Emilia 2024
Emilia 2024
Abstract— Recently, a constructive approach to averaging- efficient upper bound on the small parameter ε. This time-
based stability was proposed for linear continuous-time sys- delay approach was also employed in [3], [7] for averaging
tems with small parameter ε > 0 and rapidly-varying almost of systems with constant/time-varying delays. Moreover,
periodic coefficients. The present paper extends this approach
to discrete-time linear systems with rapidly-varying periodic input-to-state stability (ISS), L2 -gain analysis and stochastic
coefficients. We consider linear systems with state delays, where extension of the time-delay approach were presented in [19].
results on the stability via averaging are missing. Differently In [18], the time-delay approach to averaging approach was
from the continuous-time, our linear matrix inequalities (LMIs) extended to discrete-time systems, also it was extended to
are feasible for any delay (i.e. the system is exponentially stable) ISS analysis of the perturbed systems, as well as to obtain
provided ε is small enough. We introduce an efficient change
of variables that leads to a perturbed averaged system, and practical stability of discrete-time switched affine systems.
employ Lyapunov analysis to derive LMIs for finding maximum Recently, a novel constructive approach for linear
values of the small parameter ε > 0 and delay that guarantee continuous-time systems with rapidly-varying almost peri-
the exponential stability. Numerical example illustrates the odic coefficients was introduced. Differently from the time-
effectiveness of the proposed approach. delay approach, the method of [9] relies on a novel non-
delayed transformation which yields simpler analysis and
I. I NTRODUCTION essentially less conservative results in the numerical exam-
ples. This approach is applicable to averaging of systems
Time-varying control systems with almost periodic coef- with both constant and time-varying delays, where for the
ficients arise in many modern engineering applications in- discrete-time the results are missing.
cluding satellite attitude and hypersonic vehicle flight control Our objective in this paper is to extend the approach
systems [1], [4], [5], [15], [16], [17]. Over the last few of [9] to the discrete-time systems, including systems with
decades these systems received a lot of attention from the constant delays. Although the fundamental ideas are inspired
control community [8], [11],[12], [13]. One of the most by the continuous case [9], construction of the appropriate
efficient methods for stability analysis of such systems is transformations and the subsequent Lyapunov analysis are
the method of averaging [2], [11], [14]. The main idea not immediately extendable from the continuous framework,
behind the averaging method relies on the approximation but rather require significant adaptation to the discrete-time
of the solutions of a time-varying system by solutions of a case. Linear discrete-time delayed systems with periodic co-
corresponding averaged system. The exponential stability of efficients are considered. Differently from [18], we start with
the averaged system guarantees the asymptotic stability of a new presentation of the system, where the system matrix
the original time-varying system for small enough parameter is presented as a linear combination of constant matrices
ε > 0. However, one of the main disadvantages of the multiplied by scalar rapidly-varying terms with zero average.
classical averaging method is the inability of providing an We then suggest a new discrete-time transformation of the
efficient quantitative upper bound on the small parameter ε rapidly-varying coefficient, and employ a direct Lyapunov
for which stability of the original system is preserved. method leading to stability conditions in the form of LMIs.
Recently, a constructive time-delay approach has been The LMIs are accompanied by theoretical guarantees on their
introduced for the periodic averaging of continuous-time feasibility for small enough values of the system parameters.
systems [7]. By averaging the system backwards in time, the Furthermore, differently from the continuous-time delayed
system was transformed into a model with time-delays of the case, we present conditions that guarantee stability of the
length of the small parameter. Stability of the transformed discrete-time system for arbitrary delay, provided ε is small
system was shown to imply the stability of the original enough. Numerical example illustrates the efficiency of the
system [7]. Then, direct Lyapunov-Krasovkii method was suggested method.
applied to obtain LMI-based conditions that guarantee input Notation: Rn denotes the n-dimensional Euclidean space
to state stability of the transformed system and provide an with norm | · |, and | · |1 is l 1 norm, Rn×m is the set of all
n × m real matrices with the induced matrix norm || · ||, 0n
∗ This work was supported by Israel Science Foundation (grant no.
and In are the zero matrix and the identity matrix of order
673/19), ISF-NSFC joint research program (grant no. 3054/23), Chana and n, respectively. Z+ is the set of non-negative integers. The
Heinrich Manderman Chair on System Control at Tel Aviv University. notation P > 0 for P ∈ Rn×n means that P is symmetric
† These authors contributed equally. and positive definite. col{} is a column array of column
a,b,c - School of Mathematical Sciences, Tel-Aviv Univer-
sity, Tel Aviv, Israel, (A. Jbara) adamjbara@mail.tau.ac.il, (R. Katz) scalars/vectors. The sub-diagonal elements of a symmetric
ramikatz@mail.tau.ac.il, (E. Fridman ) emilia@tauex.tau.ac.il. matrix are denoted by ∗, the superscript T denotes matrix
Consider the discrete-time system: Using equations (5), (7) and (8), we obtain
xk+1 = [I + εA(k)]xk , (2) 2 2
zk+1 − zk = εAav zk − ε ∑ ∑ A j Am ρ j (k + 1)am (k)xk
where xk ∈ Rn ,
A(k) : Z+ → Rn×n ,
ε > 0 is a small parameter. j=1 m=1
System (2) (and further System (30)) can be regarded as the 2 2
discretization of the continuous-time system (2.1) ( system +ε ∑ ρ j (k)Aav A j xk − ε ∑ A j Aav ρ j (k + 1)xk . (12)
(3.1)) in [9]. j=1 j=1
ε k+T Since ρ j (k) = O(ε), one has in (15): |Hρ |1 = O(ε 2 ) and
ρ j (k + 1) − ρ j (k) = − ∑ (k + T − i)a j (i) (7) |Hρ,a |1 = O(ε 2 ). Then, for all positive diagonal matrices
T i=k+1 (1) (2)
k+T k+T
Λρ , Λρ ∈ R2×2 and Λρ,a ∈ R4×4 the following hold:
ε ε
− ∑ a j (i) + ∑ (k + T − i)a j (i) = εa j (k). (1) (1) (1) (1)
T i=k+1 T i=k
Yρ (k)T (Λρ ⊗ In )Yρ (k) ≤ |Λρ Hρ |1 |xk |2 ,
(2) (2) (2) (2)
for all j ∈ {1, 2, ..., N} and k ∈ Z+ . Yρ (k)T (Λρ ⊗ In )Yρ (k) ≤ |Λρ Hρ |1 |xk |2 , (17)
Introduce the change of variables Yρ,a
T
(k)(Λρ,a ⊗ In )Yρ,a (k) ≤ |Λρ,a Hρ,a |1 |xk |2 .
N (1) (2)
zk = xk − ∑ ρ j (k)A j xk , (8) The matrices Λρ , Λρ and Λρ,a will be decision variables
j=1 in the LMIs derived below (see (27), (28)).
1656
C. Lyapunov analysis where
For stability analysis of (14) subject to (13), we introduce φ1 φ2 φ3
the Lyapunov function Ψ ε = ∗ φ4 ε(I2 ⊗ Aav )T A T PA1 ,
∗ ∗ −(Λρ,a ⊗ In ) + εA1T PA1
V (k) = |zk |2P , P > 0 (18)
B = β2 β3 β4 ,
and a decay rate α := 1 − εθ , where 0 ≤ θ < 1/ε. Denote 2
( j)
β1 = Qθ + ∑ |Λρ Hρ |1 In + |Λρ,a Hρ,a |1 In ,
Qθ (ε) := ATav P + PAav + θ P + εATav PAav . (19) j=1
1657
B. System transformation C. Lyapunov analysis
We modify the transformation (8) to account for the delay: For stability analysis of (35) subject to (37), introduce the
Nd Lyapunov function
zk = xk − ∑ ρm (k)Am xk−d , k ≥ d. (32) 2
!
m=1 V (k) = VP (k) + ε ∑ VSm (k) +VR (k) , (40)
For simplicity of the presentation, we assume Nd = 2. Let m=1
VP (k) = |zk |2P , P > 0, (41)
ξk = xk − xk−d , k ∈ Z+ .
k−1
By employing (4), (7) and (30) - (32) we obtain VSm (k) = ∑ α k−i−1 |xi |2Sm , m = 1, 2, Sm > 0, (42)
i=k−m·d
zk+1 − zk = εAav zk − εAd ξk −1 k−1
+ε ∑2m=1 Aav Am ρm (k)xk−d VR (k) = d ∑ ∑ α k−s−1 |xs+1 − xs |2R , R > 0. (43)
−ε ∑2m=1 Am A0 ρm (k + 1)xk−d (33) i=−d s=k+i
−ε ∑2m=1 Am Ad ρm (k + 1)xk−2d and a desired decay rate α := 1 − εθ , where 0 ≤ θ < 1/ε.
−ε ∑2m=1 ∑2i=1 Am Ai ρm (k + 1)ai (k − d)xk−2d , Here VS1 and VR compensate xk−d , whereas VS2 compensates
Since ρ j (k) = O(ε), equation(33) has the form xk−2d in the stability analysis. Then,
zk+1 − zk = εAav zk − εAd ξk + O(ε 2 ). VP (k + 1) − αVP (k) = ε|zk |2Qθ + ε 2 ξkT ATd PAd ξk
Denote + ε 2 Yρ,d (k)T Aρ,d
T
PAρ,d Yρ,d (k)
+ ε 2 Yρ,a,d (k)T A1T PA1 Yρ,a,d (k)
Ya,d (k) = col{a j (k)xk−d }2j=1 ,
(m) + ε 2 Yρ,2d (k)T [A (I2 ⊗ Ad )]T P[A (I2 ⊗ Ad )]Yρ,2d (k)
Yρ,d (k) = col{ρ j (k + m − 1)xk−d }2j=1 , m = 1, 2,
− 2εzTk (I + εAav )T PAd ξk
( j)
Yρ,d (k) = col{Yρ,d (k)}2j=1 , (34) + 2εzTk (I + εAav )T PAρ,d Yρ,d (k)
Yρ,2d (k) = col{ρ j (k + 1)xk−2d }2j=1 , − 2εzTk (I + εAav )T PA (I2 ⊗ Ad )Yρ,2d (k) (44)
Yρ,a,d (k) = col{ρ1 (k + 1)a1 (k − d)xk−2d , − 2εzTk (I + εAav )T PA1 Yρ,a,d (k)
ρ1 (k + 1)a2 (k − d)xk−2d , ρ2 (k + 1)a1 (k − d)xk−2d , − 2ε 2 ξkT ATd PAρ,d Yρ,d (k) + 2ε 2 ξkT ATd PA1 Yρ,a,d (k)
ρ2 (k + 1)a2 (k − d)xk−2d }, + 2ε 2 ξkT ATd PA (I2 ⊗ Ad )Yρ,2d (k)
A = [A , 0n×n , 0n×n ], Aρ,d = [Aav A , −A (I2 ⊗ A0 )].
˜
− 2ε 2 Yρ,d (k)T Aρ,d
T
PA (I2 ⊗ Ad )Yρ,2d (k)
Then, (33) can be presented as − 2ε Yρ,d (k) Aρ,d PA1 Yρ,a,d (k)
2 T T
zk+1 − zk = εAav zk − εAd ξk + εAρ,d Yρ,d (k) + 2ε 2 Yρ,2d (k)T [A (I2 ⊗ Ad )]T PA1 Yρ,a,d (k),
− εA (I2 ⊗ Ad )Yρ,2d (k) − εA1 Yρ,a,d (k), (35)
where Qθ is defined in (19). Substituting zk = xk − A˜Yρ,d (k)
whereas due to (30), (32) we have in (44) , we get
1658
Define exists M > 0 such that for all ε ∈ (0, ε ∗ ] and 0 ≤ d ≤ d ∗ , the
solution of (30) initialized at {x j }0j=−d satisfies
W := −εηkT Πηk + ελ2d |xk−2d |2 + ελd |xk − ξk |2 ,
λd := |Λa Ha |1 + |Λρ,d (I2 ⊗ Hρ )|1 , |xk |2 ≤ M |x|2[−d,0] (1 − θ ε)k−d , ∀d ≤ k ∈ Z+ . (55)
λ2d := |Λρ,2d Hρ |1 + |Λρ,a Hρ,a |1 , (51) Moreover, if (53) and δ2 < α d hold with ε = ε ∗ , d = d ∗ and
Π = diag{0n , Π },(1) θ = 0, then (30) is exponentially stable for all ε ∈ (0, ε ∗ ]
and 0 ≤ d ≤ d ∗ . Also, given any d, the inequalities (53) and
Π(1) := diag{0, Λρ,d , Λρ,2d , Λa , Λρ,a } ⊗ In .
δ2 < α d are always feasible for small enough ε and θ .
Then, (39) implies that W ≥ 0. Using (39)-(51) and the S-
Proof. The fact that feasibility of (53) and δ2 < α d with
procedure [6], we arrive at
ε ∗ , d ∗ implies feasibility for all ε < ε ∗ , d < d ∗ , follows by
V (k + 1) − αV (k) ≤ V (k + 1) − αV (k) +W monotonicity of (53), and δ2 < α d with respect to ε < ε ∗ ,
d < d ∗ (i.e., as the small parameters decrease, the eigenvalues
≤ εηkT (Θε,d + ε 2 d 2 L T RL )ηk of (53) are non-increasing).
T
+ εxk−2d (−α 2d S2 + λ2d In )xk−2d ≤ 0, (52) Feasibility of (53), and δ2 < α d implies that for all d ≤
k ∈ Z+ ,
provided
V (k + 1) − αV (k) ≤ 0 ⇒ V (k + 1) ≤ α k+1−d V (d),
Θε,d + ε 2 d 2 L T RL < 0, −α 2d S2 + λ2d In < 0, (53) d−1
V (d) = |zd |2P + ∑2m=1 ∑i=d−m·d α d−i−1 |xi |2Sm
−1 d−1
where +d ∑i=−d ∑s=d+i α d−s−1 |xs+1 − xs |2R .
(56)
β B Also, V (k) ≥ σmin (P)|zk |2 , for any d ≤ k ∈ Z+ . Thus, there
= 1
Θε,d , B = β2 β3 β4 0n×2n β5 ,
∗ Ψε,d exists some M1 > 0 such that
ω1 ω2 ω3 0n×2n εATd PA1
∗ ω4 ω5 04n×2n ω6 |zk |2 ≤ M1 |x|2[−d,0] α k−d , d ≤ k ∈ Z+ . (57)
(1)
Ψε,d = −Π + ∗ ∗ ω7 02n×2n ω8 , To conclude the same for the solution xk of the system (30),
∗ ∗ ∗ 02n×2n 02n×4n for any i ∈ Z+ , we denote Xi = |x|2[id,(i+1)d] . From (9), (33)
∗ ∗ ∗ ∗ εA1T PA1 and (57), we find that
β1 = Qθ + (1 − α d )S1 + S2 + λd In , α = 1 − θ ε,
Xi+1 ≤ M2 α id + δ2 Xi , i ∈ Z+ ,
d T
β2 = α S1 − (In + εAav ) PAd − λd In ,
where M2 = M1 |x|2[−d,0] . Set Y1 = X1 and consider the linear
β3 = −Qθ A˜ + (In + εAav )T PAρ,d , difference equation
β4 = −(In + εAav )T PA (I2 ⊗ Ad ),
Yi+1 = M2 α id + δ2Yi , i ∈ Z+ . (58)
β5 = −(In + εAav )T PA1 ,
By induction, we obtain Xi ≤ Yi for all i ∈ Z+ . Moreover,
ω1 = εATd PAd − α d (S1 + R) + λd In ,
the solution of (58) is given by Yi = µd α (i−1)d + δ2i−1 (X1 −
ω2 = AT P(In + εAav )A˜ − εAT PAρ,d ,
d d (54) M2 α
µd ), i ∈ Z+ , where µd = α−δ . Note that Yi is decreasing.
2
ω3 = εAd PA (I2 ⊗ Ad ), ω4 = εAρ,d PAρ,d + A Qθ A
T T ˜T ˜ Let k ∈ Z+ such that k ∈ I[id, (i + 1)d]. Then
− A˜T (In + εAav )T PAρ,d − Aρ,d
T
P(In + εAav )A˜, |xk |2 ≤ X j ≤ δ2i−1 (X1 − µd ) + µd α id−d
T k−d
ω5 = −εAρ,d PA (I2 ⊗ Ad )+ ≤ δ2 d (X1δ−µd ) + µd α −d α k−d
2
A˜T (In + εAav )T PA (I2 ⊗ Ad ),
≤ (X1δ−µd ) + µd α −d α k−d ,
2
T
ω6 = −εAρ,d PA1 + A˜T (In + εAav )T PA1 ,
where the last inequality follows from δ2 < α d , which proves
(55).
ω7 = ε[A (I2 ⊗ Ad )]T P[A (I2 ⊗ Ad )].
For LMI feasibility guarantees, choose Λa = Λρ,2d = λ1 I2 ,
ω8 = ε[A (I2 ⊗ Ad )]T PA1 . Λρ,a = Λρ,d = λ1 I4 , R = λ1 In , S1 = λd In , where λ1 > λd ,
Summarizing, we arrive at: S2 = λ2 In where λ2 = 2λ2d (keep in mind that λ2d = O(ε 2 )).
For θ = 0 (so α = 1), the inequality −α 2d S2 + λ2d In < 0
Theorem 2. Consider system (30) subject to Assumption 2, hold, and also for any d there is a small enough ε such
let Ha , Hρ , Hρ,a , be defined by (38). Given positive tuning
∗
that ε 2 d 2 L T RL and δ2 < α d (since δ2 = O(ε)) are small
parameters θ , d ∗ , and ε ∗ subject to δ2 < α d and θ ε ∗ < 1. enough, while Θε,d is independent of d. Therefore, by
n×n
Let there exist 0 < P, S1 , S2 , R ∈ R , and diagonal positive choosing θ = 0 , d = 0, (so α = 1), it is enough to prove
(1) (2)
matrices Λa , Λρ , Λρ , Λρ,2d ∈ R2×2 and Λρ,a ∈ R4×4 such that Θε,d < 0. Also there is a 0 < P ∈ Rn such that β1 < 0
that (53) with notations (49), (51), (54) holds with ε = ε ∗ and for small enough ε (see (19), Assumption 2). It is easily
d = d√∗ . Then system (30) is exponentially stable with a decay seen that Ψε,d < 0 for large enough λ1 and small enough
rate 1 − θ ε for all ε ∈ (0, ε ∗ ] and 0 ≤ d ≤ d ∗ . Namely, there ε > 0. Next, we apply Schur complement with respect to
1659
Θε,d , whence Θε,d < 0 iff β1 − λ1 B(λ1−1 Ψε,d )−1 BT < 0. Note
1
that −(λ1−1 Ψε,d )−1 is bounded as λ1 → ∞ (converges to
the identity matrix), whereas B and β1 are independent of
λ1 implying the feasibility of Θε,d . Thus, for any large d,
there exist small enough ε ∗ and θ such that feasibility is
assured.
D. Numerical example
Example 2: (Stabilization by fast switching [10]) Consider
(30) with A0 = 0 and
(
A1 , k ∈ [100nε, 100(n + 0.4)ε),
AD (k) = (59)
A2 , k ∈ [100(n + 0.4)ε, 100(n + 1)ε),
Fig. 1. The trajectory of the Euclidean norm of the system state xi converges
where to the origin. Zoom plot of the graph is also provided.
0.1 0.3 −0.13 −0.16
A1 = , A2 = . (60)
0.6 −0.2 −0.33 0.03
Given k ∈ Z+ , (59) can be written as R EFERENCES
AD (k) = χ[100nε,100(n+0.4)ε) (k)A1 + [1 − χ[100nε,100(n+0.4)ε) (k)]A2 , [1] H. An, Q. Wu, H. Xia, and C. Wang. Control of a time-varying
hypersonic vehicle model subject to inlet un-start condition. Journal
where χ[100nε,100(n+0.4)ε) (k) is an indicator function. Here, of the Franklin Institute, 355(10):4164–4197, 2018.
AD (k) can be presented as (31) with [2] F. Bullo. Averaging and vibrational control of mechanical systems.
SIAM Journal on Control and Optimization, 41(2):542–562, 2002.
−0.038 0.024 [3] B. Caiazzo, E. Fridman, and X. Yang. Averaging of systems with
Ad = , fast-varying coefficients and non-small delays with application to sta-
0.042 −0.062
bilization of affine systems via time-dependent switching. Nonlinear
A1 and A2 given by (60) and Analysis: Hybrid Systems, 48:101307, 2023.
( [4] X. Cheng, Y. Tan, and I. Mareels. On robustness analysis of linear
0.6, k ∈ [100nε, 100(n + 0.4)ε), vibrational control systems. Automatica, 87:202–209, 2018.
a1 (k) = , [5] R. H. Christensen and I. Santos. Design of active controlled rotor-
−0.4, k ∈ [100(n + 0.4)ε, 100(n + 1)ε),
blade systems based on time-variant modal analysis. Journal of sound
and vibration, 280(3-5):863–882, 2005.
(
−0.6, k ∈ [100nε, 100(n + 0.4)ε), [6] E. Fridman. Introduction to time-delay systems: analysis and control.
a2 (k) = .
0.4, k ∈ [100(n + 0.4)ε, 100(n + 1)ε), Birkhauser, Systems and Control: Foundations and Applications, 2014.
[7] E. Fridman and J. Zhang. Averaging of linear systems with almost
An explicit computation of ρm (k), m = 1, 2 yields |ρm (k)| ≤ periodic coefficients: A time-delay approach. Automatica, 122:109287,
(m) 2020.
εhρ = 0.6ε, ∀k ∈ Z+ . We consider θ ∈ {0, 0.01}, ε = 0.05.
[8] C. J. Harris and M. JF. Stability of linear systems: some aspects of
Note that with some simply calculations, it is easily shown kinematic similarity. 1980.
that (9) holds for ε = 0.05. Verify the LMIs of Theorem 2 [9] R. Katz, E. Fridman, and F. Mazenc. Constructive method for
to obtain the maximal value d which preserves feasibility of averaging-based stability via a delay free transformation. To appear
in Autmatica, 2024.
the LMIs. We find the corresponding upper bounds d ∗ that [10] R. Katz, F. Mazenc, and E. Fridman. Stability by averaging via time-
guarantees the system’s exponential stability for all 0 ≤ ε ≤ varying lyapunov functions. 22nd IFAC World Congress. To appear.,
ε ∗: 2023.
[11] H. K. Khalil. Nonlinear Systems. Prentice Hall, 3rd edition, 2002.
θ = 0, d ∗ = 28; θ = 0.01, d ∗ = 5. [12] P. Li, J. Lam, R. Lu, and K.-W. Kwok. Stability and l 2 synthesis of a
class of periodic piecewise time-varying systems. IEEE Transactions
We further provide simulations of system with a fixed ε = on Automatic Control, 64(8):3378–3384, 2018.
0.05, d = 28 and an initial condition xi = [1, −0.5]T , ∀i ∈ [13] F. Mazenc, M. Malisoff, and M. S. De Queiroz. Further results on
I[−d, 0]. The results are shown in Fig. 1. We see that the strict Lyapunov functions for rapidly time-varying nonlinear systems.
Automatica, 42(10):1663–1671, 2006.
system state converges to the origin, which demonstrates the [14] S. Meerkov. Principle of vibrational control: theory and applications.
efficacy of the proposed method. IEEE Transactions on Automatic Control, 25(4):755–762, 1980.
[15] H. Sandberg and E. Möllerstedt. Periodic modelling of power systems.
IV. C ONCLUSION IFAC Proceedings Volumes, 34(12):89–94, 2001.
This paper presents a novel quantitative approach to av- [16] R. Wisniewski. Linear time-varying approach to satellite attitude
control using only electromagnetic actuation. Journal of Guidance,
eraging for stability of discrete-time systems with rapidly- Control, and Dynamics, 23(4):640–647, 2000.
varying periodic coefficients. By applying a novel system [17] X. Xie and J. Lam. Guaranteed cost control of periodic piecewise
representation, state transformation and further employing a linear time-delay systems. Automatica, 94:274–282, 2018.
[18] X. Yang, J. Zhang, and E. Fridman. Periodic averaging of discrete-time
direct Lyapunov method, explicit LMI conditions for stability systems: A time-delay approach. IEEE Transactions on Automatic
were derived. The LMIs provides upper bounds on the small Control, 2022.
parameters that preserve exponential stability of the original [19] J. Zhang and E. Fridman. L2 -gain analysis via time-delay approach to
periodic averaging with stochastic extension. Automatica, 137:110126,
system. The method was extended to linear discrete systems 2022.
with constant delay. Future work may include improvement
of the method, its extension to time-varying delays and its
control applications such as averaging-based control.
1660