Proceedings of the 41st Chinese Control Conference

July 25-27, 2022, Hefei, China

A distributed approach to seek Nash equilibrium over jointly

strongly connected switching networks
Xiongnan He, Jie Huang
Department of Mechanical and Automation Engineering, The Chinese University of Hong Kong, Shatin, N.T., Hong Kong.
E-mail: [email protected], [email protected]

Abstract: The Nash equilibrium seeking over networks of N players has been studied under the assumption that the network is
static and strongly connected or switching and every time strongly connected. In this paper, we further consider the case where
the network is jointly strongly connected. Since a jointly strongly connected network can be disconnected at every time instant,
the existing approach cannot handle such a case. Like the literature, assuming the pseudogradient dynamics has a globally
exponentially stable Nash equilibrium, we first establish a distributed estimator for actions of all players over jointly strongly
connected networks. Then we compose the pseudogradient dynamics with the distributed estimator to obtain an extended gradient
system with N + 1 subsystems. Under the assumption that the pseudogradient of the game is strongly monotone and Lipschitz
continuous, we show that, starting from any initial condition, the state of every subsystem of the extended gradient system will
exponentially converge to the Nash equilibrium of the game concerned.
Key Words: Nash equilibrium seeking, Jointly strongly connected switching graphs, Multi-agent systems, Consensus.

⎡ ⎤
1 Introduction A1
⎢ .. ⎥
the block diagonal matrix ⎣ . ⎦. ⊗ is the no-
Recently, increasing attention has been paid to the dis-
tributed Nash equilibrium seeking for the incomplete infor- An
mation case where each player may have only limited knowl- tation of the Kronecker product. 1N is the N -dimensional
edge about the actions and /or the payoff functions of other column vector with all elements 1. IN represents an N × N
players, and individual players have to exchange the infor- identity matrix.
mation of the actions and payoff functions of opponents 2 Preliminaries
through a communication network.
In this section, we introduce the foundations of non-
The distributed Nash equilibrium seeking for a non-
cooperative game theory and graph theory based on [5], [7],
cooperative N -player game was first studied over an undi-
[8] and [10].
rected and connected network in [4, 12]. In particular, by
proposing a distributed estimator for the actions of various 2.1 Game Theory
players, the approach in [12] is able to seek the Nash equi- A non-cooperative game denoted by Γ is defined by a
librium by solving a consensus problem for an extended gra- triplet as follows:
dient system. The work in [12] was further extended to ev-
ery time strongly connected switching networks in [13]. In Γ  {N , fi , Ui }, (1)
practice, the communication network can be disconnected
from time to time. Reference [6] further investigated the dis- where N is the set of N players. For each player i ∈
tributed Nash equilibrium seeking over jointly strongly con- N of player i is denoted by xi ∈ Ui ⊂
N , the action
nected networks. This paper is an abridged version of [6]. Rni . Let i=1 ni = n and U = U1 × U2 × · · · ×
The rest of the paper is organized as follows. Section UN ⊂ Rn , which is called the strategy space. Then,
2 provides preliminaries for graph theory and game theory. fi : U → R is the payoff function for player i. Let
Section 3 summarizes some existing results of [4, 12]. Sec- x = col(x1 , x2 , · · · , xN ) ∈ Rn , which is called the strat-
tion 4 presents the main result. Section 5 closes this paper egy vector, x−i  (x1 , x2 , · · · , xi−1 , xi+1 , · · · , xN ), and
with some remarks. U−i = U1 × · · · × Ui−1 × Ui+1 · · · × UN . Then, the goal of
Notation The norm ||x|| denotes 2-norm of a vector x each player i is to minimize its payoff function fi (xi , x−i )
while ||P || is the induced Euclidean norm for a matrix P . over xi ∈ Ui , that is,
Rn is the n-dimensional Euclidean space. Rm×n is the set of minimize fi (xi , x−i ) subject to xi ∈ Ui (2)
all m × n real matrices. For column vectors ai , i = 1, · · · , s,
) = [aT1 , · · · , aTs ⎤
col(a1 , · · · , as⎡ ]T . diag(a1 , ..., an ) is the di- A strategy vector x∗ = (x∗i , x∗−i ) ∈ U is said to be a Nash
a1 equilibrium point (NE) if it is such that
⎢ .. ⎥
agonal matrix ⎣ . ⎦. block diag(A1 , ..., An ) is
fi (x∗i , x∗−i ) ≤ fi (xi , x∗−i ), ∀i ∈ N . (3)
an
Thus, any unilateral change of its own strategy increases its
payoff value.
This work was supported in part by the Research Grants Council of
the Hong Kong Special Administration Region under grant No. 14202619, 2.2 Graph Theory
in part by National Natural Science Foundation of China under Project
61973260, and in part by the Projects of Major International (Regional) A time-varying graph is denoted by G(t) = (V, E(t)),
Joint Research Program NSFC under Grant 61720106011. where V = {1, . . . , N } is the node set , and E(t) ⊂ V × V

6905
Authorized licensed use limited to: University of Virginia Libraries. Downloaded on April 08,2024 at 21:35:09 UTC from IEEE Xplore. Restrictions apply.
is the edge set. For i, j = 1, 2, ..., N and i = j, if there is an col( ∂f1 (x 1 ,x−1 ) ∂f2 (x2 ,x−2 )
∂x1 , ∂x2 , · · · , ∂fN (x N ,x−N )
∂xN ), which
edge from node j to node i, then (i, j) ∈ E(t). In this case, is called the pseudogradient of f .
we say j is a neighbor of i at time t.
Assumption 3.2 The pseudogradient F : Rn → Rn is
Let Vi (t) denote the set of all neighbors of node i at time t.
strongly monotone, i.e., (x − x )T (F (x) − F (x )) > μ||x −
A subset of E(t) of the form {(i1 , i2 ), (i2 , i3 ), ..., (ik−1 , ik )}
x ||2 , ∀x, x ∈ Rn for some μ > 0, and Lipschitz continu-
is said to be a directed path of G(t) from node i1 to node ik ,
ous, i.e., ||F (x) − F (x )|| ≤ θ||x − x ||, ∀x, x ∈ Rn , for
and in this case, we say node i1 can reach node ik at time
some θ > 0.
t. The graph G(t) is said to be static if G(t) = G(0) for all
t > 0. A static graph is denoted by G = (V, E). A static Remark 3.1 The following system
graph is said to be connected if there is a node that can reach
every other node, and is said to be strongly connected if there ∂fi (xi , x−i )
ẋi = − , i∈N (4)
is a directed path between any two nodes. An edge (i, j) is ∂xi
called an undirected edge of E if (i, j) ∈ E ⇔ (j, i) ∈ E. G
is called gradient dynamics [4], which can be put into the
is called an undirected graph if every edge in E is undirected.
following compact form:
The weight adjacency matrix of a graph G is a nonnega-
tive matrix A = [aij ] ∈ RN ×N where, for i, j = 1, 2, ..., N ,
ẋ = −F (x) (5)
i = j, aij > 0 if there is an edge from node j to node i and
aij = 0 if otherwise. Since, for i = 1, 2, ..., N ,there is no By Lemma 2 of [4] or Corollary 4.2 of [1], under Assump-
such nodes as (i, i), we have aii = 0. For i = 1, 2, ..., N , let tion 3.1, the game has a pure Nash equilibrium x∗ , which
N N
din
i = j=1 aji and di
out
= j=1 aij , which are called satisfies
the in-degree and out-degree of node i, respectively. Let
D = diag (dout out out
1 , d2 , · · · , dN ), which is called the degree F (x∗ ) = 0n (6)
matrix of G. The matrix L = D − A is called Laplacian of
G corresponding to A. By Lemma 2 of [4] again, Assumptions 3.1 and 3.2, the NE
A time function σ : [0, ∞) → P = {1, 2, . . . , n0 } x∗ of (5) is exponentially stable.
with n0 being some positive integer is said to be a piece-
Thus, (5) lends itself to a gradient-based method to seek
wise constant switching signal if there exists a sequence
Nash equilibrium for games in perfect information.
{tj : j = 0, 1, 2, . . .} satisfying t0 = 0 and tj+1 − tj ≥ τ
for some positive constant τ such that, for all t ∈ [tj , tj+1 ), Remark 3.2 If xi , i = 1, · · · , N , are governed by the fol-
σ(t) = p for some p ∈ P. P is called the switching in- lowing single integrator system
dex set, tj is called the switching instant, and τ is called the
dwell time. ẋi = ui , i = 1, · · · , N (7)
A graph Gs = (Vs , Es ) is a subgraph of G = (V, E) if
Vs ⊆ N and Es ⊆ E ∩ (Vs , Vs ). Given a set of r graphs where ui is the control input. Then, system (4) can be
{Gi = (V, Ei ), i = 1, 2, · · · , r}, the graph G = (V, E) where viewed as being obtained by applying the control law ui =
r
E = i=1 Ei is called the union of graphs Gi , denoted by − ∂fi (x i ,x−i )
∂xi .439 Thus, the problem studied here can also
r
G = i=1 Gi . be considered as the dynamic Nash equilibrium seeking with
Given a piecewise constant switching signal σ(t) and a the single integrator dynamics.
set of n0 static graphs Gi = (V, Ei ), i = 1, . . . , n0 , One 3.2 Imperfect Information
can define a time-varying graph Gσ(t) = (V, Eσ(t) ), which is
called a switching graph. A switching graph Gσ(t) is said to In the imperfect information case, players are not able
be every time connected (every time strongly connected), if to observe the actions of all the other players. The com-
for t ≥ 0, the graph Gσ(t) is connected (strongly connected). munication topology of N players is described by a graph
For any t ≥ 0, s > 0, let Gσ([t,t+s)) = ti ∈[t,t+s) Gσ(ti ) . G = (V, E) with V = {1, . . . , N } and E ⊂ V ×V. E contains
We call Gσ([t,t+s)) the union graph of Gσ(t) over the time an edge (j, i) if and only if the player i is able to observe
interval [t, t + s). the action xj of player j. The Nash equilibrium seeking for
the imperfect information case was studied in several papers,
3 Existing Results say, [4], [12] and [13]. In particular, reference [12] studied
the Nash equilibrium seeking by making use of a distributed
3.1 Perfect Information estimator for the strategic vector for the special case where
In the perfect information case, each player updates its ni = 1 for i = 1, · · · , N .
action with the continuous time, i.e., xi = xi (t). This case Let yij ∈ Rnj denote the estimation of player i on player j
was studied in, say, [1], [2], [3], [4] and [9]. For simplicity, and yi = col(yi1 , yi2 , · · · , yiN ) ∈ Rn denote the estimation
we sometimes drop the explicit dependence of xi on time. of player i on the actions of all other players. Then, the
For Nash equilibrium seeking of the game in perfect infor- distributed estimator for the strategic vector proposed in [12]
mation, the following two assumptions are commonly used, is as follows:
Assumption 3.1 For all i ∈ N , Ui = Rni , the players’ N
payoff function fi : Rn → R is C 1 in its arguments, strictly ẏij = − aik (yij − ykj ) + aij (yij − xj ) (8)
convex and radially unbounded in xi for every x−i ∈ U−i . k=1

To introduce the next assumption, let F (x) = i, j = 1, · · · , N

6906
Authorized licensed use limited to: University of Virginia Libraries. Downloaded on April 08,2024 at 21:35:09 UTC from IEEE Xplore. Restrictions apply.
where aij is the element of the adjacency matrix of the graph The following result is a rephrasing of Theorem 2 of [12].
G.
Theorem 3.1 Under Assumptions 3.3 to 3.5, suppose the
The distributed Nash equilibrium seeking updating law in
[12] is given as follows, function ∂f∂x
i (x)
i
, ∀ i ∈ N , are globally Lipschitz. Then, there
exists a positive constant δ ∗ such that, for each δ ∈ (0, δ ∗ ),
∂fi (x∗ , 1N ⊗ x∗ ) is the globally exponentially stable equilib-
ẋi = −δ k̄i (yi ) (9a) rium of (10a) and (10b).
∂xi
N
Remark 3.5 The Lyapunov function of systems (10a) and
ẏij = − aik (yij − ykj ) + aij (yij − xj ) (9b) (10b) in [12] is as follows,
k=1
c
V = (x − x∗ )T k̄−1 (x − x∗ ) + (1 − c)ỹ T P1 ỹ (13)
where k̄i is a fixed positive number, δ is some positive num- 2
ber for modulating the gradient dynamics (9a), where ỹ = y − 1N ⊗ x, 0 < c < 1 is a constant, and P1 is
The concatenated form of equation (9) is as follows: a symmetric positive-definite matrix such that
ẋ = −δ k̄H(y) (10a) P1 (L ⊗ In + B0 ) + (L ⊗ In + B0 )T P1 = Q1 (14)
ẏ = −(L ⊗ In + B0 )y + B0 (1N ⊗ x) (10b)
where Q1 is any symmetric positive-definite matrix.
where L is the Laplacian matrix of the communication 4 Main Result
graph,
In [12], the graph is required to be connected and undi-
B0 = diag (a11 In1 , a12 In2 , · · · , a1N InN , a21 In1 , · · · , rected while in [13], the graph is required to be strongly con-
a2N InN , · · · , aN 1 In1 , · · · , aN N InN ) nected for all t ≥ 0. In practice, due to the change of the en-
vironment or the failure of sensors and actuators, a switching
k̄ = diag(k̄i Ii ), y = col(y1 , · · · , yN ) ∈ RN n and H(y) = graph can be disconnected from time to time. To deal with
(y1 ) f2 (y2 )
col( f1∂x 1
, ∂x2 , · · · , fN∂x
(yN )
N
). It is noted that, when ni = this more practical scenario, we assume the communication
1, i = 1, · · · , N , (10b) is in the same form as equation (8) topology of N players is described by a switching graph
of [12]. Gσ(t) = (V, Eσ(t) ) with V = {1, . . . , N } and Eσ(t) ⊂ V × V
Reference [12] presented both local and global results. for all t ≥ 0. For any t ≥ 0, Eσ(t) contains an edge (j, i)
Here we will only summarize the global result of [12], which if and only if the player i is able to observe the action xj of
is based on the following assumptions. player j at time t.
We assume the switching graph satisfies the following as-
Assumption 3.3 The graph G is connected and undirected. sumption.
Assumption 3.4 The players’ payoff functions fi (x), ∀i ∈ Assumption 4.1 There exists a subsequence {ik } of {i : i =
N are C 2 . 0, 1, 2, . . . } with tik+1 − tik < ν for some positive number
Assumption 3.5 Each player’s function fi (xi , x−i ) is con- ν such that the union graph Gσ([tik ,tik+1 )) is strongly con-
vex in xi , ∀ xi ∈ N . Furthermore, the pseudogradient nected.
F : Rn → Rn is strongly monotone. Remark 4.1 For convenience, we say a switching graph
Remark 3.3 It is noted that the Nash equilibrium in [12] is Gσ(t) satisfying Assumption 4.1 is jointly strongly connected.
defined as follows: Under Assumption 4.1, the graph can be disconnected at
every time instant. Since the matrix P1 in the Lyapunov
fi (x∗i , x∗−i ) ≥ fi (xi , x∗−i ), ∀i ∈ N function (13) is the unique positive definite solution of (14),
which relies on the strongly connected assumption of the
Thus, the original form of Assumption 3.5 in [12] is as fol- graph (Assumption 3.3), the approach in [12] does not ap-
lows: ply to this case. Similarly, since the approach in [13] relies
Each player’s function fi (xi , x−i ) is concave in xi , on every time strongly connected assumption of the graph, it
∀ xi ∈ N . Furthermore, for all x, z ∈ Rn , does not apply to this case, either.
(x − z)T (F (x) − F (z)) ≤ −m||x − z||2 (11) Let aij (t) be the element of the adjacency matrix of the
switching Gσ(t) and Lσ(t) denote the Laplacian matrix of the
where m > 0 is a constant. switching graph Gσ(t) and
Remark 3.4 It was pointed out in [12] that, under Assump- Bσ(t) = diag (a11 (t)In1 , · · · , a1N (t)InN , a21 (t)In1 ,
tion 3.3, the matrix −(L ⊗ In + B0 ) is Hurwitz, and thus,
· · · , a2N (t)InN , · · · , aN 1 (t)In1 , · · · , aN N (t)InN )
for any x, the following equation
Then system (10) can be modified as follows:
(L ⊗ In + B0 )y q + B0 (1N ⊗ x)) = 0 (12)
ẋ = −δ k̄H(y) (15a)
has a unique solution y q ∈ RN n . It can be verified that ẏ = −(Lσ(t) ⊗ In + Bσ(t) )y + Bσ(t) (1N ⊗ x) (15b)
y q = 1N ⊗ x since L1N = 0. Also, as pointed out in
Remark 4 of [12], under Assumption 3.5, the game (1) has a To obtain our Lyapunov function candidate for the systems
unique Nash equilbrium. (15a) and (15b), we first establish the following key lemma.

6907
Authorized licensed use limited to: University of Virginia Libraries. Downloaded on April 08,2024 at 21:35:09 UTC from IEEE Xplore. Restrictions apply.
Lemma 4.1 The origin of the linear switched system Let Φ(t, τ ) be the state transition matrix of system

ẏ = −(Lσ(t) ⊗ In + Bσ(t) )y (16) ẏ = −(Lσ(t) ⊗ In + Bσ(t) )y (19)

is exponentially stable. Then, differentiating both sides of Φ(t, τ )Φ(τ, t) = IN n

gives
Proof: Let ȳl = col(y1l , . . . , yN l ) with yjl being the jth
component of ȳl , l = 1, . . . , N , respectively. Also, let ∂
Δlσ(t) = diag {a1l (t), a2l (t), . . . , aN l (t)}, l = 1, . . . , N . Φ(τ, t) = Φ(τ, t)(Lσ(t) ⊗ In + Bσ(t) ) (20)
∂t
Let ȳ = col(ȳ1 , . . . , ȳN ). Then, (16) is equivalent to the
following system Let
 ∞

ȳ˙ l = Hσ(t)l
⊗ Inl ȳl , l = 1, . . . , N (17) P (t) = Φ(τ, t)T QΦ(τ, t)dτ (21)
t

l
where Hσ(t) = Lσ(t) + Δlσ(t) . Thus, (16) is exponentially where Q is some constant positive definite matrix. Clearly,
stable if and only if, for l = 1, . . . , N , (17) is exponentially P (t) is continuous for all t ≥ 0. Since the origin of the linear
stable. switched system (19) is exponentially stable, we have
Now, for each l = 1, . . . , N , define a switching graph
l
Ḡσ(t) l
= (V̄, Ēσ(t) l
) with V̄ = {0, 1, . . . , N }, and Ēσ(t) = ||Φ(τ, t)|| ≤ α1 e−λ1 (τ −t) , τ ≥ t ≥ 0 (22)
l l
Eσ(t) Eσ(t) where Eσ(t) = {(0, j) | ajl (t) = 0, j = for some α1 > 0 and λ1 > 0. Then, it can be verified that
1, . . . , N }.
Then the Laplacian L̄lσ(t) of Ḡσ(t)
l
can be partitioned as c5 ||x||2 ≤ xT P (t)x ≤ c6 ||x||2 (23)
follows:
⎡ ⎤ for some positive constants c5 and c6 . Hence, P (t) is pos-
0 0 ··· 0 itive definite and bounded, and there exists some positive
⎢ ⎥
L̄lσ(t) = ⎢ ⎥. (18) constant p such that ||P (t)|| ≤ p for all t ≥ 0. Then, for
⎣ −Δlσ(t) 1N l
Hσ(t) ⎦
t ∈ [tj , tj+1 ), j = 1, 2, · · ·
 ∞  
∂
We will show that, for all l = 1, . . . , N , the switching Ṗ (t) = Φ(τ, t)T Q Φ(τ, t) dτ
t ∂t
l
graph Ḡσ(t) has the following properties:  ∞  T
∂
Let the subsequence {ik } of {i : i = 0, 1, 2, . . . } be as de- + Φ(τ, t) QΦ(τ, t)dτ − Q
fined in Assumption 4.1. Then, for l = 1, . . . , N , the union ∂t
ik+1 −1 l t ∞
l
graph Ḡσ([t ik ,tik+1 ])
= j=i k
Ḡσ(j) contains a spanning = Φ(τ, t)T QΦ(τ, t)dτ (Lσ(t) ⊗ In + Bσ(t) )
tree with the node 0 as the root. t
 ∞
l
For this purpose, let Ēσ([t i ,ti ]) be the edge set + (Lσ(t) ⊗ In + Bσ(t) ) T
Φ(τ, t)T QΦ(τ, t)dτ − Q
k k+1
l l
of the graph Ḡσ([t i ,t ]) . Then, Ēσ([t ,tik+1 ]) = t
k ik+1 i
l l
k
= P (t)(Lσ(t) ⊗ In + Bσ(t) )
Eσ([tik ,tik+1 ]) Eσ([t i ,ti ]) where Eσ([tik ,tik+1 ]) =
ik+1 −1 k k+1 + (Lσ(t) ⊗ In + Bσ(t) )T P (t) − Q
{(0, j) | m=ik ajl (m) = 0, j = 1, . . . , N }. Under
Assumption 4.1, there exists at least one 1 ≤ j ≤ N such Next, let ỹ = y − 1N ⊗ x. Then, the derivative of ỹ along
ik+1 −1
that m=i k
ajl (m) = 0. That is, for all l = 1, . . . , N , the solution of (15) is
l
Eσ([ti ,ti ]) is nonempty. Thus, there is at least one
k k+1
edge from the node 0 to some node j with 1 ≤ j ≤ N ỹ˙ = ẏ − 1N ⊗ ẋ
l l
(24)
in Ḡσ([t . Since Gσ([t is jointly strongly = −(Lσ(t) ⊗ In + Bσ(t) )ỹ + 1N ⊗ (δ k̄H(y))
ik ,tik+1 ]) ik ,tik+1 ])

connected, the node 0 can reach every other node of the

l
graph Ḡσ([t . Thus, the union graph Ḡσ([tl Let z = col(x − x∗ , ỹ), and let the Lyapunov function
ik ,tik+1 ]) ik ,tik+1 ])
candidate for the system (15a) and (24) be as follows,
contains a spanning tree with the node 0 as the root. As
l
Hσ(t) consists of the last N rows and the last N columns V (z, t) = W (x − x∗ ) + ỹ T P (t)ỹ (25)
of (18), by Corollary 4 of [11], for l = 1, . . . , N , (17) is
exponentially stable. Hence, (16) is exponentially stable.  where
We now present our main result as follows:
1
Theorem 4.1 Under Assumptions 3.1 and 3.2, and Assump- W (x − x∗ ) = (x − x∗ )T k̄−1 (x − x∗ ) (26)
2
tion 4.1, there exists δ ∗ > 0 such that, for 0 < δ < δ ∗ ,
the equilibrium (x∗ , 1N ⊗ x∗ ) of the system (15) is globally For any t ∈ [tj , tj+1 ) with j = 1, 2, · · · , the derivative of
exponentially stable. V along the solution of (15a) and (24) is as follows:
Proof: By Remark 3.1, under Assumptions 3.1 and 3.2, the d  T 
game (1) has a unique Nash equilibrium point x∗ . V̇ (z, t) = Ẇ (x − x∗ ) + ỹ P (t)ỹ (27)
dt

6908
Authorized licensed use limited to: University of Virginia Libraries. Downloaded on April 08,2024 at 21:35:09 UTC from IEEE Xplore. Restrictions apply.
Using Assumption 3.2 and noting H(1N ⊗ x) = F (x) gives References
Ẇ (x − x∗ ) = −δ(x − x∗ )T H(y) [1] T. Basar and G. Olsder. Dynamic Noncooperative Game The-
ory. 2nd ed. Philadelphia, PA, USA: SIAM, 1999.
= −δ(x − x∗ )T F (x) − δ(x − x∗ )T (H(y) − F (x)) (28) [2] S. Flam. Equilibrium, evolutionary stability and gradient dy-
≤ −δμ||x − x∗ ||2 + δθ||x − x∗ ||||ỹ|| namics. International Game Theory Review, 4(4):357–370,
2002.
Also, we have [3] F. Francisco and J. Pang. Finite-dimensional variational in-
equalities and complementarity problems. Springer, 2003.
d  T 
ỹ P (t)ỹ = ỹ T P (t)ỹ˙ + ỹ˙ T P (t)ỹ + ỹ T Ṗ (t)ỹ
[4] D. Gadjov and L. Pavel. A passivity-based approach to Nash
dt equilibrium seeking over networks. IEEE Transactions on
= −ỹ T (((Lσ(t) ⊗ In + Bσ(t) )T P (t) Automatic Control, 64(3):1077–1092, 2019.
[5] C. Godsil and G. Royle. Algebraic Graph Theory. New York,
+ P (t)((Lσ(t) ⊗ In + Bσ(t) ) − Ṗ (t))ỹ NY, USA: Springer, 2001.
+ (1N ⊗ (δ k̄H(y)))T P (t)ỹ [6] X. He and J. Huang, “Distributed Nash equilibrium seeking
over strongly connected switching networks”, under review
+ ỹ T P (t)(1N ⊗ (δ k̄H(y))) by Automatica, 2022.
= −ỹ T Qỹ + (1N ⊗ (δ k̄H(y)))T P (t)ỹ [7] P. Morris. Introduction to Game Theory. Springer, 2012.
[8] M. J. Osborne, M. J. Rubinstein, and A. Rubinstein. A Course
+ ỹ T P (t)(1N ⊗ (δ k̄H(y))) in Game Theory. Cambridge, Mass.: MIT, 1994.
[9] G. Scutari, F. Facchinei, J.-S. Pang, and D. P. Pallomar. Real
Let k = ||k̄||. Then we have and complex monotone communication games. IEEE Trans-
actions on Information Theory, 60(7):400–409, 2014.
d  T 
ỹ P (t)ỹ ≤ −ỹ T Qỹ [10] B. Stephen, and L. Vandenberghe. Convex Optimization.
dt Cambridge, UK ; New York: Cambridge UP, 2004.
+ 2δkθ||ỹ||||P (t)||(||ỹ|| + ||x − x∗ ||) (29) [11] Y. Su and J. Huang. Cooperative output regulation with ap-
≤ −λmin (Q)||ỹ||2 + 2δkθp||ỹ||2 plication to multi-agent consensus under switching network.
IEEE Transactions on Systems, Man, and Cybernetics, Part B
+ 2δkθp||ỹ||||x − x∗ || (Cybernetics), 42(3):864–875, 2012.
[12] M. Ye and G. Hu. Distributed Nash equilibrium seeking by a
where λmin (Q) is the minimal eigenvalue of matrix Q. consensus based approach. IEEE Transactions on Automatic
Substituting (28) and (29) into (27) gives Control, 62(9):4811–4818, 2017.
[13] M. Ye and G. Hu. Distributed Nash equilibrium seeking
V̇ ≤ −δμ||x − x∗ ||2 − λmin (Q)||ȳ||2 in multiagent games under switching communication topolo-
+ (δθ + 2δkθp)||x − x∗ ||||ȳ|| + 2δkθp||ȳ||2 gies. IEEE Transactions on Cyberbetics, 48(11):3208–3217,
  (30) 2018.
 ∗
 ||x − x∗ ||
= −δ ||x − x || ||ȳ|| B ,
||ȳ||

where
 
μ − θ2 − kθp
B= λmin (Q)
− θ2 − kθp δ − 2kθp
4μλmin (Q)
Let δ ∗ = θ 2 +4k2 θ 2 p2 +4μkpθ 2 +8μkθp . Then, for all 0 <
∗
δ < δ , the matrix B is positive definite, and thus, V̇ ≤
−δλmin (B)||z||2 , where λmin (B) is the minimal eigen-
value of matrix B. Since V is a positive definite quadratic
function, applying the Lyapunov theorem on exponential sta-
bility completes the proof. 
5 Conclusions
In this paper, we have considered the distributed Nash
equilibrium seeking over jointly strongly connected commu-
nication networks, which include the connected and undi-
rected networks as studied in [4, 12] and every time con-
nected switching networks as studied in [13] as special cases.
A key technical result is Lemma 4.1 which establishes the
exponential stability of the error dynamics of the distributed
estimator (16). This result in turn leads to the construction
of the Lyapunov’s function candidate (25) for the extended
system (15).
As in [4, 12, 13], this paper only considered static games.
The extension of the result of this paper to dynamic games
is under way.

6909
Authorized licensed use limited to: University of Virginia Libraries. Downloaded on April 08,2024 at 21:35:09 UTC from IEEE Xplore. Restrictions apply.