Journal of Computational and Applied Mathematics 236 (2012) 2660–2672

Contents lists available at SciVerse ScienceDirect

Journal of Computational and Applied

Mathematics
journal homepage: www.elsevier.com/locate/cam

Mean-square stability analysis of numerical schemes for stochastic

differential systems
A. Tocino ∗ , M.J. Senosiain
University of Salamanca, Spain

article info abstract

Article history: For ordinary differential systems, the study of A-stability for a numerical method reduces
Received 14 April 2011 to the scalar case by means of a transformation that uncouples the linear test system
Received in revised form 30 December 2011 as well as the difference system provided by the method. For stochastic differential
equations (SDEs), mean-square stability (MS-stability) has been successfully proposed
MSC: as the generalization of A-stability, and numerical MS-stability has been analyzed for
60H05
one-dimensional equations. However, unlike the deterministic case, the extension of
Keywords: this analysis to multi-dimensional systems is not straightforward. In this paper we give
Stochastic differential systems necessary and sufficient conditions for the MS-stability of multi-dimensional systems
Linear test system with one Wiener noise. The criterion presented does not depend on any norm. Based
Mean square stability on the Routh–Hurwitz theorem, we offer a particular criterion of MS-stability for two-
Numerical method dimensional systems in terms of their coefficients. In addition, a counterpart criterion of
Euler method
MS-stability is given for numerical schemes applied to multi-dimensional systems. The
Stochastic theta methods
MS-stability behavior of a stochastic numerical method is determined by the comparison of
its stability region with the stability region of the system. As an application, the numerical
MS-stability of θ -methods applied to bi-dimensional systems is investigated.
© 2012 Elsevier B.V. All rights reserved.

1. Introduction

Stochastic differential equations (SDEs) have become an important tool in many scientific areas owing to their application
for modeling dynamical systems. As a consequence, there is increasing interest in developing numerical methods for solving
SDEs. The numerical stability of these methods has mainly been analyzed for scalar SDEs, see e.g. [1–6]. In this paper we are
interested in the study of the numerical stability of stochastic schemes applied to multi-dimensional SDEs, see e.g. [7]. As
for other stochastic concepts, one expects that stochastic stability would prove to be an extension of the deterministic case.
It is well-known that the study of linear A-stability for ordinary differential equations (ODEs) can be reduced to the scalar
case by means of a transformation that uncouples the test system as well as the difference system given by the method, see
e.g. [8]. For SDEs, in the scalar case Saito and Mitsui [6] have considered a linear test equation with one multiplicative noise
and have proposed the concept of numerical MS-stability as the generalization of A-stability. In addition they analyzed the
MS-stability domains of a number of schemes. Following this line, Higham [4,5] studied MS-stability of scalar θ -methods
for equations with one noise, and Buckwar and Sickenberger [1] studied those with several noise terms. However, unlike
the deterministic case, the scalar stochastic analysis cannot be directly extended to multi-dimensional systems because in
the stochastic linear test equation there are two matrices – one for the drift and one for the diffusion coefficients – and
hence, in general, at most one of them can be assumed to be diagonal. Only in the simultaneous diagonalizable case does the
analysis reduce to the study of a scalar SDE; see [9]. Several attempts have been made by Mitsui and Saito to extend their

∗ Corresponding author. Tel.: +34 923294460; fax: +34 923294583.

E-mail addresses: [email protected] (A. Tocino), [email protected] (M.J. Senosiain).

0377-0427/$ – see front matter © 2012 Elsevier B.V. All rights reserved.
doi:10.1016/j.cam.2012.01.002
 A. Tocino, M.J. Senosiain / Journal of Computational and Applied Mathematics 236 (2012) 2660–2672 2661

results to multi-dimensional systems, see [10,11], providing only sufficient conditions as the stability criterion. Another
important drawback of their solution, also found in [12], lies in the dependence of the results on the specific matrix norm
used. In this paper we propose new criteria with necessary and sufficient conditions for the MS-stability of the linear test
equation (Section 3) and for the numerical approximation (Section 5), which, in addition, do not depend on the chosen norm.
In Section 4, the criterion of MS-stability in the two-dimensional case is converted into a set of necessary and sufficient
conditions in terms of the coefficients of the linear test system matrices. As an application, in Section 6, the numerical
MS-stability of bi-dimensional θ -methods is investigated. We extend the scalar case results in [4,5], concluding that bi-
dimensional θ -methods for θ ≥ 1/2 preserve the MS-stability of the linear test equation. A similar result can be found
in [13], where a bi-dimensional linear test equation with two independent noises was used.

2. Preliminaries

In this section we introduce some notation and definitions for later use. Consider the d-dimensional SDE of Itô type
given by
dXt = f (t , Xt ) dt + g (t , Xt ) dWt ,
(1)
X t0 = c ,
where c is a constant vector; W is the standard scalar Wiener process, and the coefficients f = f (t , x) and g = g (t , x) with
t ∈ [t0 , T ], x ∈ Rd , satisfy the assumptions of the existence and uniqueness theorem, see [14], and are continuous with
respect to t. Let us also assume that f (t , 0) = 0 and g (t , 0) = 0 for t ≥ t0 . Notice that this implies that the process Xt ≡ 0,
called the equilibrium position, is the (unique) solution of (1) with c = 0.

Definition 1 ([14,15]). The equilibrium position is said to be stable in mean square if for every ε > 0 there exists a δ > 0
such that
sup E ∥Xt (c )∥2 ≤ ε for ∥c ∥ ≤ δ.
t0 ≤t <∞

And it is said to be asymptotically stable in mean square if it is stable in mean square and if for all c in a neighborhood of
x=0
lim E ∥Xt (c )∥2 = 0.
t →∞

Replacing in the above expressions E ∥Xt (c )∥2 by ∥E [Xt (c ) Xt (c )′ ]∥, one can say that the equilibrium position possesses a
stable second moment or an asymptotically stable second moment, respectively.
It can be shown, see [14], that stability in mean square (MS-stability) is equivalent to the stability of the second moment.

3. Linear MS-stability

The linear stability theory for ordinary differential equations starts from a simple test system, whose solutions tend to
zero as t tends to infinity and seek the conditions for the numerical scheme to behave similarly. The usual test system,
see [8], is of the form
x′ = Ax,
where A is a constant matrix with different eigenvalues λ1 , . . . , λd ∈ C, all of them in the negative half-plane. In this case
there exists a non-singular matrix T such that T −1 AT = diag[λ1 , . . . , λd ], and the transformation x = Tz uncouples both
the test system and the difference system produced by the use of a numerical scheme. Thus the analysis of stability for
multi-dimensional systems can be reduced to the study of scalar test equations. As an extension, in the stochastic case, a
linear SDE
dXt = A Xt dt + B Xt dWt (2)
with A, B real matrices and A fulfilling the above condition, can be considered. Notice that, in general, the transformation
T does not diagonalize B, i.e., the matrices A, B are not simultaneously diagonalizable (SD). The SD case has been analyzed
in [9].
Following Arnold [14], the second moment P (t ) = E [Xt Xt′ ] = (pij (t )) of the solution of (2) satisfies the equation
dP (t )
= AP (t ) + P (t )A′ + BP (t )B′ , (3)
dt
and the MS-stability of the equilibrium position of (2) is identical to the stability of the second moment P (t ), which in turn is
equivalent to the stability of the trivial solution of the ordinary differential system (3). Since P (t ) is symmetric, (3) reduces
to a linear system of d(d + 1)/2 differential equations of the form
dY
= MY , (4)
dt
2662 A. Tocino, M.J. Senosiain / Journal of Computational and Applied Mathematics 236 (2012) 2660–2672

j
where the components of the vector Y are the different pij = E [Xti Xt ]. For example, if d = 2, A = (aij ) and B = (bij ), then
(4) can be written as:
 dp (t ) 
11
 dt  
2(a12 + b11 b12 ) p11 (t )

2a + b2 b212

 dp22 (t )   11 2 11
 
 = b21 2a22 + b222 2(a21 + b21 b22 )  p22 (t ) . (5)
 dt 
a21 + b11 b21 a12 + b12 b22 a11 + a22 + b12 b21 + b11 b22 p12 (t )
dp12 (t )
 
dt
Therefore, the asymptotic mean-square stability of test system (2) is identical to the ordinary stability of the equilibrium
position Y ≡ 0 of the linear system (4), which is equivalent, see e.g. [16], to the condition that all the eigenvalues of M lie
in the left half-plane. Denoting by σ (M ) the spectrum of the matrix M and
ν(M) := max{ℜ(λ) : λ ∈ σ (M)}
its spectral abscissa, we arrive at the criterion:

Proposition 1. The linear test system (2) is asymptotically MS-stable if and only if ν(M ) < 0.

Example 1. Let us assume, see [9], that

α
   
0 1 0
A= , B=
β γ 0 α
with α, β, γ ∈ R. Thus

α2
 
0 2
M=0 2γ + α 2  2β
β 1 γ + α2

has eigenvalues α 2 + γ , α 2 + γ ± 4β + γ 2 . As a consequence



α + γ + 4β + γ 2 if 4β + γ 2 > 0
 2
ν(M) =
α2 + γ if 4β + γ 2 ≤ 0

 conclude that these problems would be MS-stable if and only if 4β +γ ≤ 0 and α +γ < 0, or 4β +γ > 0 and
2 2 2
and we may
α + γ + 4β + γ < 0. Komori and Mitsui [9] propose the case of α = 3, β = −100, γ = −25; since 4β + γ = 225 > 0
2 2 2

and α 2 + γ + 4β + γ 2 = −1 < 0, the problem is stable.



 in this problem, i.e. taking α = 0, the ODE x = Ax is stable when 4β + γ ≤ 0 and

′ 2
Notice that by dropping the noise
γ < 0 or 4β + γ > 0 and γ + 4β + γ < 0. Thus, once the values of γ , β have been fixed such that the ODE is stable,
2 2

the introduction of a noise gives a stable or unstable SDE, depending on its intensity α . For example, γ = −4, β = −5 gives
a stable ODE (4β + γ 2 < 0), and the SDE obtained by introducing a noise of intensity α is stable if and only if −2 < α < 2.
The set of pairs of matrices A, B that provide MS-asymptotically stable systems
DSDE = {A, B ∈ Rd×d : ν(M ) < 0}
will be called the domain of MS-stability of the test SDE (2).

Remark 1. Since M is a real matrix, using the Routh–Hurwitz criterion, see [17], the condition ν(M ) < 0 can be verified in
terms of the coefficients of M , without an explicit computation of its eigenvalues.

4. MS-stability of bi-dimensional schemes

Following [11] in this section, we shall assume that d = 2 and that A is a diagonal matrix, i.e., the 2 × 2 test system is of
the form
λ1 α1 β1
   
0
dXt = Xt dt + X dWt (6)
0 λ2 β2 α2 t
with parameters λi , αi , βi ∈ R, i = 1, 2. In this case, we have (4)–(5) with

2λ1 + α12 β12 2α1 β1

 

M= β22 2λ2 + α22 2α2 β2 . (7)

α1 β2 α2 β1 λ1 + λ2 + α1 α2 + β1 β2
 A. Tocino, M.J. Senosiain / Journal of Computational and Applied Mathematics 236 (2012) 2660–2672 2663

First, notice that the concept of MS-stability and the criterion given in Proposition 1 overcome the drawbacks of the
corresponding concept and criterion of MS-stability w.r.t. a logarithmic norm given in [11]. On the one hand, the condition in
Proposition 1 does not depend on any norm; on the other hand, it is a necessary and sufficient condition. Since ν(M ) ≤ µ(M )
for any logarithmic norm µ, one has that the condition µ(M ) < 0 for a logarithmic norm µ is sufficient (but not necessary)
for the convergence to the zero solution. In particular, the criteria based on µ1 , µ2 , µ∞ logarithm norms given in [11] may
be unable to decide in some cases, as illustrated in the following example:

Example 2. If
   
−100 0 0 2
A= , B=
0 −1 9 0

we have
−200 4 0
 
M= 81 −2 0 ,
0 0 −83
√ √ √
σ (M) = {−101 − 45 5, −83, −101 + 45 5} and ν(M) = −101 + 45 5 < 0. Accordingly, the equation is
asymptotically MS-stable. However the criteria of MS-stability w.r.t. µp in [11] do not decide for p = 1, 2, ∞ because
√
µ2 (M) = −101 + 46 429/2 > 0, µ1 (M) = 2 > 0 and µ∞ (M) = 79 > 0.
Using the Routh–Hurwitz theorem, we can give an especially easy to check criterion. The characteristic polynomial of the
matrix M given in (7) is

P (x) = det(xI − M ) = x3 + a2 x2 + a1 x + a0 (8)

where

a2 = −α12 − α1 α2 − α22 − β1 β2 − 3(λ1 + λ2 );

a1 = α13 α2 − β12 β22 + 2β1 β2 (λ1 + λ2 ) + α22 (−β1 β2 + 3λ1 + λ2 )
+ α12 α22 − β1 β2 + λ1 + 3λ2 + 2 λ21 + 4λ1 λ2 + λ22 + α1 α2 α22 + 2(λ1 + λ2 ) ;
     
(9)
a0 = β13 β23 + 2β1 β2 λ1 α22 − 2λ2 + β12 β22 (λ1 + λ2 ) − α13 α2 α22 + 2λ2
   

− 2λ1 (λ1 + λ2 ) α22 + 2λ2 − α1 α2 3β12 β22 + 2λ1 α22 + 2λ2

    

− α12 α22 (−3β1 β2 + λ1 + λ2 ) + 2λ2 (−β1 β2 + λ1 + λ2 ) .

 

According to the Routh–Hurwitz criterion and Proposition 1, we have

Proposition 2. The two-dimensional test system (6) is asymptotically MS-stable if and only if a2 > 0, a0 > 0 and a2 a1 > a0 ,
with a2 , a1 , a0 given in (9).

As an application, for the system of Example 2 we have a0 = 6308 > 0, a2 = 285 > 0, a2 a1 − a0 = 4 793 662 > 0,
confirming that the equation is MS-stable.
Particular simple expressions of the criterion can be given in the simultaneously diagonalizable (SD) case and in the
special cases introduced in [11]: the singly anti-diagonal (SAD) and the singly diagonal anti-diagonal (SDAD) systems, whose
equations are given in (6) with respective diffusion matrices

α1 β α β
     
0 0
BSD = , BSAD = , BSDAD = .
0 α2 β 0 β α
Notice that for the SD and SAD test systems the matrix M in (7) becomes

2λ1 + α12 2λ1 β2

   
0 0 0
MSD =  0 2λ2 + α22 0 , MSAD =  β 2 2λ2 0 ,
0 0 λ1 + λ2 + α1 α2 0 0 λ1 + λ2 + β 2

respectively. Thus, in agreement with [1,11], we have

Corollary 3. In the SD case, the two-dimensional test system (6) is asymptotically MS-stable if and only if maxi=1,2 {αi2 + 2λi }
< 0.
Proof. Since

σ (MSD ) = {2λ1 + α12 , 2λ2 + α22 , λ1 + λ2 + α1 α2 }

2664 A. Tocino, M.J. Senosiain / Journal of Computational and Applied Mathematics 236 (2012) 2660–2672

and λ1 + λ2 + α1 α2 ≤ max{2λ1 + α12 , 2λ2 + α22 }, then

ν(MSD ) = max{2λ1 + α12 , 2λ2 + α22 },

and Proposition 1 leads to the conclusion. 

Remark 2. Since MSD = diag[A, B, C ] with A = 2λ1 + α12 , B = 2λ2 + α22 and C = λ1 + λ2 + α1 α2 , in the SD case the
coefficients of P (x) in (8) are a0 = −ABC , a1 = AB + AC + BC , a2 = −(A + B + C ) and the Routh–Hurwitz conditions of
Proposition 2 for the SD case give:
a2 = −(A + B + C ) = −α12 − α1 α2 − α22 − 3λ1 − 3λ2 > 0;
a0 = −ABC = −(α12 + 2λ1 )(α22 + 2λ2 )(α1 α2 + λ1 + λ2 ) > 0;
a1 a2 − a0 = −(A + B)(A + C )(B + C )
= −(α12 + α22 + 2λ1 + 2λ2 )(α12 + α1 α2 + 3λ1 + λ2 )(α22 + α1 α2 + λ1 + 3λ2 ) > 0.
These three conditions are equivalent to ν(MSD ) = max{A, B, C } = max{2λ1 + α12 , 2λ2 + α22 } < 0.

Corollary 4. The system (6) in the SAD case is asymptotically MS-stable if and only if

λ1 + λ2 + β 4 + (λ1 − λ2 )2 < 0. (10)

Proof. It is easy to see that

  
σ (MSAD ) = λ1 + λ2 + β 2 , λ1 + λ2 ± β 4 + (λ1 − λ2 )2 .

Thus ν(MSAD ) = λ1 + λ2 + β 4 + (λ1 − λ2 )2 , and the assertion follows from Proposition 1.




Remark 3. In the SAD case, in agreement with the scheme

µ∞ (MSAD ) < 0 H⇒ ν(MSAD ) < 0 ⇐⇒ Condition (10) holds,
it can be shown algebraically that condition (7) of [11] implies the
 condition of MS-stability in Corollary
√ 4. The reverse,
however, is not true. Example 1 in [11] gives that λ1 + λ2 + β 4 + (λ1 − λ2 )2 = −101 + 9817 < 0, whereas
µ∞ (M) = 2 > 0. Recall that, as said above, the conditions in [11] using logarithmic norms are sufficient but not necessary
for MS-stability.
The MS-stability domain of a SAD system can be identified with the MS-stability region
SSAD = {(λ1 , λ2 , β) ∈ R3 : condition (10) holds}.
A geometrical representation of SSAD can be obtained in R3(x,y,z ) , taking

x = λ1 ∆, y = λ2 ∆, z = β 2∆ (11)
with any ∆ > 0. Condition (10) of MS-stability becomes

x+y+ z 2 + (x − y)2 < 0 (12)
where z > 0. The set of points satisfying (12) is the solid limited by the grid shown in Fig. 1.
Finally, from the Routh–Hurwitz conditions, for the SDAD case we have:

Corollary 5. The system (6) in the SDAD case is asymptotically MS-stable if and only if
(i) −3(λ1 + λ2 ) − β 2 − 3α 2 > 0;
(ii) (β 2 + λ1 + λ2 )(β 4 − 4λ1 λ2 ) − α 6 + 3α 4 (β 2 − λ1 − λ2 ) − α 2 (3β 4 − 2β 2 (λ1 + λ2 ) + 2(λ21 + 4λ1 λ2 + λ22 )) > 0;
(iii) −(λ1 + λ2 )(4β 2 (λ1 + λ2 ) + (3λ1 + λ2 )(λ1 + 3λ2 )) − 4α 6 − 12α 4 (λ1 + λ2 ) − α 2 (−4β 4 + 11λ21 + 26λ1 λ2 + 11λ22 +
4β 2 (λ1 + λ2 )) > 0.
For the general SDAD case it does not seem to be possible to calculate the eigenvalues of the MSDAD matrix. Notice also
that a visual representation of MS-stability regions would not be possible with the parameters λi , i = 1, 2, α, β appearing
in the linear test equation. To avoid this problem, we shall assume that λ1 = λ2 = λ, which does not involve any restriction
in the diffusion part of the equation. Under this hypothesis (called the SDAD* case)
2λ + α 2 β2 2αβ
 

MSDAD∗ =  β2 2λ + α 2 2αβ 
αβ αβ 2λ + α + β
2 2

and
σ (MSDAD∗ ) = {2λ + α 2 − β 2 , 2λ + (α − β)2 , 2λ + (α + β)2 }.
Since 2λ + α 2 − β 2 ≤ max{2λ + (α − β)2 , 2λ + (α + β)2 }, from Proposition 1 we conclude:
 A. Tocino, M.J. Senosiain / Journal of Computational and Applied Mathematics 236 (2012) 2660–2672 2665

Fig. 1. Geometrical representation of the MS-stability region of the test system (6) in the SAD case.

Fig. 2. Geometrical representation of the MS-stability region of the test system in the SDAD* case.

Corollary 6. The system (6) in the SDAD* case is asymptotically MS-stable if and only if

max{2λ + (α − β)2 , 2λ + (α + β)2 } < 0. (13)

The MS-stability domain of a SDAD* system can be identified with the MS-stability region

SSDAD∗ = {(λ, α, β) ∈ R3 : condition (13) holds}. (14)

A geometrical representation of SSDAD∗ can be obtained in R3(u,v,w) , taking

√ √
u = λ ∆, v = α ∆, w=β ∆ (15)

with any ∆ > 0. Condition (13) of MS-stability becomes

2u + (v − w)2 < 0; 2u + (v + w)2 < 0. (16)

The set of points satisfying (16) is the solid limited by the grid shown in Fig. 2.
2666 A. Tocino, M.J. Senosiain / Journal of Computational and Applied Mathematics 236 (2012) 2660–2672

5. Numerical MS-stability

We now aim to offer a criterion of numerical MS-stability. When a numerical method is applied with step ∆ > 0 to solve
the linear test equation (2), a recurrence of the form
Xn+1 = F (∆, 1Wn )Xn (17)
is obtained. By analogy with the definition for the SDE, we say that the numerical scheme is asymptotically stable in the
mean square sense for a selected step-size ∆ > 0 if the numerical solution produced, {Xn }, satisfies
lim E ∥Xn ∥2 = 0.
n→∞

This condition is equivalent, see [14], to the stability of the second moment of Xn :
lim ∥E [Xn Xn′ ]∥ = 0,
n→∞
j
which in turn is equivalent to Y n → 0, where Y n represent a vector with the d(d + 1)/2 different components E [Xni Xn ], i, j =
j j
1, . . . , d. Taking expected values in the relation between Xni +1 X n +1 and Xni Xn obtained from (17) gives a one-step difference
equation of the form

Y n +1 = M Y n . (18)
Since, see [16], Y n → 0 if and only if

ρ(M) < 1, (19)

where ρ(M ) := max{|λ| : λ ∈ σ (M )} denotes the spectral radius of the matrix M , M is called the stability matrix of the
scheme and we can give the following:

Proposition 7. The numerical scheme applied with step ∆ > 0 produces an MS-stable numerical solution iff the stability matrix
of the scheme satisfies ρ(M ) < 1.

Remark. In [11], the conditions of numerical MS-stability with respect to ∥ · ∥p can be found. An important drawback of
these conditions is their dependence on the chosen norm. Notice also that given a matrix A, since ρ(A) ≤ ∥A∥ for any sub-
multiplicative norm, the conditions of MS-stability w.r.t. ∥ · ∥p , ∥M ∥p < 1, given in [11] are sufficient, but not equivalent,
for numerical MS-stability.
Given ∆ > 0, the set of pairs of matrices A, B for which a method applied with step ∆ produces a MS-stable solution
SM (∆) = {A, B ∈ Rd×d : condition (19) holds}
will be called the domain of MS-stability of the method. A scheme will be called MS-stable if SM (∆) ⊆ SSDE for any ∆ > 0,
that is, whenever the SDE is stable then so is the method applied with any step-size.

6. Numerical MS-stability of θ -methods

In the autonomous scalar case, Higham [4] showed how θ -methods, obtained by the introduction of implicitness in the
drift of Euler scheme,
Xn+1 = Xn + (1 − θ )f (Xn ) + θ f (Xn+1 ) + g (Xn )1Wn ,
are semi-implicit schemes that show better stability behavior with the increase in θ ∈ [0, 1]. In the bi-dimensional case,
we shall consider the counterpart semi-implicit schemes

Xn1+1 Xn1 1 − θ1 f 1 (Xn ) θ1 f 1 (Xn+1 ) g 1 (Xn )

           
0 0
= + + ∆+ 1Wn , (20)
Xn2+1 Xn2 0 1 − θ2 f 2 (Xn ) 0 θ2 f 2 (Xn+1 ) g 2 (Xn )
with θ1 , θ2 ∈ [0, 1]. Note that θ1 = θ2 = 0 give the Euler scheme and θ1 = θ2 = 1 gives a fully implicit drift scheme.
Applying (20) to the test system (6) leads to (18), with Y n = (E [(Xn1 )2 ], E [(Xn2 )2 ], E [Xn1 Xn2 ]) and
M(θ )
α1 2 ∆ + (1 + (1 − θ1 )λ1 ∆)2 β1 2 ∆ 2α1 β1 ∆
 

 (1 − θ1 λ1 ∆)2 (1 − θ1 λ1 ∆)2 (1 − θ1 λ1 ∆)2 

 
 β 2
2
∆ α2
2
∆ + (1 + ( 1 − θ2 )λ2 ∆) 2
2 α β
2 2 ∆ .

=
(1 − θ2 λ2 ∆)2 (1 − θ2 λ2 ∆)2 (1 − θ2 λ2 ∆)2
 
 
 
 α1 β2 ∆ α2 β1 ∆ α1 α2 ∆ + β1 β2 ∆ + (1 + (1 − θ1 )λ1 ∆)(1 + (1 − θ2 )λ2 ∆) 
(1 − θ1 λ1 ∆)(1 − θ2 λ2 ∆) (1 − θ1 λ1 ∆)(1 − θ2 λ2 ∆) (1 − θ1 λ1 ∆)(1 − θ2 λ2 ∆)
(21)
 A. Tocino, M.J. Senosiain / Journal of Computational and Applied Mathematics 236 (2012) 2660–2672 2667

6.1. The SD case

In the SD case (β1 = β2 = 0) we have:

 α 2 ∆ + (1 + (1 − θ )λ ∆)2 
1 1 1
0 0
 (1 − θ1 λ1 ∆)2 
α2 ∆ + (1 + (1 − θ2 )λ2 ∆)
 
2 2
M SD (θ ) = 
 
0 0 ;
(1 − θ2 λ2 ∆)2
 
 
 α1 α2 ∆ + (1 + (1 − θ1 )λ1 ∆)(1 + (1 − θ2 )λ2 ∆) 
0 0
(1 − θ1 λ1 ∆)(1 − θ2 λ2 ∆)
since
 α1 α2 ∆ + (1 + (1 − θ1 )λ1 ∆)(1 + (1 − θ2 )λ2 ∆) 
 
 
 (1 − θ1 λ1 ∆)(1 − θ2 λ2 ∆) 
α1 ∆ + (1 + (1 − θ1 )λ1 ∆) α2 ∆ + (1 + (1 − θ2 )λ2 ∆)2
 2 2 2

≤ max , ,
(1 − θ1 λ1 ∆)2 (1 − θ2 λ2 ∆)2
condition (19) of numerical MS-stability of the θ -method in the SD case, becomes
α1 2 ∆ + (1 + (1 − θ1 )λ1 ∆)2 α2 2 ∆ + (1 + (1 − θ2 )λ2 ∆)2
 
max , < 1,
(1 − θ1 λ1 ∆)2 (1 − θ2 λ2 ∆)2
i.e. the stability reduces to the MS-stability of the one-dimensional θi -methods applied to the scalar test equations dXt =
λi Xt dt + αi Xt dWt , i = 1, 2 respectively, see [1,6,9].

6.2. The SAD case

In the SAD case (α1 = α2 = 0, β1 = β2 = β ), we have

 (1 + (1 − θ )λ ∆)2 β 2∆ 
1 1
0
 (1 − θ1 λ1 ∆)2 (1 − θ1 λ1 ∆)2 
β 2∆ (1 + (1 − θ2 )λ2 ∆)2
 
MSAD (θ ) =  ,
 
 (1 − θ λ ∆)2 0
 2 2 (1 − θ2 λ2 ∆)2 

(1 + (1 − θ1 )λ1 ∆)(1 + (1 − θ2 )λ2 ∆) + β 2 ∆
 
0 0
(1 − θ1 λ1 ∆)(1 − θ2 λ2 ∆)
with eigenvalues
(1 + (1 − θ1 )λ1 ∆)(1 + (1 − θ2 )λ2 ∆) + β 2 ∆
r1 = ,
(1 − θ1 λ1 ∆)(1 − θ2 λ2 ∆)
(1 + (1 − θ1 )λ1 ∆)2 (1 − θ2 λ2 ∆)2 + (1 + (1 − θ2 )λ2 ∆)2 (1 − θ1 λ1 ∆)2
r2± =
2(1 − θ1 λ1 ∆)2 (1 − θ2 λ2 ∆)2
[(1 + (1 − θ1 )λ1 ∆)2 (1 − θ2 λ2 ∆)2 − (1 + (1 − θ2 )λ2 ∆)2 (1 − θ1 λ1 ∆)2 ]2 + 4β 4 ∆2 (1 − θ1 λ1 ∆)2 (1 − θ2 λ2 ∆)2

± ,
2(1 − θ1 λ1 ∆)2 (1 − θ2 λ2 ∆)2
and the numerical MS-stability condition (19) becomes max{|r1 |, |r2+ |, |r2− |} < 1. Since r1 ≤ r2+ and |r2− | ≤ r2+ , the
asymptotic MS-stability of a θ -method is equivalent to conditions

− 1 < r1 , r2+ < 1. (22)

We shall analyze the stability behavior of each θ -method with θ1 = θ2 = θ , comparing its MS-stability region
S(θ,θ) (∆) = {(λ1 , λ2 , β) ∈ R3 : conditions (22) hold}
with the MS-stability region SSAD of the problem.

Theorem 8. For all ∆ > 0

1
S(θ,θ) (∆) ⊂ SSAD if 0 ≤ θ < ;
2
S  ( ∆) = SSAD ;
1 1
,
2 2

1
S(θ,θ) (∆) ⊃ SSAD if < θ ≤ 1.
2
2668 A. Tocino, M.J. Senosiain / Journal of Computational and Applied Mathematics 236 (2012) 2660–2672

Proof. For the sake of simplicity, we shall write

M = 1 + (1 − θ )λ1 ∆, N = 1 + (1 − θ )λ2 ∆, R = 1 − θ λ1 ∆, S = 1 − θ λ 2 ∆, F = β 2 ∆.

With this notation

(M 2 S 2 − N 2 R2 )2 + 4F 2 R2 S 2

MN + F M 2 S 2 + N 2 R2 +
r1 = ; +
r2 = .
RS 2R2 S 2

We shall first prove that S(θ,θ) (∆) ⊆ SSAD if 0 ≤ θ ≤ 1/2. We have to see that conditions (22) imply condition (10) for a
(λ1 , λ2 , β) ∈ R3 . Since r2+ < 1 we have that

(M 2 − R2 )(N 2 − S 2 ) > F 2 ≥ 0; (23)

S 2 (M 2 − R2 ) + R2 (N 2 − S 2 ) ≤ 0.

These inequalities lead to M 2 − R2 < 0 and N 2 − S 2 < 0, i.e.,

(1 − 2θ )λ21 ∆2 + 2λ1 ∆ < 0, (1 − 2θ )λ22 ∆2 + 2λ2 ∆ < 0.

Then λ1 < 0, λ2 < 0 and

(1 − 2θ )(λ21 + λ22 )∆ + 2(λ1 + λ2 ) < 0.

Hence

1
(1 − 2θ )λ1 λ2 ∆ + λ1 + λ2 ≤ (1 − 2θ )(λ21 + λ22 )∆ + 2(λ1 + λ2 ) < 0.

(24)
2

Rewriting (23) as

β 4 < λ1 λ2 (2 + (1 − 2θ )λ1 ∆)(2 + (1 − 2θ )λ2 ∆),

leads to

β 4 + (λ1 − λ2 )2 < λ1 λ2 (2 + (1 − 2θ )λ1 ∆)(2 + (1 − 2θ )λ2 ∆) + (λ1 − λ2 )2

= ((1 − 2θ )λ1 λ2 ∆ + (λ1 + λ2 ))2 .

From this last inequality, using (24)


β 4 + (λ1 − λ2 )2 < − ((1 − 2θ )λ1 λ2 ∆ + λ1 + λ2 ) ,

which gives

λ1 + λ2 + β 4 + (λ1 − λ2 )2 < −(1 − 2θ )λ1 λ2 ∆ < 0,

and the first part of the proof is complete.

We shall now prove the inclusion SSAD ⊆ S(θ,θ) (∆) for 1/2 ≤ θ ≤ 1. If (λ1 , λ2 , β) satisfies (10) then λ1 , λ2 < 0. Now,
condition −1 < r1 is equivalent to

−(1 − θ λ1 ∆)(1 − θ λ2 ∆) < (1 + (1 − θ )λ1 ∆)(1 + (1 − θ )λ2 ∆) + β 2 ∆,

which can be written

∆(2θ − 1)(λ1 + λ2 ) < 2 + ∆β 2 + (1 − 2θ + 2θ 2 )λ1 λ2 ∆2 . (25)

Since 1/2 ≤ θ ≤ 1 implies that 1/2 ≤ 1 − 2θ + 2θ 2 ≤ 1, it is obvious that (25) holds, because ∆(2θ − 1)(λ1 + λ2 ) ≤ 0
whereas the term on the right of (25) is positive.
To see that r2+ < 1, using the above notation, we must prove that

(M 2 S 2 − N 2 R2 )2 + 4F 2 R2 S 2 < 2R2 S 2 − (M 2 S 2 + N 2 R2 ). (26)
 A. Tocino, M.J. Senosiain / Journal of Computational and Applied Mathematics 236 (2012) 2660–2672 2669

Fig. 3. Comparison between the MS-stability region of the SAD test equation (gridded) and the stability regions of θ -methods with θ1 = θ2 = 0 (top left),
1
4
(top right), 12 (bottom left) and 1 (bottom right).

Notice that R2 −M 2 = −2λ1 ∆+(2θ −1)λ21 ∆2 > 0 and S 2 −N 2 = −2λ2 ∆+(2θ −1)λ22 ∆2 > 0. Then 2R2 S 2 −(M 2 S 2 +N 2 R2 ) =
(R2 − M 2 )S 2 + (S 2 − N 2 )R2 > 0, and (26) is equivalent to
2
(M 2 S 2 − N 2 R2 )2 + 4F 2 R2 S 2 < 2R2 S 2 − (M 2 S 2 + N 2 R2 ) ,


which can be written as:

β 4 < λ1 λ2 (2 + (1 − 2θ )λ1 ∆)(2 + (1 − 2θ )λ2 ∆). (27)

From (10), β 4 < 4λ1 λ2 , and hence:

β4 < 4λ1 λ2
4 + 2(1 − 2θ )(λ1 + λ2 )∆ + (1 − 2θ )2 λ1 λ2 ∆2 λ1 λ2
 
≤
= (2 + (1 − 2θ )λ1 ∆)(2 + (1 − 2θ )λ2 ∆)λ1 λ2 ,

which gives (27). 

Theorem 8 shows for 0 ≤ θ < 1/2 that if the SAD test equation is unstable then so is the θ -method for all ∆ > 0. In the
case θ ≥ 1/2, Theorem 8 proves that whenever the SAD equation is stable then so is the θ -method for any ∆ > 0. These
results extend the MS-stability properties of one-dimensional stochastic θ -methods, see [4,5].
Using the variables defined in (11), a geometrical representation of the MS-stability region of each θ -method can be
obtained in R3(x,y,z ) and compared with the MS-stability region of the SAD test system shown in Fig. 1. For example, if θ = 0
(Euler method), S(0,0) (∆) ⊂ SSAD for all ∆ > 0. The MS-stability region of the Euler scheme (shaded solid) contained in the
MS-stability region of the test equation (grid solid) has been plotted in Fig. 3 (top left). In a similar way, Fig. 3 confirms that
for any ∆ > 0, S( 1 , 1 ) (∆) ⊂ SSAD (top right), S( 1 , 1 ) (∆) = SSAD (bottom left) and SSAD ⊆ S(1,1) (∆) (bottom right).
4 4 2 2
2670 A. Tocino, M.J. Senosiain / Journal of Computational and Applied Mathematics 236 (2012) 2660–2672

6.3. The SDAD* case

For the SDAD* case, α1 = α2 = α, β1 = β2 = β and λ1 = λ2 = λ. If we restrict our attention to the behavior of
θ -methods with θ1 = θ2 = θ , (21) becomes
 α 2 ∆ + (1 + (1 − θ )λ∆)2 β 2∆ 2αβ ∆

 (1 − θ λ∆)2 (1 − θ λ∆)2 (1 − θ λ∆)2 
β ∆ α ∆ + (1 + (1 − θ )λ∆) 2αβ ∆
 2 2 2

MSDAD∗ (θ ) = 
 


 (1 − θ λ∆)2 (1 − θ λ∆)2 (1 − θ λ∆)2 

αβ ∆ αβ ∆ α 2 ∆ + β 2 ∆ + (1 + (1 − θ )λ∆)2
 
(1 − θ λ∆)2 (1 − θ λ∆)2 (1 − θ λ∆)2
with eigenvalues

(α 2 − β 2 )∆ + (1 + (1 − θ )λ∆)2
s1 = ;
(1 − θ λ∆)2
(α − β)2 ∆ + (1 + (1 − θ )λ∆)2
s2 = ;
(1 − θ λ∆)2
(α + β)2 ∆ + (1 + (1 − θ )λ∆)2
s3 = .
(1 − θ λ∆)2
Now, the numerical MS-stability condition (19) becomes max{|s1 |, |s2 |, |s3 |} < 1; since s2 ≥ 0, s3 ≥ 0 and |s1 | ≤ max{s2 , s3 }
the asymptotic MS-stability of a θ -method is equivalent to conditions s2 < 1, s3 < 1, i.e.,

max{(α ± β)2 ∆ + (1 + (1 − θ )λ∆)2 } < (1 − θ λ∆)2 ,

which can be written

max{(α ± β)2 + 2λ + λ2 ∆(1 − 2θ )} < 0. (28)

To analyze the stability behavior of each θ -method, we compare its MS-stability region
∗
S(θ,θ) (∆) = {(λ, α, β) ∈ R3 : condition (28) holds}
with the MS-stability region SSDAD∗ given in (14). Direct comparison between conditions (13) and (28) gives:

Theorem 9. For all ∆ > 0, we have:

1
∗
S(θ,θ) (∆) ⊂ SSDAD∗ if 0 ≤ θ < ;
2
S∗  (∆) = SSDAD∗ ;
1 1
,
2 2

1
∗
S(θ,θ) (∆) ⊃ SSDAD∗ if < θ ≤ 1.
2
For 0 ≤ θ < 1/2, given (λ, α, β) ∈ SSDAD∗ , the θ -method is MS-stable if and only if

−2λ − (α ± β)2
 
∆ < min .
λ2 (1 − 2θ )

For 0 ≤ θ < 1/2 Theorem 9 shows that (a) if the SDAD* test equation is unstable then so is the θ -method for all ∆ > 0,
and (b) if the SDAD* is stable then so is the θ -method for a sufficiently small ∆. In the case θ ≥ 1/2, Theorem 9 proves that
whenever the SDAD* equation is stable then so is the θ -method for any ∆ > 0. These properties are an extension of the
MS-stability of scalar stochastic theta methods presented in [4,5], which in turn is a direct generalization of deterministic
A-stability. A similar result in the bi-dimensional case with a test equation with two independent Wiener noises is reported
in [13].
Using the variables defined in (15), a geometrical representation of the MS-stability region of each θ -method can be
obtained in R3(u,v,w) ; it can also be compared with the MS-stability region of the SDAD* test system shown in Fig. 2. For
example, if θ = 0 (Euler method), S(∗0,0) (∆) ⊂ SSDAD∗ for all ∆ > 0. The MS-stability region of the Euler scheme (shaded
solid) contained in the MS-stability region of test equation (grid solid) has been plotted in Fig. 4, top left. In a similar way,
Fig. 4 confirms that for any ∆ > 0, S ∗1 1 (∆) ⊂ SSDAD∗ (top right), S ∗1 1 (∆) = SSDAD∗ (bottom left) and SSDAD∗ ⊆ S(∗1,1) (∆)
(4,4) (2,2)
(bottom right).
 A. Tocino, M.J. Senosiain / Journal of Computational and Applied Mathematics 236 (2012) 2660–2672 2671

Fig. 4. Comparison between the MS-stability region of the SDAD* test equation (gridded) and the stability regions of θ -methods with θ = (0, 0) (top left),
( 41 , 14 ) (top right), ( 21 , 12 ) (bottom left) and (1, 1) (bottom right).

7. Conclusions and future issues

In sum, we have studied MS-stability for multidimensional systems with scalar noise. Criteria that do not depend on any
norm have been proposed for the analysis of the MS-stability of the linear test equation and for the numerical approximation.
In addition, for two-dimensional systems, simple necessary and sufficient conditions of MS-stability depending on the
real parameters of the drift and diffusion matrices have been given. Finally, the numerical MS-stability of θ -methods has
been investigated. We plan to follow this work by extending this MS-stability analysis to systems of higher dimension and
addressing our research in stochastic linear systems with complex parameters in the drift and diffusion coefficients.

Acknowledgments

The second author was supported by Ministerio de Ciencia e Innovación (DGICYT) under the Project MTM2009-09676.

References

[1] E. Buckwar, T. Sickenberger, A comparative linear mean-square stability analysis of Maruyama—and Milstein-type methods, Math. Comput. Simul. 81
(2011) 1110–1127.
[2] K. Burrage, T. Tian, A note on the stability properties of the Euler methods for solving stochastic differential equations, New Zealand J. Math. 29 (2)
(2000) 115–127.
[3] D.B. Hernández, R. Spigler, Convergence and stability of implicit Runge–Kutta methods for systems with multiplicative noise, BIT 33 (1993) 654–669.
[4] D.J. Higham, Mean-square and asymptotic stability of the stochastic theta method, SIAM J. Numer. Anal. 38 (2000) 753–769.
[5] D.J. Higham, A-stability and stochastic mean-square stability, BIT 40 (2000) 404–409.
[6] Y. Saito, T. Mitsui, Stability analysis of numerical schemes for stochastic differential equations, SIAM J. Numer. Anal. 33 (6) (1996) 2254–2267.
[7] H. Schurz, Asymptotical mean square stability of an equilibrium point of some linear numerical solutions with multiplicative noise, Stoch. Anal. Appl.
14 (1996) 313–354.
[8] J.D. Lambert, Numerical Methods for Ordinary Differential Systems, John Wiley & Sons, Chichester, 1991.
[9] Y. Komori, T. Mitsui, Stable ROW-type weak scheme for stochastic differential equations, Monte Carlo Methods Appl. 1 (4) (1995) 279–300.
[10] T. Mitsui, Y. Saito, MS-stability analysis for numerical solutions of stochastic differential equations-beyond single-step single dim, in: Some Topics in
Industrial and Applied Mathematics, in: Ser. Contemp. Appl. Math. CAM, vol. 8, Higher Ed. Press, Beijing, 2007, pp. 181–194.
2672 A. Tocino, M.J. Senosiain / Journal of Computational and Applied Mathematics 236 (2012) 2660–2672

[11] Y. Saito, T. Mitsui, Mean-square stability of numerical schemes for stochastic differential systems, Vietnam J. Math. 30 (2002) 551–560.
[12] M.I. Abukhaled, Mean square stability of a class of Runge–Kutta methods for 2-dimensional stochastic differential systems, Appl. Numer. Anal. Comput.
Math. 1 (2004) 77–89.
[13] E. Buckwar, C. Kelly, Towards a systematic linar stability analysis of numerical methods for systems of stochastic differential equations, SIAM J. Numer.
Anal. 48 (2010) 298–321.
[14] L. Arnold, Stochastic Differential Equations: Theory and Applications, John Wiley & Sons, New York, 1974.
[15] R.Z. Has’minskii, Stochastic Stability of Differential Equations, Sijthoff & Noordhoff, Alphen aan der Rijn, 1980.
[16] P. Deuflhard, F. Bornemann, Scientific Computing with Ordinary Differential Equations, Springer, New York, 2002.
[17] F.R. Gantmacher, The Theory of Matrices, vol. 2, Chelsea Publishing Co., New York, 1959.