3210 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 68, NO.

6, JUNE 2023

Data-Driven Analysis and Controller Design

for Discrete-Time Systems Under
Aperiodic Sampling
Stefan Wildhagen , Julian Berberich , Michael Hertneck , and Frank Allgöwer

Abstract—This article is concerned with data-driven interval (MSI), i.e., the largest time span between two sampling
analysis of discrete-time systems under aperiodic sam- instants such that stability is preserved. Knowledge or at least a
pling, and in particular with a data-driven estimation of the good approximation of the MSI is often required for the analysis
maximum sampling interval (MSI). The MSI is relevant for
and design of aperiodic sampling strategies. A multitude of dif-
the analysis of and controller design for cyber-physical,
embedded and networked systems, since it gives a limit on ferent methods for analyzing aperiodically sampled systems and
the time span between sampling instants such that stability estimating the MSI have been discussed in the literature, such
is guaranteed. We propose tools to compute the MSI for a as the time-delay approach [3]–[5], impulsive/hybrid systems
given controller and to design a controller with a preferably approach [6], [7], switched systems approach [8]–[10], or robust
large MSI, both directly from a finite-length, noise-corrupted input/output approach [11], [12].
state-input trajectory of the system. We follow two distinct All of the methods abovementioned require an accurate model
approaches for stability analysis, one taking a robust con-
to be applicable, although obtaining such a model via first
trol perspective and the other a switched systems perspec-
tive on the aperiodically sampled system. In a numerical principles can be a challenging task. Measured trajectories of
example and a subsequent discussion, we demonstrate a system, by contrast, can typically be obtained easily. This fact
the efficacy of our developed tools and compare the two has been leveraged to estimate a model based on given data in the
approaches. field of system identification [13], and more recently, to perform
system analysis and controller design directly via measured
Index Terms—Data-driven control, robust control,
sampled-data control, switched systems. data [14]. An interesting stream of research, which is closely
related to the results in this article, has been sparked by the
finding that the entire behavior of a linear time-invariant (LTI)
I. INTRODUCTION system can be described by a single trajectory [15]. A number of
HE widespread and ever-increasing prevalence of cyber- different works considered topics, such as verifying dissipativity
T physical systems (CPSs) and of embedded and networked
control systems (NCSs) has sparked a large interest in the study
properties [16], [17], model predictive control [18], [19], or
state-feedback design [20]–[22], all based on measured data,
of sampled-data control systems in the recent decades [1], [2]. as well as combining data with prior knowledge for controller
Especially, aperiodically sampled systems emerged as a tool design [23].
to abstract and analyze many real-world scenarios where the The main contribution of this work is to investigate data-
sampling time points are not guaranteed to be, or are even de- driven analysis of discrete-time systems under aperiodic sam-
signed not to be, equidistant: NCSs with packet dropouts, delays, pling. In contrast to existing model-based sampled-data control
and aperiodic scheduling strategies, or event- and self-triggered approaches, which usually handle continuous-time systems, we
approaches, to name a few examples. A fundamental concept deal with discrete-time systems, since, in any practical scenario,
for systems under aperiodic sampling is the maximum sampling data can only be measured at discrete time instants. In addition,
the discrete-time perspective is justified by the fact that CPSs,
embedded control systems, and NCSs involve digital devices,
Manuscript received 23 October 2021; revised 11 February 2022;
accepted 7 June 2022. Date of publication 17 June 2022; date of cur- which dictate a clock rate for the entire control system. In par-
rent version 29 May 2023. This work was supported by the Deutsche ticular, we develop techniques 1) to estimate the MSI for a given
Forschungsgemeinschaft (DFG, German Research Foundation) under controller and 2) to design controllers with a preferably high MSI
Grant 285825138 and under Germany’s Excellence Strategy—EXC bound, both directly from a noise-corrupted, finite-length state-
2075—390740016. The work of Julian Berberich was supported by
the International Max Planck Research School for Intelligent Systems input trajectory of the unknown system. We present two different
(IMPRS-IS). Recommended by Associate Editor P. Rapisarda. (Corre- approaches to achieve these goals: The first, referred to as robust
sponding author: Stefan Wildhagen.) input/output approach, is based on splitting the aperiodically
The authors are with the University of Stuttgart, Institute for
Systems Theory and Automatic Control, 70569 Stuttgart, Germany
sampled system into an LTI system and a delay operator, and a
(e-mail: [email protected]; [email protected] subsequent stability analysis using input/output properties of the
stuttgart.de; [email protected]; frank.allgower@ delay operator and robust control tools; the second, referred to
ist.uni-stuttgart.de). as switched systems approach, comprehends the aperiodically
Color versions of one or more figures in this article are available at
https://doi.org/10.1109/TAC.2022.3183969.
sampled system with its sampling period-dependent dynamics
Digital Object Identifier 10.1109/TAC.2022.3183969 as a switched system, and leverages the well-understood analysis

0018-9286 © 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on December 18,2023 at 13:09:51 UTC from IEEE Xplore. Restrictions apply.
WILDHAGEN et al.: DATA-DRIVEN ANALYSIS AND CONTROLLER DESIGN FOR DISCRETE-TIME SYSTEMS UNDER APERIODIC SAMPLING 3211

for this system class. As we will see later, the robust input/output yT (t) = y(t) for all t ∈ N[0,T ] and yT (t) = 0 for all t ∈ N≥T +1 .
approach has a lower computational complexity and comes with We write y 2 for the 2 norm of a signal y ∈ 2 and Δ 2 :=
little conservatism even if the available data are noisy, whereas inf{γ | Δ(y)T 2 ≤ γ yT 2 , y ∈ 2e , T ∈ N0 } for the 2
the switched systems approach can outperform the latter and gain of an operator Δ : 2e → 2e .
provide very tight estimations of the MSI especially when rich
data are available. In both approaches, we leverage a recently II. SETUP AND PROBLEM STATEMENT
proposed data-driven system parameterization [22], [23] to ver-
ify the stability conditions robustly for all systems consistent A. System Under Aperiodic Sampling
with the data. Let us consider a discrete-time LTI system
To the best of our knowledge, this is the first work to offer
a data-driven analysis of discrete-time aperiodically sampled x(t + 1) = Atr x(t) + Btr u(t), x(0) = x0 ∈ Rn (1)
systems. Our recent work [24] solves a similar problem in the with state x(t) ∈ R , input u(t) ∈ R , and time instants t ∈
n m

continuous-time domain, combining data-driven control meth- N0 . A typical application for such a discrete-time setup are
ods with the time-delay approach to sampled-data systems [4]. continuous-time processes that are sensed, controlled, and ac-
Note that sampled-data systems form a subclass of time-delay tuated by digital devices, which is commonly the case in CPSs,
systems with bounded delay. For this reason, the results on embedded control systems, and NCSs. We assume throughout
data-driven control of time-delay systems in [25] could be this article that the true system matrices Atr and Btr are unknown
used to treat the problem considered in this article as well. and that only state-input measurements are available.
As another alternative, one could identify the unknown system To close the loop, (1) is sampled and controlled at aperiodic
from the measured data, and in a second step, use the identified sampling instants tk ∈ N0 , where
model in existing model-based discrete-time stability conditions t0 = 0, tk+1 − tk ≥ 1 (2)
for aperiodically sampled systems [8]–[10], [26] in a two-step
procedure. However, we note that our proposed approaches such that the sampling interval hk := tk+1 − tk is time-varying.
are more direct, allowing to estimate the MSI and to design A sampled version of the plant state is {x(tk )}∞
k=0 , from which
a controller directly from data in a single step. Furthermore, the controller computes a sequence of control values {u(tk )}∞
k=0
obtaining tight estimation bounds from noisy data of finite length via a linear state-feedback law u(tk ) = Kx(tk ), K ∈ Rm×n .
is a challenging problem when estimating a model, e.g., via We assume that the control input applied to the plant is held
least-squares estimation [27], which might be problematic since constant in between sampling instants, i.e., u(t) = u(tk ), t ∈
we ultimately aim for guarantees for the true underlying system. N[tk ,tk+1 −1] . The aperiodically sampled system in closed loop
Nonetheless, methods based on set membership estimation [28], can then be written as
[29] are indeed guaranteed to contain the true system in the x(t + 1) = Atr x(t) + Btr Kx(tk ) ∀t ∈ N[tk ,tk+1−1] ∀k ∈ N0
error bounds. We compare both mentioned alternatives, [25] tk+1 = tk + hk ∀k ∈ N0
and the two-step procedure based on set membership estima-
tion, to our proposed approaches in a numerical example in t0 = 0, x(0) = x0 .
Section V. (3)
The rest of this article is organized as follows. In Section II, Aperiodic sampling is relevant for a multitude of practical
we introduce the system under aperiodic sampling, the available scenarios. For instance, NCSs, where the communication be-
data, and the considered problems. The robust input/output tween plant and controller takes place over a shared commu-
approach and the switched systems approach can be found in nication network, exhibit several phenomena that give rise to
Sections III and IV, respectively. A numerical analysis of the aperiodic sampling: Packet dropouts, which occur especially in
proposed approaches is performed in Section V. We summarize wireless and/or congested networks, result in varying sampling
this article and give potential directions for future work in periods [9]. Further, the reception of control updates can be
Section VI. out of order due to time-varying transmission delays, in which
We denote by N the set of natural numbers, N0 :=N ∪ {0} case it might be useful to discard the outdated ones [30], [31].
and N[a,b] := N0 ∩ [a, b], N≥a := N0 ∩ [a, ∞), a, b ∈ N0 . We Another common scenario is aperiodic or contention-based net-
denote by I the identity matrix and by 0 the zero matrix of appro- work access protocols. In addition, it is often favorable not to
priate dimension, and by In the identity matrix of dimension n. sample and transmit the system’s state periodically to counteract
Let A ∈ Rn×n be a real matrix. We write A  0 (A  0) if A is overloading of the network, as is done for instance in (periodic)
symmetric and positive (semi-)definite, and we denote negative event-triggered and self-triggered approaches [32], [33]. In all
(semi-)definiteness similarly. Let σmax (A) denote the maximum these cases, it is possible to write the resulting control system
singular value of A. We write v 2 for the 2-norm of a vector with aperiodic sampling in the form (3).
v ∈ Rn , and A 2 for the induced 2-norm of A. The Hermitian In the abovementioned scenarios, although the exact sampling
transpose of a complex matrix B ∈ C n×m is denoted by B ∗ . The instants are unknown in advance, a bound on the sampling
Kronecker product of two matrices C ∈ Rn×m and D ∈ Rp×r interval is often known or can be estimated. This is the case, e.g.,
is denoted by C ⊗ D. We write  if an element in a matrix can if the number of consecutive packet losses and the maximum
be inferred from symmetry. We denote by 2 the space of square delay are upper bounded [9], [31]. In event- and self-triggered
integrable signals and by 2e the extended 2 space. For some schemes, the interevent times can be proven to be upper bounded
T ∈ N0 , we denote by ·T : 2e → 2e the truncation operator, under certain conditions [34], [35]. Further, in practical imple-
which assigns to a signal y ∈ 2e the signal yT which satisfies mentations of such schemes, it is often sensible to constrain the

Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on December 18,2023 at 13:09:51 UTC from IEEE Xplore. Restrictions apply.
 3212 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 68, NO. 6, JUNE 2023

length of the sampling interval to guarantee a certain amount of Such bounds have also been considered for the purpose of robust
attention dedicated to the process [35]. Hence, in this article, we data-driven control in [36]. The latter is especially useful for
are concerned with stability analysis and controller design for describing the practically relevant case of pointwise disturbance
(3) in case the sampling interval is arbitrary but bounded, i.e., bounds, since it provides a tight description thereof and, other
hk takes arbitrary values in the set N[1,h] for a given h ∈ N. In than quadratic full-block bounds, guarantees that the set of
addition, we are interested in estimating a preferably tight lower matrices compatible with the data does not grow when more
bound on the largest h such that (3) is asymptotically stable, i.e., data is added. Let cd (N ) denote the number of decision variables
on the MSI. involved in the respective multiplier description. Then, imple-
Remark 1: In most classical works on aperiodically sampled mentation of diagonal multipliers involves cd (N ) = N decision
systems, the sampling interval takes values in hk ∈ (0, ∞) variables compared with cd (N ) = 1 for quadratic multipliers. A
(cf., [2]–[4], [6]–[8], [11], and [12]), as continuous-time systems thorough discussion of Assumption 2 and possible choices of P d
are considered therein. In contrast, we consider hk ∈ N in this can be found in [23, Sec. II.C].
article since we work in discrete time. As a result, the considered Furthermore, we pose the following assumption on Bd .
setup can be seen as the discrete-time equivalent to the classical Assumption 3: The matrix Bd has full column rank.
approaches formulated in continuous time. This is essentially without loss of generality: Should it not
be the case, one can define another pair d, Bd with the same
B. Available Data influence on (4), but with Bd satisfying Assumption 3 [23].
Remark 4: As an alternative to the disturbance description
The main challenge in this article is that analysis and con- introduced in (4) and Assumptions 2 and 3, one could also
troller design are performed without knowledge of the true consider a formulation involving measurement noise w (cf.,
matrices Atr and Btr . Instead, we suppose that state-input data
N −1
[20, Sec. V.A]). If a multiplier description of Atr W − W + [with
{x(t)}N t=0 , {u(t)}t=0 , N ∈ N, of the perturbed system W, W + defined similarly as in (13)] was available, the results
x(t + 1) = Atr x(t) + Btr u(t) + Bd d(t) (4) presented in this article would require only minor modifications
to accommodate such a setup.
are available, where d(t) ∈ R is an unknown disturbance and
nd
Finally, we note that the measurements are taken at each of
Bd is a known matrix. Bd can be used to incorporate knowledge
the time instants. It is not restrictive to assume that such data are
on the way the disturbance enters the system, e.g., if it is
available, since in sampled-data systems, the aperiodic sampling
known that it only affects a subset of the states. If no such prior
comes into play only for closed-loop operation. By contrast, the
knowledge is available, then one may set Bd = I.
required state and input trajectories can be collected indepen-
ˆ N −1 that affected
The particular disturbance sequence {d(t)} t=0 dently of each other in an open-loop experiment. To illustrate

the measured  and written in matrix form as
data is unknown this, consider an NCS where communication with the controller
ˆ
D̂ := d(0) ˆ − 1) . As in [23], we assume that a
· · · d(N takes place via a network. In such a setup, it is nonetheless
multiplier description for the disturbance is available. possible to probe the system with an open-loop input trajectory
Assumption 2: The disturbance satisfies D̂ ∈ D, where and to record the system response at the sensor without any need
⎧   ⎫ for using the network.
⎨  D   ⎬
 D
D := D ∈ Rnd ×N  Pd  0 ∀Pd ∈ P d
⎩ I I ⎭
C. Problem Statement
and where P d is a convex cone of symmetric matrices admitting Having introduced aperiodically sampled systems and the
an LMI representation. Moreover, there exists Pd ∈ P d such
    available data, we may now formalize the problems considered
that I 0 Pd I 0 ≺ 0. in this article.
Assumption 2 encompasses various special cases, e.g., Problem 1 (Aperiodic sampling: Analysis): Given state-input
N −1
quadratic bounds on the entire sequence measurements {x(t)}N t=0 and {u(t)}t=0 of (4), a disturbance
   description D, matrix Bd , upper bound h for the sampling
D̂ Qd Sd D̂ interval, and, furthermore, controller K, determine if the origin
 0, Qd ≺ 0
I Sd Rd I of the closed-loop aperiodically sampled system (3) is asymp-
totically stable for an arbitrarily time-varying sampling interval
as used similarly, e.g., in [17], [20]–[22], and [24], via hk ∈ N[1,h] .
 
Qd Sd  Problem 2 (Aperiodic sampling: Controller design): Given
Pd = τ τ >0 . (5) the same setup as in Problem 1, design a controller K such that
Sd Rd
the origin of the closed-loop aperiodically sampled system (3)
ˆ
Alternatively, componentwise 2-norm bounds d(t) 2 ≤ d can is asymptotically stable for an arbitrarily time-varying sampling
be considered via diagonal multipliers interval hk ∈ N[1,h] .
Pd = Problem 3 (MSI): Determine the largest h, denoted by hMSI ,
  such that Problem 1, respectively, Problem 2, admits a solu-
−diag{pi }N 0  tion. In other words, determine a possibly tight lower bound
i=1 
N 2  pi ≥ 0, i ∈ N[1,N ] . of the MSI.
0 i=1 pi d Ind We will focus on Problems 1 and 2 in the technical sections of
(6) this article. This is because once we have obtained a solution to

Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on December 18,2023 at 13:09:51 UTC from IEEE Xplore. Restrictions apply.
WILDHAGEN et al.: DATA-DRIVEN ANALYSIS AND CONTROLLER DESIGN FOR DISCRETE-TIME SYSTEMS UNDER APERIODIC SAMPLING 3213

Problems 1 and 2, a solution to Problem 3 can be found via

a linear search over h [37, Ch. 6], or, often more efficiently,
via an exponential search [38]. We note that since h takes
integer values, these algorithms are guaranteed to be successful
[37, Ch. 6], i.e., to return the tightest MSI bound.
We propose two approaches to address Problems 1 and 2.
In the first approach (see Section III), we write the aperiodi-
cally sampled system (3) as an interconnection of a classical
LTI system and a delay operator. In the second approach (see
Section IV), we comprehend (3) as a discrete-time switched
Fig. 1. Sawtooth shape of τ (t).
system.
In practice, one often deals with continuous-time processes.
In such a case, the choice of discretization period H ∈ (0, ∞), properties of the delay operator and to establish stability condi-
with which the data are recorded from the underlying physical tions. A definition of an IQC in the input/output framework,
process, is of great importance. The reason is that while it is which we will use throughout this article, is given in the
often possible to stabilize the discrete-time system (4) with a following.
large discretization period, the achieved control performance Definition 5: A bounded, causal operator Δ : p2e → q2e , y →
might be very poor. Especially in a data-driven setup, i.e., when e satisfies the hard static1 IQC defined by a multiplier Π ∈
the dynamics of the process to-be-controlled are not known, R(p+q)×(p+q) if the following inequality holds for all y ∈ p2 :
it might be hard to adequately select H such that all relevant   
dynamics are reflected in the discrete-time model and a certain T
y(t) y(t)
Π ≥ 0 ∀T ∈ N0 . (7)
performance objective is met.
t=0
(Δy)(t) (Δy)(t)
However, in practice, often expert knowledge is available,
e.g., an estimate of the fastest relevant dynamics of the system, As a short-hand notation, we write Δ ∈ IQC(Π) if Δ satisfies
from which a suitable H can be chosen. Alternatively, one could the hard static IQC (7) in the sense of Definition 5.
explicitly specify a performance objective and incorporate it
into the stability conditions. Then, one could test if the desired A. Main Idea
performance objective can be satisfied, and if not, adjust the
Note that, as an intermediate step, we can rewrite the aperi-
discretization period H accordingly.
odically sampled system (3) as a time-delay system
Although we focus on the stabilization problem (i.e., Prob-
lems 1 and 2) in the main body of this article, we will demonstrate x(t + 1) = (Atr + Btr K)x(t) + Btr K(x(tk ) − x(t))
in Remark 13 that a performance objective can be directly in- = (Atr + Btr K)x(t) + Btr K(x(t − τ (t)) − x(t))
corporated in the robust input/output approach. For the switched (8)
systems approach, incorporating a performance objective is not for t ∈ N[tk ,tk+1 −1] , k ∈ N0 , where
straightforward due to the unknown intersampling behavior, and
is left to future work. τ (t) := t − tk , t ∈ N[tk ,tk+1 −1] , k ∈ N0 (9)
represents the amount of time by which the feedback information
III. ROBUST INPUT/OUTPUT APPROACH is delayed at time t. Note that τ (0) = 0 since t0 = 0. The delay
sequence {τ (t)} has a “sawtooth shape,” as illustrated in Fig. 1,
The existing (model-based) robust input/output approach to i.e., it is reset to zero at sampling instants and is subsequently
aperiodically sampled systems, pioneered in [3], [11], and increased by one in each time step where no sampling occurs.
[39] and further refined in [12], is tailored to continuous-time The maximum possible delay is h − 1.
systems. To accommodate our discrete-time setup, which we Following [40, Sec. 2], we subsequently rewrite the time-
consider because we ultimately aim for an incorporation of data, delay system (8) as an interconnection between an LTI system
we provide input/output methods for analyzing discrete-time and a delay operator. First, we note that
systems in this section.
In particular, we take the following steps. In Section III-A, e(t) := x(t − τ (t)) − x(t)
we introduce the main idea of the robust input/output approach, can be viewed as the “error” induced by aperiodic sampling. It
which is first to write the aperiodically sampled system as a can be represented by the telescopic sum
time-delay system, which exhibits a sawtooth shape of the delay
sequence. Second, following [40], we express this time-delay e(t) = x(t − τ (t)) − x(t) = x(t − τ (t))
system as an interconnection of an LTI system and a delay − x(t − τ (t) + 1) + x(t − τ (t) + 1) − · · · − x(t)
operator. In Section III-B, we provide a novel bound on the
2 gain of this delay operator. With this, model-based stability
conditions are formulated using results from robust control 1 Hard refers to the fact that the IQC holds for all T ≥ 0, whereas soft IQCs
theory in Section III-C. Finally, in Section III-D, we present hold only for T → ∞. Static refers to the fact that the signals (y, Δy) are
the data-driven stability criteria by combining the model-based multiplied with Π directly, whereas filtered versions thereof are used in dynamic
IQCs. A static IQC is equivalent to dissipativity with a general quadratic supply
conditions with the data-driven system parameterization in [23]. rate (cf., [44]–[46]). Nonetheless, we speak of IQCs in this article to be consistent
We use the well-established concept of integral quadratic con- with the literature on time-delay and aperiodically sampled systems [2], [12],
straints (IQCs) (see [40]–[44]) in order to describe input/output [40].

Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on December 18,2023 at 13:09:51 UTC from IEEE Xplore. Restrictions apply.
 3214 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 68, NO. 6, JUNE 2023


t−1 for all h [11]. In contrast, the factor of improvement is dependent
= x(i) − x(i + 1). on h in the discrete-time case.
i=t−τ (t) The 2 gain proven in Lemma 6 directly implies that Δ
If we introduce an artificial output satisfies a hard static IQC.
Corollary 9: For any X = X   0, it holds that Δ ∈
y(t) := x(t) − x(t + 1)
IQC(Π2 ), where
= (I − Atr − Btr K)x(t) − Btr Ke(t) 
h
(h − 1)X 0
and define the delay operator Δ : n2e → n2e , e = Δy, as Π2 := 2 .
0 −X

t−1
e(t) = (Δy)(t) := y(i), t ∈ N[tk ,tk+1 −1] , k ∈ N0 Proof: A positive definite X allows a factorization X =
1 1 1 1
i=t−τ (t) X 2 X 2 where X 2  0. Further, it holds that Δ(X 2 y) =
(10) 1 1 1
X 2 Δ(y) and that X 2 y = 0 if y = 0 since X 2  0. With this
we may write the aperiodically sampled system (3) as an inter- and Lemma 6, the hard static IQC follows immediately. 
connection of an LTI system and the delay operator
  
x(t + 1) Atr + Btr K Btr K x(t) C. Model-Based Stability Criteria
= (11a)
y(t) I − Atr − Btr K −Btr K e(t) In this section, we assume, in contrast to the rest of this article,
that the true system matrices Atr and Btr are known. To formulate
e(t) = (Δy)(t). (11b) stability conditions for (11), we exploit the hard static IQC for
The main idea is now, as in [40], to analyze stability of the the delay operator in Corollary 9, which allows us to directly use
feedback interconnection (11) using robust control theory. In existing results on linear systems in feedback with uncertainties
particular, the delay operator will be embedded into a class of satisfying hard static IQCs.
uncertainties acting on the LTI system (11a) by bounding its 2 Theorem 10: Suppose there exist matrices Q = Q  0 ∈
gain and thereby describing its input/output behavior as a hard R n×n
and X = X   0 ∈ Rn×n such that
static IQC. ⎡ ⎤
 Q 0 00
Atr +Btr K 
Btr K ⎢ 0 −Q 00 ⎥
B. 2 Gain of the Delay Operator
I 0 ⎢ ⎥
I−Atr −Btr K −Btr K ⎣ 0 0 h (h − 1)X 0 ⎦
0 I 2
Subsequently, we provide an estimate of the 2 gain of the 00 0 −X
delay operator Δ and express this condition in terms of a hard ⎡ ⎤
static IQC. In contrast to [40], which analyzed time-delay sys- Atr + Btr K Btr K
tems with an arbitrary bounded delay, in the case of aperiodically ⎢ I 0 ⎥
×⎣ ≺0
I − Atr − Btr K −Btr K ⎦
(12)
sampled systems, the delay {τ (t)} exhibits a sawtooth shape,
as illustrated in Fig. 1. By exploiting this information, we will 0 I
be able to derive a smaller 2 -gain estimate than if we would is satisfied. Then, the origin of (11) is asymptotically stable.
dissolve {τ (t)} into the class of bounded delays τ (t) ∈ N[0,h−1] Proof: The statement of the theorem will be a direct con-
and use the results in [40]. This is the main technical contribution sequence of [44, Corollary 11], which is a modification of the
in terms of the model-based analysis of aperiodically sampled classical circle criterion [47]. The proof of [44, Corollary 11]
discrete-time systems, as presented in Sections III-A–III-C. It relies on recognizing that (12) implies that (11a) is dissipative
is important to note in this respect that a tighter uncertainty de- w.r.t. Π2 and that (11) forms a so-called neutral intercon-
scription of Δ will lead to less conservative stability conditions nection, from which stability follows. Thus, we now verify
for the interconnection (11). the assumptions of [44, Corollary 11]. First, we note that the
Lemma 6:The 2 gain of the delay operator Δ is upper delay operator satisfies the hard static IQC Δ ∈ IQC(Π2 ), as
established in Corollary 9. Second, the delay operator satisfies
2 (h − 1), i.e., for all y ∈ n2 , we have
h
bounded by
[44, Condition (8)], as [y0] Π2 [y0] ≥ 0 holds due to X  0.

T
h T
Finally, we note that the stability conditions for continuous-

(Δy)(t) (Δy)(t) ≤ (h − 1) y(t) y(t) ∀T ∈ N0 . time systems in [44, Corollary 11] can be used for discrete-
t=0
2 t=0 time systems as well under the modification discussed in
The proof can be found in the Appendix. [44, p. 12, ramifications]. 
Remark 7: Taking the bounded-delay perspective as in [40],
one may find the bound h − 1 on the 2 gain
 of Δ. Lemma 6 D. Data-Driven Stability Criteria for Analysis and
provides a tighter bound, since it holds that h2 (h − 1) ≤ h − Controller Design
1 for all h ∈ N. In particular, equality holds for h ∈ {1, 2}, Finally, in this Section, we tackle Problems 1 and 2. Recall
whereas the strict inequality comes into play for h ∈ N≥3 . As that state-input measurements {x(t)}N N −1
t=0 and {u(t)}t=0 of the
h → ∞, the improvement approaches a factor of √12 . disturbed system (4) are available. Note that the noise-corrupted
Remark 8: In the continuous-time case, the factor of improve- data could be explained by a multitude of different matrices
ment, when taking into account the sawtooth shape compared A and B, especially since the particular disturbance sequence
with taking the bounded delay perspective, was proven to be π2 that corrupted the experiment is unknown as well. Let us arrange

Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on December 18,2023 at 13:09:51 UTC from IEEE Xplore. Restrictions apply.
WILDHAGEN et al.: DATA-DRIVEN ANALYSIS AND CONTROLLER DESIGN FOR DISCRETE-TIME SYSTEMS UNDER APERIODIC SAMPLING 3215

the measured data as follows: and PAB ∈ P AB such that

  ⎡   ⎤ ⎡ ⎤
X + = x(1) x(2) . . . x(N ) 0 I I K S 0 00 00
⎢ ⎥ ⎢ 0 −S 00 00 ⎥
  ⎢I 0 0 ⎥ ⎢ ⎥
⎢  ⎥ ⎢ ⎥
X = x(0) x(1) . . . x(N − 1) ⎢ ⎥ 00 X 0 00 ⎥
⎢ 0 0 0 K ⎥ ⎢ inv
⎢ ⎥ ⎢ 00 0 − 2
X 00 ⎥
  ⎢0 I ⎥ ⎢ h(h−1) inv ⎥
U = u(0) u(1) . . . u(N − 1) . ⎢ 0 ⎥ ⎢ ⎥
(13) ⎣0 0 I ⎦ ⎣ 00 00 ⎦
PAB
I −I 0 00 00
With this, we may define the set of all matrices compatible with ⎡  ⎤
the measured data and the disturbance bound as 0 I I K
   ⎢ ⎥
⎢I 0 0 ⎥
ΣAB = A B | X + = AX + BU + Bd D, D ∈ D . ⎢  ⎥
⎢ ⎥
×⎢⎢0 0 0 K ⎥ ⎥≺0 (16)
In order to guarantee stability for (11), we need to verify the ⎢0 I 0 ⎥
⎢ ⎥
model-based stability condition for each A B ∈ ΣAB . To ⎣0 0 I ⎦
this end, we use a data-driven parameterization of the compatible I −I 0
matrices, as developed in [23], where it was shown that it is satisfied. Then, the origin of (15) is asymptotically stable for
is possible to represent the matrices contained in ΣAB by a any [A B] ∈ ΣAB .
quadratic constraint. We define Proof: We use the full-block S-procedure [48, Lemma A.1]
⎡ ⎤ ⎡ ⎤ to conclude that (16) implies that
−X 0 −X 0 ⎡ ⎤
P AB := ⎣ −U 0 ⎦ P d ⎣ −U 0 ⎦ . (A + BK) (BK)
⎢ 0 ⎥
X + Bd X + Bd ⎢ I ⎥
⎣ (I − A − BK) −(BK) ⎦
Then, it follows directly from [23, Lemma 2] and the discussion 0 I
thereafter that the set ΣAB can be expressed in terms of a ⎡ ⎤
S 0 00
quadratic constraint as follows: ⎢ 0 −S ⎥
00
⎧ ⎡ ⎤ ⎡ ⎤ ⎫ ×⎢ ⎣ 0 0 Xinv 0
⎥ ≺ 0
⎦ (17)
⎪ A
⎨

A ⎪
⎬ 2
⎢ ⎥ ⎢ ⎥ 0 0 0 − h(h−1) Xinv
ΣAB = A B  ⎣B  ⎦ PAB ⎣B  ⎦ 0 ∀PAB ∈ P AB .    
⎪
⎩ ⎪
⎭
I I holds for all A B , which satisfy (14), i.e., for all A B ∈
(14) ΣAB . It is easy to verify that (17) is equivalent to
⎡ ⎤ ⎡ ⎤
The idea is now to rewrite the interconnection of the LTI sys- (A + BK) (BK) 0 −I 0 0
⎢ 0 ⎥
tem and the delay operator as a linear fractional transformation ⎢ I ⎥ ⎢I 0 0 0 ⎥
(LFT) with two uncertainty channels ⎣ (I − A − BK) −(BK) ⎦ ⎣ 0 0 0 −I ⎦
0 I 00 I 0
⎡ ⎤
⎡ 0
⎤ 0 I ⎡ ⎤ ⎡ ⎤
x(t + 1) S 0 00
⎢ I
⎣ y(t) ⎦ = ⎢ 0 −I ⎥ x(t)
⎥ ⎣ e(t) ⎦ ⎢ 0 −S 00 ⎥
⎣ I ⎦ (15a)
z(t)
0
0 w(t) ×⎢ ⎣ 0 0 h(h−1)2 ⎥
Xinv 0 ⎦  ≺ 0
K K 00 0 −Xinv
e(t) = (Δy)(t) (15b) which, in turn, is equivalent to
   ⎡ ⎤
w(t) = A B z (t) (15c) I 0
⎢ −(A + BK) −(BK) ⎥
  ⎢ ⎥
⎣ 0 I ⎦
where Δ ∈ IQC(Π2 ) and A B ∈ ΣAB . The first channel  
(A + BK − I) (BK)
e → y contains the delay operator as in the model-based case
⎡ ⎤
[cf., (11)], and the second channel z → w represents the uncer- S 0 00
tainty in the system matrices due to the data-driven setup and the ⎢ 0 −S 00 ⎥
×⎢ ⎣ 00
⎥  ≺ 0. (18)
disturbance. Using this system description, we are finally able 2
X
h(h−1) inv
0 ⎦
to provide a data-driven stability condition for (15), using the 00
parameterization (14) of the compatible system matrices and the 0 −Xinv
S-procedure. We are now ready to apply the dualization lemma
Theorem 11: Suppose Assumptions 2 and 3 are satisfied [45, Lemma 4.9] to (18). The required inertia assumptions
and h ≥ 2. Furthermore suppose, given a controller K, there clearly hold since the inner matrix in (18) possesses exactly 2n
exist matrices S = S   0 ∈ Rn×n , Xinv = Xinv 
 0 ∈ Rn×n positive and negative eigenvalues. Using the dualization lemma,

Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on December 18,2023 at 13:09:51 UTC from IEEE Xplore. Restrictions apply.
 3216 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 68, NO. 6, JUNE 2023

(18) directly implies that Proof: We apply the Schur complement to (20), with respect
⎡ ⎤ to its fourth diagonal component and substitute F = KS to
A + BK BK
obtain (16) with Xinv = S. This can be easily seen by calculating
⎢ I 0 ⎥
⎣ I − A − BK −BK ⎦ Π̃ for this particular choice and subsequently expanding (16).
Stability then follows from Theorem 11. 
0 I
Naturally, since the true system matrices Atr and Btr are con-
⎡ −1 ⎤ tained in ΣAB , the aperiodically
S 0 00  sampled
 system (3) is stable
⎢ 0 −S −1
00 ⎥ if (15) is robustly stable for all A B ∈ ΣAB . For complete-
×⎢⎣ −1
⎥ ≺ 0
⎦ (19)
00 h
2 (h − 1)X inv 0 ness, this statement is formalized in the following result.
00 0 −1
−Xinv Corollary 15: Suppose the conditions of Theorem 11 or of
  Corollary 14 are fulfilled. Then, the controller K renders the
holds for any A B ∈ ΣAB . With Theorem 10 and taking origin of the aperiodically sampled system (3) asymptotically
Q := S −1 and X := Xinv −1
in (12), we finally conclude asymp- stable for all hk ∈ N[1,h] .
totic stability for any A B ∈ ΣAB .  Theorem 11 may be used to address Problem 1, since it allows
to check an MSI bound for a given controller. Corollary 14,
Remark 12: Theorem 11 requires h ≥ 2 to ensure that the where Xinv is coupled to S, enables a cosearch for a controller
inverse of Π2 exists. Note that if h = 1, the aperiodically and thereby solves Problem 2.
sampled system (3) would in fact degenerate to a periodically
sampled one, in which case the results in [23] or Theorem 21 IV. SWITCHED SYSTEMS APPROACH
(see Section IV-C) can be used to analyze stability.
Remark 13: A performance objective can be incorporated In this section, we interpret the aperiodically sampled sys-
into the robust input/output approach, e.g., by adding the per- tem, whose dynamics depend on the arbitrarily time-varying
formance channel d → p to the LFT (15), where sampling interval, as a switched system. We elaborate on the
main idea of this approach before stating our contributions.
p(t) = Cp x(t) + Dp u(t).
An H2 -performance bound of γ > 0 for the channel d → p could A. Main Idea
then be enforced (cf., [22, Th. 17] and [23, Th. 1]) if (16) holds
with −S in the second diagonal element of the middle matrix Let us define for an h ∈ N
replaced by Bd Bd − S, and if there additionally exists Γ such 
h−1

that trace(Γ) < γ 2 and Ahtr := (Atr )h and Btrh := (Atr )i Btr .
 i=0
Γ (Cp + Dp K)S Then, it is immediate that the state of the aperiodically sampled
 0.
 S system (3) at sampling instants is governed by the discrete-time
switched system
An H∞ performance objective could be incorporated analo-
gously (cf., [22, Th. 20] and [23, Th. 2]). x(tk+1 ) = (Ahtr k + Btrhk K)x(tk ) (21)
Condition (16) in Theorem 11 allows for a simultaneous where the switching sequence is given by hk ∈ N[1,h] . Due to
search for S, Xinv , and PAB via a semidefinite program (SDP), linearity of the aperiodically sampled system (3), the state at
since it is an LMI in these variables. However, (16) is not intersampling instants is bounded in the sense of [49, Def. 2],
amenable to controller design yet. For a particular choice of such that asymptotic stability of (21) implies asymptotic stability
Xinv , namely Xinv = S, a simultaneous search for all involved of (3) and vice versa [49, Th. 2]. Note that since Atr and Btr are
variables and a stabilizing controller K is indeed possible via an unknown, the system matrices appearing in the switched system
SDP using the standard transformation F = KS, as we present Ahtr , Btrh , and h ∈ N[1,h] are naturally unknown as well.
in the following result.
Model-based stability analysis of switched systems is a well-
Corollary 14: Suppose Assumptions 2 and 3 are satisfied and
 understood topic [50]–[52], and the framework has been suc-
h ≥ 2. Furthermore, suppose there exist  matrices  S=S 
QAB SAB
cessfully applied to aperiodically sampled systems2 [8]–[10].
0∈R n×n
, F ∈R m×n
, and PAB = S  RAB ∈ P AB such In particular, it was shown in [9] and [51] that stability holds
AB
that, (20) shown at the bottom of this page, is satisfied. Then, the
−1
controller K :=  FS  renders the origin of (15) asymptotically 2 To be precise, Xiong and Lam [9] and Yu et al. [10] considered an NCS with
stable for any A B ∈ ΣAB . bounded packet loss, which admits an equivalent mathematical description.

⎡ ⎤
RAB − S   
⎢ h(h−1)−2 ⎥
⎢ −RAB S + RAB   ⎥
⎢ 
h(h−1)
 ⎥
⎢ ⎥
⎢ S S F ⎥≺0 (20)
⎢ SAB − SAB + QAB  ⎥
⎢ F F 0 ⎥
⎣   ⎦
0 0 0 F − 21 S

Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on December 18,2023 at 13:09:51 UTC from IEEE Xplore. Restrictions apply.
WILDHAGEN et al.: DATA-DRIVEN ANALYSIS AND CONTROLLER DESIGN FOR DISCRETE-TIME SYSTEMS UNDER APERIODIC SAMPLING 3217

if there exist matrices S h = S h  0 ∈ Rn×n and G h ∈ Rn×n , With the assumption that the measured data stem from the
h ∈ N[1,h] , which satisfy perturbed linear system (4), it is immediate that the matrices
 X+h
, X h , and U h contain state-input measurements of the lifted
G h + G h − S h  2 perturbed system
 0, (h, j) ∈ N[1,h] . (22) ⎡ ⎤
(Ahtr + Btrh K)G h S j u(t)
However, it is not straightforward to translate (22) to the data- ⎢ .. ⎥
x(t + h) = Ahtr x(t) + B htr ⎣ . ⎦
driven setup. This is mainly owed to the fact that the matrices
Ahtr and Btrh , which correspond to the different possible sampling u(t + h − 1)
intervals, appear in the stability conditions. From the given ⎡ ⎤
N −1 d(t)
trajectory {x(t)}N t=0 and {u(t)}t=0 , it is not directly possible  
to obtain data-driven parameterizations of B h , since this would ⎢ .. ⎥
+ Ah−1
tr Bd · · · Bd ⎣ . ⎦ (25)
require that the input in the recorded data is held constant over
h consecutive time instants. A possibility to obtain the sought d(t + h − 1)
parameterizations would be to conduct multiple experiments h ∈ N[1,h] . Note that the system matrices of the lifted per-
with all the sampling periods of interest. However, this could turbed system (25) are the same as in the lifted switched sys-
potentially be very laborious and costly. tem (24). We define the lifted disturbance for h = 1 as D1 =
In Section IV-B, we show that by reexpressing (21) with [d1 (0)· · ·d1 (N − 1)] := [d(0)· · ·d(N − 1)] and for h ∈ N[2,h]
a lifted input matrix, it is in fact possible to obtain parame- as  
terizations of the required matrices for all sampling periods Dh = dh (0) ··· dh (N − h)
h ∈ N[1,h] . Our solution uses the given data {x(t)}N t=0 and
{u(t)}N −1
only and requires a slightly modified assumption ⎡  ⎤ ⎡ ⎤
t=0 tr Bd )
(Ah−1 d(0) · · · d(N − h)
on the disturbance. In Section IV-C, we provide data-driven ⎢ .. ⎥ ⎢ .. .. ⎥
conditions for stability using the derived parameterizations. := ⎣ . ⎦ ⎣ . . ⎦. (26)
Bd d(h − 1) · · · d(N − 1)
B. Data-Driven Parameterization of the Lifted Systems We can now define the set of all lifted matrices consistent with
Note that from the definition of Btrh , it follows that we can the measured data and the lifted disturbance (26) as
 
rewrite its product with K as:
⎡ ⎤ ΣhAB := Ah B h ∈ Rn×(n+hm) |
K
 
⎢ .. ⎥ 
Btr K = Atr Btr · · · Btr ⎣ . ⎦ .
h h−1
h
X+ =A X +B U +
h h h h
Bdh Dh , Dh ∈D h
K
Hence, defining the lifted input matrix and the lifted controller for all h ∈ N[1,h] , where Bd1 := Bd , D1 := D, Bdh := I, and
⎡ ⎤⎫    
K ⎪ ⎬ Dh := Dh according to (26)  d(0) · · · d(N − 1) ∈ D
 
⎢ .⎥
B tr := Ah−1
h
tr Btr · · · Btr , K := ⎣ .. ⎦ h times (23)
h
⎪
⎭ for all h ∈ N[2,h] . Clearly, the particular lifted disturbance se-
K quence that affected the measurements also lies in Dh .
the switched system (21) can be rewritten as the lifted switched In order to obtain a data-driven parameterization of ΣhAB as
system in [23], we require a description of the set Dh via a quadratic
 
constraint. Denoting n1d := nd , nhd = n for h ∈ N[2,h] , and
x(tk+1 ) = Ahtr k + B htr k K hk x(tk ). (24)
Nh := N − h + 1, Dh would need to satisfy the representation
Recall that we have access to one measured state-input tra- ⎧   ⎫
N −1 ⎨   ⎬
jectory {x(t)}Nt=0 and {u(t)}t=0 of (4). The recorded data  D h
D h
Dh ∈ Rnd ×Nh 
h

were sampled at each time instant and were affected by the Pdh  0 ∀Pdh ∈ P hd
⎩ I I ⎭
unknown disturbance sequence {d(t)} ˆ N −1 . We will show in this
t=0
section that using this data, it is possibleto obtain data-driven (27)
parameterizations of the lifted matrices Ah B h appearing where P hd is a known convex cone of symmetric matrices ad-
in (24). Let us define, for an arbitrary h ∈ N[1,h] , the matrices mitting an LMI representation, for which there exists Pdh ∈ P hd
   
containing the measured data (see also [53]) as that satisfies I 0 Pdh I 0 ≺ 0. For h ≥ 2, it is generally
 
X+h
:= x(h) x(h + 1) · · · x(N ) difficult to derive a tight characterization as in (27), even if Atr
  was known.
X h := x(0) x(1) · · · x(N − h) Therefore, in the following, we will derive an overapprox-
imation of the set Dh by a quadratic constraint for all h ∈
⎡ ⎤ N[1,h] . We assume the following special case for the disturbance
u(0) u(1) ··· u(N − h)
⎢ .. .. .. ⎥ bound D.
U := ⎣
h
. . . ⎦. Assumption 16:  The components of the  unknown disturbance
u(h − 1) u(h) ··· u(N − 1) sequence D̂ = d(0) ˆ ˆ
· · · d(N − 1) are norm bounded in

Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on December 18,2023 at 13:09:51 UTC from IEEE Xplore. Restrictions apply.
 3218 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 68, NO. 6, JUNE 2023

the 2-norm with a known upper bound d ≥ 0, i.e.,

d(t) 2 ≤ d ∀t ∈ N[0,N −1] .
For such disturbance sequences, we derive the preferably tight
overapproximation
⎧  ⎫
⎨    ⎬
h h
 D̃ D̃ h
D̃h := D̃h ∈ Rnd ×Nh 
h
P̃dh  0 ∀P̃dh ∈ P̃ d
⎩ I I ⎭
(28)
h
which corresponds to determining a class of multipliers P̃ d with
h
   
P̃dh ∈ P̃ d such that I 0 P̃dh I 0 ≺ 0, for which Dh ⊆
D̃h is satisfied for all h ∈ N[1,h] . Then, if we define
 
Σ̃hAB := Ah B h ∈ Rn×(n+hm) |

h
X+ = Ah X h + B h U h + Bdh D̃h , D̃h ∈ D̃h

it is clear that ΣhAB ⊆ Σ̃hAB for all h ∈ N[1,h] , and that Σ̃hAB can
be expressed via a quadratic constraint, as discussed in [23]. In
h
the following, we will elaborate how to construct P̃ d .
Remark 17: Assumption 16 directly implies that (5) with
2
Qd = −I, Sd = 0 and Rd = d N I or (6) are valid classes of
multipliers for the overapproximation D̃1 (cf., [23, Sec. II.C]).
First, note that with Assumption 16 and (26), the 2-norm of
the components of Dh can be upper bounded by

h−1
dh (t) 2 ≤ Aitr 2 Bd 2 d(t + i) 2
i=0 Then, if we rewrite (30) as
⎡ ⎤ ⎡ ⎤⎡ ⎤

h−1
Ah −I 0 0 Ah
≤ σmax (Aitr )σmax (Bd )d (29) ⎢ h ⎥ ⎢ ⎥⎢ ⎥
i=0
⎣B ⎦ ⎣ 0 0 0 ⎦ ⎣B h ⎦  0 (33)
t ∈ {0, . . . , N − h}, h ∈ N[1,h] . Note that σmax (Aitr ) is not I 0 0 σ 2h I I
known since Atr is unknown. Therefore, we use in the following we may use the S-procedure [48, Lemma A.1] to derive the
the measured data to derive an overapproximation of σmax (Aitr ). condition
To this end, consider that if it was known that ⎡ ⎤
−I 0 0
2
tr  σ h I
Ahtr Ah (30) ⎢ ⎥
⎣0 0 0 ⎦ − P̃AB
h
0 (34)
for some σ h ≥ 0, it would follow directly that σmax (Atr ) ≤ σ h .
h
0 0 σh I 2
We can verify a bound, such as (30), only in case the same
h
bound holds for all Ah that are consistent with the data. As- which, if it holds for some P̃ABh
∈ P̃ AB , implies
h
sume, for now, that a class of multipliers P̃ d such that Dh ⊆  that the

quadratic matrix inequality (QMI) (33) holds for all Ah B h ,
D̃h was known for a given h. It is straightforward to show
(cf., [23, Lemma 2]) that then, if we define which satisfy (32). Since ΣhAB ⊆ Σ̃hAB , this provides us with
⎡ ⎤ ⎡ ⎤ a tractable condition to estimate the maximum singular value
−X h 0 −X h 0 directly from noisy data.
P̃ AB := ⎣ −U h 0 ⎦ P̃ d ⎣ −U h 0 ⎦
h h
(31) In view of (29), we can see that the componentwise bound on
h
X+ Bdh X+h
Bdh Dh depends only on the singular values σmax (Ai ) from i = 1
the set Σ̃hAB can be expressed by the quadratic constraint to h − 1. Therefore, it is possible to construct a bound on the
⎡ ⎤ ⎡ ⎤ lifted disturbance recursively: For a given h, first the quadratic
  A
h
Ah disturbance bound D̃h can be used to estimate the maximum
⎢ ⎥ h ⎢ h ⎥ singular value σ h of Ah via (34). From this, a componentwise
Σ̃hAB = Ah B h  ⎣B h ⎦ P̃AB ⎣B ⎦ 0
I I disturbance bound for Dh+1 can be constructed via (29), which
 in turn implies that (5) or (6) are valid classes of multipliers for
h
the overapproximation D̃h+1 (cf., Remark 17). After this, h is
∀P̃AB
h
∈ P̃ AB . (32) increased by one and the procedure is repeated. In Algorithm 1,
this recursive scheme is summarized and formalized.

Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on December 18,2023 at 13:09:51 UTC from IEEE Xplore. Restrictions apply.
WILDHAGEN et al.: DATA-DRIVEN ANALYSIS AND CONTROLLER DESIGN FOR DISCRETE-TIME SYSTEMS UNDER APERIODIC SAMPLING 3219

Algorithm 1 takes as inputs the data, data length N , dis- C. Data-Driven Stability Criteria for Analysis and
turbance matrix Bd , and componentwise bound on the distur- Controller Design
bance d from Assumption 16. It returns as outputs the class
h Having obtained parameterizations of the lifted system matri-
of multipliers P̃ d for the overapproximations D̃h of the lifted ces, we may now translate the model-based stability
h  conditions

disturbance bound Dh , and the class of multipliers P̃ AB pa- to the data-driven setup. Since the true matrices Ahtr B htr are
rameterizing the set Σ̃hAB . After initialization, the algorithm unknown,
 we must verify stability of (24) for all “uncertainties”
iterates from h = 1 to h to estimate the respective singular values
h A h
B ∈ ΣhAB . For controller design, we also need to make
h
and disturbance bounds. In Line 7, the class of multipliers P̃ d
is defined, which follows immediately from a componentwise sure that the controller gain matrix K h follows the stacked
disturbance bound dh (t) 2 ≤ dh (cf., Remark 17). Since we structure (23), since only then, the lifted switched system (24) is
are interested in possibly small singular values and disturbance equivalent to the aperiodically sampled system (3). Therefore,
bounds, an SDP, which minimizes σ 2h subject to (34), is solved we want to achieve stability of the origin of the uncertain lifted
in Line 8. Then, if the SDP admits a solution, an estimate for switched system
! "
the maximum singular value h
of A is determined in Lines 11 x(tk+1 ) = Ahk + B hk K hk x(tk )
2
and 13 as the minimum of σ h , which comes from the SDP,
hk ∈ N[1,h]
and of σ 1 σ h−1 . The latter is a valid upper bound for σmax (Ah )
 
simply due to the fact that σmax (CD) ≤ σmax (C)σmax (D) for
Ah B h ∈ ΣhAB , h ∈ N[1,h] . (35)
some matrices C and D. If the SDP does not admit a solution,
the singular value estimate is simply set to σ 1 σ h−1 in Line A sufficient condition for robust stability of (35) is the ex-
15. Finally, in Line 16, the componentwise disturbance bound istence of matrices S h = S h  0 ∈ Rn×n and G h ∈ Rn×n ,
for Dh+1 is computed based on the singular value estimates h ∈ N[1,h] , satisfying
and (29). 
Remark 18: We remark that for an h ∈ N[1,h] , all members G h + G h − S h  2
   0, (h, j) ∈ N[1,h] (36)
h−1 ni h−1 (Ah + B h K h )G h S j
of the set i=1 i σ | i=1 in i = h are also valid estimates  
for σmax (Ahtr ). Therefore, one might modify Lines 13 and 15 of for all Ah B h ∈ ΣhAB , h ∈ N[1,h] (cf., [51, Th. 2]).
Algorithm 1 to minimize over all members of this set or arbi- As a step toward verifying stability from data, in the following
trary subsets thereof. While this can result in tighter estimates result we derive an equivalent
 characterization
 of Condition (36)
of the singular values and, hence, of the disturbance bounds, as a QMI in the variable A h
B .h
the corresponding scheme has combinatorial complexity
in h̄. Lemma 20: Condition (36) holds if and only if the conditions
In the following result, we state a simple condition under G h + G h − S h  0, h ∈ N[1,h] , and
which Algorithm 1 indeed returns valid disturbance bounds. A ⎡ ⎤ ⎡ ⎤
Ah Ah
proof is omitted, since the statement follows immediately from ⎢ h ⎥ hj ⎢ ⎥ 2
the preceding discussions. ⎣B ⎦ M ⎣B h ⎦  0, (h, j) ∈ N[1,h] (37)
Lemma 19: Suppose Assumptions 3 and 16 are satisfied and I I
1
there exist σ 21 , P̃AB
1
∈ P̃ AB such that (34) is fulfilled for h = hold, where M hj is defined in (38) shown at the bottom of this
1. Then, it holds that Dh ⊆ D̃h and that ΣhAB ⊆ Σ̃hAB for all page.
h h Proof: We apply the Schur complement to (36) with respect
h ∈ N[1,h] , where P̃ d and P̃ AB , h ∈ N[1,h] , are the outputs of
Algorithm 1. to the first diagonal element to obtain the equivalent conditions
We summarize this section: we were able to derive a pa- G h + G h − S h  0, h ∈ N[1,h] and

rameterization of the system matrices A h
B h
involved S j − (Ah + B h K h )G h (G h + G h − S h )−1   0
2
in the lifted switched system (24) for all h ∈ N[1,h] , us- (h, j) ∈ N[1,h] . We pull the terms Ah and B h and I out of the
ing a single state-input trajectory sampled at each time in- latter condition to arrive at (37). 
stant. This was achieved by lifting the input, circumventing By virtue of Lemma 20, we have now represented the sta-
the need for several experiments with each of the sampling bility
 condition
 (36) as the QMI (37), which must hold for all
periods of interest. To this end, as an independent contri- A h
B ∈ ΣAB in order to conclude robust stability. In order
h h
bution, we derived a procedure to estimate the maximum  
singular value of (monomials of) Atr using only measured to verify (36) for all matrices Ah B h ∈ ΣhAB , we now lever-
data.  
age that ΣhAB ⊆ Σ̃hAB and verify (36) for all Ah B h ∈ Σ̃hAB

⎡ ⎤
# hj hj $ −G h (G h + G h − S h )−1 G h  
M11 M12
M hj = hj  hj := ⎣ −K h G h (G h + G h − S h )−1 G h −K h G h (G h + G h − S h )−1 G h K h  ⎦ (38)
(M12 ) M22
0 0 Sj

Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on December 18,2023 at 13:09:51 UTC from IEEE Xplore. Restrictions apply.
 3220 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 68, NO. 6, JUNE 2023

using the S-procedure [48, Lemma A.1]. The following result Conditions (39) and (40), shown at the bottom of this page,
constitutes the main result of this section and allows to verify in Theorem 21hjallow for a simultaneous search for matrices
whether the uncertain system (35) is robustly stable for a given S h , G h , and P̃ AB , but not for a stabilizing controller K. This
h. is due to the fact that we look for a single controller gain
Theorem 21: Suppose Assumptions 3 and 16 are satisfied, K satisfying (40) for each h, such that standard convexifying
1
there exist σ 21 , P̃AB
1
∈ P̃ AB such that (34) is satisfied for h = 1, variable transformations are not applicable. For the model-based
h scenario, an alternative was presented in [9]: It was demonstrated
and P̃ AB , h ∈ N[1,h] , are the outputs of Algorithm 1. Further-
that fixing the matrices G 1 = · · · = G h =: G indeed enables the
more, suppose, given a controller K, there exist matrices S h = design of a nonswitched controller, although this might come
hj h
S h  0 ∈ Rn×n , G h ∈ Rn×n , h ∈ N[1,h] , and P̃AB ∈ P̃ AB , with an increase of conservatism. We follow a corresponding
2
(h, j) ∈ N[1,h] , such that approach to allow for controller design in the data-driven case
as well.
G h + G h − S h  0, h ∈ N[1,h] (39) Corollary 24: Suppose Assumptions 3 and 16 are satisfied,
and (40) are satisfied. Then, the origin of the uncertain there exist σ 21 , P̃AB
1 1
∈ P̃AB such that (34) is satisfied for h = 1,
 switched

system (35) is asymptotically stable for any Ah B h ∈ and P̃AB , h ∈ N[1,h] , are the outputs of Algorithm 1. Fur-
h

thermore, suppose there exist matrices S h = S h  0 ∈ Rn×n ,

ΣhAB , h ∈ N[1,h] . hj
h ∈ N[1,h] , G ∈ Rn×n , F ∈ Rm×n and P̃AB ∈ P̃AB
h
, (h, j) ∈
Proof: We apply the Schur complement to (40) with respect
2
to its first diagonal element, which yields that (40) is equivalent N[1,h] , such that
to G + G − Sh  0 ∀h ∈ N[1,h] (41)
hj 2
M hj − P̃AB  0, (h, j) ∈ N[1,h] .
and (42) are satisfied. Then, the controller K h := [K  · · ·K  ]
2
Next, we use [48, Lemma A.1] for each (h, j) ∈ N[1,h] to with K := FG −1 renders the origin of the uncertain  switched 
conclude
 that
 this condition implies that (37) holds
 for all system (35) asymptotically stable for any Ah B h ∈
Ah B h which satisfy (32), i.e., for all Ah B h ∈ Σ̃hAB . ΣhAB , h ∈ N[1,h] .
 
Since ΣhAB ⊆ Σ̃hAB , (37) also holds for all Ah B h ∈ ΣhAB . Proof: Since G + G  − S h  0 and S h  0, G is of full rank
By such that the transformation K = FG −1 is uniquely defined. We
 Lemma 20, the fact that (37) and (39) hold for substitute F = KG in (42), shown at the bottom of this page, to
all Ah B h ∈ ΣhAB implies that (36) is satisfied for all obtain (40). We then conclude asymptotic stability via Theorem
 
Ah B h ∈ ΣhAB as well. From this, we have stability of the 21. 
  As in Section III, we conclude with the fact that since the true
uncertain switched system (35) for all Ah B h ∈ ΣhAB , h ∈ matrices are contained in the set of matrices that can explain
N[1,h] , by [51, Th. 2].  the data, robust stability of the uncertain lifted switched system
(35) implies stability of the true switched systems (21) and
To verify stability for a given h, one has to follow a two-
(24) and, thus, also of the aperiodically sampled system (3). A
step procedure: First, one has to invoke Algorithm 1 in order to  necessary requirement for this was that K h indeed follows the
determine parameterizations of the lifted matrices Ah B h , structure (23), which we achieved in the controller design result
and second, one uses these parameterizations in the stability Corollary 24.
conditions of Theorem 21. Corollary 25: Suppose the conditions of Theorem 21 or
Remark 22: Note that if one increases the tested sampling Corollary 24 are fulfilled. Then, the controller K renders the
interval from hold to hnew > hold , it is not required to rerun origin of the aperiodically sampled system (3) asymptotically
Algorithm 1 from h = 1. Instead, it is sufficient to iterate over stable for all hk ∈ N[1,h] .
the missing steps h ∈ N[hold ,hnew −1] due to its recursive nature. Theorem 21 provides a tool to check stability of the aperiodi-
Remark 23: For a given data length N , the highest sampling cally sampled system (3) given a controller K, and is, therefore,
interval that can be checked via Theorem 21 is h = N . suitable to tackle Problem 1. Since Corollary 24 permits a

⎡ ⎤
G h + G h − S h   
⎢ 
⎢ Gh 0  ⎥ ⎥
⎢  ⎥− 0  2
 0, (h, j) ∈ N[1,h] (40)
⎢  0 ⎥ 0 hj
⎣ K · · · K Gh 0 ⎦ P̃ AB

0 0 0 S j

⎡ ⎤
G + G − Sh   
⎢ 
⎢ G 0  ⎥⎥
⎢  ⎥− 0  2
 0, (h, j) ∈ N[1,h] (42)
⎢  ⎥ hj
0 P̃AB
⎣ F · · · F 0 0 ⎦
0 0 0 Sj

Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on December 18,2023 at 13:09:51 UTC from IEEE Xplore. Restrictions apply.
WILDHAGEN et al.: DATA-DRIVEN ANALYSIS AND CONTROLLER DESIGN FOR DISCRETE-TIME SYSTEMS UNDER APERIODIC SAMPLING 3221

cosearch for a stabilizing controller, it may be used to address

Problem 2.

V. NUMERICAL ANALYSIS
For a numerical evaluation of our approaches, we consider
the example from [54], exactly discretized with a discretization
period of H = 0.1 s. The system matrices are then
 
1 0.0995 0.0005
Atr = , Btr =
0 0.9900 0.0100
where we rounded off after four decimal places. Recall that
Atr and Btr are unknown to our data-driven approaches. We have
N state-input measurements {x(t)}N N −1 Fig. 2. MSI bounds with K = −[3.75 11.5], computed via ro-
t=0 and {u(t)}t=0 at each bust input/output (Theorem 11 —) and switched systems approach
of the time instants available, where the data-generating input (Theorem 21 —), and two-step procedure (set membership estimation
is taken uniformly from u(t) ∈ [−1, 1]. Let us assume that the and Theorem 10 - -), for three different data lengths and various distur-
ˆ N −1 , where
measurements are perturbed by a disturbance {d(t)} bance levels d. If there is no marker, the stability conditions were not
t=0
ˆ ∈ [−d, d] for some d ≥ 0, which implies that feasible for h = 1.
it holds that d(t)
Assumption 16 is fulfilled. Throughout this numerical example,
we describe this componentwise bound by the means of diagonal In contrast to the robust input/output approach, with the
multipliers (6) in Assumption 2. In addition, suppose it is known switched systems approach we observe that the MSI estimates
that the disturbance acts only on the first state, which we may improve significantly if more data are available. While the esti-
 
incorporate by setting Bd = 0.01 0 . Note that with this mates are lower than those of the robust input/output approach
for N = 5, they are significantly higher if N = 50 or N = 500
choice of Bd , Assumption 3 is fulfilled and a certain d corre- and the disturbance level is small. In some cases, even the true
sponds to an input-to-noise ratio of approximately 1/d. Further- MSI of 17 is recovered despite the data-driven setup and nonzero
more, in all of the tested cases, there existed σ 21 , P̃AB
1 1
∈ P̃AB s.t. noise.
(34) was fulfilled for h = 1, which is a precondition for Theorem When estimating the MSI from data, there are in general two
21 and Corollary 24. The numerical results were obtained using main sources of conservatism: The first one being that of the
MATLABR2019b, YALMIP [55], and Mosek [56]. data-driven parameterization, and the second one being that of
the (model-based) stability conditions. In the robust input/output
A. Comparison of Robust Input/Output and Switched approach, the conservatism of the data-driven parameterization
Systems Approach in this example seems to be rather low, since the estimated MSIs
are higher than for the switched systems approach especially
1) Comparison of Different Sizes N of the Dataset:
when the dataset is small and the disturbance level is high. On
First,
 let us consider
 Problem 1 for the given controller K =
the other hand, the conservatism induced by comprehending
− 3.75 11.5 , the same that was considered in [4]–[6], and the delay operator as a bounded disturbance seems to be quite
[11]. For this special scenario, the MSI can be computed exactly high in this example. Even if the model was known exactly
if model knowledge is available. It was shown in [6] that it we find hMSI = 12, compared with the true MSI of 17. As
amounts to 1.7 s, which corresponds to h = 17 with our chosen discussed in Section IV, the data-driven parameterizations of
discretization period. We have presented two possibilities to the lifted matrices in the switched systems approach introduce
estimate lower bounds on the MSI directly from data, either via conservatism since they rely on an overapproximation of the set
Theorem 11 using the robust input/output approach or via The- of compatible matrices. As a result, the latter is outperformed by
orem 21 using the switched systems approach. Both approaches the robust input/output approach if the introduced conservatism
were tested with three different data lengths N = 5, N = 50, is considerable, e.g., when the noise level is high and/or few data
and N = 500, and different disturbance levels to investigate are available. The model-based switched stability conditions, in
their effect on the MSI bounds. The results can be found in contrast, seem to come with little conservatism compared to the
Fig. 2. ones from the robust input/output approach in this example. As
As a first observation, we see that the guaranteed MSI bounds a result, the MSI bounds with the switched systems approach are
decrease with increasing disturbance level for all three data very tight and are much improved over the ones from the robust
lengths and for both the robust input/output and the switched input/output approach if there is little noise and sufficiently
systems approach. For the robust input/output approach, for all informative data available.
tested data lengths, the MSI estimate amounts to hMSI = 12 2) Controller Design: Second, we turn our attention to
when d = 0. We also computed the MSI using the model-based Problem 2, where the aim is to optimize for a controller that
robust input/output conditions in Theorem 10 and obtained the gives a possibly high MSI bound. In the robust input/output
same value, i.e., hMSI = 12. With increasing disturbance level d, approach, we can do so by virtue of Corollary 14, while in the
initially, the MSI estimate barely decreases. However, at a certain switched systems approach, we may use Corollary 24 for this
d, the stability conditions become infeasible, i.e., we do not find task. The corresponding MSI bounds with N = 50 and different
a solution even for h = 1. We observe that for larger datasets, d are presented in Fig. 3.
the minimal disturbance level for which we have infeasibility We observe that for both approaches, the computed MSI
increases. bounds with K as an optimization variable are increased
Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on December 18,2023 at 13:09:51 UTC from IEEE Xplore. Restrictions apply.
 3222 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 68, NO. 6, JUNE 2023

that they are independent of h for the former, whereas they grow
quadratically with h for the latter. In addition, one needs to run
Algorithm 1 in the switched systems approach, which requires
solving h SDPs whose complexity is independent of h. This is
also reflected in the required computation times in Fig. 4: They
are independent of h for the robust input/output approach and
amount to approximately 0.35 s for all experiments. By contrast,
we can observe that the required time for the switched systems
approach increases with h.
Fig. 3. MSI bounds with controller K computed via robust input/output Second, we compare the complexity of the data-driven stabil-
(Corollary 14 —) and switched systems approach (Corollary 24 —), for ity conditions Theorems 11 and 21 with their model-based coun-
N = 50 and various disturbance levels d. terparts Theorem 10 and [51, Th. 2]. We note that the number of
TABLE I
decision variables and constraints is the same for the data-driven
NUMBER OF DECISION VARIABLES AND CONSTRAINTS IN THE ROBUST conditions as for the model-based ones, with the exception that
INPUT/OUTPUT (THEOREMS 10 AND 11) AND SWITCHED SYSTEMS the former additionally include the number of decision variables
APPROACH (SEE[51, TH. 2] AND THEOREM 21), AND TWO-STEP cd (N ) involved in the chosen disturbance multiplier. As a result,
PROCEDURE (SET MEMBERSHIP ESTIMATION AND THEOREM 10)
the increase in complexity with increasing system dimension is
the same for the data-driven and model-based conditions.
Third, we discuss the influence of data length on complex-
ity and performance of the proposed data-driven approaches.
. To this end, assume that the disturbance satisfies a com-
ponentwise bound as in Assumption 16. Then, one could
choose either a quadratic or a diagonal disturbance multiplier
and
(see Remark 17). With a quadratic multiplier, one has cd (N ) = 1
such that the complexity of the proposed data-driven approaches
is independent of the data length. On the other hand, there is
no guarantee that more data will shrink the set of compatible
matrices and thereby lead to a larger MSI estimate. Quite the
contrary, in [23] and [36], it was shown that adding more data can
even decrease performance in this case. In contrast, a diagonal
multiplier provides the guarantee that more data will never
decrease performance [23], [36]. However, this comes at the
price that complexity of the stability conditions now increases
linearly with the data length as cd (N ) = N .

B. Comparison With Alternative Methods

1) Comparison With [25]: As discussed in the introduction,
Fig. 4. Computation times to check the stability conditions in for the results of [25] on data-driven control of time-delay systems
the robust input/output (Theorem 11) and switched systems approach
permit an analysis of aperiodically sampled systems as well. The
(Algorithm 1 and Theorem 21), for N = 50, disturbance level d = 0.01,
and different h.
prerequisites of [25, Th. 4.2] were fulfilled in the noise-free case
d = 0 if N ≥ 5. For any such data length, this result yielded an
MSI estimate of hMSI = 4 with the controller K = −[3.75 11.5].
compared to K = −[3.75 11.5], especially so for the robust In most cases, the data-driven methods proposed in this
input/output approach. Now, the robust input/output approach article yield a considerably better estimate of the MSI. This
yields larger MSI estimates than the switched systems approach might be due to the fact that since the latter work considers
for any of the considered disturbance levels. arbitrary bounded delays, it comes with a certain degree of
3) Comparison of Complexity: Here, we will briefly com- conservatism when applied to aperiodically sampled systems
pare the computational complexity of the robust input/output and (cf., Sections III-A and III-B).
switched systems approach. Table I tabulates an overview over 2) Comparison With Two-Step Procedure: To leverage
the number of decision variables and the number of constraints data for an estimation of the MSI, an alternative to the approaches
involved in the respective stability conditions for analysis, i.e., presented in this article is to perform a two-step procedure of sys-
Theorem 10 and [51, Th. 2] in the model-based case, and tem identification and subsequently checking the model-based
Theorems 11 and 21, and Algorithm 1 in the data-driven case. stability conditions in [5] and [8]–[10], or in Theorem 10. In
In Fig. 4, the computation times required to check the stability this section, we will briefly compare this two-step procedure to
conditions in Theorems 11 and 21, and to run Algorithm 1 are our data-driven approaches numerically. We will consider set
depicted for various h. All experiments were run on an Intel membership estimation [28], [29] for the system identification
Core i7-105110U with 1.80 GHz. step, since it delivers guaranteed error bounds despite the pres-
Comparing first the number of decision variables and con- ence of disturbances. In particular, we follow a straightforward
straints in the data-driven stability conditions of the robust approach and estimate a polytope in which the true matrices
input/output and switched systems approach in Table I, we note are guaranteed to lie, subsequently overbound this polytope by
Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on December 18,2023 at 13:09:51 UTC from IEEE Xplore. Restrictions apply.
WILDHAGEN et al.: DATA-DRIVEN ANALYSIS AND CONTROLLER DESIGN FOR DISCRETE-TIME SYSTEMS UNDER APERIODIC SAMPLING 3223

a hypercube and
 check the stability conditions in Theorem 10 Since the importance of data as well as of cyber-physical,
with K = − 3.75 11.5 for each of its vertices. Note that embedded and NCSs continues to grow, combining concepts
from data-driven control and sampled-data control is a highly
this procedure provides a guaranteed lower bound on the MSI.
relevant research direction. The data-driven analysis of ape-
The results of this method can be found in Fig. 2, next to
riodically sampled systems, as presented in this work, may
the results of the proposed data-driven method based on the
contribute to this emerging field by providing a novel approach
robust-input output approach, namely Theorem 11. The number
to model and analyze a great variety of problems at the in-
of involved decision variables and constraints for both options
tersection of data-driven and sampled-data control, such as
are summarized in Table I.
learning event-triggered control [60], learning unknown channel
It can be recognized that, for a small set of noise levels, the
conditions [61], or data-driven network access scheduling [62].
two-step procedure gives slightly better results than the pro-
posed data-driven approach. However, the number of involved
constraints grows exponentially with the square of the system APPENDIX
dimension. This is due to the fact that one constraint is added for A. Proof of Lemma 6
each vertex of the matrix hypercube, whose dimension
  is equal As done in the continuous-time case [11], we handle the
to the number of coefficients in the matrices A B , namely operator Δ in the lifted domain. For a signal g ∈ n2e , the
n(n + m). For this reason, the stability conditions in this two- corresponding lifted signal is defined as
step procedure become practically intractable even for modest ⎧⎡ ⎤ ⎡ ⎤ ⎫
⎪
⎨ g(0) g(t1 ) ⎪
⎬
system dimensions. In contrast, and as already recognized above,
⎢ .. ⎥ ⎢ .. ⎥
the proposed data-driven approaches grow on the same order as g := ⎣ . ⎦ , ⎣ . ⎦ , . . . .
⎪
⎩ ⎪
⎭
their model-based counterparts, namely quadratically with the g(t1 − 1) g(t2 − 1)
system dimension.
Further, for a T ∈ N0 , we define the lifted truncated signal
corresponding to gT as
⎧ ⎡ ⎤ ⎫
VI. SUMMARY AND OUTLOOK ⎪
⎪ ⎡ ⎤ g(tK ) ⎡ ⎤ ⎪
⎪
⎪
⎪ g(0) ⎢ . ⎥ 0 ⎪
⎪
⎨ ⎢ ⎥ ⎬
In this article, we approached a problem at the intersection ⎢ .. ⎥ ⎢
.
. ⎥ ⎢ .. ⎥
g T := ⎣ . ⎦ , . . . , ⎢ ,
⎥ . ⎣ ⎦ , . . .
of sampled-data control and data-driven control: We developed ⎪
⎪ ⎣ g(T ) ⎦ ⎪
⎪
⎪
⎪ ⎪
⎪
tools to lower bound the MSI of a system under aperiodic ⎩ g(t1 − 1) 0
0 ⎭
sampling, requiring no model knowledge and using only a
measured trajectory. Thereby, we considered both analysis and where K := max{k | tk ≤ T }. As the tk are not equidistant in
controller design directly from the available data, which may be time, the time axis in the lifted signal is split nonuniformly as
of finite length and subject to noise. In particular, we presented well. Nonetheless, the 2 norm of the original signal is preserved
two distinct approaches to achieve these goals, the first taking in the lifted domain g T 22 = gT 22 [1]. Further, we introduce
a robust control perspective and the second a switched systems a lifted operator Δ mapping y → e via
perspective on the aperiodically sampled system. The former, e(k) = (Δy)(k) := Ck y(k), k ∈ N0
dubbed robust input/output approach, comes with a lower com-
with
putational complexity and is able to produce decent estimates ⎡ ⎤
of the MSI even if the dataset is small and the noise level is 0 ··· 0 0
high. The latter, dubbed switched systems approach, can yield ⎢ .. .. ⎥
⎢I ..
.⎥
⎢ . . ⎥
very tight estimations of the MSI especially when the dataset is Ck := ⎢ . ⎥ ∈ Rhk ×hk .
sufficiently large and there is little noise. The validity of both ⎢. .. ⎥
⎣. . 0 0⎦
approaches was illustrated with a numerical example.
In the continuous-time formulation of the robust input/output I ··· I 0
approach, a passivity-like property of the delay operator was Now, let us rewrite Δ using τ (t) = t − tk as
established in addition to a bound on the 2 gain. Future work

t−1
could investigate this topic in the discrete-time case as well, (Δy)(t) := y(i), t ∈ N[tk ,tk +hk −1] , k ∈ N0
since it was shown in [12] that incorporating this property might i=tk
lead to greatly improved estimates of the MSI. In addition, it
might be possible to improve the 2 gain estimate of the delay from which it is easy to recognize that Δy T = (Δy)T for any
operator even further. In the switched systems approach, future T ∈ N0 . Since lifting preserves the signal norms, it clearly holds
work could focus on whether and how less conservative param- that Δy T 2 = ΔyT 2 . Since both Δ and Δ are causal and
eterizations of the lifted matrices could be obtained. A starting y T 22 = yT 22 , we conclude that the 2 gain of Δ is equal to
point might be to construct tighter overapproximations of the that of Δ.
lifted disturbance, e.g., by trying to incorporate information The crucial property of Δ is that it is static, i.e., e(k) depends
on its structure [cf., (26)]. A further line of future research on y(k) only. Furthermore, it merely sums the inputs y(t)
could investigate how to incorporate a performance objective in between sampling instants. As a result, its 2 gain is the
into the switched systems approach. Finally, future work could maximum 2 gain of the summation operator over the intervals
extend data-driven analysis of the MSI to more general system [0, hk − 1], k ∈ N0 , which is clearly attained in the longest
classes, e.g., polynomial systems [57], [58] or general nonlinear possible interval h − 1. To summarize, the 2 gain of Δ is
systems [59]. given by that of D : n2e [0, h − 1] → n2e [0, h − 1], y → Dy, and
Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on December 18,2023 at 13:09:51 UTC from IEEE Xplore. Restrictions apply.
 3224 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 68, NO. 6, JUNE 2023


(Dy)(t) := t−1 i=0 y(i), where 2 [0, h − 1] denotes the space of
[6] P. Naghshtabrizi, J. P. Hespanha, and A. R. Teel, “Exponential stability of
impulsive systems with application to uncertain sampled-data systems,”
bounded signals of length h. Syst. Control Lett., vol. 57, no. 5, pp. 378–385, 2008.
Dy
We will now bound D 2 = supy∈n2e ,y=0 y  2 explicitly. [7] D. Carnevale, A. R. Teel, and D. Nesic, “A Lyapunov proof of an
2 improved maximum allowable transfer interval for networked control
Let us first take a look at the square of the fraction systems,” IEEE Trans. Autom. Control, vol. 52, no. 5, pp. 892–897,
h−1 t−1 
 
t−1
 May 2007.
Dy 2 2
t=0 i=0 y(i) · i=0 y(i) [8] L. Hetel, A. Kruszewski, W. Perruquetti, and J.-P. Richard, “Discrete and
= h−1 . (43) intersample analysis of systems with aperiodic sampling,” IEEE Trans.
y 22 
y(t) y(t) Autom. Control, vol. 56, no. 7, pp. 1696–1701, Jul. 2011.
t=0
[9] J. Xiong and J. Lam, “Stabilization of linear systems over networks with
Next, we turn our attention to the product of sums in the bounded packet loss,” Automatica, vol. 43, pp. 80–87, 2007.
numerator, which can be rewritten and upper bounded for all [10] M. Yu, L. Wang, T. Chu, and G. Xie, “Stabilization of networked control
t ∈ N[0,h−1] as follows (denoting the maximum eigenvalue of a systems with data packet dropout and network delays via switching system
approach,” in Proc. Conf. Decis. Control, 2004, pp. 3539–3544.
symmetric matrix A ∈ Rn×n by λmax (A)): [11] L. Mirkin, “Some remarks on the use of time-varying delay to model
%t−1 & %t−1 & sample-and-hold circuits,” IEEE Trans. Autom. Control, vol. 52, no. 6,
 

y(i) · y(i) pp. 1109–1112, Jun. 2007.
i=0 i=0 [12] H. Fujioka, “Stability analysis of systems with aperiodic sample-and-hold
⎛ ⎞ devices,” Automatica, vol. 45, no. 3, pp. 771–775, 2009.
[13] L. Ljung, System Identification: Theory for the User. Englewood Cliffs,
⎡ ⎤  ⎜⎡ ⎤ ⎟⎡ ⎤ NJ, USA: Prentice-Hall, 1987.
y(0) ⎜ 1 ··· 1 0 ⎟ y(0) [14] Z.-S. Hou and Z. Wang, “From model-based control to data-driven con-
⎢ .. ⎥ ⎜ ⎜⎢ .. . . .. .. ⎥
⎟
⎟⎢ .. ⎥ trol: Survey, classification and perspective,” Inf. Sci., vol. 235, pp. 3–35,
⎢ ⎥ ⎜⎢ .⎥ ⎢ ⎥
⎢
=⎢ . ⎥ ⎢. . . ⎥ ⊗I ⎟ ⎢ . ⎥ 2013.
⎥ ⎜ ⎜ ⎢ ⎥ ⎟ ⎢ ⎥
0⎦ ⎟
[15] J. C. Willems, P. Rapisarda, I. Markovsky, and B. De Moor, “A note
⎣y(t − 1)⎦ ⎜⎣1 · · · 1 ⎟ ⎣ y(t − 1) ⎦ on persistency of excitation,” Syst. Control Lett., vol. 54, pp. 325–329,
⎜ 0 ··· 0 0 ⎟
y(t) ⎝ ⎠ y(t) 2005.
* +, - [16] T. M. Maupong, J. C. Mayo-Maldonado, and P. Rapisarda, “On Lyapunov
=:Et functions and data-driven dissipativity,” IFAC-PapersOnLine, vol. 50,
no. 1, pp. 7783–7788, 2017.
1 12 
t [17] A. Koch, J. Berberich, and F. Allgower, “Provably robust verification of
1 1
≤ λmax (Et ⊗ I) [y(0) · · · y(t)] 2 = t y(i) y(i). dissipativity properties from data,” IEEE Trans. Autom. Control, to be
i=0 published, doi: 10.1109/TAC.2021.3116179.
[18] J. Coulson, J. Lygeros, and F. Dörfler, “Data-enabled predictive control: In
(44) the shallows of the DeePC,” in Proc. Eur. Control Conf., 2019, pp. 307–312.
Inequality (44) follows directly from the facts that Et ⊗ I [19] J. Berberich, J. Köhler, M. A. Müller, and F. Allgöwer, “Data-driven model
predictive control with stability and robustness guarantees,” IEEE Trans.
is symmetric and that λmax (Et ⊗ I) = λmax (Et )λmax (I) = t. Autom. Control, vol. 66, no. 4, pp. 1702–1717, Apr. 2021.
Since it clearly holds that λmax (I) = 1, it remains to prove [20] C. De Persis and P. Tesi, “Formulas for data-driven control: Stabilization,
that λmax (Et ) = t: First, we check that t is indeed an eigen- optimality and robustness,” IEEE Trans. Autom. Control, vol. 65, no. 3,
value of Et with the corresponding eigenvector [1 . . . 1 0] . pp. 909–924, Mar. 2020.
[21] J. Berberich, A. Koch, C. W. Scherer, and F. Allgöwer, “Robust data-
Second, we note that Et is positive semidefinite, since it can driven state-feedback design,” in Proc. Amer. Control Conf., 2020,
be factorized as Et = [1 . . . 1 0] [∗], i.e., all eigenvalues are pp. 1532–1538.
real and greater
 than or equal to zero. Finally, we note that [22] H. van Waarde, M. K. Camlibel, and M. Mesbahi, “From noisy
tr(Et ) = t = i λi (Et ) and combine this finding with the first data to feedback controllers: Non-conservative design via a matrix S-
lemma,” IEEE Trans. Autom. Control, vol. 67, no. 1, pp. 162–175,
two to conclude that λmax (Et ) = t. Jan. 2022.
We plug the estimate (44) into (43) to obtain [23] J. Berberich, C. W. Scherer, and F. Allgöwer, “Combining prior knowledge
h−1 t and data for robust controller design,” 2020, arXiv:2009.05253.
Dy 22 t i=0 y(i) y(i) [24] J. Berberich, S. Wildhagen, M. Hertneck, and F. Allgöwer, “Data-driven
2 ≤ t=0  analysis and control of continuous-time systems under aperiodic sam-
y 2 h−1 
t=0 y(t) y(t) pling,” in Proc. IFAC Symp. Syst. Identification, 2021, pp. 210–215.
[25] J. G. Rueda-Escobedo, E. Fridman, and J. Schiffer, “Data-driven control
h−1 h−1
t i=0 y(i) y(i) 
h−1 for linear discrete-time delay systems,” IEEE Trans. Autom. Control, to be
h
≤ t=0 h−1 = t = (h − 1) published, doi: 10.1109/TAC.2021.3096896.
 2 [26] A. Seuret, “A novel stability analysis of linear systems under asynchronous
t=0 y(t) y(t) t=0
samplings,” Automatica, vol. 48, pp. 177–182, 2012.
which concludes the proof. [27] N. Matni, A. Proutiere, A. Rantzer, and S. Tu, “From self-tuning regulators
to reinforcement learning and back again,” in Proc. Conf. Decis. Control,
2019, pp. 3724–3740.
REFERENCES [28] M. Milanese and A. Vicino, “Optimal estimation theory for dynamic
systems with set membership uncertainty: An overview,” Automatica,
[1] T. Chen and B. A. Francis, Optimal Sampled-Data Control Systems. Berlin,
vol. 27, no. 6, pp. 997–1009, 1991.
Germany: Springer, 1995.
[29] G. Belforte, B. Bona, and V. Cerone, “Parameter estimation algorithms for
[2] L. Hetel et al., “Recent developments on the stability of systems with
a set-membership description of uncertainty,” Automatica, vol. 26, no. 5,
aperiodic sampling: An overview,” Automatica, vol. 76, pp. 309–335,
pp. 887–898, 1990.
2017.
[30] J. P. Hespanha, P. Naghshtabrizi, and Y. Xu, “A survey of recent results
[3] E. Fridman, A. Seuret, and J.-P. Richard, “Robust sampled-data stabiliza-
in networked control systems,” Proc. IEEE, vol. 95, no. 1, pp. 138–162,
tion of linear systems: An input delay approach,” Automatica, vol. 40,
Jan. 2007.
pp. 1441–1446, 2004.
[31] X. Zhang, Q. Han, and X. Yu, “Survey on recent advances in networked
[4] E. Fridman, “A refined input delay approach to sampled-data control,”
control systems,” IEEE Trans. Ind. Inform., vol. 12, no. 5, pp. 1740–1752,
Automatica, vol. 46, no. 2, pp. 421–427, 2010.
Oct. 2016.
[5] A. Seuret and E. Fridman, “Wirtinger-like Lyapunov–Krasovskii function-
[32] W. P. M. H. Heemels, K. H. Johansson, and P. Tabuada, “An introduction to
als for discrete-time delay systems,” IMA J. Math. Control Inf., vol. 35,
event-triggered and self-triggered control,” in Proc. Conf. Decis. Control,
no. 3, pp. 861–876, 2018.
2012, pp. 3270–3285.
Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on December 18,2023 at 13:09:51 UTC from IEEE Xplore. Restrictions apply.
WILDHAGEN et al.: DATA-DRIVEN ANALYSIS AND CONTROLLER DESIGN FOR DISCRETE-TIME SYSTEMS UNDER APERIODIC SAMPLING 3225

[33] W. P. M. H. Heemels, M. C. F. Donkers, and A. R. Teel, “Periodic event- [61] K. Gatsis and G. J. Pappas, “Sample complexity of networked control
triggered control for linear systems,” IEEE Trans. Autom. Control, vol. 58, systems over unknown channels,” in Proc. Conf. Decis. Control, 2018,
no. 4, pp. 847–861, Apr. 2013. pp. 6067–6072.
[34] R. Postoyan, R. G. Sanfelice, and W. Heemels, “Inter-event times analysis [62] A. S. Leong, A. Ramaswamy, D. E. Quevedo, H. Karl, and L. Shi, “Deep
for planar linear event-triggered controlled systems,” in Proc. Conf. Decis. reinforcement learning for wireless sensor scheduling in cyber–physical
Control, 2019, pp. 1662–1667. systems,” Automatica, vol. 113, 2020, Art. no. 108759.
[35] G. Gleizer and M. Mazo, “Towards traffic bisimulation of linear peri-
odic event-triggered controllers,” IEEE Control Syst. Lett., vol. 5, no. 1,
pp. 25–30, Jan. 2021.
[36] A. Bisoffi, C. De Persis, and P. Tesi, “Trade-offs in learning controllers Stefan Wildhagen received the master’s de-
from noisy data,” Syst. Control Lett., vol. 154, 2021, Art. no. 104985. gree in engineering cybernetics from the Uni-
[37] D. E. Knuth, The Art of Computer Programming. London, U.K.: Pearson versity of Stuttgart, Stuttgart, Germany, in 2018.
Educ., 1997. He is currently working toward the Ph.D. degree
[38] J. L. Bentley and A. C.-C. Yao, “An almost optimal algorithm for un- in networked control systems, with a focus on
bounded searching,” Inf. Process. Lett., vol. 5, no. 3, pp. 82–87, 1976. optimization-based scheduling and control as
[39] C.-Y. Kao and B. Lincoln, “Simple stability criteria for systems with time- well as on data-driven methods.
varying delays,” Automatica, vol. 40, no. 8, pp. 1429–1434, 2004. He is currently with the Institute for Systems
[40] C. Kao, “On stability of discrete-time LTI systems with varying time Theory and Automatic Control, Stuttgart, Ger-
delays,” IEEE Trans. Autom. Control, vol. 57, no. 5, pp. 1243–1248, many, under supervision of Prof. Allgöwer.
May 2012. Mr. Wildhagen is a Member of the Graduate
[41] A. Megretski and A. Rantzer, “System analysis via integral quadratic School Simulation Technology, University of Stuttgart.
constraints,” IEEE Trans. Autom. Control, vol. 42, no. 6, pp. 819–830,
Jun. 1997.
[42] J. Veenman, C. W. Scherer, and H. Köroğlu, “Robust stability and perfor-
mance analysis based on integral quadratic constraints,” Eur. J. Control, Julian Berberich received the master’s degree
vol. 31, pp. 1–32, 2016. in engineering cybernetics from the University
[43] B. Hu, M. J. Lacerda, and P. Seiler, “Robustness analysis of un- of Stuttgart, Stuttgart, Germany, in 2018. Since
certain discrete-time systems with dissipation inequalities and integral 2018, he has been working toward the Ph.D.
quadratic constraints,” Intl. J. Robust Nonlinear Control, vol. 27, no. 11, degree in the field of data-driven analysis and
pp. 1940–1962, 2017. control with the Institute for Systems Theory and
[44] C. Scherer, “Dissipativity and integral quadratic constraints: Tailored com- Automatic Control, Stuttgart, under supervision
putational robustness tests for complex interconnections,” IEEE Control of Prof. Frank Allgöwer.
Syst. Mag., vol. 42, no. 3, pp. 115–139, Jun. 2022. Mr. Berberich was the recipient of the Out-
[45] C. Scherer and S. Weiland, Linear Matrix Inequalities in Control, 3rd ed. standing Student Paper Award at the 59th Con-
New York, NY, USA: Springer, 2000. ference on Decision and Control in 2020. He is
[46] N. Kottenstette, M. J. McCourt, M. Xia, V. Gupta, and P. J. Antsaklis, “On a Member of the International Max-Planck Research School, University
relationships among passivity, positive realness, and dissipativity in linear of Stuttgart.
systems,” Automatica, vol. 50, no. 4, pp. 1003–1016, 2014.
[47] G. Zames and P. L. Falb, “Stability conditions for systems with mono-
tone and slope-restricted nonlinearities,” SIAM J. Control, vol. 6, no. 1,
pp. 89–108, 1968. Michael Hertneck received the master’s degree
[48] C. Scherer, “Robust mixed control and linear parameter-varying control in mechatronics from the University of Stuttgart,
with full-block scalings,” in Proc. Adv. Linear Matrix Inequality Methods Stuttgart, Germany, in 2019.
Control, 2000, pp. 187–207. He is currently a Research and Teaching As-
[49] D. Nešić, A. R. Teel, and E. D. Sontag, “Formulas relating KL stability sistant with the Institute for Systems Theory and
estimates of discrete-time and sampled-data nonlinear systems,” Syst. Automatic Control, Stuttgart. His research inter-
Control Lett., vol. 38, no. 1, pp. 49–60, 1999. ests focuses on networked control systems with
[50] D. Liberzon and A. S. Morse, “Basic problems in stability and design of a focus on time- and event-triggered sampling
switched systems,” IEEE Control Syst. Mag., vol. 19, no. 5, pp. 59–70, strategies.
Oct. 1999. Mr. Hertneck is a member of the Gradu-
[51] J. Daafouz, P. Riedinger, and C. Iung, “Stability analysis and control ate School Simulation Technology, University of
synthesis for switched systems: A switched Lyapunov function approach,” Stuttgart.
IEEE Trans. Autom. Control, vol. 47, no. 11, pp. 1883–1887, Nov. 2002.
[52] D. Liberzon, Switching in Systems and Control. Boston, MA, USA:
Birkhäuser, 2003.
[53] X. Wang, J. Berberich, J. Sun, G. Wang, F. Allgöwer, and J. Chen, Frank Allgöwer received the Engineering de-
“Data-driven control of event- and self-triggered discrete-time systems,” gree in cybernetics and applied mathematics
2022, arXiv:2202.08019. from the University of Stuttgart, Stuttgart, Ger-
[54] W. Zhang, M. Branicky, and S. Phillips, “Stability of networked control many, and the University of California, Los An-
systems,” IEEE Control Syst. Mag., vol. 21, no. 1, pp. 84–99, Feb. 2001. geles, CA, USA, respectively, in 1987 and the
[55] J. Löfberg, “YALMIP: A toolbox for modeling and optimization in MAT- Ph.D. degree in control engineering from the
LAB,” in Proc. IEEE Int. Conf. Comput. Aided Control Syst. Des., 2004, University of Stuttgart in 1996.
pp. 284–289. He is currently a Professor of mechanical en-
[56] MOSEK ApS, “The MOSEK optimization toolbox for MATLAB manual,” gineering with the University of Stuttgart, where
2015. [Online]. Available: https://docs.mosek.com/latest/toolbox/index. he is also the Director of the Institute for Sys-
html tems Theory and Automatic Control (IST).
[57] T. Martin and F. Allgöwer, “Dissipativity verification with guarantees for He is active in serving the community in several roles: Among others,
polynomial systems from noisy input-state data,” IEEE Control Syst. Lett., he has been the President of the International Federation of Automatic
vol. 5, no. 4, pp. 1399–1404, Oct. 2021. Control for the years 2017–2020, Vice President of the Technical Ac-
[58] M. Guo, C. De Persis, and P. Tesi, “Data-driven stabilization of nonlinear tivities of the IEEE Control Systems Society for 2013/14, and Editor of
polynomial systems with noisy data,” IEEE Trans. Autom. Control, to be the journal Automatica from 2001 to 2015. From 2012 to 2020, he was
published, doi: 10.1109/TAC.2021.3115436. the Vice President of German Research Foundation, Bonn, Germany,
[59] T. Martin and F. Allgöwer, “Data-driven system analysis of nonlinear which is Germany’s most important research funding organization. His
systems using polynomial approximation,” 2021, arXiv:2108.11298. research interests include predictive control, data-based control, net-
[60] L. Sedghi, Z. Ijaz, M. Noor-A-Rahim, K. Witheephanich, and D. Pesch, worked control, cooperative control, and nonlinear control with applica-
“Machine learning in event-triggered control: Recent advances and open tion to a wide range of fields including systems biology.
issues,” 2020, arXiv:2009.12783.
Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on December 18,2023 at 13:09:51 UTC from IEEE Xplore. Restrictions apply.