Nash Social Distancing Games with Equity Constraints: How Inequality

Aversion Affects the Spread of Epidemics?

Ioannis Kordonisa , Athanasios-Rafail Lagosa , George P. Papavassilopoulosa,b

a National Technical University of Athens, School of Electrical and Computer Engineering, 9 Iroon Polytechniou str., Athens, Postal Code 157 80,
Greece
b University of Southern California, Department of Electrical Engineering-Systems, Los Angeles, U.S.A.

Abstract
In this paper, we present a game-theoretic model describing voluntary social distancing during the spread of an
epidemic. The payoffs of the agents depend on the social distancing they practice and on the probability of getting
infected. We consider two types of agents: the non-vulnerable agents with a small cost if they get infected and
the vulnerable agents with a higher cost. For the modeling of the epidemic outbreak, we consider a variant of the
SIR (Susceptible-Infected-Removed) model involving populations of susceptible, infected, and removed persons of
vulnerable and non-vulnerable types. The Nash equilibria of this social distancing game are studied. The main
contribution of this work is the analysis of the case where the players, desiring to achieve a low social inequality,
pose a bound on the variance of the payoffs. In this case, we introduce and characterize a notion of Generalized Nash
Equilibrium (GNE) for games with a continuum of players and provide characterizations for this type of GNE. It turns
out that often there is a continuum of GNE. However, among the GNE, for a given value of the variance bound, there
is one that Pareto dominates the others. We also provide conditions under which a more restrictive variance constraint
benefits all the game participants. Furthermore, we describe a bargaining-based algorithm for choosing the variance
constraint. Through numerical studies, we show that inequality constraints result in a slower spread of the epidemic
and an improved cost for the vulnerable players. Furthermore, we present some examples where inequality constraints
are also beneficial for non-vulnerable players.
Keywords:
COVID-19 Pandemic, Nash Games, Inequality Aversion, Social Distancing

1. Introduction

Epidemics have harassed humanity for centuries, and people have investigated several strategies to contain them.
The development of medicines and vaccines and the evolution of healthcare systems with specialized personnel and
equipped hospitals have significantly affected the spread of many epidemics and eliminated some contagious diseases.
However, during the current COVID-19 pandemic, Non-Pharmaceutical Interventions (primarily social distancing)
have been among the most effective strategies to reduce the disease spread. Due to the uneven distribution of the
vaccines, the emergence of SARS-CoV-2 variants, and people’s resistance to vaccination, social distancing will likely
remain significant in a large part of the globe in the near future.
Epidemiological models are essential in designing measures and strategies to control epidemics1 . In the last
century, epidemiologists have made significant progress in the mathematical modeling of the spread of epidemics.
From the seminal works of Kermack and McKendrick [2] and Ross [3], a prevalent approach in the mathematical
modeling of epidemics is compartmental models. These models consider that each agent belongs in some

? I. Kordonis e-mail: jkordonis1920@yahoo.com, A.-R. Lagos e-mail: lagosth@mail.ntua.gr, G.P. Papavassilopoulos e-mail:
yorgos@netmode.ntua.gr
1 For example, Imperial College London’s report [1] profoundly influenced UK’s response to COVID-19 epidemic.

Preprint submitted to Applied Mathematics and Computation July 4, 2022

compartment according to her infection state (e.g., Susceptible-Infected-Recovered) and study the evolution of each
compartment’s population. The literature on these models is extensive, so for a summary, we refer to Chapter 2 of
[4]. There are also other elegant approaches to epidemics modeling, such as the ones that take into consideration
the heterogeneous networked structure of human interconnections [5]. However, the compartmental models remain a
well-studied and fruitful approach, widely used in real-life applications.
The development of epidemiological models is a valuable tool in designing protective measures against the spread
of an epidemic. Still, these measures will be adopted by agents who act in a self-interested manner, at least to some
extent. Thus, game theory is an appropriate complementary mathematical tool to be used in this field. Indeed, many
game-theoretic models have been developed to study voluntary vaccination [6, 7, 8, 9, 10, 11, 12, 13] and behavioral
changes of the agents [14, 15, 16, 17, 18, 19, 20, 21], such as social distancing, use of face masks, and better hygiene
practice. Another closely related stream of research is the study of the adoption of decentralized protection strategies
in engineered and social networks [22, 23, 24, 25]. Recently, with the emergence of the COVID-19 pandemic, there is
a renewed interest in modeling individual behaviors. Related tools include dynamic game analysis of social distancing
[26], evolutionary game theory [27, 28, 29, 30] and network game models [31].
The majority of game-theoretic models are based on the assumption that the rational behavior for an agent is
to maximize selfishly her own payoff ignoring the social impacts (externalities) of this choice. This can lead to
‘free-riding’ phenomena in vaccination games [9, 12] or in disobedience of social distancing rules in social distancing
games, which can both result in a higher prevalence of the spread of the epidemic [9, 32] and to harmful consequences
for the vulnerable members of the society. The phenomenon that the Nash equilibrium strategies result to a social
welfare less than the optimal one is well known in game theory community as the Tragedy of the Commons [33].
There are some notable exceptions [34, 35] analyzing epidemic games involving altruistic individuals. However, even
in the cases considering that the agents prefer the strategies that maximize a social welfare function, there may still
exist significant inequalities among their payoffs.
Epidemics may create vastly unequal outcomes in terms of health risks. For example, the severe illness or fatality
risk for a person infected by SARS-COV-2 varies widely with age, and underlying health conditions [36, 37]. It is,
therefore, interesting to investigate how people treat these inequality issues. There is a lot of empirical evidence that
people are often motivated by fairness considerations [38, 39]. That is, people are often willing to sacrifice some of
their own payoff to achieve a more equitable outcome. When it comes to health inequalities, people are often very
inequality averse [40]. Especially if an agent has vulnerable relatives, it is quite natural for her to alter her behavior
during an epidemic outbreak to protect them. In the context of the current COVID-19 crisis, it has been observed
that communication strategies that aim to indicate the effects of social distancing behavior on others, especially on
vulnerable persons (strategy of the identifiable victim), are very effective [41].
In this work, we employ a novel approach to model the agents’ possible desire to keep the inequality among their
payoffs below a certain threshold. Particularly, we consider that the players share a common constraint bounding
inequality, modeled as their costs’ variance. Following the literature (e.g. [42, 43, 44]), we consider a compartmental
model (SIR) for the spread of an epidemic and a social distancing game among the agents. The payoffs of the agents
consist of two terms: a cost for the social distancing and a cost proportional to the probability of getting infected.
There are two types of agents, non-vulnerable agents, who have a small cost if they get infected and vulnerable agents,
who have a higher cost. The size of the society is considered large, so the game has a continuum of players (it is a
non-atomic game). In this game, the agents determine their actions to optimize their payoff and simultaneously respect
a constraint concerning the variance of all the agents’ payoffs. Due to this constraint, the game is, in fact, a generalized
game with a non-convex constraint. The majority of the bibliography on generalized games [45, 46] does not analyze
generalized games with non-convex constraints. Furthermore, there are a few references on generalized non-atomic
games [47, 48, 49]. However, in these papers, the convexity of the constraint set is built into the definition of the
Generalized Nash Equilibrium (GNE). Another related paper by Singh and Wiszniewska-Matyszkiel [50] examines a
dynamic game with a continuum of players having state-dependent constraints. In this work, we give a new definition
and characterization of GNE for constrained non-atomic games. As it often happens in constrained games, there is
frequently a multitude (or even a continuum) of GNE. We present some results for structure of the GNE of the variance
constrained epidemics game, and some tools for the equilibrium selection. Furthermore, we provide conditions, under
which a more restrictive variance constraint benefits all the participants of the game.
Numerical results indicate that there may be many Nash equilibria inducing different costs for the players. Thus,
even in the absence of social distancing regulations, it is beneficial for the players to coordinate and choose the
2
‘best’ equilibrium. In the variance constrained case, we numerically find that the inequality constraint (bounding the
variance) can be beneficial for all the players. For vulnerable players, stricter constraints found to be always beneficial,
while for the non-vulnerable, there is an optimal intermediate value for the variance constraint.
The rest of the paper is organized as follows. Section 2 describes the compartmental model for the epidemic
outbreak and the social distancing game between the agents. In Section 3, we analyze the game and characterize its
Nash equilibria. In Section 4, we introduce the constraint that concerns the variance of the payoffs and derive an
appropriate definition of generalized Nash equilibrium for variance constrained games. Section 5 analyzes symmetric
GNE with variance constraints, and proposes some tools for equilibrium selection. In section 6, we present a
methodology for the computation of the Nash equilibrium strategies of Section 3 and then give some numerical
examples. Moreover, in the same section, we present an example of generalized Nash equilibrium computation for
the variance constrained game. In the Appendix, we collocate the proofs of several propositions of the paper.

2. Mathematical Model
This section presents a variation of a popular epidemics model which assumes a continuum of agents. The state of
each agent could be Susceptible (S), Infected (I), Recovered (R) or Dead (D). A susceptible person may be infected at
a rate proportional to the rate she meets with infected people. Infected persons either recover or die at a constant rate.
We assume that an individual recovered from the infection is immune i.e., she could not be infected again.
We distinguish between two types of agents: non-vulnerable and vulnerable. We use the index j = 1 for
non-vulnerable agents and j = 2 for vulnerable. The difference of the two types of agents is the severity of a possible
infection (including the probability to survive). An infected agent recovers with a probability rate α0j and dies with a
probability rate α j − α0j . The evolution of the individual states is presented in Figure 1.

α'j
S I R

αj-α'j
D

Figure 1: The Markov process describing the evolution of the state of each individual.

We analyze the behavior of the agents for a time interval [0, T ]. Denote by ui ∈ [um , u M ] the action of player i,
indicating the fraction of time this person spends in public places. The minimum value of the actions um describes
the minimum contact a person needs for surviving and u M describes a restriction placed by the government. In the
absence of a restriction we consider u M = 1.
Assumption 1: The actions ui of all the players are constant during a time interval [0, T ]. This interval represents
a wave of the epidemic.
Remark 1. In a more general model ui could be function of the state variables or time, but there are some arguments
in favor of this choice. First, there may be a high uncertainty for the values of the state variables. Second, ui ’s reflect
some everyday routine choices of the people and these choices may be difficult to adapt constantly.
Denote by µ1 a (Borel) measure on [um , u M ] describing the distribution of the actions of the players in category 1
i.e., for A ⊂ [um , u M ], the value of µ1 (A) denotes the mass of the players using an action u ∈ A. Similarly, denote by µ2
the distribution of actions of the players of type 2. The total mass of players of types 1 and 2 is n1 and n2 respectively
i.e., µ1 ([um , u M ]) = n1 and µ2 ([um , u M ]) = n2 . We first describe the evolution of the epidemic for general distributions
µ1 , µ2 .
Denote by S 1u (t) the probability a non-vulnerable player who plays u ∈ [um , u M ] to be susceptible at time t and by
I1u (t) the probability to be infected. Similarly define S 2u (t), I2u (t). The rate at which this person gets infected is given
by: ruI f , where r is a positive constant and I f denotes the density of infected people in ‘public places’. Each player
contributes to I f (t) proportionally to her probability of being infected at time t. The dynamics is given by:
Ṡ ju = −ruS ju I f (t)
, (1)
I˙ju = ruS ju I f (t) − α1 I ju
3
j = 1, 2 and: Z Z
I f (t) = I1u0 (t)u0 · µ1 (du0 ) + I2u0 (t)u0 · µ2 (du0 ). (2)
[um ,u M ] [um ,u M ]

The initial conditions are:

S 1u (0) = S 2u (0) = (1 − I0 ), I1u (0) = I2u (0) = I0 , (3)
where I0 is the percentage of infected persons at time 0. Here we assume, without loss of generality, that at the
beginning of the time interval [0, T ], the agents of both types are infected with the same probability I0 .
It is convenient to append differential equations (1) with an auxiliary variable z(t) which evolves according to:

ż = I f (t), (4)

with initial condition z(0) = 0.

Remark
RT 2. The reason for this extension is that the costs of the players (defined in a bit) can be expressed in terms
of 0 I (t) = z(T ).
f

We then introduce some function spaces which will be used in the proof of the existence and uniqueness of a
solution to the initial value problem (1)–(4). Let:

X = C([um , u M ], R4 ) × R,

where C([um , u M ], R4 ) be the space of continuous functions defined on [um , u M ] and values on R4 . The space X
equipped with the norm:
kxk = max{|x1u |, |x2u |, |x3u |, |x4u |, |x5 | : u ∈ [um , u M ]},
is a Banach space. We also consider the Banach space Y of signed measures on [um , u M ] with the total variation norm.

Remark 3. a) The function space X describes the space in which (S 1u (t), S 2u (t), I1u (t), I2u (t), z(t)) evolves.
Particularly, for a fixed t, S ju (t), I ju (t) are functions : [um , u M ] → R, and z(t) is a real number. b) The distributions of
the players’ actions live in Y. This function space will be also used in Section 4.

Proposition 1. The initial value problem (1)–(3) has a unique solution. Furthermore, this solution is continuous on
u.

Proof: See Appendix A. 

The cost of an agent i of type j consists of two terms. The first term is proportional to the probability of getting
infected and the vulnerability of the player. The second term, represents the benefits earned from social interactions.
The cost is given by:
Ji = G j Pi − Q j (ui , ū1 , ū2 ), (5)
where G j corresponds to the expected severity of a possible infection, Pi is the probability that i gets infected within the
time interval [0, T ]. Note that G2 > G1 . The quantity Q j (ui , ū1 , ū2 ) represents the utility derived from the interaction
with others where ū1 , ū2 are the mean actions of the players of types j = 1 and j = 2 respectively. For simplicity, we
assume that Q j has a bilinear form:
Q j (ui , ū1 , ū2 ) = s j1 ui ū1 + s j2 ui ū2 ,
where s j1 , s j2 are non-negative constants.
Remark 4. In the computation of the second term in (5), we assume that the number of people in these types is
approximately constant with time. This is a good approximation in epidemics with a low mortality rate and duration
small compared to the average human life.

4
 Let us then compute the probability of getting infected, Pi . The probability S ijui (t) that an agent i of either type
( j = 1 or j = 2) is not infected up to time t evolves according to:

Ṡ ijui = −rS ijui ui I f .

Thus, we have: " Z T !#

Pi = 1 − S ijui (T ) = I0 + (1 − I0 ) 1 − exp −rui I (t)dt .
f
0

Here we assume that at the beginning of the time interval [0, T ], all the agents are infected with a small probability I0 .
RT
Denoting by F(µ1 , µ2 ) = r 0 I f (t)dt the cost is written as:

J j (ui , µ1 , µ2 ) = G j I0 + (1 − I0 ) 1 − exp (−ui F(µ1 , µ2 )) − s j1 ui ū1 − s j2 ui ū2 .

  

Since we assume a very large population of players, each one of them is not able to affect the distributions µ1 , µ2 .
Thus, the variations of µ1 , µ2 with respect to ui are not considered. It is interesting to observe that the individual cost
Ji is concave in ui . To see this take the second derivative of Ji with respect to ui :
!2
∂2 J i T T
Z ! Z
= −G j (1 − I0 ) exp −rui I f (t)dt − I f (t)dt < 0.
∂u2i 0 0

Therefore, the possible actions minimizing the individual cost are ui = um and ui = u M . Thus, to compute the Nash
equilibria, we focus on distributions assigning the entire mass on {um , u M }.

Remark 5. The fact that the cost function Ji is concave in ui simplifies the analysis a lot, implying that players choose
either u = um or u = u M . It further allows us to describe dynamics using a finite-dimensional model.

3. Nash Equilibrium

To analyze the Nash equilibrium we focus on distributions having all the mass on {um , u M }.The dynamics is given
by:
Ṡ jum = −rum I f S jum , Ṡ juM = −ru M I f S juM ,
I˙jum = rum I f S jum − α j I jum , I˙juM = ru M I f S juM − α j I juM , (6)
ż = I ,
f

where j = 1, 2, the total mass of ‘free infected people’ I f is given by:

2
X
I f (t) = n j ((1 − ũ j )um I jum (t) + ũ j u M I juM (t)),
j=1

and ũ j = µ j ({u M }) is the percentage of players of type j using u M . The initial conditions are given by:

S jum (0) = S juM (0) = (1 − I0 ), I jum (0) = I juM (0) = I0 , z(0) = 0 (7)

Denote by φzũ1 ,ũ2 (t) the z-part of the solution of the differential equation (6) with initial conditions (7). Then it
holds:
F(µ1 , µ2 ) = F(ũ1 , ũ2 ) = rφzũ1 ,ũ2 (T ).

Remark 6. We may interpret the equilibria where some players of type j play um and some u M as equilibria in
symmetric mixed strategies. Particularly, each of type j randomizes and plays u M with probability ũ j .

Proposition 2. Consider a set of strategies characterized by ũ1 , ũ2 , where a fraction 1 − ũ j of the players of type j use
u = um and a fraction ũ j of the players of type j use u = u M . This set of strategies is a Nash equilibrium if and only if,
for each j = 1, 2, one of the following holds:
5
 (i) 0 < ũ j < 1 and:
G j (1 − I0 )(e−um F(ũ1 ,ũ2 ) − e−uM F(ũ1 ,ũ2 ) ) = (u M − um )(s j1 ū1 + s j2 ū2 ),
where ū j = um + (u M − um )ũ j .
(ii) ũ j = 0 and:
G j (1 − I0 )(e−um F(ũ1 ,ũ2 ) − e−uM F(ũ1 ,ũ2 ) ) ≥ (u M − um )(s j1 ū1 + s j2 ū2 ).

(iii) ũ j = 1 and:
G j (1 − I0 )(e−um F(ũ1 ,ũ2 ) − e−uM F(ũ1 ,ũ2 ) ) ≤ (u M − um )(s j1 ū1 + s j2 ū2 ).

Proof : The proof is immediate, observing that (i) corresponds to the case where the players of type j are indifferent
between um and u M and (ii), (iii) correspond to preference of um over u M and u M over um respectively. 
The existence of a Nash equilibrium is a consequence of Theorem 1 of [51].

Corollary 1. Assume that s11 = s21 , s12 = s22 , and that G2 > G1 . Then the possible Nash equilibria (ũ1 , ũ2 ) are in of
one of the following forms (0, 0), (ũ1 , 0), (1, 0), (1, ũ2 ), (1, 1).

Proof : Let (ũ1 , ũ2 ) be a Nash equilibrium. Then, if ũ1 < 1 it holds:

G1 (1 − I0 )(e−um F(ũ1 ,ũ2 ) − e−uM F(ũ1 ,ũ2 ) ) ≥ (u M − um )(s11 ū1 + s12 ū2 ).

Since, s11 = s21 , s12 = s22 and G2 > G1 we have:

G2 (1 − I0 )(e−um F(ũ1 ,ũ2 ) − e−uM F(ũ1 ,ũ2 ) ) > (u M − um )(s21 ū1 + s22 ū2 ).

But, since (ũ1 , ũ2 ) is a Nash equilibrium Proposition (2) implies that ũ2 = 0 
Corollary 2. If um = 0, then (ũ1 , ũ2 ) = (0, 0) is always a Nash equilibrium.

4. The Variance Constrained Game

This section analyzes a game situation, where the players pose a shared bound on the variance of their costs. To
do so, we first introduce a notion of equilibrium with a shared constraint, for non-atomic games, and then characterize
it in terms of small variations. We assume that the strategies of the players are symmetric (that is all the players of the
same type use the same strategy), allowing for randomization. Consider a pair of distributions (µ1 , µ2 ) for the actions
of the players. Then, the players of type j randomize according to µ̄ j (·) = µ j (·)/n j .
The variance of the costs is given by:

(J1 (u0 , µ1 , µ2 ) − J)
¯ 2 µ1 (du0 ) + (J2 (u0 , µ1 , µ2 ) − J)¯ 2 µ2 (du0 )
R R
V(µ1 , µ2 ) = ,
n1 + n2
(8)
n1 (J1 (u0 , µ1 , µ2 ) − J) ¯ 2 µ̄1 (du0 ) + n2 (J2 (u0 , µ1 , µ2 ) − J)¯ 2 µ̄2 (du0 )
R R
= ,
n1 + n2

where J¯ = (n1 J¯1 + n2 J¯2 )/(n1 + n2 ), and:

Z Z
1
J¯j = J j (u0 , µ1 , µ2 )µ j (du0 ) = J j (u0 , µ1 , µ2 )µ̄ j (du0 ).
nj

We then describe a notion of equilibrium for the generalized game with variance constraint. Ideally, to have an
equilibrium, the actions of each player should minimize the cost subject to the variance constraint. The difficult point
here is that, since we have a continuum of players, the variance does not depend on the actions of individual players.
To define a meaningful notion of equilibrium, instead of analyzing the effect of a deviation of a single player, we
consider the deviation of a small fraction of players of type j and see how a variation from the nominal mixed strategy

6
µ̄ j affects the cost of this small group of players and the total variance. Then, we take the limit as the total mass of the
group of players tends to zero.
Denote by j the type of players containing the deviating group and by − j the other type of players. Assume
that the total mass of deviating players is ε and that the deviating players use a mixed strategy µ̄0j (note that it holds
µ̄0j ([um , u M ]) = 1). Then, the distribution of the actions of the players of type j is given by:

µ j + ε(µ̄0j − µ̄ j ) = µ j + εδµ j ,

where we use the notation δµ j to describe the difference µ̄0j − µ̄ j .

The mean cost of the deviating players after the deviation is:
Z
Jj =
¯dev
J j (u0 , µ j + ε(µ̄0j − µ̄ j ), µ− j )µ̄0j (du0 ),

while before the deviation is J¯j . The following lemma expresses the limit of this deviation, as well as the directional
(Gateaux) derivative of the variance, in terms of linear bounded operators. This result will be used to define the
Generalized Nash equilibrium.

Lemma 1. For all µ− j it holds:

j − J j , as ε → 0, is a linear functional of δµ j = µ̄ j − µ̄ j . Particularly, it is written

(i) The limit of the variation J¯dev ¯ 0

as: Z
j
lim( J j − J j ) = Kµ1 ,µ2 δµ j =
¯dev ¯ J j (u0 , µ j , µ− j )δµ j (du0 ), (9)
ε→0

where Kµj1 ,µ2 ∈ Y ? and Y ? is the space of bounded linear functionals on Y .

(ii) The directional derivative of the variance V(µ j , µ− j ) in the direction δµ j = µ̄0j − µ̄ j is expressed as:

V(µ j + εδµ j , µ− j ) − V(µ j , µ− j )

lim = Lµj1 ,µ2 δµ j ,
ε→0 ε

where Lµj1 ,µ2 ∈ Y ? . Furthermore, Lµj1 ,µ2 can be written as:

Z
Lµj1 ,µ2 δµ j = var
f j,V,µ1 ,µ2
(u0 )δµ j (du0 ), (10)

var
with f j,V,µ1 ,µ2
(u0 ) continuous.

Proof: See Appendix B. 

Definition 1 (Generalized Nash Equilibrium). A distribution of actions described by (µ1 , µ2 ) is a Generalized Nash
Equilibrium (GNE) with variance constraint V ≤ C if either:

(i) V(µ1 , µ2 ) < C and for any j = 1, 2 and any probability measure µ̄0j , it holds:

Kµj1 ,µ2 δµ j ≥ 0,

where δµ j = µ̄0j − µ̄ j , or

(ii) V(µ1 , µ2 ) = C and for any j = 1, 2 and any probability measure µ̄0j , it holds:

Kµj1 ,µ2 δµ j < 0 =⇒ Lµj1 ,µ2 δµ j > 0,

where δµ j = µ̄0j − µ̄ j .

7
Remark 7. In the first case of the definition, any small group of players is not sufficient to increase the variance above
C. Thus, if µ1 , µ2 is an equilibrium, there is no profitable deviation, and the definition coincides with the equilibrium
of Section 3. In the second case, µ1 , µ2 is an equilibrium, if any profitable deviation for a small group of players,
increases the variance above C.
Remark 8. Some notions of GNE for games with a continuum of players were already introduced in the literature
[47, 48, 49]. However, these definitions assume a convex constraint set. Let us note that Definition 1 is neither a
generalization nor a special case of the these definitions.
We then introduce a refinement of GNE, called variance non-stationary GNE. It turns out that variance
non-stationary GNE are easier to compute.
Definition 2 (variance stationary GNE). A pair (µ1 , µ2 ) is variance stationary if either:
(i) for all δµ1 = µ̄01 − µ̄1 , with µ̄01 probability measure, it holds:

Lµ11 ,µ2 δµ1 ≥ 0,

or
(ii) for all δµ2 = µ̄02 − µ̄2 , with µ̄02 probability measure, it holds:

Lµ21 ,µ2 δµ2 ≥ 0.

We call a GNE (µ1 , µ2 ) variance non-stationary if it is not variance stationary.

Lemma 2. Assume that (µ1 , µ2 ) is not variance stationary. Then, (µ1 , µ2 ) is a Generalized Nash equilibrium with
variance constraint V ≤ C if and only if either it satisfies (i) of Definition 1 or V(µ1 , µ2 ) = C and for any probability
measure µ̄0 , it holds:
Kµj1 ,µ2 δµ j < 0 =⇒ Lµj1 ,µ2 δµ j ≥ 0, (11)
where δµ j = µ̄0j − µ̄ j .
Proof: The direct part is immediate. Assume that the converse is not true, that is, there is a probability measure µ̄0j
such that:
Kµj1 ,µ2 (µ̄0j − µ̄ j ) < 0, Lµj1 ,µ2 (µ̄0j − µ̄ j ) = 0.
Then, since (µ1 , µ2 ) is not variance stationary, there is a probability measure µ̄00j such that Lµj1 ,µ2 (µ̄00j − µ̄ j ) < 0. Hence,
there is a θ ∈ (0, 1) such that:
Kµj1 ,µ2 (θµ̄0j + (1 − θ)µ̄00j − µ̄ j ) < 0, Lµj1 ,µ2 (θµ̄0j + (1 − θ)µ̄00j − µ̄ j ) < 0.
But this contradicts (11). 
Assume that VA (µ1 , µ2 ) = C and (µ1 , µ2 ) is not variance stationary. Then, (µ1 , µ2 ) is an equilibrium if and only if
there is no δµ j = µ̄0j − µ̄ j such that:
Kµj1 ,µ2 δµ j < 0 and Lµj1 ,µ2 δµ j < 0. (12)
The following proposition characterizes the variance non-stationary GNE in terms of measures supported on at most
two points.
Proposition 3. If there is a probability measure µ̄0j such that (12) holds true, then there is another probability measure
µ̄00j supported on at most two points which also satisfies (12) with δµ = µ̄00 − µ̄.
Proof: See Appendix C 
Let us introduce the following quantities:
j j
j,µ1 ,µ2 (u) = Kµ1 ,µ2 (du − µ̄ j ),
gK j,µ1 ,µ2 (u) = Lµ1 ,µ2 (du − µ̄ j ),
gL (13)
where du is a Dirac measure supported on u. Using these quantities we have the following necessary (Corollary 3),
and necessary and sufficient conditions (Corollary 4).
8
Corollary 3. If (µ1 , µ2 ) is a GNE then for all u ∈ [um , u M ], j = 1, 2 if gK
j,µ1 ,µ2 (u) < 0 then g j,µ1 ,µ2 (u) ≥ 0.
L

Corollary 4. A variance non-stationary pair (µ1 , µ2 ) is a GNE if and only if for all u0 , u00 ∈ [um , u M ], ρ ∈ [0, 1],
j = 1, 2 if ρgK
j,µ1 ,µ2 (u ) + (1 − ρ)g j,µ1 ,µ2 (u ) < 0 then ρg j,µ1 ,µ2 (u ) + (1 − ρ)g j,µ1 ,µ2 (u ) ≥ 0.
0 K 00 L 0 L 00

Remark 9. The proposed formulation describes pro-social behaviors in terms of bounding the variance of the costs.
There are various alternative formulations. For example, people may bound the maximum number of infected
individuals, reflecting the bounded capacity of the healthcare systems. Another alternative would be to consider
altruistic players [52]. Finally, pro-social behavior can be modeled as Kantian behavior [53]. We chose to model
pro-social behavior as bounding the variance, because of the vast health inequities created by the current COVID-19
pandemic.

5. Symmetric GNE of the Variance Constrained Game and Equilibrium Selection

We then focus on the class of symmetric GNE, where players of the same type use the same deterministic strategy.
That is:
µ j = n j du j , j = 1, 2.
In this case, the cost of the players of the same type is the same. With a slight abuse of notation, let us denote this cost
by J j (u1 , u2 ). The variance for symmetric strategies is given by:
n1 n2
V(u1 , u2 ) = (J2 − J1 )2 (14)
(n1 + n2 )2
Let us note that J1 (u1 , u2 ), J2 (u1 , u2 ), and V(u1 , u2 ) are continuously differentiable (see for example [54], Section 3.3).
In the rest of the section, assume that

Proposition 4. Let µ1 , µ2 be a pair of variance non-stationary symmetric strategies. The following are equivalent:
(i) µ1 , µ2 constitutes a GNE, with variance constraint V ≤ C = V(u1 , u2 ).
(ii) It holds:
j,µ1 ,µ2 (u), g j,µ1 ,µ2 (u)) : u ∈ [um , u M ]}) ∩ Q3 = ∅,
conv({(gK j = 1, 2
L o

where Qo3 is the interior of the third quadrant of the plane and ‘conv’ stands for the convex hull.

(iii) There exist ρ1 , ρ2 ∈ [0, 1] such that:

(1 − ρ j )gK
j,µ1 ,µ2 (u) + ρ j g j,µ1 ,µ2 (u) ≥ 0,
L
(15)

for all u ∈ [um , u M ].

j,µ1 ,µ2 (u), g j,µ1 ,µ2 (u)) : u ∈ [um , u M ]} does not intersect Q3 . Furthermore, if it intersects both the second
(iv) The set {(gK L o

and the fourth quadrants Q2 and Q4 it holds:

ϑ2j + ϑ4j ≤ π/2, j = 1, 2,

where the angles ϑ2j , ϑ4j are defined as:

  K  
  |g j,µ1 ,µ2 (u)|  
ϑ2j = sup  K L
 , j = 1, 2.
 
arctan : (g (u), g (u)) ∈ Q
 
j,µ1 ,µ2 j,µ1 ,µ2 2

 L
 
 
g (u)
 
j,µ1 ,µ2
 
  L  
j
  |g j,µ1 ,µ2 (u)|  
ϑ4 = sup  K L
, j = 1, 2.
 

arctan  K  : (g j,µ ,µ (u), g j,µ ,µ (u)) ∈ Q4 
1 2 1 2


 g (u)
j,µ1 ,µ2



9
Proof : (i) is equivalent to (ii): Observe that

j,µ1 ,µ2 (u), g j,µ1 ,µ2 (u)) : u ∈ [um , u M ]}) ∩ Q3 , ∅

conv({(gK L o

if and only if there are ρ j ∈ [0, 1] and u0 , u00 ∈ [um , u M ] such that:

ρ j gK
j,µ1 ,µ2 (u ) + (1 − ρ j )g j,µ1 ,µ2 (u ) < 0,
0 K 00

ρ j gL
j,µ1 ,µ2 (u ) + (1 − ρ j )g j,µ1 ,µ2 (u ) < 0.
0 L 00

Thus, (ii) is equivalent to the conditions of Corollary 4.

(ii) implies (iii): Fix a j ∈ 1, 2. Observe that both Qo3 and A = conv({(gK
j,µ1 ,µ2 (u), g j,µ1 ,µ2 (u)) : u ∈ [um , u M ]}) are
L

convex. Thus, there is a linear function:

f (x, y) = α1 x + α2 y + α3 ,
where not both of α1 , α2 are zero, such that

f (x, y) ≥ 0, for (x, y) ∈ A, and f (x, y) ≤ 0, for (x, y) ∈ Qo3 . (16)

It is not difficult to see that f (x, y) ≤ 0 for all (x, y) ∈ Qo3 implies that α1 , α2 ≥ 0 and α3 ≤ 0. Thus, both inequalities
in (16) hold true with α3 = 0. Furthermore, since α1 and α2 are not both zero (15) holds true with ρ j = α2 /(α1 + α2 ).
(iii) implies (ii): Obvious.
(iii) implies (iv): Refer to Figure 2. Inequality (15) implies that there is a line (the dashed line of the figure) such
that the set {(gK j,µ1 ,µ2 (u), g j,µ1 ,µ2 (u)) : u ∈ [um , u M ]} lies above this line. This implies that ϑ2 + π/2 + ϑ4 ≤ π, which in
L

turn implies (iv).

(1-ρ,ρ)

Figure 2: Part (iv) of Proposition 4 states that the graph of {(gKj,µ1 ,µ2 (u), g j,µ1 ,µ2 (u)) : u ∈ [um , u M ]} in the several quadrants, should lie above the
L

dashed line. Part (iv) states that the sum of ϑ2 and ϑ4 is less than a right angle.

(iv) implies (iii): Similar. 

There is often a continuum of GNE under variance constraints V ≤ C. Denote by GNEC the set of Generalized
Nash equilibria with constraint V ≤ C. We then examine the costs of the players on all GNEs in GNEC . It turns out
that under mild conditions, there is a GNE having the minimum cost for both types of players among all GNEs in
GNEC .

Proposition 5. Assume that the set VC = {(u1 , u2 ) ∈ [um , u M ]2 : V(u1 , u2 ) = C} is described by the image of a
smooth curve γ : [0, 1] → [um , u M ]2 , and that none of the points of VC is variance stationary. Then, there is a GNE
(uC,OPT
1 , uC,OPT
2 ) ∈ GNEC such that:

J1 (uC,OPT
1 , uC,OPT
2 ) ≤ J1 (u1 , u2 ), J2 (uC,OPT
1 , uC,OPT
2 ) ≤ J2 (u1 , u2 ),

for all (u1 , u2 ) ∈ GNEC .

10
Proof: (14) implies that |J1 − J2 | has a constant value on GNEC . Since VC is connected and V is continuous, J1 − J2
has a constant value on GNEC . In Appendix E, we show that GNEC is compact. Furthermore, J1 is continuous.
Thus, there is a pair (uC,OPT
1 , uC,OPT
2 ) minimizing J1 (u1 , u2 ) within the set of (u1 , u2 ) ∈ GNEC . But, since J1 − J2 has a
constant value on GNEC , the same pair (uC,OPT
1 , uC,OPT
2 ) minimizes the value of J2 as well. 
We then examine how the costs vary when C changes.

Definition 3. We call a pair of symmetric strategies µ j = n j du j , j = 1, 2 an interior GNE if:

(i) The values of u1 , u2 are not extreme, i.e., u1 , u2 ∈ (um , u M ).

(ii) There is a neighborhood U of u1 , u2 such that for any u01 , u02 ∈ U, there are constants ρ1 , ρ2 ∈ [0, 1) such that

(1 − ρ j )gK
j,µ0 ,µ0 (u) + ρ j g j,µ0 ,µ0 (u) ≥ 0,
L
1 2 1 2

where µ0j = n j du0j .

(iii) The gradient:

n1 n2
∇(u1 ,u2 ) V = 2 (J2 − J1 )∇(u1 ,u2 ) (J2 − J1 ),
(n1 + n2 )2
is non-zero.

(iv) There is a neighborhood of (u1 , u2 ) consisting of variance non-stationary strategies.

Remark 10. Observe that (0, 0) ∈ {(gK

j,µ0 ,µ0
(u), gL
j,µ0 ,µ0
(u)) : u ∈ [um , u M ]}. Indeed, for µ0j = n j du0j , it holds gK
j,µ0 ,µ0
(u0j ) =
1 2 1 2 1 2

Kµj1 ,µ2 (du0j − µ̄ j ) = 0. Similarly, gL

j,µ0 ,µ0
(u0j ) = 0. One way to assure that condition (ii) of Definition 3 holds is to show
1 2
that the curve (gKj,µ01 ,µ02
(u), gL
j,µ01 ,µ02
(u)) is strictly convex close (0, 0) and satisfies (15) as a strict inequality away from
(0, 0). For condition (iii), observe that ∇(u1 ,u2 ) V , 0 is equivalent to J1 , J2 and ∇(u1 ,u2 ) J2 , ∇(u1 ,u2 ) J1 .

Proposition 6. Let µ j = n j du j , j = 1, 2 be an interior GNE, with variance constraint V ≤ C. Then, all pairs of
strategies µ0j = n j du0j with u0j in a neighborhood of u j are GNE with a variance constraint V ≤ C 0 , for some C 0 in
a neighborhood of C. Furthermore, for each C 0 in a neighborhood of C, there is locally an one-parameter family of
GNEs with variance constraint V ≤ C 0 .

Proof : Since ∇(u1 ,u2 ) V is non-zero, either ∂V/∂u1 , 0 or ∂V/∂u2 , 0. Without loss of generality, assume that
∂V/∂u2 , 0. Consider the relation:
F(C, u1 , u2 ) = V(u1 , u2 ) − C = 0.
Then, since ∂V/∂u2 , 0, implicit function theorem applies, and there is a neighborhood U1 of (C, u1 ), with U1 ⊂ R2 ,
and a continuous function G : U1 → [um , u M ] such that F(C 0 , u01 , G(C 0 , u01 )) = 0, for all (C 0 , u01 ) ∈ U1 . Therefore,
since G is continuous, and µ j , j = 1, 2 is an interior GNE, there is a neighborhood U2 ⊂ U1 of (C, u1 ) such that (15)
holds for all u ∈ [um , u M ], and (C 0 , u01 ) ∈ U2 . Thus, for each C 0 in a neighborhood of C there is an one parameter
family of GNE given by (u01 , G(C 0 , u01 )). 

Definition 4. An interior GNE is called locally improvable if the Jacobian:

 ∂J1 ∂J1 
J =  ∂u ∂u2 
,
 
1
∂J2 ∂J2 
∂u1 ∂u2

is non-singular.

The condition of a non-singular Jacobian is a strengthening of ∇(u1 ,u2 ) J2 , ∇(u1 ,u2 ) J1 . In the following proposition
we show that in the neighborhood of a locally improvable GNE there is another GNE that is socially improved, that is
the cost for both types of players and the variance are lower.

11
Proposition 7. Let µ j = n j du j , j = 1, 2 be a locally improvable GNE with variance constraint V ≤ C. Then, there
is a C 0 < C and a pair of strategies µ0j = n j du0j which constitutes a GNE with variance constraint V ≤ C 0 , such that
J1 (u01 , u02 ) < J1 (u1 , u2 ) and J2 (u01 , u02 ) < J2 (u1 , u2 ).
Proof : Since µ j , j = 1, 2 is a locally improvable GNE, it is also an interior GNE. Thus, J1 (u1 , u2 ) , J2 (u1 , u2 ). Without
loss of generality assume that J2 (u1 , u2 ) > J1 (u1 , u2 ). Consider, the vectors v1 = ∇(u1 ,u2 ) J1 , v2 = ∇(u1 ,u2 ) J2 . Note that
the non-singularity of the Jacobian J implies that v1 , v2 are linearly independent. The gradient of the variance is given
by:
n1 n2
∇(u1 ,u2 ) V = 2 (J2 − J1 )(v2 − v1 ),
(n1 + n2 )2
We claim that there exists a direction w such that w · v1 , w · v2 , and w · (v2 − v1 ) are all negative. To contradict
assume that such a direction does not exist. Then, due to Gordan’s Theorem (see for example [55], Theorem 5.6.1),
there are nonnegative λ1 , λ2 , λ3 , not all of them zero such that:
λ1 v1 + λ2 v2 + λ3 (v2 − v1 ) = (λ1 + λ3 )v1 + (λ2 − λ3 )v2 = 0.
But, since v1 , v2 are linearly independent we should have λ1 + λ3 = 0, λ2 − λ3 = 0. But, since λ1 , λ2 , λ3 ≥ 0, we get
λ1 = λ2 = λ3 = 0, which is a contradiction. Thus, the claim is proved.
Finally, since µ j , j = 1, 2 is an interior GNE, the pair of strategies µ0j = n j u0j , j = 1, 2, with u0j = u j + εw j is a GNE
with variance constraint C 0 = V(u1 +εw1 , u1 +εw2 ). For small ε, it holds J1 (u01 , u02 ) < J1 (u1 , u2 ), J2 (u01 , u02 ) < J2 (u1 , u2 )
and C 0 < C. 
Remark 11. An alternative interpretation of Proposition 7 is that under the local improvement conditions, the
reduction of the variance constraint C may lead to improved performance for both types of players.
Remark 12. For non-interior GNE, we may determine the existence of local improvements as follows. Fix a pair of
strategies u1 , u2 , and assume that J1 (u1 , u2 ) > J2 (u1 , u2 ) and that det |J| , 0. Let us denote the gradients of J1 , J2 ,
and V by v1 , v2 , and v3 respectively. Observe that v1 belongs to the cone generated by v2 and v3 . Thus, the directions
where both costs J1 and J2 and the variance V decrease are described by the cone:
Cu1 ,u2 = (cone{v2 , v3 })? ,
where cone{v2 , v3 } is the cone generated by v2 , v3 and Cu1 ,u2 is its polar cone. This calculation gives a simple criterion
for determining that a GNE is Pareto dominated by another.
Remark 13. Local improvements can be used as a tool for equilibrium selection. Particularly, the most interesting
pairs of strategies are the GNE which are not locally improvable.

5.1. Equilibrium Selection

Assume that in the absence of a variance constraint, the players will use a mixed Nash strategy, described by ũ1 , ũ2
(see Section 3), and that the costs and variance of the costs are given by J˜1 , J˜2 , and Ṽ. If the variance constraint
constant C is given, then there is one optimal GNE (see proposition 5). We then present a simple bargaining-based
procedure to select both the constant C and a GNE with variance constraint V ≤ C.
Denote by GNE the set of all GNE for C < Ṽ. Assume that there is a Pareto improving point in GNE. That is,
there is a (u1 , u2 ) ∈ GNE such that J1 (u1 , u2 ) ≤ J˜1 , J2 (u1 , u2 ) ≤ J˜2 and V(u1 , u2 ) ≤ Ṽ, and that at least one of the
inequalities is strict. One way to select the a single GNE out of the set GNE it to use the following procedure:
(i) Determine the Pareto improving points:
GNEPI = {(u1 , u2 ) ∈ GNE : J1 (u1 , u2 ) ≤ J˜1 , J2 (u1 , u2 ) ≤ J˜2 }

(ii) Determine the equilibrium (u1 , u2 ) ∈ GNEPI that maximizes the quantity:
( J˜1 − J1 (u1 , u2 ))( J˜2 − J2 (u1 , u2 )).

Remark 14. The selection of GNE in (ii) is based on the Nash bargaining solution. Alternative bargaining solutions
can be also used (see for example [56]).
12
6. Computational Study
We then present some numerical results. In Subsection 6.1, we compute numerically the Nash equilibria of
the unconstrained game providing two illustrative examples and in Subsection 6.2, we study some example for the
variance constrained game2 .

6.1. Computing Unconstrained Nash Equilibria

The computation of the value of F(ū1 , ū2 ) corresponds to the numerical integration of (6). The search for pure
Nash equilibria needs just the computation of F(0, 1), F(1, 0), and F(1, 1) and checking the corresponding inequalities.
Let us then describe the procedure to find equilibria in the form of the (ũ1 , 0) or (ũ1 , 1). We have first to find the
solutions of:
Hũ2 (ũ1 ) = G1 (1 − I0 )(e−um F(ũ1 ,ũ2 ) − e−uM F(ũ1 ,ũ2 ) ) − (u M − um )(s j1 (um + (u M − um )ũ1 ) + s j2 ū2 ) = 0,
with respect to ũ1 , for a fixed value of ũ2 = 0, 1. To do so we use line search (an alternative, would be to use a
multi-start Newton algorithm). Having found a solution of Hũ2 (ũ1 ) = 0 for ũ2 = 0 or ũ2 = 1 we need also to check the
corresponding inequality. The computation of possible equilibria in the form (0, ũ2 ) or (1, ũ2 ) is similar.
Let us compute any possible Nash equilibrium where both types use mixed strategies (interior Nash equilibria).
Then we should have:
G1 (1 − I0 )(e−um F(ũ1 ,ũ2 ) − e−uM F(ũ1 ,ũ2 ) ) = (u M − um )(s11 ū1 + s12 ū2 )
(17)
G2 (1 − I0 )(e−um F(ũ1 ,ũ2 ) − e−uM F(ũ1 ,ũ2 ) ) = (u M − um )(s21 ū1 + s22 ū2 )
Any solution of this equation should belong to the line:
G2 (s11 ū1 + s12 ū2 ) = G1 (s21 ū1 + s22 ū2 )
or equivalently:
G1 (s21 + s22 ) − G2 (s11 + s12 )
(G2 s11 − G1 s21 )ũ1 + (G2 s12 − G1 s22 )ũ2 = um . (18)
u M − um
Therefore, to find any interior Nash equilibria, we examine using line search if there are solutions of (17) on the
line (18). We then present some examples with concrete values for the parameters.
Example 1: The parameters are T = 100, r = 5/16, I0 = 0.01, n1 = 0.8, n2 = 0.2, α1 = α2 = 1/8. These
parameters correspond to an epidemic with basic reproduction number R0 = 2.5, where people remain infectious for
a mean time of 8 days [30]. The s parameters are s11 = s21 = 2 and s12 = s22 = 0.5. We assume that u M = 0.8
and um = 0.5. We compute the equilibria for different values of G1 and G2 , assuming that G2 /G1 = 10. This choice
roughly corresponds to the infection fatality risks of the older people compared with the infection fatality risks of
younger people [36]. The variation described corresponds to varied ways that people may weight health, money and
well being.
The equilibria of the game are presented in Figure 3. We observe for all the values of G1 , G2 with G2 /G1 = 10
there is a unique Nash equilibrium. When G1 , G2 are small, the equilibrium strategies are ũ1 = ũ2 = 1. Then, as
G1 , G2 become larger, there is a mixed Nash equilibrium (1, ũ2 ). For intermediate values of G1 , G2 , there is a unique
equilibrium with ũ1 = 1 and ũ2 = 0. Finally for large G1 , G2 there is a unique equilibrium ũ1 = 0 and ũ2 = 0.
Example 2: In this example there is a strong homophily. Particularly, s11 = s22 = 2 and s21 = s12 = 0.5. The
rest of the parameters are as in Example 1, including the fact that G2 /G1 = 10. The equilibria for various values of
are presented in Figure 4. For low values of G1 there is a unique pure Nash equilibrium, where all the players play
u M . Then around G1 = 0.292, in addition to the pure equilibrium (1, 1), a pair of equilibria appears. One of the
new equilibra is pure and the other is mixed. In the new pure equilibrium all the players of type 1 play u M and all
the players of type 2 play um . In the mixed equilibrium, players of type 2 randomize, that is the mixed equilibrium
has the form (1, ũ2 ). As G1 becomes larger the value of ũ2 increases and eventually, around G1 = 0.363, the mixed
equilibrium (1, ũ2 ) meets with the pure equilibrium (1, 1) and they both disappear. Then, for G1 ∈ [0.363, 3.8] there
is a unique Nash equilibrium where all the players of type 1 play u M and players of type 2 play um . On the interval
G1 ∈ [3.8, 40.1], there is a unique mixed Nash equilibrium where all the players of type 2 play um and the players of
type 1 randomize. For larger values of G1 all the players play um .
2 The MATLAB code for the examples is available in GitHub https://github.com/jkordonis/EpidemicsGameTheoreticModelVarianceConstraints

13
Figure 3: The upper part of the image presents the equilibria when G1 ∈ [0.2, 0.5]. In this region, players of type 1 (non-vulnerable players) play
u = u M and players of type 2 randomize between um and u M . The probability to play u M denoted by ũ2 is illustrated in the upper part. Similarly
the lower part presents the equilibria for G1 ∈ [0.5, 50]. In this region, all the players of type 2 play u = um , while the players of type 1 randomize
with probability ũ1 .

1
u2

0.5

0
0.2 0.22 0.24 0.26 0.28 0.3 0.32 0.34 0.36 0.38 0.4
G1

1
u1

0.5

0
0 5 10 15 20 25 30 35 40 45 50
G1

Figure 4: The figure presents the equilibria of the game. The upper part present ũ2 , when G1 ∈ [0.2, 0.4]. In this interval all the players of type 1
play u M , that is ũ1 = 1. There are at most three equilibria. The equilibria for G1 ∈ [0.4, 50] are presented in the lower part of the figure. In this
region all the players of type 2 play u = um , while the players of type 1 randomize with probability ũ1 .

14
6.2. Computing Constrained Equilibria
We then search for generalized Nash equilibria with variance constraints. Let us first note that if the pair (µ1 , µ2 )
satisfies Definition 1.(i) then it also is an unconstrained Nash equilibrium. Thus, it is sufficient to check if the Nash
equilibria computed in the previous section satisfy the constraint V(µ1 , µ2 ) ≤ C.
We then compute GNE, satisfying Definition 1.(ii), in the case where the policies (µ1 , µ2 ) are of the form µ1 =
n1 du1 , µ2 = n2 du2 , where du is a Dirac measure concentrated on u. We first use a grid to find the pairs (u1 , u2 ) such
that the conditions of part (iii) of Proposition 4 are satisfied. The details of the computation of gK L
j,µ1 ,µ2 (u) and g j,µ1 ,µ2 (u)
are given in Appendix D. Then, we determine the intersection of the set of points where the conditions of Proposition
4.(iv) are satisfied with the set {(u1 , u2 ) : V(u1 , u2 ) = C}.
Example 3: The parameters in this example are similar to Example 1, except the vulnerability parameters which
are G1 = 8 and G2 = 20. The unconstrained Nash equilibrium is ũ1 = 0.602, ũ2 = 0. The costs of the players and
the variance are given by J1 = 0.185, J2 = 1.67 and V = 0.089. The set of points that satisfy the conditions of
part (iv) of Proposition 4 is illustrated in Figure 5. For each value of C less than variance of the unconstrained Nash
equilibrium, the set of GNE are with constraint V ≤ C is the intersection of the contour line V = C with the set of
points satisfying conditions (iv) of Proposition 4 (blue points of the figure). The sets of GNE for three different values
of C are illustrated in the figure.
Figure 5 shows also the optimal GNE (green line) for each C and the set of GNE which is not Pareto dominated
by others (light blue line). Let us note that the set of undominated GNE is a subset of the optimal GNE given C. We
observe that, for C > 0.0411, both types of players prefer a lower value of C, that is the

0.8
Sat. Conditions
V=0.02
0.75 GNE, V=0.02
V=0.04
GNE, V=0.02
V=0.04
0.7 GNE, V=0.06
Opt. GNE, diff. C
Undominated GNE
u2

0.65

0.6

0.55

0.5
0.5 0.55 0.6 0.65 0.7 0.75 0.8
u1

Figure 5: With blue color we present the points that satisfy Proposition 4.(iv). Each of these points is a GNE for a different value of C. . The figure
also presents the contour lines V = C, for various values of C and the corresponding equilibria. For C = 0.02, there is a continuum of GNE. For
C = 0.04, there is a continuum of GNE and an isolated GNE, while for C = 0.06, there is a unique GNE.

Observe that, for some values of C there is unique GNE, while for other values there is continuum of GNE. Note
that the set GNEC is not necessarily connected.
Figure 6 presents the Nash cost, as well as the costs of the variance constrained GNE, for various values of C.
Observe that the existence of a Variance constraint is beneficial for both types of players. Furthermore, starting for
large values of the constraint C the reduction of C leads to improved costs for both players. As C continues to

15
reduce, the cost for non-vulnerable players reaches its minimum, and increases again for a smaller value of C. On the
other hand, the cost of the vulnerable players decreases monotonically as the constraint becomes more stringent. The
smallest value of J1 is attained for C = 0.0411. In other words, with respect to the choice of C, the two types of players
have similar interests to reduce C, until C = 0.0411, but for lower values their relationship becomes competitive. The
GNE selected using the procedure of Subsection 5.1 has a variance constraint V ≤ 0.0328.

0.2
Nash Cost
0 Variance Constrained Cost
Barg. Chosen Solution
J Optimal
1
J1

-0.2

-0.4

-0.6
0 0.05 0.1 0.15
C

1.5 Nash Cost

Variance Constrained Cost
1 Barg. Chosen Solution
J1 Optimal
J2

0.5

0 0.02 0.04 0.06 0.08 0.1 0.12 0.14

C

Figure 6: The upper part of the image corresponds to the cost of non-vulnerable players and the lower part to vulnerable. With the dashed line
we present the cost of the players on the unconstrained Nash equilibrium. The red curves illustrate the cost of the players, when the players use
the optimal GNE, under the constraint V ≤ C, for several C. The yellow points correspond to the costs of the players on the equilibrium selected
according to the procedure described in Subsection 5.1. Finally, the purple points correspond to the GNE where non-vulnerable players have the
lowest cost. We present this point because, non-vulnerable players would willingly agree to decrease C up to this point.

Figure 7 illustrates the evolution of the epidemic for various values of C. We observe that, as the constraint
becomes more restrictive i.e., as C decreases the prevalence of the epidemic decreases as well.
Example 4: In this example the parameters are as in Example 1, and the vulnerability parameters G1 = 8 and
G2 = 80. The unconstrained Nash equilibrium is ũ1 = 0.602, ũ2 = 0. The fact that the Nash equilibrium is the same
with Example 3 is not surprising. Players of type 2 are now more vulnerable, and thus we expect them to continue
playing ũ2 = 0, and for the players of type 1 the game has not changed. The cost for the non-vulnerable and vulnerable
players under the Nash equilibrium and the variance are J1 = 0.185, J2 = 9.12, and V = 3.19 respectively. The set of
GNE, for various values of C is presented in Figure 8. Here situation is qualitatively different compared with Example
3. Particularly, there are not interior GNE and, for each C, there is at most one GNE.
The GNE under the constraint V ≤ C, for various values of C is illustrated in Figure 5. We observe that, as the
value of the constraint C becomes smaller the cost of the vulnerable players decreases monotonically. Furthermore,
compared to the unconstrained case, the cost of the non-vulnerable players under the variance constrained is improved
as well.

7. Conclusion

We analyzed social distancing games involving vulnerable and non-vulnerable populations of players,
characterized the Nash equilibria, and investigated how inequality constraints influence the epidemic spread and the
16
Figure 7: The evolution over time for the total number of susceptible and infected persons for the Nash equilibrium and the GNE for various values
of C. The Nash equilibrium behaviors (blue line) induce a larger number of infected people, while the variance constrained GNE behaviors cause
the epidemic to attenuate. Observe that stricter constraints lead to faster attenuation.

costs of the players. We also defined a Generalized Nash equilibrium (GNE) concept for non-atomic games with
variance constraints. We then focused on symmetric GNEs and characterized them in terms of single-point-supported
deviations. There is often a multitude of GNE. We proved that, for a given variance constraint, there is an optimal
equilibrium, which Pareto dominates all the other symmetric GNEs. We also provided sufficient conditions, under
which a more stringent constraint may lead to improved performance for all the players. We observed numerically
that if players are committed to satisfying the variance constraints, vulnerable players are better-off compared to the
unconstrained Nash equilibrium. The situation can be improved for the non-vulnerable players as well. Furthermore,
the existence of inequality constraints delays the spread of the epidemics and reduces its prevalence.
There are several directions for future research. First, the model can be generalized, including many classes of
players having different vulnerabilities, minimum actions, degrees, etc. Another direction is to use real data to check
the predictions of the model, the modeling of the dynamic response of players and the study of other applications of
the variance constrained games defined.

Appendix A. Proof of Proposition 1

Equation (1) can be written as:

ẋ1u = −rux1u · (Mx)

ẋ2u = −rux2u · (Mx)
ẋ3u = rux1u · (Mx) − α1 x3u , (A.1)
ẋ4u = rux2u · (Mx) − α1 x4u
ẋ5 = Mx

17
 0.8
dominated GNE
undominated GNE
V=0.6
0.75 V=0.7
V=0.8
V=1
0.7 V=1.2
V=1.4
u2

0.65

0.6

0.55

0.5
0.5 0.55 0.6 0.65 0.7 0.75
u1

Figure 8: The GNE for various values of C and the contour lines V = C, for various values of C. We observe that for each C, there is a unique
GNE. With red color are the un-dominated GNE and with blue the dominated. That is, for each of the blue GNE, both players prefer a lower C,
that is each blue GNE is Pareto dominated from some red GNE. On the other hand, for a red GNE, vulnerable players prefer a lower C, while
non-vulnerable a higher. Thus, each of the red GNE is not Pareto dominated by any other.

where M ∈ X ? with: Z Z
Mx = x3u uµ1 (du) + x4u uµ2 (du).

Note that kMk ≤ n1 + n2 . The initial conditions are x1u (0) = x2u (0) = 1 − I0 , x3u (0) = x4u (0) = I0 , for all u ∈ [um , u M ]
and x5 (0) = 0. In compact form we write ẋ = fµ1 ,µ2 (x).

Lemma 3. Any solution of (A.1) with the given initial conditions satisfies 0 ≤ x1u , . . . , x4u ≤ 1. Let us denote this set
by X0 i.e., X0 = {x ∈ X : 0 ≤ x1u , . . . , x4u ≤ 1, for all u}.

Proof : Consider such a solution. Observe that, x1u (t), x2u (t) ≥ 0 for all u and t. Similarly, since x ≥ 0 implies Mx ≥ 0
we have x3u (t), x4u (t) ≥ 0 for all t, u. Finally, ẋ1u + ẋ3u ≤ 0 and thus, x3u (t) ≤ x1u (0) + x3u (0) = 1. Similarly, x4u (t) ≤ 1.


Lemma 4. Let x(t) be a solution of (A.1) with the given initial conditions. Then, x ju (t) is continuous on u, for
j = 1, . . . , 4.

Proof : Let h(t) = Mx(t). Then, 0 ≤ h(t) ≤ n1 + n2 . The solution of ẋ1u = −ruh(t)x1u given by:
Z t !
x1u (t) = (1 − I0 ) exp −ru h(s)ds
0

depends continuously on u, t. A similar argument shows that x2u is continuous on u, t. Then, observe that the third
equation of (A.1) can be written as:
ẋ3u = b(u, t) − α1 x3u ,

18
 0.2
Nash Cost
0 Variance Constrained Cost
Barg. Chosen Solution
J1 Optimal
J1

-0.2

-0.4

-0.6
0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 1.3 1.4 1.5
C

10

Nash Cost
Variance Constrained Cost
J2

5 Barg. Chosen Solution

J Optimal
1

0
0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 1.3 1.4 1.5
C

Figure 9: The dashed line corresponds to the Nash cost. The red lines represent the costs for vulnerable and non-vulnerable players for various
values of C. The yellow points correspond to the costs of the players on the equilibrium selected according to the procedure described in Subsection
5.1. Finally, the purple points correspond to the GNE where non-vulnerable players have the lowest cost.

where b(u, t) = rux1u · Mx is continuous on u and bounded by u M r(n1 + n2 )(1 − I0 ). The solution of this differential
equation is: Z t
x3u (t) = x3u (0)e−α1 t + eα1 (τ−t) b(u, τ)dt,
0
which is again continuous on u, t. A similar argument shows that x4u is continuous on u, t. 
We then proceed to the proof of the proposition. Consider a saturated version of (A.1):

ẋ1u = −rusat1 (x1u ) · sat2 (Mx)

ẋ2u = −rusat1 (x2u ) · sat2 (Mx)
ẋ3u = rusat1 (x1u ) · sat2 (Mx) − α1 x3u , (A.2)
ẋ4u = rusat1 (x2u ) · sat2 (Mx) − α2 x4u
ẋ5 = Mx

where sat1 (z) = max{min{z, 1}, 0} and sat2 (z) = max{min{z, n1 + n2 }, 0}. Due to Lemma 3, any solution of (A.1), with
the given initial conditions, is also a solution of the modified system. Denote this system in compact form as ẋ = f˜(x).
It is not difficult to see that f˜ : X → X is Lipschitz with constant:

L = 2ru M (n1 + n2 ) + α1 + α2 .

Thus, Theorem 7.3 of [57] implies that there exists a unique solution within the space C([0, T ], X) of continuous
functions with values in X. Lemma 4 implies that there is no solution of (A.1) not belonging to C([0, T ], X). 

Remark 15. Note that f˜ is uniformly Lipschitz. The constant is uniform in µ1 , µ2 , for positive measures of total mass
n1 , n2 .

19
Appendix B. Proof of Lemma 1

We then proceed to the proof Lemma 1 and the computation of a formula for the directional derivative of the
variance. To this end, we first consider the sensitivity of the solution of (1) with respect to the deviation εδµ j .
We first compute the directional derivative of fµ1 ,µ2 (x) in the direction δµ. The value of fµ1 ,µ2 (x) depends on µ j
through the quantity I f (t). This quantity can be written as:

I f (t) = Tµ j I j· (t) + Tµ− j I− j· (t),

where I j· (t) : [um , u M ] → R with u 7→ I ju (t) and Tµ ∈ (C([um , u M ], R))? with:

Z
Tµ I j· (t) = I ju0 (t)u0 · µ(du0 ). (B.1)

The directional derivative of fµ1 ,µ2 (x) in the direction δµ j is:

Dδµ j [ fµ1 ,µ2 (x)] =

[−ruS ju Tδµ j I j· , − ruS (− j)u Tδµ j I j· , ruS ju Tδµ j I j· , ruS (− j)u Tδµ j I j· , Tδµ j I j· ]T .

Observe that, due to the form of Tδµ j in (B.1), there is a bounded linear operator Bµ1 ,µ2 (t) : Y → X such that:

Dδµ j [ fµ1 ,µ2 (x(t))] = Bµ1 ,µ2 (t)δµ j .

The linearized version of (1), around the trajectories (S 1u , S 2u , I1u , I2u ), is given by:
0
ẋ1u = −(ruI f )x1u
0 0
− ruS 1u (Tµ1 x3· + Tµ2 x4·
0
) − ruS 1u Tδµ j I j·
0
ẋ2u = −(ruI f )x2u
0 0
− ruS 2u (Tµ1 x3· + Tµ2 x4·
0
) − ruS 2u Tδµ j I j·
0
ẋ3u = (ruI f )x1u
0
+ ruS 1u (Tµ1 x3·
0
+ Tµ2 x4·
0
) − α1 x0 3u + ruS 1u Tδµ j I j· , (B.2)
0
ẋ4u = (ruI f )x2u
0
+ ruS 2u (Tµ1 x3·
0
+ Tµ2 x4·
0
) − α2 x0 4u + ruS 2u Tδµ j I j·
ẋ50 = Tµ1 x3u
0
+ Tµ2 x4u
0
+ Tδµ j I j·

where I f = Iµf1 ,µ2 as in (2). Thus, (B.2) is a LTV system in the form:

ẋ0 = Aµ1 ,µ2 (t)x0 + Bµ1 ,µ2 (t)δµ j . (B.3)

Lemma 5. Denote by φµ1 ,µ2 (t) the solution of (1). Then, the directional derivatives xlu
0
= Dδµ j φlu
µ1 ,µ2 (t), for l = 1, . . . , 4
and x5 = Dδµ j φµ1 ,µ2 (t) satisfy (B.3), with zero initial conditions.
0 5

Proof: Consider the function: f¯ : Y × X → X with:

f¯(µ̄, x) = fµ̄,µ− j (x)

and the set D = X0 × {µ ∈ Y : kµk = n j }

The function f¯(µ̄, x) is continuously Fréchet differentiable with respect to x. Its derivative is the operator Aµ1 ,µ2 (t)
given in (B.2)–(B.3), which is continuous and bounded in D. The directional derivative:

Dδµ j [ f¯(µ̄, x)] = Bµ1 ,µ2 (t)δµ j ,

is also continuous in x and bounded in D. Furthermore, X0 is positively invariant. Thus, Theorem 1 of [58] applies
and the proof is complete. 
Therefore, F is continuous in µ j and J j (u0 , µ j , µ− j ) is continuous in µ j , uniformly in u0 . This proves (9).
var
Lemma 6. There is a continuous function f j,F,µ 1 ,µ2
(u) such that the directional derivative of F(µ1 , µ2 ) on the δµ j
direction is given by: Z
Dδµ j F(µ1 , µ2 ) = rx50 (T ) = var
f j,F,µ1 ,µ2
(u0 )δµ j (du0 ).

20
Proof: Observe that Bµ1 ,µ2 (t) : Y → X can be written as a composition of two linear operators Bµ11 ,µ2 (t) : Y → R and
Bµ21 ,µ2 (t) : R → X with: Bµ11 ,µ2 (δµ) = Tδµ j I j· and
h iT
Bµ21 ,µ2 (l) = −rux1u l −rux2u l rux1u l rux2u l l .

The solution of the LTV system (B.3) with zero initial conditions is written in the form (e.g. paragraph 1.4 of
[59]): Z T
x (T ) =
0
Ψ(T, t)Bµ1 ,µ2 (t)δµ j dt
0
where Ψ(t, s) : X → X is the state transition operator.
Denote by C : X → R the operator picking the last component i.e., C(x0 ) = x50 . Then, x50 (T ) can be written as:
Z T
x50 (T ) = [CΨ(T, t)Bµ21 ,µ2 (t)]Bµ11 ,µ2 (t)δµ j dt.
0

Note that ψt = CΨ(T, t)Bµ21 ,µ2 is a linear function ψt : R → R. Write this function as ψt (l) = ϕ(t) · l. Thus:
Z T Z Z " Z T #
x50 (T ) = ϕ(t) ux j+2,u (t)δµ j (du)dt = u ϕ(t)x j+2,u (t)dt δµ j (du)
0 0

Define: Z T
var
f j,F,µ 1 ,µ2
(u) = ru ϕ(t)x j+2,u (t)dt
0
var
Observe that, since x j+2,u (t) is continuous with respect to u, the function f j,F,µ1 ,µ2
(u) is also continuous in u. Thus:
Z
Dδµ j F(µ1 , µ2 ) = rx50 (T ) = var
f j,F,µ 1 ,µ2
(u)δµ j (du),

and the proof is complete 

The directional derivative of ū j is given by:
Z
Dδµ j ū j = u0 δµ j (du0 )/n j .

The directional derivative of the cost of a player of type j0 who uses an action u is given by:

Dδµ j J j0 (u, µ j , µ− j ) = G j0 (1 − I0 )ue−uF Dδµ j F − s j0 j uDδµ j ū j

Z
= var
f j,J j0 ,µ1 ,µ2
(u0 , u)δµ j (du0 ), (B.4)

where:
var
f j,J j0 ,µ1 ,µ2
(u0 , u) = G j0 (1 − I0 )ue−uF f j,F,µ
var
1 ,µ2
(u0 ) − s j0 j uu0 /n j ,
is a continuous function of u0 , u. The directional derivative of the mean cost of the players of type j (the same with
the type of the deviating players) is given by:
" "Z Z ##
1
Dδµ j J¯j = lim J j (u, µ j + εδµ j , µ− j )(µ j + εδµ j )(du) − J j (u, µ j , µ− j )µ j (du)
ε→0 εn j
Z " Z #
1
= Dδµ j J j (u, µ j , µ− j )µ̄ j (du) + lim J j (u, µ j + εδµ j , µ− j )δµ j (du)
ε→0 n j
Z Z
1
= Dδµ j J j (u, µ j , µ− j )µ̄ j (du) + J j (u, µ j , µ− j )δµ j (du).
nj

21
Substituting (B.4) into the last equation and using Fubini’s theorem, we get:
Z " Z #
1
Dδµ j J j =
¯ J j (u, µ j , µ− j ) + f J j ,µ1 ,µ2 (u , u)µ̄ j (du) δµ j (du0 )
var 0
nj
Z
= f j,var
J¯j ,µ1 ,µ2
(u0 )δµ j (du0 ) (B.5)

Similarly: Z "Z # Z
Dδµ j J¯− j = var
f j,J− j ,µ1 ,µ2
(u0
, u)µ̄ −j (du) δµ j (du 0
) = f j,var
J¯− j ,µ1 ,µ2
(u0 )δµ j (du0 )

Combining the last two equations we compute the variation of the mean cost:
Z
Dδµ j J¯ = (n1 Dδµ j J¯1 + n2 Dδµ j J¯2 )/(n1 + n2 ) = f j,var 0 0
¯ 1 ,µ2 (u )δµ j (du ).
J,µ

We then proceed to the computation of the directional derivative of the variance in two steps. The first part is:
"Z # "Z
Dδµ j (J j (u, µ j , µ− j ) − J) ¯ 2 µ j (du) = lim 1 (J j (u, µ j + εδµ j , µ− j ) − J(µ
¯ j + εδµ j , µ− j ))2 (µ j + εδµ j )(du)−
ε→0 ε
Z #
− (J j (u, µ j , µ− j ) − J(µ ¯ j , µ− j ))2 µ j (du)
Z Z Z
= (J j (u, µ j , µ− j ) − J) ¯ 2 δµ j (du) + 2 (J j (u, µ j , µ− j ) − J(µ ¯ j , µ− j )) [ f j,J
var
j ,µ1 ,µ2
(u0 , u) − f j,var 0 0
¯ 1 ,µ2 (u )]δµ j (du )µ j (du)
J,µ
Z
= var
f j,V j ,µ1 ,µ2
(u0 )δµ j (du0 ).

In the second equality the interchange of the integral with the limit is possible, because the convergence
limε→0 [J j (u, µ j + εδµ j , µ− j ) − J j (u, µ j , µ− j )]/ε is uniform in u. In the last equality we use again Fubini’s theorem.
var
Note that f j,V j ,µ1 ,µ2
(u0 ) is continuous in u0 .
Similarly:
"Z #
Dδµ j (J− j (u, µ j , µ− j ) − J) ¯ 2 µ j (du) =
Z Z
= 2 (J− j (u, µ j , µ− j ) − J(µ j , µ− j )) [ f j,J
¯ var
− j ,µ1 ,µ2
(u0 , u) − f j,var 0 0 0
¯ 1 ,µ2 (u )]δµ j (du )µ j (du )
J,µ
Z
= var
f j,V − j ,µ1 ,µ2
(u0 )δµ j (du0 )

Therefore, the directional derivative of the variance is written as:

Z
Dδµ j V(µ j , µ− j ) = var
f j,V,µ1 ,µ2
(u0 )δµ j (du0 ),

var
for a continuous function f j,V,µ1 ,µ2
(u0 ).

Appendix C. Proof of Proposition 3

Let us drop the dependence on j. Inequalities (12), recalling that µ̄0 is a probability measure, can be written as:
Z
f¯K (u0 )µ̄0 (du0 ) < 0,
Z
f¯L (u0 )µ̄0 (du0 ) < 0,

22
where f¯K (u0 ) = J j (u0 , µ1 , µ2 ) − J j (u0 , µ1 , µ2 )µ̄(du0 ), f¯L (u0 ) = f j,V (u0 )µ̄(du0 ).
R R
var
j ,µ1 ,µ2
(u0 ) − f j,V
var
j ,µ1 ,µ2

Using the (uniform) continuity of f¯K , f¯K we have that for a set of u1 , . . . , uN and c1 , . . . , cN > 0 it holds:

c1 f¯K (u1 ) + · · · + cN f¯K (uN ) < 0,

c1 f¯L (u1 ) + · · · + cN f¯L (uN ) < 0,

If there is a point uk where f¯K (uk ) < 0 and f¯L (uk ) ≤ 0 or f¯K (uk ) ≤ 0 and f¯L (uk ) < 0 we are done. Thus, assume
that there is no such point. Excluding from the summation the terms where both f¯K (uk ) and f¯L (uk ) are non-negative
the inequalities still hold true. Thus, assume that for any point uk either f¯K (uk ) < 0 < f¯L (uk ) or f¯L (uk ) < 0 < f¯K (uk ).
Reordering the terms, the inequalities can be written as:

x1 + · · · + xn − y1 − · · · − ym < 0
(C.1)
−z1 − · · · − zn + w1 + · · · + wm < 0

where

xi = cki f¯K (uki ), ki ∈ {k : ck f¯K (uk ) > 0},

yi = |cki f¯K (uki )|, ki ∈ {k : ck f¯K (uk ) < 0},
wi = cki f¯L (uki ), ki ∈ {k : ck f¯L (uk ) > 0},
zi = |cki f¯L (uki )|, ki ∈ {k : ck f¯L (uk ) < 0}.

Let i0 be such that yi0 /wi0 ≥ yi /wi for all i and denote α = yi0 /wi0 .
Claim 1: There is a j0 such that αz j0 > x j0 . Indeed if this in not true, and αz j ≤ x j for all j, then multiplying the
second relationship with α we get:

−αz1 − · · · − αzn + αw1 + · · · + αwm ≥ −x1 − · · · − xn + y1 + · · · + ym > 0.

This completes the proof of the claim.

Claim 2: We may choose a ρ ∈ [0, 1] such that:

ρx j0 − (1 − ρ)yi0 < 0

−ρz j0 + (1 − ρ)wi0 < 0

Indeed for ρ in the interval:
yi0 yi0
<ρ< ,
αz j0 + yi0 x j0 + y i0
both inequalities are valid. The interval is not empty, since αz j0 > x j0 .
Choose such a ρ. Then, using the form of x, y, z, w, there are k1 , k2 such that:

ρck1 f¯K (uk1 ) + (1 − ρ)ck2 f¯K (uk2 ) < 0,

ρck1 f¯L (uk1 ) + (1 − ρ)ck2 f¯K (uk2 ) < 0,

Thus, taking
ρck1 (1 − ρ)ck2
µ̄00 = duk1 + du ,
ρck1 + (1 − ρ)ck2 ρck1 + (1 − ρ)ck2 k2
where du is the Dirac measure at u, and the proof is complete.

23
Appendix D. Computation of GNE

Consider a discrete set Ud ⊂ [um , u M ] and a grid Ud × Ud of pairs (u1 , u2 ). We first find the set of pairs (u1 , u2 ) on
the grid, such that V(n1 du1 , n2 du2 ) = C approximately holds. The dynamics becomes:

Ṡ 1u1 = −ru1 I f S 1u1 , Ṡ 2u2 = −ru2 I f S 2u2 ,

I˙1u1 = ru1 I f S 1u1 − α1 I1u1 , I˙2u2 = ru2 I f S 2u2 − α2 I2u2 , (D.1)
ż = I ,f

where I f (t) = u1 n1 I1u1 (t) + u2 n2 I2u2 (t).

For every point on the grid that V(n1 du1 , n2 du2 ) = C approximately holds, we compute the function gK
j,µ1 ,µ2 (u), for
u ∈ U, as:
gKj,µ1 ,µ2 (u) = J j (u, du1 , du2 ) − J j (u j , du1 , du2 ).

For j = 1, 2 and a grid of values of u ∈ U, we solve the pair of differential equations:

Ṡ 1u = −ruI f S 1u , I˙1u = ruI f S 1u − α j I1u

where I f is the quantity computed during the solution of (D.1).

Consider the variation δµ j = du − du j The linearized system around the solution of (D.1) consists of 5 differential
equations:
0
ẋ1u 1
= −(ru1 I f )x1u
0
1
0
− ru1 S 1u1 (n1 u1 x3u 1
+ n2 u2 x4u
0
2
) − ru1 S 1u1 (I ju − I ju j )
0
ẋ2u 2
= −(ruI f )x2u
0
2
0
− ru2 S 2u2 (n1 u1 x3u 1
+ n2 u2 x4u
0
2
) − ru2 S 2u2 (I ju − I ju j )
0
ẋ3u 1
= (ru1 I f )x1u
0
1
+ ru1 S 1u1 (n1 u1 x3u
0
1
+ n2 u2 x4u
0
2
) − α1 x0 3u1 + ru1 S 1u1 (I ju − I ju j ),
0
ẋ4u 2
= (ru2 I f )x2u
0
2
+ ru2 S 2u2 (n1 u1 x3u
0
1
+ n2 u2 x4u
0
2
) − α2 x0 4u2 + ru2 S 2u2 (I ju − I ju j )
ẋ50 = n1 u1 x3u
0
1
+ n2 u2 x4u
0
2
+ I ju − I ju j
with zero initial conditions. It holds Dδµ j F(µ1 , µ2 ) = x50 (T ), and Dδµ j ū j = (u − u j )/n j . The directional derivative
Dδµ j J j0 (u, µ j , µ− j ) is given in (B.4) and the directional derivative Dδµ j J¯j by:

1
Dδµ j J¯j = Dδµ j J j (u j , µ j , µ− j ) + (J j (u, µ j , µ− j ) − J j (u j , µ j , µ− j )).
nj

Furthermore Dδµ j J¯− j = Dδµ j J− j (u− j , µ j , µ− j ) and Dδµ j J¯ = (n1 Dδµ j J¯1 + n2 Dδµ j J¯2 )/(n1 + n2 ). The directional derivative
of the variance can be written as:
Dδµ j V = (Dδµ j V1 + Dδµ j V2 )/(n1 + n2 ),
where
"Z #
Dδµ j V1 = Dδµ j (J j (u , µ j , µ− j ) − J)
0 ¯ 2 µ j (du0 )

= (J j (u, µ j , µ− j ) − J)¯ 2 − (J j (u j , µ j , µ− j ) − J)
¯ 2+
h i
+ 2n j (J j (u j , µ j , µ− j ) − J(µ
¯ j , µ− j )) Dδµ j J j (u j , µ j , µ− j ) − Dδµ j J¯ ,

and:
"Z #
Dδµ j V2 = Dδµ j (J− j (u , µ j , µ− j ) − J) µ− j (du )
0 ¯ 2 0

h i
= 2n− j (J− j (u− j , µ j , µ− j ) − J(µ
¯ j , µ− j )) Dδµ j J− j (u− j , µ j , µ− j ) − Dδµ j J¯

j,µ1 ,µ2 (u), for u ∈ U and use Corollary 4 to check if u1 , u2 is a GNE.

We then compute the values of gL

24
Appendix E. Compactness of GNEC

We show that the set of points (u1 , u2 ) ∈ VC satisfying Proposition 4.(ii) is compact. Assume that a sequence
(uν1 , uν2 ) belongs to this set and that (ū1 , ū2 ) is its limit. Due to the continuity of γ, it holds (ū1 , ū2 ) ∈ VC .
Define the functions ḡK j (u, u1 , u2 ) = g j,µ1 ,µ2 (u), ḡ j (u, u1 , u2 ) = g j,µ1 ,µ2 (u), for µ j = n j du j and observe that, due to
K L L

(9), (10), (13), these functions are continuous in all their arguments.
If (ū1 , ū2 ) does not satisfy Proposition 4.(ii), then there is a pair of points u, u0 and a ρ ∈ (0, 1) such that:

ρḡK
j (u, ū1 , ū2 ) + (1 − ρ)ḡ j (u , ū1 , ū2 ) < 0,
K 0

ρḡL
j (u, ū1 , ū2 ) + (1 − ρ)ḡ j (u , ū1 , ū2 ) < 0,
L 0

ν ν
for j = 1 or j = 2. But, since ḡK
j (u, u1 , u2 ), ḡ j (u, u1 , u2 ) are continuous the same inequality holds also for (u1 , u2 ), for
L

large ν, which is a contradiction.

References
[1] N. Ferguson, D. Laydon, G. Nedjati Gilani, N. Imai, K. Ainslie, M. Baguelin, S. Bhatia, A. Boonyasiri, Z. Cucunuba Perez,
G. Cuomo-Dannenburg, et al., Report 9: Impact of non-pharmaceutical interventions (npis) to reduce covid19 mortality and healthcare
demand (2020).
[2] W. O. Kermack, A. G. McKendrick, A contribution to the mathematical theory of epidemics, Proceedings of the royal society of london.
Series A, Containing papers of a mathematical and physical character 115 (772) (1927) 700–721.
[3] R. Ross, An application of the theory of probabilities to the study of a priori pathometry, Proceedings of the Royal Society of London. Series
A, Containing papers of a mathematical and physical character 92 (638) (1916) 204–230.
[4] L. J. Allen, F. Brauer, P. Van den Driessche, J. Wu, Mathematical epidemiology, Vol. 1945, Springer, 2008.
[5] R. Pastor-Satorras, C. Castellano, P. Van Mieghem, A. Vespignani, Epidemic processes in complex networks, Reviews of modern physics
87 (3) (2015) 925.
[6] H. Zhang, J. Zhang, C. Zhou, M. Small, B. Wang, Hub nodes inhibit the outbreak of epidemic under voluntary vaccination, New Journal of
Physics 12 (2) (2010) 023015.
[7] S. L. Chang, M. Piraveenan, M. Prokopenko, Impact of network assortativity on epidemic and vaccination behaviour, arXiv preprint
arXiv:2001.01852 (2020).
[8] C. T. Bauch, D. J. Earn, Vaccination and the theory of games, Proceedings of the National Academy of Sciences 101 (36) (2004) 13391–13394.
[9] C. T. Bauch, A. P. Galvani, D. J. Earn, Group interest versus self-interest in smallpox vaccination policy, Proceedings of the National Academy
of Sciences 100 (18) (2003) 10564–10567.
[10] T. C. Reluga, C. T. Bauch, A. P. Galvani, Evolving public perceptions and stability in vaccine uptake, Mathematical biosciences 204 (2)
(2006) 185–198.
[11] T. C. Reluga, A. P. Galvani, A general approach for population games with application to vaccination, Mathematical biosciences 230 (2)
(2011) 67–78.
[12] H. Zhang, F. Fu, W. Zhang, B. Wang, Rational behavior is a ‘double-edged sword’when considering voluntary vaccination, Physica A:
Statistical Mechanics and its Applications 391 (20) (2012) 4807–4815.
[13] P. E. M. Fine, J. A. Clarkson, Individual versus public priorities in the determination of optimal vaccination policies, American journal of
epidemiology 124 (6) (1986) 1012–1020.
[14] M. Kremer, Integrating behavioral choice into epidemiological models of aids, The Quarterly Journal of Economics 111 (2) (1996) 549–573.
[15] R. Vardavas, R. Breban, S. Blower, Can influenza epidemics be prevented by voluntary vaccination?, PLoS computational biology 3 (5)
(2007).
[16] S. Del Valle, H. Hethcote, J. M. Hyman, C. Castillo-Chavez, Effects of behavioral changes in a smallpox attack model, Mathematical
biosciences 195 (2) (2005) 228–251.
[17] F. H. Chen, Rational behavioral response and the transmission of stds, Theoretical population biology 66 (4) (2004) 307–316.
[18] S. Funk, M. Salathé, V. A. Jansen, Modelling the influence of human behaviour on the spread of infectious diseases: a review, Journal of the
Royal Society Interface 7 (50) (2010) 1247–1256.
[19] S. Funk, E. Gilad, C. Watkins, V. A. Jansen, The spread of awareness and its impact on epidemic outbreaks, Proceedings of the National
Academy of Sciences 106 (16) (2009) 6872–6877.
[20] F. H. Chen, Modeling the effect of information quality on risk behavior change and the transmission of infectious diseases, Mathematical
biosciences 217 (2) (2009) 125–133.
[21] A. d’Onofrio, P. Manfredi, Information-related changes in contact patterns may trigger oscillations in the endemic prevalence of infectious
diseases, Journal of Theoretical Biology 256 (3) (2009) 473–478.
[22] G. Theodorakopoulos, J.-Y. Le Boudec, J. S. Baras, Selfish response to epidemic propagation, IEEE Transactions on Automatic Control
58 (2) (2012) 363–376.
[23] S. Trajanovski, Y. Hayel, E. Altman, H. Wang, P. Van Mieghem, Decentralized protection strategies against sis epidemics in networks, IEEE
Transactions on Control of Network Systems 2 (4) (2015) 406–419.
[24] A. R. Hota, S. Sundaram, Game-theoretic vaccination against networked sis epidemics and impacts of human decision-making, IEEE
Transactions on Control of Network Systems 6 (4) (2019) 1461–1472.

25
[25] Y. Huang, Q. Zhu, A differential game approach to decentralized virus-resistant weight adaptation policy over complex networks, IEEE
Transactions on Control of Network Systems 7 (2) (2019) 944–955.
[26] F. Toxvaerd, Equilibrium social distancing, Cambridge working papers in Economics (2020).
[27] C.-J. Karlsson, J. Rowlett, Decisions and disease: a mechanism for the evolution of cooperation, Scientific Reports 10 (1) (2020) 1–9.
[28] M. A. Amaral, M. M. de Oliveira, M. A. Javarone, An epidemiological model with voluntary quarantine strategies governed by evolutionary
game dynamics, arXiv preprint arXiv:2008.05979 (2020).
[29] M. Ye, L. Zino, A. Rizzo, M. Cao, Modelling epidemic dynamics under collective decision making, arXiv preprint arXiv:2008.01971 (2020).
[30] K. A. Kabir, J. Tanimoto, Evolutionary game theory modelling to represent the behavioural dynamics of economic shutdowns and shield
immunity in the covid-19 pandemic, Royal Society open science 7 (9) (2021) 201095.
[31] A.-R. Lagos, I. Kordonis, G. Papavassilopoulos, Games of social distancing during an epidemic: Local vs statistical information, arXiv
preprint arXiv:2007.05185 (2020).
[32] M. van Boven, D. Klinkenberg, I. Pen, F. J. Weissing, H. Heesterbeek, Self-interest versus group-interest in antiviral control, PLoS One 3 (2)
(2008).
[33] G. Hardin, The tragedy of the commons, Science 162 (1968) 1243–1248.
[34] L. Alfaro, E. Faia, N. Lamersdorf, F. Saidi, Social interactions in pandemics: fear, altruism, and reciprocity, Tech. rep., National Bureau of
Economic Research (2020).
[35] P. N. N. Brown, B. Collins, C. Hill, G. Barboza, L. Hines, Individual altruism cannot overcome congestion effects in a global pandemic game,
arXiv preprint arXiv:2103.14538 (2020).
[36] A. T. Levin, W. P. Hanage, N. Owusu-Boaitey, K. B. Cochran, S. P. Walsh, G. Meyerowitz-Katz, Assessing the age specificity of infection
fatality rates for covid-19: systematic review, meta-analysis, and public policy implications, European journal of epidemiology (2020) 1–16.
[37] C. Hoffmann, E. Wolf, Older age groups and country-specific case fatality rates of covid-19 in europe, usa and canada, Infection 49 (1) (2021)
111–116.
[38] E. Fehr, K. M. Schmidt, A theory of fairness, competition, and cooperation, The quarterly journal of economics 114 (3) (1999) 817–868.
[39] J. H. Fowler, T. Johnson, O. Smirnov, Egalitarian motive and altruistic punishment, Nature 433 (7021) (2005) E1–E1.
[40] M. Robson, M. Asaria, R. Cookson, A. Tsuchiya, S. Ali, Eliciting the level of health inequality aversion in england, Health economics 26 (10)
(2017) 1328–1334.
[41] P. D. Lunn, S. Timmons, C. A. Belton, M. Barjaková, H. Julienne, C. Lavin, Motivating social distancing during the covid-19 pandemic: An
online experiment, Social Science & Medicine 265 (2020) 113478.
[42] T. C. Reluga, Game theory of social distancing in response to an epidemic, PLoS computational biology 6 (5) (2010).
[43] P. Poletti, M. Ajelli, S. Merler, Risk perception and effectiveness of uncoordinated behavioral responses in an emerging epidemic,
Mathematical Biosciences 238 (2) (2012) 80–89.
[44] P. Poletti, B. Caprile, M. Ajelli, A. Pugliese, S. Merler, Spontaneous behavioural changes in response to epidemics, Journal of theoretical
biology 260 (1) (2009) 31–40.
[45] F. Facchinei, C. Kanzow, Generalised nash equilibrium problems, Ann. Oper. Res. 175 (2010) 177–211.
[46] F. Facchinei, J.-S. Pang, Finite-dimensional variational inequalities and complementarity problems, Springer Science & Business Media,
2007.
[47] D. Paccagnan, B. Gentile, F. Parise, M. Kamgarpour, J. Lygeros, Nash and wardrop equilibria in aggregative games with coupling constraints,
IEEE Transactions on Automatic Control 64 (4) (2018) 1373–1388.
[48] P. Jacquot, C. Wan, Nonsmooth aggregative games with coupling constraints and infinitely many classes of players, arXiv preprint
arXiv:1806.06230 (2018).
[49] P. Jacquot, C. Wan, Nonatomic aggregative games with infinitely many types, arXiv preprint arXiv:1906.01986 (2019).
[50] R. Singh, A. Wiszniewska-Matyszkiel, et al., Linear quadratic game of exploitation of common renewable resources with inherent constraints,
Topological Methods in Nonlinear Analysis 51 (1) (2018) 23–54.
[51] A. Mas-Colell, On a theorem of Schmeidler, Journal of Mathematical Economics 13 (3) (1984) 201–206.
[52] E. Shim, G. B. Chapman, J. P. Townsend, A. P. Galvani, The influence of altruism on influenza vaccination decisions, Journal of The Royal
Society Interface 9 (74) (2012) 2234–2243.
[53] I. Kordonis, A model for partial kantian cooperation, in: Advances in Dynamic Games, Springer, 2020, pp. 317–346.
[54] H. K. Khalil, Nonlinear systems third edition.
[55] D. Bertsekas, Convex optimization theory, Vol. 1, Athena Scientific, 2009.
[56] R. B. Myerson, Game theory: analysis of conflict, Harvard university press, 1997.
[57] H. Brezis, Functional analysis, Sobolev spaces and partial differential equations, Springer Science & Business Media, 2010.
[58] H. T. Banks, H. K. Nguyen, Sensitivity of dynamical systems to banach space parameters, Journal of mathematical analysis and applications
323 (1) (2006) 146–161.
[59] K. Deimling, Ordinary differential equations in Banach spaces, Vol. 596, Springer, 2006.

26