162 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 67, NO.

1, JANUARY 2022

From Noisy Data to Feedback Controllers:

Nonconservative Design via a Matrix S-Lemma
Henk J. van Waarde , M. Kanat Camlibel , Member, IEEE, and Mehran Mesbahi , Fellow, IEEE

Abstract—In this article, we propose a new method to ob- In fact, one of the unsolved problems is to come up with
tain feedback controllers of an unknown dynamical system nonconservative control design strategies using only a finite
directly from noisy input/state data. The key ingredient of number of data samples.
our design is a new matrix S-lemma that will be proven in
this article. We provide both strict and nonstrict versions We will tackle this problem by providing necessary and
of this S-lemma, which are of interest in their own right. sufficient conditions on noisy data, under which controllers can
Thereafter, we will apply these results to data-driven con- be obtained. As a consequence, our ensuing control technique is
trol. In particular, we will derive nonconservative design nonconservative and also shown to be tractable from a compu-
methods for quadratic stabilization, H2 and H∞ control, all tational point of view. The technical ingredient that enables our
in terms of data-based linear matrix inequalities. In contrast
to previous work, the dimensions of our decision variables design is a new generalization of the classical S-lemma [1], [2],
are independent of the time horizon of the experiment. Our which will be proven in this article. We will formulate our control
approach, thus, enables control design from large datasets. problems using the general data informativity framework, as
Index Terms—Data-driven control, LMIs, robust control, introduced in [3]. As such, the results developed in this article
uncertain systems. can be seen as a natural extension of those in [3] to noisy data.

A. Literature on Data-Driven Control

I. INTRODUCTION
The literature on data-driven control is expanding rapidly. Our
N THIS article, we study the problem of designing control
I laws for an unknown dynamical system using noisy data.
This general problem exists for a long time, but has seen a re-
account of previous work is, therefore, not exhaustive, but we
note that additional references can be found in the survey [4]. We
mention contributions to data-driven optimal control [5]–[16],
newed surge of interest over the past few years. The problem can proportional–integral–derivative control [17], [18], predictive
be approached via different angles, for example, using combined control [19]–[23], and nonlinear control [24]–[28]. Some of
system identification and model-based control, or by computing these techniques are iterative in nature: the controller is updated
control laws from data without the intermediate modeling step. online when new data are presented. Examples of this include
We will contribute to the second category of methods, aiming at policy iteration methods [29] and iterative feedback tuning [30].
control design directly from noisy data. Other methods are one-shot in the sense that the controller is con-
One of the main challenges in this area is to come up with structed offline from a batch of data. We mention, for instance,
robust control laws that guarantee stability and performance of virtual reference feedback tuning [31] and methods based on
the unknown system despite the inherent uncertainty caused by Willems’ fundamental lemma [32] (see also [33]). The latter
noisy data. Even though there are several recent contributions line of work has been quite fruitful, with contributions ranging
addressing this issue, there are multiple open questions. from output matching [34] and control by interconnection [35],
to data-enabled predictive control [36] and a data-based closed-
Manuscript received November 12, 2020; accepted December 19, loop system parameterization [37]. This parameterization has
2020. Date of publication December 25, 2020; date of current version been used for stabilizing and optimal control design using data-
December 29, 2021. This work was supported by the Data Science and
Systems Complexity Centre, University of Groningen, The Netherlands.
based linear matrix inequalities (LMIs) [37]. We also mention
Recommended by Associate Editor T. Iwasaki. (Corresponding author: the extension [38] studying linear–quadratic regulator using
Henk J. van Waarde.) noisy data and the article [39] for a closed-loop parameterization
Henk J. van Waarde is with the Bernoulli Institute for Mathemat-
ics, Computer Science, and Artificial Intelligence and the Engineering
using noisy data. Additional recent research directions include
and Technology Institute Groningen, University of Groningen, 9747 AG data-driven control of networks [40], [41] and the interplay
Groningen, The Netherlands (e-mail: [email protected]). between data-guided control and model reduction [42].
M. Kanat Camlibel is with the Bernoulli Institute for Mathematics,
Computer Science, and Artificial Intelligence, University of Groningen,
9747 AG Groningen, The Netherlands (e-mail: [email protected]). B. Review of the S-Lemma
Mehran Mesbahi is with the William E. Boeing Department of Aero-
nautics and Astronautics, University of Washington, Seattle, WA 98195 First proven by Yakubovich [1], the S-lemma is a classical
USA (e-mail: [email protected]). result in control theory and optimization [2]. The result revolves
Color versions of one or more figures in this article are available at
https://doi.org/10.1109/TAC.2020.3047577. around the question when the nonnegativity of one quadratic
Digital Object Identifier 10.1109/TAC.2020.3047577 function implies that of another. The crux of the S-lemma is that
0018-9286 © 2020 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 13:08:25 UTC from IEEE Xplore. Restrictions apply.
H. J. VAN WAARDE et al.: FROM NOISY DATA TO FEEDBACK CONTROLLERS: NONCONSERVATIVE DESIGN VIA A MATRIX S-LEMMA 163

this seemingly difficult implication is equivalent to the feasibility formulations in [37] and [39] that provide sufficient con-
of an LMI in a scalar variable, called a multiplier. The act of ditions for controller design.
replacing the implication by an LMI is often referred to as the 3) Last but not least, the variables involved in our method
S-procedure. The S-procedure is more generally applicable in are independent of the time horizon of the experiment.
situations where one quadratic inequality is implied by multiple Our approach is, thus, applicable to large datasets. This
quadratic inequalities [43]. In this case, multiple scalar multi- is an advantage over closed-loop system parameteri-
pliers are used. It is well known, however, that the S-procedure zations [37], [39], which become computationally in-
is conservative in general, although there exist special cases in tractable when applied to big data.
which losslessness (according to S-lemma) can be shown [2,
Sec. 3]. A generalized S-lemma involving more general types D. Outline of This Article
of multipliers was considered in [44] and [45] and was shown In Section II, we will formulate the problem. Section III
to be nonconservative under extra assumptions. contains our results on the matrix S-lemma. These results are
The classical S-lemma and the contributions mentioned above then applied to data-driven stabilization in Section IV and
all deal with vector variables. However, for reasons that will to data-driven H2 control and H∞ control in Section V. In
become clear in Section II, we need a type of S-lemma that is ap- Section VI, we provide simulation examples. Finally, Sec-
plicable to quadratic functions of matrix variables. Such a result tion VII concludes this article.
has been reported for specific quadratic functions in [46, Th. 3.3].
In the special case that variables are bounded, an S-lemma for II. PROBLEM OF DATA-DRIVEN STABILIZATION
matrix variables can also be derived by combining the so-called
full block S-procedure [47]–[50] with results from the literature Consider the linear time-invariant system
on LMI relaxations [51]–[53]. These specific results are, how- x(t + 1) = As x(t) + Bs u(t) + w(t) (1)
ever, less suited for the application of data-driven control that we
have in mind. Therefore, in this article, we derive general matrix where x ∈ Rn denotes the state, u ∈ Rm is the input, and
S-lemmas for both strict and nonstrict inequalities. Our matrix w ∈ Rn is an unknown noise term. The matrices As ∈ Rn×n
S-lemma for nonstrict inequalities is a direct generalization of and Bs ∈ Rn×m denote the unknown state and input matrices,
the classical S-lemma [1]. It also recovers the result from [46] as respectively. Our goal is to design stabilizing controllers for (1)
a special case. As a corollary of our matrix S-lemma for a strict on the basis of a finite number of measurements of the state and
inequality, we recover the S-lemma derived from the general input of the system. To this end, suppose that we measure state
theory on LMI relaxations [51]. and input data on a time interval1 and collect these samples in
the matrices

C. Our Contributions X := [x(0) x(1) · · · x(T )]

The core of our approach is to formulate data-driven control as U− := [u(0) u(1) · · · u(T − 1)].
the problem of deciding whether one quadratic matrix inequality By defining the matrices
(QMI) is implied by another one. Our first contribution is to
extend the classical S-lemma to quadratic matrix inequalities. X+ := [x(1) x(2) · · · x(T )]
Our second contribution is to apply these results to data-driven X− := [x(0) x(1) · · · x(T − 1)]
control. In particular, we come up with design procedures for
quadratic stabilization, H2 control, and H∞ control. W− := [w(0) w(1) · · · w(T − 1)]
Throughout this article, we will assume no statistics on the we clearly have
noise, but we will work with general bounded disturbances. We
are thereby inspired by recent papers [37], [39] that formalize X+ = As X− + Bs U− + W− . (2)
the assumption of bounded disturbances in terms of quadratic We emphasize that the system matrices As and Bs as well as the
matrix inequalities. In fact, we will work with an assumption noise term W− are unknown, while X and U− are measured. Be-
on the noise that is closely related to that of [39] and is more fore we introduce the problem, we will explain our assumption
general than the assumption in [37]. In terms of control design, on the noise W− .
our approach completely differs from the above papers. In fact,
instead of working with data-based parameterizations of closed- A. Assumption on the Noise
loop systems [37]–[39], we will work with a representation of
We will formalize our assumption on the noise in terms of a
all open-loop systems explaining the data, akin to the framework
QMI.
of [3]. We believe that our approach is attractive for the following
Assumption 1: The noise samples w(0), w(1), . . . ,
reasons.
w(T − 1), collected in the matrix W− , satisfy the bound
1) We provide robust guarantees on the stability and per-
    
formance of the unknown data-generating system. The I Φ11 Φ12 I
design involves data-guided LMIs that are tractable from 0 (3)
W− Φ
12 Φ22 W−
a computational point of view and are easy to implement.
2) By virtue of our matrix S-lemma, the design method is 1 All our results are still true for data collected on multiple intervals, see [3,
nonconservative. This is in contrast with previous LMI Ex. 2] for more details on how to arrange the data matrices in this case.
Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on December 18,2023 at 13:08:25 UTC from IEEE Xplore. Restrictions apply.
 164 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 67, NO. 1, JANUARY 2022

for known matrices Φ11 = Φ 

11 , Φ12 and Φ22 = Φ22 < 0.
B. Problem Formulation
Note that the negative definiteness of Φ22 ensures that the set We will follow the general framework for data-driven analysis
of noise matrices W− satisfying (3) is bounded. In the special and control in [3]. To this end, we define the set of all systems
case Φ12 = 0 and Φ22 = −I, (3) reduces to (A, B) explaining the data (U− , X), i.e., all (A, B) satisfying
T
 −1
W− W− = w(t)w(t)  Φ11 . (4) X+ = AX− + BU− + W− (5)
t=0
for some W− satisfying (3). We denote this set by Σ:
Inequality (4) has the interpretation that the energy of w is
bounded on the finite-time interval [0, T − 1]. If w is a random Σ := {(A, B) | (5) holds for some W− satisfying(3)}.
variable, its sample covariance matrix is given by
  We can only guarantee that a state feedback u = Kx stabilizes
1 1 the true system (As , Bs ) if it stabilizes all systems in Σ. This
W− I − J W−
T T motivates the following definition of informative data. Loosely
speaking, data are called informative if they enable the design
where J is the matrix of ones. Thus, (3) can also capture known
of a controller that stabilizes all systems in Σ (and thus, the
bounds on the sample covariance by the choices Φ12 = 0 and
unknown (As , Bs )).
Φ22 = − T1 (I − T1 J). We emphasize, however, that we do not
Definition 3: Assume that (U− , X) satisfies (2) for some W−
make any assumptions on the statistics of w and work with the
satisfying Assumption 1. The data (U− , X) are called informa-
general bound (3) instead. Note that [37, Asm. 5] is a special case
 tive for quadratic stabilization if there exists a feedback gain K
of Assumption 1 for the choices Φ11 = γX+ X+ with γ > 0,
and a matrix P = P  > 0 such that
Φ12 = 0, and Φ22 = −I. We remark that norm bounds on the
individual noise samples w(t) also give rise to bounds of the P − (A + BK)P (A + BK) > 0 (6)
form (3), although this may lead to some conservatism. Indeed,
note that w(t)22   implies that w(t)w(t)  I for all t. As for all (A, B) ∈ Σ.
such, the bound (4) is satisfied for Φ11 = T I. Note that, in particular, we are interested in quadratic sta-
Remark 1: Note that the noise model in Assumption 1 is the bilization, and we ask for a common Lyapunov matrix P for
“transposed” of the model in [39], in the sense that we penalize, all (A, B) ∈ Σ. We will not treat (A, B)-dependent Lyapunov
e.g., the term W− Φ22 W− instead of a term W− Qw W− . In some matrices in this article, but consider this case for future work
cases, these two different noise models are actually equivalent. instead.
For example, if Φ11 > 0 and Φ12 = 0, then (3) can be written Definition 3 leads to two natural problems. First, we are
via a Schur complement argument as interested in the question under which conditions the data are
  informative. We formalize this in the following problem.
Φ11 W− Problem 1 (Informativity): Assume that (U− , X) satisfies (2)
 0.
W− −Φ−1 22
for some W− satisfying Assumption 1. Find necessary and suf-
ficient conditions under which the data (U− , X) are informative
In turn, this is equivalent to −Φ−1  −1
22 − W− Φ11 W−  0, which is for quadratic stabilization.
of the same form as [39]. The second problem is a design issue: we are interested in
Remark 2: In some cases, we may know a priori that the noise procedures to come up with a feedback that stabilizes all systems
w does not directly affect the entire state space, but is contained in Σ.
in a subspace. This prior knowledge can be captured by the noise Problem 2 (Control design): Assume that (U− , X) satisfies
model in Assumption 1. Indeed, W− is of the form W− = E Ŵ− (2) for some W− satisfying Assumption 1. If the data (U− , X)
for some Ŵ− ∈ Rr×T satisfying are informative for quadratic stabilization, find a stabilizing
     feedback gain K such that (6) is satisfied for all (A, B) ∈ Σ.
I Φ̂11 Φ̂12 I In addition to data-driven stabilization, we are also interested
0
Ŵ− Φ̂
12 Φ̂22 Ŵ− in including performance specifications. Natural extensions to
Problems 1 and 2 will be discussed in Section V.
if and only if
    
I E Φ̂11 E  E Φ̂12 I C. Our Approach
 0.
W− Φ̂
12 E

Φ̂22 W− In what follows, we will outline our strategy for solving
Problems 1 and 2. Let (A, B) ∈ Σ and rewrite (5) as
Thus, the conclusion is that we can incorporate the knowledge
that W− ∈ im E by appropriate choices of the Φ-matrices in W− = X+ − AX− − BU− . (7)
(3). Showing the above claim is straightforward: note that the
“only if” statement follows by pre- and postmultiplication with Recall that by Assumption 1, we have
E and E  , respectively. The “if” part follows by noting that x ∈     
ker E  implies x W− Φ̂22 W− x  0; thus, W− x = 0. Hence, I Φ11 Φ12 I
 0.
ker E  ⊆ ker W− ; equivalently, im W− ⊆ im E. W− Φ
12 Φ22 W−

Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on December 18,2023 at 13:08:25 UTC from IEEE Xplore. Restrictions apply.
H. J. VAN WAARDE et al.: FROM NOISY DATA TO FEEDBACK CONTROLLERS: NONCONSERVATIVE DESIGN VIA A MATRIX S-LEMMA 165

By substitution of (7), this yields for some M11 ∈ R, M12 ∈ R1×n , and M22 = M22 
∈ Rn×n . A
⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤ homogeneous quadratic function of the form f (x) = x M22 x
I I X+   I X+ I is called a quadratic form. The following theorem describes the
⎢ ⎥ ⎢ ⎥ Φ11 Φ12 ⎢ ⎥ ⎢ ⎥
⎣ A ⎦ ⎣0 −X− ⎦  ⎣0 −X− ⎦ ⎣ A ⎦  0. celebrated S-lemma, proven by Yakubovich [1] (see also [2, Th.
Φ12 Φ22
B 0 −U− 0 −U− B 2.2]).
(8) Theorem 5 (S-lemma): Let f, g : Rn → R be quadratic func-
This shows that A and B satisfy a QMI of the form (8).2 In fact, tions. Suppose that there exists x̄ ∈ Rn such that g(x̄) > 0.
the set Σ of all systems explaining the data can be equivalently Then, f (x)  0 for all x ∈ Rn such that g(x)  0 if and only if
characterized in terms of (8), as asserted in the following lemma. there exists a scalar α  0 such that
Lemma 4: We have that Σ = {(A, B) | (8) is satisfied}.
f (x) − αg(x)  0 ∀x ∈ Rn . (11)
Proof: Suppose that (A, B) ∈ Σ. Then, (7) is satisfied for
some W− satisfying (3). This means that (8) holds. As such We note that the functions f and g are not assumed to be
convex. As such, it appears to be difficult to check the condition
Σ ⊆ {(A, B) | (8) is satisfied} .
f (x)  0 for all x ∈ Rn satisfying g(x)  0. The importance
To prove the reverse inclusion, let (A, B) be such that (8) is sat- of the S-lemma lies in the fact that the characterization (11) of
isfied. Define W− := X+ − AX− − BU− . By (8), W− satisfies this condition is equivalent to an LMI
the assumption (3). Since (5) holds for (A, B) by construction,    
we conclude that (A, B) ∈ Σ.  M11 M12 N11 N12

−α 
0
By Lemma 4, the set Σ of systems explaining the data is M12 M22 N12 N22
characterized by a QMI in (A, B). Next, we turn our attention to
in the scalar variable α  0. Here, the matrices N11 ∈ R,
the design condition (6). Suppose that we fix3 a Lyapunov matrix
N12 ∈ R1×n , and N22 ∈ Rn×n define the quadratic function g
P = P  > 0 and a feedback gain K. Note that inequality (6) is
analogous to (10).
equivalent to
The scalar α is called a multiplier, and the assumption g(x̄) >
⎡ ⎤ ⎡ ⎤⎡ ⎤
I P 0 0 I 0 for some x̄ ∈ Rn is often referred to as the Slater condition.
⎢ ⎥ ⎢  ⎥ ⎢ ⎥ > 0 This assumption is necessary in the sense that Theorem 5 is false
⎣A ⎦ ⎣ 0 −P −P K ⎦ ⎣A ⎦ (9)
  
without it. To show this by means of an example, one can take,
B 0 −KP −KP K B e.g., f (x) = x Ax and g(x) = −x Bx with A and B as in the
which is yet another QMI in A and B. Therefore, Problem example of [57, p. 4476]. A version of the S-lemma, where g
1 essentially boils down to understanding under which condi- satisfies a strict inequality, has been presented in [2, Th. 7.8].
tions the QMI (9) holds for all (A, B) satisfying the QMI (8). We will reformulate the result in the following theorem.
Data-driven stabilization, thus, naturally leads to the following Theorem 6 (Strict S-lemma): Let f, g : Rn → R be quadratic
fundamental question: forms. Suppose that there exists an x̄ ∈ Rn such that g(x̄) > 0.
Then, f (x)  0 for all x ∈ Rn such that g(x) > 0 if and only if
When does one QMI imply another QMI? there exists a scalar α  0 such that
The familiar reader will immediately recognize the similarity
f (x) − αg(x)  0 ∀x ∈ Rn .
between the above question and the statement of the so-called
S-lemma [2]. In fact, the S-lemma provides conditions under Note that Theorem 6 is stated with two multipliers in [2,
which the nonnegativity of one quadratic function implies that Th. 7.8]. However, the inclusion of the Slater condition allows
of another one. This motivates the following section, in which us to state Theorem 6 with a single multiplier α.
we generalize the S-lemma to matrix variables.
B. S-Lemma With Matrix Variables
III. MATRIX-VALUED S-LEMMA
Next, we aim at generalizing Theorems 5 and 6 to quadratic
In this section, we present a new S-lemma with matrix vari- functions of the form
ables. Before we do so, we provide a brief recap on the classical          
S-lemma. I M11 M12 I I N11 N12 I

and 
X M12 M22 X X N12 N22 X
A. Recap of the Classical S-Lemma (12)
where X ∈ Rn×k is a matrix variable, and the partitioned matri-
A function f : Rn → R is called quadratic if it can be written
ces M, N ∈ R(k+n)×(k+n) are real and symmetric. As our first
in the form
     step, the following theorem provides an S-lemma for homoge-
1 M11 M12 1 neous quadratic functions of the form X  M X and X  N X.
f (x) = 
(10) Naturally, instead of the nonnegativity of functions in the clas-
x M12 M22 x
sical S-lemma, we now consider the positive (semi)definiteness
2 We note that quadratic uncertainty descriptions have also arisen in the
of quadratic functions of matrix variables.
papers [54]–[56] studying data-driven control under the assumption that w is a Theorem 7 (Homogeneous matrix S-lemma): Let M, N ∈
normally distributed process noise.
3 We make this hypothesis purely to explain the ideas behind our approach. In Rn×n be symmetric matrices and assume that X̄  N X̄ > 0 for
fact, in Section IV, we show how P and K can be computed from data. some X̄ ∈ Rn×k . The following statements are equivalent.
Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on December 18,2023 at 13:08:25 UTC from IEEE Xplore. Restrictions apply.
 166 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 67, NO. 1, JANUARY 2022

       
i) X  M X  0 for all X ∈ Rn×k such that X  N X  0. I I I I
n×k
ii) X  M X  0 for all X ∈ Rn×k such that X  N X > 0. I) M  0 ∀Z ∈ R with N  0.
Z Z Z Z
iii) There exists a scalar α  0 such that M − αN  0.        
Remark 8: The assumption on the existence of X̄ such that I I n×k I I
X̄  N X̄ > 0 is a natural generalization of the Slater condition II) M  0 ∀Z ∈ R with N > 0.
Z Z Z Z
in Theorems 5 and 6. The assumption is again necessary in the
sense that Theorem 7 is false without it. Nonetheless, it can be III) There exists a scalar α  0 such that M − αN  0.
shown that the assumption can be weakened if one is interested Note that for k = 1, assumption (14) reduces to the standard
only in the equivalence of (i) and (iii). In fact, one can show using Slater condition. In this case, Theorem 9 recovers Theorems 5
similar arguments as in the proof of Theorem 7 that (i) ⇐⇒ and 6 in the following sense: the equivalence of (I) and (III) is
(iii) under the assumption that ∃x̄ ∈ Rn such that x̄ N x̄ > 0, the statement of Theorem 5. The equivalence of (II) and (III)
i.e., under the “standard” Slater condition. generalizes Theorem 6 for quadratic forms to general quadratic
Proof of Theorem 7: It is clear that (i) ⇒ (ii) and (iii) ⇒ (i). As functions.
such, it suffices to prove the implication (ii) ⇒ (iii). To this end, Proof of Theorem 9: Clearly, (I) ⇒ (II) and (III) ⇒ (I). Thus,
suppose that (ii) holds. Let x ∈ Rn be such that x N x > 0. We it suffices to prove that (II) ⇒ (III). Our strategy will be to show
want to prove that x M x  0 so that we can apply Theorem 6. that (II) implies statement (ii) of Theorem 7. To this end, suppose
Choose a vector v ∈ Rk such that v = 1. Next, we define that (II) holds, and let X ∈ R(k+n)×k be such that X  N X > 0.
the matrix X ∈ Rn×k as X := X̄ + xv  for  = 0. Clearly, Partition X as
 
X  N X is equal to X1
 X=
2 X̄  N X̄ +  X̄  N xv  + vx N X̄ + (x N x)vv  . X2

We claim that X  N X is positive definite for  sufficiently small. where X1 ∈ Rk×k and X2 ∈ Rn×k . Clearly, for all sufficiently
To prove this claim, first, suppose that y ∈ Rk is nonzero and small  > 0, we have
v  y = 0. Then, we obtain    
X1 + I X1 + I
y  X  N Xy = 2 y  X̄  N X̄y > 0. N > 0.
X2 X2
Second, suppose that y ∈ Rk is nonzero and v  y =: β = 0. Also note that X1 + I is nonsingular for all sufficiently small
Then, y  X  N Xy is equal to  > 0. This implies that
    
y  2 X̄  N X̄ +  X̄  N xv  + vx N X̄ y + (x N x)β 2
I I
N > 0.
which is positive for  sufficiently small since β = 0 and X2 (X1 + I)−1 X2 (X1 + I)−1
x N x > 0. We conclude that X  N X > 0 for  sufficiently
small. Now, by (ii), we conclude that X  M X  0. Multiplica- By (II), we have
tion of the latter inequality from left by v  and right by v yields    
I I
the inequality M 0
 X2 (X1 + I)−1 X2 (X1 + I)−1
2 v  X̄  M X̄v +  v  X̄  M x + x M X̄v + x M x  0.
(13) equivalently
This implies that x M x  0. Indeed, if x M x < 0, then there    
exists a sufficiently small  = 0 such that X1 + I X1 + I
M 0
 X2 X2
2 v  X̄  M X̄v +  v  X̄  M x + x M X̄v + x M x < 0
for all  > 0 sufficiently small. By taking the limit  ↓ 0, we con-
which contradicts (13). To conclude, we have shown that clude that X  M X  0. Therefore, statement (ii) [equivalently,
x M x  0 for all x ∈ Rn such that x N x > 0. By Theorem 6, statement (iii)] of Theorem 7 is satisfied. This means that (III)
condition (iii) is satisfied. This proves the theorem.  holds, which proves the theorem. 
Next, we build on Theorem 7 by introducing a general (in- As a special case of Theorem 9, we recover [46, Th. 3.3].
homogeneous) S-lemma with matrix variables. The following Corollary 10: The QMI M11 + M12 Z + Z  M12 
+
theorem is one of the main results of this section. Z  M22 Z  0 holds for all Z ∈ Rn×k satisfying I − Z  DZ 
Theorem 9 (Matrix S-lemma): Let M, N ∈ R(k+n)×(k+n) be 0 if and only if there exists a scalar α  0 such that
symmetric matrices, and assume that there exists some matrix    
Z̄ ∈ Rn×k such that M11 M12 I 0
−α  0.
    M12
M22 0 −D
I I
N > 0. (14) Proof: Note that the generalized Slater condition (14) is sat-
Z̄ Z̄
isfied (one can choose e.g., Z̄ = 0). Thus, the statement follows
Then, the following statements are equivalent: from Theorem 9. 

Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on December 18,2023 at 13:08:25 UTC from IEEE Xplore. Restrictions apply.
H. J. VAN WAARDE et al.: FROM NOISY DATA TO FEEDBACK CONTROLLERS: NONCONSERVATIVE DESIGN VIA A MATRIX S-LEMMA 167

Theorem 9 provides a natural generalization of the S-lemma to matrix S-lemma is also a corollary of Theorem 11. To obtain the
matrix variables. However, note that for the application that we result, we substitute A = 0, B = I, W (x) = −M , and
have in mind, we need a slightly different version of the theorem. ⎧     ⎫
Indeed, note that in the data-driven context of Section II, a strict ⎨ Δ Δ ⎬
inequality (9) must hold for all (A, B) satisfying a nonstrict Δ= Δ| N 0
⎩ I I ⎭
inequality (8). As such, we need to extend Theorem 9 to the
case when the inequality involving M is strict. Before we do so, into [51, eq. (1.2)]. Then, we combine the full block S-procedure
we introduce the shorthand notation (cf. [49], [51, p. 367]) with the fact that the LMI relaxation
⎧     ⎫
⎨ I I ⎬ of [51] is exact for a single full block [51, Th. 5.3]. This yields
SN := Z ∈ Rn×k | N 0 . the following result.
⎩ Z Z ⎭ Corollary 12: Let M, N ∈ R(k+n)×(k+n) be symmetric ma-
trices, partitioned as in (12). Assume that N is nonsingular,
Theorem 11 (Strict matrix S-lemma): Let M and N by sym- N11  0, and N22 < 0. Then, we have that
metric matrices in R(k+n)×(k+n) . Assume that SN is bounded
   
and that there exists some matrix Z̄ ∈ Rn×k satisfying (14). I I
Then, we have that M > 0 for all Z ∈ SN
Z Z
   
I I if and only if there exists α  0 such that M − αN > 0.
M > 0 for all Z ∈ SN (15)
Z Z Proof: We will show that the assumptions on N imply the
assumptions of Theorem 11. First of all, N22 < 0 implies that
if and only if there exists α  0 such that M − αN > 0.
SN is bounded. Second, the nonsingularity of N and N22
Proof: The “if” part is clear, so we focus on proving the “only −1 
implies that the Schur complement N11 − N12 N22 N12 is non-
if” part. Suppose that (15) holds. We claim that there exists an
singular, and since N11  0 and N22 < 0, we have N11 −
 > 0 such that −1  −1 
N12 N22 N12 > 0. Thus, the matrix Z̄ := −N22 N12 satisfies
   
I I the Slater condition (14). The result now follows from Theorem
(M − I) > 0 for all Z ∈ SN . (16) 11. 
Z Z
It turns out that we can even state Theorem 11 without the
Suppose that this is not the case. Then, there exists a sequence boundedness assumption if some more structure on the matrices
{i } with i → 0 as i → ∞ with the property that for each i, M and N is given. In fact, we have the following result.
there exists Zi ∈ SN such that Theorem 13: Let M, N ∈ R(k+n)×(k+n) be symmetric ma-
    trices, partitioned as in (12). Assume that M22  0, N22  0,
I I and ker N22 ⊆ ker N12 . Suppose that there exists some matrix
(M − i I) > 0. Z̄ ∈ Rn×k satisfying (14). Then, we have that
Zi Zi
   
Since SN is bounded, the sequence {Zi } is bounded. As such, I I
by the Bolzano–Weierstrass theorem, it contains a converging M > 0 for all Z ∈ SN (17)
Z Z
subsequence with limit, say, Z ∗ . We conclude that
    if and only if there exist α  0 and β > 0 such that
I I  
∗
M > 0. βI 0
Z Z∗ M − αN  .
0 0
Note that SN is closed, and thus, Z ∗ ∈ SN . Since (15) holds,
we arrive at a contradiction. Therefore, we conclude that there Proof: The “if” part is clear, so we focus on proving the “only
exists an  > 0 such that (16) holds. In particular, this implies if” statement. Suppose that (17) holds. We will first prove that
the existence of  > 0 such that ker N22 ⊆ ker M22 and ker N22 ⊆ ker M12 . Let Z ∈ SN and
    Ẑ ∈ Rn×k be such that N22 Ẑ = 0. By the hypothesis ker N22 ⊆
I I ker N12 we have Z + γ Ẑ ∈ SN for any γ ∈ R. Thus, we obtain
(M − I)  0 for all Z ∈ SN .
Z Z    
I I
Now, by Theorem 9, there exists an α  0 such that M + γ(M12 Ẑ + (M12 Ẑ) ) + γ 2 Ẑ  M22 Ẑ > 0.
Z Z
(M − I) − αN  0. (18)
This implies that M22 Ẑ = 0. Indeed, recall that M22  0. Thus,
We conclude that M − αN > 0, which proves the theorem.  if M22 Ẑ = 0, then there exists a sufficiently large γ such that
At this point, it is worthwhile to point out a relation between (18) is violated. Similarly, we conclude that M12 Ẑ = 0. There-
Theorem 11 and the literature on LMI relaxations in robust fore, we have shown that
control (see [51]–[53]). In fact, we can derive a type of matrix
S-lemma from the general theory in [51]. As it turns out, this ker N22 ⊆ ker M22 and ker N22 ⊆ ker M12 . (19)

Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on December 18,2023 at 13:08:25 UTC from IEEE Xplore. Restrictions apply.
 168 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 67, NO. 1, JANUARY 2022

Subsequently, we claim that there exists a β > 0 such that IV. DATA-DRIVEN STABILIZATION REVISITED
      
I βI 0 I In this section, we apply the theory from Section III to
M− > 0 for all Z ∈ SN . (20) data-driven stabilization, i.e., to Problems 1 and 2 defined in
Z 0 0 Z
Section II. To this end, for given P = P  > 0 and K, we define
If this claim is not true, then there exists a sequence {βi } such the partitioned matrices
that βi → 0 and for all i there exists Zi ∈ SN such that ⎡ ⎤
  P 0 0
       M11 M12
I βi I 0 I M= 
:= ⎣ 0 −P −P K  ⎦ (22)
M− > 0. (21) M12 M22 
Zi 0 0 Zi 0 −KP −KP K
 
N11 N12
Define V := {Z ∈ Rn×k | N22 Z = 0}. Write Zi as Zi = N= 
N12 N22
Zi1 + Zi2 , where Zi1 ∈ V ⊥ and Zi2 ∈ V. By the hypothesis
ker N22 ⊆ ker N12 , we see that Zi1 ∈ SN . Next, we claim that ⎡ ⎤ ⎡I X ⎤
I X+  +
the sequence {Zi1 } is bounded. We will prove this claim by Φ11 Φ12 ⎣
:= ⎣ 0 −X− ⎦  −0 −X− ⎦ . (23)
contradiction. Thus, suppose that {Zi1 } is unbounded. Clearly, Φ12 Φ22 0 −U
0 −U− −
the sequence
 1  Recall from Section II that data-driven stabilization entails
Zi
deciding whether (9) holds for all (A, B) satisfying (8). In terms
Zi1 
of the matrices M and N as defined above, we thus have to decide
is bounded. By the Bolzano–Weierstrass theorem, it thus has a whether
convergent subsequence with limit, say Z∗ . Note that        
1  I I I I
N11 + N12 Zi1 + (N12 Zi1 ) + (Zi1 ) N22 Zi1  0. M > 0 ∀Z ∈ R(n+m)×n with N  0.
Zi1 
2 Z Z Z Z
(24)
By taking the limit along the subsequence as i → ∞, we get
Here, Z is given by
Z∗ N22 Z∗  0. Using the fact that N22  0, we conclude that
Z1
 
Z∗ ∈ V. Since Zi1 ∈ V ⊥ for all i, also Zi1  ∈ V ⊥ , and thus, A
i Z := .
Z∗ ∈ V ⊥ . Therefore, we conclude that both Z∗ ∈ V and Z∗ ∈ B
Z1
V ⊥ , i.e., Z∗ = 0. This is a contradiction since Zi1  has norm
i The idea is now to apply Theorem 13. To this end, we have to
1 for all i. We conclude that the sequence {Zi1 } is bounded. verify its assumptions. In particular, we will check that M22  0,
It, thus, contains a convergent subsequence with limit, say Z∞ . N22  0 and ker N22 ⊆ ker N12 . Note that
Note that SN is closed, and thus, Z∞ ∈ SN . By (19) and (21),
       
we conclude that I I X− X−
       M22 = − P  0, N22 = Φ22 0
I βi I 0 I K K U− U−
M− >0
Zi1 0 0 Zi1
because P > 0 and Φ22 < 0. Since Φ22 is nonsingular, we also
for all i. We take the limit as i → ∞, which yields see that
     
I I X−
M > 0. ker N22 = ker
Z∞ Z∞ U−
As Z∞ ∈ SN this contradicts (17). As such, we conclude that    
X−
there exists β > 0 such that (20) holds. In particular, there exists ker N12 = ker (Φ12 + X+ Φ22 )
β > 0 such that U−
      
I βI 0 I and thus ker N22 ⊆ ker N12 . We conclude that the assumptions
M−  0 for all Z ∈ SN . of Theorem 13 are satisfied. We assume that the generalized
Z 0 0 Z
Slater condition (14) holds for N in (23). Then, Theorem 13
The theorem now follows by application of Theorem 9.  asserts that (24) holds if and only if there exist scalars α  0

⎡ ⎤ ⎡ ⎤ ⎡ ⎤
P − βI 0 0 0 I X+   I X+
⎢ 0 −P −L 0⎥ ⎢0 −X− ⎥ Φ12 ⎢ −X− ⎥
⎢ ⎥ ⎢ ⎥ Φ11 ⎢0 ⎥
⎢ ⎥ − α⎢ ⎥ ⎢ ⎥  0. (FS)
⎣ 0 −L 0 L⎦ ⎣0 −U− ⎦ Φ
12 Φ22 ⎣0 −U− ⎦
0 0 L P 0 0 0 0

Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on December 18,2023 at 13:08:25 UTC from IEEE Xplore. Restrictions apply.
H. J. VAN WAARDE et al.: FROM NOISY DATA TO FEEDBACK CONTROLLERS: NONCONSERVATIVE DESIGN VIA A MATRIX S-LEMMA 169

and β > 0 such that Th. 2] and [39, Th. 4]. Indeed, in the case of noisy data,
  it was recognized that certain linear constraints [39, eq.
βI 0 (3)] defining these closed-loop systems were difficult to
M − αN  . (25)
0 0 incorporate in the control design.5 Our design procedure
does not suffer from the above problem. In fact, the
From a design point of view, the matrices P and K that appear in constraint [39, eq. (3)] is automatically incorporated in
M are not given. However, the idea is now to compute matrices our control design approach.
P , K and scalars α and β such that (25) holds. In fact, by the 3) The variables P, L, α, and β are independent of the time
above discussion, the data (U− , X) are informative for quadratic horizon T of the experiment. In fact, note that P ∈ Rn×n ,
stabilization if and only if there exists an n × n matrix P = L ∈ Rm×n , and α, β ∈ R. Also, the LMI (FS) is of di-
P  > 0, a K ∈ Rm×n , and two scalars α  0 and β > 0 such mension (3n + m) × (3n + m) and thus independent of
that (25) holds. We note that (25) (in particular, M ) is not linear in T . As such, our approach fundamentally differs from the
P and K. Nonetheless, by a rather standard change of variables design methods in [37] and [39], where certain decision
and a Schur complement argument, we can transform (25) into variables have dimension T × n; cf. [37, Th. 6] and [39,
an LMI. We summarize our progress in the following theorem, Cor. 6]. We believe that our T -independent design method
which is the main result of this section. will play a crucial role in control design from larger
Theorem 14: Assume that the generalized Slater condition datasets. We note that the collection of big datasets is often
(14) holds for N in (23) and some Z̄ ∈ R(n+m)×n . Then, the unavoidable, for example, because the signal-to-noise
data (U− , X) are informative for quadratic stabilization if and ratio is small, or because the data-generating system is
only if there exists an n × n matrix P = P  > 0, an L ∈ Rm×n , large scale.
and scalars α  0 and β > 0 satisfying (FS) shown at the bottom Remark 15: We note that under the extra assumption
of the previous page.  
Moreover, if P and L satisfy (FS), then K := LP −1 is a X−
rank =n+m (26)
stabilizing feedback gain for all (A, B) ∈ Σ. U−
Proof: To prove the “if” statement, suppose that there exist P ,
L, α, and β satisfying (FS). Define K := LP −1 . By computing it is possible to prove a variant Theorem 14, in which the
the Schur complement of (FS) with respect to its fourth diagonal nonstrict inequality is replaced by a strict inequality, and the
block, we obtain (25). As such, (24) holds. We conclude that term −βI is removed. This can be done by invoking Theo-
the data (U− , X) are informative for quadratic stabilization, and rem 11, which is possible since (26) implies that the set Σ is
K = LP −1 is indeed a stabilizing controller for all (A, B) ∈ Σ. bounded. The reason is that the coefficient matrix N22 defining
Conversely, to prove the “only if” statement, suppose that the the quadratic term in (8) is negative definite if (26) holds.
data (U− , X) are informative for quadratic stabilization. This Here, we chose to state and prove Theorem 14 in the slightly
means that there exist P = P  > 0 and K such that (24) holds. more general setting without assuming (26). In the discussion
By Theorem 13, there exist α  0 and β > 0 satisfying (25). preceding Theorem 14, we have verified the assumptions of
Finally, by defining L := KP and using a Schur complement Theorem 13 for M and N in (22) and (23). In particular,
argument, we conclude that (FS) is feasible.  this implies that the subspace inclusions (19) hold, and thus,
Theorem 14 provides a powerful necessary and sufficient ker[X− U− ] ⊆ ker[I K  ], equivalently
condition under which quadratically stabilizing controllers can    
I X−
be obtained from noisy data. The assumption (14) puts a mild im ⊆ im . (27)
condition on the data matrices appearing in (23). It is satisfied K U−
whenever N has at least n positive eigenvalues, a condition that Therefore, any controller K that stabilizes the systems in Σ is
is simple to verify from given data. So far, this condition was necessarily of the form (27). This generalizes [3, Lem. 15] to
satisfied in all of our numerical experiments; see Section VI for the case of noisy data.
more details.4 Theorem 14 leads to an effective design procedure
for obtaining stabilizing controllers directly from data. Indeed, V. INCLUSION OF PERFORMANCE SPECIFICATIONS
the approach entails solving the LMI (FS) for P, L, α, and β and
computing a controller as K = LP −1 . We now discuss some of In this section, we extend our data-driven stabilization result
the features of our control design procedure. by including different performance specifications. In particular,
1) First of all, we note that the procedure is nonconservative we will treat the H2 and H∞ control problems, thereby illustrat-
since Theorem 14 provides a necessary and sufficient con- ing the general applicability of the theory in Section III.
dition for obtaining quadratically stabilizing controllers
from data. A. H2 Control
2) We believe that our approach based on the set Σ of As before, consider the unknown system (1). We associate to
open-loop systems provides a valuable alternative to the (1) a performance output
data-based closed-loop system parameterizations of [37,
z(t) = Cx(t) + Du(t) (28)

4 In addition, we remark that even if the generalized Slater condition does not 5 In fact, it was mentioned in [58] that involving the condition [39, eq. (3)] in
hold, the ‘if’ statement of Theorem 14 remains true. design procedures is still an open problem.
Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on December 18,2023 at 13:08:25 UTC from IEEE Xplore. Restrictions apply.
 170 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 67, NO. 1, JANUARY 2022

where z ∈ Rp , and C and D are known matrices that specify Note that (32) is independent of A and B. In turn, we can write
the performance. For any (A, B) ∈ Σ explaining the data, the (33) as
feedback law u = Kx yields the closed-loop system ⎡ ⎤
⎡ ⎤ Y 0 ⎡ ⎤
I ⎢     ⎥ I
x(t + 1) = (A + BK)x(t) + w(t) ⎣ A ⎦ ⎢ ⎥⎣ A ⎦ > 0.
⎣ 0 − Y (Y − C  C )−1 Y ⎦

z(t) = (C + DK)x(t). (29) B L Y,L Y,L
L B
  
The transfer matrix from w to z of (29) is given by =:M
(34)
G(z) := (C + DK)(zI − (A + BK))−1
Note that inequality (34) is of a form where A and B appear
and its H2 norm is denoted by G(z)H2 . Let γ > 0. It is well on the left and their transposes appear on the right, analogous to
known that A + BK is stable and G(z)H2 < γ if and only if (9). As such, we are in a position to apply Theorem 13. In fact,
there exists a matrix P = P  > 0 such that we derive the following theorem.
P > (A + BK) P (A + BK) + (C + DK) (C + DK) Theorem 17: Assume that the generalized Slater condition
(14) holds for N in (23) and some Z̄ ∈ R(n+m)×n . Then, the
tr P < γ 2 data (U− , X) are informative for H2 control with performance
(30) γ if and only if there exist matrices Y = Y  > 0, Z = Z  and
where tr denotes trace. The data-driven H2 problem entails L, and scalars α  0 and β > 0 satisfying (H2 ).
the computation of a feedback gain K from data such that Moreover, if Y and L satisfy (H2 ), then K := LY −1 is such
G(z)H2 < γ for all (A, B) ∈ Σ. Similar to our results for that A + BK is stable and G(z)H2 < γ for all (A, B) ∈ Σ.
quadratic stabilization, we restrict the attention to a matrix P Proof: Suppose that (H2 ) is feasible and define P := Y −1
that is common for all (A, B). This leads to the following natural and K := LP . The last two inequalities of (H∈ ) imply that
definition. tr P < γ 2 . We now compute the Schur complement of the first
Definition 16: Assume that (U− , X) satisfies (2) for some LMI in (H∈ ) with respect to the diagonal block
W− satisfying Assumption 1. The data (U− , X) are informative  

for H2 control with performance γ if there exist matrices P = Y CY,L
.
P  > 0 and K such that (30) holds for all (A, B) ∈ Σ. CY,L I
With the theory of Section III in place, characterizing infor-
We thereby make use of the fact that this block is nonsingular
mativity for H2 control essentially boils down to massaging the
by the second LMI of (H∈ ). The computation of the Schur
inequalities (30) such that they are amenable to design. To this
complement results in
end, note that the first inequality of (30) is equivalent to  
βI 0
Y − A 
Y,L P AY,L − CY,L CY,L > 0 M − αN  (35)
0 0
where we defined AY,L := AY + BL and CY,L := CY + DL
where M is defined in (34) and N is defined in (23). We thus
with Y := P −1 and L := KY . Using a Schur complement
conclude that inequality (34) is satisfied for all (A, B) ∈ Σ. As
argument, this is equivalent to
such, (33) holds for all (A, B) ∈ Σ. Note that (32) holds by the
 

Y − CY,L CY,L A second LMI of (H∈ ). Therefore, we conclude that (30) holds for
Y,L
>0 (31) all (A, B) ∈ Σ. In other words, the data (U− , X) are informative
AY,L Y
for H2 control with performance γ, and K = LY −1 is a suitable
Now, (31) holds if and only if controller.
Conversely, suppose that the data (U− , X) are informative

Y − CY,L CY,L > 0 (32) for H2 control with performance γ. Then, there exist matrices

P = P  > 0 and K such that (30) holds for all (A, B) ∈ Σ.
Y − AY,L (Y − CY,L CY,L )−1 A
Y,L > 0. (33) Define Y := P −1 , L := KY and Z := P . Clearly, the last two

⎡ ⎤ ⎡ ⎤ ⎡ ⎤
Y − βI 0 0 0 0 I X+ I X+
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ 0 0 0 Y 0 ⎥ ⎢0 −X− ⎥  ⎢0 −X− ⎥  
⎢ ⎥ ⎢ ⎥ Φ11 Φ12 ⎢ ⎥ Y 
CY,L
⎢ 0 0 0 L 0 ⎥ − α ⎢0 −U− ⎥ ⎢0 ⎥
−U− ⎥  0, >0
⎢ ⎥ ⎢ ⎥ Φ Φ22 ⎢ CY,L I
⎢  ⎥ ⎢ ⎥ ⎢ ⎥
⎣ 0 Y L 
Y CY,L ⎦ ⎣0 0 ⎦ 12 ⎣0 0 ⎦
0 0 0 CY,L I 0 0 0 0
 
Z I
 0, tr Z < γ 2 (H2 )
I Y

Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on December 18,2023 at 13:08:25 UTC from IEEE Xplore. Restrictions apply.
H. J. VAN WAARDE et al.: FROM NOISY DATA TO FEEDBACK CONTROLLERS: NONCONSERVATIVE DESIGN VIA A MATRIX S-LEMMA 171

inequalities of (H∈ ) are satisfied by definition of Z. In addition, Definition 19: Assume that (U− , X) satisfies (2) for some
we know that (32) and (33) hold for all (A, B) ∈ Σ. By (32), W− satisfying Assumption 1. The data (U− , X) are informative
the second LMI of (H∈ ) is satisfied. To prove that the first LMI for H∞ control with performance γ if there exist matrices P =
of (H∈ ) also holds, we want to apply Theorem 13. Note that P  > 0 and K such that (36) and (37) hold for all (A, B) ∈ Σ.
we have already verified the assumptions of this theorem for the By pre- and postmultiplication of (36) by P −1 , we obtain
matrix N in (23); see the discussion preceding Theorem 14. In  −1
1
addition, we note that Y − A Y,L Y − I AY,L − CY,L
CY,L > 0
    γ2
Y  −1 Y 1
M22 = − (Y − CY,L CY,L ) 0 Y − I>0
L L γ2

since Y − CY,L CY,L > 0. Hence, Theorem 13 is applicable. We where the matrices Y := P −1 , L := KY , AY,L := AY + BL,
conclude that there exist α  0 and β > 0 such that (35) holds. and CY,L := CY + DL are defined as in the H2 problem. Note
Using a Schur complement argument, we see that Y , L, α, and that the first of these inequalities can again be written in the—by
β satisfy the first LMI of (H∈ ). Thus, (H∈ ) is feasible, which now familiar—form
 ⎡ ⎤
proves the theorem. ⎡ ⎤ Y − C  CY,L 0 ⎡ ⎤
Remark 18: If we know a priori that the noise w is contained I ⎢
Y,L
    ⎥ I
in a subspace, say im E, then this information can easily be ⎣ A ⎦ ⎢ Y Y ⎥ ⎣ A ⎦ > 0
⎣ 0 − Z ⎦

exploited in the H2 controller design. In fact, we only need to B L L B
replace the LMI involving Z by
  where Z := (Y − γ12 I)−1 . We thus have the following theorem.
Z E Theorem 20: Assume that the generalized Slater condition
 0.
E Y (14) holds for N in (23) and some Z̄ ∈ R(n+m)×n . Then, the
data (U− , X) are informative for H∞ control with performance
We recall that prior knowledge of w ∈ im E, if available, can γ if and only if there exist matrices Y = Y  > 0 and L, and
also be captured by our noise model (see Remark 2). A natural scalars α  0 and β > 0 satisfying (H∞ ).
choice is thus to use E both in the noise model (3) and in the Moreover, if Y and L satisfy (H∞ ), then K := LY −1 is such
LMI (H∈ ). However, we remark that this is not necessary: the that A + BK is stable and G(z)H∞ < γ for all (A, B) ∈ Σ.
noise in the experiment may come from a different subspace The proof of Theorem 20 is based on Theorem 13. It follows
than the disturbances that are attenuated by the H2 controller. similar steps as the proof of Theorem 17 and is, therefore, not
reported here.
B. H∞ Control
In this section, we will turn our attention to the H∞ con- VI. EXAMPLES
trol problem. As before, consider system (1) with performance In this section, we illustrate our theoretical results by exam-
output (28). For any (A, B) ∈ Σ, the feedback u = Kx yields ples and numerical simulations.
the system (29) with transfer matrix from w to z given by
G(z). We will denote the H∞ norm of G(z) by G(z)H∞ . Let
A. Stabilization Using Bounds on the Noise Samples
γ > 0. By [59, Th. 4.6.6(iii)], the matrix A + BK is stable and
G(z)H∞ < γ if and only if there exists a matrix P = P  > 0 Consider an unstable system of the form (1) with As and Bs
such that given by
 −1 ⎡ ⎤
 −1 1  0.850 −0.038 −0.380
P − AK P − 2 I AK − CK CK > 0 (36) ⎢ ⎥
γ As = ⎣ 0.735 0.815 1.594 ⎦
1 −0.664 0.697 −0.064
P −1 − I>0 (37)
γ2 ⎡ ⎤
1.431 0.705
where we have defined AK := A + BK and CK := C + DK. ⎢ ⎥
Bs = ⎣1.620 −1.129⎦ .
We now have the following definition of informativity for H∞
control. 0.913 0.369

⎡ ⎤ ⎡ ⎤ ⎡ ⎤

Y − βI 0 0 0 CY,L I X+ I X+
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ 0 0 0 Y 0 ⎥ ⎢0 −X− ⎥  ⎢0 −X− ⎥
⎢ ⎥ ⎢ ⎥ Φ11 Φ12 ⎢ ⎥ 1
⎢ 0 0 ⎥ ⎢ −U− ⎥ ⎢0 −U− ⎥
⎢ 0 0 L ⎥ − α ⎢0 ⎥ Φ Φ22 ⎢ ⎥  0, Y − γ 2 I > 0. (H∞ )
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎣ 0 Y L Y − γ12 I 0 ⎦ ⎣0 0 ⎦ 12 ⎣0 0 ⎦
CY,L 0 0 0 I 0 0 0 0

Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on December 18,2023 at 13:08:25 UTC from IEEE Xplore. Restrictions apply.
 172 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 67, NO. 1, JANUARY 2022

In this example, we assume that the noise samples w(t) are

bounded in norm as w(t)22   for all t. As explained in
Section II, we can capture this prior knowledge using the noise
model (3) with Φ11 = T I, Φ12 = 0, and Φ22 − I. We pick a
time horizon of T = 20 and draw the entries of the inputs and
initial state randomly from a Gaussian distribution with zero
mean and unit variance. The noise samples are drawn uniformly
at random from the ball {w ∈ R3 | w22  }. We aim at
constructing stabilizing controllers from the input/state data for
various values of . In particular, we investigate six different
noise levels:  ∈ {0.5, 1, 1.5, 2, 2.2, 2.4}. For each noise level,
we generate 100 datasets using the method described above. We
check the generalized Slater condition (14) by verifying that N
in (23) has three positive eigenvalues; this turns out to be true for
Fig. 1. Achieved H2 performance of the true system in feedback with a
all 600 datasets. For each noise level, we record the percentage data-based controller (blue) and the optimal (model-based) performance
of datasets from which a stabilizing controller was found for of the true system (red).
(As , Bs ) using the formulation (FS). We display the results in
the following table:
750 input and state samples of (1). The entries of the inputs and
initial state were drawn randomly from a Gaussian distribution
For  = 0.5, we find a stabilizing controller in all 100 cases. with zero mean and unit variance. Also, the noise samples were
When the noise level increases, the percentage of datasets for drawn randomly from a Gaussian distribution, with zero mean
which the LMI (FS) is feasible decreases. The interpretation is and variance σ 2 with σ = 0.005. In this example, we assume
that by increasing the noise we enlarge the set of explaining knowledge of a bound on the energy of the noise as
systems Σ. It thus becomes harder to simultaneously stabilize
W− W−  1.35 T σ 2 I. (38)
the systems in Σ. Nonetheless, even for the larger noise level
of  = 2.4, we find a stabilizing controller in 73 out of the 100 We verified that this bound is satisfied for the generated noise
datasets. sequence. In addition, we verified that the matrix N in (23) has
six positive eigenvalues; thus, the generalized Slater condition
B. H2 Control of a Fighter Aircraft (14) holds.
We consider a state-space model of a fighter aircraft [59, Ex. Next, we want to compute an H2 controller for the unknown
10.1.2]. In particular, we discretize the model of [59] using a system using the generated data. We do so by minimizing γ
sampling time of 0.01, which results in the (unstable) system of subject to (H∈ ). This is a semidefinite program that we solve in
the form (1) with As and Bs given by MATLAB, using Yalmip [60] with Mosek as an LMI solver. The
⎡ ⎤ obtained controller K stabilizes the original system (As , Bs ).
1.000 −0.374 −0.190 −0.321 0.056 −0.026 In addition, the system, in feedback with K, has an H2 norm of
⎢ ⎥
⎢0.000 0.982 0.010 −0.000 −0.003 0.001 ⎥ γs , where γs2 = 1.007. We note that this is almost identical to
⎢ ⎥
⎢0.000 0.115 0.975 −0.000 −0.269 0.191 ⎥ the smallest possible H2 norm of 1.000.
⎢ ⎥
⎢0.000 0.001 0.010 1.000 −0.001 0.001 ⎥
Subsequently, we repeat the above experiment using only a
⎢ ⎥
⎢ ⎥ part of our dataset. In particular, we compute an H2 controller
⎣0.000 0.000 0.000 0.000 0.741 0.000 ⎦ via the semidefinite program as before, using only the first i
0.000 0.000 0.000 0.000 0.000 0.741 samples of X+ , X− , and U− for i = 50 100, . . . , 750. We display
  the results in Fig. 1.
0.007 0.000 −0.043 0.000 0.259 0.000 In each of the cases, a stabilizing controller was found from
−0.003 0.000 0.030 0.000 0.000 0.259 data. However, the performance of these controllers when ap-
plied to the true system varies and is quite poor for i < 500.
respectively. We consider the performance output as in (28) with
 Starting from i = 500 and onward, the performance is close to
C= 0 0 0 0 0 1 the optimal performance of the true system.
Next, we investigate what happens when we increase the
and D = 0. First, we look for the smallest γ such that (30) is variance σ 2 of the noise. First, we take σ = 0.05. We again
feasible for (As , Bs ). This minimum value of γ is 1.000 and can generate 750 data samples and assume the same bound on the
be regarded as a benchmark: no data-driven method can perform noise. The H2 controller achieves a performance of γs2 = 1.146
better than the model-based solution using full knowledge of when interconnected to the true system. Increasing the variance
(As , Bs ). of the noise has the effect that the set Σ of explaining systems
Of course, our goal is not to use the knowledge of (As , Bs ) but becomes larger. As such, it is more difficult to control all systems
to seek a data-driven solution instead. Therefore, we collect T = in Σ resulting in a slightly larger γs . This behavior becomes

Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on December 18,2023 at 13:08:25 UTC from IEEE Xplore. Restrictions apply.
H. J. VAN WAARDE et al.: FROM NOISY DATA TO FEEDBACK CONTROLLERS: NONCONSERVATIVE DESIGN VIA A MATRIX S-LEMMA 173

even more apparent when increasing the variance of the noise We start by applying Theorem 14 to the data in this exam-
to σ = 0.5. In this case, we obtain a controller that yields a ple. Note that X− = [0 0 1] and X+ = [0 1 0]. It can be
performance of γs2 = 3.579. Increasing σ even more to σ = 1 easily verified that (P, L, α, β) = (0.9, −1.35, 1.1, 0.18) is a
results in infeasibility of the LMI’s (H∈ ) for any γ; the set of solution to (FS). In addition, the Slater condition is satisfied
explaining systems has become too large for a quadratically in this example. As such, we conclude by Theorem 14 that the
stabilizing controller to exist. controller K = L/P = −1.5 is stabilizing for all (A, B) ∈ Σ.
We remark that the size of the set Σ does not only depend In particular, we see that the true closed-loop system matrix
on the variance of the noise, but also on the available bound on −0.5 is stable.
the noise. Throughout this example, we have used the bound We will now investigate the design method of [37, Th. 6]. This
(38). However, if we reconsider the case of σ = 0.5 with the approach involves finding a matrix Q and a scalar α > 0 such
tighter bound W− W−  1.22 T σ 2 I (which is also satisfied in that X− Q is symmetric and
this example), we obtain a controller with better performance    

γs2 = 2.706. This illustrates the simple fact that data-driven X− Q − αX+ X+ X+ Q I Q
> 0, > 0 (39)
controllers not only depend on the particular design strategy, Q X+ 
X− Q Q X− Q
but also on the prior knowledge on the noise.
We conclude the example with a remark on the dimension of α2
>γ (40)
the variables involved in the formulation (H∈ ). The symmetric 4 + 2α
matrices Y and Z both have 21 free variables. The matrix L where we recall that γ = 1 in this example. We will now show
contains 12 variables, and α and β are both scalar variables. that (39) and (40) do not have a solution (Q, α), and thus, the
Thus, the total number of variables is 56. The size of the largest design procedure from [37, Th. 6] cannot find a controller that is
LMI in (H∈ ) is 21 × 21. We emphasize that our approach is guaranteed to stabilize the true
√ system. To see this, note that (40)
based directly on the set Σ of open-loop systems and avoids the and α > 0 imply α > 1 + 5. We write Q = [q1 q2 q3 ] and
parameterization of closed-loop systems, as employed in [37] note that by the upper left block of the first matrix in (39), we have
and [39]. Such parameterizations involve decision variables of q3 > α and, thus, q3 > 1. Now, by taking the Schur complement
dimension T × n, which would result in at least 4500 variables of the second matrix in (39) with respect to the upper left block,
in this example. we obtain q3 > q12 + q22 + q32 . However, this inequality cannot
be satisfied since q3 > 1. As such, we conclude that (39) and
C. Comparison With Related Results (40) do not have a solution.
As we have mentioned before, one of the advantages of our Next, we turn our attention to the design procedure of [39,
approach compared to existing work is that our LMI condition Rem. 7]. For Sw = 0, this procedure boils down to finding a
provides necessary and sufficient conditions for data-driven solution (Y, M ) to
quadratic stabilization. The purpose of this example is to demon- ⎡ ⎤
−Y 0 M  X+ 
M
strate that our results can thus lead to stabilizing controllers even ⎢ 0
in situations where the results from [37] and [39] cannot. ⎢ Qw 1 0 ⎥ ⎥
⎢ ⎥<0 (41)
Consider the system (1), where As = 1 and Bs = 1. Suppose ⎣ X+ M 1 −Y 0 ⎦
that T = 3 and the noise matrix is given by M 0 0 −1
−Rw

W− = 12 12 12 . X− M = Y (42)
We collect the data samples where we recall that Qw = −1 and Rw = I in this example.
 We will now show that (41) and (42) do not have a solution
X= 0 0 1 0 (Y, M ). To see this, note that (42) implies M = [m1 m2 Y]
 with m1 , m2 ∈ R. The negative definiteness of the submatrix of
U− = − 12 1
2 − 32 . (41) consisting of the second and third rows and columns imply
Y > 1. Moreover, by inspection of the first and fourth row and
Throughout the example, we assume that we have access to
column block of (41), we see that −Y + M  M < 0 and, thus,
the noise bound W− W−  1. Note that this bound is indeed
−Y + m21 + m22 + Y 2 < 0. This inequality, however, cannot be
satisfied, and that it can be captured using Assumption 1 by the
satisfied as Y > 1. As such, we see that (41) and (42) do not have
choices Φ11 = 1, Φ12 = 0, and Φ22 = −I. We also note that
a solution (Y, M ).
by Remark 1, this noise bound can be captured by [39, As. 3]
We conclude that Theorem 14 can be successfully applied
with the choices Qw = −1, Sw = 0, and Rw = I. Finally, we
even in situations, in which the design procedures of [39] and
note that the noise bound can be captured by [37, Asm. 5] with
[37] do not lead to controllers that are guaranteed to stabilize
the choice γ = 1. As such, we can compare the design methods
the true system.
reported in Theorem 14 of this article with the approaches in [39,
Cor. 6, Rem. 7] and [37, Th. 6].6
this comparison. In fact, our results (as well as those in [39]) are applicable
note that [37] studies stabilization in the setting that w represents a
6 We to general bounded disturbances, hence also to disturbances resulting from
bounded nonlinearity. The interpretation of w, however, is not important for bounded nonlinearities.

Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on December 18,2023 at 13:08:25 UTC from IEEE Xplore. Restrictions apply.
 174 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 67, NO. 1, JANUARY 2022

VII. DISCUSSION AND CONCLUSION REFERENCES

We have studied the problem of obtaining feedback con- [1] V. A. Yakubovich, “S-procedure in nonlinear control theory,” Vestnik
trollers from noisy data. The essence of our approach has been Leningrad Univ. Math., vol. 4, pp. 73–93, 1977.
[2] I. Pòlik and T. Terlaky, “A survey of the S-lemma,” SIAM Rev., vol. 49,
to formulate data-driven control as the problem of determining no. 3, pp. 371–418, 2007.
when one QMI implies another one. To get a grip on this funda- [3] H. J. van Waarde, J. Eising, H. L. Trentelman, and M. K. Camlibel,
“Data informativity: A new perspective on data-driven analysis and
mental question, we have generalized the classical S-lemma [1] control,” IEEE Trans. Autom. Control, vol. 65, no. 11, pp. 4753–4768,
to matrix variables. The implication involving quadratic matrix Nov. 2020.
inequalities is thereby equivalent to an LMI in a scalar variable. [4] Z. Hou and Z. Wang, “From model-based control to data-driven control:
Survey, classification and perspective,” Inf. Sci., vol. 235, pp. 3–35, 2013.
We have established several versions of the matrix S-lemma, for [5] R. E. Skelton and G. Shi, “The data-based LQG control problem,” in Proc.
both strict and nonstrict inequalities. These matrix S-lemmas IEEE Conf. Decis. Control, Dec. 1994, pp. 1447–1452.
are interesting in their own right and generalize existing S- [6] K. Furuta and M. Wongsaisuwan, “Discrete-time LQG dynamic con-
troller design using plant Markov parameters,” Automatica, vol. 31, no. 9,
lemmas [2] as well as a theorem involving quadratic matrix pp. 1317–1324, 1995.
inequalities [46]. [7] G. Shi and R. Skelton, “Markov data-based LQG control,” ASME J. Dyn.
We have followed up by applying our matrix S-lemma to Syst., Meas., Control, vol. 122, no. 3, pp. 551–559, 1998.
[8] W. Favoreel, B. De Moor, P. Van Overschee, and M. Gevers, “Model-free
data-driven control. In particular, we have given necessary and subspace-based LQG-design,” in Proc. Amer. Control Conf., Jun. 1999,
sufficient conditions under which stabilizing, H2 , and H∞ con- vol. 5, pp. 3372–3376.
trollers can be obtained from noisy data. Our control design re- [9] W. Aangenent, D. Kostic, B. de Jager, R. van de Molengraft, and
M. Steinbuch, “Data-based optimal control,” in Proc. Amer. Control Conf.,
volves around data-guided LMIs, which can be solved efficiently Jun. 2005, pp. 1460–1465.
using modern LMI solvers. In addition to being nonconservative, [10] B. Pang, T. Bian, and Z. Jiang, “Data-driven finite-horizon optimal control
an attractive feature of our design procedure is that decision for linear time-varying discrete-time systems,” in Proc. IEEE Conf. Decis.
Control, Dec. 2018, pp. 861–866.
variables are independent of the time horizon of the experiment. [11] G. R. Gonçalves da Silva, A. S. Bazanella, C. Lorenzi, and L. Campestrini,
So far, we have only applied the matrix S-lemma involving a “Data-driven LQR control design,” IEEE Control Syst. Lett., vol. 3, no. 1,
strict inequality (see Theorems 11 and 13) to data-driven control. pp. 180–185, Jan. 2019.
[12] G. Baggio, V. Katewa, and F. Pasqualetti, “Data-driven minimum-energy
However, we are convinced that also the matrix S-lemma with controls for linear systems,” IEEE Control Syst. Lett., vol. 3, no. 3, pp. 589–
nonstrict inequalities (see Theorem 9) will find applications, 594, Jul. 2019.
for example, in the verification of dissipativity properties from [13] S. Mukherjee, H. Bai, and A. Chakrabortty, “On model-free reinforce-
ment learning of reduced-order optimal control for singularly perturbed
data [58]. systems,” in Proc. IEEE Conf. Decis. Control, Dec. 2018, pp. 5288–5293.
The noise model that we have employed is flexible and can [14] S. Alemzadeh and M. Mesbahi, “Distributed q-learning for dynamically
describe, e.g., constant disturbances, energy bounded noise, and decoupled systems,” in Proc. Amer. Control Conf., Jul. 2019, pp. 772–777.
[15] S. Dean, H. Mania, N. Matni, B. Recht, and S. Tu, “On the sample
norm bounds on noise samples. If one is only interested in the complexity of the linear quadratic regulator,” Found. Comput. Math.,
latter, however, we expect that more specific control techniques vol. 20, pp. 633–679, Aug. 2019.
are possible. In fact, analogous to (8), we can write the inequality [16] H. J. van Waarde and M. Mesbahi, “Data-driven parameterizations of
w(t) w(t)   as suboptimal LQR and H2 controllers,” 2020, arXiv:1912.07671.
[17] L. H. Keel and S. P. Bhattacharyya, “Controller synthesis free of analytical
⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤
 I x(t + 1)  I
models: Three term controllers,” IEEE Trans. Autom. Control, vol. 53,
I I x(t + 1)  no. 6, pp. 1353–1369, Jul. 2008.
⎢ ⎥ ⎢ ⎥ I 0 ⎢ ⎥ ⎢ ⎥
⎣0 −x(t) ⎦ ⎣A ⎦  0.
[18] M. Fliess and C. Join, “Model-free control,” Int. J. Control, vol. 86, no. 12,
⎣A ⎦ ⎣0 −x(t) ⎦
0 −I pp. 2228–2252, 2013.
B 0 −u(t) 0 −u(t) B [19] W. Favoreel, B. De Moor, and M. Gevers, “SPC: Subspace predictive
   control,” IFAC Proc. Vol., vol. 32, no. 2, pp. 4004–4009, 1999.
:=Nt [20] J. R. Salvador, D. Muñoz de la Peña, T. Alamo, and A. Bemporad,
“Data-based predictive control via direct weight optimization,” IFAC-
In the spirit of the S-procedure, one could thus design a stabi- PapersOnLine, vol. 51, no. 20, pp. 356–361, 2018.
lizing controller by computing7 matrices P = P  > 0 and K, [21] L. Huang, J. Coulson, J. Lygeros, and F. Dörfler, “Data-enabled predictive
control for grid-connected power converters,” in Proc. IEEE Conf. Decis.
and multiple nonnegative scalars α0 , α1 , . . . , αT −1 such that Control, 2019, pp. 8130–8135.
T −1 [22] D. Alpago, F. Dörfler, and J. Lygeros, “Extended Kalman filter for
 data-enabled predictive control,” IEEE Control Syst. Lett., vol. 4, no. 4,
M− α t Nt > 0 pp. 994–999, Oct. 2020.
t=0 [23] L. Hewing, K. P. Wabersich, M. Menner, and M. N. Zeilinger, “Learning-
based model predictive control: Toward safe learning in control,” Annu.
with M given by (22). We will consider norm-bounded noise Rev. Control, Robot., Auton. Syst., vol. 3, no. 1, pp. 269–296, 2020.
samples in more detail in future work. [24] P. Tabuada, W. Ma, J. Grizzle, and A. D. Ames, “Data-driven control for
Yet, another idea for future work is to extend the current feedback linearizable single-input systems,” in Proc. IEEE Conf. Decis.
Control, 2017, pp. 6265–6270.
results for state-feedback design to data-driven dynamic output [25] T. Dai and M. Sznaier, “A moments based approach to designing MIMO
feedback design. Specifically, it would be interesting to see data driven controllers for switched systems,” in Proc. IEEE Conf. Decis.
whether the matrix S-lemmas can be applied to obtain dynamic Control, 2018, pp. 5652–5657.
[26] P. Tabuada and L. Fraile, “Data-driven stabilization of SISO feedback
output feedback controllers from a finite set of noisy input– linearizable systems,” 2020, arXiv:2003.14240.
output samples. [27] A. Bisoffi, C. De Persis, and P. Tesi, “Data-based stabilization of unknown
bilinear systems with guaranteed basin of attraction,” Syst. Control Lett.,
vol. 145, 2020, Art. no. 104788.
7 This procedure is likely to be conservative, however, since the classical S-
[28] M. Guo, C. De Persis, and P. Tesi, “Learning control for polynomial
lemma is in general conservative for more than two quadratic functions [2]. systems using sum of squares,” 2020, arXiv:2004.00850.

Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on December 18,2023 at 13:08:25 UTC from IEEE Xplore. Restrictions apply.
H. J. VAN WAARDE et al.: FROM NOISY DATA TO FEEDBACK CONTROLLERS: NONCONSERVATIVE DESIGN VIA A MATRIX S-LEMMA 175

[29] S. J. Bradtke, “Reinforcement learning applied to linear quadratic regula- [57] Z.-Z. Yan and J.-H. Guo, “Some equivalent results with Yakubovich’s
tion,” in Proc. Int. Conf. Inf. Process. Syst., 1993, pp. 295–302. S-lemma,” SIAM J. Control Optim., vol. 48, no. 7, pp. 4474–4480, 2010.
[30] H. Hjalmarsson, M. Gevers, S. Gunnarsson, and O. Lequin, “Iterative [58] A. Koch, J. Berberich, and F. Allgöwer, “Verifying dissipativity properties
feedback tuning: Theory and applications,” IEEE Control Syst. Mag., from noise-corrupted input-state data,” 2020, arXiv:2004.07270.
vol. 18, no. 4, pp. 26–41, Aug. 1998. [59] R. Skelton, T. Iwasaki, and D. Grigoriadis, Unified Algebraic Approach to
[31] M. C. Campi, A. Lecchini, and S. M. Savaresi, “Virtual reference feedback Control Design. New York, NY, USA: Taylor & Francis, 1997.
tuning: A direct method for the design of feedback controllers,” Automat- [60] J. Lofberg, “YALMIP: A toolbox for modeling and optimization in MAT-
ica, vol. 38, no. 8, pp. 1337–1346, 2002. LAB,” in Proc. IEEE Int. Conf. Robot. Autom., 2004, pp. 284–289.
[32] J. C. Willems, P. Rapisarda, I. Markovsky, and B. L. M. De Moor, “A note
on persistency of excitation,” Syst. Control Lett., vol. 54, no. 4, pp. 325–
329, 2005.
[33] H. J. van Waarde, C. De Persis, M. K. Camlibel, and P. Tesi, “Willems’
fundamental lemma for state-space systems and its extension to multiple
datasets,” IEEE Control Syst. Lett., vol. 4, no. 3, pp. 602–607, Jul. 2020. Henk J. van Waarde received the master’s
[34] I. Markovsky and P. Rapisarda, “Data-driven simulation and control,” Int. (summa cum laude) and Ph.D. (cum laude) de-
J. Control, vol. 81, no. 12, pp. 1946–1959, 2008. grees in applied mathematics from the Univer-
[35] T. M. Maupong and P. Rapisarda, “Data-driven control: A behavioral sity of Groningen, Groningen, The Netherlands,
approach,” Syst. Control Lett., vol. 101, pp. 37–43, 2017. in 2017 and 2020, respectively.
[36] J. Coulson, J. Lygeros, and F. Dörfler, “Data-enabled predictive control: He was a Visiting Researcher with the Univer-
In the shallows of the DeePC,” in Proc. Eur. Control Conf., Jun. 2019, sity of Washington, Seattle, WA USA, in 2019–
pp. 307–312. 2020. His research interests include data-driven
[37] C. De Persis and P. Tesi, “Formulas for data-driven control: Stabilization, control, networks of dynamical systems, system
optimality, and robustness,” IEEE Trans. Autom. Control, vol. 65, no. 3, identification and identifiability, robust and opti-
pp. 909–924, Mar. 2020. mal control, and the geometric theory of linear
[38] C. De Persis and P. Tesi, “low-complexity learning of linear quadratic systems.
regulators from noisy data,” 2020, arXiv:2005.01082.
[39] J. Berberich, A. Koch, C. W. Scherer, and F. Allgöwer, “Robust data-driven
state-feedback design,” in Proc. Amer. Control Conf., 2020, pp. 1532–
1538.
[40] A. Allibhoy and J. Cortés, “Data-based receding horizon control of linear
network systems,” IEEE Control Syst. Lett., vol. 5, no. 4, pp. 1207–1212,
Oct. 2021.
[41] G. Baggio, D. S. Bassett, and F. Pasqualetti, “Data-driven control of
complex networks,” 2020, arXiv:2003.12189. M. Kanat Camlibel (Member, IEEE) received
[42] N. Monshizadeh, “Amidst data-driven model reduction and control,” IEEE the Ph.D. degree in mathematical theory of sys-
Control Syst. Lett., vol. 4, no. 4, pp. 833–838, Oct. 2020. tems and control from Tilburg University, Tilburg,
[43] S. Boyd, L. E. Ghaoui, E. Feron, and V. Balakrishnan, Linear Matrix The Netherlands, in 2001.
Inequalities in System and Control Theory (ser. Studies in Applied and He is currently an Associate Professor with
Numerical Mathematics). Philadelphia, PA, USA: SIAM, 1994. the Bernoulli Institute for Mathematics, Com-
[44] T. Iwasaki, G. Meinsma, and M. Fu, “Generalized s-procedure and finite puter Science, and Artificial Intelligence, Univer-
frequency KYP lemma,” Math. Problems Eng., vol. 6, pp. 305–320, Jan. sity of Groningen, Groningen, The Netherlands,
2000. where he was an Assistant Professor between
[45] T. Iwasaki and S. Hara, “Generalized KYP lemma: Unified frequency do- 2007 and 2013. From 2001 to 2007, he held
main inequalities with design applications,” IEEE Trans. Autom. Control, Postdoctoral/Assistant Professor positions with
vol. 50, no. 1, pp. 41–59, Jan. 2005. the University of Groningen, Tilburg University, and Eindhoven Technical
[46] Z.-Q. Luo, J. F. Sturm, and S. Zhang, “Multivariate nonnegative quadratic University, Eindhoven, The Netherlands. His research interests include
mappings,” SIAM J. Optim., vol. 14, no. 4, pp. 1140–1162, 2004. differential variational inequalities, complementarity problems, optimiza-
[47] C. W. Scherer, “A full block S-procedure with applications,” in Proc. IEEE tion, piecewise affine dynamical systems, switched linear systems, con-
Conf. Decis. Control, 1997, vol. 3, pp. 2602–2607. strained linear systems, multiagent systems, model reduction, geometric
[48] T. Iwasaki, “Robust stability analysis with quadratic separator: Parametric theory of linear systems, and data-driven control.
time-varying uncertainty case,” in Proc. IEEE Conf. Decis. Control, 1997,
vol. 5, pp. 4880–4885.
[49] C. W. Scherer, “LPV control and full block multipliers,” Automatica,
vol. 37, no. 3, pp. 361–375, 2001.
[50] T. Iwasaki and G. Shibata, “LPV system analysis via quadratic separator Mehran Mesbahi (Fellow, IEEE) received the
for uncertain implicit systems,” IEEE Trans. Autom. Control, vol. 46, no. 8, Ph.D. degree in electrical engineering from the
pp. 1195–1208, Aug. 2001. University of Southern California, Los Angeles,
[51] C. W. Scherer, “Relaxations for robust linear matrix inequality problems CA, USA, in 1996.
with verifications for exactness,” SIAM J. Matrix Anal. Appl., vol. 27, no. 2, He was a member of the Guidance, Naviga-
pp. 365–395, 2005. tion, and Analysis Group, Jet Propulsion Labo-
[52] C. W. Scherer, “LMI relaxations in robust control,” Eur. J. Control, vol. 12, ratory, Pasadena, CA, from 1996 to 2000, and
no. 1, pp. 3–29, 2006. an Assistant Professor of Aerospace Engineer-
[53] C. W. Scherer and C. W. J. Hol, “Matrix sum-of-squares relaxations for ing and Mechanics with the University of Min-
robust semi-definite programs,” Math. Program., vol. 107, pp. 189–211, nesota, Minneapolis, MN, USA, from 2000 to
2006. 2002. He is currently a Professor of Aeronau-
[54] J. Umenberger, M. Ferizbegovic, T. B. Schön, and H. Hjalmarsson, “Ro- tics and Astronautics, an Adjunct Professor of Electrical and Computer
bust exploration in linear quadratic reinforcement learning,” in Proc. Int. Engineering and Mathematics, and the Executive Director of the Joint
Conf. Neural Inf. Process. Syst., 2019, pp. 15310–15320. Center for Aerospace Technology Innovation, University of Washington,
[55] M. Ferizbegovic, J. Umenberger, H. Hjalmarsson, and T. B. Schön, “Learn- Seattle, WA, USA. His research interests include aerospace systems,
ing robust LQ-controllers using application oriented exploration,” IEEE control and learning, and networks.
Control Syst. Lett., vol. 4, no. 1, pp. 19–24, Jan. 2020. Dr. Mesbahi was the recipient of the NSF CAREER Award, the NASA
[56] A. Iannelli and R. S. Smith, “A multiobjective LQR synthesis approach to Space Act Award, the UW Distinguished Teaching Award, the Graduate
dual control for uncertain plants,” IEEE Control Syst. Lett., vol. 4, no. 4, Instructor of the Year Award, and the UW College of Engineering Inno-
pp. 952–957, Oct. 2020. vator Award.

Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on December 18,2023 at 13:08:25 UTC from IEEE Xplore. Restrictions apply.