tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
the optimal insulin infusion problem
Index Terms Conewise linear systems Lyapunov sta
bility optimizationbased control insulin infusion Farkas
lemma
I INTRODUCTION
Conewise linear systems CLS are dynamical systems for
which the state space is partitioned into a finite number of non
overlapping polyhedral cones 13 The dynamics within
each cone is governed by a linear timeinvariant dynamical
system called a mode These systems pose significant chal
lenges due to their piecewise linear nature In particular it
has been shown in 4 that the stability analysis of CLS is a
NPhard problem Hence algorithms for deciding stability of
CLS are inherently inefficient We can mention the converse
Lyapunov results in 3 which lead to necessary and sufficient
stability conditions The results in 3 state that the origin of
a CLS is globally exponentially stable GES if and only if it
admits a conewise linear Lyapunov function whose associated
conic partition does not coincide with the original system
partition in general As a consequence the method proposed
in 3 may be undecidable or computationally intractable
We can also mention that CLS are known to be equivalent
The authors are with Université de Lorraine CNRS CRAN F54000
Nancy France emails firstnamesurnameunivlorrainefr
This work is supported by the ANR through the projects NICETWEET
ANR20CE480009 TRECOS ANR20CE400009 and OL YMPIA
ANR23CE480006to a class of linear complementarity systems 1 for which
conecopositive Lyapunov functions may be synthesized as
done recently in 5 in continuoustime The advantage is the
derivation of converse results with polynomial approximations
but again there is no guarantee of computational tractability
in general To alleviate the computational obstruction of the
CLS stability analysis an alternative approach is to exploit
additional properties for classes of CLS This is the approach
pursued in this work where we focus on discretetime CLS
whose solutions all exhibit a finite number of switches By
switches we mean the transition of a solution from one cone
to another
Our motivation to study CLS with finitely many switches
comes from the application of the optimizationbased control
approach of 6 namely quadratic controlLyapunov policy
QCLP for optimal insulin infusion in 7 The primary
objective is to minimize peak blood glucose level BGL
caused by a food impulse while adhering to the constraint
that insulin flow must be positive 7 Under the assumption
that the response of the meal lasts longer than the response
of an insulin impulse as observed in specific situations eg
low glycemic index or highfatprotein meals it is proved
in 7 that the optimal infusion policy is given by an open
loop policy combining an insulin bolus applied with the
meal and a specific form of decaying insulin flow thereafter
This strategy exhibits the shortcomings of being openloop
and of requiring the perfect knowledge of the meals This
justifies the need for alternative efficient suboptimal closed
loop feedback policies It appears that the problem can be
formalized as a Linear Quadratic Regulator LQR problem
with positive control inputs as we show This constrained
optimal LQR problem arises in numerous realworld systems
and leads to major methodological challenges 812 Nec
essary and sufficient optimality conditions for the continuous
time case are available in 9 Nevertheless the associated
numerical algorithm suffers from the curse of dimensionality
approximations are thus required to determine the stationary
infinite horizon optimal feedback moreover and importantly
no stability guarantees are provided This is the reason why we
opted for QCLP 11 instead which is easy to implement and
shows remarkable performances in simulations when applied
to the model in 7 see Figure 1 in Section II We show that
the obtained closedloop system can be modelled as a CLS
but ensuring its stability remains challenging Indeed despite
our attempts to apply existing methods for stability analysis of
CLS 3 or optimizationbased control systems such as 13
and its extensions 1416 as well as sumofsquaresbasedarXiv240612530v1 eessSY 18 Jun 20242 GENERIC COLORIZED JOURNAL VOL XX NO XX XXXX 2021
methods like 17 or the discretetime counterpart of 5 we
were unable to obtain a stability certificate for the CLS under
consideration
We realized that the abovementioned motivating example
satisfies a distinctive feature the number of switches of any
solution to the closedloop system is uniformly bounded This
property is extremely useful to investigate stability as any so
lution is eventually given by a linear dynamical system As for
this example each mode has a Schur state matrix the global
exponential stability of the origin can then be established Our
goal in this work is to formalize and generalize these findings
We start by assuming that a given general discretetime
CLS is such that all its solutions exhibit a finite number of
switches and we give a necessary and sufficient condition for
the origin to be globally exponentially stable This stability
property relies on the assumption that solutions switch finitely
many times which is nonrobust to exogenous disturbances
a priori We might thus deduce that the ensured stability
property is not robust We prove that this is not the case by
establishing that for a CLS global exponential stability of the
origin implies exponential inputtostate stability with respect
to a general class of additive disturbances Afterwards we
focus on the main challenge of this work that is to derive
conditions under which all solutions to a CLS exhibit a finite
number of switches We are not aware of such results in the
literature We can mention eg 18 which deals with linear
complementarity systems with continuoustime dynamics and

the optimal insulin infusion problem

Index Terms Conewise linear systems Lyapunov sta

bility optimizationbased control insulin infusion Farkas

lemma

I INTRODUCTION

Conewise linear systems CLS are dynamical systems for

which the state space is partitioned into a finite number of non

overlapping polyhedral cones 13 The dynamics within

each cone is governed by a linear timeinvariant dynamical

system called a mode These systems pose significant chal

lenges due to their piecewise linear nature In particular it

has been shown in 4 that the stability analysis of CLS is a

NPhard problem Hence algorithms for deciding stability of

CLS are inherently inefficient We can mention the converse

Lyapunov results in 3 which lead to necessary and sufficient

stability conditions The results in 3 state that the origin of

a CLS is globally exponentially stable GES if and only if it

admits a conewise linear Lyapunov function whose associated

conic partition does not coincide with the original system

partition in general As a consequence the method proposed

in 3 may be undecidable or computationally intractable

We can also mention that CLS are known to be equivalent

The authors are with Université de Lorraine CNRS CRAN F54000

Nancy France emails firstnamesurnameunivlorrainefr

This work is supported by the ANR through the projects NICETWEET

ANR20CE480009 TRECOS ANR20CE400009 and OL YMPIA

ANR23CE480006to a class of linear complementarity systems 1 for which

conecopositive Lyapunov functions may be synthesized as

done recently in 5 in continuoustime The advantage is the

derivation of converse results with polynomial approximations

but again there is no guarantee of computational tractability

in general To alleviate the computational obstruction of the

CLS stability analysis an alternative approach is to exploit

additional properties for classes of CLS This is the approach

pursued in this work where we focus on discretetime CLS

whose solutions all exhibit a finite number of switches By

switches we mean the transition of a solution from one cone

to another

Our motivation to study CLS with finitely many switches

comes from the application of the optimizationbased control

approach of 6 namely quadratic controlLyapunov policy

QCLP for optimal insulin infusion in 7 The primary

objective is to minimize peak blood glucose level BGL

caused by a food impulse while adhering to the constraint

that insulin flow must be positive 7 Under the assumption

that the response of the meal lasts longer than the response

of an insulin impulse as observed in specific situations eg

low glycemic index or highfatprotein meals it is proved

in 7 that the optimal infusion policy is given by an open

loop policy combining an insulin bolus applied with the

meal and a specific form of decaying insulin flow thereafter

This strategy exhibits the shortcomings of being openloop

and of requiring the perfect knowledge of the meals This

justifies the need for alternative efficient suboptimal closed

loop feedback policies It appears that the problem can be

formalized as a Linear Quadratic Regulator LQR problem

with positive control inputs as we show This constrained

optimal LQR problem arises in numerous realworld systems

and leads to major methodological challenges 812 Nec

essary and sufficient optimality conditions for the continuous

time case are available in 9 Nevertheless the associated

numerical algorithm suffers from the curse of dimensionality

approximations are thus required to determine the stationary

infinite horizon optimal feedback moreover and importantly

no stability guarantees are provided This is the reason why we

opted for QCLP 11 instead which is easy to implement and

shows remarkable performances in simulations when applied

to the model in 7 see Figure 1 in Section II We show that

the obtained closedloop system can be modelled as a CLS

but ensuring its stability remains challenging Indeed despite

our attempts to apply existing methods for stability analysis of

CLS 3 or optimizationbased control systems such as 13

and its extensions 1416 as well as sumofsquaresbasedarXiv240612530v1 eessSY 18 Jun 20242 GENERIC COLORIZED JOURNAL VOL XX NO XX XXXX 2021

methods like 17 or the discretetime counterpart of 5 we

were unable to obtain a stability certificate for the CLS under

consideration

We realized that the abovementioned motivating example

satisfies a distinctive feature the number of switches of any

solution to the closedloop system is uniformly bounded This

property is extremely useful to investigate stability as any so

lution is eventually given by a linear dynamical system As for

this example each mode has a Schur state matrix the global

exponential stability of the origin can then be established Our

goal in this work is to formalize and generalize these findings

We start by assuming that a given general discretetime

CLS is such that all its solutions exhibit a finite number of

switches and we give a necessary and sufficient condition for

the origin to be globally exponentially stable This stability

property relies on the assumption that solutions switch finitely

many times which is nonrobust to exogenous disturbances

a priori We might thus deduce that the ensured stability

property is not robust We prove that this is not the case by

establishing that for a CLS global exponential stability of the

origin implies exponential inputtostate stability with respect

to a general class of additive disturbances Afterwards we

focus on the main challenge of this work that is to derive

conditions under which all solutions to a CLS exhibit a finite

number of switches We are not aware of such results in the

literature We can mention eg 18 which deals with linear

complementarity systems with continuoustime dynamics and