J. Math. Anal. Appl.

348 (2008) 540–554

Contents lists available at ScienceDirect

Journal of Mathematical Analysis and Applications

www.elsevier.com/locate/jmaa

A highly sensitive mean-reverting process in finance and the

Euler–Maruyama approximations
Fuke Wu a,∗ , Xuerong Mao b , Kan Chen b
a
Department of Mathematics, Huazhong University of Science and Technology, Wuhan, Hubei 430074, PR China
b
Department of Statistics and Modelling Science, 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: Empirical studies show that the most successful continuous-time models of the short-
Received 12 July 2007 term rate in capturing the dynamics are those that allow the volatility of interest
Available online 3 August 2008 changes to be highly sensitive to the level of the rate. However, from the mathematics,
Submitted by L. Guo
the high sensitivity to the level implies that the coefficients do not satisfy the linear
Keywords:
growth condition, so we can not examine its properties by traditional techniques. This
Structure of interest rate paper overcomes the mathematical difficulties due to the nonlinear growth and examines
Stochastic differential equation its analytical properties and the convergence of numerical solutions in probability. The
Convergence in probability convergence result can be used to justify the method within Monte Carlo simulations that
Euler–Maruyama method compute the expected payoff of financial products. For illustration, we apply our results
Monte Carlo simulation compute the value of a bond with interest rate given by the highly sensitive mean-reverting
process as well as the value of a single barrier call option with the asset price governed by
this process.
© 2008 Elsevier Inc. All rights reserved.

1. Introduction

The short-term riskless interest rate is one of the most fundamental and important quantities in financial markets. Many
models have been put forward to explain its behavior. Letting R (t ) represent the short-term interest rate, we list some
well-known models as follows:
1. Merton [11] dR (t ) = μ dt + σ dw (t ),
2. Vasicek [13] dR (t ) = (μ + λ R (t )) dt + σ √ dw (t ),
3. Cox, Ingersoll and Ross [5] dR (t ) = (μ + λ R (t )) dt + σ R (t ) dw (t ),
4. Dothan [6] dR (t ) = σ R (t ) dw (t ),
5. Geometric Brownian motion dR (t ) = λ R (t ) dt + σ R (t ) dw (t ),
6. Brennan and Schwartz [1] dR (t ) = (μ + λ R (t )) dt + σ R (t ) dw (t ),
3
7. Cox, Ingersoll and Ross [4] dR (t ) = σ R (t ) 2 dw (t ),
8. Constant Elasticity of Variance [3] dR (t ) = λ R (t ) dt + σ R (t )γ dw (t )
where λ, μ and σ are constants.
As in [2] and [12], these eight models may be nested within the following stochastic differential equation:
 
dR (t ) = λ μ − R (t ) dt + σ R γ (t ) dw (t ) (1.1)
by simply placing the appropriate restrictions on the three parameters λ, μ and γ .

* Corresponding author. Fax: +86 27 87543231.

E-mail addresses: [email protected] (F. Wu), [email protected] (X. Mao), [email protected] (K. Chen).

0022-247X/$ – see front matter © 2008 Elsevier Inc. All rights reserved.
doi:10.1016/j.jmaa.2008.07.069
 F. Wu et al. / J. Math. Anal. Appl. 348 (2008) 540–554 541

When γ = 1/2, the solution of Eq. (1.1) is the well-known mean-reverting square root process [5]. It is widely used
to model volatility, interest rates and other financial quantities. Many papers (e.g. [8,9]) discuss its analytical properties.
Higham and Mao [7] examine strong convergence of its Monte Carlo simulation. When γ ∈ [1/2, 1], Mao et al. [10] discuss
its analytical properties and strong convergence of numerical solutions.
Some empirical studies show that the most successful continuous-time models of the short-term rate in capturing the
dynamics are those that allow the volatility of interest changes to be highly sensitive to the level of the rate. By χ 2 tests to
US T-bill data, the above models which assume γ < 1 are rejected and those which assume γ  1 are not rejected. Applying
the Generalized Method Moment, Chan et al. [2] give γ = 1.449. Using the same data, by the Gaussian Estimation methods,
Nowman [12] estimates γ = 1.361.1 Therefore, it is more evident to consider γ  1.
However, γ > 1 implies that the diffusion coefficient does not satisfy the linear growth condition so we cannot apply the
classical results (e.g. [9]) on the existence and uniqueness of the solutions, boundedness of the moment, the convergence of
its Euler–Maruyama approximate solutions, and so on. This paper develops new techniques to overcome these difficulties.
Many empirical studies, including [2] and [12], estimate the parameters of the continuous-time model by the Euler–
Maruyama discrete approximation. However, they can not ensure that these parameters are precise enough because it has
not been proved yet that the approximate solution will converge to the exact solution when the discrete step size tends to
zero. This paper will fill the gap.
In the next section, we first consider the existence and nonnegativity of the solution of Eq. (1.1). This is a natural
requirement since Eq. (1.1) is frequently used to model the interest rate. In section three, we consider various boundedness
of the solution of Eq. (1.1), including the moment boundedness, the stochastic boundedness and the pathwise estimations.
In Section 4, we introduce the Euler–Maruyama approximations to the solution of Eq. (1.1) and examine its convergence in
probability. Finally, we choose bonds and a single barrier call option to show that the numerical solution can be used to
compute the expected payoffs.

2. Positive and global solutions

Throughout this paper, let (Ω, F , P) be a complete probability space with a filtration {Ft }t 0 satisfying the usual condi-
tions (namely, it is right continuous and increasing while F0 contains all P-null sets). Let w (t ) be a scalar Brownian motion
defined on the probability space. We consider the mean-reverting γ -process
 
dR (t ) = λ μ − R (t ) dt + σ R γ (t ) dw (t ) (2.1)
with the initial value R (0) > 0, where λ, μ and σ are positive and γ > 1.
In order for a stochastic differential equation to have a unique global (i.e., no explosion in a finite time) solution for
any given initial value, the coefficients of the equation are generally required to satisfy the linear growth condition and the
local Lipschitz condition (see [9]). However, the diffusion coefficient of Eq. (2.1) does not satisfy the linear growth condition,
though it is locally Lipschitz continuous. We wonder if the solution of Eq. (2.1) may explode at a finite time. Furthermore,
since Eq. (2.1) is used to model interest rate and other quantities, it is critical that the solution R (t ) will never become
negative. The following theorem reveals the existence of the positive solution.

Theorem 2.1. For any given initial value R (0) > 0, λ, μ and σ > 0, there exists a unique positive global solution R (t ) to Eq. (2.1) on
t  0.

Proof. Since the coefficients of (2.1) satisfy the local Lipschitz condition, for any given initial value R (0) > 0, there must
exist a unique local solution R (t ) ∈ [0, τe ], where τe is the explosion time. To show this solution is global, we need to show
that τe = ∞ a.s. For a sufficient large integer k > 0, namely 1/k < R (0) < k, define the following stopping time,
 
τk = inf t ∈ [0, τe ]: R (t ) ∈/ [1/k, k] ,
where throughout this paper we set inf ∅ = ∞ (as usual ∅ denotes the empty set). Clearly, τk is increasing as k → ∞. Set
τ∞ = limk→∞ τk , whence τ∞  τe a.s. If we can prove τk → ∞ a.s. as k → ∞, then τe = ∞ a.s. and R (t ) > 0 a.s. for all
t  0. In other words, to complete the proof what we need to show is that τ∞ = ∞ a.s. To prove this, for any constant T , if
P{τk  T } → 0 as k → ∞, then we have P{τ∞ = ∞} = 1, which is the required assertion.
For θ ∈ (0, 1), define a C 2 -function V : (0, ∞) → (0, ∞) by
V ( R ) = R θ − 1 − θ log R . (2.2)
It is easy to see that V (·)  0 and V ( R ) → ∞ as R → ∞ or R → 0. Applying the Itô formula yields

    θσ 2  
dV R (t ) = λθ μ R θ−1 (t ) − R θ (t ) − μ R −1 (t ) + 1 + (θ − 1) R θ+2(γ −1) (t ) + R 2(γ −1) (t ) dt
2
 
+ σ θ R θ+γ −1 (t ) − R γ −1 (t ) dw (t ). (2.3)

1
In this paper, using one month British sterling rate data, Norman obtain γ = 0.2898. The difference between US and British data shows the volatility
of short-term rates is highly sensitive to the level of rates in the US, while it is not in UK.
542 F. Wu et al. / J. Math. Anal. Appl. 348 (2008) 540–554

By boundedness of polynomial, for θ ∈ (0, 1), there exists a constant K 1 such that

  θσ 2  
λθ μ R θ−1 − R θ − μ R −1 + 1 + (θ − 1) R θ+2(γ −1) + R 2(γ −1)  K 1 .
2
Therefore, for any t ∈ [0, T ],
   
E V R (t ∧ τk )  V R (0) + K 1 T ,

so
     
P(τk  T ) V (1/k) ∧ V (k)  E V R ( T ∧ τk )  V R (0) + K 1 T .

Therefore P(τk  T ) → 0 since V (1/k) ∧ V (k) → ∞ as k → ∞. This implies P(τ∞ = ∞) = 1, as required. 2

3. Boundedness

For the interest rates and other assets’ prices, boundedness is a natural requirement. In this section, we will establish
various boundedness for the solution to Eq. (2.1).

3.1. Moment boundedness

We mainly focus on the boundedness of the first moment and the second moment.

Theorem 3.1. The solution of Eq. (2.1) obeys

E R (t )  R (0) + μ, ∀t  0, (3.1)

and

lim sup E R (t )  μ. (3.2)

t →∞

Proof. Applying the Itô formula yields

   
d e λt R (t ) = e λt λμ dt + σ R (t )γ dw (t ) .

For any positive number n, define a stopping time

 
τn = inf t: R (t ) > n .
Then
t ∧τn
 
E e λ(t ∧τn ) R (t ∧ τn ) = R (0) + λμE e λs ds
0
  λt
 R (0) + μ e − 1 .

Letting n → ∞, the Fatou theorem yields

 
e λt E R (t )  R (0) + μ e λt − 1 ,

which implies the required assertions. 2

From the proof of Theorem 2.1, we observe that the average in time of the moment of the solutions will be bounded,
which is described as follows.

Theorem 3.2. For any θ ∈ (0, 1), there exists a positive constant K θ such that for any initial value R (0) > 0, the solution of Eq. (2.1)
has the property
t 
1 θ+2(γ −1)
lim sup ER (s) ds  K θ . (3.3)
t →∞ t
0
 F. Wu et al. / J. Math. Anal. Appl. 348 (2008) 540–554 543

Proof. It is obvious that there exists a constant K 2 such that

 
  θσ 2 1
λθ μ R θ−1 − R θ − μ R −1 + 1 + (θ − 1) R θ+2(γ −1) + R 2(γ −1)  K 2 .
2 2

It then follows from (2.3) that

t ∧τk
θ(1 − θ)σ 2    
E R θ+2(γ −1) (s) ds + E V R (t ∧ τk )  V R (0) + K 2 t .
4
0

Letting k → ∞ and applying the Fatou theorem, we have

t
θ(1 − θ)σ 2  
E R θ+2(γ −1) (s) ds  V R (0) + K 2 t ,
4
0

which implies the required assertion. 2

Corollary 3.1. If γ > 3/2, then there is a constant K > 0 such that for any R (0) > 0, the solution of Eq. (2.1) obeys

t
1
lim sup E R 2 (s) ds  K . (3.4)
t →∞ t
0

Proof. As γ > 3/2, we can choose θ ∈ (0, 1) for θ + 2(γ − 1)  2. By the Hölder inequality, we compute
t t 2
θ+2(γ −1)
1 2 1 θ+2(γ −1)
E R (s) ds  ER (s) ds .
t t
0 0

Letting t → ∞ and applying Theorem 3.2 give the assertion. 2

3.2. Stochastic boundedness

The following result shows that R (t ) will stay in a belt area with a large probability.

Theorem 3.3. If γ ∈ (1, 2), for any ε ∈ (0, 1) and R (0) > 0, there exists a pair of positive constants H = H (ε, R (0)) and
h = h(ε , R (0)) such that
 
P h  R (t )  H  1 − ε , for ∀t  0. (3.5)

Proof. For any ε > 0, let H = 2( R (0) + μ)/ε . Then by the Chebyshev inequality and Theorem 3.1,
  R (0) + μ ε
P R (t ) > H  = . (3.6)
H 2

Set y (t ) = R −1 (t ). By the Itô formula,

 
dy (t ) = −λμ y 2 (t ) + λ y (t ) + σ 2 R 2γ −3 (t ) dt − σ y 2−γ (t ) dw (t ). (3.7)

Fix any constant θ ∈ (0, 1). Then

   
d e θ t y (t ) = e θ t −λμ y 2 (t ) + (λ + θ) y (t ) + σ 2 R 2γ −3 (t ) dt − σ e θ t y 2−γ (t ) dw (t ).

Noting γ ∈ (1, 2), we have

R 2γ −3 = 1{γ = 3 } + y 3−2γ 1{1<γ < 3 } + R 2γ −3 1{ 3 <γ <2} .
2 2 2

Noting λμ > 0 and 0 < 3 − 2γ < 1 when γ ∈ (1, 3/2), then we conclude that there exists a constant K 3 such that
 
−λμ y 2 + (λ + θ) y + σ 1 + y 3−2γ 1{1<γ < 3 }  K 3 , 2
2
544 F. Wu et al. / J. Math. Anal. Appl. 348 (2008) 540–554

so, by the Jensen inequality and Theorem 3.1, we may compute

t t
1
e θ t E y (t )  + K3 e θ s ds + σ 2 E e θ s R 2γ −3 (s)1{ 3 <γ <2} (s) ds
R (0) 2
0 0
t
1 K3  θt   2γ −3
 + e − 1 + σ2 e θ s E R (s) 1{ 3 <γ <2} ds
R (0) θ 2
0

1 K3  θt  σ2  2γ −3  
 + e −1 + R (0) + μ eθ t − 1 .
R (0) θ θ
Therefore, there exists a constant K 4 such that

E y (t )  K 4 .
Similarly, by the Chebyshev inequality, there exists a positive constant h = h(ε , R (0)) such that
    ε
P R (t )  h = P y (t )  h−1  1 − .
2
This, together with (3.6), implies
     
P h  R (t )  H = P R (t )  h − P R (t ) > H  1 − ε ,
as required. 2

3.3. Pathwise estimations

Theorem 3.4. If γ ∈ (1, 2), then for any initial value R (0) > 0,
log R (t )
lim inf  −1 a.s. (3.8)
t →∞ log t

Proof. From Eq. (3.7),

t +1 t +1 
λμ λμ
E y (t + 1) + E 2
y (s) ds = E y (t ) + E − y 2 (s) + λ y (s) + σ 2 R 2γ −3 (s) ds. (3.9)
2 2
t t

There exists a constant K 5 such that

λμ
− y 2 + λ y + σ 2 + σ 2 y 3−2γ 1{1<γ < 3 } < K 5 ,
2 2

so we have
t +1 t +1
λμ
E y 2 (s) ds  E y (t ) + K 5 + σ 2 E R 2γ −3 (s)1{ 3 <γ <2} ds
2 2
t t
2
 2γ −3
 K4 + K5 + σ R (0) + μ . (3.10)

By Eq. (3.7), for any u ∈ [t , t + 1],

u u u u

y (u ) = y (t ) − λμ y 2 (s) ds + λ y (s) ds + σ 2 R 2γ −3 (s) ds − σ y 2−γ (s) dw (s),

t t t t

so we have
t +1 t +1
 
E sup y (u )  E y (t ) + λ E y (s) ds + σ 2 E R 2γ −3 (s) ds
t u t +1
t t
 u 
 
 2−γ 
+ σE sup  y (s) dw (s) . (3.11)
t u t +1 
t
 F. Wu et al. / J. Math. Anal. Appl. 348 (2008) 540–554 545

By the Lyapunov inequality, we have

 3−2γ  2γ −3
E R 2γ −3 (t )  1 + E y (t ) 1{1<γ < 3 } + E R (t ) 1{ 3 <γ <2} .
2 2

Applying the Burkholder–Davis–Gundy inequality and the Jensen inequality, etc. yields
 u  t +1 1
  2
 2−γ  2(2−γ )
E sup  y (s) dw (s)  6E y (s) ds
t u t +1 
t t
t +1 1
  2
2

6 E γ − 1 + (2 − γ ) y (s) ds
t
t +1 1
2
2
6 (γ − 1) + (2 − γ )E y (s) ds .
t
 t +1
By the boundedness of E t
y 2 (s) ds and E y (t ), we therefore see from (3.11) there exists a constant K 6 such that
 
E sup y (u )  K 6 . (3.12)
t u t +1

Let ε > 0 be arbitrary. By the Chebyshev inequality, we have

  K6
P sup y (t ) > k1+ε  , k = 1, 2, . . . .
kt k+1 k1+ε

Applying the Borel–Cantelli lemma yields, for almost all ω ∈ Ω,

1+ε
sup y (t )  k (3.13)
kt k+1

holds for all but finitely many k. Hence, there exists a k0 (ω), for almost all ω ∈ Ω , for which (3.13) holds whenever k  k0 .
Consequently, for almost all ω ∈ Ω , if k  k0 and k  t  k + 1,
log y (t ) (1 + ε ) log k
 = 1 + ε. (3.14)
log t log k
That is,
log R (t )
lim inf  −(1 + ε ). (3.15)
t →∞ log t
Letting ε → 0, we obtain the desired assertion (3.8). The proof is therefore complete. 2

This theorem shows that for any ε > 0, there exists a positive random variable T ε such that, with probability one,
−(1+ε )
R (t )  t , for ∀t  T ε . (3.16)
In other words, with probability one, the solution will not decay faster than t −(1+ε) . The following theorem describes the
growth constraint.

Theorem 3.5. If γ ∈ (1, 2), then for any initial value R (0) > 0,
log R (t )
lim sup  0 a.s., (3.17)
t →∞ t 1+ε
where ε is an arbitrary positive constant.

Proof. For any positive constant θ , applying the Itô formula to e θ t log R (t ) results in
t  
σ2
e θ t log R (t ) = log R (0) + e θ s θ log R (s) − λ − R 2(γ −1) (s) + λμ y (s) ds + M (t ), (3.18)
2
0

where
t

M (t ) = σ e θ s R γ −1 (s) dw (s)
0
546 F. Wu et al. / J. Math. Anal. Appl. 348 (2008) 540–554

is a real-valued continuous local martingale vanishing at t = 0 with quadratic variation

t
 
M (t ), M (t ) = σ 2 e 2θ s R 2(γ −1) (s) ds.
0

Fix any ∈ (0, 1) and ξ > 1. For every integer k  1, using the exponential martingale inequality we have
  
  ξ eθ t 1
P sup M (t ) − M (t ), M (t ) > log k  ξ.
0t k 2 k

By the Borel–Cantelli lemma we observe that there exists an integer k = k(ω) such that
  ξ eθ t
M (t )  e −θ k M (t ), M (t ) + log k
2
for almost all ω ∈ Ω . Thus Eq. (3.18) leads to
t  
σ 2 (1 − e −γ (k−s) )
e θ t log R (t )  log R (0) + e θ s θ log R (s) − λ − R 2(γ −1) (s) ds
2
0
t
ξ eθ t
+ λμ e θ s y (s) ds + log k (3.19)
0

for 0  t  k and k  k(ω). We may easily observe that there exists a constant K 7 such that

σ 2 (1 − e −γ (k−s) )
θ log R − λ − R 2(γ −1)  K 7 ,
2
so we have, for t > 1,
t
K7  θt  ξ eθ t
e θ t log R (t )  log R (0) + e − 1 + λμ e θ s y (s) ds + + log k
θ
0
T ε ∨1
K7  θt 
 log R (0) + e − 1 + λμ e θ s y (s) ds
θ
0
t
ξ eθ t
+ λμ e θ s y (s) ds + log k (3.20)
T ε ∨1

where T ε is defined in (3.16). Compute

t t

e θ s y (s) ds  e θ s s1+ε ds
T ε ∨1 T ε ∧1
t
1 t 1+ε
 s1+ε e θ s  T ∧1 − e θ s sε ds
θ ε θ
T ε ∧1

1
 t 1+ε e θ t .
θ
Hence,
T ε ∨1
θt K7  θt  λμ 1+ε θ t ξ e θ t
e log R (t )  log R (0) + e − 1 + λμ e θ s y (s) ds + t e + log k.
θ θ
0

Thus
T ε ∨1
  K7 λμ 1+ε ξ
log R (t )  log R (0) + + t + log k + λμe −θ t e θ s y (s) ds (3.21)
θ θ
0
 F. Wu et al. / J. Math. Anal. Appl. 348 (2008) 540–554 547

for 1  t  k and k  k(ω). If k − 1  t  k, it follows that

  T ε ∨1
 
log R (t )

1 log R (0) + K 7 + λμ + ξ log k λμ
+ θ t 1+ε e θ s y (s) ds, (3.22)
t 1+ε t 1+ε θ θ (k − 1)1+ε e t
0

then we obtain
log R (t ) λμ
lim sup  a.s.
t →∞ t 1+ε θ
Letting θ → ∞ yields the required assertion (3.20). 2

4. The Euler–Maruyama method

This section deals with the regime where the time interval, [0, T ], is fixed. There is so far no explicit solution to Eq. (2.1)
so we consider its numerical solution. We refer to it as the Euler–Maruyama (EM) method. Now we define the discrete EM
approximate solution to (2.1) for a given fixed timestep ∈ (0, 1) and r0 = R (0),

rk+1 = rk + λ(μ − rk ) + σ |rk |γ wk , (4.1)

where w k = w (tk+1 ) − w (tk ) is a Brownian motion increment.

In our analysis, it will be more convenient to use continuous-time approximation. Letting [t / ] be the integer part of
t / , we hence introduce the step process

[ T /]−1
r̄ (t ) = rk 1[k ,(k+1) ) (t ) (4.2)
k=0

and define the continuous approximation

discretation
t t
 
r (t ) = r0 + λ μ − r̄ (s) ds + σ r̄ (s)γ dw (s). (4.3)
0 0

We know that r (t ) is not computable because it requires knowledge of the entire Brownian path, not just its -
increments. However, since rk = r (tk ), an error bound for r (t ) will automatically implies the error bound for {rk }k0 .
Therefore, we mainly investigate the error bound for r (t ). For this error bound, we have the following theorem.

Theorem 4.1. For R (t ) in (2.1) and r (t ) in (4.3),

   
sup  R (t ) − r (t ) = 0,
2
lim in probability. (4.4)
→0 0t  T

Proof. We divide the whole proof into four steps.

Step 1. We use the same notation as in the proof of Theorem 2.1, where we have shown that for the stopping time τk ,
1    
P(τk  T )  V R (0) + K 1 T . (4.5)
V (1/k) ∧ V (k)

Step 2. We define the similar stopping time,

 
ρk = inf t ∈ [0, T ]: r (t ) ∈/ [1/k, k] .
To prove that ρ has the same property as τk , we choose θ = 1/2 in (2.2), namely in the definition of the function V (·).
Applying the Itô formula to (4.3) yields

t ∧ρk
    λμ  1  λ 1 
E V r (t ∧ ρk ) = V R (0) + E r − 2 (s) − r −1 (s) − r − 2 (s)r̄ (s) − r −1 (s)r̄ (s)
2 2
0
2 
σ σ2 3
+ r −2 (s)r̄ 2γ (s) − r − 2 (s)r̄ 2γ (s) ds. (4.6)
4 8
548 F. Wu et al. / J. Math. Anal. Appl. 348 (2008) 540–554

Rearranging the terms on the right-hand side gives

t ∧ρk
    λμ  1  λ 1 
E V r (t ∧ ρk ) = V R (0) + E r − 2 (s) − r −1 (s) − r − 2 (s) − 1
2 2
0

σ2 σ2 3 λ 1  
+ r 2(γ −1) (s) − r 2γ − 2 ( s ) + r − 2 (s) − r −1 (s) r (s) − r̄ (s)
4 8 2

σ2  3  
+ r − 2 (s) − 2r −2 (s) r 2γ (s) − r̄ 2γ (s) ds. (4.7)
8
Note that there exists a constant c 1 such that
λμ  1  λ 1  σ2 σ2 3
r − 2 − r −1 − r− 2 − 1 + r 2(γ −1) − r 2γ − 2  c 1
2 2 4 8
and a constant c 2 (k) such that
 2γ 
r − r̄ 2γ   c 2 (k)|r − r̄ |2γ .
When r ∈ [1/k, k], which implies r̄ ∈ [1/k, k], we have
t ∧ρk
    λ   
r − 2 (s) + r −1 (s)r (s) − r̄ (s)
1
E V r (t ∧ ρk )  V R (0) + c 1 T + E
2
0

σ 2 c2 (k)  − 32 −2

 
 2γ

+ r (s) − 2r (s) r (s) − r̄ (s) ds
8
t ∧ρk
  λ 1   
 V R (0) + c 1 T + + k 2 + k E r (s) − r̄ (s) ds
2
0
t ∧ρk
σ 2 c2 (k)    
+
3
k 2 + 2k2 E r (s) − r̄ (s)2γ ds. (4.8)
8
0
For γ > 1,
        
r (s) − r̄ (s)2γ  r (s) − r̄ (s) r (s) + r̄ (s) 2γ −1  (2k)2γ −1 r (s) − r̄ (s). (4.9)
Substituting (4.9) into (4.8) yields that there exists a constant c 3 (k) such that
t ∧ρk
     
E V r (t ∧ ρk )  V R (0) + c 1 T + c 3 (k)E r (s) − r̄ (s) ds. (4.10)
0
 t ∧ρk
Now we compute E 0 |r (s) − r̄ (s)| ds. For s ∈ [0, t ∧ ρk ], by definition (4.3),
    
r (s) − r̄ (s) = λ(μ − r[s/ ] ) s − [s/ ] + σ |r[s/ ] |γ w (s) − w [s/ ]
  
 λ(μ + k) + σ kγ  w (s) − w [s/ ] , (4.11)
so, noting ∈ (0, 1), we have
t ∧ρk t ∧ρk
    
E r (s) − r̄ (s) dr  λ(μ + k) T + σ kγ E  w (s) − w [s/ ]  ds
0 0
T
  
 λ(μ + k) T + σ kγ E  w (s) − w [s/ ]  ds
0
T
  
 λ(μ + k) T + σ kγ E w (s) − w [s/ ]  ds
0
1
 λ(μ + k) T + σ kγ T 2

  1
 λ(μ + k) + σ kγ T 2

1
=: D 2 . (4.12)
 F. Wu et al. / J. Math. Anal. Appl. 348 (2008) 540–554 549

Substituting (4.12) into (4.10) yields

    1
E V r (t ∧ ρk )  V R (0) + c 1 T + c 3 (k) D 2 . (4.13)

Therefore, we can show

1    1 
P(ρk  T )  V R (0) + c 1 T + c 3 (k) D 2 . (4.14)
V (1/k) ∧ V (k)
Step 3. Let θk = τk ∧ ρk . We claim that there exists a constant c 4 (k) such that
   
E sup  R (t ) − r (t )2  c 4 (k) . (4.15)
0t θk ∧ T

For any 0  t 1  T , from (4.3) and (2.1), we have

t 1 ∧θk t 1 ∧θk
   γ  γ 
R (t 1 ∧ θk ) − r (t 1 ∧ θk ) = −λ R (s) − r̄ (s) ds + σ R (s) − r̄ (s) ds.
0 0

Therefore, for any t ∈ [0, T ], by the Hölder inequality and the Doob martingale inequality (cf. Mao [9]), we have

t ∧θk t ∧θk
       γ 
sup  R (t 1 ∧ θk ) − r (t 1 ∧ θk )  2λ2 t E  R (s) − r̄ (s)2 ds + 8σ 2 E  R (s) − r̄ γ (s)2 ds.
2
E (4.16)
0t 1 t
0 0

Note that the function h(x) = xγ for any x > 0 and γ  1 is locally Lipschitz continuous, which implies that for R (s), r̄ (s) ∈
[1/k, k], there exists a constant c 5 (k) such that
 γ   
 R (s) − r̄ γ (s)2  c 5 (k) R (s) − r̄ (s)2γ
 2  2γ −2
 c 5 (k) R (s) − r̄ (s)  R (s) − r̄ (s)
 2
 c 5 (k)(2k)2(γ −1)  R (s) − r̄ (s)
 2  2 
 2c 5 (k)(2k)2(γ −1)  R (s) − r (s) + r (s) − r̄ (s) . (4.17)

Substituting (4.17) into (4.16) yields

   
sup  R (t 1 ∧ θk ) − r (t 1 ∧ θk )
2
E
0t 1 t

t ∧θk t ∧θk
     
2
 4 λ t + 4σ c 5 (k)(2k) 2 2(γ −1)
E  R (s) − r (s)2 ds + E r (s) − r̄ (s)2 ds
0 0
t t ∧θk
   2  
2
 4 λ t + 4σ c 5 (k)(2k) 2 2(γ −1)
E R (s ∧ θk ) − r (s ∧ θk ) ds + E r (s) − r̄ (s)2 ds . (4.18)
0 0

In the same way as the computation of (4.12), there exists a constant D̄ such that
t ∧θk
 
E r (s) − r̄ (s)2 ds  D̄ . (4.19)
0

Substituting (4.19) into (4.18) gives

  2 
E sup R (t 1 ∧ θk ) − r (t 1 ∧ θk )
0t 1 t
t
 2 2 2(γ −1)
  2
 4 λ t + 4σ c 5 (k)(2k) E R (s ∧ θk ) − r (s ∧ θk ) ds
0
 
+ 4T λ2 t + 4σ 2 c 5 (k)(2k)2(γ −1) D̄ .
Using the Gronwell inequality yields (4.15).
550 F. Wu et al. / J. Math. Anal. Appl. 348 (2008) 540–554

Step 4. Let ε , δ ∈ (0, 1) be arbitrarily small. Set

  2 
Ω̄ = ω: sup  R (t ) − r (t )  δ .
0t  T

Using (4.15),
 
δP Ω̄ ∩ {θk  T } = E[1{θk T } 1{Ω̄} ]
   
 E 1{θk T } sup  R (t ) − r (t )2
0t  T ∧θk
   
E sup  R (t ) − r (t )2
0t  T ∧θk

 c 4 (k) .
This, together with (4.5) and (4.14), implies
 
P(Ω̄)  P Ω̄ ∩ {θk  T } + P(θk  T )
 
 P Ω̄ ∩ {θk  T } + P(τk  T ) + P(ρk  T )
1
c 4 (k) 2V ( R (0)) + K 1 T + c 1 T + Dc 3 (k) 2
 + . (4.20)
δ V (1/k) ∧ V (k)
Recalling that k → ∞, V (1/k) ∧ V (k) → ∞, we can choose k sufficiently large for
2V ( R (0)) + K 1 T + c 1 T ε
<
V (1/k) ∧ V (k) 2
and then choose sufficiently small for
1
c 4 (k) Dc 3 (k) 2 ε
+ <
δ V (1/k) ∧ V (k) 2
to obtain
  2 
P(Ω̄) = P sup  R (t ) − r (t )  δ < ε , (4.21)
0t  T

which is the desired assertion (4.4). 2

5. Valuation of bonds and options

In this section we will show that the EM method can be used to compute financial quantities. Typically, we choose bonds
and barrier options to demonstrate our theory.

5.1. Bonds

In the case where R (t ) in (2.1) models the short-term interest rate dynamics, the price of a bond at the end of period is
given by
T 
B ( T ) = E exp − R (t ) dt . (5.1)
0

Using the step function r̄ (t ) in (4.2), a natural approximation to B ( T ) is

T 
 
B̄ ( T ) = E exp − r̄ (t ) dt . (5.2)
0

The following result shows the convergence of this approximation.

Theorem 5.1. In the notations above,

 
lim  B ( T ) − B̄ ( T ) = 0. (5.3)
→0

To prove this assertion, we need the following lemmas.

 F. Wu et al. / J. Math. Anal. Appl. 348 (2008) 540–554 551

Lemma 5.1. For r (t ) in (4.3) and r̄ (t ) in (4.2),

   
sup r (t ) − r̄ (t ) = 0
2
lim in probability. (5.4)
→0 0t  T

Proof. We first prove

  
lim E sup r (t ) − r̄ (t ) = 0.
→0 0t  T ∧ρk

By (4.11), for any t ∈ [0, T ∧ ρk ], we have

        
E sup r (t ) − r̄ (t )2  2λ2 (μ + k)2 2
+ 2σ 2 k2γ E sup  w (t ) − w [t / ] 2
0t  T ∧ρk 0t  T ∧ρk
   2 
 2λ2 (μ + k)2 2
+ 2σ 2 k2γ E sup  w (t ) − w [t / ]  . (5.5)
0t  T

By the Doob martingale inequality,

   4     
E sup  w (t ) − w [t / ]  =E sup sup  w (t ) − w (k )4
0t  T 0k[ T / ]−1 k t (k+1)

[ T /]−1    
 E sup  w (t ) − w (k )4
k t (k+1)
k=0
[ T /]−1
   4
 E w (k + 1) − w (k )
k=0
[ T /]−1
2
3
k=0
= 3T .
Hence, by the Lyapunov inequality, we have
   2        1
E sup  w (t ) − w [t / ]   E sup  w (t ) − w [t / ] 4 2
0t  T 0t  T
1 1
 (3T ) 2 2 . (5.6)

Substituting (5.6) into (5.5) and noting ∈ (0, 1) yield that there exists a constant c 5 (k) such that
   
E sup r (t ) − r̄ (t )2  c 5 (k) 1
2 . (5.7)
0t  T ∧ρk

For arbitrarily small constants δ, ε ∈ (0, 1), set

  2 
Ω̃ = ω: sup r (t ) − r̄ (t )  δ .
0t  T

Then,
 
δP Ω̃ ∩ {ρk  T } = δE(1{ρk T } 1{Ω̃} )
   
 E 1{ρk T } sup r (t ) − r̄ (t )2
0t  T ∧ρk
   
E sup r (t ) − r̄ (t )2
0t  T ∧ρk
1
 c 5 (k) 2 . (5.8)

This, together with (4.14), yields that

 
P(Ω̃)  P Ω̃ ∩ {ρk  T } + P(ρk  T )
1
c 5 (k) 1 V ( R (0)) + c 1 T + Dc 3 (k) 2
 2 + . (5.9)
δ V (1/k) ∧ V (k)
552 F. Wu et al. / J. Math. Anal. Appl. 348 (2008) 540–554

Choose k sufficiently large such that

V (r (0)) + c 1 T ε
<
V (1/k) ∧ V (k) 2
and then choose sufficiently small such that
1
c 5 (k) 1 Dc 3 (k) 2 ε
2 + < .
δ V (1/k) ∧ V (k) 2
Hence we have
  2 
P sup r (t ) − r̄ (t )  δ < ε , (5.10)
0t  T

which is the desired assertion. 2

Lemma 5.2. For R (t ) in (2.1) and r̄ (t ) in (4.2),

  
lim sup  R (t ) − r̄ (t ) = 0 in probability. (5.11)
→0 0t  T

Proof. For sufficiently small ε , δ ∈ (0, 1),

         
P sup  R (t ) − r̄ (t )  δ  P sup  R (t ) − r (t ) + sup r (t ) − r̄ (t )  δ
0t  T 0t  T 0t  T
 
      δ
 P sup  R (t ) − r (t ) + sup r (t ) − r̄ (t )  δ, sup r (t ) − r̄ (t ) 
0t  T 0t  T 0t  T 2
 
      δ
+ P sup  R (t ) − r (t ) + sup r (t ) − r̄ (t )  δ, sup r (t ) − r̄ (t ) 
0t  T 0t  T 0t  T 2
   
  δ   δ
 P sup  R (t ) − r (t )  + P sup r (t ) − r̄ (t )  .
0t  T 2 0t  T 2

Theorem 4.1 and Lemma 5.1 therefore yield the desired assertion. 2

Proof of Theorem 5.1. It is sufficient to prove

T T
  P  
exp − r̄ (t ) −→ exp −  R (t ) .
0 0

In other words, we need to prove that, for arbitrarily small constants δ, ε ∈ (0, 1),
 T T  
   
 
r̄ (t ) dt   δ < ε .
P exp − R (t ) dt − exp − (5.12)
 
0 0

Using e −|x| − e −| y|  |x − y | and the nonnegativity of R (t ), we have

 T T   T 
       
 r̄ (t ) dt   
exp − R (t ) dt − exp − R (t ) − r̄ (t ) dt 
   
0 0 0
 
 T sup  R (t ) − r̄ (t ). (5.13)
0t  T

Applying Lemma 5.2 yields the desired assertion. 2

5.2. A path-dependent option

We now consider the case where Eq. (2.1) models a single barrier call option, which, at expiry time T , pays the European
call value if R (t ) never exceeded the fixed barrier B, and pays zero otherwise. We suppose that the expected payoff is
computed from a Monte Carlo simulation based on the step function method (4.2). The following theorem shows that the
expected payoff computed by numerical method will converge to the real expected payoff as → 0.
 F. Wu et al. / J. Math. Anal. Appl. 348 (2008) 540–554 553

Theorem 5.2. Let R (t ) and r̄ (t ) be defined by (2.1) and (4.2) respectively. Let E be the exercise price and let B be a barrier. Define
 + 
V := E R ( T ) − E 1{0 R (t ) B , 0t  T } ; (5.14)
 + 
V := E r̄ ( T ) − E 1{0r̄ (t ) B , 0t  T } . (5.15)

Then

lim | V − V | = 0. (5.16)
→0

Proof. Set A = {0  R (t )  B , 0  t  T }, Ā = {0  r̄ (t )  B , 0  t  T }. We will complete the proof if we can prove

 + P  +
r̄ ( T ) − E 1 Ā −→ R ( T ) − E 1A ,

which equivalent to that, for any arbitrarily small constants ε , δ ∈ (0, 1),
 +  +  
P  r̄ ( T ) − E 1 Ā − R (t ) − E 1 A   δ < ε . (5.17)

Making use of the inequality

      
 r̄ ( T ) − E + − R (t ) − E +    R (t ) − r̄ (t ),

we have
 +  +  
P  r̄ ( T ) − E 1 Ā − R (t ) − E 1 A   δ
 +  +   
= P  r̄ ( T ) − E 1 Ā − R (t ) − E 1 A   δ ∩ ( A ∩ Ā )
 +  +    
+ P  r̄ ( T ) − E 1 Ā − R (t ) − E 1 A   δ ∩ A ∩ Ā c
 +  +    
+ P  r̄ ( T ) − E 1 Ā − R (t ) − E 1 A   δ ∩ A c ∩ Ā
      
 P  R (t ) − r̄ (t )  δ + P A ∩ Ā c + P A c ∩ Ā .
By Lemma 5.2, we have P(| R (t ) − r̄ (t )|  δ)  ε /3. We will then complete the proof if we can prove
  ε
P A ∩ Ā c  (5.18)
3
and
  ε
P A c ∩ Ā  . (5.19)
3
For any sufficient small κ , we have
 
A = sup R (t )  B
0t  T
   
= sup R (t )  B − κ ∪ B − κ  sup R (t )  B
0t  T 0t  T

=: A 1 ∪ A 2 .
Hence,
       
A ∩ Ā c = A 1 ∩ Ā c ∪ A 2 ∩ Ā c ⊆ sup  R (t ) − r̄ (t )  κ ∪ A 2 .
0t  T

So
     
P A ∩ Ā c  P sup  R (t ) − r̄ (t )  κ + P( A 2 ).
0t  T

Now we may choose κ so small that P( A 2 ) < ε /6, then by Lemma 5.2, choose so small that P(sup0t  T | R (t ) − r̄ (t )| 
δ) < ε /6, so (5.18) holds.
Now, for any κ > 0, we write
 
Ac = sup R (t ) > B
0t  T
   
= sup R (t ) > B + κ ∪ B < sup R (t )  B + κ
0t  T 0t  T

=: A c1 ∪ A c2 .
554 F. Wu et al. / J. Math. Anal. Appl. 348 (2008) 540–554

So
     
P A c ∩ Ā  P A c1 ∩ Ā + P A c2
     
 P sup  R (t ) − r̄ (t ) > κ + P A c . 2
0t  T

Repeating the above process, (5.19) also holds. The proof is complete. 2

Acknowledgments

The authors would like to thank the financial support from Chinese Scholarship Council. They also wish to thank the referees for their detailed com-
ments and helpful suggestions.

References

[1] M.J. Brennan, E.S. Schwartz, Analyzing convertible bonds, J. Finan. Quant. Anal. 15 (1980) 907–929.
[2] K.C. Chan, G.A. Karolyi, F.A. Longstaff, A.B. Sanders, An empirical comparison of alternative models of the short-term interest rate, J. Finance 47 (3)
(1992) 1209–1227.
[3] J.C. Cox, Notes on option pricing I: Constant elasticity of variance diffusion, Working paper, Stanford University, 1975.
[4] J.C. Cox, J.E. Ingersoll Jr., S.A. Ross, An analysis of variable rate loan contracts, J. Finance 35 (1980) 389–403.
[5] J.C. Cox, J.E. Ingersoll Jr., S.A. Ross, A theory of the term structure of interest rates, Econometrica 53 (1985) 385–407.
[6] U.L. Dothan, On the term structure of interest rates, J. Finan. Econom. 6 (1978) 59–69.
[7] D.J. Higham, X. Mao, Convergence of Monte Carlo simulations involving the mean-reverting square root, J. Comput. Finance 8 (2005) 35–61.
[8] Y.K. Kwok, Mathematical Models of Financial Derivatives, Springer, Berlin, 1998.
[9] X. Mao, Stochastic Differential Equations and Applications, Horwood, 1997.
[10] X. Mao, A. Truman, C. Yuan, Euler–Maruyama approximations in mean-reverting stochastic volatility model under regime-switching, J. Appl. Math.
Stoch. Anal. 2006 (2006) 1–20.
[11] R.C. Merton, Theory of rational option pricing, Bell J. Econom. Management Sci. 4 (1973) 141–183.
[12] K.B. Nowman, Gaussian estimation of single-factor continuous time models of the term structure of interest rate, J. Finance 52 (4) (1997) 1695–1706.
[13] O. Vasicek, An equilibrium characterization of the term structure, J. Finan. Econom. 5 (1977) 177–188.