1.

Numerical solution of 1D PDE

To begin with, suppose that U ≝ U (t , x ) is some well-defined function on ( t , x ) ∈ [ 0 , T ] × [ x , x ] satisfying

the following 1D PDE

( ∂t + μ ( t , x ) ∂x +σ ( t , x ) ∂2x −r ( t , x ) ) U =0 , (1)

defined on ( t , x ) ∈ ( 0 , T ) × ( x , x ), with some well-defined terminal condition U ( T , x )=f ( x ) for x ∈ [ x , x ].

The conditions of U ( t , x ) and U ( t , x ), where t ∈ ( 0 , T ) , are determined by the boundary conditions.
To numerically compute the solution of Equation (1), we start with the discretization of the time domain and
n
the space domain. Discretize the time domain with the time grid { t k }k=1
t
, where t 1=0 and t n =T , and the
t

nx
space domain with the space grid { x }
j j=1
, where x 1=x and x nx =x . Note that we allow non-uniform grids.
−1
Now, call x 1 the lower boundary, x n the upper boundary, and the set of points { x j }nj=2
x
x
the interior region. For
k k k k
any time point t k , U j ≝ U ( t k , x j ), μ j ≝ μ ( t k , x j ), σ j ≝ σ ( t k , x j ) and r j ≝ r ( t k , x j ), where j=1 ,… , n x and
k =1 , … ,n t .
To approximate Equation (1), it is necessary to approximate the first-order and second-order partial derivatives
through discretization. Let Δ x j ≝ x j+ 1−x j , where j=1 ,… , n x −1. For j=2 , … ,n x −1, the first-order
partial derivative at x j is approximated by

δ x U j =α j−1 U j−1 +α j U j +α j +1 U j+1 , (2)

where
Δ xj
α j−1∧¿− ,
Δ x j−1 ( Δ x j −1 + Δ x j )
Δ x j−1
α j +1∧¿ ,
Δ x j ( Δ x j −1 + Δ x j )
α j∧¿−( α j −1 +α j +1 ) ,
while the second-order partial derivative at x j is approximated by
2
δ x U j =β j−1 U j−1 + β j U j + β j +1 U j+1 , (3)

where
2
β j−1∧¿ ,
Δ x j−1 ( Δ x j−1+ Δ x j )
2
β j +1∧¿ ,
Δ x j ( Δ x j−1+ Δ x j )
β j∧¿−( β j−1+ β j+1 ) .
Note that α j ’s and β j ’s depend only on j but not time.
Substitute the approximated partial derivatives into the PDE and obtain the following discretization equation
2
∂t U j∧¿−μ j δ x U j−σ j δ x U j +r j U j
¿=−μ j ( α j−1 U j −1 +α j U j+ α j+1 U j +1 )−σ j ( β j−1 U j−1 + β j U j + β j+1 U j +1 )+ r j U j
¿=−( μ j α j−1+ σ j β j−1 ) U j−1−( μ j α j +σ j β j−r j ) U j −( μ j α j +1+ σ j β j +1 ) U j+1 ,
for j=2 , … ,n x −1. For convenience, let

1
 c j∧≝ μ j α j−1+ σ j β j−1 ,
a j∧ ≝ μ j α j + σ j β j−r j ,
b j∧ ≝ μ j α j+1 +σ j β j+1 ,
which results in
∂t U j=−c j U j−1−a j U j−b j U j+ 1 . (4)
The time index will be specified later as it depends on the time discretization scheme used.
Next, we consider the boundary conditions.

1.1 Null-gamma condition

The Null-gamma condition states that on boundary, that is, for t ∈ ( 0 , T ) and, x=x or x=x ,
2 2
∂ x U =∂ x U when x is the log-spot or ∂ x U =0 when x is the spot.
Assume that x is the spot. On the lower boundary x=x , the Null gamma condition is given by
2
δ x U 2∧¿ 0
β β
U 1∧¿− 2 U 2− 3 U 3 .
β1 β1
For convenience, write
U 1 = λ2 U 2 + λ 3 U 3 . (5)

The key observation is that U 1 is a linear combination of U 2 and U 3.

Again, on the upper boundary x=x , the Null gamma condition is given by
2
δ U n −1∧¿ 0 x

β n −1 β n −2
U n ∧¿− U n −1−
x
U n −2 . x

x
βn x
βn x
x
x

For convenience, write

U n =λn −1 U n −1 + λ n −2 U n −2 .
x x x x x
(6)

The key observation is that U n x is a linear combination of U n x −1 and U n x −2 .

Given that the solution at the current time point t k+1 is known, we are ready to calculate the solution at the
previous point t k . We could proceed with either the explicit scheme or the implicit scheme. First, we consider
the explicit scheme.
Explicit scheme
After discretization in time domain, the explicit scheme for Equation (4) is given by
k+1 k
U j −U j k +1 k +1 k+1 k+1 k+ 1 k +1
=−c j U j−1−a j U j −b j U j+1 ,
Δtk
for j=2 , … ,n x −1, where Δ t k =t k +1−t k .
k +1 k +1
Since U 1 and U n x could be calculated by the boundary conditions as in Equations (5) and (6) above,

j + Δ t k ( c j U j−1 +a j U j +b j U j +1 ) ,
U kj =U k+ 1 k+1 k+1 k+1 k+1 k+1 k+1
(7)

for j=2 , … ,n x −1.

2
 Next, we consider the implicit scheme.
Implicit scheme
After discretization in time domain, the implicit scheme for Equation (4) is given by
k+1 k
U j −U j k k k k k k
=−c j U j−1−a j U j −b j U j +1 .
Δtk
Now, Equation (7) could no longer be applied because
k k k
U 1∧¿ λ 2 U 2 + λ3 U 3 ,
U n ∧¿ λn −1 U kn −1 + λ n −2 U nk −2 ,
k
x x x x x

are not known at the current time point t k . However, we can use these two equations to modify the
coefficients on the boundaries and arrive at a solvable system of linear equations.
Set j=2 and apply Equation (5). The discretized PDE becomes
k+1 k
U 2 −U 2 k k k k
∧¿−c 2 U 1−a 2 U 2 −b2 U 3
Δtk
¿=−( a2 + λ 2 c 2 ) U 2 −( b 2+ λ3 c 2 ) U 3 .
k k k k k k

For convenience, denote

~ k k
a 2∧ ≝ a2 + λ2 c2 ,
k

~k
b 2∧ ≝ b2k + λ3 c k2 .
Set j=n x −1 and apply Equation (6). The discretized PDE becomes
k +1 k
U n −1−U n −1 k k k k k k
x
∧¿−cn −1 U n −2−an −1 U n −1−b n −1 U n
x

Δ tk x x x x x x

¿=−( cn −1 + λ n −2 bn −1 ) U n −2 + ( an −1 + λ n −1 b n −1 ) U n −1 .
k k k k k k
x x x x x x x x

For convenience, denote

~c k ∧≝ c k + λ bk ,
n −1 x
n −1 n −2 n −1 x x x

~
a n −1∧≝ an −1 + λ n −1 b kn −1 .
k k
x x x x

Therefore, the coefficients on the boundaries are modified.

Now, the discretized PDE could be written in matrix form

[ ][ ]
k k+1
U2 U2
k k+1

[ ])
U3 U3

(
~ak ~ k
¿ ¿c
k k k
¿ ~k
2 b2 3 a 3 b 3
I n −2− Δ t k k k
an −1 ¿ ⋮ = ⋮ , (8)
x
⋱ ⋱ ⋱ ¿c a ¿ ¿ x
k k +1
n x −2 n x −2
U nx −2 U nx −2
k k +1
U nx −1 U nx −1

k ~k
a 2, b 2 , ~
I n −2 is the identity matrix of order n x −2. Note the modified coefficients ~ c n −1 and ~
k k
where x
a n −1. The x x

key observation is that the matrix in the bracket on the left-hand side is tri-diagonal. Hence, the above system
of equations could be easily solved by standard tri-diagonal matrix solver.

1.2 Dirichlet conditions

The treatment of Dirichlet conditions is much simpler and the discretized form is given by

3
 k k
U 1 ∧¿ g ,
U kn ∧¿ hk ,
x

k k
where g ’s and h ’s are known for all k ’s.
Explicit scheme
Since the solution values on the boundaries are known at the current time point t k+1, the explicit scheme
k +1 k +1 k +1 k+1
proceeds as in Equation (7) , where U 1 =g and U n x =h .

Implicit scheme
The implicit scheme for Equation (4) is given by
k+1 k
U j −U j k k k k k k
=−c j U j−1−a j U j −b j U j +1 ,
Δtk
for j=2 , … ,n x −1.

Set j=2. The discretized PDE becomes

k+1 k
U 2 −U 2 k k k k k k
∧¿−c 2 U 1−a 2 U 2 −b2 U 3
Δtk
¿=−c k2 g k −a k2 U k2−b k2 U k3 .
Set j=n x −1. The discretized PDE becomes
k+1 k
U n −1−U n −1 k k k k k k
x
∧¿−c n −1 U n −2−a n −1 U n −1−bn −1 U n
x

Δtk x x x x x x

k k k k k k
¿=−c nx −1 U nx −2 −a nx −1 U nx −1 −b n x −1 h .
Now, the discretized PDE could be written in matrix form

[ ][ ] [ ]
k k+1
U2 U2 c2 g
k k

U k3 U 3k+1

( [ ])
ak2 b k2 ¿ ¿ ck3 ak3 bk3 ¿ k 0
I n −2− Δ t k an −1 ¿ ⋮ = ⋮ + Δtk ⋮ . (9)
x
⋱ ⋱ ⋱ ¿ ckn −2 akn −2 ¿ ¿ x
k k +1
0
x x
U nx −2 U nx −2
k
U k
U k +1 b nx −1 hk
nx −1 nx −1

k k k k
Note that the coefficients of the matrix near boundaries (that is, a 2, b 2, c nx −1 and a nx −1) need not be modified
but a source term matrix is added to the left-hand side. Similar to the case of Null-gamma condition, the
resulting system of equations could be easily solved by standard tri-diagonal matrix solver.

1.3 θ-scheme
For convenience, let

[] [ ]
U k2 k
c2 g
k

U k3

[ ]
ak2 bk2 ¿ ¿ c k3 a3k b3k ¿ k 0
k k k
U = ⋮ ,L = k k
an −1 ¿ and F = ⋮ .
k ⋱ ⋱ ⋱ ¿ c n −2 a n −2 ¿ ¿ x

0
U n x −2
x x

U k b kn −1 hk
n x −1 x

4
where the matrices F k is the source term matrix added for Dirichlet conditions (if applicable).
Now, the explicit scheme could be written as

U k =U k+ 1+ Δ t k ( Lk +1 U k+1 + F k+1 ) ,
while the implicit scheme could be written as

( I n −2 −Δ t k Lk ) U k =U k+1 + Δ t k F k .
x

The θ -scheme is a mixture of explicit scheme and implicit scheme, and is given by
k+1 k
U −U
=−( 1−θ ) ( L U + F )−θ ( L U + F ) ,
k+1 k+1 k+1 k k k
Δtk
where the matrices F k and F k+1 are the source term matrices added for Dirichlet conditions (if applicable).
After rearrangement,

( I −θ Δ t k Lk ) U k −θ Δ t k F k =U k+1 + ( 1−θ ) Δ tk ( Lk +1 U k +1 + Fk +1 ) .
The above problem could be solved by using first the explicit scheme and then the implicit scheme.
In the first step, the right-hand side is equivalent to solving the following explicit scheme:
~k
U =U k+1 + ( 1−θ ) Δ t k ( Lk +1 U k+1 + F k+1 ) ,
where the size of the time step is scaled by ( 1−θ ).
In the second step, the left-hand side is equivalent to solving the following implicit scheme:

( I n −2 −θ Δ t k Lk ) U k −θ Δ t k F k =~
x
Uk ,
where the size of the time step is scaled by θ .

5
2. Numerical solution of 2D PDE

Now, suppose that U ≝ U (t , x , y ) is some well-defined function on ( t , x , y ) ∈ [ 0 , T ] × [ x , x ] × [ y , y ]

satisfying the following 2D PDE

( ∂t + μ x ( t , x , y ) ∂ x +σ x ( t , x , y ) ∂ 2x + μ y ( t , x , y ) ∂ y +σ y ( t , x , y ) ∂ 2y + ρ ( t , x , y ) ∂x ∂ y −r ( t , x , y ) ) U=0
(10) ,

defined on ( t , x ) ∈ ( 0 , T ) × ( x , x ) × ( y , y ) , with some well-defined terminal condition U ( T , x , y ) =f ( x )

for ( x , y ) ∈ [ x , x ] × [ y , y ]. The conditions of U ( t , x , y ) , U ( t , x , y ) , U ( t , x , y ) and U ( t , x , y ) , where
t ∈ ( 0 , T ) , are determined by the boundary conditions. Note that in our application, variable x refers to the
spot / log-forward moneyness, and variable y refers to variance.
To numerically compute the solution of Equation (10), we start with the discretization of the time domain and
n
the space domains. Discretize the time domain with the time grid { t k }k=1
t
, where t 1=0 and t n =T , and the
t

nx ny
space domains with the x -grid { x } where x 1=x and x n =x , and y -grid { y }
j j=1
, where y 1= y and
x l l =1
,
y n = y . Note that we allow non-uniform grids. Now, call x 1 the lower boundary of x -grid, x n the upper
y x

boundary of x -grid, y 1 the lower boundary of y -grid, y n , the upper boundary of y -grid, and the set of points
y

( x j , y l ), where j=2 , … ,n x −1 and l=2 , … , n y −1 , the interior region. For any time point t k ,
k x, k x x ,k x y,k y y ,k y
U j ,l ≝ U ( t k , x j , y l ), μ j ,l ≝ μ ( t k , x j , y l ), σ j ,l ≝ σ ( t k , x j , y l ), μ j ,l ≝ μ ( t k , x j , y l ), σ j ,l ≝ σ ( t k , x j , y l ),
k k
ρ j ,l ≝ ρ ( t k , x j , y l ) and r j ,l ≝ r ( t k , x j , y l ), where j=1 ,… , n x , l=1 , … , n y and k =1 , … ,n t .
Similar to the case of 1D PDE, to approximate Equation (10), it is necessary to approximate the first-order and
second-order partial derivatives through discretization. Let Δ x j ≝ x j+ 1−x j for j=1 ,… , n x −1, and
Δ yl ≝ y l+1− yl for l=1 , … , n y −1. For j=2 , … ,n x −1 and l=2 , … , n y −1 , the approximated first-
order partial derivative at x j is given by
x x x
δ x U j , l=α j−1 U j−1 , l+ α j U j ,l +α j +1 U j+1 , l , (11)

where

x Δ xj
α j−1∧¿− ,
Δ x j−1 ( Δ x j −1 + Δ x j )
x Δ x j−1
α j +1∧¿ ,
Δ x j ( Δ x j −1 + Δ x j )
α j ∧¿−( α j−1+ α j+1 ) ,
x x x

the approximated second-order partial derivative at x j is given by

x x x
δ x U j , l=β j−1 U j−1 , l+ β j U j ,l + β j +1 U j+1 , l , (12)

where
x 2
β j−1∧¿ ,
Δ x j−1 ( Δ x j−1+ Δ x j )
2
β xj +1∧¿ ,
Δ x j ( Δ x j−1+ Δ x j )
β xj ∧¿−( β xj−1 + β xj+1 ) ,
the approximated first-order partial derivative at y l is given by
y y y
δ y U j ,l =α l−1 U j ,l−1 +α l U j ,l +α l+1 U j ,l+1 , (13)

6
where

y Δ yl
α l−1∧¿− ,
Δ y l−1 ( Δ y l−1+ Δ y l )
y Δ y l −1
α l+1∧¿ ,
Δ y l ( Δ y l−1+ Δ y l )
α l ∧¿−( α l−1 +α l+1 ) ,
y y y

and the approximated second-order partial derivative at y l is given by

y y y
δ y U j ,l =β l−1 U j ,l−1 + β l U j ,l + β l+1 U j ,l+1 , (14)

where
y 2
β l−1∧¿ ,
Δ y l−1 ( Δ y l−1 + Δ y l )
y 2
β l+1∧¿ ,
Δ x j ( Δ y l−1+ Δ y l )
β l ∧¿−( βl −1 + β l+1 ) .
y y y

For the cross derivative, write down explicitly

δ x δ y U j ,l=¿ δ x ( α l−1
y
U j , l−1 + α ly U j , l+ α ly+1 U j ,l +1 )

[ ] ( [ ] [ ] [ ])
x '
α j−1 U j−1 ,l −1 U j−1 ,l U j−1 ,l+1
x y y y
¿∧ α j
α l−1 U j ,l−1 + α l U j , l +α l+1 U j ,l +1
x
(15)
α j+1
U j+1 , l−1 U j+1 ,l U j +1 ,l +1
j+1 l+1
¿∧
'
∑ ∑ '
x
α j α l U j ,l .
'
y
' ' '

j = j−1 l =l −1

With the approximated partial derivatives above, for j=2 , … ,n x −1 and l=2 , … , n y −1 , re-write Equation
(10) at point ( x j , y l ) as

( ∂t + μ xj ,l δ x + σ xj ,l δ2x + μ yj , l δ y +σ yj ,l δ 2y + ρ j , l δ x δ y −r j ,l ) U=0.
Substitute (11), (12), (13), (14) and (15) into the above equation and re-arrange the resulting equation in
vector

([ ] [ ]) [ ]
xy '
lc j ,l 0 U j−1 , l−1
∂t U j ,l =−¿ mc xy
j ,l
+ c yj ,l U j ,l−1
uc
xy
j ,l
0 U j+ 1, l−1

([ ] [ ] [ ]) [ ]
xy x '
la j ,l 0 c j ,l U j−1 ,l
xy
−¿ ma j,l
+ a j ,l + a xj ,l
y
U j ,l (16)
xy x
ua j,l
0 b j ,l U j+1 , l

([ ] [ ]) [ ]
'
lb xyj ,l 0 U j−1 ,l+1
−¿ mb xy
j ,l
+ b yj , l U j , l+1 ,
ub xy
j ,l
0 U j+1 ,l +1

7
 where, for clarity, we have separated the coefficients into those terms involving derivatives of x only, those
terms involving derivatives of y only, and those terms involving cross-derivatives. Specifically, the terms
involving derivatives of x only are
x x x x x
c j ,l =¿ α j−1 μ j ,l + β j−1 σ j ,l ,
a xj ,l =¿ α xj μ xj ,l + β xj σ xj ,l−r j , l ,
x x x x x
b j ,l=¿ α j+1 μ j ,l + β j+1 σ j ,l .
The terms involving derivatives of y only are
y y y y y
c j ,l=¿ α l−1 μ j ,l + β l−1 σ j , l ,
a yj ,l=¿ α ly μ yj ,l + β ly σ yj ,l ,
y y y y y
b j , l=¿ α l+1 μ j , l+ βl +1 σ j , l .
The terms involving cross-derivatives are
xy x y
lc j ,l∧¿ ρ j ,l α j−1 α l−1 ,
mcxyj , l∧¿ ρ j ,l α xj α l−1
y
,
xy x y
uc j , l∧¿ ρ j ,l α j+1 α l −1 ,
la xyj ,l∧¿ ρ j ,l α xj−1 α ly ,
xy x y
ma j , l∧¿ ρ j ,l α j α l ,
xy x y
ua j , l∧¿ ρ j ,l α j+1 α l ,
xy x y
lb j ,l∧¿ ρ j ,l α j−1 α l +1 ,
xy x y
mb j , l∧¿ ρ j ,l α j α l+ 1 ,
xy x y
ub j , l∧¿ ρ j ,l α j+1 α l +1 .
Next, consider the numerical solution at the corners and on the boundaries.

2.1 Numerical solutions at corners

Given limited information about the behavior of solution at extreme values of x and y , the condition at the
corners of the grid is that the cross derivative is zero. This condition provides the foundation of computing the
solution values at the 4 corners, namely, U 1 ,1 , U n x , 1, U 1 , ny and U n x , n y, given the solution values on the
boundaries and in the interior region.
Upper left-hand corner
To calculate U 1 ,1 , set ( j, l )=( 2 , 2 ) in Equation (15).
3 3
δ x δ y U 2 , 2=∑ ∑ α xj α ly U j ,l =0. ' ' ' '

j ' =1 l ' =1

Upon solving,
−1
U 1 ,1 = x y
α1 α 1
[
α 1 ( α 2 U 1 ,2 +α 3 U 1 ,3 ) + α 2 ( α 1 U 2 ,1 +α 2 U 2 ,2 +α 3 U 2 , 3 ) +α 3 ( α 1 U 3 ,1 +α 2 U 3 , 2+ α 3 (17)
x y k y k x y y y x y y y k
]
U 3 , 3) .

where U 1 , 2, U 1 ,3 , U 2 , 1 and U 3 , 1 are calculated using boundary conditions.

Lower left-hand corner

To calculate U n x , 1, set ( j, l )=( n x −1 , 2 ) in Equation (15).

nx 3
δ x δ y U n −1 , 2=
x
'
∑ ∑ α xj αly U j , l =0.
'
' ' ' '

j =nx −2 l =1

8
 Upon solving,
−1
U n , 1=
x x y
αn α 1
x
[
α n −2 ( α 1 U n −2 , 1+ α 2 U n −2 ,2 + α 3 U n −2 ,3 )+ α n −1 ( α 1 U n −1 , 1+ α 2 U n −1 ,2 +α 3 U n(18)
y
x
y
x
y
x
x y y
x
y
−1 ,3 ) + α n ( α 2 U n ,2+ α
x
x
y
x x x x x

where U n x , 2, U n x , 3, U n x −1 , 1 and U n x −2 , 1 are calculated using boundary conditions.

Upper right-hand corner

To calculate U n x , 1, set ( j, l )=( 2 , n y −1 ) in Equation (15).

3 ny
δ x δ y U 2 , n −1=∑ ∑ x y
y
α j α l U j ,l =0.
' ' ' '
' '
j =1 l =n y −2

Upon solving,
−1
U1, n =y x y
α1 αn
x y
[
α 1 ( α n −2 U 1 ,n −2+ α n −1 U 1 ,n −1 ) +α 2 ( α n −2 U 2 ,n −2+ α n −1 U 2 ,n −1+ α n U 2 ,n ) +α(19)
y
y y
x y
y y
y y
y 3 ( α n −2 U 3 , n −2 +α n −1
x y y
y y y y y y y y

where U 1 , ny −1, U 1 , ny −2, U 2 , ny and U 3 , ny are calculated using boundary conditions.

Lower right-hand corner

To calculate U n x , n y, set ( j, l )=( n x −1 , n y −1 ) in Equation (15).

nx ny
δ x δ y U n −1 , n −1=
x y
'
∑ ∑ '
x y
α j α l U j ,l =0.
' ' ' '

j =nx −2 l =n y −2

Upon solving,
−1
Un ,n =
x y x y
αn α n
x y
[
α n −2 ( α n −2 U n −2 ,n −2 +α n −1 U n −2 ,n −1+ α n U n −2 ,n )+ α n −1 ( α n −2 U n −1 , n −2 +α
x
y
y x y
y
y
x y
x y
y
(20)
k
y
y
n −1 U n −1 ,n −1 +α n U x y x y x y y x y y

x y

where U n x , n y −1 , U n x 1 , ny −2, U n x −1 , ny and U n x −2 , ny are calculated using boundary conditions.

2.2 Null-gamma conditions

Since the function U has two space varaibles ( x and y ), the Null-gamma conditions could be applied to either
one or both of the boundaries of the two space grids.

2.2.1 On the boundaries of x -grid

The Null-gamma condition states that on boundary, that is, for t ∈ ( 0 , T ) and, x=x or x=x ,
2 2
∂ x U =∂ x U when x is the log-spot or ∂ x U =0 when x is the spot.
Assume that x is the spot. For l=2 , … , n y −1 , on the lower boundary x=x , the Null-gamma condition is
given by
2
δ x U 2 ,l∧¿ 0
β x2 β x3
U 1 ,l ∧¿− x U 2 ,l− x U 3 , l .
β1 β1
For convenience, write
x x
U 1 ,l =λ2 U 2, l+ λ3 U 3 , l . (21)

The key observation is that U 1 ,l is a linear combination of U 2 , l and U 3 , l.

Again, for l=2 , … , n y −1 , on the upper boundary x=x , the Null-gamma condition is given by

9
 2
δ x U n −1 ,l ∧¿ 0
x
x x
β n −1 β n −2
U n , l∧¿− x

x
U n −1 ,l− x

x
U n −2 ,l .
x
β nx
x
βn x
x

For convenience, write

x x
U n , l=λ n −1 U n −1 ,l + λn −2 U n −2 , l .
x x x x x
(22)

The key observation is that U n x , l is a linear combination of U n x −1 , l and U n x −2 , l.

Explicit scheme
After discretization in time domain, the explicit scheme for x in Equation (16) is written as

[ ][ ]
x ,k+1 ' k+1
k+1c j ,l U j−1 ,l
k
U −U j,l x ,k +1 j,l
=− a j ,l U k+1
j ,l
,
Δ tk
b xj ,l,k +1 U kj+1
+1
,l

where j=2 , … ,n x −1 and l=2 , … , n y −1 , which could be rewritten as

j ,l + Δ t k ( c j ,l j +1 ,l ) ,
+1
U kj ,l=U k+1 x, k+1
U kj−1 x ,k+1
, l + a j ,l U kj +1 x, k+1
,l + b j ,l U k+ 1
(23)
k +1 k +1
where U 1 ,l is calculated using Equation (21) and U n x , l is calculated using Equation (22).

Implicit scheme
Consider the lower boundary of x -grid. Set j=2 and apply Equation (21). The space-discretized 2D PDE (16)
becomes

([ ] [ ]) [ ]
xy , k ' x k x k
k+1 k lc 2 ,l 0 λ2 U 2, l−1+ λ3 U 3 , l−1
U −U
2 ,l 2 ,l
=−¿ mc2xy, l, k + c 2y ,l, k k
U 2 ,l −1
Δtk xy , k
uc 2, l 0 k
U 3 ,l −1

([ ] [ ] [ ]) [ ]
xy , k x ,k ' x k x k
la 2 ,l 0 c 2 ,l λ 2 U 2 ,l + λ3 U 3 ,l
−¿ ma2xy, l, k + a2y,,kl + ax2 ,,lk k
U 2 ,l
xy , k x,k k
ua 2, l 0 b2 ,l U 3 ,l

([ ] [ ]) [ ]
'
lb xy
2 ,l
,k
0 λ2x U k2 ,l +1+ λ3x U 3k , l+1
−¿ mb xy
2, l
,k
+ b 2y ,l,k U kj ,l+ 1 ,
ub xy ,k
2 ,l
0 U k
j+ 1, l+1

where l=2 , … , n y −1 , which could be rewritten as

10
 ([ ] [ ]) [ ]
'
U k+1 k
2 ,l −U 2 ,l
xy , k x xy ,k
mc 2, l + λ2 lc 2 ,l y ,k
c 2 ,l U 2 ,l−1
k
=−¿ xy , k x xy ,k
+ k
Δtk uc 2 ,l + λ3 lc 2 , l 0 U 3 ,l−1

([ ][ ][ ]) [ ]
xy ,k x xy ,k x,k x x, k ' k
ma2 ,l + λ 2 la2 , l a y, k a + λ c U 2, l
−¿ xy ,k x xy , k
+ 2 , l + 2x ,l, k 2x 2x,,lk k (24)
ua2 , l + λ3 la2 , l 0 b 2 ,l + λ3 c 2 ,l U3,l

([ ] [ ]) [ ]
xy ,k x xy , k ' k
mb2 ,l + λ 2 lb2 , l b y ,k U j ,l+ 1
−¿ xy ,k x xy , k
+ 2, l k
,
ub2 , l + λ 3 lb2 , l 0 U j+1 , l+1

Consider the upper boundary of x -grid. Set j=n x −1 and apply Equation (22). The space-discretized 2D PDE
(16) becomes

([ ] [ ]) [ ]
xy , k ' k
k+1 k lc n −1 ,l 0 U n −2 ,l −1
U nx −1 ,l −U nx −1 ,l
x x

=¿ mc xy ,k
n −1, l
+ cny,−1
k
,l
k
U
nx −1 ,l −1
Δtk x x

xy , k
uc n −1 ,l 0 λ x
nx −1 U k
n x −1 , l−1
x
nx −2 +λ U kn −2 ,l−1
x x

([ ] [ ] [ ]) [ ]
xy , k x ,k ' k
lα n −1 ,l
x
0 c n −1 ,l x
U n −2 , l x
y ,k
+¿ mα nxy−1
,k
,l
+ a n −1 ,l + axn ,−1
x
k
,l x x
U kn −1 , l x
xy ,k
uα n −1 , l 0 x,k
b n −1 ,l
x k
λn −1 U n −1 ,l + λ n −2 U n −2 ,l
x k
x x x x x x

([ ] [ ]) [ ]
xy ,k ' k
lb n −1 ,l x
0 U n −2 ,l +1
x

+ ¿ mbxy ,k
n −1 ,l
+ b ny ,k−1 ,l x
k
U n −1 ,l +1 ,
x x

ub xy ,k
n −1 ,l
0 λnx −1 U kn −1 ,l +1+ λnx −2 U kn −2 ,l +1
x x x x x

where l=2 , … , n y −1 , which could be rewritten as

([ ] [ ]) [ ]
'
U k+1 k
n −1 ,l −U n −1 ,l
xy , k
lc n −1 ,l + λ n −2 uc n −1 ,l
x xy , k
0
k
U n −2 , l−1
x x
=¿ x x x
+ y,k
x

Δ tk mc
xy ,k
n x −1 , l +λ
x
nx −1
xy , k
uc
n x −1, l
c n −1 , l x
U
k
n x −1 , l−1

([ ][ ][ ]) [ ]
'
xy , k x xy , k 0 x ,k x x ,k k
lα nx −1 ,l +λ nx −2 uα nx −1 ,l y ,k c nx −1 ,l +λ nx −2 b nx −1 ,l U n −2 ,l
+¿ xy ,k x xy ,k
+ a nx −1 ,l + x,k x x,k
x
(25)
mα +λ uα a +λ b U kn −1 ,l
n x −1 , l nx −1 nx −1 , l
0 nx −1 ,l nx −1 nx −1 ,l x

([ ] [ ]) [ ]
'
lb xy ,k x xy , k
n −1 ,l + λ n −2 ub n −1, l 0 U kn −2 , l+1
+¿ x
xy , k
x
x
x
xy ,k
+ y, k k
x
,
mb n x −1 , l +λ n x −1 ub nx −1 ,l
b n x −1 , l U n x −1 , l+1

Using Equations (24) and (25), the implicit scheme on x -grid is formed by rewriting Equation (33) as

[ ][ ]
U k2 , l U k+1
2, l

U k3 , l U k+1

( [ ])
~ax , k ~
b2x ,,kl ¿ ¿ c3x ,,kl a x3 ,,lk b3x ,,kl ¿ ~x , k 3 ,l
2 ,l
I n −2− Δ t k an −1 ,l ¿ ⋮ = ⋮ , (26)
x
⋱ ⋱ ⋱ ¿ c xn ,−2
k
axn ,−2
k
¿ ¿ x
k k+1
,l x ,l x
U n x −2 , l U nx −2 ,l
k k+1
U n x −1 , l U nx −1 ,l

11
 ~x , k
a 2 ,l , b 2 ,l , ~
where l=2 , … , n y −1 . Note the modified coefficients ~ c nx−1 , l and ~
x,k x ,k x,k
a nx −1 ,l . Similar to the case of 1D
PDE in Section 1.1, the above system of equations could be easily solved by standard tri-diagonal matrix solver.

2.2.2 On the boundaries of y -grid

The Null-gamma condition states that on boundary, that is, for t ∈ ( 0 , T ) and, y= y or y= y ,
2
∂ y U =0.
For j=2 , … ,n x −1, on the lower boundary y= y , the Null-gamma condition is given by
2
δ y U j ,2∧¿ 0
βy βy
U j ,1∧¿− 2y U j , 2− 3y U j ,3 .
β1 β1
For convenience, write
y y
U j ,1=λ 2 U j , 2+ λ 3 U j , 3 . (27)

The key observation is that U j ,1 is a linear combination of U j ,2 and U j ,3.

Again, for j=2 , … ,n y −1, on the upper boundary y= y , the Null-gamma condition is given by
2
δ y U j , n −1∧¿ 0 y
y y
β n −1 β n −2
U j ,n ∧¿− y

y
U j , n −1− y

y
U j , n −2 .
y
β ny
y
βn y
y

For convenience, write

U j ,n =λ ny −1 U j , n −1 + λ ny −2 U j , n −2 .
y y y y y
(28)

The key observation is that U j ,n y is a linear combination of U j ,n y −1 and U j ,n y −2.

Explicit scheme
After discretization in time domain, the explicit scheme for y in Equation (16) is written as

[ ][ ] [ ][ ] [ ][ ]
+1
0
'
U k+1 0
'
U k+1 0
'
U kj−1,
U k+1
j,l −U k
j,l
j−1 ,l −1 j−1 ,l l+1
=− c yj ,,lk+1 U kj +1 − y ,k +1
a j ,l k+1
U j ,l − y, k+1
b j ,l k+1
U j ,l +1 ,
Δ tk ,l−1
0 U k+1
j+1 , l−1
0 +1
U kj+1 ,l
0 U k+1
j +1 ,l+ 1

where j=2 , … ,n x −1 and l=2 , … , n y −1 , which could be rewritten as

j ,l + Δ t k ( c j ,l U j ,l+ 1) ,
y ,k+ 1 k +1
U kj ,l=U k+1 y , k+1
U k+1 y , k+1
j , l−1+ a j , l U k+1
j ,l +b j ,l
(29)
k +1 k +1
where U j ,1 is calculated using Equation (27) and U j ,n y is calculated using Equation (28).

Implicit scheme
Consider the lower boundary of y -grid. Set l=2 and apply Equation (27). The space-discretized 2D PDE (16)
becomes

12
 ([ ] [ ]) [ ]
xy ,k ' y k y k
k+1 k lc j ,2 0 λ 2 U j−1 ,2 + λ 3 U j−1 ,3
U j,2 −U j,2
=−¿ mcxyj , 2,k + c yj ,2,k y k
λ 2 U j ,2 + λ 2 U j ,3
y k
Δtk xy ,k
uc j , 2 0 y k
λ2 U j+1 ,2 + λ 2 U j+1 , 3
y k

([ ] [ ] [ ]) [ ]
xy ,k x, k ' k
la j ,2 0 c j ,2 U j−1 ,2
−¿ maxyj , 2, k + a j , 2 + a xj ,2,k
y,k
U j ,2
k

xy ,k x ,k k
ua j , 2 0 b j ,2 U j+1 , 2

([ ] [ ]) [ ]
'
lb xyj ,2, k 0 U kj−1 ,3
−¿ mb xyj ,2,k + b y,j ,2k U kj , 3 ,
ub xyj ,2, k 0 U kj+1 ,3

where j=2 , … ,n x −1, which could be rewritten as

([ ][ ] [ ]) [ ]
xy , k y xy , k x, k ' k
k+1 k la j ,2 + λ2 lc j , 2 0 c j ,2 U j−1 ,2
U j,2−U j,2
=−¿ maxyj ,2, k + λ2y mc xyj ,2,k + a j ,2 + λ2 c j , 2 + a xj ,2,k
y,k y y ,k
U j ,2
k
Δtk xy , k
ua j ,2 + λ2 uc j ,2
y xy ,k
0 x ,k
b j ,2
k
U j+1 , 2
(30)

([ ][ ]) [ ]
xy , k y xy ,k ' k
lb j,2 + λ lc
3 j ,2 0 U j−1, 3
xy , k y xy , k y ,k y y,k k
−¿ mb j,2 + λ mc
3 j ,2
+ b j ,2 + λ3 c j ,2 U j ,3
.
xy , k y xy , k k
ub j,2 + λ uc
3 j ,2
0 U j +1 ,3

Consider the upper boundary of y -grid. Set l=n y −1 and apply Equation (28). The space-discretized 2D PDE
(16) becomes

([ ] [ ]) [ ]
xy , k ' k
k +1 k lc j , n −1 0 U j−1 , n −2
U j ,n y −1 −U j , ny −1
y y

=−¿ mc xyj ,n, k−1 + c yj ,n, k −1 k

U j ,n −2
Δtk y y y

uc xyj , n, k−1 0 U k
j +1 ,n y −2
y

([ ] [ ] [ ]) [ ]
xy ,k x ,k ' k
la j ,n −1y
0 c j ,n −1 y
U j−1 ,n −1 y
y,k
−¿ ma xy ,k
j , n −1
+ a j ,n −1 + a xj ,n,k −1
y y y
U kj , n −1 y
xy ,k
ua j ,n −1 0 x ,k
b j ,n −1 U
k
j +1 ,n y −1
y y

([ ] [ ]) [ ]
xy ,k ' y k y k
lb j ,n −1
y
0 λn −1 U j−1 ,n −1+ λn −2 U j−1 ,n −2
y y y y

−¿ mb xy ,k
j , n −1
+ b yj ,n, k −1 y
y
λn −1 U j ,n −1+ λn −2 U j ,n −2
k y k
,
y y y y y

ub xyj ,n,k−1 0 λ ny −1 U kj+1 ,n −1+ λny −2 U kj+1 , n −2

y y y y y

where j=2 , … ,n x −1, which could be rewritten as

13
 ([ ][ ]) [ ]
xy ,k y xy , k ' k
k+1 k lc j ,n −1+ λn −2 lb j , n −1 0 U j −1 , n −2
U j , ny −1 −U j , ny −1
y y y y

=−¿ mc xy ,k
j ,n y −1 +λ
y
n y −2 mb
xy ,k
j ,n y −1
+ c j , n −1 + λ ny −2 b yj ,n, k −1
y ,k
U
k
j ,n y −2
Δ tk y y y

uc xy ,k
j ,n y −1 +λ y
n y −2 ub xy , k
j ,n y −1
0 U kj +1 ,n −2 y

(31)

([ ][ ] [ ]) [ ]
xy ,k y xy , k x ,k ' k
la j , ny −1 +λ ny −1 lb j ,n y −1 0 c j ,n y −1 U j−1 ,n y −1
xy ,k y xy , k y,k y y,k x ,k k
−¿ ma j , ny −1 +λ ny −1 mb j , ny −1
+ a j , n y −1 + λ n −1 b j ,n −1 + a
y y j ,n y −1 U j , ny −1
.
ua
xy ,k
j , ny −1 +λ
y
ny −1 ub
xy ,k
j , n y −1
0 b
x ,k
j ,n y −1 U
k
j+1 ,n y −1

Using Equations (30) and (31), the implicit scheme on y -grid is given by

[ ][ ]
k k+1
U j,2 U j,2
k k+1

[ ])
U j,3 U j,3

(
~a y , k ~ y ,k
b j ,2 ¿ ¿ c j,3
y,k y,k
a j ,3
y,k
bj , 3 ¿ ~ y, k
j ,2
I n −2−Δ t k a j ,n −1¿ ⋮ = ⋮ , (32)
y
⋱ ⋱ ⋱ ¿ c yj ,,kn −2 a yj ,,kn −2 ¿ ¿ y
k k+1
y y
U j , ny −2 U j , ny −2
k k+1
U j , ny −1 U j , ny −1

y ,k ~ y ,k
a j,2, b j,2, ~
where j=2 , … ,n x −1. Note the modified coefficients ~ c j ,n y−1 and ~
y,k y ,k
a j , ny−1. Similar to the case of
1D PDE in Section 1.1, the above system of equations could be easily solved by standard tri-diagonal matrix
solver.

2.3 Dirichlet conditions

In pricing most of the derivatives, Dirichlet conditions are used only on the x -grid because event-based payoff
(for example, knock-out event) is based on the observed value of the underlying spot price rather than on the
unobserved value of the stochastic variance.
The Dirichlet conditions are given by
k k
U 1, l∧¿ gl ,
U kn , l∧¿ hlk ,
x

are known for all l ’s, and l=2 , … , n y −1 .

k k
where g l ’s and h l ’s

Explicit scheme
Since the solution values on the boundaries are known at the current time point t k+1, the explicit scheme
k +1 k +1 k +1 k+1
proceeds as in Equation (23), where U 1 ,l =gl and U n x , l =hl .

Implicit scheme
The implicit scheme is obtained by slightly modifying Equation (26) as

[ ][ ] [ ]
U k2 , l U k+1
2, l c x2 ,,lk gkl
U k3 , l U k+1

( [ ])
ax2 ,,lk b2x ,,kl ¿ ¿ c3x ,,kl a x3 ,,lk b3x ,,kl ¿ x , k 3 ,l 0
I n −2− Δ t k an −1 ,l ¿ ⋮ = ⋮ + Δ tk ⋮
(33) ,
x
⋱ ⋱ ⋱ ¿ c xn ,−2
k
axn ,−2
k
¿ ¿ x
k k+1
0
,l x ,l x
U n x −2 , l U nx −2 ,l
x ,k
U k
U k+1 b nx −1 ,l hkl
n x −1 , l nx −1 ,l

where l=2 , … , n y −1 . Similar to the case of 1D PDE in Section 1.1, the above system of equations could be
easily solved by standard tri-diagonal matrix solver.

14
2.4 Calculation of cross term
Assume the solution values on the boundaries are known at the current time point t k+1 . As shown later, in the
predictor-corrector scheme, the cross term in Equation (16) must be calculated, whose calculation is given by

[ ][ ] [ ][ ] [ ][ ]
xy ,k +1 ' k +1 xy , k+1 ' k+1 xy , k+1 ' k+1
lc j ,l U j−1, l−1 la j ,l U j−1 ,l lb j ,l U j−1 ,l +1
xy , k+1 k+1 xy , k+1 k+1 xy ,k +1
− mc j , l U j ,l −1 − ma j ,l U j , l − mb j ,l U k+ 1
j , l+1
, (34)
xy ,k+1 k+ 1 xy , k+1 k+1 xy ,k +1 k +1
uc j,l U j +1 ,l−1 ua j ,l U j+1 , l ub j ,l U j+1 ,l +1

k +1 k +1 k +1
where j=2 , … ,n x −1 and l=2 , … , n y −1 , U 1 ,l , U n , l , U kj +1
,1 and U j ,n are calculated using boundary
x y
k +1 k +1 k +1 k +1
conditions in Sections 2.2 and 2.3, and U 1 ,1 , U n , 1, U 1 , n and U n , n are calculated using Section 2.1.
x y x y

2.5 Matrix form of the numerical solution

Now, re-write the space-discretized 2D PDE (16) in matrix form.
∂t U =−LU −F ,
where

U =[ U 2 ,2 , … ,U n −1 ,2 ,U 2 ,3 ,… ,U n −1, 3 ,… , U 2 , n −1 , … ,U n −1 ,n −1 ] ,
'
x x y x y

which is formed by simply stacking the following (n x −2)-column vectors

U l =[ U 2 ,l , … ,U n −1 ,l ] ,
'
x

for l=2 , … , n y −1 , and

L=
[ A2 B2 ¿ ¿
C 3 A3 B 3 ¿
¿ ⋱ ¿ ⋱ ¿ ⋱ ¿ ¿ ¿ Cl ¿ A l¿ Bl ¿ ¿ ⋱ ¿ ⋱ ¿ ⋱ ¿ C n −1¿ A n −1 ¿ , y y
]
where Al , Bl and C l are tri-diagonal matrices, and F is a matrix of source terms (only applicable to Dirichlet
conditions).
It would help understand the “geometry” of the matrix if we write down explicitly the dependence of time-
derivative of U j ,l on the values of U j−1 ,l−1, U j ,l−1, U j+1 ,l −1 ,U j−1 ,l, U j ,l , U j+1 ,l ,U j−1 ,l+1 , U j ,l+ 1 and
U j+1 ,l +1, where j=3 , … , nx −2 and l=3 , … , n y −2.
Assume that Null-gamma conditions are applied on all boundaries. Consider the element U j ,l . Using Equation
(16), the matrix multiplication could be explicitly written as

[ ]
⋮
∂
U =−¿
∂ t j,l
⋮
Note that the elements
xy y xy xy
lc j ,l , c j ,l +mc j , l , uc j ,l
are in the ( l−1 ) -th block of matrix L, the elements
x xy x y xy x xy
c j ,l +la j ,l , a j ,l +a j ,l +ma j ,l , b j , l+ ua j , l
are in the l -th block of matrix L, and the elements
xy y xy xy
lb j , l , b j ,l +mb j ,l , ub j , l
are in the ( l+1 )-th block of matrix L.

15
Therefore, L could be written as
x y xy
L=L + L + L ,
where L x is given by

[ ]
x
A2 0 ¿ ¿
L x= x
x x
¿ ⋱ ¿ ⋱ ¿ ⋱ ¿ ¿ ¿ 0 ¿ A l ¿ 0 ¿ ¿ ⋱ ¿ ⋱ ¿ ⋱ ¿ 0 ¿ An −1 ¿ ,
0 A 0 ¿
y
3

x
with tri-diagonal matrix Al given by

[ ]
x x
a2 ,l b2 ,l ¿ ¿
Alx = x x x
x x x x x
¿ ⋱ ¿ ⋱ ¿ ⋱ ¿ ¿ ¿ c j ,l ¿ a j , l ¿ b j ,l ¿ ¿ ⋱ ¿ ⋱ ¿ ⋱ ¿ c n −1 ,l ¿ an −1 ,l ¿ ,
c a b ¿
x x
3, l 3 ,l 3 ,l

y
L is given by

[ ]
y y
A2 B2 ¿ ¿
L y= y y y
y y y y y
¿ ⋱ ¿ ⋱ ¿ ⋱ ¿ ¿ ¿ C l ¿ A l ¿ Bl ¿ ¿ ⋱ ¿ ⋱ ¿ ⋱ ¿ C n −1 ¿ An −1 ¿ ,
C A B ¿
y y
3 3 3

y y y
with diagonal matrices Al , Bl , C l

[ ]
y
y a 2, l ¿ ¿ y
Al = ¿¿ ⋱ ¿¿ an −1 , l ¿ ,
⋱ ¿ ¿ x

[ ]
y
y b2 ,l ¿ ¿ y
Bl = ¿¿ ⋱ ¿¿ b n −1 ,l ¿ ,
⋱ ¿ ¿ x

[ ]
y
y c2 , l ¿ ¿ y
Cl = ¿¿ ⋱ ¿¿ c n −1 ,l ¿ ,
⋱ ¿ ¿ x

and L xy is given by

[ ]
xy xy
A2 B2 ¿ ¿
L xy = xy xy xy
xy xy xy
¿ ⋱ ¿ ⋱ ¿ ⋱ ¿ ¿ ¿ C l ¿ A l ¿ Bl ¿ ¿ ⋱ ¿ ⋱ ¿ ⋱ ¿ C n −1 ¿ A n −1¿
xy xy

C A B ¿
y y
3 3 3

xy xy xy
with tri-diagonal matrices Al , Bl and C l

Al ∧¿
xy
[ ⋱
la xy
j ,l
⋱
ma xy
j ,l
⋱
ua xy
j ,l
¿
¿
⋱¿ ⋱¿ ,
]
Bl ∧¿
xy
[ ⋱
lb
xy
j ,l
⋱
mb
xy
j ,l
⋱
ub
xy
j,l
¿
¿
⋱ ¿⋱¿ ,
]
C l ∧¿
xy
[ ⋱
lc
xy
j ,l
⋱
mc
xy
j,l
⋱
uc
xy
j ,l
¿
¿
⋱¿ ⋱¿ .
]
Finally, discretize the space-discretized PDE (16) in time domain with the θ -scheme as
k+1 k
U −U
=−( 1−θ ) ( L U + F )−θ ( L U + F ) ,
k+1 k+1 k+1 k k k
Δtk
where θ ∈ [ 0 , 1 ] . Re-arranging the terms products

( I −θ Δ t k Lk ) U k −θ Δ t k F k =( I + (1−θ ) Δ t k Lk +1 ) U k+ 1+ ( 1−θ ) Δ t k F k+1 ,

16
 where I is an identity matrix of order ( n x −2 ) ( n y −2 ). After the matrices Lk and Lk+ 1 are split, the above
equation could be written more generally as

( I −θ Δ t k ( L x ,k + L y , k )−θ Δ t k Lxy , k ) U k −θ Δ t k F k =( I +( 1−θ ) Δ t k ( L x ,k+1 + L y , k+1 ) + ( 1−θ ) Δ t k(35)

L
xy ,k +1
) U k+1 + ( 1−θ )
where θc is the weight for the cross term. Nevertheless, Murex sets θ=0.5 for non-implicit time points and
θ=1 for implicit time points.
2.6 Predictor-corrector scheme
The left-hand side of Equation (35) can be rewritten as

I −θ Δ t k ( Lx , k+1 + L y ,k +1 )=( I −θ Δ t k L x, k+1 )( I −θ Δ t k L y , k+1 ) −θ2 ( Δ t k ) L x ,k+1 L y ,k +1 .

2

As the last term is of order 2 in time discretization, it corresponds to the same error level as the PDE
discretization and could be ignored. Apply the same approximation on the right-hand side of Equation (35).
Therefore, Equation (35) could be approximated by

( I −θ Δ t k L x, k )( I −θ Δ t k L y ,k ) U k =( I + ( 1−θ ) Δ tk Lx , k+1 )( I + ( 1−θ ) Δ tk L y ,k +1 ) U k+1 + ( 1−θ ) Δ(36)

t k L xy ,k +1 U k+1 +θ Δ t k
Since U k +1 on the right-hand side of the above equation is unknown, Murex applies the predictor-corrector
method.
Predictor step
Replace L xy ,k U k by L xy ,k+ 1 U k +1 on the right-hand side of Equation (36). Define
~
U k +1 ≝ [ I + ( 1−θ ) Δ t k Lx ,k +1 ][ I + ( 1−θ ) Δ t k L y , k+1 ] U k +1+ (1−θ ) Δ t k F k+ 1 . (37)

The predictor is calculated as

^ k =( I −θ Δ t k L y ,k )−1 ( I −θ Δ t k L x, k )−1 ( ~
U [k+1
U + Δ tk L
xy ,k+1
U ) +θ Δ t k F .
k +1 k
] (38)

^ k is an approximated value of U k .
Now, U
Corrector step
^ k . By incorporating (37), the value of U k is
Next, replace U k on the right-hand side of Equation (36) by U
calculated by

) ( I −θ Δ t k L x, k ) ((~
y ,k −1
F ).
−1
U =( I −θ Δ t k L ^ k ) +θ Δ t k(39)
k k +1 xy , k+1 k+1 xy , k k
U + ( 1−θ ) Δ t k L U +θ Δ t k L U

2.7 Computing the numerical solution

To implement the predictor-corrector scheme for 2D PDE, the following numerical procedure is implemented
in our code.
Predictor step

Step 1: Perform explicit scheme in y on U k +1 according to (29) (by replacing Δ t k by ( 1−θ ) Δ t k)

Step 2: Perform explicit scheme in x on the results of Step 1 according to (23)1 (by replacing Δ t k by
( 1−θ ) Δ t k) to obtain ~k +1
U in (37).
xy , k+1 k+ 1
Step 3: Perform the calculation of cross term on U k +1 at t k+1 according to (34) to obtain Δ t k L U .
~k +1 xy , k+1 k+1
Step 4: Calculate U −Δ t k L U .

1
In the discussion that follows, the actual form of F is not important because it is already taken care in
calculation discussed in Sections 2.2 and 2.2.

17
Step 5: Perform implicit scheme in x on the results of Step 4 according to (26) for Null-gamma condition or
(33) for Dirichlet condition (by replacing Δ t k by θ Δ t k ).

Step 6: Perform implicit scheme in y on the results of Step 5 according to (32) (by replacing Δ t k by θ Δ t k ) to
^ k in (38).
obtain the predictor U
Corrector step
xy ,k+ 1 k +1
Step 7: Multiply the results of Step 3 by ( 1−θ ) to obtain ( 1−θ ) Δ t k L U .

^ k at t k according to (34) (by replacing Δ t k by θ Δ t k ) to

Step 8: Perform the calculation of cross term on U
xy , k ^ k
obtain θ Δ t k L U .
~k +1
Step 9: Calculate the results U + ( 1−θ ) Δ t k L
xy , k+1 k+1 xy , k ^ k , where the first term is from
U +θ Δ t k L U
Step 6, the second term is from Step 7 and the third term is from Step 8.
Step 10: Perform implicit scheme in x on the results of Step 9 according to (26) for Null-gamma condition or
(33) for Dirichlet condition (by replacing Δ t k by θ Δ t k ).

Step 11: Perform implicit scheme in y on the results of Step 10 according to (32) (by replacing Δ t k by θ Δ t k )
to obtain the solution U k in (39).

18