This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TAC.2015.2422451, IEEE Transactions on Automatic Control
Limited circulation. For review only
1
IEEE-TAC Submission no.: 14-0202.4

Implicit Lyapunov-Krasovski Functionals For

Stability Analysis and Control Design of
Time-Delay Systems
A. Polyakov, D. Efimov, W. Perruquetti, J.-P. Richard

Abstract—The method of Implicit Lyapunov-Krasovski Func- Recently [23], the Implicit Lyapunov Function (ILF) method
tional (ILKF) for stability analysis of time-delay systems is has been developed for non-asymptotic (finite-time and fixed-
introduced. Theorems on Lyapunov, asymptotic, (hyper) expo- time) analysis. This method allows the control to be designed
nential, finite-time and fixed-time stability analysis using ILKF
are presented. The hyper exponential stabilization algorithm for a together with a Lyapunov function of the closed-loop system.
time-delay system is presented. The theoretical result is supported A simple procedure for parameter tuning based on Linear
by numerical simulation. Matrix Inequalities (LMIs) was developed for linear plants.
The ideas, which underline the ILF method, were ini-
tially introduced for control synthesis problems in [16]. The
I. I NTRODUCTION corresponding design methodology was called controllability
function method. In order to be more precise we follow more
The celebrated Second Lyapunov Method was originally
recent terminology of [1]. The ILF method uses Lyapunov
founded on the so-called energetic approach to stability anal-
function defined in the implicit form by some algebraic equa-
ysis. It considers any positive definite function as possible
tions. Stability analysis in this case does not require solution
energetic characteristic (”energy”) of a dynamic system and
of this equation, since the Implicit function theorem (see,
the evolution of this ”energy” in time is studied. If a dynamic
for example, [6]) helps to check all stability conditions by
system has an energetic function, which is decreasing (or
analyzing the algebraic equation directly.
strictly decreasing) along any trajectory of the system, then
This technical note extends the concept of ILF to the case
this system has some stability property and the corresponding
of time-delay systems. The stability theorems presented in the
energetic function is called a Lyapunov function.
paper introduce the ILKF method for Lyapunov, asymptotic,
The Lyapunov function method is one of the main the- (hyper) exponential and non-asymptotic (finite-time and fixed-
oretical tools for stability analysis and control synthesis of time) stability analysis of time-delay systems. They allow us
nonlinear systems ([2], [21]). This method allows the stability to define the Lyapunov-Krasovski functionals in the implicit
properties of the system to be studied without finding its form. The developed method is demonstrated on the control
solutions. The extension of the Lyapunov’s ideas to the case design problem for hyper exponential stabilization of a special
of time-delay systems is given by Lyapunov-Razumikhin and time-delay linear system. The procedure to design of control
Lyapunov-Krasovski theorems, see, for example, [15], [10], parameters is presented in the LMI form.
[25], [9], [17]. The corresponding energetic function becomes
a functional in the last case. Finding the appropriate Lyapunov
II. N OTATION
(-Krasovski) functions is a difficult problem in general case.
In addition to qualitative analysis of stability, control theory Through the paper the following notation will be used:
is also interested in quantitative analysis of the convergence • R is the field of real numbers, R+ = {x ∈ R : x > 0};
n
rate. In linear case the concept of exponential stability can • k · k is the Euclidian norm in R ;
n
be utilized for this purpose. Nonlinear control systems may • diag{λi }i=1 is the diagonal matrix with the elements λi
demonstrate faster (e.g. hyper exponential) convergence. In on the main diagonal;
particular, nonlinear systems can be finite-time [26] and fixed- • a positive definite continuous function σ : R → R
time stable [22]. Finite-time control problem is a subject of belongs to the class K if it is strictly increasing on R+
intensive research in the last years; e.g., see [11], [3], [13], and σ(0) = 0; if additionally it is radially unbounded
[19], [21]. The method of Lyapunov(-Krasovski) functions then σ belongs to K∞ ;
helps to estimate the rate of convergence and to adjust the • Ch is the space of continuous functions [−h, 0] →
control parameters in order to provide the desired rate. Rn with the norm k · kh defined as follows kϕkh =
max kϕ(t)k for ϕ ∈ Ch ;
t∈[−h,0]
Andrey Polyakov, Denis Efimov, Wilfrid Perruquetti and J.-P. Richard
are with Non-A INRIA - LNE, Parc Scientifique de la Haute Borne 40, • C0h := {ϕ ∈ Ch : ϕ(0) = 0} is a subspace of Ch .
1/2
avenue Halley Bat.A, Park Plaza 59650 Villeneuve d’Ascq, France (e-mail: • for P > 0 the matrix
 P := B is such that B 2 = P
[email protected], [email protected], [email protected], −1/2 1/2 −1
and P = P ;
[email protected]). They are also with LAGIS UMR CNRS 8219,
Ecole Centrale de Lille, BP 48, Cité Scientifique, 59651 Villeneuve-d’Ascq, • λmin (P ) and λmax (P ) are minimal and maximal eigen-
France. values of symmetric matrix P ∈ Rn×n , respectively.
Preprint submitted to IEEE Transactions on Automatic Control. Received: April 3, 2015 13:47:47 PST
0018-9286 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TAC.2015.2422451, IEEE Transactions on Automatic Control
Limited circulation. For review only
2
IEEE-TAC Submission no.: 14-0202.4

• if P ∈ Rn×n is symmetric then the inequalities P > 0 Note that the Lyapunov-Razumikhin approach to finite-time
(P < 0) and P ≥ 0 (P ≤ 0) means that P is positive stability analysis is inconsistent in the general case [8].
(negative) definite and semidefinite, respectively; The next definition extends the concept of fixed-time sta-
bility [22] to time-delay systems.
III. S TABILITY D EFINITIONS Definition 3: The origin of the system (1) is said to be
fixed-time stable, if it is finite-time stable and the settling-
Consider the system of the form
time functional T : Ch → R+ is bounded, i.e. there exists
ẋ(t) = f (t, xh (t)), t ∈ R+ Tmax ∈ R+ such that sup T (ϕ) = Tmax for some
(1) ϕ∈Ch :kϕkh <δ
x(t) = ϕ(t) ∈ Ch , t ∈ [−h, 0]
δ ∈ R+ . If δ = +∞ then the origin of the system (1) is said
where x ∈ Rn , xh (t) ∈ Ch is the state function defined by to be globally fixed-time stable.
means of the formula xh (t) := x(t + τ ) with τ ∈ [−h, 0] with The right-hand sides of finite-time and fixed-time stable
h ∈ R+ (time delay) and f : R+ × Ch → Rn is a continuous systems are non-Lipschitz continuous [20] or even discontin-
operator. Note that the continuity of f implies, at least, local uous [22]. Such models appear for mechanical systems with
existence of solutions of the system (1) due to Theorem 2.1, dry friction (see, for example, [4]). The phyper exponentially
page 41 from [12]. The advanced detailed analysis of the stable system presented above is also non-Lipschitz at the
existence or uniqueness properties of the solutions goes out origin.
of the scope of this paper. Assume that the origin is an
equilibrium point of the system (1), i.e. f (t, 0) = 0 for all IV. S TABILITY A NALYSIS OF T IME -D ELAY S YSTEMS
t ∈ R+ . A solution of the system (4) with the initial function U SING I MPLICIT LYAPUNOV-K RASOVSKI F UNCTIONAL
ϕ ∈ Ch is denoted by x(t, ϕ).
A. Implicit functions and operators
The definitions of Lyapunov, asymptotic and exponential
stability are well-known, we refer the reader, for example, to Below we utilize a special class of functions introduced by
[12]. Interested in quantitative analysis of stability character- the following definition.
ized by rate of convergence, we briefly discuss some related Definition 4: The function q : R2+ → R, (σ, s) → q(σ, s)
notions. Note that this paper deals only with strong uniform is said to be of the class IK∞ iff 1) q is continuous on R2+ ; 2)
stability properties of the system (1). The words ”strong for any s ∈ R+ there exists σ ∈ R+ such that q(σ, s) = 0; 3)
uniform” are omitted below for simplicity of presentation. for any fixed s ∈ R+ the function q(·, s) is strictly decreasing
Let us introduce the functions ρr : R → R by the following on R+ ; 4) for any fixed σ ∈ R+ the function q(σ, ·) is strictly
recursive formula increasing on R+ ; 5) lim σ = 0, lim s = 0, lim σ = +∞,
s→0+ σ→0+ s→+∞
(σ,s)∈Γ (σ,s)∈Γ (σ,s)∈Γ
ρr−1 (z)
ρ0 (z) = z, ρr (z) = e , 
where Γ = (σ, s) ∈ R2+ : q(σ, s) = 0 .

where r ≥ 1 is integer number. The next definition presents If q ∈ IK∞ then there exists a unique function σ ∈ K∞ ,
hyper exponential asymptotic stability [24]. It follows the ideas such that q(σ(s), s) = 0 for all s ∈ R+ . Indeed, the condition
of hyperpower (see, [18]). 2) claims that there exists a function σ : R+ → R+ such
Definition 1: The origin of the system (1) is said to be that q(σ(s), s) = 0 for any s ∈ R+ . The condition 3) implies
hyper exponentially stable of degree r ≥ 1, if it is asymp- the uniqueness of this function. The implicit function theorem
totically stable, ∃σ ∈ K∞ and ∃α ∈ R+ such that the [14] and the conditions 1) - 4) guarantee continuity of σ in
inequality kx(t, ϕ)k ≤ σ(kϕkh )e−ρr (αt) , t ∈ R+ holds for R+ . In order to show that σ is strictly increasing let us take
all ϕ ∈ Ch : kϕkh < δ. If δ = +∞ then the origin of the two arbitrary numbers s1 , s2 ∈ R+ such that s1 < s2 and
system (1) is said to be globally hyper exponentially stable. denote σ1 = σ(s1 ) and σ2 = σ(s2 ). The condition 4) implies
For r = 0 the hyper exponential stability becomes exponen- q(σ1 , s1 ) = 0 < q(σ1 , s2 ). Hence, q(σ2 , s2 ) = 0 < q(σ1 , s2 )
tial one [12]. The number α is called the decay rate of hyper and due to the condition 3) we have σ1 < σ2 . Finally, the
exponential convergence. For example, the following delay- condition 5) guarantees that σ(s) → +∞ as s → +∞ and
free system ẋ(t) = −(1 − ln(|x(t)|))x(t), x ∈ R is locally the function σ can be prolonged by continuity to the origin as
hyper exponentially stable with the rate α = 1 and degree follows σ(0) = 0.
r = 1, i.e. kx(t)k ≤ σ(|x0 |)e−ρ1 (t) , where ρ1 (t) = et . Recall that the operator g : Z → Y, where Z and Y are
The next two definitions present a stability with non- Banach spaces, is called F-differentiable (Frèchet differen-
asymptotic reaching phase. tiable) at z0 ∈ Z if there exists a linear bounded operator
Definition 2 ([19]): The origin of the system (1) is said to Dgz0 : Z → Y
be finite-time stable, if it is Lyapunov stable and for any ϕ ∈ kg(z)−g(z0 )−Dgz0 (z−z0 )kY
→0 as kz − z0 kZ → 0
kz−z0 kZ
Ch : kϕkh < δ there exists 0 ≤ T0 (ϕ) < +∞, such tha
kx(t, ϕ)k = 0 for t ≥ T0 (ϕ). If δ = +∞ then the origin of where k · kZ and k · kY are norms in the Banach spaces Z and
the system (1) is said to be globally finite-time stable. Y, respectively.
The functional T : C[−h,0] → R+ defined as follows Theorem 1 ([7]): Let Z, Y and Q be Banach spaces. Let g :
T (ϕ) = inf{T0 (ϕ) ≥ 0 : x(t, ϕ) = 0, ∀t > T (ϕ)} is called Z×Y → Q be a continuous operator such that 1) g(z0 , y0 ) = 0
the settling-time functional [19] of the system (1). Examples for some z0 ∈ Z and y0 ∈ Y; 2) g is F-differentiable in
of finite-time stable systems of the form (1) are given in [19]. the neighborhood of the point (z0 , y0 ); 3) Dgz0 0 is invertible,
Preprint submitted to IEEE Transactions on Automatic Control. Received: April 3, 2015 13:47:47 PST
0018-9286 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TAC.2015.2422451, IEEE Transactions on Automatic Control
Limited circulation. For review only
3
IEEE-TAC Submission no.: 14-0202.4

where g 0 (z) := g(z, y0 ). Then there exists a unique operator t ∈ R+ . Obviously, Q̃(Ṽ (t), t) = 0 for all t ∈ R+ such
f : Y → Z defined in an open neighborhood U0 of y0 such that that xh (t) 6= 0, and the classical implicit function
h i theorem [6] −1
g(f (y), y) = 0 for all y ∈ U0 . Moreover, if g is continuously for Euclidean spaces implies dt d
Ṽ (t) = − ∂V ∂ Q̃ ∂ Q̃
∂t . Taking
F-differentiable at a neighborhood of (z0 , y0 ) then f is also
into account continuity of the functional V at 0 ∈ Ch and the
continuously differentiable at some neighborhood of y0 .
condition C5) we derive
h i−1
B. Theorems on Implicit Lyapunov-Krasovski Functional V̇ (xh (t)) = − ∂Q(V,x
∂V
h (t)) ∂Q(V,xh (t))
∂t ≤ −σ̃(kx(t)k),
The next theorem provides a background for stability anal- h i−1
∂Q
ysis of time-delay systems using ILKF. where σ̃ = − ∂V σ and t ∈ R+ . Finally, applying
Theorem 2: If there exists a continuous functional Q : R+ × Theorem 2.1, page 105 from [12] we finish the proof.
Ch → R such that: Theorem 2 introduces an additional flexibility to the design
C1) Q is continuously F-differentiable on R+ × Ch ; of Lyapunov-Krasovski functionals provided the possibility of
C2) for any y ∈ Ch there exists V ∈ R+ such that Q(V, y) = their implicit definition. However, the implicit design requires
0; a more complicated analysis based on conditions C1)-C5).
C3) ∂Q(V,y)
∂V < 0 for all V ∈ R+ and y ∈ Ch ; Note that the condition C3) can be relaxed as follows
C4) there exist qi ∈ IK∞ , i = 1, 2 such that for all V ∈ R+ ∂Q(V,y)
∂V < 0 for all (V, y) ∈ Ω. It is worth stressing out
q1 (V, ky(0)k) ≤ Q(V, y), ∀y ∈ Ch \C0h , that, similarly to the classical method of Lyapunov-Krasovski
Q(V, y) ≤ q2 (V, kykh ), ∀y ∈ Ch \{0}, functionals, the condition C5) can be checked using the right-
hand side of the equation (1) only (see the next section).
C5) for all (V, xh (t)) ∈ Ω such that x(t) satisfies (1) we have Theorem 3: If there exists a continuous functional Q :
∂Q(V,xh (t)) R+ × Ch → R, which satisfies the conditions C1)-C4) of
∂t ≤ −σ(kx(t)k), ∀t ∈ R+ ,
Theorem 2 and the condition:
where σ : R → R and C5bis) for all (V, xh (t)) ∈ Ω such that 0 < V < V max and
xh (t) satisfies (1) we have
Ω = {(V, y) ∈ R+ × Ch : Q(V, y) = 0} . (2)
∂Q(V,xh (t))
∂t ≤ αV 1−µ ∂Q(V,x
∂V
h (t))
, t ∈ R+ ,
Then the origin of the system (1) is Lyapunov stable if σ is a
nonnegative function and asymptotically stable if σ ∈ K. where α, V max ∈ R+ , µ ∈ (0, 1] and Ω is defined by (2).
Proof: In order to prove the theorem let us show that there Then the origin of system (1) is finite-time stable (globally
exists a functional V : Ch → R+ ∪ {0} that satisfies condi- finite-time stable if V max = +∞) with the following settling
tions of classical stability theorem for functional-differential time estimate:
equations (for example, Theorem 2.1 on page 105 from [12]). V0µ
I. Given ỹ ∈ Ch let us consider the scalar function Q(·, ỹ) T (ϕ) ≤ αµ , V0 ∈ R+ : Q(V0 , ϕ) = 0. (3)
and show that it has exactly one zero on R+ . The condition Proof: Repeating the proof of Theorem 2 we show
C2) claims that there exists at least one Ṽ ∈ R+ such that existence of a proper Lyapunov-Krasovski functional V :
Q(Ṽ , ỹ) = 0. The conditions C1) and C3) imply that the Ch → R+ ∪ {0} implicitly defined by the equation Q(V, y) =
function Q(·, ỹ) is monotone, i.e. Ṽ is the unique zero in R+ . 0. The condition C5bis) and hthe impliciti function theorem
Therefore, there exists a unique functional V : Ch \{0} → −1
[6] provide V̇ (xh (t)) = − ∂Q(V,x h (t)) ∂Q(V,xh (t))
≤
R+ such that Q(V (y), y) = 0 for any y ∈ Ch . Theorem 1 ∂V ∂t
1−µ
guarantee that V is continuously F-differentiable functional −αV (xh (t))). Hence, Proposition 4 presented in [19]
on Ch \{0}. implies finite-time stability of the origin of the system (1)
II. Since q1 , q2 ∈ IK∞ then there exists σ1 , σ2 ∈ K∞ with the settling time estimate (3).
such that q1 (σ1 (s), s) = 0 and q2 (σ2 (s), s) = 0 for all s ∈ Theorem 4: Let there exist two continuous functionals Q1 :
R+ . Note that the functions σ1 and σ2 can be extended by R+ × Ch → R and Q2 : R+ × Ch → R, which satisfy the
continuity to R+ ∪ {0} as follows σ1 (0) = 0 and σ2 (0) = 0. conditions C1)-C4) of Theorem 2. If the functional Q1 satisfies
The condition C4) implies q1 (V (y), ky(0)k) ≤ the condition C5bis) of Theorem 3 with σ(V ) = αV 1−µ , α ∈
Q(V (y), y) = 0 = q1 (σ1 (ky(0)k), ky(0)k) for all R+ , µ ∈ (0, 1], V max = 1, and
y ∈ Ch \C0h and q2 (σ2 (kykh ), kykh ) = 0 = Q(V (y), y) ≤ C6) Q1 (1, y) = Q2 (1, y) for all y ∈ Ch : kyk < δ, where
q2 (V (y), kykh ) for all y ∈ Ch \{0}. Due to the condition δ ∈ R+ is some constant;
3) of Definition 4 the obtained inequalities guarantee C7) for all (V, xh (t)) ∈ Ω2 such that V ≥ 1 and xh (t) satisfies
σ1 (ky(0)k) ≤ V (y) for all y ∈ Ch \C0h and V (y) ≤ σ2 (kykh ) (1) we have
for all y ∈ Ch \{0}. Hence, the functional V can be ∂Q2 (V,xh (t))
≤ βV 1+ν ∂Q2 (V,xh (t))
, t ∈ R+
∂t ∂V
extended by continuity to Ch as follows V (0) = 0. Taking
into account V (y) ∈ R+ for any y ∈ C0h \{0} we derive where β, ν ∈ R+ and Ω2 is defined by (2) with Q = Q2 ,
σ1 (ky(0)k) ≤ V (y) ≤ σ2 (kykh ) for all y ∈ Ch . then the origin of system (1) is (globally) fixed-time stable (if
1 1
III. Let x(t) be a solution of (1) and xh (t) ∈ Ch be defined δ = +∞) with the settling time estimate: T (ϕ) ≤ αµ + βν .
by xh (t) := x(t + τ ) with τ ∈ [−h, 0]. Let us consider the Proof: The conditions C1)-C4) of Theorem 2 guarantees
functions Ṽ (t) := V (xh (t)) and Q̃(V, t) := Q(V, xh (t)) for existence of proper Lyapunov-Krasovski functionals V1 :
Preprint submitted to IEEE Transactions on Automatic Control. Received: April 3, 2015 13:47:47 PST
0018-9286 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TAC.2015.2422451, IEEE Transactions on Automatic Control
Limited circulation. For review only
4
IEEE-TAC Submission no.: 14-0202.4

Ch → R+ ∪ {0} and V2 : Ch → R+ ∪ {0} implicitly V. E XAMPLE : H YPER E XPONENTIAL S TABILIZATION

defined by the equations Q1 (V1 , y) = 0 and Q2 (V2 , y) = 0 A. Problem formulation
for any y ∈ Ch . Given y0 ∈ Ch the inequality V1 (y0 ) ≥ 1
Let us consider a single input control system of the form
holds if and only if V2 (y0 ) ≥ 1. Indeed, if V1 (y0 ) ≥ 1 then 
the conditions C3) and C6) imply Q2 (1, y0 ) = Q1 (1, y0 ) ≥ ẋ(t) = A0 x(t) + A1 x(t − h) + bu(t), t > 0,
(4)
Q1 (V1 (y0 ), y0 ) = 0 = Q2 (V2 (y0 ), y0 ) and V2 (y0 ) ≥ 1. x(t) = ϕ(t) ∈ Ch , t ∈ [−h, 0],
Analogously, given y0 ∈ Ch the inequality V1 (y0 ) < 1 holds where x ∈ Rn is the state vector, u ∈ R is a control input,
if and only if V2 (y0 ) < 1. A0 , A1 ∈ Rn×n
 are the system matrices of the form A0 =
The condition C7) implies V̇2 (xh (t)) ≤ −βV21+ν (xh (t)) for

0 In−1
V2 ≥ 1. Hence, for V2 (xh (0)) > 1 we have V2 (xh (t)) < 1, , A1 = diag{ai }ni=1 , ai ∈ R, i = 1, 2, ..., n and
0 0
1
∀t > T2 = βν . the vector b ∈ Rn×1 is of the form b = (0, 0, ..., 1)> .
The condition C5bis) of Theorem 3 implies that if The problem is to stabilize hyper exponentially the origin
V1 (xh (t1 )) < 1 for some t1 ∈ R+ ∪ {0}, then we have of the system (4).
1
V1 (xh (t)) = 0, ∀t > t1 + T1 , where T1 = αµ .
Implicit Lyapunov-Krasovski functional-based analysis of
B. Hyper Exponential Control Design
exponential and hyper exponential stability requires some
additional restrictions to the class of functionals to be used. Introduce the Implicit Lyapunov-Krasovski Functional
Theorem 5: If there exist a continuous functional Q : R+ × (ILKF) by the equality
Q(V, y) := −1 + y > (0)D V1 P D V1 y(0)+



Ch → R, which satisfies the conditions C1) - C4) of Theorem
2 and the conditions: R0 e
µ 2τh+h > (5)
y (τ )D V1 SD V1 y(τ )dτ,



C4∗ ) there exists c ∈ R+ such that q1 (cs, s) ≥ 0 for 0 < cs < V
−h
Vmax , where q1 is defined by the condition C4) of Theorem 2
and V max ∈ R+ is some constant number; where µ ∈ (0,  1), the diagonal
n matrix D(·) is defined by
C5∗ ) for all (V, xh (t)) ∈ Ω and 0 < V < V max such that D(λ) = diag λ1+µ(n−i) i=1 for λ ∈ R+ and 0 < P =
x(t) satisfies (1) we have P > ∈ Rn×n , 0 < S = S > ∈ Rn×n . Denote H :=
diag{1 + µ(n − i)}ni=1 .
∂Q(V,xh (t)) ∂Q(V,xh (t))
∂t ≤ αp(V )V ∂V , ∀t ∈ R+ , If µ = 0 then the equation Q(V, y) = 0 provides
q well-known RLyapunov-Krasovski functional: V (y) =
the
where α ∈ R+ is some constant number, then the origin of 0
y > (0)P y(0) + −h y > (τ )Sy(τ )dτ .
system (1) is
The next proposition presents an algorithm of the ILKF-
• exponentially stable if p(V ) = 1; based control design.
• hyper exponentially stable of degree r = 1 if Proposition 1: If X ∈ Rn×n , Z ∈ Rn×n and y ∈ R1×n
 satisfy the following system of LMIs:
1 + ln(V ) for V > 1,
p(V ) = 
A0 X+XA> > > 2eµ µ >

1 − ln(V ) for V ≤ 1, 0 +by+y b + h X+e A1 ZA1 X
≤ 0,
−1
X eµ Z
with the convergence rate α. The global (hyper) exponential (6)
stability is provided if V max = +∞. γX ≥ XH +HX > 0, X > 0, ZH +HZ −µZ > 0, Z > 0.
Proof: Repeating the proof of Theorem 2 we show exis-
then the functional Q defined by (5) with P = X −1 and S =
tence of a proper Lyapunov-Krasovski functional V : Ch →
Z −1 satisfies the conditions C1)-C4) of Theorem 2 and the
R+ ∪ {0} implicitly defined by the equation Q(V, y) = 0.
control defined by
The condition C4∗ implies that q1 (cs, s) ≥ q1 (σ1 , s) for
all (σ1 , s) ∈ R2+ : q1 (σ1 , s) = 0. Hence, due to the u(V, x(t)) = V 1−µ kD(V −1 )x(t), (7)
condition 3) of Definition 4 a function σ1 ∈ K∞ such that
q1 (σ1 (s), s) = 0, s ∈ R+ satisfies the inequality cs ≤ σ1 (s) with k = yX −1 , V ∈ R+ : Q(V, xh (t)) = 0
for all s ∈ R+ : σ1 (s) ≤ Vmax . stabilizes the origin of the system (4) hyper
Let us denote for shortness V (t) = V (xh (t)) and exponentially with degree r = 1, the decay rate
2µ


V0 = V (xh (0)). The condition C5∗ ) implies V̇ (t) ≤ α = γh , where γ = max λmax X 1/2 H+HX 1/2 ,


−αp(V (t))V (t) along the trajectories of (1). For p(V ) = 1 µ+λmax Z 1/2 H+HZ 1/2 , and the domain of
we have V (t) ≤ V0 e−αt . For p(V ) = 1 + ln(V ) and initial conditions Π = {ϕ ∈ Ch : Q(1, ϕ) < 0}. The
−αt
V0 > 1 we derive V (t) ≤ ee ln(eV0 )−1 ≤ V0 e1−e , 0 ≤
αt
control (7) is a continuous p function of time and
t ≤ T1 , where T1 := α−1 ln(ln(eV0 )) and V (T1 ) ≤ 1. ku(V, xh (t))k ≤ V 1−µ λmax (P −1/2 k T kP −1/2 ) for
For p(V ) = 1 − ln(V ) and V0 ≤ 1 we obtain V (t) ≤ Q(V, xh (t)) = 0.
αt αt
ee ln(V0 /e)+1 ≤ V0 e1−e , ∀t > 0. Taking into account Proof: I. The functional Q(V, y(·)) defined by (5) is
cky(0)k ≤ σ1 (ky(0)k) ≤ V (y) we finish the proof of the continuously F-differentiable in R+ ×Ch . Since P > 0, Q > 0
(hyper) exponential stability. and µ ∈ (0, 1) then qi ∈ IK∞ , i = 1, 2 and the inequalities
2τ +h
The next section presents an example of control design R0 e
λmin (P )ky(0)k2 +λmin (S) -h ( V ) hµ-1 ky(τ )k2 dτ
using Theorem 5. max{V 2+2µ(n−1) ,V 2 }
≤ Q(V, y(·))+1 ≤
Preprint submitted to IEEE Transactions on Automatic Control. Received: April 3, 2015 13:47:47 PST
0018-9286 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TAC.2015.2422451, IEEE Transactions on Automatic Control
Limited circulation. For review only
5
IEEE-TAC Submission no.: 14-0202.4
R 0 e µ 2τ +h
λmax (P )ky(0)k2 +λmax (S) -h ( V ) h ky(τ )k2 dτ Hence, for Q(V, y) = 0 we have ku(V, x(t))k2 ≤
2+2µ(n−1) ,V 2 }
min{V
λmax (P −1/2 k T kP −1/2 )V 2−2µ . Since V (·) is a continuous
hold for all V ∈ R+ and y ∈ Ch . Hence, it is easy to see that functional then the function u(·) := u(V (xh (·)), xh (·)) is the
for any y ∈ Ch there exists V ∈ R+ such that Q(V, y) = 0. continuous function and u(t) → 0 as V (xh (t)) → 0.
Introduce the functions q1 : R+ × R+ → R and q2 : R+ × The presented control scheme admits LMI constraints to
2
R+ → R by the formulas q1 (σ1 , s) = n λmin (P )s o − parameters of nonlinear hyper exponential (fast) feedback. Any
2 2+2µ(n−1)
max σ1 , σ1
existing LMI solver can be used in order to adjust the control
(λmax (Pn)+λmax (S)eµ σ2µoh)s2 parameters. Note that introducing an additional LMI
1 and q2 (σ2 , s) = 2+2µ(n−1) − 1, where
min σ22 , σ2
X yT
 
σ1 , σ2 , s ∈ R+ . The obtained estimates also guarantee that
≥ 0, u0 ∈ R+ (8)
q1 (V, ky(0)k) ≤ Q(V, y(·)) ≤ q2 (V, kykh ) for all V ∈ R+ y u20
and y ∈ Ch . The derivative of Q(·, y) for y ∈ Ch is
to (6) we guarantees ku(t)k ≤ u0 for V ≤ 1. Indeed, the
> −1 −1
∂Q(V,y)
∂V = − y (0)D(V )(HPV+P H)D(V )y(0) − inequality λmax (P −1/2 k T kP −1/2 ) ≤ u20 can be equivalently
R0 e µ 2τh+h >
y (τ )D(V −1 )(HS+SH)D(V −1 )y(τ )dτ rewritten as y T y ≤ u20 X and Schur Complement implies (8).
−h ( V )
V −
R 0 2τ +h e µ 2τ +h > −1 −1
µ −h h ( V ) h y (τ )D(V )SD(V )y(τ )dτ
V . C. Practical implementation
∂Q In order to apply the control (7) in practice it is necessary
The inequalities (6) for P := X −1 and S = Z −1 imply < ∂V
0 for all V ∈ R+ and y ∈ Ch \{0}. Therefore, the conditions to present the procedure for calculation of the value V for any
C1)-C4) of Theorem 2 hold. Obviously, the condition C4*) of given y(·) ∈ Ch such that Q(V, y(·)) = 0. Since Q satisfies
the condition C3) of Theorem 2 then for any y(·) the ILF
p
Theorem 5 also holds for c = λmax (P ) and V max = 1.
II. If x(t) is a solution to the system (4), (7) then equation has a unique solution, which can be easily found by
means of, for example, the bisection method.
∂Q(V,xh (t))
= ∂t An alternative approach is to realize the ILKF-based control
> −1 −1
2x (t)D(V )P D(V + A1 x(t − h) + bu(t))+
)(A0 x(t) scheme using a Newton method for numerical calculation of
eµ V −µ x> (t)D(V )SD(V −1 )x(t)−
−1
the solution of the equation Q(V, y(·)) = 0, for example,
V e x (t − h)D(V −1 )SD(V −1 )x(t − h)−
µ −µ >
h i−1
0
2µ ln( V ) R
e
2τh+h > V̇n = −pQ(Vn , y(·)) ∂Q(V∂Vnn,y(·)) , p ∈ R+ . (9)
h
e
V x (t+τ )D(V −1 )SD(V −1 )x(t+τ )dτ.
−h
For any fixed y(·) the presented gradient method guarantees
Taking into account that D(V −1 )A0 D−1 (V −1 ) = V −1 A0 , Vn → V : Q(V, y(·)) = 0. In order to boost the convergence
D(V −1 )A1 D−1 (V −1 ) = A1 , D(V −1 )bu = of Newton method its fast modification can be considered (see,
−µ −1
V bkD(V )x(t) we obtain for example, [5]). In practice, the ILKF control can be realized
∂Q(V,xh (t)) 2µ ln( Ve ) in a sufficiently fast digital control device.
∂t =− h + q > W q+
eµ
2
µ ln V − V µ xT (t)D(V −1 )P D(V −1 )x(t),
e



h
D. Numerical Simulation
where
 −µ  Consider the problem of hyper exponential stabilization of
P (A0 +bk) + (A0 +bk)> P + S + h2 P P A1
 
e the system (4) with n = 2, the matrix A1 = diag{1, 0}, µ =
W= ,
A> 1P −S 0.75, u0 = 50 and the delay h = 0.5. By solving the
LMI
 q µ  system (6) we obtain K = −82.3072 −14.6158 and
e −1
V µ D(V )x(t)
q= q µ .    
V −1
6.3785 0.7197 1.5866 0.0032
D(V )x(t − h) P = , S= .
eµ 0.7197 0.1081 0.0032 0.0008
µ
Taking into account µ ln Ve ≤ Ve µ for V ∈ R+ the


h i−1 The simulations for the ILKF control application have
LMIs (6) and Schur Complement imply − ∂V ∂Q
≥ been made in MATLAB Simulink. The value of ILKF was
∂Q(V,xh (t)) −1 calculated on-line using the Newton method (9) with p =
V
γ and ∂t ≤ − 2µ ln(eV
h
)
. Therefore, V̇ = 103 . All integrals, which appear in (9), were approximated
i−1
2µ ln(eV −1 )V
h
− ∂V ∂Q ∂Q(V,xh (t))
≤− , for V (y) ≤ 1. The- by Simpson formula. In order to avoid singularities during
∂t γh
orem 5 implies that the system (4), (7) is hyper exponentially calculation of Q(V, y(·)) the admissible value of V is also
stable with the domain of hyper exponential attraction Π (or bounded from below by Vmin = 0.01. In this case the hyper
equivalently for V (·) ≤ 1). The estimate exponential convergence can only be guaranteed to a zone
p of the hyper expo- defined by inequality V (y) ≤ Vmin . However, such assumption
nential convergence rate follows from λmin (P )ky(0)k ≤ V .
is not restrictive from practical point of view, since due to
III. The control function u is locally bounded. Indeed,
measurement noises and uncertainties just a practical stability
ku(V, x(t))k2 = V 2−2µ xT (t)D(V −1 )k T kD(V −1 )x(t) is possible in real-life systems. The results of simulations are
≤ λmax (P -1/2 k T kP -1/2 )V 2−2µ xT (t)D(V -1 )P D(V -1 )x(t). depicted in Figure 1.
Preprint submitted to IEEE Transactions on Automatic Control. Received: April 3, 2015 13:47:47 PST
0018-9286 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TAC.2015.2422451, IEEE Transactions on Automatic Control
Limited circulation. For review only
6
IEEE-TAC Submission no.: 14-0202.4

VI. C ONCLUSIONS [7] B.K. Driver. Analysis Tools with Applications. Springer,
This technical note presents the method of the implicit 2003.
Lyapunov-Krasovski functionals and demonstrates its possible [8] D. Efimov, A. Polyakov, E. Fridman, W. Perruquetti,
application to control design. The obtained stability theorems and J.-P. Richard. Comments on finite-time sta-
extend the results of the paper [23] to time-delay systems. bility of time-delay systems. Automatica, (DOI:
The main difficulty of this extension is the representation 10.1016/j.automatica.2014.05.010), 2014.
of all conditions of Lyapunov-Krasovski theorems using the [9] E. Fridman. Descriptor discretized lyapunov functional
implicit functions in Banach spaces (see, conditions C1)-C5) method: Analysis and design. IEEE Transactions on
of Theorem 2). The special class of implicit functions IK∞ Automatic Control, 51(5):890–897, 2006.
is introduced for this purpose. [10] K. Gu, V. Kharitonov, and J. Chen. Stability of Time-
The application ILKF method for fast (hyper exponential) Delay Systems. Boston, MA: Birkhauser, 2003.
stabilization of a time-delay system is demonstrated on an [11] V.T. Haimo. Finite time controllers. SIAM Journal of
example. The corresponding ILKF has been obtained by means Control and Optimization, 24(4):760–770, 1986.
of an appropriate parametrization of the classical explicit [12] J. Hale. Theory of Functional Differential Equations.
Lyapunov-Krasovski functional. The hyper exponential sta- Springer-Verlag, 1977.
bilization of a more general time-delay model is a possible [13] Y. Hong. Finite-time stabilization and stabilizability of a
direction of future research. class of controllable systems. Systems & Control Letters,
46(4):231–236, 2002.
1 3
[14] K. Jittorntrum. An implicit function theorem. Journal
2 of Optimization Theory and Applications, 25(4):575–577,
0.5 1978.
1
[15] V.L. Kharitonov. Robust stability analysis of time delay
systems: A survey. Annual Review in Control, 23:185–
x1

x2

0 0

196, 1999.
−1
−0.5
[16] V.I. Korobov. Solution of a synthesis problem using a
−2 controllability function. Doklady Academii Nauk SSSR,
248:1051–1055, 1979.
−1 −3
0 0.5 1 1.5 2
t
0 0.5 1
t
1.5 2 [17] M. Krstic. Delay Compensation for Nonlinear, Adaptive,
50
and PDE Systems. Birkhauser Boston, 2009.
[18] J. Macdonnell. Some critical points on the hyperpower
x...
25 function n x = xx . International Journal of Mathe-
matical Eduction, 20(2):297–305, 1989.
0 [19] E. Moulay, M. Dambrine, N. Yeganefar, and W. Perru-
u

quetti. Finite-time stability and stabilization of time-delay

−25
systems. System & Control Letters, 57(7):561–566, 2008.
[20] E. Moulay and W. Perruquetti. Finite-time stability and
stabilization: State of the art. Lecture Notes in Control
−50
0 0.5 1 1.5 2
t and Information Sciences, 334:23–41, 2006.
[21] Y. Orlov. Discontinous systems: Lyapunov analysis and
Fig. 1. The simulation results (x1 - blue line; x2 - green line; u - red line).
robust synthesis under uncertainty conditions. Springer-
Verlag, 2009.
[22] A. Polyakov. Nonlinear feedback design for fixed-time
R EFERENCES stabilization of linear control systems. IEEE Transactions
[1] J. Adamy and A. Flemming. Soft variable-structure on Automatic Control, 57(8):2106–2110, 2012.
controls: a survey. Automatica, 40:1821–1844, 2004. [23] A. Polyakov, D. Efimov, and W. Perruquetti. Finite-time
[2] A. Bacciotti and L. Rosier. Lyapunov Functions and stabilization using implicit lyapunov function technique.
Stability in Control Theory. Springer, 2nd edition, 2005. In 9th Symposium on Nonlinear Control Systems,, pages
[3] S.P. Bhat and D.S. Bernstein. Finite-time stability of con- 140–145, 2013.
tinuous autonomous systems. SIAM Journal of Control [24] A. Polyakov, D. Efimov, W. Perruquetti, and J.-P.
and Optimization, 38(3):751–766, 2000. Richard. Implicit Lyapunov-Krasovski Functionals for
[4] B. Biemond, N. van de Wouw, and H. Nijmeijer. Bifur- Time Delay Systems. In 53rd Conference on Decision
cations of equilibrium sets in mechanical systems with and Control, 2014.
dry friction. Physica D, 241:1882–1894, 2012. [25] J.-P. Richard. Time-delay systems: An overview of
[5] J. Cortes. Finite-time convergent gradient flows some recent advances and open problems. Automatica,
with applications to network consensus. Automatica, 39(10):1667–1694, 2003.
42(11):1993–2000, 2006. [26] E. Roxin. On finite stability in control systems. Rendi-
[6] R. Courant and F. John. Introduction to calculus and conti del Circolo Matematico di Palermo, 15(3):273–283,
analysis (Vol. II/1). New York: Springer, 2000. 1966.
Preprint submitted to IEEE Transactions on Automatic Control. Received: April 3, 2015 13:47:47 PST
0018-9286 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.