J. Math. Anal. Appl.

474 (2019) 1536–1550

Contents lists available at ScienceDirect

Journal of Mathematical Analysis and Applications

www.elsevier.com/locate/jmaa

A stochastic differential equation SIS epidemic model with two

independent Brownian motions
Siyang Cai, Yongmei Cai ∗ , Xuerong Mao
Department of Mathematics and Statistics, University of Strathclyde, Glasgow G1 1XH, UK

a r t i c l e i n f o a b s t r a c t

Article history: In this paper, we introduce two perturbations in the classical deterministic
Received 6 June 2018 susceptible–infected–susceptible epidemic model. Greenhalgh and Gray [4] in 2011
Available online 13 February 2019 use a perturbation on β in SIS model. Based on their previous work, we consider
Submitted by Y. Du
another perturbation on the parameter μ + γ and formulate the original model as
Keywords: a stochastic differential equation (SDE) with two independent Brownian Motions
SIS model for the number of infected population. We then prove that our Model has a unique
Independent Brownian motion and bounded global solution I(t). Also we establish conditions for extinction and
Extinction persistence of the infected population I(t). Under the conditions of persistence, we
Persistence show that there is a unique stationary distribution and derive its mean and variance.
Stationary distribution Computer simulations illustrate our results and provide evidence to back up our
theory.
© 2019 Elsevier Inc. All rights reserved.

1. Introduction

Research on epidemics modelled by introducing deterministic compartmental models makes great con-
tribution to understanding the behaviour of epidemics and helping control of deadly diseases [3,14]. For
example, Capasso [3] introduces the Kermack–Mckendrick model to describe diseases that offer permanent
immunity after an individual catching the diseases for a period of time. However, some diseases such as
sexually transmitted and bacterial disease do not have permanent immunity. Susceptible individuals will
catch the disease at some time to become infected, while after a short period of time infected individuals will
become susceptible again. Susceptible–infected–susceptible (SIS) model is a very simple but also commonly
used model to describe such epidemic problems [9]. S(t) and I(t) are used to represent the numbers of
susceptible and infected populations at time t. The deterministic model is
 dS(t)
dt = μN − βS(t)I(t) + γI(t) − μS(t)
dI(t)
(1.1)
dt = βS(t)I(t) − (μ + γ)I(t)

* Corresponding author.
E-mail address: [email protected] (Y. Cai).

https://doi.org/10.1016/j.jmaa.2019.02.039
0022-247X/© 2019 Elsevier Inc. All rights reserved.
 S. Cai et al. / J. Math. Anal. Appl. 474 (2019) 1536–1550 1537

with initial values S0 + I0 = N and here N is the total size of population. μ is the per capita death rate
and γ is the rate at which infected individuals become cured. β is the disease transmission coefficient. With
the condition S + I = N , we can rewrite the original two ODEs (1.1) into

dI(t) = [β(N − I(t))I(t) − (μ + γ)I(t)] dt (1.2)

While deterministic models are not enough to describe problems in real world because parameters are
easily influenced by all kinds of circumstances with uncertainty. Thus stochastic models with different
environmental noises are more appropriate in epidemic problems. For example, A. Gray et al. [4] consider
the perturbation on β in deterministic SIS model. They firstly analyse (1.2) in a small time interval [t, t + dt)
with the d notation for small change in any quantity. Hence we have dI(t) = I(t + dt) −I(t) in (1.2). Then the
disease transmission coefficient β can be regarded as the rate at which each infected individual make contacts
with other individuals and the total number of new infections in the small time interval is βI(t)S(t) dt and
also, a single infected individual makes β dt potentially infectious contacts with other individuals in the small
time interval. Consequently, when some stochastic environmental factor is introduced on each individual in
the population, they replace β by a random variable β̃

β̃ dt = β dt + σ1 dB1 (t) (1.3)

Here dB1 (t) = B1 (t + dt) − B1 (t) is the increment of a standard Brownian motion. Hence the potentially
infectious contacts made by a single infected individual with another individual in the population in the
small time interval [t, t + dt) are normally distributed with mean β dt and variance σ12 dt. Also, Y. Zhao
et al. [15] use the same perturbation in SIS model with a vaccination and then find the conditions for the
disease to become extinct and persist. There are also many other contributions on different epidemic models
using multiple environmental noises [6,7,10,11].
Now based on the previous work of A. Gray et al. [4], we now consider another perturbation on (μ + γ)
with (1.3) existing in traditional SIS model. Within the same small time interval [t, t + dt), we regard
(μ + γ)I(t) dt as the total number of infected individuals becoming cured or pass away in this time interval.
In other words, this is the total reduction of infections. Hence each single individual contributes (μ + γ) dt
in the reduction of infections in the small time interval [t, t + dt).
Then we introduce stochasticity on (μ + γ). (μ + γ) is replaced by a random variable (μ̃ + γ̃)

(μ̃ + γ̃) dt = (μ + γ) dt + σ2 N − I(t) dB2 (t) (1.4)

Here we do not simply set (μ̃+ γ̃) dt = (μ +γ) dt +σ2 dB2 (t) to be the second perturbation. When susceptible
population S(t) = N − I(t) is large, which means there are few infected individuals, the error of estimating
μ and γ will be large. Thus we suppose that the variance of estimating μ + γ is proportional to the number
of susceptible population. As a result, the reduction of infections caused by medical care and death of a
single infected individual in the small time interval [t, t + dt) is normally distributed with mean (μ + γ) dt
and variance σ2 2 (N − I(t)) dt. This is also a biologically reasonable model because the variance trends to 0
when dt goes to 0.
Such a diffusion coefficient in square root form is widely used in financial stochastic differential equations
such as Square Root Process. Mao [12] indicates that Square Root Process may be more appropriate if the
asset price volatility does not increase dramatically when S(t) increases (S(t) greater than 1), because the
variance of error term is proportional to S(t). Meanwhile, in epidemic modelling, Liang and Greenhalgh
et al. introduce demographic stochasticity [6] in thedeterministic SIS model based on Allen’s work [1].
The diffusion coefficient of their SDE SIS model is βI(t)(N − I(t)) + (μ + γ) which is very similar to
ours. However, to the best of our knowledge, there is not enough work on incorporating white noise with
square-root diffusion into the epidemic models. As a result, this paper aims to fill the gap.
1538 S. Cai et al. / J. Math. Anal. Appl. 474 (2019) 1536–1550

As a result, we assume that two Brownian motions B1 (t) and B2 (t) are independent. We then substitute
two perturbations in our SIS ODE (1.2). We have

dI(t) = [β(N − I(t))I(t) − (μ + γ)I(t)] dt + σ1 I(t)(N − I(t)) dB1 (t)


− σ2 I(t) N − I(t) dB2 (t) (1.5)

with initial value I(0) = I0 ∈ (0, N ). In the following sections we will concentrate on giving some properties
of the solution I(t) of this SDE.

2. Existence of unique positive solution

In order for the model to make sense, we need to prove that the solution of our SDE has a unique global
solution which remain within (0, N ), with the initial value I0 ∈ (0, N ).

Theorem 2.1. If μ + γ ≥ 12 σ22 N , then for any given initial value I(0) = I0 ∈ (0, N ), the SDE has a unique
global positive solution I(t) ∈ (0, N ) for all t ≥ 0 with probability one, namely,

P{I(t) ∈ (0, N ), ∀t ≥ 0} = 1

Proof. The coefficients of our SDE are locally Lipschitz continuous and for any given initial value, there
is a unique maximal local solution I(t) on t ∈ [0, τe ), where τe is the explosion time [12]. Let k0 ≥ 0 be
sufficient large to satisfy k10 < I0 < N − k10 . For each integer k ≥ k0 , define the stopping time

τk = inf{t ∈ [0, τe ) : I(t) ∈

/ (1/k, N − 1/k)}

In this paper we set inf∅ = ∞. Obviously, τk is increasing when k → ∞. And we set τ∞ = limk→∞ τk . It is
clear that τ∞ ≤ τe almost sure. So if we can show that τ∞ = ∞ a.s., then τe = ∞ a.s. and I(t) ∈ (0, N ) a.s.
for all t ≥ 0.
Here we assume τ∞ = ∞ a.s. is not true. Then we can find a pair of constants T > 0 and  ∈ (0, 1) such
that

P{τ∞ ≤ T } > 

So we can find an integer k1 ≥ k0 large enough, such that

P{τk ≤ T } ≥  ∀k ≥ k1 (2.1)

Define a function V : (0, N ) → R+ by

N2
V (x) = − log x − log (N − x) + log
4

and

1 1 1 1
Vx = − + , Vxx = 2 +
x N −x x (N − x)2

Let f (t) = β(N − I(t))I(t) − (μ + γ)I(t), g(t) = (σ1 I(t)(N − I(t)), −σ2 N − I(t)I(t)) and dB(t) =
( dB1 (t), dB2 (t)).
 S. Cai et al. / J. Math. Anal. Appl. 474 (2019) 1536–1550 1539

By Ito formula [12], we have, for any t ∈ [0, T ] and k ≥ k1

 k
t∧τ  k
t∧τ

EV (I(t ∧ τk )) = V (I0 ) + E LV (I(s)) ds + E Vx g(s) dB(s) (2.2)

0 0

 t∧τk
E 0
Vx g(s) dB(s) = 0. Also it is easy to show that

x
LV (x) = −β(N − x) + (μ + γ) + βx − (μ + γ)
N −x
1 x2
+ (σ12 (N − x)2 + σ12 x2 + σ22 (N − x) + σ22 )
2 N −x
1
≤ −β(N − x) + (μ + γ) + βx + [σ12 (N − x)2 + σ12 x2 + σ22 (N − x)]
2
≤C (2.3)

C is a constant when μ + γ ≥ 12 σ22 N and x ∈ (0, N ).

Then we have

 k
t∧τ

EV (I(t ∧ τk )) ≤ V (I0 ) + E C ds
0

≤ V (I0 ) + Ct (2.4)

which yields that

EV (I(T ∧ τk )) ≤ V (I0 ) + CT (2.5)

Set Ωk = {τk ≤ T } for k ≥ k1 and we have P(Ωk ) ≥ . For every ω ∈ Ωk , I(τk , ω) equals either 1/k or
N − 1/k and we have

N2
V (I(τk , ω)) = log
4(N − 1/k)1/k

Hence

∞ > V (I0 ) + CT ≥ E[IΩk (ω)V (I(τk , ω))]

N2
≥ P(Ωk ) log
4(N − 1/k)1/k
N2
=  log
4(N − 1/k)1/k

letting k → ∞ will lead to the contradiction

∞ > V (I0 ) + CT = ∞

So the assumption is wrong and we must have τ∞ = ∞ almost sure, whence the proof is now completed. 2
1540 S. Cai et al. / J. Math. Anal. Appl. 474 (2019) 1536–1550

However, the condition for our model to have bounded positive solution μ +γ ≥ 12 σ22 N might be confusing
to readers. There are two different ways to understand this condition. In [4] there is no constraint on σ1
but after adding second perturbation on μ + γ, the square root term will trend to infinity very fast when
I(t) → N . So there must be a condition on σ2 to neutralise it. Also, by the classical Feller test in Mao’s book
[12] on Mean Reverting Square Root Process, there is a very similar result on constraining the coefficient
before square root term in order to make the solution always positive.

3. Extinction

In this section, we will discuss the conditions for the disease to die out in our SDE model (1.5). Here we
give the conditions for the solution I(t) of our SDE becoming extinction.

σ 2 N 2 +σ 2 N σ12 N 2 +σ22 N
Theorem 3.1. Given that R0S := R0D − 12(μ+γ)2 = βN
μ+γ − 2(μ+γ) < 1, then for any given initial value
I(0) = I0 ∈ (0, N ), the solution of SDE obeys

1
lim sup log I(t) < 0 a.s. (3.1)
t→∞ t

if one of the following three conditions is satisfied

• σ12 N + 12 σ22 ≤ β or
• 12 σ22 ≥ βor
• (β − σ1 2(μ + γ)) ∨ (β − σ12 N ) < 12 σ22 < β

namely, I(t) will trend to zero exponentially a.s. And the disease will die out with probability one.

Proof. Here we use Ito formula

t t
log I(t) log I0 1 1
= + LṼ (I(s)) ds+ σ1 (N − I(s)) dB1 (s)
t t t t
0 0

t 
1
− σ2 (N − I(s)) dB2 (s) (3.2)
t
0

LṼ is defined by

1
LṼ (x) = β(N − x) − (μ + γ) − [σ12 (N − x)2 + σ22 (N − x)], x ∈ (0, N ) (3.3)
2
According to the large number theorem for martingales[12], we must have

t t 
1
lim sup { σ1 (N − I(s)) dB1 (s) − σ2 (N − I(s)) dB2 (s)} = 0 (3.4)
t→∞ t
0 0

So if we can prove LṼ ≤ C̃ < 0, then lim supt→∞ 1t log I(t) < 0 a.s. (C̃ is a constant).
We first examine LṼ at 0 and N. LṼ (N ) = −(μ + γ) < 0 and LṼ (0) = βN − (μ + γ) − 12 (σ12 N 2 + σ22 N )
so we must have firstly

LṼ (0) < 0, which is ensured by R0S < 1 (3.5)

 S. Cai et al. / J. Math. Anal. Appl. 474 (2019) 1536–1550 1541

LṼ (x) has the maximal value when

−β + σ12 N + 12 σ22 2 σ2 − β
1 2
x = x̂ = = N + (3.6)
σ12 σ12

and

1 (β − 12 σ22 )2
LṼ (x̂) = − (μ + γ) (3.7)
2 σ12

is the maximal value of LṼ when x ∈ R.

So we need to discuss with the following three different cases.
Case 1. x̂ ≤ 0. With LṼ < 0 at 0 and N, if we have x̂ ≤ 0 Then LṼ < 0 for all x ∈ (0, N ). Consequently,

1
σ12 N + σ22 ≤ β (3.8)
2

Case 2. x̂ ≥ N . This is similar with Case 1. LṼ < 0 for all x ∈ (0, N ). So we must have

1 2
σ ≥β (3.9)
2 2

Case 3. x̂ ∈ (0, N ). In this case we need to make sure the maximal value LṼ (x̂) < 0. So we have

1 (β − 12 σ22 )2
LṼ (x̂) = − (μ + γ) < 0 (3.10)
2 σ12

Also,

1 2
σ <β (3.11)
2 2
and
1
σ12 N + σ22 > β (3.12)
2

is required for x̂ within (0, N ). Rearrange and we therefore have the result for Case 3

 1 2
(β − σ1 2(μ + γ)) ∨ (β − σ12 N ) < σ <β (3.13)
2 2

Hence when any of the three cases is satisfied, we must have LṼ ≤ C̃ < 0 (C̃ is a constant). It then follows
that

log I(t) log I0 1

lim sup ≤ lim sup + lim sup C̃t + 0 < 0 a.s.
t→∞ t t→∞ t t→∞ t

Therefore we now have obtained the proof of Theorem 3.1. 2

Simulation. In this paper we assume that the unit of time is one day and the population size is measured
in units of 1 million. Consequently, our parameters are given by the following values in this section.

N = 100, β = 0.4, μ + γ = 0.45, σ1 = 0.03

1542 S. Cai et al. / J. Math. Anal. Appl. 474 (2019) 1536–1550

Fig. 1. Extinction with condition 1.

Fig. 2. Extinction with condition 2.

In order to find the value of σ2 , we initially need the model to make sense, so we have

σ2 ≤ 2(μ + γ)/N = 0.94868 (3.14)

and also if there is extinction in our model, we need

R0S < 1, which results in σ2 ≥ 0 (3.15)

Using these parameters in the other three conditions, we have the corresponding σ2 to satisfy the three
conditions in extinction.

• condition 1: σ2 ≤ 0.78740078 or,

• condition 2: σ2 ≥ 0.8944271 or,
• condition 3: 0.78740078 ≤ σ2 ≤ 0.8944271

Here we choose 0.3, 0.9 and 0.82 respectively and plot our model by using Euler–Maruyama (EM) Method
[12,13] in R, with step size Δ = 0.001 and both large and small initial values. The computer simulations
are presented in Figs. 1, 2 and 3. Clearly, our results in this section are illustrated and supported by the
simulations. With the values of parameters, the disease will die out.

4. Persistence

In this section we want to discuss the conditions for the disease to persist in our model. However, there
are many definitions of persistence in stochastic dynamic problems [2,4,5,7,11–13]. For example, in Mao’s
 S. Cai et al. / J. Math. Anal. Appl. 474 (2019) 1536–1550 1543

Fig. 3. Extinction with condition 3.

book [12] he gives a very general definition of persistence, which only needs the disease to never become
extinction with probability 1, such that

lim inf I(t) > 0

t→∞

While Greenhalgh and Gray [4] define persistence of their model as oscillations around a positive level. This
is a very strong result in epidemic problem. As our works is an extension of [4], we give Theorem 4.1 as
follows
σ 2 N 2 +σ 2 N σ12 N 2 +σ22 N
Theorem 4.1. If R0S = R0D − 12(μ+γ)2 = βN
μ+γ − 2(μ+γ) > 1, then for any given initial value I(0) = I0 ∈
(0, N ), the solution of (1.5) follows

lim sup I(t) ≥ ξ and lim inf I(t) ≤ ξ a.s. (4.1)

t→∞ t→∞

where

−β + σ12 N + 12 σ22 + β 2 − σ22 β − 2σ12 (μ + γ) + 14 σ24
ξ= (4.2)
σ12

which is the only positive root of LṼ = 0 in (0, N ). I(t) will be above or below the level ξ infinitely often
with probability one.

Proof. When R0S > 1, recall that

1
LṼ (x) = β(N − x) − (μ + γ) − [σ12 (N − x)2 + σ22 (N − x)], x ∈ (0, N )
2
−β+σ12 N + 12 σ22
and we have LṼ (0) > 0, LṼ (N ) = −(μ + γ) < 0 and ξ > x̂ = σ12
. So LṼ (x) is strictly increasing
in (0, 0 ∨ x̂) and strictly decreasing in (0 ∨ x̂, N ).
Here we recall (3.2)

t t
log I(t) log I0 1 1
= + LṼ (I(s)) ds + σ1 (N − I(s)) dB1 (s)
t t t t
0 0

t 
1
− σ2 (N − I(s)) dB2 (s)
t
0
1544 S. Cai et al. / J. Math. Anal. Appl. 474 (2019) 1536–1550

By the large number theorem for martingales [12], there is an Ω2 ⊂ Ω with P{Ω2 } = 1 such that for every
ω ∈ Ω2

t t 
1
lim { σ1 (N − I(s)) dB1 (s) − σ2 (N − I(s)) dB2 (s)} = 0 (4.3)
t→∞ t
0 0

Now we assume that lim supt→∞ I(t) ≥ ξ a.s. is not true. Then there must be a small  ∈ (0, 1) such that

P{lim sup I(t) ≤ ξ − 2} >  (4.4)

t→∞

Let Ω1 = {lim supt→∞ I(t) ≤ ξ − 2}, then for every ω ∈ Ω1 , there exist T = T (ω) large enough, such that

I(t, ω) ≤ ξ − 2 +  = ξ − , when t ≥ T (ω) (4.5)

which means when t ≥ T (ω), LṼ (I(t, ω)) ≥ LṼ (ξ − ). So we have for any fixed ω ∈ Ω1 ∩ Ω2 and t ≥ T (ω)

T(ω)
1 1 1
lim inf log I(t, ω) ≥ 0 + lim LṼ (I(s, ω)) ds + lim LṼ (ξ − )(t − T (ω))
t→∞ t t→∞ t t→∞ t
0

≥ LṼ (ξ − ) > 0

which yields

lim I(t, ω) = ∞ (4.6)

t→∞

and this contradicts with the assumption (4.4). So we must have lim supt→∞ I(t) ≥ ξ almost sure.
Similarly, if we assume that lim inf t→∞ I(t) ≤ ξ a.s. is not true. Then there must be a small δ ∈ (0, 1)
such that

P{lim inf I(t) ≥ ξ + 2δ} > δ (4.7)

t→∞

Let Ω3 = {lim inf t→∞ I(t) ≥ ξ + 2δ}, then for every ω ∈ Ω3 , there exist T  = T  (ω) large enough, such that

I(t, ω) ≥ ξ + 2δ − δ = ξ + δ, when t ≥ T  (ω) (4.8)

Now we fix any ω ∈ Ω3 ∩ Ω2 and t ≥ T  (ω) in (3.2) and we have

T (ω)
1 1 1
lim sup log I(t, ω) ≤ 0 + lim LṼ (I(s, ω)) ds + lim LṼ (ξ + δ)(t − T  (ω))
t→∞ t t→∞ t t→∞ t
0

≤ LṼ (ξ + δ) < 0

which yields

lim I(t, ω) = 0 (4.9)

t→∞

and this contradicts the assumption (4.7). So we must have lim inf t→∞ I(t) ≤ ξ almost sure. 2
 S. Cai et al. / J. Math. Anal. Appl. 474 (2019) 1536–1550 1545

Fig. 4. Persistence.

Simulation. In this section we choose the values of our parameter as follows

N = 100, β = 0.5, μ + γ = 0.45, σ1 = 0.03

With R0S > 1, we have σ2 < 0.1. Hence here we choose σ2 = 0.05 and the level ξ = 0.916056. Similarly,
as the level ξ is very closed to zero, we use both large and small initial values and plot the level ξ in the
simulation plots to illustrate the results. From Fig. 4, it is clear that the number of infectious population
will fluctuated around the level ξ. Thus the disease will not die out or explode, which means the disease
will persist.

5. Stationary distribution

In this section we will prove that there exists a unique stationary distribution of our SDE model (1.5)
when the solution persists between 0 and N . So we give the first theorem this section.

Theorem 5.1. If R0S > 1, then our SDE model (1.5) has a unique stationary distribution.

In order to complete our proof, we need to initially use a well-known result from Khaminskii as a
lemma [8].

Lemma 5.2. The SDE model (1.3) has a unique stationary distribution if there is a strictly proper subinterval
(a,b) of (0,N) such that E(τ ) < ∞ for all I0 ∈ (0, a] ∪ [b, N ), where

τ = inf{t ≥ 0 : I(t) ∈ (a, b)} (5.1)

also,

sup E(τ ) < ∞ (5.2)

I0 ∈[ā,b̄]

for every interval [ā, b̄] ⊂ (0, N ).

Now we can prove Theorem 5.1 using Lemma 5.2.

Proof. Firstly we need to fix any (a, b) such that,

0<a<ξ<b<N (5.3)
1546 S. Cai et al. / J. Math. Anal. Appl. 474 (2019) 1536–1550

recall LṼ in last section, we can see that

LṼ (x) ≥ LṼ (0) ∧ LṼ (a), if 0 < x ≤ a (5.4)

LṼ (x) ≤ LṼ (b), if b ≤ x < N (5.5)

also, recall (3.2)

t t
log I(t) = log I0 + LṼ (I(s)) ds+ σ1 (N − I(s)) dB1 (s)
0 0

t 
− σ2 (N − I(s)) dB2 (s)
0

and define

τ = inf{t ≥ 0 : I(t) ∈ (a, b)} (5.6)

Case 1. For all t ≥ 0 and any I0 ∈ (0, a), from (5.4), we have

t∧τ
E log I(t ∧ τ ) = E log I0 + E LṼ (I(s)) ds + 0
0

≥ log I0 + E(LṼ (0) ∧ LṼ (a))(t ∧ τ ) (5.7)

From definition of τ , we know that

log a ≥ E log I(t ∧ τ ) when I0 ∈ (0, a] (5.8)

Rearrange and we have

log ( Ia0 )
E(t ∧ τ ) ≤
LṼ (0) ∧ LṼ (a)

when t → ∞

log ( Ia0 )
E(τ ) ≤ < ∞, ∀I0 ∈ (0, a] (5.9)
LṼ (0) ∧ LṼ (a)

Case 2. For all t ≥ 0 and any I0 ∈ (b, N ), from (5.5), we have

t∧τ
E log I(t ∧ τ ) = E log I0 + E LṼ (I(s)) ds + 0
0

≤ log I0 + E(LṼ (b))(t ∧ τ ) (5.10)

From definition of τ , we know that

log b ≤ E log I(t ∧ τ ) when I0 ∈ (b, N ] (5.11)

 S. Cai et al. / J. Math. Anal. Appl. 474 (2019) 1536–1550 1547

Rearrange and we have

log b ≤ log I0 + LṼ (b)E(t ∧ τ )

log ( Ib0 )
E(t ∧ τ ) ≤ −
| LṼ (b) |

when t → ∞

log ( Ib0 )
E(τ ) ≤ − ≤ ∞ ∀I0 ∈ (b, N ] (5.12)
| LṼ (b) |

Combine the results from both Case 1 and Case 2 and we complete the proof of Theorem 5.1. 2

Now we need to give the mean and variance of the stationary distribution.

Theorem 5.3. If R0S > 1 and denote m and v as the mean and variance of the stationary distribution of SDE
model (1.5). Then we have

2β(R0S − 1)(μ + γ)
m= (5.13)
2β 2 − σ12 (βN + μ + γ) − σ22 β

and
βN − μ − γ
v= m − m2 (5.14)
β

Proof. For any I0 ∈ (0, N ), we firstly recall (1.5) in the integral form

t t
I(t) = I0 + [β(N − I(s))I(s) − (μ + γ)I(s)] ds + σ1 I(s)(N − I(s)) dB1 (s)
0 0

t 
− σ2 I(s) N − I(s) dB2 (s) (5.15)
0

Dividing both sides by t and when t → ∞, applying the ergodic property of the stationary distribution [8]
and also the large number theorem of martingales, we have the result that

0 = (βN − μ − γ)m − βm2 (5.16)

where m, m2 are the mean and second moment of the stationary distribution. Also, we need to consider
(3.2) as well

t t
log I(t) log I0 1 1
= + LṼ (I(s)) ds + σ1 (N − I(s)) dB1 (s)
t t t t
0 0

t 
1
− σ2 (N − I(s)) dB2 (s) (5.17)
t
0

when t → ∞. We have
1548 S. Cai et al. / J. Math. Anal. Appl. 474 (2019) 1536–1550

Fig. 5. Stationary distribution.

1 2 1 1 1
σ1 m2 − (σ12 N + σ22 − β)m = βN − μ − γ − σ12 N 2 − σ22 N (5.18)
2 2 2 2

Note that βN − μ − γ − 12 σ12 N 2 − 12 σ22 N = (R0S − 1)(μ + γ). Rewrite this

1 2 1
σ m2 − (σ12 N + σ22 − β)m = (R0S − 1)(μ + γ) (5.19)
2 1 2

Rearrange and we have

2β(R0S − 1)(μ + γ)
m= (5.20)
2β 2 − σ12 (βN + μ + γ) − σ22 β

Hence

βN − μ − γ
v = m 2 − m2 = m − m2 2 (5.21)
β

Simulation. In this section we choose the values of our parameter as follows

N = 100, β = 0.5, μ + γ = 0.45, σ1 = 0.02, σ2 = 0.05

R0S = 1.06389 > 1 so the disease will persist and there is a stationary distribution of our model. And for
these parameters, the mean and variance of the stationary distribution of our model is

m = 6.23982, v = 23.4628

In order to reach the stationary distribution in our simulation, we set a long run of 20000 iterations with
step size Δ = 0.001 and then calculate the mean and variance for the last 1000 iterations. The results from
simulations show that

m = 6.531954, v = 23.2428

Fig. 5 also displays the path of I(t) and the empirical cumulative distribution functions for the last 1000
samples of the simulation.
 S. Cai et al. / J. Math. Anal. Appl. 474 (2019) 1536–1550 1549

6. Conclusion

In this paper we introduce another perturbation on μ + γ based on Greenhalgh and Gray’s research
[4] with a different form. This SIS SDE model with two independent Brownian motion has very similar
properties as theirs [4]. We firstly prove that our model has a unique and positive solution which is bounded
with (0, N ) with probability 1. Then we define the Stochastic Reproduction Number of our model, which
needs a weaker condition for the model to be extinction compared to the classical deterministic model and
the previous model with one perturbation. When R0S < 1, we find the further three conditions for the disease
to die out. As long as one of these is satisfied, the disease will die out with probability one. When R0S > 1,
we prove that the solution of our model will oscillate around a positive level ξ almost surely. Under this
circumstance, we find the unique positive stationary distribution of our SDE model with the expression of
mean and variance. Importantly, simulations with different values of parameters are produced to illustrate
and support our theoretical results.
Our new perturbation clearly needs σ2 not too large from Theorem 2.1 to ensure a unique bounded
positive solution of (1.5). However this perturbation extends the requirements for R0S < 1 compared to
the deterministic SIS model and the results in [4]. This means for those parameters that will not cause the
disease to die out in the deterministic model as well as Gray’s model [4], extinction will become possible if we
add the new perturbation. Meanwhile, we find the unique stationary distribution with no extra conditions,
which means that adding our new perturbation in Gray’s model [4] will have similar results.

Acknowledgments

The authors would like to thank the reviewers and the editors for their very professional comments
and suggestions. The authors would like to thank the Leverhulme Trust (RF-2015-385), the Royal Soci-
ety (WM160014, Royal Society Wolfson Research Merit Award), the Royal Society and the Newton Fund
(NA160317, Royal Society-Newton Advanced Fellowship) and the EPSRC (EP/K503174/1) for their fi-
nancial support. The first two authors would like to thank the University of Strathclyde for the financial
support. The second author would also like to thank China Scholarship Council and the MASTS pooling
initiative (The Marine Alliance for Science and Technology for Scotland) for their financial support. MASTS
is funded by the Scottish Funding Council (HR09011).

References

[1] E.J. Allen, Modelling with Ito Stochastic Differential Equations, Springer-Verlag, Dordecht, 2007.
[2] Y. Cai, X. Mao, Stochastic prey–predator system with foraging arena scheme, Appl. Math. Model. 64 (2018) 357–371.
[3] V. Capasso, A generalization of the Kermack–McKendrick deterministic epidemic model, Math. Biosci. 42 (1–2) (2005)
43–61.
[4] A. Gray, D. Greenhalgh, L. Hu, X. Mao, J. Pan, A stochastic differential equation SIS epidemic model, SIAM J. Appl.
Math. 71 (3) (2011) 876–902.
[5] A. Gray, D. Greenhalgh, X. Mao, J. Pan, The SIS epidemic model with Markovian switching, J. Math. Anal. Appl. 394 (2)
(2011) 876–902.
[6] D. Greenhalgh, Y. Liang, X. Mao, Demographic stochasticity in the SDE SIS epidemic model, Discrete Contin. Dyn. Syst.
Ser. B 20 (9) (2015) 2859–2884.
[7] D. Greenhalgh, Y. Liang, X. Mao, SDE SIS epidemic model with demographic stochasticity and varying population size,
Int. J. Appl. Math. Comput. 276 (2016) 218–238.
[8] R.Z. Hasminskii, Stochastic Stability of Differential Equations, Sijthoff & Noordhoff, Alphenaan den Rijn, The Netherlands,
Germantown, MD, 1980.
[9] H.W. Hethcote, J.A. Yorke, Gonorrhea Transmission Dynamics and Control, Lecture Notes in Biomath, vol. 56, Springer-
Verlag, Berlin, 1984.
[10] G. Hu, M. Liu, K. Wang, The asymptotic behaviours of an epidemic model with two correlated stochastic perturbations,
Appl. Math. Comput. 218 (21) (2012) 10520–10532.
1550 S. Cai et al. / J. Math. Anal. Appl. 474 (2019) 1536–1550

[11] Y. Liang, D. Greenhalgh, X. Mao, A stochastic differential equation model for the spread of HIV amongst people who
inject drugs, Comput. Math. Methods Med. 2016 (2016) 6757928.
[12] X. Mao, Stochastic Differential Equations and Applications, 2nd ed., Horwood, Chichester, UK, 2008.
[13] X. Mao, C. Yuan, Stochastic Differential Equations with Markovian Switching, Imperial College Press, London, UK, 2006.
[14] B. Shugin, L. Stone, Z. Agur, Pulse vaccination strategy in the SIR epidemic model, Bull. Math. Biol. 60 (1998) 1123–1148.
[15] Y. Zhao, D. Jiang, D. O’Regan, The extinction and persistence of the stochastic SIS epidemic model with vaccination,
Phys. A 392 (20) (2013) 4916–4927.