tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
not discretetime dynamics as we do and ensures finitely many
switches for any solution on finite time intervals only while we
are seeking for results over the whole domain of the solutions
There is therefore a need for novel methodological tools that
allows establishing that any solution to given discretetime
CLS switch a given maximum number of times
In this context we present general conditions in terms of
set intersections To ease the testing of these conditions we
derive alternative tractable conditions which boil down to
verifying whether a specific solution to an auxiliary discrete
time system is nonnegative thanks to the use of Farkas lemma
These conditions are derived for the case of a partition of the
state space made of two cones only to avoid over complicating
the used notation It is interesting to note that the required
nonnegativity property differs from the abundant literature
on positive systems eg 19 which concentrate on systems
for which allsolutions take nonnegative values component
wise In our case a single solution has to be nonnegative
we therefore present tailored conditions for this purpose which
have their own interest and which are successfully applied to
the optimal insulin infusion problem
The rest of the paper is organized as follows The moti
vating example inspired by 7 and the problem statement
are presented in Section II Stability results are established
in Section III assuming all the solutions of the considered
CLS have a finite number of switches Section IV provides
conditions under which solutions to a CLS exhibit no more
than a given maximum number of switches Section V is
dedicated to the nonnegativity analysis of the auxiliary system
derived from Farkas Lemma The results are finally applied
to examples including the insulin infusion problem in Section VI Section VII concludes the paper Some results are
postponed to the appendix to avoid breaking the flow of
exposition
Notations Rdenotes the set of real numbers Rthe set
of nonnegative real numbers Nthe set of nonnegative
integers NN0 andCthe set of complex numbers For
real matrices or vectors indicates transpose The identity
matrix of the considered set of matrices is denoted I For
any symmetric matrix X0X0 means that Xis
positive semidefinite Given a1 a nRwithnN we
usediaga1 a nto denote the diagonal matrix whose
diagonal components are a1 a n For any real square
matrix MσMdenotes its spectrum For any vectors v we
write v0when all its entries are nonnegative The notation
sforsRstands for the integer part of s and recall that
s1s The interior of a set Sis denoted intS its closure
clSand we use B0 rfor the closed ball centered at the
origin of radius r0of the considered Euclidean space Also
cardSis the cardinal of the set S Given nN we say
that set CRnis a closed convex cone of RnifCis closed
and for any x yCand any a bRaxbyC Given
two sets AandBABstands for the set of functions defined
onBwith values in A
II M OTIVATION AND PROBLEM FORMULATION
We first focus on the LQR problem with scalar positive in
puts Section IIA which covers the optimal insulin infusion
problem presented in more details afterwards Section IIB
After having illustrated the potential of QCLP for the near
optimal insulin infusion we formalize the stability analysis
of the obtained closedloop system as the stability problem
of a CLS Section IIC There we also discuss possible
but unfortunately unsuccessful Lyapunovbased approaches
to establish stability properties for the considered example
thereby motivating the problem statement Section IID
A LQR with positive inputs
The envisioned optimal insulin infusion problem is modeled
as a LQR problem with positive inputs We thus consider the
deterministic discretetime linear system
xt1AxtBut 1
where xtx1
t xn
tRnis the state and utR
is the nonnegative scalar control input at time tN with
nN We denote the solution to 1 initialized at state
xRnwith input sequence uRN
at time tNas
ϕt xuandϕ0 xx The cost function is given for
anyxRnand infinitelength sequence of nonnegative inputs
uu0 u1 RN
by
Jxu

t0ℓϕt xu ut 2
where ℓz vzQz2zSvRv2for any z vRn
RQRnnsuch that QQ0R0SRn
andQ S
SR0 We know from Bellman equation that theDAAFOUZ et al CONEWISE LINEAR SYSTEMS WITH FINITEL Y MANY SWITCHES 3
optimal value function associated with 2 namely Vx
minuRN
Jxusatisfies for any xRn
Vxmin
uRℓx uVAxBu 3
based on which the optimal feedback policy is given by
gxmin
uRℓx uVAxBu 4
Hence to construct the optimal policy g we need to know V
which is very challenging in general because of the constraint
thatuhas to be nonnegative see 9 for results in the
continuoustime case Consequently we exploit an alternative
suboptimal policy proposed in 11 known as QCLP The idea

not discretetime dynamics as we do and ensures finitely many

switches for any solution on finite time intervals only while we

are seeking for results over the whole domain of the solutions

There is therefore a need for novel methodological tools that

allows establishing that any solution to given discretetime

CLS switch a given maximum number of times

In this context we present general conditions in terms of

set intersections To ease the testing of these conditions we

derive alternative tractable conditions which boil down to

verifying whether a specific solution to an auxiliary discrete

time system is nonnegative thanks to the use of Farkas lemma

These conditions are derived for the case of a partition of the

state space made of two cones only to avoid over complicating

the used notation It is interesting to note that the required

nonnegativity property differs from the abundant literature

on positive systems eg 19 which concentrate on systems

for which allsolutions take nonnegative values component

wise In our case a single solution has to be nonnegative

we therefore present tailored conditions for this purpose which

have their own interest and which are successfully applied to

the optimal insulin infusion problem

The rest of the paper is organized as follows The moti

vating example inspired by 7 and the problem statement

are presented in Section II Stability results are established

in Section III assuming all the solutions of the considered

CLS have a finite number of switches Section IV provides

conditions under which solutions to a CLS exhibit no more

than a given maximum number of switches Section V is

dedicated to the nonnegativity analysis of the auxiliary system

derived from Farkas Lemma The results are finally applied

to examples including the insulin infusion problem in Section VI Section VII concludes the paper Some results are

postponed to the appendix to avoid breaking the flow of

exposition

Notations Rdenotes the set of real numbers Rthe set

of nonnegative real numbers Nthe set of nonnegative

integers NN0 andCthe set of complex numbers For

real matrices or vectors indicates transpose The identity

matrix of the considered set of matrices is denoted I For

any symmetric matrix X0X0 means that Xis

positive semidefinite Given a1 a nRwithnN we

usediaga1 a nto denote the diagonal matrix whose

diagonal components are a1 a n For any real square

matrix MσMdenotes its spectrum For any vectors v we

write v0when all its entries are nonnegative The notation

sforsRstands for the integer part of s and recall that

s1s The interior of a set Sis denoted intS its closure

clSand we use B0 rfor the closed ball centered at the

origin of radius r0of the considered Euclidean space Also

cardSis the cardinal of the set S Given nN we say

that set CRnis a closed convex cone of RnifCis closed

and for any x yCand any a bRaxbyC Given

two sets AandBABstands for the set of functions defined

onBwith values in A

II M OTIVATION AND PROBLEM FORMULATION

We first focus on the LQR problem with scalar positive in

puts Section IIA which covers the optimal insulin infusion

problem presented in more details afterwards Section IIB

After having illustrated the potential of QCLP for the near

optimal insulin infusion we formalize the stability analysis

of the obtained closedloop system as the stability problem

of a CLS Section IIC There we also discuss possible

but unfortunately unsuccessful Lyapunovbased approaches

to establish stability properties for the considered example

thereby motivating the problem statement Section IID

A LQR with positive inputs

The envisioned optimal insulin infusion problem is modeled

as a LQR problem with positive inputs We thus consider the

deterministic discretetime linear system

xt1AxtBut 1

where xtx1

t xn

tRnis the state and utR

is the nonnegative scalar control input at time tN with

nN We denote the solution to 1 initialized at state

xRnwith input sequence uRN

at time tNas

ϕt xuandϕ0 xx The cost function is given for

anyxRnand infinitelength sequence of nonnegative inputs

uu0 u1 RN

by

Jxu

t0ℓϕt xu ut 2

where ℓz vzQz2zSvRv2for any z vRn

RQRnnsuch that QQ0R0SRn

andQ S

SR0 We know from Bellman equation that theDAAFOUZ et al CONEWISE LINEAR SYSTEMS WITH FINITEL Y MANY SWITCHES 3

optimal value function associated with 2 namely Vx

minuRN

Jxusatisfies for any xRn

Vxmin

uRℓx uVAxBu 3

based on which the optimal feedback policy is given by

gxmin

uRℓx uVAxBu 4

Hence to construct the optimal policy g we need to know V

which is very challenging in general because of the constraint

thatuhas to be nonnegative see 9 for results in the

continuoustime case Consequently we exploit an alternative

suboptimal policy proposed in 11 known as QCLP The idea