Proceedings of IEEE CDC 2002

On the Construction of Lyapunov Functions using

the Sum of Squares Decomposition1
Antonis Papachristodoulou2 Stephen Prajna2
Control and Dynamical Systems
California Institute of Technology
Pasadena, CA 91125 – USA

Abstract A technique developed recently in [3] for constructing

Lyapunov functions is based on the following compu-
A relaxation of Lyapunov’s direct method has been pro- tational relaxation: instead of asking for V to be pos-
posed recently that allows for an algorithmic construc- itive definite, and for its time derivative negated to
tion of Lyapunov functions to prove stability of equi- be positive semidefinite, it requires them to be sums
libria in nonlinear systems, but the search is restricted of squares. This relaxation circumvents the NP hard-
to systems with polynomial vector fields. In this paper, ness of proving positivity of a polynomial and uses the
the above technique is extended to include systems with computational tractability of proving the existence of
equality, inequality, and integral constraints. This al- a sum of squares decomposition, therefore formulating
lows certain non-polynomial nonlinearities in the vector the Lyapunov function search algorithmically. In this
field to be handled exactly and the constructed Lya- way polynomial nonlinearities can be handled exactly.
punov functions to contain non-polynomial terms. It
also allows robustness analysis to be performed. Some In this paper the class of systems for which construc-
examples are given to illustrate how this is done. tion is possible will be extended to nonlinear systems
with equality, inequality, and integral constraints. Con-
straints arise naturally in many physical systems: me-
1 Introduction chanical systems with equality constraints, systems
with parametric uncertainty (which can be accommo-
Stability of dynamical systems plays a very important dated by inequality constraints), and systems contain-
role in control system analysis and design. Unlike the ing dynamic uncertainty (which can be modelled by
case of linear systems, proving stability of equilibria IQCs). Sufficient conditions for proving stability under
of nonlinear systems is more complicated. A sufficient such constraints will be developed in Section 2. In the
condition is the existence of a Lyapunov function [1]: same section, the conditions will be computationally
a positive definite function V defined in some region relaxed to the existence of a sum of squares decompo-
of the state space containing the equilibrium point sition, to allow for their algorithmic verification.
whose derivative along the system trajectories is neg-
ative semi-definite. This is Lyapunov’s direct method, As a side benefit, we will be able to construct Lyapunov
which even though addresses exactly and in a simple functions for nonlinear systems with non-polynomial
way the important issue of stability, it does not provide vector fields, but which can be transformed to equiva-
any coherent methodology for constructing such a func- lent systems with polynomial vector fields under equal-
tion. Lyapunov’s indirect method that investigates the ity and inequality constraints on the state variables; the
local stability of the equilibria, is inconclusive when search for a Lyapunov function will be performed in-
the linearized system has imaginary axis eigenvalues. directly in terms of these new, non-polynomial terms,
Other methodologies to determine the stability prop- and so the resulting Lyapunov functions will not neces-
erties of the equilibria of nonlinear systems (such as ex- sarily be polynomial. Examples of this will be given in
haustive simulations, Linear Parameter Varying (LPV) Section 3. The sum of squares programs correspond-
techniques, Integral Quadratic Constraint (IQC) for- ing to these examples are converted into semidefinite
mulations [2] etc) are sometimes quite conservative. programs using SOSTOOLS [4], and are solved using
1 Work financially supported by AFOSR MURI “Mathemat-
SeDuMi [5], a semidefinite programming solver.
ical Infrastructure for Robust Virtual Engineering” and “Uni- 2 Constrained Dynamical Systems and
fied Theory for Complex Biological and Engineering Networks”,
NIH/NIGMS AfCS (Alliance for Cellular Signalling), DARPA Stability
“Enlightened multiscale simulation of biochemical networks”, the
Kitano ERATO Systems Biology Project, and URI “Protecting Consider the nonlinear system
Infrastructures from Themselves.”
2 The authors contributed equally to this work. ẋ = fx (x, u), (1)

1
with the following constraints: where we have used the fact that ai1 (x, u), bi2 (x, u)
and ci3 (x, u) satisfy (2)–(4). On the other hand, if we
ai1 (x, u) ≤ 0, for i1 = 1, ..., N1 , (2) have condition (6) fulfilled, then the inequality can be
bi2 (x, u) = 0, for i2 = 1, ..., N2 , (3) divided by w(x, u) and integrated from 0 to T to obtain
RT
0
ci3 (x, u)dt ≤ 0, for i3 = 1, ..., N3 , and ∀T ≥ 0. (4) XZ T
pi1 (x, u)
V (0) − V (T ) ≥ − ai (x, u)dt ≥ 0.
n
Here x ∈ R is the state of the system, and u ∈ R m 0 w(x, u) 1
is a collection of auxiliary variables (such as inputs, The rest of the proof is similar to the proof of Lya-
non-polynomial functions of states, uncertain parame- punov’s theorem, which can be found in many standard
ters (for robustness analysis), etc.). We assume that textbooks, e.g., [1].
the ai1 ’s, bi2 ’s, and ci3 ’s are polynomial functions in
(x, u), and fx (x, u) is a vector of polynomial or ratio-
nal functions in (x, u) with no singularity in D, where Remark 2 For the case in which fx (x, u) is a ratio-
D ⊂ Rm+n is defined as nal vector field such as n(x,u)
d(x,u) , the multiplier w(x, u)
D = {(x, u) ∈ Rm+n | ai1 (x, u) ≤ 0, bi2 (x, u) = 0, should be chosen such that w(x, u) ∂V ∂x fx (x, u) is a poly-
for all i1 and i2 }. nomial. However, w(x, u) cannot be used if we want
to make use of the integral constraints, in which case
Without loss of generality, it is assumed that fx (x, u) =
we have to use condition (5). In the latter case, we
0 for x = 0 and u ∈ Du0 , where
introduce a set of auxiliary variables v defined to be
Du0 = {u ∈ Rm |(0, u) ∈ D}. equal to the derivative of x, and we rewrite the sys-
tem ẋ = fx (x, u) as a system with a polynomial vector
The following theorem is an extension of Lyapunov’s
field and some polynomial equality constraints as fol-
stability theorem, and can be used to prove that the
lows: ẋ = v; d(x, u)v = n(x, u).
origin is a stable equilibrium of the above system.
It uses a technique reminiscent of the well-known S-
procedure [6] in nonlinear and robust control theory.
Remark 3 We will term a Lyapunov function V for
which V is positive definite and − dV
dt ≥ 0 global if no
Theorem 1 Suppose that for the above system there restrictions on the state-space have been made. Other-
exist polynomial functions1 V (x), w(x, u), pi1 (x, u), wise, we term it local.
qi2 (x, u), and constants ri3 ≥ 0 such that

• V (x) is positive definite2 in a neighborhood of the To construct Lyapunov functions for systems with non-
origin. polynomial vector fields, the vector field is rendered
• w(x, u) > 0 and pi1 (x, u) ≥ 0 in D. polynomial using auxiliary variables u that encompass
the non-polynomial terms. Let u = g(x), where g(x)
Then is a vector of non-polynomial functions that appear in
∂V X the vector field. Then u will have its own dynamics,
− fx (x, u) + pi1 (x, u)ai1 (x, u)
∂x ∂g
X X u̇ = ẋ , fu (x, u).
+ qi2 (x, u)bi2 (x, u) + ri3 ci3 (x, u) ≥ 0 (5) ∂x
or Generally, this transformation will induce some equal-
∂V X ity and inequality constraints, as we will see in Sec-
− w(x, u) fx (x, u) + pi1 (x, u)ai1 (x, u) tion 3. To prove stability in this case, we search
∂x
X for a Lyapunov function Ṽ (x, u), such that V (x) =
+ qi2 (x, u)bi2 (x, u) ≥ 0 (6)
Ṽ (x, u)|u=g(x) is positive definite around the origin.
will guarantee that the origin of the state space is a Theorem 1 can be applied directly in this case, by re-
stable equilibrium of the system. placing all occurrence of ∂V ∂x fx (x, u) in (5) and (6) by
∂ Ṽ ∂ Ṽ
∂x xf (x, u)+ ∂u u f (x, u). Note that V (x) contains non-
Proof: If condition (5) is fulfilled, then the inequality polynomial terms, so this technique can also be used
can simply be integrated from time t = 0 to t = T to to construct non-polynomial Lyapunov functions, even
obtain when the original system has a polynomial vector field.
XZ T
V (0) − V (T ) ≥ − {pi1 (x, u)ai1 (x, u) − ri3 ci3 (x, u)}dt The challenge lies in how the above conditions can be
0
≥ 0, verified algorithmically. To this end, we take advantage
1 Although
of the computational tractability of the sum of squares
not written explicitly here, we assume that we keep
track of the indices.
decomposition in order to avoid the NP hardness of
2 Strictly speaking, it is enough to require V to have a local proving that a polynomial function is positive definite
minimum at the origin. or positive semi-definite. The condition that p(x) is a

2
sum of squares is more strict, yet more verifiable, than parameterized by some decision variables, on which re-
p(x) ≥ 0. strictions may be applied to render it positive definite.
When dealing with non-polynomial terms, it may be a
A multivariate polynomial p(x1 , ..., xn ) , p(x) is a sum function of both x and u. This will be made clearer in
of squares (SOS, for brevity), if there exist polynomials the examples in Section 3.
f1 (x), ..., fm (x) such that
m
X If such a V (x) that fulfills the conditions of Program 4
p(x) = fi2 (x). is not found, one of higher order will be sought. Fail-
i=1 ure to find a Lyapunov function does not necessarily
This in turn is equivalent to the existence of a positive mean that the equilibrium is unstable, as all the above
semidefinite matrix Q [3], and a properly chosen vector conditions are sufficient.
of monomials Z(x) such that
Using equality and inequality constraints, the class of
T
p(x) = Z (x)QZ(x). (7) systems for which Lyapunov function synthesis is possi-
ble has been extended dramatically: in many cases it is
Notice that p(x) being an SOS implies p(x) ≥ 0. How- possible to transform systems that originally contained
ever, we note that he converse is not true, except for radical, trigonometric and other terms to systems with
some special cases. In fact there are famous counterex- polynomial vector fields under equality and inequality
amples to this effect (such as the Motzkin form). For a constraints. Furthermore, Lyapunov functions that can
more extensive account on SOS polynomials and their be constructed in this way can be extended to include
applications, the readers should consult [3] and the ref- non-polynomial terms. This will be seen in the next
erences therein. section.

We can search for such a Q by solving a semidefinite 3 Examples

program. This is done using SOSTOOLS, a software
developed for this purpose [4] which uses SeDuMi [5] Three examples will be presented to illustrate how Lya-
as the semidefinite programming solver. punov functions can be constructed using the method
described in the previous section.
Therefore, by deliberately opting to work with polyno-
mial and rational functions, the positive definite condi- 3.1 Example 1: A Simple System with a Poly-
tions in Theorem 1 can be relaxed to the existence of an nomial Vector Field
SOS decomposition, and the problem can be cast as an As a first example, consider the system
SOS program [4]. Under this relaxation, the search for
a bounded degree Lyapunov function V (x) and multi- ẋ1 = −x31 + 4x32 − 6x3 x4
pliers pi1 (x, u), qi1 (x, u), and ri3 can be efficiently per- ẋ2 = −x1 − x2 + x35
formed. The SOS program can therefore be formulated
ẋ3 = x1 x4 − x3 + x4 x6
as follows.
ẋ4 = x1 x3 + x3 x6 − x34
ẋ5 = −2x32 − x5 + x6
Program 4 Suppose that we are given the system (1)–
(4). For a polynomial function W (x) with a predeter- ẋ6 = −3x3 x4 − x35 − x6
mined form that is locally positive definite, find bounded
degree polynomials V (x), pi1 (x, u)’s, qi2 (x, u)’s and which has an equilibrium at the origin. As a first at-
constants ri3 ’s such that tempt, we will try to construct
P6 P a quadratic Lyapunov
6
function of the form V = i=1 j=i aij xi xj where the
1. V (x) − W (x) is a sum of squares (implying ≥ 0), aij ’s are the unknowns. To guarantee positive definite-
2. The left-hand side of inequality (5) or (6) is a ness of V , we impose the condition
sum of squares,
(V − W ) is a sum of squares,
3. pi1 (x, u) are sums of squares, P6
where W = k=1 ǫk x2k , with ǫk ’s positive decision vari-
4. ri3 ≥ 0. ables that satisfy ǫk ≥ 0.1 ∀ k.

In this SOS program, Condition 2 is a computa- A Lyapunov function of this form is not found by SOS-
tional relaxation to inequality (5) or (6) in Theorem 1, TOOLS, so we will aim for a 4th order Lyapunov func-
whereas Condition 1 is required to impose strict pos- tion; in this case, V will contain all monomials
P6 that
itive definiteness on V (x), as required by Theorem 1. have order between 2 and 4, and W = k=1 ǫ1k x2k +
P6 4
Using W (x) restricts V (x) to be positive definite, or at k=1 ǫ2k xk with the ǫ1k ’s and ǫ2k ’s positive decision
least have a local minimum at the origin. W (x) may be variables that satisfy ǫ1k + ǫ2k ≥ 0.1 ∀ k.

3
 2
We translate the equilibrium to the origin using a state
1.8
transformation u → x1 , v → x2 , x1 = u− ū, x2 = v − v̄,
1.6
to get an equivalent system
1.4

1.2 ẋ1 = a − (x1 + ū) + (x1 + ū)2 (x2 + v̄), (12)

= b − (x1 + ū)2 (x2 + v̄),
Region 1 Intersection of regions 1 and 2
1 ẋ2 (13)
a

0.8

0.6
whose equilibrium is at the origin. Now suppose that
0.4
the parameters a and b are not exactly known, but
belong to the set {a ≥ a, b ≥ b}. Notice that when the
0.2

Unstable Equilibrium Point

Region 2 parameters a and b are changed, the equilibrium (ū, v̄)
0
0 0.2 0.4 0.6 0.8 1
b
1.2 1.4 1.6 1.8 2
also changes, see Equations (10)–(11). So there are four
parameters in the state equations (12)-(13) - namely a,
Figure 1: Stability region for the chemical oscillator prob- b, ū, and v̄ - that are coupled via the algebraic equality
lem, Regions 1 and 2 defined by equations (14) constraints (10)-(11). Denote these parameters (which
and (15) can be regarded as auxiliary variables) by u1 through
u4 , in accordance with the notation in Theorem 1.
In this case, SOSTOOLS returns Moreover, there exist inherent constraints on the state
V = 0.7257x21 +1.3x42 +2.325x23 +1.575x24 +0.65x45 +1.3x26 variables, as the concentrations of the reactants has
to be positive. Furthermore, for our purpose, it is
as a Lyapunov function for this system. enough to find a Lyapunov function that has non-
positive derivative in a local region around the equi-
3.2 Example 2: A Chemical Oscillator librium. In this case, we can impose the inequality
Two species models of interacting populations can ex- constraints x21 ≤ γu23 , x22 ≤ γu24 where 0 < γ ≤ 1.
hibit limit cycle periodic oscillations [7]. The simplest, Thus, our system (complete with the equality and in-
but chemically plausible tri-molecular reaction that ad- equality constraints) is described by
mits periodic solutions is
ẋ1 = u1 − (x1 + u3 ) + (x1 + u3 )2 (x2 + u4 )
k1
X ⇋ A,
k
B →2 Y, 2X + Y →3 3X,
k ẋ2 = u2 − (x1 + u3 )2 (x2 + u4 )
k−1
0 ≥ x21 − γu23
in which species X is in dynamical equilibrium with 0 ≥ x22 − γu24
species A with a forward rate of reaction k1 and a back- 0 ≥ a − u1
ward rate of reaction k−1 , and so on. Using the law of
mass action, and non-dimensionalising the equations, 0 ≥ b − u2
we get 0 = u1 − u3 + u23 u4
0 = u2 − u23 u4 .
u̇ = a − u + u2 v, (8)
v̇ = b − u2 v, (9) For this example, two quartic Lyapunov functions,
which prove stability of all the dynamical systems
where u, v are the non-dimensional concentrations of X within the following ranges of a and b and which are
and Y , and a, b are non-negative constant parameters parameterized by u1 and u2 have been constructed.
that depend on the concentrations of A and B. It is
known that for a and b satisfying Region 1: (a, b) = (0.25, 0) (14)
Region 2: (a, b) = (0, 1.05) (15)
(b − a) ≥ (b + a)3 ,
The SOS program is the same as Program 4, and the
the system exhibits a stable limit cycle and the equi-
positive definite function W is similar to the one in
librium point is unstable (see Figure 1).
the previous example. Each of the resulting Lyapunov
A Lyapunov function will be constructed for a region functions has more than 30 terms in it, and is therefore
of the rest of the parameter space, to prove robust sta- not listed here. However, the level curves of these Lya-
bility of the equilibrium. punov functions are shown for two different parameter
values in Figures 2 and 3.
The equilibrium of the above system is a (ū, v̄) pair
that satisfies 3.3 Example 3: A Whirling Pendulum
Consider the whirling pendulum [8] shown in Figure 4.
0 = a − ū + ū2 v̄, (10) It is a pendulum of length lp whose suspension end is
0 = b − ū2 v̄. (11) attached to a rigid arm of length la , with a mass mb

4
 is satisfied, the only equilibria in the system are (x1 , x2 )
0.54 satisfying sin x1 = 0, x2 = 0. One equilibrium corre-
sponds to x1 = 0, i.e., the pendulum is hanging verti-
0.52 cally downward (stable), and the other equilibrium cor-
responds to x1 = π, i.e., the vertically upward position
0.5
(unstable). As θ̇a2 is increased beyond g/lp , a supercrit-
v

ical pitchfork bifurcation of equilibria occurs [9]. The

0.48
(x1 , x2 ) = (0, 0) equilibrium becomes unstable, and two
0.46
other equilibria appear. These equilibria correspond to
cos x1 = l gθ̇2 , x2 = 0.
0.9 1 1.1 p a
u

We will now prove the stability of the equilibrium point

Figure 2: Example 2: a = 0.5, b = 0.5. Solid lines show at the origin for θ̇a satisfying (18), by constructing a
trajectories, dotted curves are level curves of
Lyapunov function. Obviously the energy of this me-
the Lyapunov function
chanical system can be used as a Lyapunov function
0.64
but since our purpose is to show that a Lyapunov func-
0.62
tion can be found using the SOS decomposition, we will
assume that our knowledge is limited to the state equa-
0.6
tions describing the system and that we know nothing
about the underlying energy.
v

0.58

0.56
Since the vector field (16)–(17) is not polynomial, a
0.54
transformation to a polynomial vector field must be
performed before we are able to construct a Lyapunov
1.45 1.5 1.55 1.6 1.65 1.7 1.75
u
function using the SOS decomposition. For this pur-
pose, introduce u1 = sin x1 and u2 = cos x1 to get:
Figure 3: Example 2: a = 0.1, b = 1.5. Solid lines show
trajectories, dotted curves are level curves of
ẋ1 = x2 , (19)
the Lyapunov function g
. ẋ2 = θ̇a2 u1 u2 − u1 , (20)
θa lp
u̇1 = x2 u2 , (21)
la u̇2 = −x2 u1 . (22)

lp
In addition, we have the algebraic constraint
θp

u21 + u22 − 1 = 0. (23)

mb

The whirling pendulum system will now be described

Figure 4: The whirling pendulum by Equations (19)–(23). Notice that all the functions
here are polynomial, so that Theorem 1 can be used to
attached to its free end. The arm rotates with angular prove stability.
velocity θ̇a . The pendulum can oscillate with angular
We will perform the analysis with the parameters of the
velocity θ̇p in a plane normal to the arm, making an
system set at some fixed values. Assume that all the
angle θp with the vertical in the instantaneous plane
parameters except g are equal to 1, and g itself is equal
of motion. We will ignore frictional effects and assume
to 10, for which condition (18) is satisfied. For a me-
that all links are slender so that their moment of inertia
chanical system like this, we expect that some trigono-
can be neglected.
metric terms will be needed in the Lyapunov function.
Using x1 = θp and x2 = θ̇p as state variables, we obtain Thus we will try to find a Lyapunov function of the
the following state equations for the system: following form:

ẋ1 = x2 , (16) V = a1 x22 + a2 u21 + a3 u22 + a4 u2 + a5 , (24)

g 2 2 2
= a1 x2 + a2 sin x1 + a3 cos x1 + a4 cos x1 + a5
ẋ2 = θ̇a2 sin x1 cos x1 − sin x1 . (17)
lp
The number and stability properties of equilibria in this where the ai ’s are the unknown coefficients. These co-
system depend on the value of θ̇a . When the condition efficients must satisfy

θ̇a2 < g/lp (18) a3 + a4 + a5 = 0, (25)

5
 for V to be equal to zero at (x1 , x2 ) = (0, 0). To guar- functions for dynamical systems with polynomial vec-
antee that V is positive definite, we search for V s that tor fields. This was based on a relaxation of the con-
satisfy dition that a function is positive semidefinite to the
V − ǫ1 (1 − u2 ) − ǫ2 x22 ≥ 0, (26) condition that it is a sum of squares.
where ǫ1 and ǫ2 are positive constants (we set ǫ1 ≥ 0.1,
The class of dynamical systems for which Lyapunov
ǫ2 ≥ 0.1). Positive definiteness holds as
functions can be synthesized with the above method
ǫ1 (1 − u2 ) + ǫ2 x22 = ǫ1 (1 − cos x1 ) + ǫ2 x22 has been extended in this paper to accommodate sys-
tems with equality, inequality and integral constraints.
is a positive definite function in the (x1 , x2 )–space (as- In doing so, certain nonpolynomial nonlinearities can
suming all x1 that differ by 2π are in the same equiva- be handled, as shown in the examples in Section 3,
lence class). and Lyapunov functions that are nonpolynomial in the
An example of Lyapunov function for this whirling pen- state variables can be constructed.
dulum system is given by
Future research directions in this area include charac-
V = 0.33445x22 +1.4615u21 +1.7959u22 −6.689u2 +4.8931. terizing the terms that are needed in the Lyapunov
function, and choosing the shaping function W . Such
In fact, we can take things a bit further, and assume a characterization is crucial, as in most cases the size
that the parameters of the system are fixed at the same of the resulting semidefinite program can be reduced
values as before except for θ̇a , which is assumed to drastically by using a proper choice of W and search-
be unknown, aiming to construct a Lyapunov function ing for Lyapunov functions containing only a minimal
that is parameterized by θ̇a . It is expected that the set of terms that are really needed.
function will be a Lyapunov function when θ̇a satisfies
condition (18), and not otherwise. Such a candidate
Lyapunov function has been found: References
V = 7.3601 + 0.33033x22 − 6.6066u2 + 7.0698θ̇a2 + 0.3304u22 θ̇a2 [1] H. K. Khalil. Nonlinear Systems. Prentice Hall,
Inc., second edition, 1996.
In this case, we cannot impose a condition like (25) to
[2] A. Megretski and A. Rantzer. System analysis
make V = 0 at (x1 , x2 ) = (0, 0), because the value of V
via integral quadratic constraints. IEEE Transactions
at that point also depends on θ̇a . Nevertheless, notice
on Automatic Control, 42(6):819–830, 1997.
that (x1 , x2 ) = (0, 0) will always be a stationary point
of V . [3] P. A. Parrilo. Structured Semidefinite Programs
and Semialgebraic Geometry Methods in Robustness
By construction, the derivative of this function along and Optimization. PhD thesis, Caltech, Pasadena, CA,
the trajectories of the system is nonpositive, but its 2000.
shape near the origin of the state space depends on the
[4] S. Prajna, A. Papachristodoulou, and P. A. Par-
value of θ̇a . The origin will be a minimum of V (at
least locally) when the Hessian of V evaluated at the rilo. Introducing SOSTOOLS: A general purpose sum
origin, of squares programming solver. To appear in Proc.
  41st IEEE Conf. on Decision and Control, 2002.
∂2V ∂2 V  
 ∂x
2 ∂x1 x2 6.6066 − 0.6608θ̇a2 0 [5] J. F. Sturm. Using SeDuMi 1.02, a MATLAB
∂2V
1
∂2 V
 = ,
∂x1 x2 2
0 0.33033 toolbox for optimization over symmetric cones. Opti-
∂x2
x1 =0,x2 =0
mization Methods and Software, 11–12:625–653, 1999.
is positive definite, i.e., when θ̇a2 < 9.9979 ≈ 10. If [6] S. Boyd, L. El Ghaoui, E. Feron, and V. Balakr-
this condition holds, then V will qualify as a Lyapunov ishnan. Linear Matrix Inequalities in System and Con-
function for the system. Notice that the condition we trol Theory. Society for Industrial and Applied Math-
obtain here agrees exactly with condition (18): When ematics (SIAM), 1994.
θ̇a2 ≥ 10, the equilibrium point undergoes a pitchfork
bifurcation and becomes unstable; at the same point, [7] J. D. Murray. Mathematical Biology. Springer-
the origin changes from being a minimum of V to a Verlag, second edition, 1993.
saddle point for it, and hence V stops being a Lyapunov [8] K. Furuta, M. Yamakita, and S. Kobayashi.
function. Swing-up control of inverted pendulum using pseudo-
state feedback. Journal of Systems and Control Engi-
4 Conclusions neering, 206:263–269, 1992.
[9] J.E. Marsden and T. Ratiu. Introduction to Me-
Constructing Lyapunov functions has always been a chanics and Symmetry. Springer-Verlag, NY, second
challenging task and an important problem in dynam- edition, 1999.
ical systems and control theory. An algorithmic ap-
proach was developed recently to construct Lyapunov

View publication stats