You are on page 1of 11

Automatica 125 (2021) 109437

Contents lists available at ScienceDirect

Automatica
journal homepage: www.elsevier.com/locate/automatica

Robust control Minkowski–Lyapunov functions✩


Saša V. Raković
Beijing Institute of Technology, Beijing, China

article info a b s t r a c t

Article history: This article introduces, and studies, the robust control Minkowski–Lyapunov inequality, which is
Received 17 January 2020 a natural generalization of the recently introduced control Minkowski–Lyapunov inequality. The
Received in revised form 21 August 2020 generalization considers the setting of parametrically uncertain linear discrete time systems. The
Accepted 4 October 2020
necessary and sufficient conditions for the characterization and existence of Minkowski functions that
Available online 4 January 2021
verify the robust control Minkowski–Lyapunov inequality are derived. The article also characterizes,
Keywords: and establishes topological properties of, the corresponding robustly stabilizing set-valued control map.
Linear systems © 2020 Elsevier Ltd. All rights reserved.
Parametric uncertainty
Stability
Minkowski functions
Robust control Lyapunov inequality

1. Introduction Mayne, 2009; Sontag, 1998) and excellent survey paper (Giesl &
Hafstein, 2015) provide a more complete overview of the related
Lyapunov functions (Lyapunov, 1992) play a major role in research activity. Naturally, theory, computation and applications
the stability analysis of dynamics. Likewise, control Lyapunov of control Lyapunov functions lead to more precise research
functions (Artstein, 1983) are fundamentally important for the programs when concrete classes of dynamical systems are stud-
design of stabilizing control laws for dynamical systems. The ied. In this sense, stability analysis of parametrically uncertain
foundations of control Lyapunov functions have been created dynamics, in which the uncertainty is merely bounded, is best
by a profound Artstein’s theorem (Artstein, 1983), while theory, performed with the aid of robust Lyapunov functions. In the
computation and applications of control Lyapunov functions have same vein, control synthesis for parametrically uncertain control
formed an important research field. From a topological point of systems under set-membership description of the uncertainty
view, the main challenges are to establish necessary and sufficient can be advantageously executed with the aid of robust control
conditions for the existence, and to characterize theoretically Lyapunov functions. Indeed, the robust stability analysis of para-
correct classes, of control Lyapunov functions for a given class metrically uncertain linear discrete time dynamics (Barabanov,
of control systems (Barabanov, 1988; Blanchini & Miani, 2008; 1988; Blanchini & Miani, 2008; Goebel et al., 2006; Lazar, 2010)
Clarke et al., 1997, 1998; Molchanov & Pyatnitskii, 1986; Sontag, and the robust stabilization of parametrically uncertain linear
1989, 1998; Sontag & Sussmann, 1996). From a numerical point discrete time systems (Blanchini & Miani, 2008; Molchanov &
of view, the major chores are the efficient computation, represen- Pyatnitskii, 1986; Nguyen et al., 2011) are topical research prob-
tation and utilization of such control Lyapunov functions (Baier lems. Solutions to these two problems can be obtained by using,
et al., 2018; Boyd et al., 1994; Giesl & Hafstein, 2015; Goebel et al., respectively, strictly convex quadratic robust Lyapunov functions,
2006). The beneficial applications of control Lyapunov functions and robustly stabilizing linear control functions together with
include, inter alia, the design of stabilizing control (Blanchini & strictly convex quadratic robust control Lyapunov functions. Such
Miani, 2008; Freeman & Kokotović, 2008; Krstić & Deng, 1998) solutions are computationally very convenient, due to their re-
as well as consistently improving optimal and model predictive ductions to linear matrix inequalities (Boyd et al., 1994), but are
control (Rawlings & Mayne, 2009). bound to be very conservative (Barabanov, 1988; Blanchini &
The comprehensive monographs (Blanchini & Miani, 2008; Miani, 2008; Goebel et al., 2006; Raković, 2020c). Indeed, it is
Clarke et al., 1998; Freeman & Kokotović, 2008; Rawlings & known that the use of polyhedral Lyapunov functions (Barabanov,
1988; Blanchini & Miani, 2008; Lazar, 2010) offers an improved
✩ The material in this paper was not presented at any conference. This paper solution to the robust stability analysis problem, while the inter-
was recommended for publication in revised form by Associate Editor Denis
polation based control strategies (Nguyen et al., 2011) result in
Arzelier under the direction of Editor Sophie Tarbouriech. a considerably more flexible solution to the robust stabilization
E-mail address: sasa.v.rakovic@gmail.com. problem.

https://doi.org/10.1016/j.automatica.2020.109437
0005-1098/© 2020 Elsevier Ltd. All rights reserved.
S.V. Raković Automatica 125 (2021) 109437

1.1. Theme, novelty and relevance provides a novel, and complete, topological characterization of
the robustly stabilizing set-valued control map associated with a
The theme of our study is the robust control Lyapunov in- robust control Minkowski–Lyapunov function. Theorem 5 estab-
equality with a priori prescribed decrease condition lishes a novel necessary and sufficient condition for the existence
of robust control Minkowski–Lyapunov functions. Corollaries 2
∀x, ∃u : sup V (f (x, u)) + ℓ(x, u) ≤ V (x). (1.1) and 3 are novel results that provide considerable refinements of
f ∈F
Theorem 3 and Theorem 4 for polyhedral Minkowski decrease
In the above robust control Lyapunov inequality, the constraint and robust control Minkowski–Lyapunov functions.
f ∈ F reflects the set-membership representation of the para- The developed framework is primarily and strongly driven by
metric uncertainty of the controlled state transition map (x, u) ↦ → its utilization for the design of robustly stabilizing control laws
f (x, u) generating the system x+ = f (x, u) for which x and u are as well as the synthesis of consistently improving finite horizon
the current state and control, and x+ is the uncertain successor robust optimal control and stabilizing robust model predictive
state, (x, u) ↦ → ℓ(x, u) is an a priori given decrease function, and control. The application of the developed framework to the ro-
x ↦ → V (x) that verifies the above inequality is a robust control bust stabilization of parametrically uncertain linear discrete time
Lyapunov function. In general, the uncertainty bounding set F systems is outlined in Section 6.1. The robust control Minkowski–
and the decrease and robust control Lyapunov functions, ℓ (·, ·) Lyapunov inequality is of paramount importance for consistent
and V (·), are subject to mild technical requirements in order to improvement of finite horizon robust optimal control and sta-
render the above inequality a regular robust control Lyapunov bility of robust model predictive control. Indeed, this inequality
inequality. represents a conventional terminal stabilizing condition (Rawl-
The specific object of our study is the above robust control ings & Mayne, 2009) in robust model predictive control without
Lyapunov inequality (1.1) over the space of Minkowski functions constraints and a crucial component of the terminal stabilizing
of compact convex sets that contain the origin in the interior conditions (Rawlings & Mayne, 2009) in robust model predictive
within the context of parametrically uncertain linear discrete control under constraints. Thus, our framework provides novel
time systems. Minkowski function of a compact convex set that results within the context of finite horizon robust optimal control
contains the origin in the interior is formally defined in Sec- and stabilizing robust model predictive control, the utility of
tion 1.2. The class of parametrically uncertain linear discrete which is discussed in Sections 6.2 and 6.3 . An illustrative example
time systems is formally defined in Section 2, and it represents in Section 7.3 highlights that the proposed framework can be also
a mathematical model in which a realized matrix pair (A, B) of applied within the setting of constrained parametrically uncertain
the uncertain controlled state transition map (x, u) ↦ → Ax + Bu linear discrete time systems.
is allowed to be any element of the convex hull of the finite Analysis in what follows is made as close as possible to the one
collection {(Ai , Bi ) : i ∈ I } of known matrix pairs. The decrease performed in the predecessor article (Raković, 2020a) concerned
function ℓ (·, ·) is Minkowski function of a compact convex set with deterministic linear discrete time systems; This is done
that contains the origin in the interior, and it is referred to as deliberately so as to utilize results derived in Raković (2020a) as
the Minkowski decrease function. The considered robust control a natural base for the generalizations reported herein.
Lyapunov functions V (·) are Minkowski functions of compact
1.2. Article structure and contributions overview
convex sets that contain the origin in the interior, and they are
referred to as the robust control Minkowski–Lyapunov functions.
Section 2 provides a concise formulation of the robust con-
Such a setting of the robust control Lyapunov inequality (1.1) is
trol Minkowski–Lyapunov inequality, and it details the setting
referred to as the robust control Minkowski–Lyapunov inequality,
utilized for the subsequent technical analysis.
and it is formalized in Section 2.
Section 3 derives an equivalent, and conveniently compact, re-
The above terminology is compatible with the notions of
formulation of the robust control Minkowski–Lyapunov inequal-
Minkowski–Lyapunov functions (Raković, 2017, 2020b; Raković
ity, which provides a foundational platform for establishing our
& Lazar, 2014), robust Minkowski–Lyapunov functions (Raković,
main contributions in Section 4 and their further refinements in
2020c), control Minkowski–Lyapunov functions (Raković, 2020a)
Section 5.
and Minkowski–Bellman functions (Raković, 2019). These recent Section 4 establishes the necessary and sufficient conditions
and advanced frameworks have been obtained via a methodolog- for the characterization and existence of robust control
ical fusion of Lyapunov and control Lyapunov functions (Artstein, Minkowski–Lyapunov functions. The necessary and sufficient
1983; Lyapunov, 1992) on the one side and Minkowski and condition characterizing robust control Minkowski–Lyapunov
support functions of compact convex sets that contain the origin functions is provided in terms of a conveniently compact set
in the interior (Rockafellar, 1970; Schneider, 1993) on the other inclusion. This condition is a novel contribution, and it also
side. admits a relevant interpretation in the jargon of set invariance,
To the best of the author’s knowledge, the above described ro- which reveals a correspondence of robust control Minkowski–
bust control Minkowski–Lyapunov inequality has not been studied Lyapunov functions and robust control invariant sets. The nec-
in the open literature. The robust control Minkowski–Lyapunov essary and sufficient condition for the existence of robust control
inequality is a natural generalization of the deterministic con- Minkowski–Lyapunov functions is the strict stabilizability of the
trol Minkowski–Lyapunov inequality introduced, and studied, considered parametrically uncertain linear discrete time sys-
recently in Raković (2020a). The study of the robust control tem, which confirms the topological compatibility and theoretical
Minkowski–Lyapunov inequality demonstrates that the design exactness of the proposed framework. Section 4 also demon-
of robustly stabilizing control laws for parametrically uncertain strates that any robust control Minkowski–Lyapunov function
linear discrete time systems is nonconservative when it is guided is a regular robust control Lyapunov function, and that the ro-
with Minkowski functions of compact convex sets that contain bustly stabilizing set-valued control map associated with a robust
the origin in the interior, and it is numerically potent when such control Minkowski–Lyapunov function is an outer semicontinu-
sets are also polytopic. ous, locally bounded, compact-valued, convex-valued, positively
This article reports several novel technical results. Theorem 3 homogeneous of the first degree set-valued map.
delivers a novel necessary and sufficient condition that character- Section 5 documents a refinement of the fundamental prop-
izes all robust control Minkowski–Lyapunov functions. Theorem 4 erty of robust control Minkowski–Lyapunov functions and the
2
S.V. Raković Automatica 125 (2021) 109437

topological structure of the associated robustly stabilizing set- 2. Robust control Minkowski–Lyapunov inequality
valued control map within a more structured polytopic setting.
The necessary and sufficient condition for the characterization We consider parametrically uncertain linear discrete time sys-
of robust control Minkowski–Lyapunov functions is refined to a tems, as modeled by
number of simpler polyhedral set inclusions, while the robustly
x+ ∈ F (x, u) (2.1)
stabilizing set-valued control map is demonstrated to be polyhe-
n m
dral, polytopic-valued, and Lipschitz continuous with respect to where x ∈ R and u ∈ R are the current state and control,
the Hausdorff distance. respectively, and x+ ∈ Rn is the successor state. The controlled
Section 6 outlines relevant control applications, and Section 7 state transition map F (·, ·) : Rn × Rm ⇒ Rn is specified, for all
concludes the article by discussing the computational aspects and (x, u) ∈ Rn × Rm , by
by delivering an academic illustration. F (x, u) := {Ax + Bu : (A, B) ∈ M}, with
Basic Notation and Facts: The Minkowski sum of nonempty sets M := convh(MD ) and
X and Y in Rn is denoted by
MD := {(Ai , Bi ) ∈ Rn×n × Rn×m : i ∈ I }, (2.2)
X ⊕ Y := {x + y : x ∈ X , y ∈ Y }.
where I = {1, 2, . . . , p} is a finite index set.
The image of a nonempty set X under a matrix of compatible The main object of our study is the robust control Minkowski–
dimensions (or a scalar) M is denoted by Lyapunov inequality

M X := {Mx : x ∈ X }. ∀x ∈ R n , ∃u ∈ Rm :
max g(S , Ax + Bu) + g(V , (x, u)) ≤ g(S , x). (2.3)
A compact convex subset X of Rn that contains the origin is (A,B)∈M
a C -set in Rn . A compact convex subset X of Rn that contains
Minkowski function g(V , (·, ·)) of a proper C -set V in Rn+m is
the origin in its interior is a proper C -set in Rn . The Minkowski
referred to as the Minkowski decrease function. Minkowski func-
function g(X , ·) of a proper C -set X in Rn is given, for all y ∈ Rn , tion g(S , ·) of a proper C -set S in Rn that verifies the robust
by control Minkowski–Lyapunov inequality (2.3) is referred to as a
g(X , y) := min{γ : y ∈ γ X , γ ≥ 0}. robust control Minkowski–Lyapunov function.
γ We also study the set-valued map K(S , ·) : Rn ⇒ Rm
The support function h(X , ·) of a nonempty closed convex subset specified, for all x ∈ Rn , by
X of Rn is specified, for all y ∈ Rn , by
K(S , x) := {u ∈ Rm : max g(S , Ax+Bu)+g(V , (x, u)) ≤ g(S , x)}.
(A,B)∈M
h(X , y) := sup{y x : x ∈ X }.
T
x (2.4)
The polar set X ∗ of a set X in Rn with 0 ∈ X is For any robust control Minkowski–Lyapunov function g(S , ·), the
set-valued map K(S , ·) is the robustly stabilizing set-valued con-
X ∗ := {y ∈ Rn : ∀x ∈ X , xT y ≤ 1}.
trol map, since its values K(S , x) characterize all robustly stabi-
lizing control actions guaranteeing the required decrease at any
Theorem 1 (Schneider, 1993, Theorems 1.6.1 and 1.7.6). (i) For any
x ∈ Rn .
proper C -set X in Rn , the polar set X ∗ is a proper C -set in Rn and
The robust control Minkowski–Lyapunov inequality (2.3) is
(X ∗ )∗ = X .
analyzed for an arbitrary, but fixed, matrix collection {(Ai , Bi ) :
(ii) Minkowski function g(X , ·) of a proper C -set X in Rn is equal to
i ∈ I } and a proper C -set V . Unless stated otherwise, the standing
support function h(X ∗ , ·) of the polar set X ∗ of the set X , i.e., assumption herein is:
∀y ∈ R n , g(X , y) = h(X ∗ , y).
Assumption 1. (Ai , Bi ) ∈ Rn×n × Rn×m for each i ∈ I where
Proposition 1 (Rockafellar & Wets, 2009, Corollary 11.24). (i) For I := {1, 2, . . . , p} is a finite index set. V is a proper C -set in Rn+m .
any nonempty closed convex set X in Rn and any matrix M ∈ Rm×n , S is a proper C -set in Rn .

The main objectives of this article are to:


∀y ∈ R m , h(M X , y) = h(X , M T y).
(i) derive a necessary and sufficient condition that a proper C -
(ii) For any nonempty closed convex sets X and Y in Rn ,
set S needs to satisfy in order for its Minkowski function
∀z ∈ Rn , h(X ⊕ Y , z) = h(X , z) + h(Y , z). g(S , ·) to verify the robust control Minkowski–Lyapunov
inequality (2.3),
The convex hull of a set X is denoted by convh(X ); For a finite (ii) establish the relevant topological properties of the robustly
set of matrices MD = {Mi ∈ Rn×(n+m) : i ∈ I }, stabilizing set-valued control map K(S , ·), and
∑ ∑ (iii) provide a necessary and sufficient condition for the exis-
convh(MD ) = { λi Mi : ∀i ∈ I , λi ≥ 0, λi = 1}. tence of robust control Minkowski–Lyapunov functions.
i∈I i∈I

Px := (I O), I ∈ Rn×n and O ∈ Rn×m , denotes the (x, u) ↦ → x 3. Equivalent reformulation


projection matrix.
Our first step is to derive an equivalent, and a more con-
Conventions, Complete Terminology and Proofs: We do not dis- venient, form of the robust control Minkowski–Lyapunov in-
tinguish between a variable and its vectorized form in algebraic equality (2.3). This chore can be accomplished by characterizing
expressions. No confusion should arise. The interested reader is explicitly the function (x, u) ↦ → max(A,B)∈M g(S , Ax + Bu) +
referred to Appendix A for a complete overview of the technical def- g(V , (x, u)). To this end, let, for all (x, u) ∈ Rn × Rm ,
initions and terminology utilized throughout this article. Appendix B
provides technically more involved proofs. J(x, u) := max g(S , Ax + Bu) + g(V , (x, u)). (3.1)
(A,B)∈M

3
S.V. Raković Automatica 125 (2021) 109437

Since Minkowski function g(X , ·) of a proper C -set X is convex, the equivalent reformulation of the ordinary control Minkowski–
(A, B) ↦ → g(S , Ax + Bu) + g(V , (x, u)) is a convex function for any Lyapunov inequality studied in Raković (2020a). Because S and R
arbitrary (x, u). Thus, since M = convh({(Ai , Bi ) : i ∈ I }), for all are proper C -sets in Rn and Rn+m , respectively, the relation (3.5)
(x, u) ∈ Rn × Rm , holds true if and only if S ⊆ Px R. The set inclusion S ⊆ Px R is,
hence, a necessary and sufficient condition that a proper C -set S
J(x, u) = max g(S , Ai x + Bi u) + g(V , (x, u)). (3.2)
i∈I in Rn needs to satisfy in order for its Minkowski function g(S , ·)
to verify the robust control Minkowski–Lyapunov inequality (2.3).
As formally shown in Appendix B.1, the following result ex-
This set inclusion characterizes all robust control Minkowski–
tends slightly a helpful characterization of the function (x, u) ↦ →
Lyapunov functions and, hence, it constitutes their fundamental
g(S , Ax + Bu) + g(V , (x, u)) derived originally in Raković (2019),
property. This fact yields our first main result.
and also used in Raković (2020a, Proposition 3) within the deter-
ministic setting.
Theorem 3. Suppose Assumption 1 holds. Minkowski function
g(S , ·) verifies the robust control Minkowski–Lyapunov inequal-
Proposition 2. Under Assumption 1, for any i ∈ I ,
ity (2.3) if and only if
∀(x, u) ∈ R × R ,
n m ⋂( )∗
S ⊆ Px R with R = (Ai Bi )T S ∗ ⊕ V ∗ . (4.1)
g(S , Ai x + Bi u) + g(V , (x, u)) = g(Ri , (x, u))
i∈I
( )∗
for a proper C -set Ri := (Ai Bi )T S ∗ ⊕ V ∗ in Rn+m . The fundamental property of robust control
Minkowski–Lyapunov functions admits a relevant interpretation
Due to (3.2) and Proposition 2, for all (x, u) ∈ R × R , n m
in terminology of set invariance. First, we recall that a nonempty
J(x, u) = max g(Ri , (x, u)), (3.3) set X is a robust control invariant set for the system x+ ∈ F (x, u)
i∈I
with F (·, ·) given by (2.2) and the state dependent constraints
∗ ∗ (x, u) ∈ Z if and only if for all x ∈ X there exists a u such that
( )
where, for all i ∈ I , Ri = (Ai Bi )T S ∗ ⊕ V . The following result,
the proof of which is provided in Appendix B.2, yields the explicit (x, u) ∈ Z and F (x, u) ⊆ X . When the robust control invariant set
characterization of J (·, ·). X and the state dependent constraint set Z are convex, the latter
condition F (x, u) ⊆ X is equivalent to the requirement that for
Proposition 3. For any finite collection {Zi : i ∈ I } of proper all i ∈ I it holds that Ai x + Bi u ∈ X .
C -sets Zi in Rn+m , The generator set S of any robust control Minkowski–
Lyapunov function g(S , ·) is a proper C -set in Rn , which satisfies
∀(x, u) ∈ Rn × Rm , max g(Zi , (x, u)) = g(Z , (x, u)) the derived
i∈I ⋂ ( necessary T ∗
and )sufficient condition S ⊆ Px R with

R = i∈I (A i Bi ) S ⊕ V ∗ . Hence, for all x ∈ S there exists
Zi in Rn+m .

for a proper C -set Z :=
a u ∈ Rm such that (x, u) ∈ R, which, in view of the robust
i∈I

Due to (3.3) and Proposition 3, for all (x, u) ∈ Rn × Rm , control Minkowski–Lyapunov inequality (2.3), ensures a fortiori
that, for all i ∈ I , Ai x + Bi u ∈ S . (This conclusion follows since,
J(x, u) = g(R, (x, u)), (3.4)
for all i ∈ I , g(S , Ai x + Bi u) + g(V , (x, u)) ≤ g(S , x).) Consequently,
∗ ∗
⋂ ( T ∗
)
where R = i∈I Ri with Ri = (Ai Bi ) S ⊕ V for each i ∈ I . the generator set S of any robust control Minkowski–Lyapunov
Summarizing, the preceding technical analysis has revealed the function g(S , ·) is a proper C -set S in Rn , which is a robust
explicit characterization of the function J (·, ·), and it has formally control invariant set for the system x+ ∈ F (x, u) with F (·, ·)
established the following. given⋂ by (2.2)
( and the state)dependent constraints (x, u) ∈ R with

R = i∈I (Ai Bi )T S ∗ ⊕ V ∗ .
Theorem 2. Suppose Assumption 1 holds. On the other hand, if a proper C -set S in Rn is a robust control
invariant set for the system x+ ∈ F (x, u) with F (·, ·) given
∀(x, u) ∈ R × Rm ,
n
by (2.2) constraints (x, u) ∈ R with
⋂ and( the state dependent
max g(S , Ax + Bu) + g(V , (x, u)) = g(R, (x, u)), ∗ ∗
T ∗
)
R = i∈I (Ai B i ) S ⊕ V , then it holds that S ⊆ Px R. In
(A,B)∈M
( )∗ turn, Minkowski function g(S , ·) is a robust control Minkowski–
(Ai Bi )T S ∗ ⊕ V ∗ in Rn+m .

for a proper C -set R := i∈I
Lyapunov function.
In view of the above technical discussion, we have formally
Evidently, with Theorem 2 in mind, the robust control established the following interpretation of the fundamental prop-
Minkowski–Lyapunov inequality (2.3) admits an equivalent, and erty of robust control Minkowski–Lyapunov functions.
a conveniently compact, reformulation as

∀x ∈ Rn , ∃u ∈ Rm : g(R, (x, u)) ≤ g(S , x), (3.5) Corollary 1. Suppose Assumption 1 holds. Minkowski function
g(S , ·) verifies the robust control Minkowski–Lyapunov inequal-
which can be further equivalently rewritten as ity (2.3) if and only if its generator set S is a robust control invariant
∀x ∈ Rn , ∃u ∈ Rm : (x, u) ∈ g(S , x)R, (3.6) set for the system x+ ∈ F (x, u) with F (·, ·) given by (2.2)⋂and
n+m
the
( state )∗ constraints (x, u) ∈ R with R =
dependent i∈I
where the proper C -set R in R is given by (Ai Bi )T S ∗ ⊕ V ∗ .
∗ ∗
⋂(
(Ai Bi )T S ∗ ⊕ V .
)
R := (3.7)
4.2. Robustly stabilizing set-valued control map
i∈I

4. Proper C -generator sets In light of the equivalent reformulation (3.6) of the robust
control Minkowski–Lyapunov inequality (2.3) and the definition
4.1. Fundamental property of the robustly stabilizing set-valued control map K(S , ·) in (2.4),
the topological characterization of the set-valued map K(S , ·)
The equivalent reformulation (3.5) of the robust control follows directly from Raković (2020a, Theorem 3), as formally
Minkowski–Lyapunov inequality (2.3) is form-wise identical to summarized by the following.

4
S.V. Raković Automatica 125 (2021) 109437

Theorem 4. Suppose Assumption 1 holds. Let R be given by (3.7). We have established that robust control Minkowski–Lyapunov
The set-valued map K(S , ·) specified by (2.4) is equivalently given, functions exist if and only if the finite collection {(Ai , Bi ) :
for all x ∈ Rn , by i ∈ I } of matrix pairs is strictly stabilizable. This fact yields the
following.
K(S , x) := {u ∈ Rm : (x, u) ∈ g(S , x)R}. (4.2)

Furthermore, when Minkowski function g(S , ·) verifies the robust Theorem 5. Take any finite collection {(Ai , Bi ) ∈ Rn×n × Rn×m :
control Minkowski–Lyapunov inequality (2.3), the robustly stabi- i ∈ I } of matrix pairs and any proper C -set V in Rn+m . There exists a
lizing set-valued control map K(S , ·) is nonempty relative to Rn , proper C -set S in Rn whose Minkowski function g(S , ·) verifies the
and it is an outer semicontinuous, locally bounded, compact-valued, robust control Minkowski–Lyapunov inequality (2.3) if and only if
convex-valued, positively homogeneous of the first degree set- the finite collection {(Ai , Bi ) ∈ Rn×n × Rn×m : i ∈ I } of matrix pairs
is strictly stabilizable.
valued map.
5. Proper C -polytopic generator sets
4.3. Robust controlled Lyapunov property
Similarly to what is established in Raković (2020a) for the case
As documented in Appendix B.3, any robust control of ordinary control Minkowski–Lyapunov inequality, the poly-
Minkowski–Lyapunov function is a regular robust control Lya- topic setting, characterized by Assumption 2, results in a consid-
punov function. This fact follows from the topological properties erable refinement of the fundamental property of robust control
of Minkowski functions generated by proper C -sets, and it is Minkowski–Lyapunov functions and the topological structure of
formally stated by the following. the related robustly stabilizing set-valued control map.

Proposition 4. Suppose Assumption 1 holds. Suppose also that Assumption 2. (Ai , Bi ) ∈ Rn×n × Rn×m for each i ∈ I where
Minkowski function g(S , ·) verifies the robust control Minkowski– I = {1, 2, . . . , p} is a finite index set. V is a proper C -polytopic
Lyapunov inequality (2.3). For any vector norm |·|N on Rn there exist set in Rn+m represented irreducibly by
strictly positive and finite scalars c1 and c2 such that
V = {(x, u) : ∀j ∈ JV , vx Tj x + vu Tj u ≤ 1}, (5.1)
∀x ∈ R ,n
c1 |x|N ≤ g(S , x) ≤ c2 |x|N , and (4.3a)
where (vxj , vuj ) ∈ R × R for all j ∈ JV , and JV is a finite
n m

∀x ∈ Rn , ∃u ∈ Rm : index set, and S is a proper C -polytopic set in Rn represented


max g(S , Ax + Bu) − g(S , x) ≤ −c1 |x|N . (4.3b) irreducibly by
(A,B)∈M
S = {x : ∀j ∈ JS , sTj x ≤ 1}, (5.2)
4.4. Existence n
where sj ∈ R for all j ∈ JS , and JS is a finite index set.

The map F (·, ·) : Rn × Rm ⇒ Rn specified by (2.2) is a con- ( Appendix B.4) demonstrates that ⋂

the proper C -sets Ri =
tinuous, with respect to the Hausdorff distance, locally bounded, (Ai Bi )T S ∗ ⊕ V ∗ , i ∈ I and R = i∈I Ri are also polytopic
sets, as formally stated by the following.
compact-valued, convex-valued, positively homogeneous of the
first degree set-valued map, in view of which the system x+ ∈
Proposition 5. Suppose
)∗ Assumption 2 holds. For any i ∈ I ,
F (x, u) is strictly stabilizable if and only if there exist a proper (
C -set L in Rn and a scalar λ ∈ [0, 1) such that Ri = (Ai Bi )T S ∗ ⊕ V ∗ is a proper C -polytopic set in Rn+m with,
possibly redundant, representation
∀x ∈ L , ∃u ∈ Rm :
Ri = {(x, u) : ∀(j, k) ∈ JS × JV ,
∀(A, B) ∈ M, Ax + Bu ∈ λL. (4.4)
(sTj Ai + vx Tk )x + (sTj Bi + vu Tk )u ≤ 1}. (5.3)
Since M is the convex hull of the finite collection {(Ai , Bi ) : i ∈ I }
Furthermore, R = i∈I Ri is also a proper C -polytopic set in Rn+m

of matrix pairs, relation (4.4) can be equivalently expressed as
with, possibly redundant, representation
∀x ∈ L , ∃u ∈ Rm :
R = {(x, u) : ∀(i, j, k) ∈ I × JS × JV ,
∀i ∈ I , Ai x + Bi u ∈ λL. (4.5)
(sTj Ai + vx Tk )x + (sTj Bi + vu Tk )u ≤ 1}. (5.4)
Thus, we say that the finite collection {(Ai , Bi ) : i ∈ I } of matrix The irreducible representation of the proper C -polytopic set R
pairs is strictly stabilizable if and only if there exists a proper C - can be obtained by solving at most cardinality(I ) · cardinality(JS ) ·
set L in Rn and a scalar λ ∈ [0, 1) such that the condition (4.5) cardinality(JV ) linear programming problems, and it takes the
holds true. following form
Given any proper C -set V in Rn+m , the existence of a proper
C -set S in Rn whose Minkowski function g(S , ·) is a robust control R = {(x, u) : ∀j ∈ JR , rx Tj x + ru Tj u ≤ 1}. (5.5)
Minkowski–Lyapunov function, implies in view of Proposition 4
The actual inequalities rx Tl x
+ ru Tl u
≤ 1 of the set R are indexed
that there exists a λ ∈ [0, 1) such that L = S and λ verify the
by a finite index set JR , and each (rxl , rul ), l ∈ JR satisfies
condition (4.4) or, equivalently, the condition (4.5). This fact, in
(rxl , rul ) = (ATi sj + vxk , BTi sj + vuk ) for a relevant index triplet
turn, implies that the finite collection {(Ai , Bi ) : i ∈ I } of matrix
(i, j, k) ∈ I × JS × JV determined by solving the mentioned linear
pairs is strictly stabilizable. programming problems.
On the other hand, given any proper C -set V in Rn+m , the strict Minkowski function g(S , ·) of a proper C -polytopic set S , sat-
stabilizability of the finite collection {(Ai , Bi ) : i ∈ I } of matrix isfying Assumption 2, is given, for all x ∈ Rn , by
pairs guarantees the existence of a Minkowski function g(S , ·) of
a proper C -set S in Rn that verifies the robust control Minkowski– g(S , x) = sTj x when x ∈ Sj , (5.6)
Lyapunov inequality (2.3). In particular, one can take S = θ L for where Sj , j ∈ JS are proper polyhedral cones in R n
a suitable θ ∈ (0, ∞) where L, together with some λ ∈ [0, 1),
verifies the condition (4.5) or, equivalently, the condition (4.4). Sj := {x : ∀k ∈ JS \ {j}, (sk − sj )T x ≤ 0}, (5.7)
5
S.V. Raković Automatica 125 (2021) 109437

and where the finite collection {Sj : j ∈ JS } forms a proper implemented by evaluating γ = g(S , x) and by solving convex
polyhedral–conical partition of Rn . Hence, the fundamental prop- minimization
erty of robust control Minkowski–Lyapunov functions can be
minimize uT u
refined to a finite number of set inclusions involving polyhedral
cones as per the following fact, the proof of which is given in with respect to u, α ≥ 0 and β ≥ 0
Appendix B.5. subject to ∀i ∈ I , Ai x + Bi u ∈ α S ,
(x, u) ∈ β V ,
Corollary 2. Invoke Assumption 2. Let R be given by (5.5). Let also,
for all j ∈ JS , Sj be given by (5.7) and α + β ≤ γ. (6.3)

YSj := {(x, u) : ∀k ∈ JR , (rxk − sj ) x + T


ru Tk u ≤ 0}. (5.8) In the polytopic setting, the robustly stabilizing set-valued map
K(S , ·) is explicitly given by (5.10)–(5.11) and its values can be
Minkowski function g(S , ·) verifies the robust control Minkowski– evaluated by solving a set of linear inequalities. In this case, the
Lyapunov inequality (2.3) if and only if minimum Euclidean norm control action can be, for any given
∀j ∈ J S , Sj ⊆ Px YSj . (5.9) x ∈ Rn , computed efficiently by determining any j ∈ JS such
that x ∈ Sj and by solving a strictly convex quadratic programme
In this case, the set-valued map K(S , ·) is polyhedral,
polytopic-valued and Lipschitz continuous with respect to the minimize uT u
Hausdorff distance, as discussed in Appendix B.6. with respect to u
subject to ∀k ∈ J R , ru Tk u ≤ (sj − rxk )T x. (6.4)
Corollary 3. Invoke Assumption 2. Let R be given by (5.5). Let also,
for all j ∈ JS , Sj and YSj be given by (5.7) and (5.8), respectively.
6.2. Consistently improving robust optimal control
The set-valued map K(S , ·) specified by (2.4) is given explicitly, for
all x ∈ Rn , by
The finite horizon N robust optimal control for the system
K(S , x) = K(Sj , x) when x ∈ Sj , (5.10) x+ ∈ F (x, u) with F (·, ·) given by (2.2) and with Minkowski stage
and terminal cost functions, g(V , (·, ·)) and g(S , ·), takes the form
where, for all j ∈ JS and all x ∈ Sj , of an infinite dimensional minimaximization problem specified,
K(Sj , x) := {u : ∀k ∈ JR , ru Tk u ≤ (sj − rxk )T x}. (5.11) for any x ∈ Rn , by
N −1
Furthermore, when Minkowski function g(S , ·) verifies the robust ∑
minimaximize g(V , (xk , uk )) + g(S , xN )
control Minkowski–Lyapunov inequality (2.3), the robustly stabiliz-
ing set-valued control map K(S , ·) is nonempty and polytopic-valued k=0

relative to Rn , its graph is the finite union of polyhedral cones in with respect to {xk }Nk=0 and {uk }Nk=−01
Rn+m subject to ∀k ∈ {0, 1, . . . , N − 1}, xk+1 ∈ F (xk , uk )

{(x, u) ∈ YSj : x ∈ Sj }, (5.12) with x0 = x, (6.5)
j∈JS
where, for typographical convenience, for each relevant k, uk :=
and it is Lipschitz continuous with respect to the Hausdorff distance. uk (xk ) and xk+1 := xk+1 (xk , uk ) are actually control and state func-
tions, and where the formulated minimaximization involves the
6. Selected control applications maximization with respect to sequences of such state functions
{xk }Nk=0 , which is followed by the minimization with respect to
6.1. Robust stabilization sequences of such control functions {uk }Nk=−01 . Within our setting,
the outlined minimaximization is well defined for all N ≥ 1,
For any robust control Minkowski–Lyapunov function g(S , ·),
albeit this exact formulation is computationally very complex.
the related set-valued map K(S , ·) is, by its definition, robustly
When Minkowski terminal cost function g(S , ·) is a robust control
stabilizing in the strong sense, i.e.,
Minkowski–Lyapunov function, the finite horizon N robust opti-
∀x ∈ R n , ∀u ∈ K(S , x), mal control is consistently improving in the sense that, for each
integer N ≥ 0,
max g(S , Ax + Bu) + g(V , (x, u)) ≤ g(S , x). (6.1)
(A,B)∈M
∀x ∈ Rn , VN0+1 (x) ≤ VN0 (x), (6.6)
Any control function κ (S , ·) : Rn → Rm that is a selection of
the robustly stabilizing set-valued control map K(S , ·) is robustly where VN0 (·) denotes the value function of the corresponding
stabilizing. A particularly interesting selection is the minimum minimaximization problem (6.5) with, by convention, V00 (x) :=
Euclidean norm control function given, for all x ∈ Rn , by g(S , x) for all x ∈ Rn .

κ (S , x) := arg min{uT u : u ∈ K(S , x)} (6.2) 6.3. Stabilizing robust model predictive control
u

due to its uniqueness for any x ∈ Rn and its regularity properties. The previously discussed exact finite horizon N robust optimal
Namely, when g(S , ·) is a robust control Minkowski–Lyapunov control also results in stabilizing robust model predictive control.
function, the minimum Euclidean norm control function κ (S , ·) In particular, the corresponding robust model predictive control
specified by (6.2) is a continuous and positively homogeneous law is u00 (x), which is, in general, set-valued map and which is
of the first degree single-valued function if Assumption 1 holds, obtained from the first term of the minimaximizing sequence
which is, in addition, piecewise linear if Assumption 2 holds. of control functions {uk }Nk=−01 with respect to the minimaximiza-
Given a robust control Minkowski–Lyapunov function g(S , ·), tion problem (6.5). The corresponding robust model predictive
the robustly stabilizing control actions, at any given x ∈ Rn , can controlled uncertain dynamics take the form
be computed numerically via convex optimization. The minimum
Euclidean norm control action can be, for any given x ∈ Rn , x+ ∈ {z : z ∈ F (x, u), u ∈ u00 (x)}. (6.7)
6
S.V. Raković Automatica 125 (2021) 109437

When Minkowski terminal cost function g(S , ·) is a robust con- In light of Theorem 6, the verification problem is also equiv-
trol Minkowski–Lyapunov function, the global robust stability of alent to the problem of checking if the set inclusion S ⊆ G −1 (S )
the origin for the robust model predictive controlled uncertain holds true, while, for instance, the set iteration given, for all
dynamics (6.7) is guaranteed. In particular, the related value integers k ≥ 0, by
function VN0 (·) satisfies the usual decrease condition in the strong
Sk+1 := G −1 (Sk ) with S0 := Px V (7.3)
sense, i.e.,
allows for a systematic design of robust control Minkowski–
∀x ∈ Rn , ∀u ∈ u00 (x), ∀x+ ∈ F (x, u), Lyapunov functions. Under Assumption 1 and when the finite
VN0 (x+ ) ≤ VN0 (x) − g(V , (x, u)). (6.8) collection {(Ai , Bi ) : i ∈ I } of the matrix pairs is strictly stabiliz-
able, the set sequence {Sk }k≥0 generated by the set iteration (7.3)
7. Computations, illustration and conclusions is the sequence of monotonically nonincreasing proper C -sets in
Rn that is inner bounded by a proper C -set in Rn . Thus, {Sk }k≥0
is convergent and its limit S is a proper C -set in Rn that satisfies
The computational methods discussed in Raković (2020a, Sec-
S = G −1 (S ). Under Assumption 2 and when the finite collection
tion 7) apply with suitable generalizations to the verification
{(Ai , Bi ) : i ∈ I } of the matrix pairs is strictly stabilizable, the sets
of the robust control Minkowski–Lyapunov inequality (2.3) and Sk are proper C -polytopic sets in Rn , and the limit S is a proper
the design of robust control Minkowski–Lyapunov functions, as C -polytopic set in Rn when it is finitely determined, i.e., when
outlined below. Sk∗ = Sk∗ +1 for a finite integer k∗ .

7.1. Set-valued verification and design 7.2. Linear programming verification and design

Given a finite collection {(Ai , Bi ) : i ∈ I } of matrix pairs In the polytopic setting, Corollary 3 and the results of Walkup
and proper C -sets V and S , the problem of checking if Minkowski and Wets (1969) on lifting projections of convex polyhedral sets
function g(S , ·) verifies the robust control Minkowski–Lyapunov guarantee the existence of a continuous piecewise linear control
inequality (2.3) is known as the verification problem. By The- function κ (S , ·) such that
orem 3, the verification problem is equivalent to the problem ∀x ∈ Rn , κ (S , x) ∈ K(S , x). (7.4)
of checking if the set inclusion S ⊆ Px R holds true, while, by
Corollary 1, the verification problem is equivalent to the problem Such a (Lipschitz) continuous piecewise linear control function
of checking if a given proper C -set S is a robust control invariant κ (S , ·) can be represented, for all x ∈ Rn , by
set for the system x+ ∈ F (x, u) with F (·, ·) given by (2.2)
κ (S , x) = Kj x when x ∈ Kj , (7.5)
and the( state dependent constraints (x, u) ∈ R, where R =
∗ ∗
⋂ T ∗
)
i∈I (A i Bi ) S ⊕ V is a proper C -set. where {Kj : j ∈ JK } is a finite family of proper polyhedral
Given a finite collection {(Ai , Bi ) : i ∈ I } of matrix pairs cones forming a proper polyhedral–conical partition of Rn , and
and a proper C -set V , the problem of constructing a proper C -set {Kj ∈ Rm×n : j ∈ JK } is a finite collection of matrices.
S whose Minkowski function g(S , ·) verifies the robust control Hence, it makes perfect sense to consider a problem of checking
Minkowski–Lyapunov inequality (2.3) is known as the design if the robust control Minkowski–Lyapunov inequality (2.3) holds
problem. In light of Theorem 3, the design problem is equivalent true with u = κ (S , x). This verification problem applies to any
given proper polyhedral–conical partitions {Sj : j ∈ JS } and
to the problem of constructing a proper C -set S for which the
set inclusion S ⊆ Px R holds true, while, in view of Corollary 1,
{Kj : j ∈ JK } of Rn and collections of vectors {sj ∈ Rn : j ∈ JS }
and matrices {Kj ∈ Rm×n : j ∈ JK } such that (5.6), (5.7) and
the design problem is equivalent to the problem of constructing a
(7.5) hold true. In light of Corollaries 2 and 3, the answer to this
proper C -set S that is a robust control invariant set for the system
variant of the verification problem is positive if and only if
x+ ∈ F (x, u) with F (·, ·) given by (2.2)
⋂ and( the state dependent
∗ ∗
constraints (x, u) ∈ R, where R = ∀j ∈ JS , ∀k ∈ J K , ∀l ∈ J R ,
T ∗
)
i∈I (Ai Bi ) S ⊕ V is a
proper C -set.
∀x ∈ Sj ∩ Kk , (rxl − sj + KkT rul )T x ≤ 0. (7.6)
The verification and design problems can be addressed by
utilizing already existing algorithms in set invariance (Blanchini As discussed in Raković (2020a, Section 7.2), the relations (7.6)
& Miani, 2008) or their suitable modifications, the computa- are equivalent to membership relations
tional efficiency of which depends on specific representations of
∀j ∈ JS , ∀k ∈ J K , ∀l ∈ J R ,
the proper C -sets V and S . Inter alia, one such an approach is
based on the following result, the proof of which is provided in (rxl − sj + KkT rul ) ∈ (Sj ∩ Kk )∗ , (7.7)
Appendix B.7. ∗
where the polar cone C of a convex cone C is by definition
C ∗ := {x : ∀y ∈ C , yT x ≤ 0}, and where each polar cone
Theorem 6. Suppose Assumption 1 holds. Minkowski function
(Sj ∩ Kk )∗ is polyhedral. Hence, the above verification problem
g(S , ·) verifies the robust control Minkowski–Lyapunov inequal- can be efficiently addressed by checking feasibility of a system of
ity (2.3) if and only if affine inequalities and equalities representing the direct algebraic
S ⊆ G −1 (S ), (7.1) reformulation of (7.7).
The preceding verification problem can be made more flexible
where G −1
(·) is given, for any nonempty set X in R , by
n
by coupling it with a simultaneous design of a collection of
matrices {Kj ∈ Rm×n : j ∈ JK }. In this case, the continuity
G −1 (X ) := (I O){(x, u, γ ) : ∀i ∈ I , Ai x + Bi u ∈ γ X of the control function κ (S , ·) has to be ensured, which, in light
(x, u) ∈ (1 − γ )V , of (7.5), requires that, for all ∀(j, k) ∈ P (JK ), and all x ∈ Kj ∩
γ ∈ [0, 1]}, (7.2) Kk , it holds that Kj x = Kk x. P (JK ) is the set of pairs (j, k) ∈
JK × JK such that j ̸ = k and in which pairs (j, k) and (k, j) are,
with I ∈ R n×n
and O ∈ R n×(m+1)
. due to the symmetry, considered only once as the pair (j, k). As
7
S.V. Raković Automatica 125 (2021) 109437

also discussed in Raković (2020a, Section 7.2), such a continuity


condition can be equivalently expressed as
∀(j, k) ∈ P (JK ), ∀l ∈ {1, . . . , m},
(Kj l − Kkl ) ∈ (Kj ∩ Kk )∗ ∩ (−(Kj ∩ Kk )∗ ),
T
(7.8)
where Kj l denotes the lth row of the matrix Kj and where each
(Kj ∩ Kk )∗ (−(Ki ∩ Kj )∗ ) is a polyhedral cone. Hence, the ver-

ification with simultaneous design of a collection of matrices
{Kj ∈ Rm×n : j ∈ JK } problem can be efficiently addressed
by means of feasibility of a system of affine inequalities and
equalities representing the direct algebraic reformulations of (7.7)
and (7.8).
As also discussed in Raković (2020a, Section 7.2), the verifi- Fig. 1. Sets S , S 0 , S # and X .
cation problem can be made more flexible by coupling it with
simultaneous design of a collection of vectors {sj ∈ Rn : j ∈
JS }. In this case, it has to be ensured that S = {x : ∀j ∈
JS , sTj x ≤ 1} is a proper C -polytopic set whose Minkowski
function g(S , ·) satisfies (5.6) and (5.7). Such conditions can be
fulfilled by considering any given proper C -polytopic set T :=
{x : ∀j ∈ JT , θjT x ≤ 1} whose Minkowski function g(T , ·)
is given, for all x ∈ Rn , by g(T , x) = θjT x when x ∈ Tj where
Tj := {x : ∀k ∈ JT \ {j}, (θk − θj )T x ≤ 0}, and by setting
JS = JT and, for each j ∈ JS , Sj = Tj . In which case, it suffices
to ensure that α −1 T ⊆ S ⊆ β −1 T with (α, β ) ∈ (0, ∞) × (0, ∞)
and that g(S , x) = sTj x when x ∈ Sj . The latter conditions
can be equivalently rewritten as polyhedral–conical membership
conditions
Fig. 2. Restrictions of function g(S , ·) and Map K(S , ·) to S .
∀j ∈ JS , (sj − αθj ) ∈ S∗j with α ∈ (0, ∞) and
(βθj − sj ) ∈ S∗j with β ∈ (0, ∞), and
under hard state and control constraints
∀j ∈ JS , ∀k ∈ JS \ {j}, (sk − sj ) ∈ Sj .

(7.9)
x ∈ X = [−10, 10] × [−10, 10] and u ∈ U = [−4, 4]
Thus, the verification with simultaneous design of a collection of
vectors {sj ∈ Rn : j ∈ JS } problem can be also efficiently was considered for an assigned set of initial conditions X0 :=
addressed via feasibility of a system of affine inequalities and 0.5X = [−5, 5] × [−5, 5]. In Blanchini and Miani (1999), the
equalities representing the direct algebraic reformulations of (7.7) solution to this robust stabilization under constraints problem
and (7.9). (In this case, sj , j ∈ JS are decision variables, and it is was obtained by constructing a robust λ-contractive set S # , which
necessary to utilize representation (5.4) of the set R in (7.7).) is a proper C -polytope satisfying X0 ⊆ S # ⊆ X and, with λ = 0.5,
The most flexible verification is the one coupled with simul-
taneous design of collections of vectors {sj ∈ Rn : j ∈ JS } and ∀x ∈ S # , ∃u ∈ U : ∀i ∈ {1, 2}, Ai x + Bi u ∈ λS # .
matrices {Kj ∈ Rm×n : j ∈ JK }; It can be performed via an
As observed in Blanchini and Miani (1999), the set X is not a ro-
iteration alternating between the verification with simultaneous
bust λ-contractive set for any λ ∈ [0, 1), and, hence, the method
design of a collection of vectors {sj ∈ Rn : j ∈ JS } and the
for robust stabilization under constraints considered in Blanchini
verification with simultaneous design of a collection of matrices
and Miani (1999) is not applicable for the assigned set of initial
{Kj ∈ Rm×n : j ∈ JK }. Furthermore, the proper polyhedral–
conditions X0 = X . The sets S # and X are plotted in Fig. 1 together
conical partitions {Sj : j ∈ JS } and {Kj : j ∈ JK } of
Rn can be refined iteratively so as to make the related verifica- with the sets 10S and 10S 0 . The sets S and S 0 are the generator
tion and design more flexible. Naturally, a more detailed treatise sets of robust control Minkowski–Lyapunov functions g(S , ·) and
of all intricate computational finesses is a research topic in its g(S 0 , ·) for the case when
own right, and, thus, it is the main theme of the companion V = [−1, 1] × [−1, 1] × [−1, 1].
paper, which unifies computational aspects for the verification
and design of control Minkowski–Lyapunov functions (Raković, The set S is constructed by a simplified version of the set it-
2020a) and robust control Minkowski–Lyapunov functions. (Note eration (7.3), in which the fixed value of γ = 0.6 is utilized
that even simpler computations of polyhedral Lyapunov functions and which converges in k∗ = 2 steps resulting in a proper C -
form a research theme in its own right (Giesl & Hafstein, 2015; polytopic set S with an irreducible representation (5.2) in which
Lazar, 2010).) cardinality(IS ) = 6. The set S 0 is constructed by the actual set
iteration (7.3), which converges in k∗ = 11 steps resulting in a
7.3. Academic illustration proper C -polytopic set S 0 with an irreducible representation (5.2)
in which cardinality(IS ) = 18. In light of Theorem 6, the sets S
We close this article by illustrating an application of our and S 0 satisfy S ⊆ G −1 (S ), S 0 = G −1 (S 0 ) and S ⊆ S 0 . Fig. 2
results to the tutorial example of Blanchini and Miani (1999), illustrates the polyhedral structure and Lipschitz continuity of
wherein the robust stabilization of the origin for a parametrically the robust control Minkowski–Lyapunov function g(S , ·) and the
uncertain linear system given via (2.1)–(2.2) with two pairs of associated robustly stabilizing set-valued control map K(S , ·).
system matrices (A1 , B1 ) and (A2 , B2 ) and with B1 = B2 = B and In this example, the set X is, however, the maximal robust
( ) ( ) ( )
1 1 0 1 0 control invariant set for the considered parametrically uncertain
A1 = 0.5 , A2 = 0.5 and B = ,
−1 1 −1 0 1 linear system and hard state and control constraints. In other
8
S.V. Raković Automatica 125 (2021) 109437

Fig. 3. Function g(S , X , ·) and Map K(S , X , ·).


Fig. 4. Extreme state sequences: Space plots.

words, the set X is the maximal set Z , w.r.t. set inclusion, such
that Z ⊆ X and
∀x ∈ Z , ∃u ∈ U : ∀i ∈ {1, 2}, Ai x + Bi u ∈ Z ,
and the related robust control invariance inducing set-valued
map I(X , ·) : X ⇒ U given, for all x ∈ X , by
I(X , x) := {u ∈ U : ∀i ∈ {1, 2}, Ai x + Bi u ∈ X },
is nonempty for all x ∈ X . Consequently, the set X is guaranteed
to be the maximal robust controlled domain of attraction of the
origin guaranteed by the robust control Minkowski–Lyapunov
function g(S , ·) if the set-valued map K(S , X , ·) given, for all x ∈ Fig. 5. Extreme state and control sequences: Time plots.
X , by

K(S , X , x) = K(S , x) I(X , x)
Appendix A. Definitions and terminology
is nonempty for all x ∈ X . Fig. 3 depicts the restriction g(S , X , ·)
of the function g(S , ·) to the set X and the set-valued map A function f (·) : Rn → Rm is:
K(S , X , ·), and it illustrates that (a) positively homogeneous of the first degree if, for all x ∈ Rn
∀x ∈ X , K(S , X , x) ̸= ∅, and all η ∈ R≥0 , f (ηx) = ηf (x),
(b) subadditive if, for all x ∈ Rn and all y ∈ Rn , f (x + y) ≤
which means that, for all x ∈ X , there exists a control u ∈ U that f (x) + f (y), and
ensures simultaneously the robust Lyapunov decrease of g(S , ·) (c) sublinear if it is positively homogeneous of the first degree
and robust control invariance of X . and subadditive.
Thus, the set X is, indeed, the maximal robust controlled do- A set-valued map F (·) associates subsets F (x) of Rm to points
main of attraction of the origin guaranteed by the robust control x in Rn . The graph of a set valued map F (·) : Rn ⇒ Rm is the set
Minkowski–Lyapunov function g(S , ·). Figs. 4 and 5 illustrate the {(x, y) : y ∈ F (x), x ∈ Rn }. A set-valued map F (·) : Rn ⇒ Rm
robust controlled stability of the origin with the maximal robust is:
controlled domain of attraction under constraints by depicting a (i) outer semicontinuous over Rn , if, for all x ∈ Rn , every
sample of controlled extreme state and control actions sequences. convergent sequence xk → x, and every convergent sequence
The plotted state sequences start from vertices of the set X yk → y with yk ∈ F (xk ) it holds that y ∈ F (x),
and are controlled by the constrained version of the minimum (ii) locally bounded over Rn , if, for all x ∈ Rn , there is a
Euclidean norm control function (6.4) in which K(S , X , x) is used neighborhood X of x such that the set F (X ) := ∪x∈X F (x) is
instead of K(S , x), while the realization of matrix pairs (A, B) takes bounded,
extreme values (i.e. either (A1 , B1 ) or (A2 , B2 )) at each time step. (iii) compact-valued over Rn , if, for all x ∈ Rn , F (x) is a compact
The simulation time is N = 9, since the possible states and control subset of Rm ,
actions at time 9 are almost equal to 0. (iv ) convex-valued over Rn , if, for all x ∈ Rn , F (x) is a convex
This academic example illustrates that our results are advanta- subset of Rm , and
geously usable for the robust stabilization under constraints. Such (v ) positively homogeneous of the first degree, if, for all x ∈ Rn
a beneficial utility is a consequence of the inherent topological and all η ∈ R≥0 , F (ηx) = ηF (x).
compatibility with the theory of robust control Lyapunov func- Given any two nonempty compact subsets X and Y of Rn , their
tions, which can be obtained by utilizing different sets, S and D, Hausdorff distance is
as the generator set S of a robust control Minkowski–Lyapunov
function and a related robust controlled domain of attraction of HBn (X , Y ) := min{η : X ⊆ Y ⊕ ηBn and Y ⊆ X ⊕ ηBn },
η≥0
the origin D. (The use of the robust λ-contractive sets (Blanchini
& Miani, 1999) enforces S = D, which is bound to be conservative where B is the closed unit Euclidean norm ball in Rn .
n

in the constrained setting.) A set-valued map F (·) : Rn ⇒ Rm is continuous, w.r.t. the


Hausdorff distance, over Rn if, for all x ∈ Rn , for all ε > 0 there
Acknowledgments exists a δ > 0 such that for all y ∈ Rn such that ∥y − x∥ ≤ δ it
holds that HBn (F (y), F (x)) ≤ ε .
The author is grateful to the Editor, Associate Editor and Ref- A set-valued map F (·) : Rn ⇒ Rm is Lipschitz continuous,
erees for timely, professional and constructive review. w.r.t. the Hausdorff distance, if there exists a constant ~ ∈ [0, ∞)
9
S.V. Raković Automatica 125 (2021) 109437

such that |·|N = g(N , ·) for a symmetric, with respect to the origin, proper
C -set N in Rn . Since S is a proper C -set in Rn , it follows that
∀x ∈ R ,n
∀y ∈ R ,
n
HBn (F (x), F (y)) ≤ ~∥x − y∥.
d1 := min{d : S ⊆ dN , d ≥ 0} and
d
The intersection of finitely many closed half-spaces is a poly-
hedral set. A polytopic set is a bounded polyhedral set. A d2 := max{d : dN ⊆ S , d ≥ 0}
d
nonempty convex subset X of Rn such that x ∈ X implies that
are finite and strictly positive scalars such that
λx ∈ X for all λ ≥ 0 is a convex cone in Rn . A convex cone X
in Rn is a polyhedral cone if X is, in addition, a polyhedral set. ∀x ∈ R n , 1 g(N , x) ≤ g(S , x) ≤ d2 g(N , x).
d− 1 −1

A convex cone X is a proper cone if and only if X is closed, full


Since V is a proper C -set in Rn+m , Px V is a proper C -set in Rn , and
dimensional, and pointed (i.e., x ∈ X and −x ∈ X holds only for
x = 0). d3 := min{d : Px V ⊆ dN , d ≥ 0}
d
A finite collection of sets {Xi ⊆ Rn : i ∈ IX } forms a proper
polyhedral–conical partition of Rn if and only if each Xi is a proper is a finite and strictly positive scalar such that, for all x ∈
3 g(N , x) ≤ g(Px V , x). By the virtue of Raković (2019,
1
polyhedral cone in Rn , interior(Xi ) ∩ interior(Xj ) = ∅ for all i ∈ IX Rn , d−
and all j ∈ IX \ {i}, and ∪i∈IX Xi = Rn . Theorem 4), for all x ∈ Rn ,
g(Px V , x) = min g(V , (x, u)).
u∈Rm
Appendix B. Proofs
Hence,
∀x ∈ R n , ∀u ∈ Rm , 3 g(N , x) ≤ g(V , (x, u)).
d−1

B.1. Proof of Proposition 2.


By (2.3), for all x ∈ Rn there exists a u ∈ Rm such that

The proof is provided in a compactified form since it is an max g(S , Ax + Bu) + g(V , (x, u)) ≤ g(S , x),
(A,B)∈M
adaptation of Raković (2020a, Proof of Proposition 3) to our
setting. Take any i ∈ I . Sets S ∗ and V ∗ are proper C -sets in Rn and and, consequently, for all (x, u) ∈ Rn × Rm such that
Rn+m , respectively, being polar sets of proper C -sets S and V . Set max g(S , Ax + Bu) + g(V , (x, u)) ≤ g(S , x),
(Ai Bi )T S ∗ is, thus, a C -set in Rn+m being a linear transformation (A,B)∈M

of a proper C -set S ∗ . Hence, R∗i = (Ai Bi )T S ∗ ⊕ V ∗ is a proper in fact, it holds that


C -set in Rn+m being the Minkowski sum of a C -set (Ai Bi )T S ∗ and
a proper C -set V ∗ . Finally, Ri is a proper C -set in Rn+m being polar max g(S , Ax + Bu) − g(S , x) ≤ −d−
3 g(N , x).
1
(A,B)∈M
set of a proper C -set R∗i . The proof is completed by noting that,
Summarizing, for any vector norm |·|N on Rn and any robust
for all (x, u) ∈ Rn × Rm ,
control Minkowski–Lyapunov function g(S , ·), with c1 := min
g(Ri , (x, u)) = h(R∗i , (x, u)) 1 , d3 } and c2 := d2 ,
1 −1 −1
{d−
= h((Ai Bi )T S ∗ ⊕ V ∗ , (x, u)) ∀x ∈ R n , c1 |x|N ≤ g(S , x) ≤ c2 |x|N , and
= h((Ai Bi ) S , (x, u)) + h(V , (x, u))
T ∗ ∗
∀x ∈ R n ,∃u ∈ Rm :
= h(S ∗ , Ai x + Bi u) + h(V ∗ , (x, u)) max g(S , Ax + Bu) − g(S , x) ≤ −c1 |x|N .
(A,B)∈M
= g(S , Ai x + Bi u) + g(V , (x, u)).
B.4. Proof of Proposition 5.
B.2. Proof of Proposition 3.
In view of representations (5.1) and (5.2), applying the defini-
tion of the polar set to V and S yields,
First, since {Zi : i ∈ I } is a finite collection of proper
V ∗ = convh({(vxj , vuj ) : j ∈ JV }), and
C -sets Zi in Rn+m , their intersection Z :=

i∈I Z i is a proper
C -set in Rn+m . Next, take any arbitrary (x, u) ∈ Rn × Rm . Let S ∗ = convh({sj : j ∈ JS }).
α = maxi∈I g(Zi , (x, u)) and β = g(Z , (x, u)). By construction, Obviously, for any i ∈ I ,
α ∈ [0, ∞) and β ∈ [0, ∞), and we show that α = β .
For each i ∈ I , g(Zi , (x, u))⋂≤ α so that,⋂for each i ∈ I , (Ai Bi )T S ∗ = convh({(ATi sj , BTi sj ) : j ∈ JS }).
(x, u) ∈ α Zi . In turn, (x, u) ∈ i∈I α Zi = α i∈I Zi = α Z so Since convh(X ) ⊕ convh(Y ) = convh(X ⊕ Y ) by the virtue of
that (x, u) ∈ α Z . Hence, β ≤ α . Schneider (1993, Theorem 1.1.2), it follows that, for any i ∈ I ,
Likewise, (x, u) ∈ β Z and, since β Z = β i∈I Zi = i∈I β Zi ,
⋂ ⋂
R∗i = (Ai Bi )T S ∗ ⊕ V ∗ is given by
it follows that for each i ∈ I , (x, u) ∈ β Zi . In turn, for each i ∈ I ,
g(Zi , (x, u)) ≤ β , and maxi∈I g(Zi , (x, u)) ≤ β . Hence, α ≤ β . R∗i = convh{(ATi sj + vxk , BTi sj + vuk ) : (j, k) ∈ JS × JV }.
Consequently, since α ≤ β and β ≤ α , α = β . Hence, since In light of the definition of the polar set, as claimed, for any i ∈ I ,
(x, u) ∈ Rn × Rm was arbitrary, Ri = ((Ai Bi )T S ∗ ⊕ V ∗ )∗ is given by
∀(x, u) ∈ Rn × Rm , max g(Zi , (x, u)) = g(Z , (x, u)). Ri = {(x, u) : ∀(j, k) ∈ JS × JV ,
i∈I
(sTj Ai + vx Tk )x + (sTj Bi + vu Tk )u ≤ 1}.

B.3. Proof of Proposition 4. Since R = i∈I Ri , it follows that, as also claimed,

R = {(x, u) : ∀(i, j, k) ∈ I × JS × JV ,
The proof is an extension of Raković (2020a, Proof of Propo-
sition 2) to our setting. Take any vector norm |·|N on Rn . Then (sTj Ai + vx Tk )x + (sTj Bi + vu Tk )u ≤ 1}.

10
S.V. Raković Automatica 125 (2021) 109437

B.5. Proof of Corollary 2. Giesl, P., & Hafstein, S. (2015). Review on computational methods for Lyapunov
functions. Discrete and Continuous Dynamical Systems, Series B, 20.
Goebel, R., Hu, T., & Teel, A. (2006). In L. Menini, L. Zaccarian, & C. T. Abdallah
In view of the relations (5.6) and (5.7), the necessary and
(Eds.), Current trends in nonlinear systems and control, Dual matrix inequalities
sufficient condition (3.6) equivalently reads as, in stability and performance analysis of linear differential/difference inclusions
(pp. 103–122). Birkhauser.
∀j ∈ J S , ∀x ∈ Sj , ∃u ∈ Rm : (x, u) ∈ (sTj x)R,
Krstić, M., & Deng, H. (1998). Stabilization of nonlinear uncertain systems. Springer
which, in light of the relation (5.8), equivalently reads as London.
Lazar, M. (2010). On infinity norms as Lyapunov functions: Alternative necessary
∀j ∈ J S , ∀x ∈ Sj , ∃u ∈ Rm : (x, u) ∈ YSj , and sufficient conditions. In 49th IEEE conference on decision and control (CDC)
(pp. 5936–5942).
which yields the claim. Lyapunov, A. M. (1992). The general problem of the stability of motion. Taylor &
Francis.
Molchanov, A. P., & Pyatnitskii, E. S. (1986). Lyapunov functions that specify
B.6. Proof of Corollary 3.
necessary and sufficient conditions of absolute stability of nonlinear nonsta-
tionary control systems I, II, and III. Automation and Remote Control, 3,4,4,
The proof of this claim follows along the lines of Raković 63–73, 5–15, 38–49.
(2020a, Proof of Theorem 6), and it is, hence, omitted. Nguyen, H. N., Olaru, S., Gutman, P. O., & Hovd, M. (2011). Constrained
interpolation–based control for polytopic uncertain systems. In 50th IEEE
conference on decision and control and European control conference (pp.
B.7. Proof of Theorem 6.
4961–4966).
Raković, S. V. (2017). The Minkowski–Lyapunov equation. Automatica, 75, 32–36.
The proof is an extension of Raković (2020a, Proof of Theo- Raković, S. V. (2019). The Minkowski–Bellman equation: Technical report, Beijing,
rem 12) to our setting. Due to the positive homogeneity of the China: Beijing Institute of Technology, Publicly available at ResearchGate:
set-valued map F (·, ·) given by (2.2) as well as the involved https://www.researchgate.net/publication/332446880_The_Minkowski-
Bellman_Equation.
Minkowski functions and proper C - structure of the associated
Raković, S. V. (2020a). Control Minkowski–Lyapunov functions. Automatica,
generator sets, it is without loss of generality to consider arbitrary (submitted for publication).
x such that g(S , x) = 1. Raković, S. V. (2020b). Polarity of stability and robust positive invariance.
If g(S , ·) is a robust control Minkowski–Lyapunov function, Automatica, 118, Article 109010.
there exists a u such that (2.3) holds true so that g(V , (x, u)) ≤ Raković, S. V. (2020c). Robust Minkowski–Lyapunov functions. Automatica, 120,
1 − maxi∈I g(S , Ai x + Bi u). Let γ := maxi∈I g(S , Ai x + Bi u) so Article 109168.
Raković, S. V., & Lazar, M. (2014). The Minkowski–Lyapunov equation for linear
that γ ∈ [0, 1] and, for all i ∈ I , g(S , Ai x + Bi u) ≤ γ and dynamics: Theoretical foundations. Automatica, 50(8), 2015–2024.
g(V , (x, u)) ≤ 1 − γ . Thus, for all i ∈ I , Ai x + Bi u ∈ γ S and Rawlings, J. B., & Mayne, D. Q. (2009). Model predictive control: Theory and design.
(x, u) ∈ (1 − γ )V with γ ∈ [0, 1]. Hence, x ∈ G −1 (S ) and Madison: Nob Hill Publishing.
S ⊆ G −1 (S ). Rockafellar, R. T. (1970). Convex analysis. Princeton University Press.
If S ⊆ G −1 (S ) then there exists (u, γ ) such that γ ∈ [0, 1], Rockafellar, R. T., & Wets, R. J.-B. (2009). A series of comprehensive studies in
mathematics: vol. 317, Variational analysis. Springer.
for all i ∈ I , Ai x + Bi u ∈ γ S and (x, u) ∈ (1 − γ )V . Hence
Schneider, R. (1993). Encyclopedia of mathematics and its applications: vol. 44,
maxi∈I g(S , Ai x + Bi u) ≤ γ and g(V , (x, u)) ≤ 1 − γ . Clearly, Convex bodies: The Brunn–Minkowski theory. Cambridge University Press.
max(A,B)∈M g(S , Ax + Bu) = maxi∈I g(S , Ai x + Bi u) and, conse- Sontag, E. D. (1989). A universal construction of Artstein’s theorem on nonlinear
quently, max(A,B)∈M g(S , Ax + Bu) ≤ γ and g(V , (x, u)) ≤ 1 − γ . stabilization. System and Control Letters, 13(2), 117–123.
Hence, max(A,B)∈M g(S , Ax + Bu) + g(V , (x, u)) ≤ 1 = g(S , x), and Sontag, E. D. (1998). Textbooks in applied mathematics: number 6, Mathematical
control theory: Deterministic finite dimensional systems (2nd ed.). New York:
g(S , ·) is a robust control Minkowski–Lyapunov function.
Springer.
Sontag, E. D., & Sussmann, H. J. (1996). General classes of control–Lyapunov
References functions. In International series of numerical mathematics: vol. 121, Stability
theory. Ascona: Birkhaüser, Basel.
Artstein, Z. (1983). Stabilization with relaxed controls. Nonlinear Analysis, Theory, Walkup, D. W., & Wets, R. J.-B. (1969). Lifting projections of convex polyhedra.
Methods & Applications, 7(11), 1163–1173. Pacific Journal of Mathematics, 28(2), 465–475.
Baier, R., Braun, P., Grüne, L., & Kellett, C. M. (2018). Numerical construction
of nonsmooth control Lyapunov functions. In P. Giselsson, & A. Rantzer
(Eds.), Large-scale and distributed optimization (pp. 343–373). Cham: Springer
International Publishing. Saša V. Raković received the Ph.D. degree in Control
Barabanov, N. (1988). Lyapunov indicators of discrete inclusions I, II, and III. Theory from Imperial College London. His Ph.D. the-
Automation and Remote Control, 49, 152–157, 283–287, 558–565. sis, entitled ‘‘Robust Control of Constrained Discrete
Blanchini, F., & Miani, S. (1999). Numerical computation of polyhedral Lyapunov Time Systems: Characterization and Implementation’’,
functions for robust synthesis. In 14th IFAC world congress, IFAC 1999 (pp. was awarded the Eryl Cadwaladr Davies Prize as the
2065–2070). Beijing, China. best Ph.D. thesis in the Department of Electrical and
Blanchini, F., & Miani, S. (2008). Set–theoretic methods in control. Birkhauser. Electronic Engineering at Imperial College London in
Boyd, S., Ghaoui, L. E., Feron, E., & Balakrishnan, V. (1994). Linear matrix in- 2005.
equalities in system and control theory. SIAM studies in applied mathematics: Saša V. Raković was affiliated with a number of
15. the worldwide well-known universities, including, in-
Clarke, F. H., Ledyaev, Y. S., Sontag, E. D., & Subbotin, A. I. (1997). Asymptotic ter alia, Imperial College London, ETH Zürich, Oxford
controllability implies feedback stabilization. IEEE Transactions on Automatic University, the University of Maryland at College Park and the University of
Control, 42(10), 1394–1407. Texas at Austin. He is currently a Full Professor with the School of Automation
Clarke, F. H., Ledyaev, Y. S., Stern, R. J., & Wolenski, P. R. (1998). Nonsmooth of Beijing Institute of Technology, Beijing, China.
analysis and control theory. New York: Springer Verlag. Saša V. Raković’s main research interests and contributions lie within the
Freeman, R. A., & Kokotović, P. V. (2008). Robust nonlinear control design: areas of synthesis of control systems, analysis of dynamical systems and decision
State–space and Lyapunov techniques. Springer Science & Business Media. making under constraints and uncertainty.

11

You might also like