826 IEEE CONTROL SYSTEMS LETTERS, VOL.

8, 2024

Data Informativity for Distributed

Positive Stabilization
Takumi Iwata , Member, IEEE , Shun-Ichi Azuma , Senior Member, IEEE ,
Ryo Ariizumi , Member, IEEE , and Toru Asai , Member, IEEE

Abstract—This letter investigates data informativity for data that are not sufficient to identify the system has been
distributed positive stabilization. Under the situation that illustrated.
the system model is unavailable but the measurement So far, [7] has discussed on data informativity for stability
data are available, we address the problem of finding a
analysis, identification, controllability analysis, stabilizability
distributed controller such that the closed-loop system
by state feedback is positive and stable. We clarify that analysis, stabilization by state feedback, deadbeat control,
a necessary and sufficient condition for the problem to linear quadratic regulation, and stabilization by dynamic
be solvable is characterized by linear matrix inequalities measurement feedback. In addition, reachability analysis [8],
(LMIs), and derive an LMI-based solution to the problem. To output synchronization [9], H2 and H∞ control [10], model
demonstrate our results, a numerical example is provided. reduction [11], and model predictive control [12] have also
Moreover, an application to the vehicle formation is given. been studied.
Index Terms—Data informativity, positive systems, dis- On the other hand, research on positive systems, while
tributed control, stabilization, LMI. classical, has been actively conducted even in recent years.
Real-world applications related to positive systems include
analysis of the population dynamics of a pest described by
a Leslie model [13], structural stability analysis of biochem-
I. I NTRODUCTION ical reaction networks [14], fluid network and congestion
ODEL-BASED control theory, which relies on the control [15], and stability analysis for SIS epidemiological
M system model, is traditional and many methods on
it have been developed. On the other hand, data-driven
models [16]. In latest works on positive systems, for example,
identification of externally positive systems [17], distributed
control theory, which is based on the measurement data of model predictive control of positive systems [18], and sta-
system, is relatively new [1]. On this theory, iterative feedback bilization for a class of positive bilinear systems [19] have
tuning [2], virtual reference feedback tuning [3], fictitious been considered. One of the studies on positive systems is
reference iterative tuning [4], data-driven stabilization by state distributed positive stabilization [20], [21], which considers
feedback and linear quadratic regulation [5], data-driven model stabilizing the system state to always be nonnegative using
predictive control [6], etc., have been studied. distributed controllers. To the best of our knowledge, data
In data-driven analysis and design of systems, data infor- informativity for distributed positive stabilization has not been
mativity has been known. This is a notion which represents investigated in works on data informativity, including [7], [8],
the sufficiency of data to solve some analysis or design [9], [10], [11], [12].
problem. Based on data informativity, it has been shown In this letter, we investigate data informativity for dis-
that several analysis and designs may be possible even if tributed positive stabilization. Under the situation that the
the data are not sufficient to uniquely determine the system. system model is unavailable but the measurement data are
In [7], an example of stabilization by state feedback from the available, we address the problem of finding a distributed
controller such that the closed-loop system by state feed-
Manuscript received 7 March 2024; revised 29 April 2024; accepted back is positive and stable. In the problem, a distributed
16 May 2024. Date of publication 22 May 2024; date of current version controller is designed based only on the available data. In
10 June 2024. This work was supported in part by the JST FOREST this sense, it is more difficult than the model-based design
Program under Grant JPMJFR2123. Recommended by Senior Editor
A. P. Aguiar. (Corresponding author: Takumi Iwata.)
problem [20], [21]. Our results are described as follows. First,
Takumi Iwata is with the School of Informatics and Data we clarify a necessary and sufficient condition for the problem
Science, Hiroshima University, Higashihiroshima 739-8521, Japan to be solvable. Second, a solution to the problem is derived.
(e-mail: [email protected]). To demonstrate our results, we provide a numerical exam-
Shun-Ichi Azuma is with the Graduate School of Informatics, Kyoto
University, Kyoto 606-8501, Japan (e-mail: [email protected]). ple. Furthermore, an application to the vehicle formation is
Ryo Ariizumi is with the Faculty of Engineering, Tokyo University given.
of Agriculture and Technology, Koganei 184-8588, Japan (e-mail: This letter is related to [7], [22]. However, it differs from
[email protected]).
Toru Asai si with the College of Engineering, Chubu University,
their results in the following aspects. In [7] extending the result
Kasugai 487-8501, Japan (e-mail: [email protected]). of [5], it has shown that the necessary and sufficient condition
Digital Object Identifier 10.1109/LCSYS.2024.3404219 on data informativity for stabilization can be represented by

c 2024 The Authors. This work is licensed under a Creative Commons Attribution 4.0 License.
For more information, see https://creativecommons.org/licenses/by/4.0/
IWATA et al.: DATA INFORMATIVITY FOR DISTRIBUTED POSITIVE STABILIZATION 827

the linear matrix inequalities (LMIs). Even by straightfor-

wardly applying the techniques from [7], the necessary and
sufficient condition for our problem to be solvable cannot
be expressed as LMIs due to the structural constraint of
feedback gain derived from distributed control. To address
this challenge, we utilize the fact that there exists a positive
definite “diagonal” solution to the Lyapunov inequality for
stable positive systems [21]. In [20], this fact has been used
Fig. 1. The control of the system (1) by the distributed controllers K1 ,
to propose the LMI-based design method of distributed H∞ K2 , and K3 . In this case, the distributed indices set is given by (3).
controller based on the model-based approach. Based on
the fact, in this letter, we derive a necessary and sufficient
condition on data informativity for distributed positive sta- where F ∈ Rm×n is a feedback gain.
bilization using LMIs, and provide a data-driven distributed Now, assume that the system (1) is controlled by m
controller design method using the LMI technique. On the distributed controllers. Let xi and ui (t) be the i-th component
other hand, in [22], data-driven positive stabilization has been of x(t) and u(t), respectively. We represent the controller
considered under the assumption of the persistent excitation determining ui (t) as Ki , and suppose that each controller Ki
(PE) condition [23]. Data satisfying the PE condition are can access to partial information of the state x(t). This can be
informative for identification. Thus, our results have the represented by F where I ⊆ {1, 2, . . . , m} × {1, 2, . . . , n} is
following advantages over [22]. First, we propose a more the distributed indices set, and
generalized version of data-driven positive stabilization termed
data-driven “distributed” positive stabilization. Second, we S(I) := {M ∈ Rm×n | [M]i,j = 0 (∀(i, j) ∈
/ I)}
show that data satisfying the PE condition are not essential
for data-driven distributed positive stabilization. This suggests is the distributed gain set. Under the structural constraint F ∈
that distributed positive stabilization may be achieved even S(I), [F]i,j is constrained to be zero for all (i, j) ∈
/ I.
without informative data for identification, and that positive Therefore, if (i, j) ∈
/ I, the controller Ki cannot have access
distributed stabilization may be achieved with less data than the information on xj (t) to determine ui (t).
is required to satisfy the PE condition. Let us provide an example of the state feedback under the
The remainder of this letter is organized as follows: Problem structural constraint F ∈ S(I). Consider the system (1) with
formulation is presented in Section II. Our main results are 3-dimensional state and 3-dimensional input. Suppose that the
given in Section III. In Section IV, a numerical example is distributed indices set
provided. Moreover, an application to the vehicle formation is
I = {(1, 1), (2, 1), (2, 2), (3, 2), (3, 3)} (3)
also given. Section V concludes this letter.
Notation: Let R and R≥0 be the real number field and is given. In this case, the input u(t) is determined by
the set of nonnegative real numbers, respectively. Let PDn be ⎡ ⎤ ⎡ ⎤⎡ ⎤
the set of n × n real diagonal matrices whose all diagonal u1 (t) [F]1,1 0 0 x1 (t)
components are positive. The symbol I represents the identity ⎣u2 (t)⎦ = ⎣[F]2,1 [F]2,2 0 ⎦⎣x2 (t)⎦. (4)
matrix of appropriate size. For a matrix A, we denote the u3 (t) 0 [F]3,2 [F]3,3 x3 (t)
(i, j)-component of A by [A]i,j . A right inverse of matrix A is
On the state feedback (4), the available state information of
denoted by A† . We describe the rank of matrix A as rank(A).
Ki (i = 1, 2, 3) can be seen in Fig. 1.
If all components of matrix A are nonnegative, we say that A
We here provide the notion of S(I)-distributed positive
is nonnegative and it is denoted by A ≥ 0. For a symmetric
stabilizing gains.
matrix S, we denote by S  0 (S ≺ 0) that S is a positive
Definition 1: Consider the system (1). Suppose that a feed-
definite matrix (a negative definite matrix).
back gain F ∈ S(I) is given. If the closed-loop system by the
state feedback (2)
II. P ROBLEM F ORMULATION
x(t + 1) = (A + BF)x(t) (5)
In this section, we formulate our problems. To do this,
several definitions are provided. is stable and x(t) ∈ Rn≥0 (t = 1, 2, . . .) for any nonnegative
initial state x(0) ∈ Rn≥0 , F is said to be an S(I)-distributed
A. Distributed Positive Feedback Systems positive stabilizing gain for (A, B).
The closed-loop system (5) by an S(I)-distributed positive
In this subsection, we explain distributed positive feedback
stabilizing gain is here called the S(I)-distributed positive
systems.
feedback system.
Consider the system
x(t + 1) = Ax(t) + Bu(t), (1) B. Data Informativity for Distributed Positive Stabilization
where x(t) ∈ Rn and u(t) ∈ Rmare the state and input, In this subsection, consider the situation that the system
respectively, and A ∈ Rn×n and B ∈ Rn×m are constant model is unknown but the measurement data are given. Under
matrices. Suppose that the input u(t) is given by this situation, we address the problem of finding an S(I)-
distributed positive stabilizing gain for the system using the
u(t) = Fx(t), (2) given data.
828 IEEE CONTROL SYSTEMS LETTERS, VOL. 8, 2024

Consider the system III. M AIN R ESULTS

x(t + 1) = As x(t) + Bs u(t), (6) In this section, we present our main results for Problems 1
and 2.
where As ∈ Rn×n and Bs ∈ Rn×m are unknown constant
matrices. For the system (6), suppose that the following input
A. Solutions to Problems 1 and 2
and state data are given:
  This subsection provides the theorems that correspond to
U− := u(0) u(1) · · · u(T − 1) , solutions to Problems 1 and 2.
 
X := x(0) x(1) · · · x(T − 1) x(T) . First, we provide a solution to Problem 1.
Theorem 1: Consider the system (6). Suppose that the
The above represents the time series of input and state for the
dataset D and the distributed indices set I ⊆ {1, 2, . . . , m} ×
system (6) over the time interval {0, 1, . . . , T − 1}.
{1, 2, . . . , n} are given. The dataset D is informative for S(I)-
The pair (U− , X) is here called the dataset, and denoted by
distributed positive stabilization if and only if there exists a
D. For the dataset D, we define the set
matrix  ∈ RT×n satisfying
D := {(A, B) ∈ Rn×n × Rn×m | X+ = AX− + BU− }, 
  X−  X+ 
where X− :=  x(0) x(1) · · · x(T − 1) and X+ := T X   0,
T X+
(7)
 −
x(1) x(2) · · · x(T) are submatrices of X. Now, D is the set X−  ∈ PDn , (8)
of all (A, B) that are consistent with D. We see that (As , Bs ) ∈
X+  ≥ 0, (9)
D since the dataset D is generated from (6).
On the dataset D, we introduce the notion called data U−  ∈ S(I). (10)
informativity for S(I)-distributed positive stabilization. This Next, we also provide a solution to Problem 2.
notion represents the sufficiency of the given dataset D to Theorem 2: Consider Problem 2. If there exists a matrix
design an S(I)-distributed positive stabilizing gain.  ∈ RT×n satisfying (7)–(10), then the matrix F =
Definition 2: Consider the system (6). Suppose that the U− (X− )−1 is an S(I)-distributed positive stabilizing gain
dataset D and the distributed indices set I ⊆ {1, 2, . . . , m} × for any (A, B) ∈ D .
{1, 2, . . . , n} are given. If there exists an S(I)-distributed Here, the conditions (7)–(10) are LMIs. Therefore, both
positive stabilizing gain F for any (A, B) ∈ D , the dataset Problems 1 and 2 can be solved as convex problems.
D is said to be informative for S(I)-distributed positive If there exists no matrix  satisfying (7)–(10) for the given
stabilization. D and I, D is not informative for S(I)-distributed positive
If the condition of Definition 2 holds, it is possible to design stabilization. Then, one strategy to achieve S(I)-distributed
an S(I)-distributed positive stabilizing gain for (As , Bs ) even positive stabilization in a data-driven manner is to repeatedly
if (As , Bs ) cannot be uniquely determined from the dataset D. collect additional data until the LMI condition is satisfied.
Otherwise, additional data are required to design.
Now, our problems are formulated as follows.
Problem 1: Consider the system (6). Suppose that the B. Proofs of Theorems 1 and 2
dataset D and the distributed indices set I ⊆ {1, 2, . . . , m} × This subsection provides the proofs of Theorems 1 and 2.
{1, 2, . . . , n} are given. Then, determine whether the dataset In order to do this, we define the set
D is informative for S(I)-distributed positive stabilization or
not. (F) := {(A, B) ∈ Rn×n × Rn×m | A + BF is Schur}.
Problem 2: Consider the system (6). Suppose that the For the given matrix F ∈ Rm×n , this represents the set of
dataset D and the distributed indices set I ⊆ {1, 2, . . . , m} × (A, B) for which the closed-loop system (5) is stable.
{1, 2, . . . , n} are given. Then, find an S(I)-distributed positive Moreover, we also define the set
stabilizing gain F for any (A, B) ∈ D .
As seen from Definition 2, Problem 1 is the problem to ≥0 (F) := (A, B) ∈ Rn×n × Rn×m |
verify the feasibility of Problem 2 on the given dataset D. A + BF is nonnegative and Schur}.
Remark 1: For example, in [24], the problem imposing the
constraint on the available information for each controller, This is a subset of (F) for which the state of closed-loop
such as F ∈ S(I), has been considered. This literature has system (5) satisfies that x(t) ∈ Rn≥0 (t = 1, 2, . . .) for any
noted that in some complex systems, the use of distributed nonnegative initial state x(0) ∈ Rn≥0 (see [26]).
controllers is important and they may not have access to all If there exists an S(I)-distributed positive stabilizing gain F
information on the system. As examples of such systems, for any (A, B) ∈ D , the following relation holds:
power grids, automobiles on highways, and paper machining
D ⊆ ≥0 (F).
have been provided. The application to Eulerian models of air
traffic flows in [25] indicates the applicability of distributed Moreover, we introduce the following result, which is well-
controllers. In particular, distributed positive stabilization has known in stability analysis of positive systems.
been considered in [20], [21], all of which are model-based Lemma 1 (See [21]): For the given matrix A ∈ Rn×n ≥0 , the
approaches. However, data-driven distributed positive stabi- following statements are equivalent.
lization and data informativity for it have not been studied. 1) The matrix A is Schur.
These are expected to be useful when models are difficult to 2) There exists a diagonal matrix P ∈ PDn such that
obtain, such as in complex systems. APAT − P ≺ 0.
IWATA et al.: DATA INFORMATIVITY FOR DISTRIBUTED POSITIVE STABILIZATION 829

Now, the proofs of Theorem 1 and 2 are provided as follows. Applying Schur complement to (16), we obtain
[Proof of Theorem 1] To prove the ‘if’ part of the statement, 
P X+ 
suppose that there exists a matrix  ∈ RT×n satisfying  0. (17)
(7)–(10). Then the proof can be shown from the facts “(7)–(9) T X+
T P
⇒ (a),” “(a) ⇒ (b),” and “(a), (b), and (10) ⇒ (c)” for the Multiplying X− from the left-hand side of (15), we see that
following statements.
1) There exists the inverse (X− )−1 ∈ PDn and X+ X− †
is P = X− . (18)
†
nonnegative and Schur, where X− := (X− ) . −1
Substituting (18) into (17), we have (7). Because P = X−  ∈
† †
2) X+ X− = A + BU− X− holds for any (A, B) ∈ D . PDn , we also have (8).
† †
3) F := U− X− satisfies F ∈ S(I) and D ⊆ ≥0 (F). Next, we prove (9). From X+ X− ≥ 0, P = X−  ∈ PDn ,
First, we prove “(7)–(9) ⇒ (a).” From (8), there exists the †
and (15), we see that X+ X− P = X+  ≥ 0.
inverse (X− )−1 of X−  and (X− )−1 ∈ PDn . Applying Finally, we prove (10). Because P ∈ PDn , there exists the
Schur complement to (7), we have inverse P−1 of P. From this and (15), we have
X−  − X+ (X− )−1 T X+
T
 0. (11) †
X− = P−1 . (19)
Because X− (X− )−1 = I, (11) can also be represented as Substituting (18) into (19), we obtain X−†
= (X− )−1 . Since
†
X+ (X− )−1 X− (X− )−1 T X+
T
− X−  ≺ 0. (12) this holds and F is of form F = U− X− , we see that

We here define the matrix X− †

:= (X− )−1 . It follows from F = U− (X− )−1 . (20)
(X− )−1 ∈ PDn and (9) that X+ X− †
= X+ (X− )−1 is a It follows from F ∈ S(I), (8), and (20) that F(X− ) = U−  ∈
nonnegative matrix. Moreover, from this and (8), we can apply S(I). This completes the proof.
†
Lemma 1 to (12). Therefore, we see that X+ X− is Schur. [Proof of Theorem 2] The proof can be provided by the
Second, we prove “(a) ⇒ (b).” Because X− := (X− )−1
†
proof of the ‘if’ part of the statement of Theorem 1.
is a right inverse of X− , we see that

†   X− † † IV. N UMERICAL E XAMPLES
X+ X− = AB X = A + BU− X− (13)
U− − In this section, we provide the following examples to
holds for any (A, B) ∈ D . demonstrate our results. Examples 1 and 2 show that identifi-
Finally, we prove “(a), (b), and (10) ⇒ (c).” We define cation is not required for distributed positive stabilization.
the matrix F := U− X− †
. It follows from (X− )−1 ∈ PDn Example 1: Consider the system (6), where
⎡ ⎤ ⎡ ⎤
and (10) that F = U− X− †
= U− (X− )−1 ∈ S(I). Since 1.1 0.1 0.3 0.9 0.0 0.1
†
(a) and (b) hold, X+ X− = A + BF is nonnegative and Schur As = ⎣0.1 0.3 0.1⎦, Bs = ⎣0.1 0.6 0.1⎦. (21)
for any (A, B) ∈ D . From the above, F satisfies F ∈ S(I) 0.2 0.2 1.0 0.0 0.2 0.8
and D ⊆ ≥0 (F).
We see that the matrix As in (21) is not Schur because its
Next, to prove the ‘only if’ part of the statement, suppose
eigenvalues are 0.2708, 0.8, and 1.3292. Suppose that As and
that F is an S(I)-distributed positive stabilizing gain for any
Bs are unknown but the following data are available:
(A, B) ∈ D . Then we prove that there exists a matrix  ∈ ⎡ ⎤
RT×n satisfying (7)–(10). 1.0 0.8 0.6 0.5 0.4
Since F is an S(I)-distributed positive stabilizing gain for U− = ⎣0.8 0.6 0.4 0.3 0.2⎦,
any (A, B) ∈ D , the relation D ⊆ ≥0 (F) holds. From 0.5 0.4 0.3 0.2 0.1
≥0 (F) ⊆ (F), we can apply Lemma 2 in Appendix. ⎡ ⎤
0 0.95 2.036 3.2864 4.8383 6.7794
Therefore, we see that X− has full row rank and F is of
† † † X = ⎣0 0.63 0.820 0.9112 1.0727 1.3023⎦. (22)
the form F = U− X− for X− satisfying that X+ X− is Schur.
† 0 0.56 1.316 2.2072 3.2667 4.5689
Because X− is a right inverse of X− and D ⊆ ≥0 (F), (13)
holds and this is nonnegative and Schur for any (A, B) ∈ D . We here give the remark for the dataset D = (U− , X).
†
Applying Lemma 1 in Appendix to X+ X− , there exists a The dataset
 T TD T does not satisfy the PE condition because
diagonal matrix P ∈ PDn such that rank( X− U− ) = 5 < n + m. In other words, D is not
T informative for identification [7, Proposition 6]. Therefore, it
† †
X+ X− P X+ X− − P ≺ 0. (14) is not possible to design S(I)-distributed positive stabilizing
gains based on (As , Bs ) identified from D.
We here define the matrix Under the situation that the dataset D and the distributed
†
 := X− P (15) indices set I in (3) are given, we first consider Problem 1. By
solving LMIs in (7)–(10), we obtain
for some diagonal matrix P ∈ PDn satisfying (14). Now we ⎡ ⎤
prove that (7)–(10) hold for the matrix . 5.5097 − 4.6275 − 15.8172
First, we prove (7) and (8). Using the matrix , (14) can ⎢ −11.9852 12.7322 53.7382⎥
⎢ ⎥
also be expressed as  = ⎢ 124.0195 − 21.9785 − 178.7098⎥. (23)
⎣−229.3669 21.5912 237.3701⎦
X+ P−1 T X+
T
− P ≺ 0. (16) 107.0679 − 7.9169 − 96.5813
830 IEEE CONTROL SYSTEMS LETTERS, VOL. 8, 2024

Therefore, from Theorem 1, we see that the dataset D is x(t) = [x1 (t) x2 (t) x3 (t)]T , and u(t) = [u1 (t) u2 (t) u3 (t)]T .
informative for S(I)-distributed positive stabilization. For this system, the scenario is given as follows. Consider
Next, we consider Problem 2. From Theorem 2, we obtain stabilizing the state of each follower to the reference value
⎡ ⎤ xk0 by state feedback. While, note that each distance between
−0.2844 0 0
vehicles must not be closer than xk0 to avoid vehicle collisions.
F = U− (X− )−1 = ⎣−0.1068 0.0338 0⎦ (24)
This corresponds to guaranteeing that x(t) ∈ Rn≥0 (t =
0 − 0.2439 − 0.6637
1, 2, . . .) for any nonnegative initial state x(0) ∈ Rn≥0 . From
using  in (23). For the distributed indices set I in (3), the the limitation that each follower can access the available state
matrix in (24) satisfies F ∈ S(I). information, Since each follower can only access limited state
In fact, applying the S(I)-distributed positive stabilizing gain information, it is necessary to design a stabilizing feedback
F in (24) to the system (6), we obtain gain F such that F ∈ S(I), where the distributed indices set I
⎡ ⎤ is given by (3). Considering the situation that As and Bs are
0.8440 0.0756 0.2336 unknown but the dataset D is given, we address the problem
As + Bs F = ⎣0.0075 0.2959 0.0336⎦. (25) of finding an S(I)-distributed positive stabilizing gain F for
0.1786 0.0116 0.4691 (As , Bs ). This is exactly a case of Problem 2.
This matrix is nonnegative and Schur from that its eigenvalues In this example, let α1 = 0, α2 = 0.1, and α3 = 0.2, and
are 0.2409, 0.3961, and 0.9556. Therefore, we see that the suppose that the following data are given:
close-loop system with the matrix in (25) is an S(I)-distributed ⎡ ⎤
0 0.9 1.2 1.5 1.6
positive feedback system.
U− = ⎣0 0.7 1.1 1.3 1.4⎦,
Moreover, we conduct an experiment to investigate the
0 0.5 1.0 1.2 1.3
minimal informative dataset for S(I)-distributed positive sta- ⎡ ⎤
bilization. In this experiment, we check for the existence of 3 3.0 3.90 5.100 6.6000 8.2000
a matrix  that satisfies the LMI condition of Theorem 1 for X = ⎣2 1.8 2.32 3.188 4.1692 5.1523⎦. (28)
1000 randomly generated datasets with T = 4. As a result, 1 0.8 1.14 1.912 2.7296 3.4837
there is no matrix  satisfying the LMI condition. Thus, the
As in Example 1, this dataset D = (U− , X) also does not
experiment implies that the dataset in (22) with T = 5 is a
satisfy the PE condition.
minimal dataset in Example 1.
Under this situation, we first consider Problem 1 to verify
Example 2: Consider the system composed of one leader
the feasibility of Problem 2. By solving LMIs in (7)–(10), we
vehicle and three follower vehicles on a straight line, as shown
obtain the matrix
in Fig. 2. The leader runs autonomously. The follower #k ⎡ ⎤
(k = 1, 2, 3) has the given reference value of distance to the −0.0053 0.0262 − 0.0197
one forward vehicle, and it is denoted by xk0 . Suppose that ⎢ 0.0014 − 0.0169 0.0045⎥
⎢ ⎥
the vehicle #k (k = 1, 2, 3) can obtain the distance between  = ⎢ 0.0270 − 0.0542 − 0.0051⎥. (29)
⎣−0.0519 0.1357 − 0.0523⎦
itself and the one forward vehicle using the distance sensor.
Then the distance can be expressed as xk0 + xk , where xk 0.0266 − 0.0771 0.0503
is the deviation between the position of vehicle #k and xk0 . It follows from Theorem 1 that the dataset D is informative
Based on the information xk0 + xk and xk0 , the vehicle #k can for S(I)-distributed positive stabilization.
obtain xk . Additionally, the vehicle #k (k = 2, 3) can obtain Moreover, we consider Problem 2. From Theorem 2, we
the information xk−1 through digital communication with the obtain a solution
one forward follower vehicle #k − 1. Therefore, the vehicle ⎡ ⎤
#1 has only x1 , while the vehicle #k (k = 2, 3) can obtain −0.3682 0 0
xk and xk−1 . Moreover, the input uk of follower #k (k = F = U− (X− )−1 = ⎣ 0.0995 − 0.3871 0⎦ (30)
1, 2, 3) is the adjustment value determined by the available 0 0.0001 −0.0128
state information.
using  in (29). For the distributed indices set I in (3), we
For the above, suppose that the dynamics of follower #k
see that (30) satisfies F ∈ S(I).
(k = 1, 2, 3) is given as follows:
In fact, applying the S(I)-distributed positive stabilizing gain
xk (t + 1) = xk (t) + (uk (t) − αk xk (t)), (26) F in (30) to the system (6), we obtain the matrix
⎡ ⎤
where αk ∈ [0, 1) (k = 1, 2, 3) is the parameter that represents 0.6318 0 0
wind resistivity, and the value of αk xk (t) becomes smaller as As + Bs F = ⎣0.0995 0.5129 0 ⎦. (31)
the distance between vehicles decreases. The dynamics in (26) 0 0.0001 0.7872
can also be described as
We can verify this matrix is nonnegative and Schur from that
xk (t + 1) = (1 − αk )xk (t) + uk (t). (27) its eigenvalues are 0.7010, 0.7863, and 0.7872. Therefore, we
see that the close-loop system with the matrix in (31) is an
Moreover, based on (27), the system model for the vehicle S(I)-distributed positive feedback system.
formation can be denoted by (6), where An example of vehicle formation represented by the dis-
⎡ ⎤
1 − α1 0 0 tributed positive feedback system with the matrix in (31) is
As = ⎣ 0 1 − α2 0 ⎦, Bs = I, provided as follows. Here, let x10 = x20 = x30 = 5, and x(0) =
 T
0 0 1 − α3 4 8 12 . Fig. 3 illustrates the time evolution of each
IWATA et al.: DATA INFORMATIVITY FOR DISTRIBUTED POSITIVE STABILIZATION 831

[3] M. C. Campi, A. Lecchini, and S. M. Savaresi, “Virtual reference

feedback tuning: A direct method for the design of feedback controllers,”
Automatica, vol. 38, no. 8, pp. 1337–1346, 2002.
[4] S. Soma, O. Kaneko, and T. Fujii, “A new method of a controller
parameter tuning based on input-output data-fictitious reference iterative
tuning-,” in Proc. 8th IFAC Workshop Adapt. Learn. Control Signal
Fig. 2. The formation of one leader vehicle and three follower vehicles. Process., Yokohama, Japan, 2004, pp. 789–794.
[5] C. D. Persis and P. Tesi, “Formulas for data-driven control: Stabilization,
optimality, and robustness,” IEEE Trans. Autom. Control, vol. 65, no. 3,
pp. 909–924, Mar. 2020.
[6] J. Berberich, J. Köhler, M. A. Müller, and F. Allgöwer, “Data-driven
model predictive control with stability and robustness guarantees,” IEEE
Trans. Autom. Control, vol. 66, no. 4, pp. 1702–1717, Apr. 2021.
[7] H. J. van Waarde, J. Eising, H. L. Trentelman, and M. K. Camlibel, “Data
informativity: A new perspective on data-driven analysis and control,”
IEEE Trans. Autom. Control, vol. 65, no. 11, pp. 4753–4768, Nov. 2020.
[8] J. Eising and M. K. Camlibel, “Data informativity for analysis of linear
systems with convex conic constraints,” IEEE Control Syst. Lett., vol. 6,
Fig. 3. The time evolution of each distance between vehicles xk0 + xk (t ) pp. 391–396, 2021.
(k = 1, 2, 3). [9] J. Jiao, H. J. van Waarde, H. L. Trentelman, M. K. Camlibel, and
S. Hirche, “Data-driven output synchronization of heterogeneous leader-
follower multi-agent systems,” in Proc. 60th IEEE Conf. Decis. Control,
Austin, TX, USA, 2021, pp. 466–471.
distance between vehicles xk0 + xk (t) (k = 1, 2, 3). Thus, we [10] H. J. van Waarde, M. K. Camlibel, J. Eising, and H. L. Trentelman,
see that all distances converge to the reference values. “Quadratic matrix inequalities with applications to data-based control,”
As a similar manner to Example 1, we conduct an exper- SIAM J. Control Optim., vol. 61, no. 4, pp. 2251–2281, 2023.
[11] A. M. Burohman, B. Besselink, J. M. A. Scherpen, and M. K. Camlibel,
iment to investigate the minimal informative dataset for “From data to reduced-order models via generalized balanced trunca-
S(I)-distributed positive stabilization. The result implies that tion,” IEEE Trans. Autom. Control, vol. 68, no. 10, pp. 6160–6175,
the dataset in (28) with T = 5 is a minimal dataset. Oct. 2023.
[12] H. H. Nguyen, M. Friedel, and R. Findeisen, “LMI-based data-driven
robust model predictive control,” 2023, arXiv:2303.04777.
V. C ONCLUSION [13] X. Chen, J. Lam, P. Li, and Z. Shu, “1 -induced norm and controller syn-
thesis of positive systems,” Automatica, vol. 49, no. 5, pp. 1377–1385,
In this letter, we discussed on data informativity for dis- 2013.
tributed positive stabilization. For the data-driven distributed [14] F. Blanchini and G. Giordano, “Piecewise-linear Lyapunov functions
positive stabilization, we clarified a necessary and sufficient for structural stability of biochemical networks,” Automatica, vol. 50,
condition for the problem to be solvable. It has been shown no. 10, pp. 2482–2493, 2014.
[15] F. Blanchini, P. Colaneri, and M. E. Valcher, “Switched positive linear
that this condition is characterized by LMIs. Moreover, an systems,” Found. Trends Syst. Control, vol. 2, no. 2, pp. 101–273, 2015.
LMI-based solution to the problem has been provided. Our [16] M. A. Rami, V. S. Bokharaie, O. Mason, and F. R. Wirth, “Stability
main results have been demonstrated by numerical examples. criteria for SIS epidemiological models under switching policies,” 2014,
arXiv:1306.0135.
[17] C. Grussler, J. Umenberger, and I. R. Manchester, “Identification of
A PPENDIX externally positive systems,” in Proc. 56th IEEE Conf. Decis. Control,
Melbourne, VIC, Australia, 2017, pp. 6549–6554.
To prove Theorems 1 and 2, the following definition and [18] J. Zhang, L. Zhang, and T. Raïssi, “A linear framework on the distributed
lemma are given. model predictive control of positive systems,” Syst. Control Lett.,
Definition 3 (See [7]): Consider the system (6). Suppose vol. 138, Apr. 2020, Art. no. 104665.
[19] Y. Kawano and M. Cucuzzella, “Stabilization for a class of positive
that the dataset D is given. If there exists an F ∈ Rm×n such bilinear systems,” IEEE Control Syst. Lett., vol. 7, pp. 3519–3524, 2023.
that (5) is stable for any (A, B) ∈ D , the dataset D is said to [20] T. Tanaka and C. Langbort, “KYP lemma for internally positive systems
be informative for stabilization. and a tractable class of distributed H-infinity control problems,” in Proc.
Lemma 2 (See [7]): Consider the system (6). Suppose that Am. Control Conf., Baltimore, MD, USA, 2010, pp. 6238–6243.
[21] A. Rantzer, “Distributed control of positive systems,” in Proc. 50th
the dataset D is given. Then, the following statements hold. IEEE Conf. Decis. Control Eur. Control Conf., Orlando, FL, USA, 2011,
1) The dataset D is informative for stabilization if and only pp. 6608–6611.
if X− has full row rank and there exists a right inverse [22] B. Shafai, A. Moradmand, and M. Siami, “Data-driven positive stabi-
† † lization of linear systems,” in Proc. 8th Int. Conf. Control, Decis. Inf.
X− of X− such that X+ X− is Schur. Technol., Istanbul, Turkey, 2022, pp. 1031–1036.
2) The matrix F satisfies D ⊆ (F) if and only if F is of [23] J. C. Willems, P. Rapisarda, I. Markovsky, and B. L. D. Moor, “A
† †
the form F = U− X− for the matrix X− satisfying that note on persistency of excitation,” Syst. Control Lett., vol. 54, no. 4,
† pp. 325–329, 2005.
X+ X− is Schur. [24] M. Rotkowitz and S. Lall, “A characterization of convex problems in
decentralized control,” IEEE Trans. Autom. Control, vol. 50, no. 12,
R EFERENCES pp. 1984–1996, Feb. 2006.
[25] H. M. Arneson and C. Langbort, “Distributed control design for a class
[1] Z. Hou and J. Xu, “On data-driven control theory: The state of the art of compartmental systems and application to Eulerian models of air
and perspective,” ACTA Automatica Sinica, vol. 35, no. 6, pp. 650–667, traffic flows,” in Proc. 46th IEEE Conf. Decis. Control, New Orleans,
2009. LA, USA, 2007, pp. 2876–2881.
[2] H. Hjalmarsson, M. Gevers, S. Gunnarsson, and O. Lequin, “Iterative [26] M. A. Rami and F. Tadeo, “Positive observation problem for linear
feedback tuning: Theory and applications,” IEEE Control Syst. Mag., discrete positive systems,” in Proc. 45th IEEE Conf. Decis. Control, San
vol. 18, no. 4, pp. 26–41, Aug. 1988. Diego, CA, USA, 2006, pp. 4729–4733.