyapunov
framework for Contraction Analysis. IEEE Transac-
tions on Automatic Control, 59(3):614–628.
Giesl, P. (2007). Construction of Global Lyapunov Func-
tions Using Radial Basis Functions. Lecture Notes in
Math. 1904, Springer.
Giesl, P. (2015). Converse theorems on contraction metrics
for an equilibrium. J. Math. Anal. Appl., (424):1380–
1403.
Giesl, P. (2019). Computation of a contraction metric for
a periodic orbit using meshfree collocation. SIAM J.
Appl. Dyn. Syst., 18(3):1536–1564.
Giesl, P. and Hafstein, S. (2013). Construction of a CPA
contraction metric for periodic orbits using semidefi-
nite optimization. Nonlinear Anal., 86:114–134.
Giesl, P. and Hafstein, S. (2014). Revised CPA method to
compute Lyapunov functions for nonlinear systems. J.
Math. Anal. Appl., 410:292–306.
Giesl, P. and Hafstein, S. (2015a). Computation and veri-
fication of Lyapunov functions. SIAM J. Appl. Dyn.
Syst., 14(4):1663–1698.
Giesl, P. and Hafstein, S. (2015b). Review of computa-
tional methods for Lyapunov functions. Discrete Con-
tin. Dyn. Syst. Ser. B, 20(8):2291–2331.
Giesl, P. and Hafstein, S. (2023). Dynamical Systems in
Theoretical Perspective, chapter Lyapunov Functions
by Interpolating Numerical Quadrature: Proof of Con-
vergence, page (to appear). Springer. Springer Pro-
ceedings in Mathematics and Statistics.
Giesl, P., Hafstein, S., Haraldsdottir, M., Thorsteinsson, D.,
and Kawan, C. (2023a). Subgradient algorithm for
computing contraction metrics for equilibria. J. Com-
put. Dynamics, 10(2):281–303.
Giesl, P., Hafstein, S., and Kawan, C. (2023b). Review on
contraction analysis and computation of contraction
metrics. J. Comput. Dynamics, 10(1):1–47.
Giesl, P., Hafstein, S., and Mehrabinezhad, I. (2021). Com-
putation and verification of contraction metrics for pe-
riodic orbits. J. Math. Anal. Appl., 503(2):Paper No.
125309, 32.
Giesl, P., Hafstein, S., and Mehrabinezhad, I. (2023c). Con-
traction metrics by numerical integration and quadra-
ture: Uniform error estimate. In Proceedings of the
20th International Conference on Informatics in Con-
trol, Automation and Robotics (ICINCO), page (sub-
mitted).
Gudmundsson, S. and Hafstein, S. (2015). Lyapunov func-
tion verification: MATLAB implementation. In Pro-
ceedings of the 1st Conference on Modelling, Identifi-
cation and Control of Nonlinear Systems (MICNON),
number 0235, pages 806–811.
Hafstein, S. (2004). A constructive converse Lyapunov the-
orem on exponential stability. Discrete Contin. Dyn.
Syst. Ser. A, 10(3):657–678.
Hafstein, S. (2005). A constructive converse Lyapunov
theorem on asymptotic stability for nonlinear au-
tonomous ordinary differential equations. Dynamical
Systems: An International Journal, 20(3):281–299.
Hafstein, S. (2019a). Computational Science - ICCS 2019:
19th International Conference, Faro, Portugal, June
12-14, 2019, Proceedings, Part V, chapter Numerical
Analysis Project in ODEs for Undergraduate Students,
pages 412–434. Springer.
Hafstein, S. (2019b). Numerical ODE solvers and integra-
tion methods in the computation of CPA Lyapunov
functions. In Proceedings of the 18th European Con-
trol Conference (ECC), pages 1136–1141.
Hafstein, S. (2020). CPA Lyapunov functions: Switched
systems vs. differential inclusions. In Proceedings of
the 17th International Conference on Informatics in
Control, Automation and Robotics (ICINCO), pages
745–753.
Hafstein, S., Kellett, C., and Li, H. (2014a). Computation
of Lyapunov functions for discrete-time systems using
the Yoshizawa construction. In Proceedings of 53rd
IEEE Conference on Decision and Control (CDC).
Hafstein, S., Kellett, C., and Li, H. (2014b). Continu-
ous and piecewise affine Lyapunov functions using the
Yoshizawa construction. In Proceedings of the 2014
American Control Conference (ACC), pages 548–553
(no. 0170), Portland (OR), USA.
Hafstein, S., Kellett, C., and Li, H. (2015). Computing con-
tinuous and piecewise affine Lyapunov functions for
nonlinear systems. J. Comput. Dyn, 2(2):227 – 246.
Hafstein, S. and Valfells, A. (2017). Study of dynamical
systems by fast numerical computation of Lyapunov
functions. In Proceedings of the 14th International
Conference on Dynamical Systems: Theory and Ap-
plications (DSTA), volume Mathematical and Numer-
ical Aspects of Dynamical System Analysis, pages
220–240.
Hafstein, S. and Valfells, A. (2019). Efficient computation
of Lyapunov functions for nonlinear systems by in-
tegrating numerical solutions. Nonlinear Dynamics,
97(3):1895–1910.
Hahn, W. (1967). Stability of Motion. Springer, Berlin.
Hartman, P. (1961). On stability in the large for systems
of ordinary differential equations. Canadian J. Math.,
13:480–492.
Hartman, P. (1964). Ordinary Differential Equations. Wi-
ley, New York.
Julian, P., Guivant, J., and Desages, A. (1999). A
parametrization of piecewise linear Lyapunov func-
tions via linear programming. Int. J. Control, 72(7-
8):702–715.
Kamyar, R. and Peet, M. (2015). Polynomial optimization
with applications to stability analysis and control – an
alternative to sum of squares. Discrete Contin. Dyn.
Syst. Ser. B, 20(8):2383–2417.
Khalil, H. (2002). Nonlinear Systems. Pearson, 3. edition.
Krasovski
˘
i, N. N. (1963). Problems of the Theory of Stabil-
ity of Motion. Mir, Moskow, 1959. English translation
by Stanford University Press.
Lewis, D. (1951). Differential equations referred to a vari-
able metric. Amer. J. Math., 73:48–58.
Lewis, D. C. (1949). Metric properties of differential equa-
tions. Amer. J. Math., 71:294–312.
Li, H., Hafstein, S., and Kellett, C. (2015). Computation of
continuous and piecewise affine Lyapunov functions
ICINCO 2023 - 20th International Conference on Informatics in Control, Automation and Robotics
52
for discrete-time systems. J. Difference Equ. Appl.,
21(6):486–511.
Lohmiller, W. and Slotine, J.-J. (1998). On Contrac-
tion Analysis for Non-linear Systems. Automatica,
34:683–696.
Marin
´
osson, S. (2002). Lyapunov function construction for
ordinary differential equations with linear program-
ming. Dynamical Systems: An International Journal,
17:137–150.
Parrilo, P. (2000). Structured Semidefinite Programs and
Semialgebraic Geometry Methods in Robustness and
Optimization. PhD thesis: California Institute of
Technology Pasadena, California.
Ratschan, S. and She, Z. (2010). Providing a basin of attrac-
tion to a target region of polynomial systems by com-
putation of Lyapunov-like functions. SIAM J. Control
Optim., 48(7):4377–4394.
Sastry, S. (1999). Nonlinear Systems: Analysis, Stability,
and Control. Springer.
Sauer, T. (2012). Numerical Analysis. Pearson, 2nd edition.
Scheeffer, L. (1884). Zur Theorie der stetigen Funktionen
einer reellen Ver
¨
anderlichen. Acta Math., 5(1):183–
194, 279–296.
Simpson-Porco, J. and Bullo, F. (2014). Contraction the-
ory on Riemannian manifolds. Systems Control Lett.,
65:74–80.
Valmorbida, G. and Anderson, J. (2017). Region of attrac-
tion estimation using invariant sets and rational Lya-
punov functions. Automatica, 75:37–45.
Vannelli, A. and Vidyasagar, M. (1985). Maximal Lya-
punov functions and domains of attraction for au-
tonomous nonlinear systems. Automatica, 21(1):69–
80.
Vidyasagar, M. (2002). Nonlinear System Analysis. Clas-
sics in Applied Mathematics. SIAM, 2. edition.
Walter, W. (2004). Analysis 1. Springer, 7th edition.
Yoshizawa, T. (1966). Stability theory by Liapunov’s sec-
ond method. Publications of the Mathematical Society
of Japan, No. 9. The Mathematical Society of Japan,
Tokyo.
Zubov, V. I. (1964). Methods of A. M. Lyapunov and their
application. Translation prepared under the auspices
of the United States Atomic Energy Commission;
edited by Leo F. Boron. P. Noordhoff Ltd, Groningen.
Positively Invariant Sets for ODEs and Numerical Integration
53