Constructing Lyapunov-Krasovskii Functionals

For Linear Time Delay Systems

Antonis Papachristodoulou, Matthew Peet and Sanjay Lall
AbstractWe present an algorithmic methodology for
constructing Lyapunov-Krasovskii (L-K) functionals for
linear time-delay systems, using the sum of squares decom-
position of multivariate polynomials to solve the related in-
nite dimensional Linear Matrix Inequalities (LMIs). The
resulting functionals retain the structure of the complete
L-K functional and yield conditions that approach the true
delay-dependent stability bounds. The method can also be
used to construct parameter-dependent L-K functionals for
certifying stability under parametric uncertainty.
I. INTRODUCTION
Functional Differential Equations (FDEs) are an im-
portant modeling tool for systems involving transport
and propagation of data or with aftereffect. Recently,
research in the area [5], [8] has been intensied as the
simplest adequate models for Internet congestion control
schemes [9], [16] are in the form of nonlinear FDEs.
The presence of delays can have an effect on system
stability and performance, so ignoring them may lead to
design aws and incorrect analysis conclusions. Stability
is classied as delay-independent if it is retained irre-
spective of the size of the delays, and delay-dependent
if it is lost at a certain delay value. In general the former
condition is more conservative as in most cases bounds
on the expected value of the delay exist.
The investigation of the stability properties of lin-
ear time-delay systems is usually performed using
frequency-domain tests which are suitable for systems
with a small number of heterogeneous delays. When
there are many heterogeneous delays involved, an at-
tractive alternative is the use of time-domain (Lyapunov-
based) methodologies, which amount to construct-
ing simple Lyapunov-Krasovskii (L-K) or Lyapunov-
Razumikhin (L-R) certicates by solving an appropriate
set of Linear Matrix Inequalities (LMIs) [6]. The stabil-
ity conditions that can be obtained in this way are often
Work nancially supported by AFOSR MURI, NIH/NIGMS AfCS
(Alliance for Cellular Signalling), DARPA, the Kitano ERATO Sys-
tems Biology Project, and URI Protecting Infrastructures from Them-
selves.
A. Papachristodoulou is with Control and Dynamical Systems,
California Institute of Technology, Pasadena, CA 91125, USA.
Email: [email protected]. Matthew Peet and San-
jay Lall are with the Department of Aeronautics and Astro-
nautics, Stanford University, Stanford, CA 94305, USA. Email:
[email protected],[email protected].
conservative - even though the existence of complete
quadratic L-K functionals necessary and sufcient for
stability is known, and so is their structure [4]. The
reason is that use of the complete functional yields
innite dimensional LMI conditions that are difcult
to verify algorithmically with current tools; researchers
have concentrated on other structures that yield simple
nite dimensional LMI conditions but which inevitably
produce conservative conditions for stability [12]. A
discretization scheme of the innite dimensional LMIs,
based on the complete Lyapunov functional was in-
troduced by Gu [4], wherein the resulting sufcient
conditions were written as a set of nite dimensional
LMIs. This approach carries a high computational cost
as delays closer to the stability boundary are tested.
Moreover the method is quite complicated to set up and
cannot be generalized to nonlinear time delay systems.
In this paper we investigate delay-dependent stability
of linear time-delay systems with or without parametric
uncertainty by solving the innite dimensional LMIs
that the complete L-K functional conditions produce
algorithmically using the sum of squares (SOS) [13]
methodology. The introduction of SOS techniques has
paved the way for analysis of nonlinear systems with
polynomial vector elds [11], and the construction of
parameter-dependent Lyapunov functions for linear sys-
tems with parameters. For nonlinear time delay systems,
this topic has been addressed in [10], [14]. The case of
linear time-delay systems merits special consideration
due to the existence of a known complete quadratic
Lyapunov functional structure necessary and sufcient
for stability. We show that the existence of a complete
L-K functional is equivalent to feasibility of certain
innite dimensional LMIs which can be solved using
the SOS approach with relatively little conservatism. By
increasing the order of the polynomial variables, our
estimate on the delay bound approaches the true value.
In section II, we review the stability conditions for
linear time-delay systems, and present the complete L-
K functional. In section III we give a brief introduction
to the SOS decomposition and how it can be computed.
In section IV we present the main result that can be used
to test delay-dependent stability. In section V we show
how robustness analysis for linear time-delay systems
with parametric uncertainty can be performed. We close
the paper with some conclusions.
A. Notation
Notation is standard; R is the set of real numbers,
R
n
denotes the n-dimensional Euclidean space. For
[0, +), C = C([, 0], R
n
) denotes the Banach space
of continuous functions dened on an interval [, 0],
taking values in R
n
with the topology of uniform con-
vergence, and with a norm = max
[,0]
|()|,
where | | is an arbitrary norm in R
n
.
II. PROBLEM STATEMENT AND PAST RESULTS
Consider the following linear system with delayed
states:
x(t) = A
0
x(t) +
k

i=1
A
i
x(t
i
) (1)
with an initial condition x() = (), [, 0], where
= max{
1
, . . . ,
k
}, C. The inhomogeneous
delays
i
are assumed to be constant, x(t) R
n
and A
i
,
i = 1, . . . , k are known real constant matrices of appro-
priate dimensions. In a later section we will investigate
the case in which the A
i
are parameter-dependent. The
following theorem (Lyapunov-Krasovskii) can be used
to prove asymptotic stability of the above system.
Theorem 1: [4] The system described by Equation
(1) is asymptotically stable if there exists a bounded
quadratic Lyapunov functional V (x
t
) such that for some
> 0, it satises:
V (x
t
) x
t
(0)
2
(2)
and its derivative along the system trajectory satises

V (x
t
) x
t
(0)
2
. (3)
As mentioned in the Introduction, we identify
two types of stability: delay-independent and delay-
dependent. In the rst case the stability property is
retained irrespective of the size of the delay, whereas
in the second one the stability property is a function of
the delay size, seen as a parameter.
This theorem is a natural extension of the Lyapunov
theory for ordinary differential equations (ODEs) to
systems described by FDEs. It respects the fact that the
state is innite dimensional and proposes a certicate
for stability that is a functional rather than a function,
which is the case for systems described by ODEs. It
is well known that a Lyapunov function necessary and
sufcient for stability of the generic linear ODE system
of the form x = Ax is V = x
T
Px where P is a positive
denite matrix satisfying A
T
P + PA < Q for some
positive denite matrix Q. Given a particular Q > 0, the
P that satises the above conditions is unique, and can
be found numerically by solving an LMI, also known as
a feasibility semidenite program [2]. In the same spirit,
the search for structures that are complete, i.e produce
necessary and sufcient conditions for stability in the
case of time-delay systems has produced some important
results in the past few years. For the case of what is
called strong delay-independent stability (for denitions
and details see [1]), the class of such Lyapunov functions
has been completely characterized [1].
Example 2: For system (1) with k = 1, a L-K can-
didate that would yield a delay-independent condition
is
V (x
t
) = x
t
(0)
T
Px
t
(0) +
_
0

x
t
()
T
Sx
t
()d
Sufcient conditions on V (x
t
) to be positive denite are
P > 0, S 0. For

V (x
t
) < 0 we require
_
A
T
P + PA + S PA
1
A
T
1
P S
_
< 0,
i.e. the conditions for stability (see Theorem 1) can be
written as an LMI with P and S as the unknowns.
The above structure may not be adequate to prove
delay-independent stability of a particular system. The
structure presented in [1] would be the next choice.
As far as delay-dependent stability is concerned, the
structure of the L-K functional necessary and sufcient
for delay-dependent stability is known, but difcult to
construct. For this reason, researchers have concen-
trated their attention on nding structures for which an
algorithmic approach can be used [6]. Inevitably the
maximum delays that could be tested in this way were
conservative.
The complete Lyapunov functional, which is neces-
sary and sufcient for delay-dependent stability of the
linear system is known [4] and has the following form:
V (x
t
) = x
T
t
(0)Px
t
(0) + x
T
t
(0)
_
0

P
1
()x
t
()d
+
_
0

x
T
t
()P
T
1
()dx
t
(0)
+
_
0

_
0

x
T
t
()P
2
(, )x
t
()dd
+
_
0

x
T
t
()Qx
t
()d (4)
with appropriate continuity conditions on P, P
1
and P
2
.
The conditions V > 0 and

V < 0 can be thought of
as innite-dimensional versions of the standard nite
dimensional LMI problem where x
t
() is now the
state at time t and summations have been replaced
by integrations. An approach to solve the innite di-
mensional LMIs was considered in [4] by discretising
the kernels, yielding a set of nite-dimensional LMIs
with size dependent on the discretisation level: ner
discretisation can test delay approaching the true bounds
at the expense of increasing computational effort. Setting
up the discretisation scheme in this algorithm is quite
involved; moreover, it returns a complicated discretised
certicate, and there exists no obvious way to generalise
the procedure to non-linear time-delay systems, issues
that the approach we propose in this paper can address.
In this paper, we consider a structure which is a slight
generalization of the complete quadratic L-K functional
(4). We construct certicates in which the kernels (ma-
trices P
1
, P
2
etc) are polynomials in the variables , .
As the order of the kernels is increased, the delay-
dependent stability conditions obtained analytically (i.e.
using frequency domain methods) can be approached. In
order to solve the resulting innite dimensional LMIs,
we propose a method using the SOS decomposition of
multivariate polynomials.
III. THE SUM OF SQUARES DECOMPOSITION
In this section we give a brief introduction to sum
of squares (SOS) polynomials, their use, and how the
existence of a SOS decomposition can be veried algo-
rithmically. A more detailed description can be found in
[13].
Denition 3: A multivariate polynomial p(x), x
R
n
is a Sum of Squares if there exist polynomials f
i
(x),
i = 1, . . . , M such that p(x) =

M
i=1
f
2
i
(x).
An equivalent characterization of SOS polynomials is
given in the following proposition.
Proposition 4: [13] A polynomial p(x) of degree 2d
is SOS if and only if there exists a positive semidenite
matrix Q and a vector Z(x) containing monomials in x
of degree d so that
p = Z(x)
T
QZ(x)
In general, the monomials in Z(x) are not algebraically
independent. Expanding Z(x)
T
QZ(x) and equating the
coefcients of the resulting monomials to the ones in
p(x), we obtain a set of afne relations in the elements
of Q. Since p(x) being SOS is equivalent to Q 0,
the problem of nding a Q which proves that p(x) is an
SOS can be cast as a semidenite program (SDP) [13].
Therefore the problem of seeking a Q such that p is a
SOS can be formulated as an LMI. Note that if a polyno-
mial p(x) is a SOS, then it is globally nonnegative. The
converse is not always true: not all positive semi-denite
polynomials can be written as SOS, apart from 3 special
cases (see [13]) - in fact, testing global non-negativity
of a polynomial p(x) is known to be NP-hard when the
degree of p(x) is greater than 4 [7], whereas checking
whether p can be written as a SOS is computationally
tractable - it can be formulated as an SDP which has a
worst-case polynomial-time complexity. The construc-
tion of the SDP related to the SOS conditions can be
performed efciently using SOSTOOLS [15], a software
that formulates general SOS programmes as SDPs and
calls semidenite programming solvers to solve them.
If the monomials in the p(x) have unknown coef-
cients then the search for feasible values of those
coefcients such that p(x) is nonnegative is also an SDP,
a fact that is important for the construction of Lyapunov
functions and other S-procedure type multipliers.
IV. MAIN RESULTS
In this section we consider the system (1) with k = 1:
x = A
0
x(t) + A
1
x(t ) f(x
t
) (5)
where x(t) R
n
and x
t
() = x(t + ), [, 0].
The results we present here can be easily extended to
the multiple delay case. We are interested in delay-
dependent conditions for stability of this system.
Here we consider structures similar to the complete
quadratic L-K functional (4) for which we construct
certicates in which the kernels (matrices P
1
, P
2
etc)
are polynomials in the variables (, ). To proceed, we
use Theorem 1 and consider the following candidate
Lyapunov functional, which is a slight generalization of
the complete Lyapunov functional (4):
V (x
t
) = a
0
(x
t
(0)) +
_
0

_
0

a
2
(x
t
())dd
+
_
0

_
0

a
1
(, , x
t
(0), x
t
(), x
t
())dd
+
_
0

_
0

a
3
(x
t
())dd (6)
where the a
i
are polynomials in the indicated variables
with bound on the degree. The polynomials are restricted
to be quadratic with respect to x
t
(0), x
t
() and x
t
()
and allowed to be any order with respect to variables
and . Such polynomials are called bipartite and their
SOS decomposition has a special structure [3]. Writing
the expression with a
1
as an innite dimensional LMI,
we have
_
0

_
0

a
1
(, , x
t
(0), x
t
(), x
t
())dd
=
_
0

_
0

x
t
()
T
a
1
(, )x
t
()dd
where a
1
may contain -functions. Positivity of this
expression is an LMI in the state x
t
with the integral
taking the place of the summation and where a
1
(i, j)
would correspond to

A
i,j
for a nite-dimensional matrix
A. In order to express sufcient conditions for the
positivity of this LMI in terms of the polynomial a
1
,
we can rewrite a
1
as follows.
a
1
(, , x
t
(0), x
t
(), x
t
())
=
_
_
x
t
(0)
x
t
()
x
t
()
_
_
T
_
_
a
1
(, )
_
_
_
_
x
t
(0)
x
t
()
x
t
()
_
_
where a
1
(, ) is a polynomial matrix in (, ). Since
a
1
is now expressed as a quadratic form with kernel
a
1
, pointwise positivity of a
1
will now be sufcient for
positivity of the expression. Now denote x
d
= x
t
(),
y = x
t
(), z = x
t
() and x = x
t
(0) for brevity. The
time derivative of V (x
t
) along f given by (5) is:

2

V (x
t
) =
_
0

_
0

v(, , x
t
(0), x
t
(), x
t
())dd
v(, , x
t
(0), x
t
(), x
t
()) =
da
0
dx
+ a
1
(0, , x, x, z)
+
2
_
a
1
x
f
a
1

a
1

_
+ (a
3
(x) a
3
(z))
a
1
(, , x, x
d
, z) + a
1
(, 0, x, y, x)
a
1
(, , x, y, x
d
) + (a
2
(x) a
2
(y))
The kernel of this expression is also quadratic in variables
x, x
d
, y and z and can be written similarly to a
1
. The
conditions of positive deniteness of V and negative
deniteness of

V are innite dimensional LMIs. To
create sufcient conditions for feasibility, we express
the LMIs using a quadratic form with kernel similar to
a
1
. One can impose positivity and negativity conditions
on these kernels for all and . Positivity of this
kernel implies positivity of the quadratic form which
implies positivity of the integral. However, enforcing this
pointwise positivity condition can be conservative. Later
on, we will show how to reduce the conservativeness
through the use of special functions.
By structuring the polynomials and testing positivity
of V and negativity of

V as explained above, it is easy
to see that the resulting sufcient conditions will be
parameterized nite dimensional LMIs in (, ). How-
ever, for notational simplicity we will be working at the
polynomial level - we multiply out the quadratic form
and search for a polynomial certicate using the bipartite
structure of the resulting expression to simplify the
search. Sufcient conditions for stability of the system
can be found in the following proposition:
Proposition 5: Consider the system given by Equa-
tion (5). Suppose we can nd polynomials a
0
(x)
a
1
(, , x, y, z), a
2
(x
t
()) and a
3
(x
t
()) and a positive
constant such that the following conditions hold:
1) a
0
(x) x
2
0,
2) a
1
(, , x, y, z) 0, , [, 0],
3) a
2
(x
t
()) 0, a
3
(x
t
()) 0,
4)
da0
dx
f +
2 a1
x
f
2 a1

2 a1

+ a
2
(x)
a
2
(y) + a
3
(x) a
3
(z) + a
1
(0, , x, x, z)
a
1
(, , x, x
d
, z) + a
1
(, 0, x, y, x)
a
1
(, , x, y, x
d
) x
2
, , [, 0].
Then the system described by Equation (5) is asymptot-
ically stable.
Proof: The rst three conditions impose that:
V (x
t
) x
t
(0)
2
.
Similarly, the fourth condition, and the discussion before
the statement of the proposition imply that

V (x
t
) x
t
(0)
2
for some > 0. Therefore from the statement of
Theorem 1, the system (5) is asymptotically stable.
Condition (2) in the above proposition asks for a
1
to
be non-negative only on a certain interval of and , as
does condition (4). To restrict ourselves to the intervals
[, 0] and [, 0], we use a process similar
to the S-procedure. The polynomial a
1
is required to
be non-negative only when g
1
( + ) 0 and
g
2
( + ) 0 are satised, which can be tested as
follows:
a
1
+ p
1
g
1
+ p
2
g
2
0 (7)
where p
1
and p
2
are sums of squares of degree 2 in x, y
and z and of bounded degree in and this will
retain the bipartite structure of the whole expression,
which will be taken advantage of in the computation.
The same can be done with Constraint (4) in the above
proposition.
In order to reduce conservativeness of the positivity
condition on the kernel of the quadratic form, we can
now add polynomial terms which integrate to zero to the
kernels. For example, we may test pointwise positivity
of a
1
+ b with the constraint
_
0

_
0

b(, )dd = 0
where b has the following structure
b =
_
_
b
11
(, ) b
12
() b
13
()
b
12
() b
22
() 0
b
13
() 0 b
33
()
_
_
(8)
Likewise, for

V , we may use c where
_
0

_
0

c(, )dd = 0 and where

c =
_

_
c
11
(, ) c
12
(, ) c
13
() c
14
()
c
12
(, ) c
22
(, ) c
23
() c
24
()
c
13
() c
23
() c
33
() 0
c
14
() c
24
() 0 c
44
()
_

_
. (9)
The computational complexity of this method in-
creases as the order of the polynomials a
1
with respect
to and is increased. Although the SOS algorithm uses
polynomials in x, y and z as well as and , since the
order of the variables x, y and z is xed at 2 one can take
advantage of the bipartite structure of the expression. For
more details on how sparsity can be exploited, see [13]
and the references therein. SOSTOOLS has routines that
exploit the sparse bipartite structure to reduce the size
of the resulting LMIs.
The system with delay can therefore be tested for
stability by solving a SOS program (10). Note that the
above procedure can be extended to handle systems with
more than one delay. A different Lyapunov structure
may then be required which allows the polynomials to
be discontinuous at discrete points. We now present an
example which investigates stability of a linear time-
delay system.
Example 6: The following two dimensional system
has been analyzed extensively in the past, and various
LMI tests developed by other researchers have been
tested against this example.
x
1
(t) = 2x
1
(t) x
1
(t ) f
1
x
2
(t) = 0.9x
2
(t) x
1
(t ) x
2
(t ) f
2
The system has been shown analytically to be asymptot-
ically stable for [0, 6.17]. The best bound on that
can be obtained without using the discretisation method
in [4] was
max
= 4.3588 in [12]. Using V (x
t
) given
by (6) we can test the maximum delays given in Table I.
We see that as the order of the polynomials is increased,
the LMI conditions we can obtain are better, but this is
at the expense of increasing computational effort.
V. ROBUST STABILITY UNDER PARAMETRIC
UNCERTAINTY
An important issue in control theory is robust stability,
i.e. ensuring stability under uncertainty. In this section
we consider robust stability of linear time-delay systems
under parametric uncertainty. Consider a time-delay
system of the form (5) with an uncertain parameter p:
x(t) = A
0
(p)x(t) + A
1
(p)x(t ) = f(x
t
, p) (11)
This kind of uncertainty can be handled directly using
the above tools as follows. Suppose one is interested in
the parameter set
P = {p R
m
| q
i
(p) 0, i = 1, . . . , N}
i..e the parametric uncertainty is captured by certain
inequalities. We will be proving robust stability for the
above system, by constructing a Parameter Dependent
Lyapunov functional, as follows:
V (x
t
, p) = a
0
(x
t
(0), p)
+
_
0

_
0

a
1
(, , x
t
(0), x
t
(), x
t
(), p)dd+
+
_
0

_
0

a
2
(x
t
(), p)dd +
_
0

_
0

a
3
(x
t
(), p)dd.
(12)
Then the conditions for stability are similar to the ones
in Proposition 5:
Proposition 7: Suppose there exist polynomials
a
0
(x, p), a
1
(, , x, y, z, p), a
2
(x
t
(), p) and
a
3
(x
t
(), p), and a positive constant such that
the following hold for all p P:
1) a
0
(x, p) x
2
0,
2) a
1
(, , x, y, z, p) 0 [, 0], [, 0],
3) a
2
(x
t
(), p) 0, a
3
(x
t
(), p) 0
4) a
1
(0, , x, x, z, p) a
1
(, , x, x
d
, z, p) +
a
1
(, 0, x, y, x, p) a
1
(, , x, y, x
d
, p) +
da0
dx(t)
f + a
2
(x, p) + a
3
(x, p) a
2
(y, p)
a
3
(z, p)) +
2 a1
x
f
2 a1

2 a1

x
2
[, 0], [, 0].
Then the system given by Equation (5) is stable for
all p P.
The conditions p P, which are captured by the
inequality constraints q
i
(p) 0 can be adjoined to the
above inequality constraints using SOS multipliers in
the similar manner that the inequalities ( +) 0 are
adjoined. A similar SOS program to (10) can easily be
constructed.
VI. CONCLUSION
In this paper we have developed a new methodology
for constructing the complete L-K functionals necessary
for stability of linear time-delay systems based on the
SOS decomposition. The results presented in this paper
complement the results already obtained for nonlinear
delay systems using SOS [10], [14]. A signicant contri-
bution of this paper is that the innite dimensional LMIs
obtained when considering complete L-K functionals for
delay-dependent stability of linear time delay systems
can be solved algorithmically using the SOS decompo-
sition. The bipartite structure of the the resulting SOS
conditions, and the use of special techniques discussed
SOS program to test stability of system (5) for a delay of size
Find polynomials a
0
(x), a
1
(, , x, y, z), a
2
(x()), a
3
(x()), > 0 and SOS q
i,j
(, , x, y, z) for i, j = 1, 2
and polynomial matrices b(, ) and c(, ) of the form (8) and (9) such that
a
0
(x) x
2
is SOS, a
1
(, , x, y, z) +
2

j=1
q
1,j
g
i
(, ) + [x, y, z]b(, )[x, y, z]
T
is SOS,
a
2
(x
t
()) is SOS, a
3
(x
t
()) is SOS, (10)
_

da0
dx
f(x
t
)
2 a1
x
f(x
t
) +
2 a1

+
2 a1

a
1
(0, , x, x, z)
+a
1
(, , x, x
d
, z) a
1
(, 0, x, y, x)
+a
1
(, , x, y, x
d
) a
2
(x) a
3
(x) + a
2
(y) + a
3
(z)
x
2
+

2
j=1
q
2,j
g
i
(, ) + [x, x
d
, y, z]c(, )[x, x
d
, y, z]
T
_

_
is SOS,
_
0

_
0

b(, )dd = 0,
_
0

_
0

c(, )dd = 0.
Order of a
1
in and 0 1 2 3 4 5 6
max 4.472 5.17 5.75 6.02 6.09 6.15 6.16
TABLE I
max FOR DIFFERENT DEGREE POLYNOMIALS a
1
IN AND IN V (xt) FOR EXAMPLE 6
allows the solution of the innite dimensional LMIs with
little conservatism. This same idea was extended to the
case of proving stability under parametric uncertainty.
The complexity of this method increases as higher
order polynomials and delays closer to the analytical
ones are tested. However, one can take advantage of
sparsity in the resulting polynomials therefore reducing
signicantly the computational complexity. Also, if the
presence of certain functions (exponential, trigonomet-
ric, etc) in these functionals is known, then it is some-
times possible to search certicates that contain them
[11]. Moreover, the procedure presented in this paper
can be extended to many other cases that appear for
time-delay systems: stability of neutral systems, systems
with distributed delays, synthesis etc. More importantly,
it can be extended to the analysis of nonlinear time delay
systems, which was the subject of earlier work [10],
[14].
REFERENCES
[1] P.-A. Bliman. Lyapunov equation for the stability of linear delay
systems of retarded and neutral type. IEEE Trans. Automat.
Contr., 47(2):327335, 2002.
[2] S. Boyd, L. El Ghaoui, E. Feron, and V. Balakrishnan. Linear
Matrix Inequalities in System and Control Theory. Society for
Industrial and Applied Mathematics (SIAM), 1994.
[3] K. Gatermann and P. A. Parrilo. Symmetry groups, semidenite
programs, and sums of squares. Journal of Pure and Applied
Algebra, 192(1-3):95128, 2004.
[4] K. Gu, V. L. Kharitonov, and J. Chen. Stability of Time-Delay
systems. Birkhauser, 2003.
[5] V. Kolmanovskii and A. Myshkis. Introduction to the Theory
and Applications of Functional Differential Equations. Kluwer
Academic Publishers, 1999.
[6] V. B. Kolmanovskii, S.-I. Niculescu, and D. Richard. On the
Lyapunov-Krasovskii functionals for stability analysis of linear
delay systems. International Journal of Control, 72(4):374384,
1999.
[7] K. G. Murty and S. N. Kabadi. Some NP-complete problems in
quadratic and nonlinear programming. Mathematical Program-
ming, 39:117129, 1987.
[8] S.-I. Niculescu. Delay Effects on Stability: A Robust Control
Approach. Lecture Notes in Control and Information Sciences
(269). Springer-Verlag, 2001.
[9] F. Paganini, J. Doyle, and S. Low. Scalable laws for stable
network congestion control. In Proceedings of the 40th IEEE
Conf. on Decision and Control, Orlando, FL, 2001.
[10] A. Papachristodoulou. Analysis of nonlinear time delay systems
using the sum of squares decomposition. In Proceedings of the
American Control Conference, 2004.
[11] A. Papachristodoulou and S. Prajna. On the construction of
Lyapunov functions using the sum of squares decomposition.
In Proceedings of the 41st IEEE Conf. on Decision and Control,
pages 34823487, 2002.
[12] P. Park. A delay-dependent stability criterion for systems with
uncertain time-invariant delays. IEEE Trans. Automat. Contr.,
44(2):876877, 1999.
[13] P. A. Parrilo. Structured Semidenite Programs and Semialge-
braic Geometry Methods in Robustness and Optimization. PhD
thesis, California Institute of Technology, Pasadena, CA, 2000.
[14] M. Peet and S. Lall. Constructing Lyapunov functions for nonlin-
ear delay-differential equations using semidenite programming.
In Proceedings of NOLCOS, 2004.
[15] S. Prajna, A. Papachristodoulou, and P. A. Parrilo. SOSTOOLS
Sum of Squares Optimization Toolbox, Users Guide. Available
at http://www.cds.caltech.edu/sostools, 2002.
[16] R. Srikant. The Mathematics of Internet Congestion Control.
Birkh auser, 2003.