Journal of the Franklin Institute 361 (2024) 106915

Contents lists available at ScienceDirect

Journal of the Franklin Institute

journal homepage: www.elsevier.com/locate/fi

Stabilization of polytopic discrete-time varying systems with rate

and magnitude saturating actuators and bounded disturbances
Lucas A.L. Oliveira a,c , Valter J.S. Leite b ,∗, Luís F.P. Silva b , Kevin Guelton c
a Graduate Program on Mathematical and Computational Modeling (PPGMMC) of the CEFET-MG, Av. Amazonas, 5253, 30510-000, Belo
Horizonte, MG, Brazil
b
Department of Mechatronics Engineering, CEFET–MG, R. Álvares Azevedo, 400, 35503-822, Divinópolis, MG, Brazil
c
CReSTIC EA3804 – Université de Reims Champagne-Ardenne, Moulin de la Housse BP1039, 51687 Reims cedex 2, France

ARTICLE INFO ABSTRACT

Keywords: This study tackles the critical challenge of designing parameter-dependent controllers for the
Polytopic modeling wide class of discrete-time polytopic systems under rate and magnitude saturating actuators.
LPV systems Such a class finds widespread use in diverse applications spanning from medicine to industrial
T–S fuzzy
engineering. Our central contribution lies in developing novel convex parameter-dependent
Saturating actuators
state feedback controller synthesis conditions that ensure regional and input-to-state stability,
Rate saturation
Slew rate
effectively mitigating the adverse effects of rate and magnitude-saturating actuators. Our
ISS approach considers the 𝓁2 ∕𝓁∞ gain between disturbance input and controlled output, offering
Region of attraction performance specifications and retaining validity for specific initial conditions and amplitude-
limited input disturbances. Moreover, our methodology is highly adaptable and suitable for a
broad spectrum of systems, including LPV/quasi-LPV and T–S fuzzy systems. We leverage the
position-type feedback model with speed limitation (PMSL) and the generalized sector condition
to derive our controller synthesis method, resulting in a parallel-distributed-compensation
strategy under standard assumptions, ensuring practicality and applicability to diverse system
requirements. To highlight the effectiveness of our approach, we present numerical examples
for comparative evaluation concerning the existing literature. Furthermore, we validate our
methodology through real-time experiments conducted on a nonlinear coupled tank system,
providing concrete evidence of its efficacy and feasibility for real-world implementation.

1. Introduction

Polytopic modeling is a widely employed technique for representing uncertain or nonlinear systems [1]. Such an approach uses
a set of local linear systems weighted by parameters or membership functions, allowing for the consideration of nonlinearities
or uncertainties while maintaining a relatively simple mathematical representation. Various classes of systems can be cast into
polytopic modeling, including Takagi–Sugeno (T–S) fuzzy systems [2] and linear parameter varying (LPV) or quasi-LPV systems [3].
LPV models are commonly used to handle uncertainties or changes in system dynamics that depend on time-varying parameters,
which are often supposed to be available online. On the other hand, T–S fuzzy and quasi-LPV models approximate the nonlinearities
present in the modeled system. In fact, T–S fuzzy and quasi-LPV systems can be viewed as a particular case of LPV systems when
the time-varying parameters depend on the system’s state [4, pp. 5]. Therefore, polytopic systems provide a broad framework

∗ Corresponding author.
E-mail addresses: arantesl@hotmail.com (L.A.L. Oliveira), valter@ieee.org (V.J.S. Leite), luis@cefetmg.br (L.F.P. Silva), kevin.guelton@univ-reims.fr
(K. Guelton).

https://doi.org/10.1016/j.jfranklin.2024.106915
Received 30 September 2023; Received in revised form 1 March 2024; Accepted 8 May 2024
Available online 11 May 2024
0016-0032/© 2024 The Franklin Institute. Published by Elsevier Inc. All rights are reserved, including those for text and data mining, AI training,
and similar technologies.
L.A.L. Oliveira et al. Journal of the Franklin Institute 361 (2024) 106915

for representing the dynamics of a wide range of systems, from medical applications to industrial engineering implementations,
including cyber–physical systems, new event-triggering techniques, and systems under constraints requiring high fidelity models [5–
11]. The key issue in both cases is the availability of information concerning the state and/or the parameter, allowing for a more
elaborate controller design and achieving higher performance indexes on the closed-loop system. The polytopic approach to the LPV
or quasi-LPV framework has several works, mostly regarding only magnitude saturating actuators. Among the different contexts,
we may cite the input-to-state stability approach presented in [12], where a filtered parameter-dependent Lyapunov function is
used allowing a smooth change on controller gains even when the LPV parameters of continuous-time systems under magnitude
saturating actuators present discontinuous jumps. The sampled-data control is considered in [13], for fuzzy Takagi–Sugeno systems,
and in [14], for LPV systems, assuming only magnitude saturating actuators and parameter-dependent Lyapunov functions. The
problem of fault tolerant control for quasi-LPV discrete-time systems is considered in [15], assuming magnitude saturating actuators,
and in [16], considering unmeasured premise variables. The advent of control via network brings new problems such as discussed
in [17] where parameter-dependent observers are developed for quasi-LPV systems under magnitude saturating actuators. In the
context of network control, strategies for saving data transmission on the stabilization of LPV systems under magnitude saturating
actuators are investigated in [18] for discrete-time case and in [19] for fuzzy Takagi–Sugeno continuous-time ones.
The importance of addressing specific phenomena within control systems becomes evident despite the versatility of polytopic
models in capturing nonlinear dynamics. This study directs its attention to the prevalent issue of rate and magnitude limitations
in actuators, a common occurrence in real-world control systems due to various constraints such as physical limitations, safety
regulations, or power constraints. Failure to appropriately handle these limitations can lead to performance degradation, the
emergence of spurious equilibrium points, and even instability in closed-loop systems [20, pp. 5]. Notable incidents, including
air crashes [21,22] and the catastrophic Chernobyl power plant explosion [23], have been attributed to the effects of rate-
limiting saturation. Additionally, in power electronics, slew-rate saturation poses a significant concern as it can restrict device
performance [24]. Despite the acknowledged significance of considering rate-of-change saturation, the literature has not addressed
it to the same extent as magnitude saturation, leaving ample room for further research investigations and motivating this work.
In the context of handling saturating actuators in (slew-)rate and magnitude, two main methods have been proposed [25]. The
first one involves designing a controller that incorporates the nonlinear actuator model, ensuring that the control signal does not
exceed the rate and magnitude limits of the actuator. For continuous-time systems, we can find in [26] and [27] conditions based
on such a method. Similarly, in [28] for discrete-time systems, the basic sector condition [20] helps to represent the saturation
nonlinearity phenomenon, allowing a convex formulation. A method to design a dynamic controller with anti-windup action and
integrators has been proposed in [29] for discrete-time systems by using the generalized sector condition [30]. In such a work, the
authors also consider the size of the region of attraction to ensure the local asymptotic stability of the closed-loop system. A similar
approach to ensure continuous systems’ internal and external stability subject to 𝓁2 disturbances is addressed in [31]. Sliding mode-
based controllers for continuous and discrete-time systems, while considering rate-saturating actuators, are investigated in [25].
Additional works in the context of continuous-time systems include [32] and [33]. In these works, the system control affected by
rate and magnitude saturating actuators is improved with the aid of an anti-windup action.
The second method, known as the position-type feedback model with speed limitation (PMSL), employs a first-order system with a
magnitude-saturating input signal and the rate saturation modeled as a constraint on the actuator state. In such a case, nested
saturations are obtained, which requires a more involved mathematical approach [34]. The authors in [35] design a dynamic
compensator to address independent rate and magnitude saturation, while in [36] and [37] semi-global and local stabilization
methods are proposed, respectively. General conditions for systems with nested saturations are given in [38] for continuous-time
systems and [39,40] for discrete-time systems. These methods involve the utilization of the polytopic model for saturating signals.
It is worth noting that the generalized sector condition is used in [34] to handle nested saturations in continuous-time systems.
Consequently, the results achieved are usually less conservative than those in [26] and [27]. Also, considering nested saturation,
the authors in [41] provide a design method for stabilizing rate-saturating sampled data systems. In [42], the class of discrete-time
polytopic LPV systems under rate and magnitude saturation is investigated, and a convex design method is proposed. However,
the effects of disturbances on closed-loop trajectories are not considered. More recent works in such a context include [43], where
authors introduce an output feedback controller design method based on positively invariant polyhedrons obtained by bilinear
optimization for discrete-time systems. In [44,45], researchers have designed robust gain-scheduled controllers using a modified
full-block S-procedure method, respectively for uncertain discrete-time and continuous-time systems, with bounded disturbances.
Additionally, using the same approach, the authors propose an extension for robust output and state feedback control of uncertain
discrete-time systems in [46]. In [47], the input rate saturation is norm bounded to handle continuous-time T–S fuzzy systems
with rate and magnitude saturation, without considering external disturbances. Additionally, [48] propose a sliding mode control
for T–S fuzzy discrete-time systems subject to rate and magnitude limited inputs designing a controller that does not exceed the
rate limit, also without perturbations effects. Thus, there is clearly a gap in discrete-time polytopic systems with rate and magnitude
saturation, where the presence of amplitude-bounded perturbations may affect the regional stability of the closed-loop system. The
present work intends to fill such a gap by proposing a novel parameter-dependent state feedback controller for such conditions.
The main contribution of this work is the development of new convex parameter-dependent state feedback controller synthesis
conditions. These conditions ensure the regional and input-to-state stability of time-varying discrete-time systems that are subject
to rate and magnitude saturating actuators. This contribution provides a practical controller design solution for a class of polytopic
models that addresses a significant practical issue in control systems.
Furthermore, our proposed conditions consider the 𝓁2 ∕𝓁∞ gain between the disturbance input and the controlled output, thereby
providing a specific performance specification. It is worth noting that these conditions remain valid for specific initial conditions

2
L.A.L. Oliveira et al. Journal of the Franklin Institute 361 (2024) 106915

and amplitude-limited input disturbances. As a result, these conditions represent an extension of our previous work [42], which
focused solely on regional-exponential stabilization.
Additionally, this general formulation applies to a broader class of systems, including LPV/quasi-LPV and T–S fuzzy systems. To
achieve the proposed methods, we employ the PMSL and the generalized sector condition, which form the basis of our proposed
method. Such an approach leads to the development of a parallel-distributed-compensation (PDC) strategy, which, under standard
assumptions, ensures our control strategy’s practicality, meeting most systems’ requirements.
To demonstrate the effectiveness of our approach, we provide numerical examples that compare our findings with the results
of the existing literature. Furthermore, we conduct real-time experiments on a nonlinear coupled tank system. These experiments
highlight the effectiveness of our proposed framework and illustrate its feasibility for implementation in industry-based devices.
In summary, this work bridges a significant gap in the field, addressing discrete-time polytopic systems with rate and magnitude
saturation, particularly in the presence of amplitude-limited perturbations, offering a robust and versatile solution with wide-ranging
practical implications.
In the next section, we formalize the main problem addressed in this paper. Following that, in Section 3, we give some preliminary
results that provide the basis for our main contributions, which are presented in Section 4. Additionally, within Section 4.1, we
elucidate certain convex optimization methods that empower control designers to address intriguing objectives. Moving on to
Section 5, we present numerical comparisons with achievements from the existing literature, along with a real-time experiment that
illustrates the applicability of our proposal in typical industrial processes that employ conventional devices. Finally, we conclude
our paper with Section 6, summarizing our findings and insights.
Notations. The set of real numbers is denoted by R, while 𝑀 ∈ R𝑛×𝑚 and 𝑥 ∈ R𝑛 are, respectively, a matrix with dimension
𝑛 × 𝑚 with real entries and the vector with 𝑛 positions and real entries. R+ stands for positive real numbers, excluding the zero.
The 𝑀 transpose is represented by 𝑀 ⊤ and 𝑀(𝑖) (𝑀𝑖𝑖 ) denotes the 𝑖th row (diagonal element (𝑖, 𝑖)) of the matrix 𝑀. 𝑀(𝑖) ⊤ stands for

(𝑀(𝑖) ) . The Euclidean norm of 𝑥 is denoted by ‖𝑥‖. The symbol ⋆ replaces symmetric transposed blocks in square matrices. Identity
⊤

and null matrices, 𝐈 and 𝟎, are assumed with appropriate dimensions. Given 𝑃 > 𝟎 and 𝜁 ∈ R+ , an ellipsoidal set is defined as
(𝑃 , 𝜁) = {𝑥(𝑘) ∈ R𝑛 ∶ 𝑥⊤ (𝑘)𝑃 𝑥(𝑘) ≤ 𝜁 −1 }.

2. Problem formulation

The class of systems treated in this paper concerns nonlinear discrete-time systems subject to rate and magnitude saturating
actuators and amplitude-bounded disturbances described by:
𝑥(𝑘 + 1) = 𝑓 (𝑥(𝑘)) + 𝐵(𝑥(𝑘))𝛷(𝑘) + 𝐵𝜔 (𝑥(𝑘))𝜔(𝑘),
(1)
𝑧(𝑘) = 𝑔(𝑥(𝑘)) + 𝐷(𝑥(𝑘))𝜔(𝑘).
where 𝑓 ∶ R𝑛 → R𝑛 , 𝑔 ∶ R𝑛 → R𝑛𝑧 , 𝐵 ∶ R𝑛 → R𝑛×𝑛𝑢 , 𝐵𝜔 ∶ R𝑛 → R𝑛×𝑛𝜔 , and 𝐷 ∶ R𝑛 → R𝑛𝑧 ×𝑛𝜔 are nonlinear functions depending on
the states entries, 𝑥(𝑘) ∈ R𝑛 is the system’s states vector, 𝛷(𝑢(𝑘)) ∈ R𝑛𝑢 is the input delivered by the actuator, 𝑧(𝑘) ∈ R𝑛𝑧 is the
performance output, and 𝜔(𝑘) ∈ R𝑛𝜔 is an 𝓁∞ exogenous signal at instant 𝑘 belonging to the following set

 = {𝜔 ∈ R𝑛𝜔 ∶ 𝜔(𝑘)⊤ 𝜔(𝑘) ≤ 𝛿 −1 }, (2)

√
−1
𝛿 > 0, with 𝛿 ∈ R+ representing the maximum amplitude of the disturbance signals. Observe that 𝜔(𝑘) captures a wide class of
amplitude-bounded signals that may be stochastic, deterministic, persistent, or limited over time.
We suppose a state feedback control law where the (linear) computed control signal 𝑢(𝑘) acts over a saturating actuator, which
delivers a signal 𝛷(𝑢(𝑘)) to the controlled system. Fig. 1 presents a diagram that illustrates such an interconnection. Moreover, Fig. 1
highlights the actuator modeling and the feedback control structure. We discuss in the sequel these structures and the closed-loop
interconnection.
The vector-valued function 𝛷(𝑢(𝑘)) ∶ R𝑛𝑢 → R𝑛𝑢 represents the rate and magnitude saturation imposed by the actuators and can
be interpreted as a discrete-time version of the PMSL used in [35]. From the Actuator block in Fig. 1 one gets:
( )
̂ + 1) = 𝑥(𝑘)
𝑥(𝑘 ̂ + 𝚜𝚊𝚝𝚛 𝛯𝚜𝚊𝚝𝚖 (𝑢(𝑘)) − 𝛯 𝑥(𝑘)
̂ , (3)
𝛷(𝑢(𝑘)) = 𝑥(𝑘),
̂ (4)
where 𝑥(𝑘)
̂ ∈ R𝑛𝑢 is the actuators’ state, the elements of the diagonal matrix 𝛯 ∈ R𝑛𝑢 ×𝑛𝑢 are related to the actuators’ eigenvalues,
and for a (linear) control signal 𝑢(𝑘) ∈ R𝑛𝑢 the symmetric saturation functions of rate and magnitude are written as
( ) ( )
𝚜𝚊𝚝𝚛 𝑢(𝑘)(𝓁) = 𝚜𝚒𝚐𝚗(𝑢(𝑘)(𝓁) ) min(|𝑢(𝑘)(𝓁) |, 𝜇𝚛(𝓁) ) and 𝚜𝚊𝚝𝚖 𝑢(𝑘)(𝓁) = 𝚜𝚒𝚐𝚗(𝑢(𝑘)(𝓁) ) min(|𝑢(𝑘)(𝓁) |, 𝜇𝚖(𝓁) )
for the symmetrical limits 𝜇𝚛 ∈ R𝑛𝑢 and 𝜇𝚖 ∈ R𝑛𝑢 , respectively. Thus, the saturated control signal belongs to the interval [−𝜇𝚖 , 𝜇𝚖 ],
i.e. −𝜇𝚖 ≤ 𝛷(𝑢(𝑘)) ≤ 𝜇𝚖 , while the signal change between two consecutive samplings is constrained to [−𝜇𝚛 , 𝜇𝚛 ], meaning that
−𝜇𝚛 ≤ 𝛷(𝑢(𝑘 + 1)) − 𝛷(𝑢(𝑘)) ≤ 𝜇𝚛 .
Therefore, we can associate an 𝓁2 ∕𝓁∞ gain between the performance (controlled) output and the disturbance input to address
the closed-loop performance in terms of disturbance attenuation given by:
‖𝑧(𝑘)‖2
𝛾 = sup . (5)
𝑤(𝑘)≠0 ‖𝜔(𝑘)‖2
Thus, a 𝛾 > 0 verifying (5) is an upper bound ensuring a minimal mitigation degree of the disturbance signal on the performance
output signal. In this sense, designing a control signal minimizing the 𝛾 value ensures a more efficient closed-loop system.

3
L.A.L. Oliveira et al. Journal of the Franklin Institute 361 (2024) 106915

Fig. 1. Schematic diagram of the closed-looped system where the main blocks are highlighted: the controlled system (upper), actuator (middle), and the controller
(bottom).

Remark 1. Note that, because of the saturating actuators, the controller design approach would provide local closed-loop
stabilization properties, especially for unstable open-loop systems. Consequently, it is reasonable to assume the nonlinear system
(1)–(4) belongs to a valid subset of the state space given by

S𝑥 = {𝑥(𝑘) ∈ R𝑛 ; |𝐿(𝑟) 𝑥(𝑘)| ≤ 𝜌(𝑟) } (6)

with 𝜌(𝑟) > 0, 𝐿(𝑟) ∈ R1×𝑛 and 𝑟 = 1, … , 𝜅, where 𝜅 is the number of constraints.

Hence, assuming 𝑥(𝑘) belonging to a valid subset of R𝑛 where the modeling is verified, i.e. for all 𝑥(𝑘) ∈ S𝑥 , and applying the
sector nonlinearity approach [49], the nonlinear system (1)–(4) can be exactly represented by the following compact polytopic
model:
𝑥(𝑘 + 1) = 𝐴(𝛼(𝑘))𝑥(𝑘) + 𝐵(𝛼(𝑘))𝛷(𝑢(𝑘)) + 𝐵𝑤 (𝛼(𝑘))𝜔(𝑘),
(7)
𝑧(𝑘) = 𝐶𝑧 (𝛼(𝑘))𝑥(𝑘) + 𝐷𝑧 (𝛼(𝑘))𝜔(𝑘),
where the weighting function 𝛼(𝑘) may depend on the state vector, which is supposed to be available or measured online. Moreover,
the weighting function belongs to the unitary simplex
{ ∑
𝑁 }
 = 𝛼 ∈ R𝑁 ∶ 𝛼(𝑖) = 1, 𝛼(𝑖) ≥ 0, 𝑖 = 1, … , 𝑁 (8)
𝑖=1

where 𝑁 denotes the number of vertices of the polytopic representation. Furthermore, we denote by 𝛺 the sequence of values
assumed by the vector valued function 𝛼(𝑘) ∈  from an initial instant 𝑘0 :
{ }
𝛺 = 𝛼(𝑘0 ), 𝛼(𝑘0 + 1), … . (9)

Note that the sequence 𝛺 may depend on the system’s state, which is the case for quasi-LPV [3] or T–S fuzzy models [2], casting
the system as a set of 𝑁 local (non)linear models weighted by time-varying functions, or it can assume any sequence as it occurs

4
L.A.L. Oliveira et al. Journal of the Franklin Institute 361 (2024) 106915

in the LPV modeling [50]. Moreover, without loss of generality, we can assume in general 𝑘0 = 0 in (9). Therefore, the matrices
𝐴(𝛼(𝑘)) ∈ R𝑛×𝑛 , 𝐵(𝛼(𝑘)) ∈ R𝑛×𝑛𝑢 , 𝐵𝑤 (𝛼(𝑘)) ∈ R𝑛×𝑛𝑤 , 𝐶𝑧 (𝛼(𝑘)) ∈ R𝑛𝑧 ×𝑛 , and 𝐷𝑧 (𝛼(𝑘)) ∈ R𝑛𝑧 ×𝑛𝑤 exist within a polytopic domain defined
as the convex combination of 𝑁 known vertices:
[ ] ∑𝑁
[ ]
𝐴(𝛼(𝑘)) 𝐵(𝛼(𝑘)) 𝐵𝑤 (𝛼(𝑘)) 𝐶𝑧 (𝛼(𝑘)) 𝐷𝑧 (𝛼(𝑘)) = 𝛼(𝑘)(𝑖) 𝐴𝑖 𝐵𝑖 𝐵𝑤𝑖 𝐶𝑧𝑖 𝐷𝑧𝑖 . (10)
𝑖=1

Considering the Controller block in Fig. 1, we propose the following polytopic state feedback control law
̂
𝑢(𝑘) = 𝐾(𝛼(𝑘))𝑥(𝑘) + 𝐾(𝛼(𝑘))𝑥(𝑘)
̂ = K(𝛼(𝑘))𝜉(𝑘), (11)

where the controller gain matrices 𝐾(𝛼(𝑘)) ∈ R𝑛𝑢 ×𝑛 ̂

and 𝐾(𝛼(𝑘)) ∈ R𝑛𝑢 ×𝑛𝑢 depend on the weighing function 𝛼(𝑘), and also belong to
a polytopic domain, such that

[ ] ∑𝑁
[ ]
K(𝛼(𝑘)) = 𝐾(𝛼(𝑘)) ̂
𝐾(𝛼(𝑘)) = 𝛼(𝑘)(𝑖) 𝐾𝑖 𝐾̂ 𝑖 . (12)
𝑖=1
[ ]⊤
Moreover, the state vector is augmented by the actuator ones, leading to 𝜉(𝑘) = 𝑥(𝑘)⊤ ̂ ⊤
𝑥(𝑘) ∈ R𝑛+𝑛𝑢 . The resulting closed-loop
system is represented by the block diagram in Fig. 1, which is given by:
𝑥(𝑘 + 1) = 𝐴(𝛼(𝑘))𝑥(𝑘) + 𝐵(𝛼(𝑘))𝛷(K(𝛼(𝑘))𝜉(𝑘)) + 𝐵𝑤 (𝛼(𝑘))𝜔(𝑘),
(13)
𝑧(𝑘) = 𝐶𝑧 (𝛼(𝑘))𝑥(𝑘) + 𝐷𝑧 (𝛼(𝑘))𝜔(𝑘).
Because the actuators have bounds in rate and magnitude, the resulting closed-loop system (13) presents a nonlinear behavior.
Thus, global stability is generally unachievable, which is the case when 𝐴(𝛼(𝑘)) is not exponentially stable. Consequently, a
fundamental issue concerns the determination of the initial conditions set 𝑎 ⊆ R𝑛+𝑛𝑢 , called the region of attraction, from which
the emanating trajectories, under 𝜔 = 𝟎, converge asymptotically to the origin. Moreover, assuming the disturbance signal 𝜔 ∈ 
with 𝛿 > 0, it must be determined 𝑎 ⊆ 𝑎 from which the outgoing closed-loop trajectories do not leave the region of attraction
(𝑎 ). Such a region can be non-convex and eventually not limited in certain directions (see [20, pp. 14]), yielding the challenge of
determining the region of attraction estimation 𝑎 ⊆ 𝑎 as large as possible. Thus, to ensure regional stability in the presence of
amplitude-bounded exogenous signals, the following definition adapted from [51] is considered.

Definition 1. Consider a positive scalar 𝛿 > 0 and any sequence 𝜔 ∈ . The resulting closed-loop system is said to be regional
stable if for any initial state belonging to 𝑎 the resulting state trajectories remain bounded in 𝑎 for all 𝑘 ≥ 0. In this case, the
set 𝑎 is positively invariant. Moreover, in case the disturbance is vanishing and the state trajectories converge towards the origin,
then the system is said input-to-state stable (ISS) and the set 𝑎 is contractive.

The objective of this paper is to develop convex methods to design the control gains in (11) ensuring the local stability of
nonlinear systems represented by polytopic discrete-time models under the presence of amplitude and rate saturating actuators and
subject to amplitude disturbance signals satisfying (2). Therefore, two fundamental problems can be formulated as follows:

Problem 1. Given the polytopic discrete-time system representation (7), the saturating actuators (3)–(4), the control law (11), and
assuming no disturbances, i.e., 𝜔 = 𝟎, determine the state feedback control gains K(𝛼(𝑘)) and an estimate of the initial conditions set
𝑎 ⊆ 𝑎 such that the resulting closed-loop system has its origin stable, with trajectories converging towards the origin without
leaving 𝑎 .

Problem 2. Given the polytopic discrete-time system representation (7), the saturating actuators (3)–(4), the control law (11), and
disturbance signals 𝜔 ∈  given in (2), determine the state feedback control gains K(𝛼(𝑘)) and an estimate of the initial conditions
set 𝑎 ⊆ 𝑎 such that the resulting closed-loop system has its origin stable and the trajectories do not leave 𝑎 . Moreover, the
designed controller must guarantee an upper limit to the induced 𝓁2 /𝓁∞ gain, see 𝛾 in (5), between the disturbance signal 𝜔(𝑘) and
the output 𝑧(𝑘), such that
√
‖𝑧(𝑘)‖2 = 𝛾(‖𝜔(𝑘)‖2 + N), (14)

where N refers to the non-null initial conditions.

3. Preliminary results

The dynamical Eq. (3) represents the model for the actuator block depicted in Fig. 1, while the corresponding output is given by
(4). To explore the (local) stability of the entire system, it is essential to incorporate the actuator states into the system dynamics. The
objective is to transform the closed-loop system into a Lur’e-type system, enabling the treatment of rate and magnitude saturation
nonlinearities. This can be achieved by employing decentralized dead-zone functions, yielding the following representation.

𝛹𝚖 (𝑘) = 𝚜𝚊𝚝𝚖 (𝑢(𝑘)) − 𝑢(𝑘)

5
L.A.L. Oliveira et al. Journal of the Franklin Institute 361 (2024) 106915

and
( ) ( )
𝛹𝚛 (𝑘) = 𝚜𝚊𝚝𝚛 𝛯𝚜𝚊𝚝𝚖 (𝑢(𝑘)) − 𝛯 𝑥(𝑘)
̂ − 𝛯𝚜𝚊𝚝𝚖 (𝑢(𝑘)) − 𝛯 𝑥(𝑘)
̂ .
Taking the control law as (11), these dead-zone functions are rewritten, respectively, as
𝛹𝚖 (𝑘) = 𝚜𝚊𝚝𝚖 (K𝜉(𝑘)) − K𝜉(𝑘)
and
(( ) ) (( ) )
𝛹𝚛 (𝑘) = 𝚜𝚊𝚝𝚛 𝛯̂ + 𝛯K(𝛼(𝑘)) 𝜉(𝑘) + 𝛯𝛹𝚖 (𝑘) − 𝛯̂ + 𝛯K(𝛼(𝑘)) 𝜉(𝑘) + 𝛯𝛹𝚖 (𝑘) ,
[ ]
with 𝛯̂ = 𝟎 −𝛯 . Finally, the resulting closed-loop system can be expressed in terms of the system’s and actuators’ states as:
( )
𝜉(𝑘 + 1) = A (𝛼(𝑘)) + B1 K (𝛼(𝑘)) 𝜉(𝑘) + B1 𝛹𝚖 (𝑘) + B2 𝛹𝚛 (𝑘) + B𝑤 (𝛼(𝑘))𝜔(𝑘),
(15)
𝑧(𝑘) = C𝑧 (𝛼(𝑘))𝜉(𝑘) + D𝑧 (𝛼(𝑘))𝜔(𝑘)
with
[ ] [ ] [ ] [ ]
𝐴(𝛼(𝑘)) 𝐵(𝛼(𝑘)) 𝟎 𝟎 𝐵 (𝛼(𝑘))
A(𝛼(𝑘)) = , B1 = , B2 = , B𝑤 (𝛼(𝑘)) = 𝑤 ,
𝟎 𝐈−𝛯 𝛯 𝐈 𝟎 (16)
[ ] [ ]
C𝑧 (𝛼(𝑘)) = 𝐶𝑧 (𝛼(𝑘)) 𝟎 , D𝑧 (𝛼(𝑘)) = 𝐷𝑧 (𝛼(𝑘)) 𝟎 .
In this study, we introduce two polyhedral sets, namely S𝚛 and S𝚖 , which capture the characteristics of rate and magnitude
saturation, respectively. Notably, S𝚛 exhibits a nested saturating function as the rate saturation depends on the magnitude. The
definitions of these sets are given in the sequence:
{ }
|(( ) )|
S𝚖 = 𝜉(𝑘) ∈ R𝑛+𝑛𝑢 ∶ || K(𝛼(𝑘)) − 𝐺m(𝛼(𝑘)) (𝓁) 𝜉(𝑘) || ≤ 𝜇𝚖(𝓁) , (17)
| |
{ }
|(([ ] )
[ ])|
| 𝜉(𝑘) |
S𝚛 = 𝜉(𝑘) ∈ R𝑛+𝑛𝑢 , 𝛹 m(𝑘) ∈ R𝑛𝑢 ∶ | 𝛯̂ + 𝛯K(𝛼(𝑘)) 𝛯 − 𝐺𝚛 (𝛼(𝑘)) (𝓁) | ≤ 𝜇 , (18)
| 𝛹 m(𝑘) | 𝚛(𝓁)
| |
[ ]
for all 𝓁 = 1, … , 𝑛𝑢 . Here, 𝐺𝚖 (𝛼(𝑘)) ∈ R𝑛𝑢 ×(𝑛+𝑛𝑢 ) and 𝐺𝚛 (𝛼(𝑘)) = 𝐺𝚛1 (𝛼(𝑘)) 𝐺𝚛2 (𝛼(𝑘)) are time-varying parameter-dependent
𝑛
matrices, where 𝐺𝚛1 (𝛼(𝑘)) ∈ R 𝑢 ×(𝑛+𝑛 ) 𝑛 ×𝑛
𝑢 and 𝐺 (𝛼(𝑘)) ∈ R 𝑢 𝑢 .
𝚛2
By employing the sets S𝚖 and S𝚛 in conjunction with the generalized sector condition [52], the control gain design is relaxed
to accommodate a certain level of saturation in the control signal 𝑢(𝑘). We leverage the lemma in the sequence to address the
challenges posed by the saturating actuators. The conditions in this lemma were used in [42] as an adaptation of [34, Lemma 1]
by considering parameter dependent matrices 𝐺𝚖 (𝛼(𝑘)), 𝐺𝚛1 (𝛼(𝑘)), 𝐺𝚛2 (𝛼(𝑘)) ∈ R𝑛𝑢 ×(𝑛+𝑛𝑢 ) .

Lemma 1. If 𝜉(𝑘) ∈ S𝚖 , 𝜉(𝑘) and 𝛹𝚖 (𝛼(𝑘)) ∈ S𝚛 then the nonlinearities 𝛹𝚖 (𝑘) and 𝛹𝚛 (𝑘) satisfy the following inequality:
⊤
( )
𝛹𝚖 (𝑘) 𝑇𝚖 𝛹𝚖 (𝑘) + 𝐺𝚖 (𝛼(𝑘))𝜉(𝑘) ≤ 0, (19)
( [ ]⊤ )
𝛹𝚛 (𝑘)⊤ 𝑇𝚛 𝛹𝚛 (𝑘) + 𝐺𝚛 (𝛼(𝑘)) 𝜉(𝑘)⊤ 𝛹𝚖 (𝑘)⊤ ≤ 0, (20)

for any diagonal matrices 𝑇𝚖 > 0 and 𝑇𝚛 > 0 belonging to R𝑛𝑢 ×𝑛𝑢 .

Proof. First, regarding (19), assuming that 𝑢(𝑘) and 𝐺𝚖 (𝛼(𝑘))𝜉(𝑘) belong to S𝚖 , we have that 𝜇𝚖(𝓁) − 𝑢(𝑘)(𝓁) + 𝐺𝚖 (𝛼(𝑘))𝜉(𝑘)(𝓁) ≥ 0
and −𝜇𝚖(𝓁) − 𝑢(𝑘)(𝓁) + 𝐺𝚖 (𝛼(𝑘))𝜉(𝑘)(𝓁) ≤ 0, 𝓁 =(1, … , 𝑛𝑢 . Consequently, 3) cases must be analyzed: ( Case 1: if 𝑢(𝑘)(𝓁) > 𝜇𝚖(𝓁)) , hence
𝛹𝚖 (𝑘)(𝓁) = 𝜇𝚖(𝓁) − 𝑢(𝑘)(𝓁) < 0 and 𝛹𝚖 (𝑘)(𝓁) 𝑇𝚖(𝓁,𝓁) 𝛹𝚖 (𝑘)(𝓁) + 𝐺𝚖 (𝛼(𝑘))𝜉(𝑘)(𝓁) = 𝛹𝚖 (𝑘)((𝓁) 𝑇𝚖(𝓁,𝓁) 𝜇𝚖(𝓁) − 𝑢(𝑘)(𝓁) + )𝐺𝚖 (𝛼(𝑘))𝜉(𝑘)(𝓁) ≤ 0 with
𝑇𝚖(𝓁,𝓁) > 0; Case 2: if −𝜇𝚖(𝓁) ≤ 𝑢(𝑘)(𝓁) ≤ 𝜇𝚖(𝓁) , hence 𝛹𝚖 (𝑘)(𝓁) = 0 and 𝛹𝚖 (𝑘)(𝓁) 𝑇𝚖(𝓁,𝓁) 𝛹𝚖 (𝑘)(𝓁) + 𝐺𝚖 (𝛼(𝑘))𝜉(𝑘)(𝓁) = 0, ∀𝑇𝚖(𝓁,𝓁) > 0; Case 3:
otherwise, 𝑢(𝑘)(𝓁) < −𝜇𝚖(𝓁) , hence 𝛹𝚖 (𝑘)(𝓁) = −𝜇𝚖(𝓁) − 𝑢(𝑘)(𝓁) > 0 and
( ) ( )
𝛹𝚖 (𝑘)(𝓁) 𝑇𝚖(𝓁,𝓁) 𝛹𝚖 (𝑘)(𝓁) + 𝐺𝚖 (𝛼(𝑘))𝜉(𝑘)(𝓁) = 𝛹𝚖 (𝑘)(𝓁) 𝑇𝚖(𝓁,𝓁) −𝜇𝚖(𝓁) − 𝑢(𝑘)(𝓁) + 𝐺𝚖 (𝛼(𝑘))𝜉(𝑘)(𝓁) ≤ 0,
with 𝑇𝚖(𝓁,𝓁) > 0. Consequently, since 𝑢(𝑘) and 𝐺𝚖 (𝛼(𝑘))𝜉(𝑘) are elements of S𝚖 , we have that
( )
𝛹𝚖 (𝑘)(𝓁) 𝑇𝚖(𝓁,𝓁) 𝛹𝚖 (𝑘)(𝓁) + 𝐺𝚖 (𝛼(𝑘))𝜉(𝑘)(𝓁) ≤ 0,
for all 𝑇𝚖(𝓁,𝓁) > 0, 𝓁 = 1, … , 𝑛𝑢 , verifying (19).
Similarly, the same steps are done with (20) for S𝚛 , changing 𝑢(𝑘) by 𝛯𝚜𝚊𝚝𝚖 (𝑢(𝑘)) − 𝛯 𝑥(𝑘),
̂ 𝛹𝚖 (𝑘) by 𝛹𝚛 (𝑘), 𝐺𝚖 (𝛼(𝑘))𝜉(𝑘) by
[ ]⊤
𝐺𝚛 (𝛼(𝑘)) 𝜉(𝑘)⊤ 𝛹𝚖 (𝑘)⊤ , and 𝜇𝚛 by 𝜇𝚖 . This concludes the proof. □
It is worth to say that a key idea behind Lemma 1 is the existence of signals given by 𝐺𝚖 (𝛼(𝑘))𝜉(𝑘) and 𝐺𝚛1 (𝛼(𝑘))𝜉(𝑘)+𝐺𝚛2 (𝛼(𝑘))𝛹𝚖 (𝑘)
compensating the control signal, 𝑢(𝑘), such that sets S𝚖 and S𝚛 are verified.
In what follows, we denote  (𝜂) the level set concerning a Lyapunov function 𝑉 (𝜉(𝑘)), where 𝜉(𝑘) ∈ R𝑛 is the state, parameter
𝛼(𝑘) ∈ R𝑁 belongs to the convex hull  given in (8), and 𝜂 is a positive scalar. Thus, we have
{ }
 (𝜂) = 𝜉(𝑘) ∈ R𝑛 ∶ 𝑉 (𝜉(𝑘)) ≤ 𝜂 −1 , 𝛼(𝑘) ∈ 𝛺 .

6
L.A.L. Oliveira et al. Journal of the Franklin Institute 361 (2024) 106915

4. Main results

This section presents convex conditions allowing the design of state feedback gains that provide solutions to Problems 1 and 2.
The conditions are formulated in terms of linear matrix inequalities (LMIs) which can be solved efficiently by numerical interior point
methods freely available. The proposed solution explores the polytopic structure of the model, offering an equally polytopic-based
control gain to stabilize the closed-loop system locally.

Theorem 1. Consider the polytopic discrete-time system under rate and magnitude saturating actuators given by (7)–(11), matrices A𝑖 ,
B1 , B2 , B𝑤𝑖 , C𝑧𝑖 , and D𝑧𝑖 obtained as in (16), and the known symmetric saturation limits 𝜇𝚖 and 𝜇𝚛 for the actuators’ rate and magnitude,
respectively. Assume that there exist positive scalars 𝜂, 𝛿, 𝜏2 , and 0 < 𝜏1 < 1, positive definite and symmetric matrices 𝑃̃𝑖 ∈ R(𝑛+𝑛𝑢 )×(𝑛+𝑛𝑢 ) ,
diagonal matrices 𝑆𝚖 , 𝑆𝚛 ∈ R𝑛𝑢 ×𝑛𝑢 , matrices 𝑋𝑖 , 𝑊𝚖𝑖 , 𝑊𝚛1 𝑖 ∈ R𝑛𝑢 ×(𝑛+𝑛𝑢 ) , 𝑊𝚛2 𝑖 ∈ R𝑛𝑢 ×𝑛𝑢 , 𝑈̃ ∈ R(𝑛+𝑛𝑢 )×(𝑛+𝑛𝑢 ) , for 𝑖, 𝑗 = 1, … , 𝑁, such that
the following LMIs
⎡−𝜏1 𝑃̃𝑖 −𝑈̃ ⊤ A⊤ ⊤ ⊤
𝑖 − 𝑋𝑖 B1 −𝑊𝚖𝑖⊤ −𝑊𝚛⊤ 𝑖 𝟎 𝑈̃ ⊤ C⊤𝑧𝑖
⎤
⎢ 1
⎥
⎢ ⋆ 𝑃̃𝑗 + 𝑈̃ + 𝑈̃ ⊤ −B1 𝑆𝚖 −B2 𝑆𝚛 −B𝑤𝑖 𝟎 ⎥
⎢ ⎥
⎢ ⋆ ⋆ −2𝑆𝚖 −𝑊𝚛⊤ 𝑖 𝟎 𝟎 ⎥
2 (21)
⎢ ⎥ < 𝟎,
⎢ ⋆ ⋆ ⋆ −2𝑆𝚛 𝟎 𝟎 ⎥
⎢ ⎥
⎢ ⋆ ⋆ ⋆ ⋆ −𝜏2 𝐈 D⊤ 𝑧𝑖 ⎥
⎢ ⎥
⎣ ⋆ ⋆ ⋆ ⋆ ⋆ −𝛾𝐈 ⎦
[ ⊤ −𝑊⊤ ]
−𝑃̃𝑖 𝑋𝑖(𝓁) 𝚖𝑖(𝓁)
2
≤ 𝟎, (22)
⋆ −𝜂𝜇𝚖(𝓁)
⎡−𝑃̃𝑖 𝑊𝚖𝑖⊤ 𝑈̃ ⊤ 𝛯̂ (𝓁)
⊤ + (𝛯𝑋 )⊤ − 𝑊 ⊤
𝑖 (𝓁)
⎤
⎢ 𝚛1 𝑖(𝓁)
⎥
⎢ ⋆ −2𝑆𝚖 ⊤ −𝑊⊤
𝑆𝚖 𝛯(𝓁) ⎥ ≤ 𝟎, (23)
⎢ 𝚛2 𝑖(𝓁) ⎥
⎢ ⋆ ⋆ 2
−𝜂𝜇𝚛(𝓁) ⎥
⎣ ⎦
− (1 − 𝜏1 )𝛿 + 𝜏2 𝜂 < 0, (24)

and
[ ]
−𝑃̃𝑖 𝑈̃ ⊤ 𝐿⊤(𝑟)
≤ 𝟎, (25)
⋆ −𝜌2(𝑟)

are verified for ∀𝓁 = 1, … , 𝑛𝑢 and ∀𝑟 = 1, … , 𝜅. Then, the control law given by

̂
𝑢(𝑘) = 𝐾(𝛼(𝑘))𝑥(𝑘) + 𝐾(𝛼(𝑘))𝑥(𝑘)
̂ = K(𝛼(𝑘))𝜉(𝑘),

with control gain matrices computed as

∑
𝑁
K(𝛼(𝑘)) = 𝛼(𝑘)(𝑖) 𝑋𝑖 𝑈̃ −1 , (26)
𝑖=1

guarantees that

1. the region 𝑎 =  (𝜂) ⊆ 𝑎 is a region of asymptotic stability for every initial condition belonging to it whenever 𝜔 = 𝟎;
2. the trajectories of the closed-loop system do not leave the set 𝑎 for every initial state inside it, for all 𝑘 ≥ 0 and 𝛼(𝑘) ∈ , for
𝜔 ≠ 𝟎 with 𝜔 ∈ ;
√ √
3. an upper limit of the induced 𝓁2 ∕𝓁∞ gain such that ‖𝑧(𝑘)‖2 ≤ 𝛾(‖𝜔(𝑘)‖2 + 𝑉 (𝜉(0))) ∀𝑘 → ∞ is given by 𝛾.

The estimate of the region of attraction 𝑎 can be obtained by replacing 𝛼𝑖 (𝑘) by their respective expressions depending on 𝑥(𝑘), which
leads to the constraints

𝜉(𝑘)⊤ 𝑃 (𝛼(𝑘))𝜉(𝑘) ≤ 𝜂 −1 = (𝑃 (𝛼(𝑘)), 𝜂). (27)

Remark 2. When considering the LPV case, the region of attraction can be obtained by the level set  (𝜂) defined as
⋂ ⋂
 (𝜂) = (𝑃 (𝛼(𝑘)), 𝜂) = (𝑃𝑖 , 𝜂). (28)
𝛼(𝑘)∈ 𝑖∈1,…,𝑁

This can be seen as a particular condition of [53, Lemma 2, with 𝑔 = 1] that recovers [54] for a polyquadratic Lyapunov function.
By this lemma, the estimate of the region of attraction 𝑎 is computed through finite-dimensional conditions using the level
sets intersection of 𝑃𝑖 , ∀𝑖 = 1, … , 𝑁. Otherwise, in the quasi-LPV case, the weighting functions 𝛼(𝑘) should be replaced by their
expressions depending on 𝑥(𝑘) into (27) to obtain a less conservative estimate of 𝑎 . For details, please see [55, pp. 373, Example
1].

7
L.A.L. Oliveira et al. Journal of the Franklin Institute 361 (2024) 106915

Proof. The local stability can be investigated considering the following polyquadratic Lyapunov function candidate:

𝑉 (𝜉(𝑘)) = 𝜉(𝑘)⊤ 𝑃 (𝛼(𝑘))𝜉(𝑘), (29)

with
∑
𝑁
𝑃 (𝛼(𝑘)) = 𝛼(𝑘)(𝑖) 𝑃𝑖 > 𝟎, (30)
𝑖=1

and 𝛼(𝑘) ∈  must satisfy the sequence set 𝛺 given in (9). Hence, the conditions (21)–(25) must ensure that
𝜆1 (𝜉(𝑘)) ≤ 𝑉 (𝜉(𝑘)) ≤ 𝜆2 (𝜉(𝑘)),
(31)
𝛥𝑉 (𝜉(𝑘)) = 𝑉 (𝜉(𝑘 + 1)) − 𝑉 (𝜉(𝑘)) ≤ −𝜆3 (𝜉(𝑘)),
are satisfied for 𝜉(𝑘) ∈ 𝑎 ⊆ R𝑛+𝑛𝑢 , with class- functions 𝜆𝑖 (𝜉(𝑘)), 𝑖 ∈ {1, 2, 3}, verified for the sequences of 𝛼(𝑘) ∈ 𝛺. Such
functions are given later in this proof. Thus, admitting that the inequality (21) is feasible, we have the nonsingular matrices 𝑃̃𝑖 > 𝟎,
𝑆𝚖 > 𝟎, and 𝑆𝚛 > 𝟎. Moreover, 𝑈̃ is a full-rank matrix due to the negativity of the block (2, 2) in (21). The closed-loop system’s
local stability can be ensured as follows: First, replace 𝑃̃𝑖 , 𝑋𝑖 , 𝑆𝚖 , 𝑆𝚛 , 𝑊𝚖𝑖 , 𝑊𝚛1 𝑖 , and 𝑊𝚛2 𝑖 , by 𝑈̃ ⊤ 𝑃𝑖 𝑈̃ , K𝑖 𝑈̃ , 𝑇𝚖−1 , 𝑇𝚛−1 , 𝐺𝚖𝑖 𝑈̃ ,
𝐺𝚛1 𝑖 𝑈̃ , and 𝐺𝚛2 𝑖 𝑆𝚖 , respectively. Next, multiply the result by 𝛼(𝑘)(𝑖) , 𝛼(𝑘 + 1)(𝑗) , 𝛼(𝑘) ∈  and sum it up for 𝑖, 𝑗 = 1, … , 𝑁 to get an
inequality depending on the weighting function at the instants 𝑘 and 𝑘 + 1. Thanks to the regularity of 𝑈̃ , one can apply a similarity
{ }
transformation by pre-multiplying the resulting inequality by 𝛶 ⊤ = 𝚍𝚒𝚊𝚐 𝑈̃ −⊤ , 𝑈̃ −⊤ , 𝑇𝚖 , 𝑇𝚛 , 𝐈, 𝐈 and post-multiplying it by 𝛶 . In the
sequence, apply the Schur complement to reduce the inequality dimensions by eliminating the last row and column. Subsequently,
[ ]⊤
pre-multiply the resulting inequality by 𝜈 ⊤ and post-multiply it by 𝜈, with 𝜈 ⊤ = 𝜉(𝑘)⊤ 𝜉(𝑘 + 1)⊤ 𝛹𝚖 (𝑘)⊤ 𝛹𝚛 (𝑘)⊤ 𝜔(𝑘)⊤ . The
resulting expression depends on the system’s augmented state and can be related to the Lyapunov candidate function by denoting
𝛥𝑉 (𝜉(𝑘)) = 𝜉(𝑘 + 1)⊤ 𝑃 (𝛼(𝑘 + 1))𝜉(𝑘 + 1) − 𝜉(𝑘)⊤ 𝑃 (𝛼(𝑘))𝜉(𝑘), and from (15) replacing C𝑧 (𝛼(𝑘))𝜉(𝑘) + D𝑧 (𝛼(𝑘))𝜔(𝑘) by 𝑧(𝑘), which yields

𝛥𝑉 (𝜉(𝑘)) + (1 − 𝜏1 )(𝜉(𝑘)⊤ 𝑃 (𝛼(𝑘))𝜉(𝑘) − 𝜂 −1 ) − 2𝛹𝚖 (𝑘)⊤ 𝑇𝚖 𝛹𝚖 (𝑘) − 2𝛹𝚖 (𝑘)⊤ 𝑇𝚖 𝐺𝚖 (𝛼(𝑘))𝜉(𝑘)

−2𝛹𝚛 (𝑘)⊤ 𝑇𝚛 𝛹𝚛 (𝑘) − 2𝛹𝚛 (𝑘)⊤ 𝑇𝚛 𝐺𝚛1 (𝛼(𝑘))𝜉(𝑘) − 2𝛹𝚛 (𝑘)⊤ 𝑇𝚛 𝐺𝚛2 (𝛼(𝑘))𝛹𝚖 (𝑘)
+𝜏2 (𝛿 −1 − 𝜔(𝑘)⊤ 𝜔(𝑘)) + 𝛾 −1 𝑧(𝑘)⊤ 𝑧(𝑘) ≤ 0. (32)

where 𝜏1 and 𝜏2 come as multipliers from the S-procedure.

Thus, with (21) verified and assuming that 𝜉(𝑘) belongs to sets S(𝜇𝚖 ) and S(𝜇𝚛 ), the conditions of Lemma 1 are fulfilled (which
is demonstrated in the sequel) and then (32) can be used to compute 𝜆1 (𝜉(𝑘)) = ℎ1 ‖𝜉(𝑘)‖, with ℎ1 = min𝑖=1,…,𝑁 𝚎𝚒𝚐(𝑈̃ −⊤ 𝑃̃𝑖 𝑈̃ −1 ),
𝜆2 (𝜉(𝑘)) = ℎ2 ‖𝜉(𝑘)‖, with ℎ2 = max𝑖=1,…,𝑁 𝚎𝚒𝚐(𝑈̃ −⊤ 𝑃̃𝑖 𝑈̃ −1 ). Additionally, considering 𝜆3 (𝜉(𝑘)) = ℎ3 ‖𝜉(𝑘)‖, with a small enough ℎ3 > 0,
one gets that Eq. (31) is verified, thus ensuring that 𝑉 (𝜉(𝑘)) is a polyquadratic Lyapunov function. Consequently, under the control
law (11), the state trajectories of the closed-loop system (13) remain bounded for any amplitude-bounded disturbance signal that
satisfies (2). Furthermore, Eq. (32) also ensures that there is an upper bound on the 𝓁2 ∕𝓁∞ gain between the disturbance and Eq. (14)
is verified under null initial conditions, which means N = 0. Thus, the closed-loop performance is guaranteed for the 𝓁2 ∕𝓁∞ gain
for all admissible 𝛼(𝑘) ∈ .
Assuming the feasibility of (22), the inclusion of the contractive set  (𝜂) in the region defined by S𝚖 ∩ S𝚛 can be demonstrated
as follows: First, we replace 𝑃̃𝑖 , 𝑋𝑖 , and 𝑊𝚖𝑖 , by 𝑈̃ ⊤ 𝑃𝑖 𝑈̃ , K𝑖 𝑈̃ , and 𝐺𝚖𝑖 𝑈̃ , respectively, in (22). In the sequel, we multiply the
resulting inequality by 𝛼(𝑘)(𝑖) , 𝛼(𝑘) ∈ , and sum it up for 𝑖 = 1, … , 𝑁. Then, we pre- and post-multiply the parameter-dependent
{ }
inequality by 𝛶 = 𝚍𝚒𝚊𝚐 𝑈̃ −⊤ , 1 and its transpose to apply Schur’s complement. By pre- and post-multiplying it by 𝜉(𝑘)⊤ and 𝜉(𝑘),
we get:
−2
𝜉(𝑘)⊤ (K(𝛼(𝑘)) − 𝐺𝚖 (𝛼(𝑘)))⊤
(𝓁)
(𝜂𝜇𝚖(𝓁) )(K(𝛼(𝑘)) − 𝐺𝚖 (𝛼(𝑘)))(𝓁) 𝜉(𝑘) − 𝜉(𝑘)⊤ 𝑃 (𝛼(𝑘))𝜉(𝑘) ≤ 0, ∀𝓁 = 1, … , 𝑛𝑢 ,

which implies that (𝑃 (𝛼(𝑘)), 𝜂) ⊂ S𝚖 , as in (17). Then, to guarantee the inclusion in S𝚛 , we assume the feasibility of (23) and follow
similar steps: we replace 𝑃̃𝑖 , 𝑋𝑖 , 𝑆𝚖 , 𝑊𝚖𝑖 , 𝑊𝚛1 𝑖 , and 𝑊𝚛2 𝑖 , by 𝑈̃ ⊤ 𝑃𝑖 𝑈̃ , K𝑖 𝑈̃ , 𝑇𝚖−1 , 𝐺𝚖𝑖 𝑈̃ , 𝐺𝚛1 𝑖 𝑈̃ , and 𝐺𝚛2 𝑖 𝑆𝚖 in (23), respectively;
next we multiply the resulting inequality by 𝛼(𝑘)(𝑖) , 𝛼(𝑘) ∈  and sum it up for 𝑖 = 1, … , 𝑁 to get a parameter-dependent inequality
that, thanks the convexity, is satisfied for all 𝛼(𝑘) ∈ . Next, we apply the congruence transformation pre- and post-multiplying it
{ } [ ]
by 𝚍𝚒𝚊𝚐 𝑈̃ −⊤ , 𝑇𝚖 , 1 and its transpose. With Schur’s complement and pre- and post-multiply the result by 𝜗⊤ = 𝜉(𝑘)⊤ 𝛹𝚖 (𝑘)⊤ and
𝜗, we get
( [ ])
−2 𝑃 (𝛼(𝑘)) 𝟎
𝜗⊤ 𝛩⊤ (𝜂𝜇𝚛(𝓁) )𝛩 − 𝜗 ≤ 0, ∀𝓁 = 1, … , 𝑛𝑢 ,
𝟎 𝟎
[ ]
with 𝛩 = (𝛯̂ + 𝛯K(𝛼(𝑘)) − 𝐺𝚛1 )(𝓁) (𝛯 − 𝐺𝚛2 )(𝓁) , 𝓁 = 1, … , 𝑛𝑢 . Therefore, we can conclude that if 𝜉(0) belongs to 𝑎 ≡  (𝜂), then
it is contained in S𝚖 ∩ S𝚛 and will remain inside such a set for all disturbances signals belonging to (2).
Lastly, assuming that (25) is verified, we: replace 𝑃̃𝑖 by 𝑈̃ ⊤ 𝑃𝑖 𝑈̃ , multiply the resulting inequality by 𝛼(𝑘)(𝑖) and sum it up for
{ }
𝑖 = 1, … , 𝑁. With 𝚍𝚒𝚊𝚐 𝑈̃ −⊤ , 1 as a congruence transformation matrix, applying Schur’s complement, and pre- and post-multiplying
the resulting inequality by 𝜉(𝑘)⊤ and 𝜉(𝑘), we get

𝜉(𝑘)⊤ 𝐿⊤
(𝑟)
(𝜂𝜌−2
(𝑟)
)𝐿(𝑟) 𝜉(𝑘) − 𝜉(𝑘)⊤ 𝑃 (𝛼(𝑘))𝜉(𝑘) ≤ 0, ∀𝑟 = 1, … , 𝜅.

Therefore, we can conclude that  (𝜂) ⊆ S𝑥 , ensuring the local stability in the validity domain of the polytopic model. Consequently,
we have  (𝜂) ⊆ (S𝑥 ∩ S𝚖 ∩ S𝚛 ), concluding the proof. □

8
L.A.L. Oliveira et al. Journal of the Franklin Institute 361 (2024) 106915

Remark 3. It is important to highlight the significance of the variable 𝜏1 , which governs the contractivity rate of the Lyapunov
function. By incorporating such a performance specification into the closed-loop system, we can ensure the inequality 𝛥𝑉 (𝜉(𝑘)) +
(1 − 𝜏1 )𝑉 (𝜉(𝑘)) ≤ 0, which is further simplified to 𝑉 (𝜉(𝑘 + 1)) − 𝜏1 𝑉 (𝜉(𝑘)) ≤ 0 with 𝜏1 ∈ (0, 1]. This criterion implies that a smaller
value of 𝜏1 leads to faster system convergence.

When considering null input disturbance, i.e., for 𝜔 = 𝟎, we can get the following particular case, recovering the conditions
contributed in [56]:

Corollary 1. Assume that there exist positive definite and symmetric matrices 𝑃̃𝑖 ∈ R(𝑛+𝑛𝑢 )×(𝑛+𝑛𝑢 ) , diagonal matrices 𝑆𝚖 , 𝑆𝚛 ∈ R𝑛𝑢 ×𝑛𝑢 ,
matrices 𝑋𝑖 , 𝑊𝚖𝑖 , 𝑊𝚛1 𝑖 ∈ R𝑛𝑢 ×(𝑛+𝑛𝑢 ) , 𝑊𝚛2 𝑖 ∈ R𝑛𝑢 ×𝑛𝑢 , and 𝑈̃ ∈ R(𝑛+𝑛𝑢 )×(𝑛+𝑛𝑢 ) such that inequalities (22), (23), (25), and

⎡−𝑃̃𝑖 −𝐻̃ ⊤ A⊤ ⊤ ⊤
𝑖 − 𝑋𝑖 B1 −𝑊𝚖𝑖⊤ −𝑊𝚛⊤ 𝑖 ⎤
1
⎢ ⎥
⎢ ⋆ 𝑃̃𝑗 + 𝑈̃ + 𝑈̃ ⊤ −B1 𝑆𝚖 −B2 𝑆𝚛 ⎥
⎢ ⎥ < 𝟎, (33)
⎢ ⋆ ⋆ −2𝑆𝚖 −𝑊𝚛⊤ 𝑖 ⎥
2
⎢ ⎥
⎣ ⋆ ⋆ ⋆ −2𝑆𝚛 ⎦
are verified for 𝑖, 𝑗 = 1, … , 𝑁, 𝓁 = 1, … , 𝑛𝑢 , and 𝑟 = 1, … , 𝜅. Then the control gain matrices given by (26) yield the control law (11)
which guarantees asymptotic stability in the closed-loop system for initial conditions within the estimated region of attraction denoted as
𝑎 =  (𝜂).

Remark 4. It is worth mentioning that Theorem 1 and Corollary 1 conditions can be cast in the robust control framework whenever
the parameter 𝛼(𝑘) is unavailable online. Thus, it is enough to get a robust version of these conditions with 𝑋𝑖 = 𝑋, 𝑖 = 1, … , 𝑁 in
both conditions. In such a case the control law (11) can be simplified as

𝑢(𝑘) = 𝐾𝑥(𝑘) + 𝐾̂ 𝑥(𝑘)

̂ = K𝜉(𝑘),
̂ correspond to the robust version of that in (26) and can be computed as
where the gain K (or equivalently 𝐾 and 𝐾)

K = 𝑋 𝑈̃ −1 .

4.1. Optimization procedures

Based on Theorem 1, we can obtain sufficient solutions for Problems 1 and 2. However, it is possible to design controllers that
optimize specific objectives, such as the region of attraction estimate size, the allowable disturbance amplitude, or the disturbance
rejection. In the following, we present convex optimization procedures to achieve these objectives.
Maximization of 𝑎 : Assuming a maximum tolerable amplitude bound for the disturbance signal, 𝛿 −1 , we can find a control
law that maximizes the region of attraction estimate, 𝑎 , while ensuring the corresponding closed-loop trajectories initiated within
it are bounded.
Hence, the size of 𝑎 can be enlarged using the following convex optimization approach:
⎧ min 𝚝𝚛𝚊𝚌𝚎(𝑅)
⎪
⎪𝑃̃𝑖,𝑗 , 𝑆𝚖 , 𝑆𝚛 , 𝑋𝑖 , 𝛾
⎪ ̃
⎪𝑊𝚖 , 𝑊𝚛1 , 𝑊𝚛2 , 𝑈
1 ∶ ⎨ (34)
⎪subject to: LMIs (21)–(25), and
[ ]
⎪
⎪ −𝑅 𝐈
≥ 0,
⎪ ⋆ (𝑃̃𝑖 + 𝑈̃ + 𝑈̃ ⊤ )
⎩
∀𝑖, 𝑗 = 1, … , 𝑁, 𝓁 = 1, … , 𝑛𝑢 , and 𝑟 = 1, … , 𝜅, where the positive definite matrix 𝑅 ∈ R𝑛 concerns the ellipsoidal set (𝑅, 𝜂) ⊆  (𝜂)
[ ]
and 𝐻 = 𝐈𝑛 𝟎𝑛×𝑛𝑢 .

Remark 5. Enhancing the results obtained in 1 by applying the relaxation algorithm described in [57], inspired by the Frank–Wolfe
algorithm. This approach increases the size of the estimated regions of attraction when calculated through (28).

Maximization of the allowable disturbance: Another possibility is to maximize the amplitude of the allowed disturbance while
maintaining system stability. In this case, the goal is to design the control gain vectors that maximize the amplitude of admissible
perturbations 𝜔(𝑘) ∈ , ensuring that the trajectories inside 𝑎 do not leave it. For this purpose, the following convex optimization
procedure can be applied:
⎧ min 𝛿
⎪ 𝑃̃ , 𝑆 , 𝑆 , 𝑋 , 𝛾
⎪ 𝑖,𝑗 𝚖 𝚛 𝑖
2 ∶ ⎨𝑊 , 𝑊 , 𝑊 , 𝑈̃ , 𝜂 (35)
⎪ 𝚖 𝚛1 𝚛2
⎪subject to: LMIs (21)–(25),
⎩

9
L.A.L. Oliveira et al. Journal of the Franklin Institute 361 (2024) 106915

∀𝑖, 𝑗 = 1, … , 𝑁, 𝓁 = 1, … , 𝑛𝑢 , and 𝑟 = 1, … , 𝜅.
Minimization of the 𝓁2 ∕𝓁∞ gain: Lastly, we can minimize the induced 𝓁2 ∕𝓁∞ gain between the disturbance signal and the
regulated or performance output, which corresponds to maximizing the disturbance rejection for a given amplitude disturbance
bound. Therefore, the 𝓁2 ∕𝓁∞ gain can be minimized by the following convex optimization procedure:
⎧ min 𝛾
⎪ 𝑃̃ , 𝑆 , 𝑆 , 𝑋
⎪ 𝑖,𝑗 𝚖 𝚛 𝑖
3 ∶ ⎨𝑊 , 𝑊 , 𝑊 , 𝑈̃ , 𝜂 (36)
⎪ 𝚖 𝚛1 𝚛2
⎪subject to: LMIs (21)–(25),
⎩
∀𝑖, 𝑗 = 1, … , 𝑁, 𝓁 = 1, … , 𝑛𝑢 , and 𝑟 = 1, … , 𝜅.

5. Numerical examples

We present 4 examples to illustrate the effectiveness of the proposed design controller conditions. The first one explores the
estimated region of attraction of a polytopic LPV model. Additionally, we investigated the changes in such a region if the model is
interpreted as a T–S fuzzy one. An interesting feature in this example is to show that the proposed approach matches the conditions
to apply the Algorithm 1 in [57], allowing for larger estimates of the region of attraction. The second example investigates the
tracking problem in a polytopic LPV model where the exogenous signal is a reference to be followed. In this example, we explore
the 𝓁2 ∕𝓁∞ gain and demonstrate that our approach leads to a more performant closed-loop system than the literature results. The
third example concerns the use of our approach in the context of slew rate specification for electronic devices. We close this section
by validating our methodology through real-time experiments conducted on a nonlinear coupled tank system, providing concrete
evidence of its efficacy and feasibility for real-world implementation.

5.1. Estimated region of attraction of LPV and T-S Fuzzy models

Consider a T–S fuzzy model, borrowed from [58], subject to control inputs limited in rate and magnitude as described by (15)
with matrices
[ ] [ ] [ ] [ ]
1 −𝜎 1 𝜎 5+𝜎 5−𝜎
𝐴1 = , 𝐴2 = , 𝐵1 = , 𝐵2 = , (37)
−1 −0.5 −1 −0.5 2𝜎 −2𝜎
and 𝜎 > 0. The bounds for rate and magnitude saturation, given by 𝜇𝚖 = 0.4 and 𝜇𝚛 = 0.2 respectively, are symmetric as discussed in
the actuator modeling, Section 2. Additionally, we assume the presence of an ideal rate limiter with 𝛯 = 1. The proposed model is
[ ]
valid for |𝑥1 (𝑘)| ≤ 𝜎. According to Remark 1, we can characterize the region S𝑥 as stated in (6), with 𝐿 = 1 0 0 and 𝜌 = 𝜎. We
assume null input disturbance (𝜔 = 𝟎). Thus, we aim to evaluate the size of the region of attraction 𝑎 . Due to the characteristics
of the model, it represents a typical LPV system, enabling the weighting functions 𝛼𝑖 (𝑘), 𝑖 = 1, 2, to assume arbitrary sequence values
within 𝛺, as long as 𝛼(𝑘) ∈  in (8). Accepting a value of 𝜎 = 0.3, by utilizing Corollary 1 (which recovers the conditions in [42])
along with the optimization procedure 1 given in (34), we obtain the controller gains:
[ ] [ ]
𝐾1 𝐾2 = −0.1872 0.0435 −0.8981 −0.1887 −0.0624 −0.8130 , (38)

and matrices associated with the polyquadratic Lyapunov function:

] ⎡
20.2347 −2.5778 55.3227 21.6573 2.2637 54.9110 ⎤
[
𝑃1 𝑃2 = ⎢−2.5778 4.5102 −20.1163 2.2637 3.8310 8.3013 ⎥ . (39)
⎢ ⎥
⎣ 55.3227 −20.1163 345.7926 54.9110 8.3013 293.4345⎦
By using (39), we can determine an estimate for the region of attraction 𝑎 =  (1). Such an estimate involves the intersection of
ellipses characterized by 𝑃𝑖 , where 𝑖 = 1, 2, as outlined in (28). However, it is possible to further expand this region by employing
an algorithm inspired by Frank–Wolfe methods [57, Algorithm 1]. We have run the [57, Algorithm 1] 8 interactions, leading the
difference between the objective function on the two last interactions to fall below 10−4 . Consequently, we obtain the controller
gain:
[ ] [ ]
𝐾1 𝐾2 = −0.1905 0.0521 −0.9722 −0.1905 −0.0584 −0.8694 , (40)

and the Lyapunov matrices:

] ⎡
19.8947 −3.1811 53.2379 21.8870 3.2360 51.5366 ⎤
[
𝑃1 𝑃2 = ⎢−3.1811 1.7724 −20.3602 3.2360 1.3520 14.8350 ⎥ . (41)
⎢ ⎥
⎣ 53.2379 −20.3602 324.5559 51.5366 14.8350 247.5603⎦
Fig. 2 shows a cross-section of the estimated region of attraction  at 𝑥̂ = 𝟎 in the phase plane for two cases: on the left, we
have the cross-section after applying the optimization procedure 1 , and on the right, the achieved region after 8 iterations of [57,
Algorithm 1], as applied in this contribution. The area of the intersection of regions related to (39) (solid black line, left plot) is
equal to 0.2889, while its counterpart on the right-hand side of Fig. 2 is 0.3534, representing an increase of 22.31%.

10
L.A.L. Oliveira et al. Journal of the Franklin Institute 361 (2024) 106915

Fig. 2. Cut of the estimate region of attraction 𝑎 at 𝑥̂ = 𝟎 obtained with optimization procedure 1 given in (34) (left) and the enhanced result achieved
with the use of [57, Algorithm 1] (right). The intersection area has been augmented by 22.31% in the right plot.

Table 1
Comparison of the estimated (intersection) areas achieved with optimization procedure 1 and with the application of [57,
Algorithm 1] for different values of 𝜎.
𝜎 1 1 with the application of [57, Algorithm 1] Area augmentation
area 𝑎 area 𝑎 (%)
0.999 0.7373 0.7535 2.68
0.5 0.5990 0.6237 4.13
0.3 0.2889 0.3534 22.31
0.1 0.0635 0.1153 81.59

Table 1 allows for the analysis of area enlargement with respect to the 𝜎 value (column 1). The results are obtained using the
optimization procedure 1 (column 2) and the application of [57, Algorithm 1], employing the same stopping criteria as before: a
reduction of the objective function by less than 10−4 . It can be observed that, as the 𝜎 value decreases, resulting in the convergence
of polytope vertices, the effectiveness of applying [57, Algorithm 1] becomes increasingly apparent.
Therefore, using [57, Algorithm 1] shows a simple and worthwhile tool to maximize the estimate of the region of attraction for
the present proposal.

Remark 6. The presented example assesses the feasibility of our conditions for 𝜎 < 1. In [58], an alternative approach, which
relaxes the feasibility conditions, is proposed by addressing a double-sum term in the closed-loop system. However, our modeling
approach, employed to characterize the rate saturation, eliminates such an issue. Moreover, the work in [58] does not address
or consider any form of actuator saturation. Hence, our intention is not to compare the results in terms of the feasibility of the
parameter 𝜎 with the findings of that study, but rather to investigate the effects of saturation on the closed-loop stability.

In conclusion, let us consider that the membership function 𝛼(𝑘) ∈  depends on the states, tailoring the model into a T–S fuzzy
or quasi-LPV approach, as discussed in the previous sections. Following [58], we assume 𝛼1 (𝑘) = (𝜎 +𝑥1 (𝑘))∕(2𝜎) and 𝛼2 (𝑘) = 1−𝛼1 (𝑘).
Consequently, the estimated region of attraction obtained through (27) results in 𝜉(𝑘)⊤ (𝑃1 𝛼1 (𝑘) + 𝑃2 𝛼2 (𝑘))𝜉(𝑘) ≤ 1. Incorporating the
membership functions into determining the region 𝑎 allows us to capture more information about the system, which could result in
a larger area. The cut of 𝑎 at 𝑥̂ = 0 for this case with 𝜎 = 0.999 is shown in Fig. 3: the reader can see the trajectories starting at the
boundary of the intersected region (blue lines), which converge asymptotically to the origin. On the contrary, unstable trajectories
with initial conditions outside 𝑎 are depicted (red lines). Moreover, the area of the intersection of the attraction region in this cut
is equal to 0.9331, an increase of 23.8% compared to the best result in Table 1 when the polytopic model is assumed of LPV-type.
Furthermore, despite the stable trajectories leaving the cut of the region of attraction’s estimate, they do not leave the contractive
set 𝑎 for 𝛼 ∈ 𝛺. This claim can be verified in Fig. 4 where the Lyapunov function 𝑉 (𝜉(𝑘)) for each stable trajectory is presented,
monotonically decreasing and less than 1.

5.2. Reference tracking of an LPV model

The objective of this example is to illustrate the performance of the proposed method for the control of a saturating system
under amplitude disturbances and compare the achievements with those reported in the literature. We consider the lateral motion
control of an F-4 fighter aircraft, as presented in [59] and further investigated in [46]. The aircraft actuators are subject to rate and

11
L.A.L. Oliveira et al. Journal of the Franklin Institute 361 (2024) 106915

Fig. 3. Cut of the estimate region of attraction for 𝜎 = 0.999 at 𝑥̂ = 0 and assuming membership functions 𝛼1 (𝑘) = (𝜎 + 𝑥1 (𝑘))∕(2𝜎) and 𝛼2 (𝑘) = 1 − 𝛼1 (𝑘).

Fig. 4. The temporal evolution of the Lyapunov function for the trajectories with initial conditions located on the boundary of the region of attraction, 𝑎 ,
illustrating that all trajectories remain in such a set.

magnitude saturation effects. The Eqs. (7) can describe the F-4 fighter aircraft time-varying model with matrices

⎡0.6740 1.3676 −4.1392 −0.0463 −0.1089 0.7915 ⎤

⎢0.0130 0.4646 + 0.2𝛥(𝑘) 1.6521 0.0173 −0.1049 −0.0431⎥
⎢ ⎥
0.0034 −0.3894 0.4932 0.0151 0.0592 0.0247 ⎥
𝐴=⎢ ,
⎢0.4141 0.2624 −1.2543 0.9917 0.0030 0.3213 ⎥
⎢ 0 0 0 0 0.0067 0 ⎥ ⎥
⎢
⎣ 0 0 0 0 0 0.0821 ⎦ (42)
[ ]⊤
0.0596 −1.1801 0.2764 0.0727 1.9865 0
𝐵= ,
3.2129 −0.2088 0.0509 0.6627 0 1.8358
[ ] [ ]
0 0 1 0 0 0 0
𝐶𝑧 = , 𝐷𝑧 = ,
0 0 0 1 0 0 0

where the uncertain parameter |𝛥(𝑘)| ≤ 1 varies arbitrarily over time and the matrix 𝐶𝑧 selects as outputs the side-slip and
roll angles from the aircraft. The aircraft actuators are under symmetric rate and magnitude saturation, with bounds given by
[ ]⊤ [ ]⊤
𝜇𝚖 = 0.7854 0.5236 and 𝜇𝚛 = 0.6545 0.5673 . We assume that the actuator’s parameters are given by 𝛯 = 𝚍𝚒𝚊𝚐{1, 1}, i.e., an
ideal rate limiter, leading to the augmented description given in (15)–(16).
To align with [46], we consider the reference tracking framework by defining the signal error as 𝑒(𝑘) = 𝑟(𝑘) − 𝑦(𝑘),̂ where 𝑟(𝑘)
[ ]
represents the reference signal and 𝑦(𝑘) ̂ = C(𝛼(𝑘))𝜉(𝑘) denotes the system’s output with C(𝛼(𝑘)) = 𝐶(𝛼(𝑘)) 𝟎 . Thus, we introduce
the integral control action over the errors by augmenting the system (15)–(16) as follows:
( )
̄ + 1) = A
𝜉(𝑘 ̄ (𝛼(𝑘)) + B̄ 1 K
̄ (𝛼(𝑘)) 𝜉(𝑘)
̄ + B̄ 1 𝛹𝚖 (𝑘) + B̄ 2 𝛹𝚛 (𝑘) + B̄ 𝑤 (𝛼(𝑘))𝑟(𝑘),
𝑦(𝑘)
̄ ̄
= C(𝛼(𝑘)) ̄
𝜉(𝑘), (43)
𝑧(𝑘)
̄ ̄ 𝑧 (𝛼(𝑘))𝜉(𝑘)
=C ̄ +D ̄ 𝑧 (𝛼(𝑘))𝑟(𝑘),

12
L.A.L. Oliveira et al. Journal of the Franklin Institute 361 (2024) 106915

Fig. 5. System’s outputs (blue lines) tracking the references signals (black dashed-lines). A change in the side-slip angle causes a deviation in the roll angle due
to the system dynamics, but the integral controller acts to correct it.

with
[ ] [ ] [ ] [ ]
A(𝛼(𝑘)) 𝟎 B B 𝟎
̄
A(𝛼(𝑘)) = , B̄ 1 = 1 , B̄ 2 = 2 , B̄ 𝑤 (𝛼(𝑘)) = ,
−C(𝛼(𝑘)) 𝐈 𝟎 𝟎 𝐈 (44)
[ ] [ ] [ ]
̄
C(𝛼(𝑘)) = C(𝛼(𝑘)) 𝟎 , C ̄ 𝑧 (𝛼(𝑘)) = −W C(𝛼(𝑘)) W , D𝑧 (𝛼(𝑘)) = W .
2 1 2
[ ]⊤
Thus, we have the extended state vector 𝜉(𝑘) ̄ = 𝜉(𝑘)⊤ 𝑒(𝑘)⊤ and, by consequence, the controller gain matrices with K(𝛼(𝑘)) ̄ =
[ ]
̂
𝐾(𝛼(𝑘)) 𝐾(𝛼(𝑘)) 𝐾𝐼 (𝛼(𝑘)) . Observe that 𝐾𝐼 (𝛼(𝑘)) refers to the LPV integral control gains. Moreover, in our approach, we admit
a full matrix 𝐾𝐼 (𝛼(𝑘)) ∈ R2×2 , which is more general than the usual diagonal integral gain. It is important to note that the designer
specifically chose the diagonal weighting matrices W1 and W2 to achieve the desired system performance. Following [46], we choose
the same weights W1 = 0.3 × 𝚍𝚒𝚊𝚐{1∕20, 1∕20} and W2 = 0.7 × 𝚍𝚒𝚊𝚐{1∕20, 1∕10}, to make a fair comparison.
The objective is to guarantee the stability of the time-varying closed-loop system during the setpoint tracking, ensuring a maximal
[ ]⊤
𝓁2 ∕𝓁∞ gain. In that case, we consider a reference signal 𝑟(𝑘) = 𝑟1 (𝑘) 𝑟2 (𝑘) with

⎧ 0.35, 6 ≤ 𝑘 < 41, {

⎪ 0.35, 6 ≤ 𝑘 ≤ 80,
𝑟1 (𝑘) = ⎨−0.35, 41 ≤ 𝑘 ≤ 80, 𝑟2 (𝑘) = (45)
⎪ 0, otherwise.
⎩ 0, otherwise;

Since we are considering 𝑟(𝑘) = 𝜔(𝑘), from (45) and (2), we have 𝛿 −1 = 0.245. We run the optimization procedure 3 in (36) where
the matrices in Theorem 1 were replaced by the corresponding ones in the system description (43)–(44). We performed a coarse
grid search in 𝜏1 and 𝜏2 to obtain the optimum value for the induced 𝓁2 ∕𝓁∞ gain. By setting 𝜏1 = 0.826 and 𝜏2 = 29, we achieved the
√
minimum gain of 𝛾 = 0.0403. In contrast to our proposed method, the same conditions were run with [46, Theorem 6] leading to
√
an optimum gain that is 98% (worst) than ours, i.e., 𝛾 = 0.0799. Therefore, our method’s better performance for this case is clear.

Our approach yields the following gain matrices:

⎡ 0.1354 0.5354 0.0118 0.1748 −0.1089 0.0910 −0.9724 0.3653 0.2507 −0.0263⎤
[ ] ⎢
̄1
K ⎢−0.0853 −0.1989 0.1451 −0.0915 0.0265 −0.0821 0.2362 −0.3850 0.1374 0.0096 ⎥⎥
̄ 2 = ⎢ 0.1227
K 0.7452 0.0523 0.1575 −0.1018 0.0807 −1.0515 0.3192 0.2625 −0.0238⎥
⎢−0.0838 −0.2750 0.1237 −0.0893 0.0318 −0.0786 0.2150 −0.3747 0.1398 0.0093 ⎥⎦
⎣
where the last two columns refer to 𝐾𝐼 (𝛼(𝑘)). Observe that the control gains of this reference tracking approach retain the algebraic
structure of (26), which yields a simple PDC implementation scheme. To evaluate the performance of the discrete-time system (43)
with matrices (42), we conducted simulations using the reference tracking signals given by (45). The evolution of the side-slip and
roll angles of the aircraft, along with their reference signals, are depicted in Fig. 5, which illustrates their behavior throughout the
simulation. It is worth mentioning that the outputs closely adhere to the desired setpoint, as indicated by the reference signals.

13
L.A.L. Oliveira et al. Journal of the Franklin Institute 361 (2024) 106915

Fig. 6. Control signals applied to the F-4 fighter aircraft model, where the saturation bounds are respected.

In particular, a perturbation in the side-slip angle reference signal (𝑟1 (𝑘)) at sample 𝑘 = 40 resulted in oscillations in the roll
angle (𝑦2 (𝑘)). Nevertheless, both outputs are interconnected due to the inherent characteristics of the system dynamics while still
maintaining stability. The time-variation of the parameters related to the uncertainty 𝛥(𝑘) used in the simulations presented in this
example is such that 𝛼1 (𝑘) = | sin(0.1𝑘)| and 𝛼2 (𝑘) = 1 − 𝛼1 (𝑘). For the reader’s reference, Fig. 6 presents the corresponding control
signals where the saturation limits are respected.
This example shows how the proposed method can be used in reference tracking, ensuring respect for the actuator’s saturating
limits. Moreover, the achievements suggest the better performance of our proposal regarding a recent literature method.

5.3. Slew rate selection in electronic circuits

We give a simple example to illustrate how other fields, such as power electronics, can leverage our proposal. In this example,
we investigate the minimal admissible slew rate that a voltage source must attain to ensure circuit stability. Consider the low-pass
filter circuit shown in Fig. 7, borrowed from [60], which is commonly found in power electronic converters. The control signal
is given by 𝑢(𝑡), and the system’s states are represented by 𝑥𝐿 (𝑡) and 𝑥𝐶 (𝑡), the voltage across the inductor and the capacitor. The
time-varying parameters are the load resistance 𝑅(𝑡), filter inductance 𝐿(𝑡), filter capacitance 𝐶(𝑡), and capacitor internal resistance
𝑟𝐶 (𝑡). The model of the circuit is given by

[ ] ⎡ −1 𝑅(𝑡) ⎤[ ]
⎡ 0 ⎤
𝑥̇ 𝐶 (𝑡) ⎢ (𝑅(𝑡) + 𝑟𝐶 (𝑡))𝐶(𝑡) (𝑅(𝑡) + 𝑟𝐶 (𝑡))𝐶(𝑡) ⎥ 𝑥𝐶 (𝑡)
=⎢ ⎥ + ⎢ 1 ⎥ 𝑢(𝑘). (46)
𝑥̇ 𝐿 (𝑡) ⎢ −𝑅(𝑡) −𝑅(𝑡)𝑟𝐶 (𝑡)
⎥ 𝑥𝐿 (𝑡) ⎢ ⎥
⎣ (𝑅(𝑡) + 𝑟𝐶 (𝑡))𝐿(𝑡) ⎣ 𝐿(𝑡) ⎦
(𝑅(𝑡) + 𝑟𝐶 (𝑡))𝐿(𝑡) ⎦
Following [60], we assume that all parameters are subject to uncertainties caused by factors such as measurement inaccuracies
and changes in operational conditions, lying inside the following intervals: 24 Ω ≤ 𝑅(𝑡) ≤ 106 Ω; 0.4 mH ≤ 𝐿(𝑡) ≤ 0.6 mH;
640 μF ≤ 𝐶(𝑡) ≤ 960 μF; and 0.5 mΩ ≤ 𝑟𝐶 (𝑡) ≤ 1.5 mΩ. We use the Euler discretization method on (46) with a sampling time
of 𝑇 = 10−6 s, and with the combinations of maximum and minimal values of the parameters, we obtain the discrete-time polytopic
model as (10) with 𝑁 = 16.
In [60], rate and magnitude input saturations are not considered, meaning that the voltage source 𝑢(𝑡) can provide infinite √
amplitude and power. Here, we address a more realistic issue where the voltage source has a magnitude bound as 𝜇𝑚 = 127 2 V
and look for the minimal allowable rate saturation. We assume null disturbance (𝜔 = 𝟎) in such a case. By using Corollary 1 we
find 𝜇𝑟 = 2.22 × 10−6 V as the minimal slew rate required for the voltage source. It means that, for each sampling period, the voltage
source must be able to increase (or decrease) its output signal by at least 2.22 V/s, to ensure local stability. Therefore, our approach
can be leveraged to study and specify performance limits required in electronic circuits.

14
L.A.L. Oliveira et al. Journal of the Franklin Institute 361 (2024) 106915

Fig. 7. Filter component and load where the parameters are time-varying.

Fig. 8. Diagram of the coupled tank system used to perform a real-time experiment. The water flows from 𝑇1 to 𝑇2 recirculated by a pump motor. Tank 𝑇1 has
a structure inside, yielding nonlinear dynamics.

5.4. Real-time control of a quasi-LPV system

To demonstrate the efficacy of our proposal, we conducted a real-time experiment using a nonlinear tank system operated by
conventional industrial devices, which served as a benchmark in related studies [12,56,61,62]. In the sequel, we present the system’s
description, its quasi-LPV modeling, and finally the controller design with real-time experiments.

5.4.1. System’s description

The experimental setup comprises two tanks, designated as 𝑇1 and 𝑇2 , each with a capacity of 200 L. Fig. 8 provides an illustrative
schematic diagram depicting the system’s configuration. Tanks 𝑇1 and 𝑇2 are interconnected by a manual valve that remains fixed
throughout the experiment. Differential pressure sensors, specifically the Honeywell model 26PCBFA6D, are employed to measure
the water levels ℎ1 (𝑡) and ℎ2 (𝑡) in the tanks. Notably, inside tank 𝑇1 , a nonlinear solid is inserted, assuming a Gaussian curve
revolution shape and made of expanded polystyrene. Consequently, the volume of liquid within tank 𝑇1 is not directly proportional
to its height, introducing significant nonlinear dynamics to the system. Thus, we model this interconnected system as a quasi-LPV
system.
In our experimental configuration, water is supplied from two 400-L reservoirs to tank 𝑇1 (𝑞𝑖𝑛 (𝑡)) using a 1 horsepower motor
pump manufactured by Dancor, model CAM-W6. Subsequently, the fluid flows from tank 𝑇1 to tank 𝑇2 through their connecting
pathway (𝑞12 (𝑡)). Finally, the outflow (𝑞𝑜𝑢𝑡 (𝑡)) in tank 𝑇2 directs the water back to the reservoirs, where the motor pump recirculates
it.
The control system integrates an open-source Python interface [63] that interacts with a Siemens Simatic 𝑆7−300 programmable
logic controller (PLC). The PLC monitors the sensor signals and transmits control commands to a WEG 𝐶𝐹 𝑊 09 frequency inverter
that, in turn, regulates the power supplied to the pump following the decisions made by the controller.

15
L.A.L. Oliveira et al. Journal of the Franklin Institute 361 (2024) 106915

5.4.2. Quasi-LPV modeling

In the sequence, we present the physical modeling of the system, its discretization, and the corresponding quasi-LPV model
cast in a polytopic description as presented in (7). Next, we use our proposal to design a controller considering the actuator’s rate
and magnitude saturation limits. Finally, we conduct a real-time level control experiment that illustrates a full application of our
methods.
The mathematical modeling of the level dynamics is based on mass balance equations, yielding
𝑞𝑖𝑛 (𝑡) ℎ1 (𝑡) − ℎ2 (𝑡)
ℎ̇ 1 (𝑡) = − ,
𝐴1 (ℎ1 (𝑡)) 𝑅12 (ℎ1 (𝑡), ℎ2 (𝑡))𝐴1 (ℎ1 (𝑡))
(47)
ℎ1 (𝑡) − ℎ2 (𝑡) 𝑞 (𝑡)
ℎ̇ 2 (𝑡) = − 𝑜𝑢𝑡 ,
𝑅12 (ℎ1 (𝑡), ℎ2 (𝑡))𝐴2 𝐴2
where ℎ1 (𝑡) represents the level of tank 𝑇1 and ℎ2 (𝑡) represents the level of tank 𝑇2 , the flow functions are experimentally determined
̂ cm3 ∕s, where the input flow is controlled by the signal 𝑢(𝑡),
as follows: 𝑞𝑖𝑛 (𝑡) = 15.0953𝑢(𝑡) ̂ and 𝑞𝑜𝑢𝑡 (𝑡) = −0.1758ℎ2 (𝑡)2 + 20.6553ℎ2 (𝑡) +
3
222.5698 cm ∕s, representing the output flow in tank 𝑇2 . The flow resistance between the tanks is modeled as 𝑅12 (ℎ1 (𝑡), ℎ2 (𝑡)) =
(8.4874 + 1.4297(ℎ1 (𝑡), ℎ2 (𝑡)) − 0.0324(ℎ1 (𝑡), ℎ2 (𝑡))2 ) × 10−3 . The transversal area of tank 𝑇2 is constant and given by 𝐴2 = 3019 cm2 ,
while the corresponding area in tank 𝑇1 exhibits a nonlinear dependency on ℎ1 (𝑡) as follows:
⎛ ((ℎ1 (𝑡) − 8) × 10−2 − 𝜇)2 ⎞
3𝑟 ⎜ cos (2.5𝜋((ℎ1 (𝑡) − 8) × 10−2 − 𝜇)) −
2𝜎 2
⎟ 4 2
√ ⎟ × 10 cm (48)
5 ⎜⎜
𝐴1 (ℎ1 (𝑡)) = 2.7𝑟 − ×𝑒
𝜎 2𝜋 ⎟
⎝ ⎠
with 𝑟 = 0.31 𝑚, 𝜎 = 0.55, and 𝜇 = 0.4. Following previous studies, we use a sampling time of 𝑇 = 4 s to discretize model (47) with
the Euler discretization, resulting in:
[ ] ⎡1 − 𝑧1 (𝑘)𝑧3 (𝑘) 𝑧1 (𝑘)𝑧3 (𝑘) ⎤[ ] [ ]
ℎ1 (𝑘 + 1) ℎ (𝑘) 𝑇 𝐾𝑏 𝑧3 (𝑘)
=⎢ 𝑧1 (𝑘) 𝑧1 (𝑘) + 𝑧2 (𝑘) ⎥ 1 + 𝑢(𝑘)
̂ (49)
ℎ2 (𝑘 + 1) ⎢ 1− ⎥ ℎ2 (𝑘) 0
⎣ 𝐴2 𝐴2 ⎦
with the functions given by
𝑇 𝑇 𝑞𝑜𝑢𝑡 1
𝑧1 (𝑘) = , 𝑧 (𝑘) = , and 𝑧3 (𝑘) = .
𝑅12 (ℎ1 (𝑡), ℎ2 (𝑡)) 2 ℎ2 (𝑘) 𝐴1 (ℎ1 (𝑘))
We consider the operational intervals 22.5 ≤ ℎ1 ≤ 46.5 cm and 12.46 ≤ ℎ2 ≤ 32.46 cm and consequently, the ranges of functions 𝑧𝑖 (𝑘)
are obtained as
180.164 ≤ 𝑧1 (𝑘) ≤ 471.141,
83.682 ≤ 𝑧2 (𝑘) ≤ 139.394, and
36.606 × 10−5 ≤ 𝑧3 (𝑘) ≤ 150.595 × 10−5 .

Employing the sector nonlinearity approach [49], we can rewrite 𝑧𝑖 (𝑘) = 𝑤1𝑖 (𝑘)𝑧̄ 𝑖 + 𝑤2𝑖 (𝑘)𝑧𝑖 , 𝑖 = 1, … , 3, with 𝑤1𝑖 (𝑘) ≥ 0, 𝑤2𝑖 (𝑘) ≥ 0,
̄
𝑤1𝑖 (𝑘) + 𝑤2𝑖 (𝑘) = 1, 𝑤1𝑖 (𝑘) = (𝑧̄ 𝑖 − 𝑧𝑖 (𝑘))∕(𝑧̄ 𝑖 − 𝑧𝑖 ), and 𝑤2𝑖 (𝑘) = (𝑧𝑖 (𝑘) − 𝑧𝑖 )∕(𝑧̄ 𝑖 − 𝑧𝑖 ) and obtain a discrete-time quasi-LPV model, whose
̄ ̄ ̄
vertices concerning (7) with 𝑁 = 8 are given by:
[ ]
𝐴1 𝐴2 𝐴3 𝐴4
=
𝐴5 𝐴6 𝐴7 𝐴8
⎡1 − 𝑧̄ 1 𝑧̄ 3 𝑧1 𝑧3
̄ 𝑧̄ +𝑧
1 − 𝑧1 𝑧̄ 3 𝑧1 𝑧̄ 3 1 − 𝑧1 𝑧3 𝑧1 𝑧3 1 − 𝑧1 𝑧̄ 3 𝑧1 𝑧̄ 3 ⎤
⎢ 𝑧̄ 1 𝑧̄1 ̄ 𝑧 +𝑧 𝑧̄1 ̄ ̄ 𝑧̄ +𝑧̄ 𝑧̄1 ̄ 𝑧 +𝑧̄
1 − ̄ 1𝐴 ̄ 2 ̄ 1 − ̄ 1𝐴 ̄ 2 ̄ 1 − ̄ 1𝐴 2 ̄ 1 − ̄ 1𝐴 2 ⎥
⎢ 𝐴2 2 𝐴2 2 𝐴2 2 𝐴2 2 ⎥, (50)
⎢1 − 𝑧̄ 1 𝑧3 𝑧̄ 1 𝑧3 1 − 𝑧̄ 1 𝑧̄ 3 𝑧̄ 1 𝑧̄ 3 1 − 𝑧̄ 1 𝑧3 𝑧̄ 1 𝑧3 1 − 𝑧̄ 1 𝑧̄ 3 𝑧̄ 1 𝑧̄ 3 ⎥
⎢ 𝑧̄ 1 ̄ 𝑧̄̄ +𝑧 𝑧̄ 1 𝑧̄ +𝑧 𝑧̄ 1 ̄ 𝑧̄̄ +𝑧̄ 𝑧̄ 1 𝑧̄ 1 +𝑧̄ 2 ⎥
⎣ 𝐴 1 − 1𝐴 ̄ 2 𝐴2
1 − 1𝐴 ̄ 2 𝐴2
1 − 1𝐴 2 𝐴2
1− 𝐴 ⎦
2 2 2 2 2
]⊤ [ [ ]⊤
with 𝐵1 = 𝐵3 = 𝐵5 = 𝐵7 = 𝑇 𝐾𝑏 𝑧3 0 , 𝐵2 = 𝐵4 = 𝐵6 = 𝐵8 = 𝑇 𝐾𝑏 𝑧̄ 3 0 , and the time-varying parameter vector given
[ 1 1 1 ̄2 2 2
]
by 𝛼(𝑘) = 𝑤1 (𝑘)𝑤2 (𝑘)𝑤3 (𝑘) ⋯ 𝑤1 (𝑘)𝑤2 (𝑘)𝑤3 (𝑘) . Therefore, we have a discrete-time polytopic description of the experimental
setup, allowing the application of the proposed methods in this paper.

5.4.3. Controller design and experiment

We must select an equilibrium point within the operating system’s range to run the experiment. Thus, with a control signal of
𝑢∗ = 38%, we experimentally obtained ℎ∗1 = 34.51 cm and ℎ∗2 = 22.46 cm. As a consequence of this operational point, the state vector
[ ]⊤
can be expressed as 𝑥(𝑘) = ℎ1 (𝑘) − ℎ∗1 ℎ2 (𝑘) − ℎ∗2 , and the control signal as 𝑢(𝑘) = 𝑢(𝑘)̂ − 𝑢∗ . Following the reference tracking
[ ]
framework presented in (43)–(44), the output is defined as 𝑦(𝑘) = 𝐶𝑥(𝑘), with 𝐶 = 1 0 , indicating a reference tracking of the
ℎ1 (𝑘) level. Rate and magnitude saturation are symmetrically bounded by 𝜇𝚛 = 1 and 𝜇𝚖 = 15, respectively, meaning that 𝑢(𝑘)
can vary of ±15% around the operational point 𝑢∗ = 38% and −1% ≤ 𝑢(𝑘 + 1) − 𝑢(𝑘) = 𝑢(𝑘 ̂ + 1) − 𝑢(𝑘)
̂ ≤ 1% or, in a temporal basis,
the actuator rate is constrained by ±0.25%∕s, since the sampling period is 𝑇 = 4 s. Thus, these choices mean that the control signal
𝑢(𝑘) can vary from 23% to 53%, i.e., −15% ≤ 𝑢(𝑘) ̂ ≤ 15%, and its change between two samples is constrained to ±1%. Note that
such a variation limit is a constraint inserted in the frequency inverter device that actuates the motor pump, avoiding, for instance,

16
L.A.L. Oliveira et al. Journal of the Franklin Institute 361 (2024) 106915

Fig. 9. Real-time reference track experiment performed with the designed controller: (a) Top plot: time response of level ℎ1 (𝑘) (blue lines) with the respective
reference signal (black dashed-lines). Bottom plot: time response of the control signal during the experiment. (b) Top plot: behavior of the Lyapunov function,
demonstrating that the function never exceeds the unitary value. Bottom plot: time-varying parameters of the quasi-LPV model throughout the experiment.

electrical and mechanical efforts on the motor pump. Additionally, we assume an ideal rate limiter with 𝛯 = 1. The validity domain
of the model (50) is given by |𝑥1 (𝑘)| ≤ 12 and |𝑥2 (𝑘)| ≤ 10, meaning that the level ℎ1 (𝑘) may vary from 22.51 cm up to 46.51 cm. Such
a choice leads the system to operate in a range where the nonlinearity in the tank 𝑇1 has stronger
[ effects] over the level dynamics.
1 0 0 0 [ ]
Consequently, by using Eq. (6) in Remark 1, we characterize such a S𝑥 -region with 𝐿 = and 𝜌 = 12 10 . Note
0 1 0 0
that, with such a change of coordinates, 𝑟(𝑘) = 0 means 𝑥1 (𝑘) = 0, or equivalently ℎ1 (𝑘) = ℎ∗1 . Therefore, 𝑟(𝑘) denotes the desired
variation of ℎ1 (𝑘) around ℎ∗1 .
Our objective is to obtain the maximum allowable amplitude of 𝑟(𝑘) ∈ , i.e., to find the minimum 𝛿 in (2), concerning
the variation of the reference signal 𝑟(𝑘), such that the state trajectories do not leave the region of attraction. Thus, we used the
optimization
√ procedure 2 given in (35) with 𝜂 = 1 and 𝜏1 = 0.985, yielding 𝜏2 = 3.818 × 10−4 and the maximum bound for 𝑟(𝑘)
of ± 𝛿 = ±6.267 cm around the equilibrium point ℎ∗1 . The resulting controller gains are implemented in the real system using
−1

the PDC structure found in (26) for a reference signal respecting the maximum bounds. The achievements are reported in Fig. 9,
where the left-hand side illustrates the time response of the controlled level ℎ1 (𝑘) (top) and the corresponding control signal sent
to the motor pump (bottom). √ The system starts close to the equilibrium condition, and we chose the reference signal reaching the
estimated limit, i.e., ℎ∗1 ± 𝛿 −1 or equivalently 28.24 ≤ ℎ1 (𝑘) ≤ 40.78 cm. The measurements’ noise comes mainly from the water
oscillation in the tanks. In Fig. 9(b), at the right-hand side, we see the behavior of the parameter-dependent Lyapunov function
(top) and the time-varying parameters of the quasi-LPV controlled system (bottom). As expected, the parameter-dependent Lyapunov
function starts near zero because the system is close to equilibrium. Moreover, it does not violate the unitary value, meaning the
state trajectory remains inside the region of attraction’s estimate, 𝑎 respecting the level set. The bottom part of Fig. 9(b) shows
parameters 𝛼(𝑘) components behavior recovered from the quasi-LPV system.
Fig. 10 shows the experimental data in a phase-plane portrait. The region of attraction 𝑎 has been computed regarding the
time-varying functions, as described in (27). Note that the phase plane is shown in two dimensions; therefore, the correspondent
region of 𝑎 is cut at 𝑥̂ = 0. The black dashed-line region represents the projection of the three-dimensional 𝑎 on the phase plan.
Therefore, we can check that, as stated in Theorem 1, the trajectories of the closed-loop system (blue lines) do not leave the set
𝑎 , for all 𝑘 ≥ 0 once the exogenous signal, 𝑟(𝑘), remains in the set . Because this is a real-time experiment where we explored
the limits of 𝑟(𝑘), the measured signals are subject to noises, justifying some points barely outside the projection region.
In summary, we have used the optimization procedure 2 to get the maximum variation on the reference signal (𝑟(𝑘), the
exogenous signal) such that the closed-loop trajectories remain inside the ensured region of attraction for the nonlinear coupled
tanks modeled as a polytopic quasi-LPV system. The experimental data suggest that our proposal is suitable for industrial applications
and effective.

6. Conclusions

We have provided a convex solution to handle polytopic model representations under saturating actuators and regional modeling
validity. The generality of the polytopic representation proves to be highly advantageous, as it accommodates a wide spectrum
of models, including LPV, quasi-LPV, and T–S fuzzy models. Both rate and magnitude saturations are addressed, leading to a
characterization of the set of initial conditions for which the respective trajectories converge to the equilibrium point. The controller
is based on a state feedback control gain that depends on system parameters, and it implements a parallel distributed compensation
strategy, ensuring the local stabilization of the saturated system. Moreover, we consider the effects of exogenous disturbance signals

17
L.A.L. Oliveira et al. Journal of the Franklin Institute 361 (2024) 106915

Fig. 10. Phase plane portrait with experimental data (blue line), the cut of the region of attraction 𝑎 at 𝑥(𝑘)
̂ = 0 (solid black line), and the projection of 𝑎
(dashed black line), all of them included in the region of validity of the model, S𝑥 (dashed magenta line).

belonging to the class of bounded amplitude signals. In such cases, we estimate the 𝓁2 ∕𝓁∞ gain between the exogenous input and
the regulated or performance output. Consequently, our approach can handle various interesting problems, including reference
tracking. We leverage the convexity of our formulation to propose some optimization procedures, allowing the design of local
stabilizing control gains while maximizing the estimate of the region of attraction, the allowable disturbance, or minimizing the
𝓁2 ∕𝓁∞ gain. We provide a set of examples to illustrate the effectiveness of our approach and to compare it with recent literature
results. Additionally, we illustrate how to use our approach in the context of polytopic LPV or T–S fuzzy modeling. Furthermore, we
have validated our approach in a real-time experiment. The polytopic modeling allows for a quasi-LPV representation of a nonlinear
process, where the actuator is a velocity-varying motor pump with magnitude and slew rate bounds. We plan to extend this research
to new contexts, such as network control systems characterized by constrained information traffic.

CRediT authorship contribution statement

Lucas A.L. Oliveira: Conceptualization, Methodology, Software, Writing – original draft, Formal analysis. Valter J.S. Leite:
Funding acquisition, Supervision, Project administration, Conceptualization, Methodology, Writing – review & editing, Formal
analysis. Luís F.P. Silva: Conceptualization, Methodology, Writing – review & editing, Investigation, Formal analysis. Kevin
Guelton: Conceptualization, Methodology, Writing – review & editing, Investigation, Formal analysis.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared
to influence the work reported in this paper.

Acknowledgments

This work was supported by Centro Federal de Educação Tecnológica de Minas Gerais (CEFET-MG), and Brazilian Agencies
FAPEMIG (APQ-02100-22), CNPq (311891/2022-5), and Project Stic-Amsud/CAPES NetConHybSDP, code 22-STIC-09.

References

[1] A. Karimi, H. Khatibi, R. Longchamp, Robust control of polytopic systems by convex optimization, Automatica 43 (2007) 1395–1402, http://dx.doi.org/
10.1016/j.automatica.2007.01.022.
[2] T. Takagi, M. Sugeno, Fuzzy identification of systems and its applications to modeling and control, IEEE Trans. Syst., Man, Cybern. SMC-15 (1985)
116–132, http://dx.doi.org/10.1109/TSMC.1985.6313399.
[3] J.S. Shamma, J.R. Cloutier, Gain-scheduled missile autopilot design using linear parameter varying transformations, J. Guid. Control Dyn. 16 (1993)
256–263, http://dx.doi.org/10.2514/3.20997.
[4] C. Briat, Linear Parameter-Varying and Time-Delay Systems — Analysis, Observation, Filtering and Control, in: Advances in Delays and Dynamics, Vol. 3,
Springer-Verlag, Berlin Heidelberg, 2015, http://dx.doi.org/10.1007/978-3-662-44050-6.
[5] Oliveira T.G., R.M. Palhares, V.C.S. Campos, P.S. Queiroz, E.N. Gonçalves, Improved takagi-sugeno fuzzy output tracking control for nonlinear networked
control systems, J. Franklin Inst. 354 (2017) 7280–7305, http://dx.doi.org/10.1016/j.jfranklin.2017.08.042.

18
L.A.L. Oliveira et al. Journal of the Franklin Institute 361 (2024) 106915

[6] S. Tasoujian, S. Salavati, M.A. Franchek, K.M. Grigoriadis, Robust delay-dependent LPV synthesis for blood pressure control with real-time bayesian
parameter estimation, IET Control Theory Appl. 14 (2020) 1334–1345, http://dx.doi.org/10.1049/iet-cta.2019.0651.
[7] P.H. Colmegna, F.D. Bianchi, R.S. Sánchez-Peña, Automatic glucose control during meals and exercise in type 1 diabetes: Proof-of-concept in silico tests
using a switched LPV approach, IEEE Control Syst. Lett. 5 (2021) 1489–1494, http://dx.doi.org/10.1109/LCSYS.2020.3041211.
[8] N. Vafamand, M.M. Arefi, M.H. Khooban, T. Dragičević, F. Blaabjerg, Nonlinear model predictive speed control of electric vehicles represented by linear
parameter varying models with bias terms, IEEE J. Emerg. Sel. Top. Power Electron. 7 (2019) 2081–2089, http://dx.doi.org/10.1109/JESTPE.2018.2884346.
[9] P.S.P. Pessim, M.J. Lacerda, On the robustness of cyber–physical LPV systems under DoS attacks, J. Franklin Inst. 359 (2022) 677–696, http://dx.doi.org/
10.1016/j.jfranklin.2021.01.043.
[10] S. De Souza, C. Tarbouriech, V.J.S. Leite, E.B. Castelan, Co-design of an event-triggered dynamic output feedback controller for discrete-time LPV systems
with constraints, J. Franklin Inst. 359 (2022) 697–718, http://dx.doi.org/10.1016/j.jfranklin.2021.04.028.
[11] M.J. Lacerda, High fidelity LPV systems under constraints, J. Franklin Inst. 359 (2022) 2755–2757, http://dx.doi.org/10.1016/j.jfranklin.2022.02.029.
[12] G.O. Ferreira, L.S. Figueiredo, M.J. Lacerda, V.J.S. Leite, ISS control for continuous-time systems with filtered time-varying parameter and saturating
actuators, Asian J. Control 24 (2022) 2888–2900, http://dx.doi.org/10.1002/asjc.2747.
[13] A.N.D. Lopes, K. Guelton, L. Arcese, V.J.S. Leite, Local sampled-data controller design for T-S fuzzy systems with saturated actuators, IEEE Control Syst.
Lett. 5 (2021) 1169–1174, http://dx.doi.org/10.1109/LCSYS.2020.3019215.
[14] A.H.K. Palmeira, J.M. Gomes da Silva Jr., J.V. Flores, Aperiodic sampled-data control for LPV systems under input saturation, IFAC-PapersOnLine 51
(2018) 130–136, http://dx.doi.org/10.1016/j.ifacol.2018.11.156.
[15] M.M. Morato, M. Jungers, J.E. Normey-Rico, O. Sename, A predictive fault tolerant control method for qLPV systems subject to input faults and constraints,
J. Franklin Inst. 359 (2022) 9129–9167, http://dx.doi.org/10.1016/j.jfranklin.2022.09.011.
[16] M.M. Quadros, V.J.S. Leite, R.M. Palhares, Robust fault hiding approach for T–S fuzzy systems with unmeasured premise variables, Inform. Sci. 589 (2022)
690–715, http://dx.doi.org/10.1016/j.ins.2021.12.129.
[17] S. Xu, H. Wen, X. Wang, Observer-based robust fuzzy control of nonlinear networked systems with actuator saturation, ISA Trans. 123 (2022) 122–135,
http://dx.doi.org/10.1016/j.isatra.2021.05.037.
[18] C. De Souza, V.J.S. Leite, E.B. Castelan, Partially parameter-dependent dynamic output controller with event-triggered for discrete-time saturated LPV
systems, J. Control, Autom. Electr. Syst. 35 (2024) 1–11, http://dx.doi.org/10.1007/s40313-023-01046-w.
[19] J. Yang, Y. Chen, L. Ma, J. Yang, Event-triggered regional stabilization for T–S fuzzy systems subject to communication delays and actuator saturations,
J. Franklin Inst. 361 (2024) 106653, http://dx.doi.org/10.1016/j.jfranklin.2024.106653.
[20] S. Tarbouriech, G. Garcia, J.M. Gomes da Silva Jr., I. Queinnec, Stability and Stabilization of Linear Systems with Saturating Actuators, Springer, London,
UK, 2011, http://dx.doi.org/10.1007/978-0-85729-941-3.
[21] D.H. Klyde, D.T. McRuer, T.T. Myers, Pilot-induced oscillation analysis and prediction with actuator rate limiting, J. Guid., Control Dyn. 20 (1997) 81–89,
http://dx.doi.org/10.2514/2.3998.
[22] H. Duda, Prediction of pilot-in-the-loop oscillations due to rate saturation, J. Guid., Control Dyn. 20 (1997) 581–587, http://dx.doi.org/10.2514/2.4080.
[23] G. Stein, Respect the unstable, IEEE Control Syst. Mag. 23 (2003) 12–25, http://dx.doi.org/10.1109/MCS.2003.1213600.
[24] D. Gacio, J.M. Alonso, J. Garcia, L. Campa, M.J. Crespo, M. Rico-Secades, PWM series dimming for slow-dynamics HPF LED drivers: the high-frequency
approach, IEEE Trans. Ind. Electron. 59 (2012) 1717–1727, http://dx.doi.org/10.1109/TIE.2011.2130503.
[25] N. Baiomy, R. Kikuuwe, An amplitude- and rate-saturated controller for linear plants, Asian J. Control 22 (2020) 77–91, http://dx.doi.org/10.1002/asjc.
1851.
[26] Kapila V., W.M. Haddad, Fixed-structure controller design for systems with actuator amplitude and rate nonlinearities, in: Proceedings of the 37th IEEE
Conference on Decision & Control, Tampa, FL, USA, 1998, pp. 909–914, http://dx.doi.org/10.1109/CDC.1998.760809.
[27] V. Kapila, H. Pan, M.S. de Queiroz, LMI-based control of linear systems with actuator amplitude and rate nonlinearities, in: Proceedings of the 38th IEEE
Conference on Decision & Control, Phoenix, AZ, USA, 1999, pp. 1413–1418, http://dx.doi.org/10.1109/CDC.1999.830166.
[28] Pan H., V. Kapila, LMI-based control of discrete-time systems with actuator amplitude and rate nonlinearities, in: Proceedings of the 2001 American
Control Conference, 2001, pp. 4140–4145, http://dx.doi.org/10.1109/ACC.2001.946388.
[29] J.M. Gomes da Silva Jr., D. Limón, T. Alamo, E.F. Camacho, Dynamic output feedback for discrete-time systems under amplitude and rate actuator
constraints, IEEE Trans. Autom. Control 53 (2008) 2367–2372, http://dx.doi.org/10.1109/TAC.2008.2007521.
[30] J.M. Gomes da Silva Jr., S. Tarbouriech, Antiwindup design with guaranteed regions of stability: An LMI-based approach, IEEE Trans. Autom. Control 50
(2005) 106–111, http://dx.doi.org/10.1109/TAC.2004.841128.
[31] F.A. Bender, J.M. Gomes da Silva Jr., Output feedback controller design for systems with amplitude and rate control constraints, Asian J. Control 14
(2012) 1–5, http://dx.doi.org/10.1002/asjc.483.
[32] S. Galeani, S. Onori, A.R. Teel, L. Zaccarian, A magnitude and rate saturation model and its use in the solution of a static anti-windup problem, Syst.
Control Lett. 57 (2008) 1–9, http://dx.doi.org/10.1016/j.sysconle.2007.06.011.
[33] F. Forni, S. Galeani, L. Zaccarian, Model recovery anti-windup for continuous-time rate and magnitude saturated linear plants, Automatica 48 (2012)
1502–1513, http://dx.doi.org/10.1016/j.automatica.2012.05.019.
[34] S. Tarbouriech, C. Prieur, J.M. Gomes da Silva Jr., Stability analysis and stabilization of systems presenting nested saturations, IEEE Trans. Autom. Control
51 (2006) 1364–1371, http://dx.doi.org/10.1109/TAC.2006.878743.
[35] F. Tyan, D.S. Bernstein, Dynamic output feedback compensation for linear systems with independent amplitude and rate saturations, Int. J. Control 67
(1997) 89–116, http://dx.doi.org/10.1080/002071797224379.
[36] Z. Lin, Semi-global stabilization of linear systems with position and rate-limited actuators, Systems Control Lett. 30 (1997) 1–11, http://dx.doi.org/10.
1016/S0167-6911(96)00082-5.
[37] J.M. Gomes da Silva Jr., S. Tarbouriech, G. Garcia, Local stabilization of linear systems under amplitude and rate saturating actuators, IEEE Trans. Autom.
Control 48 (2003) 842–847, http://dx.doi.org/10.1109/TAC.2003.811265.
[38] A. Bateman, Z. Lin, An analysis and design method for linear systems under nested saturation, Syst. Control Lett. 48 (2003) 41–52, http://dx.doi.org/10.
1016/S0167-6911(02)00246-3.
[39] A. Bateman, Z. Lin, An analysis and design method for discrete-time linear systems under nested saturation, IEEE Trans. Autom. Control 47 (2002)
1305–1310, http://dx.doi.org/10.1109/TAC.2002.800764.
[40] B. Zhou, Analysis and design of discrete-time linear systems with nested actuator saturations, Syst. Control Lett. 62 (2013) 871–879, http://dx.doi.org/
10.1016/j.sysconle.2013.06.012.
[41] A.H.K. Palmeira, J.M. Gomes da Silva Jr., J, S. Tarbouriech, I.M.F. Ghiggi, Sampled-data control under magnitude and rate saturating actuators, Int. J.
Robust Nonlinear Control 26 (2016) 3232–3252, http://dx.doi.org/10.1002/rnc.3503.
[42] L.A.L. Oliveira, M.V.C. Barbosa, L.F.P. Silva, V.J.S. Leite, Exponential stabilization of LPV systems under magnitude and rate saturating actuators, IEEE
Control Syst. Lett. 6 (2022) 1418–1423, http://dx.doi.org/10.1109/LCSYS.2021.3099443.
[43] J.G. Ernesto, E.B. Castelan, W. Lucia, G.A. Franca Dos Santos, Alternative implementation to an incremental output-feedback design approach for
constrained discrete-time parameter-varying systems, in: 5th IFAC Workshop on Linear Parameter Varying Systems LPVS 2022, 2022, pp. 25–30,
http://dx.doi.org/10.1016/j.ifacol.2022.11.285.

19
L.A.L. Oliveira et al. Journal of the Franklin Institute 361 (2024) 106915

[44] I.B. Kucukdemiral, Robust disturbance rejection for discrete-time systems having magnitude and rate bounded inputs, J. Franklin Inst. 357 (2020)
8252–8276, http://dx.doi.org/10.1016/j.jfranklin.2020.06.001.
[45] I.B. Kucukdemiral, H. Yazici, Robust gain-scheduling h∞ control of uncertain continuous-time systems having magnitude- and rate-bounded actuators: An
application of full block S-procedure, J. Franklin Inst. 358 (2021) 8226–8249, http://dx.doi.org/10.1016/j.jfranklin.2021.08.017.
[46] I. Kucukdemiral, X. Han, M.S. Erden, Robust induced 𝓁2 –𝓁∞ optimal control of discrete-time systems having magnitude and rate-bounded actuators, ISA
Trans. 129 (2022) 73–87, http://dx.doi.org/10.1016/j.isatra.2022.02.025.
[47] J. Zhang, W.B. Xie, M.Q. Shen, L. Huang, State augmented feedback controller design approach for T-S fuzzy system with complex actuator saturations,
Int. J. Control Autom. Syst. 15 (2017) 2395–2405, http://dx.doi.org/10.1007/s12555-016-0599-0.
[48] Z. Zhang, J. Li, Y. Niu, J. Song, Local sliding mode control design for T-S fuzzy systems with magnitude and rate limited input, in: 2021 33rd Chinese
Control and Decision Conference, CCDC, IEEE, Kunming, China, 2021, pp. 5818–5823, http://dx.doi.org/10.1109/CCDC52312.2021.9602250.
[49] K. Tanaka, H. Wang, Fuzzy Control Systems Design and Analysis: A Linear Matrix Inequality Approach, John Wiley & Sons, New York, NY, 2001,
http://dx.doi.org/10.1002/0471224596.
[50] J.S. Shamma, An overview of LPV systems, in: J. Mohammadpour, C.W. Scherer (Eds.), Control of Linear Parameter Varying Systems with Applications,
Springer, New York, 2012, pp. 3–26, http://dx.doi.org/10.1007/978-1-4614-1833-7_1.
[51] E.D. Sontag, Input to state stability: Basic concepts and results, in: P. Nistri, G. Stefani (Eds.), Nonlinear and Optimal Control Theory, Springer, 2008, pp.
163–220, http://dx.doi.org/10.1007/978-3-540-77653-6_3.
[52] J.M. Gomes da Silva Jr., S. Tarbouriech, Stability regions for linear systems with saturating controls, in: Proceedings of the 5th European Control Conference,
Karlsruhe, Germany, 1999, pp. 1–6, http://dx.doi.org/10.23919/ECC.1999.7099761.
[53] L.S. Figueiredo, M.J. Lacerda, V.J.S. Leite, Design of saturating state feedback control laws for discrete-time linear parameter varying systems through
homogeneous polynomial parameter-dependent functions, Int. J. Robust Nonlinear Control 31 (2021) 6585–6601, http://dx.doi.org/10.1002/rnc.5625.
[54] M. Jungers, E.B. Castelan, Gain-scheduled output control design for a class of discrete-time nonlinear systems with saturating actuators, Syst. Control Lett.
60 (2011) 169–173, http://dx.doi.org/10.1016/j.sysconle.2010.11.006.
[55] V.J.S. Leite, E.B. Castelan, L.F.P. Silva, C. de Souza, Control of constrained discrete-time systems with time-varying state delay, in: M.H. Khooban,
T. Dragicevic (Eds.), Control Strategy for Time-Delay Systems, in: Emerging Methodologies and Applications in Modelling, Identification and Control,
Academic Press, 2021, pp. 347–381, http://dx.doi.org/10.1016/B978-0-12-820599-0.00015-X.
[56] L. Oliveira, A. Bento, V.J.S. Leite, F. Gomide, Evolving granular feedback linearization: Design, analysis, and applications, Appl. Soft Comput. 86 (2020)
105927, http://dx.doi.org/10.1016/j.asoc.2019.105927.
[57] L.F.P. Silva, V.J.S. Leite, E.B. Castelan, M. Klug, K. Guelton, Local stabilization of nonlinear discrete-time systems with time-varying delay in the states
and saturating actuators, Inform. Sci. 518 (2020) 272–285, http://dx.doi.org/10.1016/j.ins.2020.01.029.
[58] T.M. Guerra, L. Vermeiren, LMI-based relaxed nonquadratic stabilization conditions for nonlinear systems in the Takagi–Sugeno’s form, Automatica 40
(2004) 823–829, http://dx.doi.org/10.1016/j.automatica.2003.12.014.
[59] Y. Ochi, K. Kanai, Pole placement in optimal regulator by continuous pole-shifting, J. Guid. Control Dyn. 18 (1995) 1253–1258, http://dx.doi.org/10.
2514/3.21538.
[60] V.F. Montagner, R.C.L.F. Oliveira, V.J.S. Leite, P.L.D. Peres, LMI approach for ∞ linear parameter-varying state feedback control, IEE Proc. — Control
Theory Appl. 152 (2005) 195–201, http://dx.doi.org/10.1049/ip-cta:20045117.
[61] A.N.D. Lopes, V.J.S. Leite, L.F.P. Silva, K. Guelton, Anti-windup TS fuzzy PI-like control for discrete-time nonlinear systems with saturated actuators, Int.
J. Fuzzy Syst. 20 (2020) 46–61, http://dx.doi.org/10.1007/s40815-019-00781-0.
[62] M.M. Quadros, I.V. Bessa, V.J.S. Leite, R.M. Palhares, Fault tolerant control for linear parameter varying systems: An improved robust virtual actuator
and sensor approach, ISA Trans. 104 (2020) 356–369, http://dx.doi.org/10.1016/j.isatra.2020.05.010.
[63] A.C. Sousa, V.J.S. Leite, I. Rubio Scola, Affordable control platform with MPC application, Stud. Inform. Control 27 (2018) 265–274, http://dx.doi.org/
10.24846/v27i3y201802.

20