DOI: 10.1214/EJP.

v11-311

a
rn l
u o
o f
J
P
c r
i ob
on abili
E lectr ty

Vol. 11 (2006), Paper no. 9, pages 249275.

Journal URL
http://www.math.washington.edu/∼ejpecp/

On random walk simulation of one-dimensional

diusion processes with discontinuous coecients
∗
Pierre Étoré

Projet OMEGA, IECN, BP 239, F-54506 Vand÷uvre-lès-Nancy CEDEX

[email protected]

http://www.iecn.u-nancy.fr/~etore/

Abstract
In this paper, we provide a scheme for simulating one-dimensional processes generated by divergence
or non-divergence form operators with discontinuous coecients. We use a space bijection to
transform such a process in another one that behaves locally like a Skew Brownian motion. Indeed
the behavior of the Skew Brownian motion can easily be approached by an asymmetric random walk.

Key words: Monte Carlo methods, random walk, Skew Brownian motion, one-dimensional process,
divergence form operator, local time.

AMS 2000 Subject Classication: Primary 660J60, 65C.

Submitted to EJP on May 11, 2005. Final version accepted on February 21, 2006.

∗ Support acknowledgement: This work has been supported by the GdR MOMAS.

249
1 Introduction
In this paper we provide a random walk based scheme for simulating one-dimensional processes generated
by operators of type
ρ  
L= ∇ a∇ + b∇. (1.1)
2
These operators appear in the modelisation of a wide variety of diusion phenomena, for instance in uid
mechanics, in ecology, in nance (see [DDG05]). If a, ρ, and b are measurable and bounded and if a and
ρ are uniformly elliptic it can be shown that L is the innitesimal generator of a Markov process X . Note
that these operators contain the case of operators of type L = a2 ∆ + b∇ whose interpretation in terms of
Stochastic Dierential Equation is well known. In the general case there is still such an interpretation.
Indeed if we assume for simplicity that b = 0 it can be shown (see Section 4) that X solves
Z tp t
ρ(Xs )a0 (Xs ) X a(x+) − a(x−)
Z
Xt = X0 + a(Xs )ρ(Xs )dWs + ds + Lx (X), (1.2)
0 0 2 a(x+) + a(x−) t
x∈I

where I is the set of the points of discontinuity of a and Lxt (X) is the symmetric local time of X in x.
For the coecients a, ρ, and b may be discontinuous, providing a scheme to simulate trajectories of X
is challenging: we cannot use the panel of Monte-Carlo methods available for smooth coecients (see
[KP92]). However, some authors have recently provided schemes to simulate X in the case of coecients
having some discontinuities.
In [Mar04] (see also [MT06]) M. Martinez treated the case of a coecient a having one point of
discontinuity. He applied an Euler scheme after a space transformation that allows to get rid of the
local time in (1.2). To estimate the speed of convergence of his method he needs a to be C 6 outside
the point of discontinuity. The initial condition has to be C 4 almost everywhere and to satisfy other
restrictive conditions.
In [LM06] (see also [Mar04]) A. Lejay and M. Martinez proposed a dierent scheme. After a piecewise
constant approximation of the coecients and a dierent space transformation, they propose to use an
exact simulation method of the Skew Brownian Motion (SBM). This one is based on the simulation of
the exit times of a Brownian motion. In general the whole algorithm is slow and costly but allows to
treat the case of coecients a, ρ and b being right continuous with left limits, and of class C 1 except on
countable set of points, without cluster point. Besides the initial condition can be taken in H1 , and the
algorithm is well adapted to the case of coecients being at on large intervals outside their points of
discontinuity.
Here, under the same hypotheses on a, ρ and b, but with the initial condition in W1,∞ ∩ H1 ∩ C0 we
propose a new scheme based mostly on random walks.
Roughly the idea is the following: assume for simplicity that b = 0. First, like in [LM06], we replace a
and ρ by piecewise constant an and ρn in order to obtain X n that is a good weak approximation of X
and that solves
Z tp X a(xn +) − a(xn −) xn
i i
Xtn = X0n + an (Xsn )ρn (Xsn )dWs + L i (X n ),
0 n n
a(xni +) + a(xni −) t
xi ∈I

where I n is the set of the points of discontinuity of an . Second by a proper change of scale we transform
X n into Y n that solves

k/n
X
Ytn = Y0n + Wt + βkn Lt (Y n ),
xn
k ∈I
n

250
where the βkn 's are explicitely known. Thus Y n behaves around each k/n like a SBM of parameter βkn
(see Subsection 5.2 for a brief presentation of the SBM). That is, heuristically, Y n when in k/n moves up
with probability (βkn + 1)/2 and down with probability (βkn − 1)/2, and behaves like a standard Brownian
motion elsewhere. Thus a random walk on the grid {k/n : k ∈ Z} can reect the behaviour of Y n as was
shown in [Leg85]. We use this to nally construct an approximation Yb n of Y n .
We obtain a very easy to implement algorithm that only requires simulations of Bernoulli random
variables. We estimate the speed of weak convergence of our algorithm by mixing an estimate of a weak
error and an estimate of a strong error. Indeed computing the strong error of the algorithm presents
diculties we were not able to overcome. On the other hand computing directly the weak error without
using a strong error estimate would lead to more complicated computations without any improvement
of the speed of convergence: basically our approach relies on the Donsker theorem and we cannot get
better than an error in O(n−1/2 ). Moreover to make such computations we should require additionnal
smoothness on the data (see [Mar04]).
We nally make numerical experiments: the proposed scheme appears to be satisfying compared to the
ones proposed in [Mar04] or [LM06].

Hypothesis. We make some assumptions from now till the end of the paper, for the sake of simplicity
but without loss of generality.
(A1) b = 0.
Indeed, as explained in [LM06] Section 2, if we can treat the case
ρ  
L= ∇ a∇ , (1.3)
2
we can treat the case (1.1) for any measurable bounded b by dening the coecients ρ and a in (1.3) in
the following manner:

If ab := a expΨ and ρb := ρ exp−Ψ ,

Z x
b(x)
with Ψ(x) = h(y)dy and h(x) = 2 ,
0 ρ(x)a(x)
ρb   ρ  
then ∇ ab ∇ = ∇ a∇ + b∇.
2 2

(A2) Let be G = (l, r) an open bounded interval of R. We will assume the process X starts from x ∈ G
and is killed when reaching {l, r}. >From a PDEs point of view this means the parabolic PDEs involving
L we will study are submitted to uniform Dirichlet boundary conditions. We could treat Neumann
boundary conditions (thanks to the results of [BC05] for instance) and, by localization arguments, the
case of an unbounded domain G (see [LM06]). But this assumption will make the material of the paper
simpler and clearer.

Outline of the paper. In Section 2 we dene precisely Divergence Form Operators (DFO) and recall
some of their properties. In Section 3 we speak of Stochastic Dierential Equations involving Local Time
(SDELT): we state a general change of scale formula and recall some convergence results established
by J.F. Le Gall in [Leg85]. In Section 4 we link DFO and SDELT: a process generated by a DFO of
coecients a and ρ having countable discontinuities without cluster points is solution of a SDELT with
coecients determined by a and ρ. In Section 5 we present our scheme. In Section 6 we estimate the
speed of convergence of this scheme. Section 7 is devoted to numerical experiments.

251
Some notations. For 1 ≤ p < ∞ we denote by Lp (G) the set of measurable functions f on G such that
Z

1/p
kf kp := |f (x)|p dx < ∞.
G

Let be 0 < T < ∞ xed. For 1 ≤ p, q < ∞ we denote by Lq (0, T ; Lp (G)) the set of measurable functions
f on (0, T ) × G such that
Z T
1/q
q
|||f |||p,q := kf kp dt < ∞.
0

For u ∈ Lp (G) we denote by du dx the rst derivative of u in the distributional sense. It is standard to
denote by W1,p (G) the space of functions u ∈ Lp (G) such that du dx ∈ L (G), and by W0 (G) the closure of
p 1,p
p
Cc∞ (G) in W1,p (G) equipped with the norm (kukp + du . We denote by H1 (G) the space W1,2 (G),
p 1/p
dx p )
and by H10 (G) the space W0 (G).
1,2

For u ∈ L2 (0, T ; L2R(G)) we denote by ∂t u the distribution such that for all ϕ ∈ Cc∞ ((0, T ) × G), we
have, h∂t u, ϕi = − 0 G u∂t ϕ. We still denote by du
T R
dx the rst derivative of u with respect to x in the
distributional sense.
We will classicaly denote by k.k∞ and |||.|||∞,∞ the supremum norms.
For the use of probability we will denote by C0 (G) the set of continuous bounded functions on G. The
symbol ∼= will denote equality in law.

2 On divergence form operators

For 0 < λ < Λ < ∞ let us denote by Ell(λ, Λ) the set of functions f on G that are measurable and such
that

∀x ∈ G, λ ≤ f (x) ≤ Λ.

For ρ ∈ Ell(λ, Λ) let us dene the measure mρ (dx) := ρ−1 (x)dx.

For any measure m with a bounded density with respect to the Lebesgue measure we then denote by
L2 (G, m) the Hilbert space of functions in L2 (G) equipped with the scalar product
Z
(f, g) 7−→ f (x)g(x)m(dx).
G

This is done in order that the operator we dene below is symmetric on L2 (G, mρ ).

Denition 2.1 Let a and ρ be in Ell(λ, Λ) for some 0 < λ < Λ < ∞. We call Divergence form operator
of coecients a and ρ, and we note L(a, ρ), the operator (L, D(L)) on L2 (G, mρ ) dened by
ρ d  d 
L= a ,
2 dx dx
and

D(L) = { u ∈ H10 (G) , Lu ∈ L2 (G)}.

Actually if a, ρ ∈ Ell(λ, Λ) the operator L(a, ρ) has sucient properties to generate a continuous Markov
process. We sum up these properties in the next theorem.

252
Theorem 2.1 Let a and ρ be in Ell(λ, Λ) for some 0 < λ < Λ < ∞. Then we have:
i) The operator L(a, ρ) on L2 (G, mρ ) is closed and self-adjoint, with dense domain.
ii) This operator is the innitesimal generator of a strongly continuous semigroup of contraction (St )t≥0
on L2 (G, mρ ).
iii) Moreover (St )t≥0 is a Feller semigroup. Thus L(a, ρ) is the innitesimal generator of a Markov
process (Xt , t ≥ 0).
iv) The process (Xt , t ≥ 0) has continuous trajectories.
Proof. We give the great lines of the proof and refer the reader to [Lej00] and [Str82] for details.
We set (L, D(L)) = L(a, ρ). First it is possible to build a symmetric bilinear form E on L2 (G, mρ ) dened
by
Z
a du dv
E(u, v) = dx, ∀(u, v) ∈ D(E) × D(E), and D(E) = H10 (G),
G 2 dx dx
that veries,

E(u, v) = − hLu, viL2 (G,mρ ) , ∀(u, v) ∈ D(L) × D(E).

Thus the resolvent of (L, D(L)) can be built and we get i). An application of the Hille-Yosida theorem
then leads to ii).
Further it is a classical result of PDEs that the semigroup (Pt )t≥0 has a density p(t, x, y) with respect to
the measure mρ such that
Z
u(t, x) = p(t, x, y)f (y)ρ−1 (y)dy (2.1)
G

is a continuous version of Pt f (x). Then for f in C0 (G), Pt f belongs to C0 (G). By the use of the maximum
principle it can be shown that (Pt )t≥0 is semi-markovian and we get iii).
Finally Aronson estimates on the density p(t, x, y) can be used to show for example that
Z
1
∀ε > 0, ∀x ∈ G, lim p(t, x, y)ρ−1 (y)dy = 0,
t↓0 t |y−x|>ε

and thus we get iv) (see Proposition 2.9 in chapter 4 of [EK86]).

We have a consistency theorem.

Theorem 2.2 Let be 0 < λ < Λ < ∞. Let a and ρ be in Ell(λ, Λ) and (an , ρn ) be a sequence of
Ell(λ, Λ) × Ell(λ, Λ).
Let us denote by S and X respectively the semigroup and the process generated by L(a, ρ) and by (S n )
and (X n ) the sequences of semigroups and processes generated by the sequence of operators L(an , ρn ).
Assume that
1 L2 (G) 1 1 L2 (G) 1
−
− − −* , and −−−−* .
an n→∞ a ρn n→∞ ρ

Then for any T > 0 and any f ∈ L2 (G) we have :

253
i) The function Stn f (x) converges weakly in L2 (0, T ; H10 (G)) to St f (x).
ii) The continuous version of Stn f (x) given by (2.1) with p replaced by pn converges uniformly on each
compact of (0, T ) × G to the continuous version of St f (x) given by (2.1).
iii)
L
(Xtn , t ≥ 0) −−−−→ (Xt , t ≥ 0).
n→∞

Proof. See in [LM06] the proofs of Propositions 3 and 4.

3 On SDE involving local time

First we introduce a new class of coecients. For 0 < λ < Λ < ∞ we denote by Coeff(λ, Λ) the set of
the elements f of Ell(λ, Λ) that verify:

i) f is right continuous with left limits (r.c.l.l.).

ii) f belongs to C 1 (G \ I), where I is a countable set without cluster point.

Let us also denote by M the space of all bounded measures ν on G such that |ν({x})| < 1 for all x in G.

Denition 3.1 Let σ be in Coeff(λ, Λ) for some 0 < λ < Λ < ∞, and ν be in M. We call Stochastic
Dierential Equation with Local Time of coecients σ and ν , and we note Sde(σ, ν), the following SDE
Z t Z
Xt = X0 + σ(Xs )dWs + ν(dx)Lxt (X),
0 R

where Lxt (X) is the symmetric local time of the unknown process X .
In [Leg85] J.F. Le Gall studied some properties of SDEs of the type Sde(σ, ν). We will recall here some
results of this work we will use in the sequel.
We will see below that σ ∈ Coeff(λ, Λ) and ν ∈ M is a sucient condition to have a unique strong
solution to Sde(σ, ν). We rst x some additionnal notations.

For f in Coeff(λ, Λ) we denote by f 0 (dx) the bounded measure corresponding to the rst derivative of f
in the generalized sense. We denote by f (x+) and f (x−) respectively the right and left limits of f in x.
We will also denote by f 0 (x) the r.c.l.l. density of the absolutely continuous part of f 0 (dx) (that it is to
say we take for f 0 (x) the right derivative of f in x).
For ν in M we denote by ν c the absolutely continuous part of ν .

3.1 A change of scale formula

Let us dene the class of bijections we will use in our change of scale.
For 0 < λ < Λ < ∞ we denote by T(λ, Λ) the set of all functions Φ on G that have a rst derivative Φ0
that belongs to Coeff(λ, Λ). The assumption made on the bijection is minimal and we can then state a
very general change of scale formula.

254
Proposition 3.1 Let σ be in Coeff(λ, Λ) for some 0 < λ < Λ < ∞. Let
X
ν(dx) = b(x)dx + cxi δxi (dx),
xi ∈I

be in M (i.e., b is measurable and bounded, and each |cxi | < 1).

Let Φ be in T(λ0 , Λ0 ) for some 0 < λ0 < Λ0 < ∞ and let J be the set of the points of discontinuity of Φ0 .
Then the next statements are equivalent:
i) The process X solves Sde(σ, ν).
ii) The process Y := Φ(X) solves Sde(γ, µ) with

γ(y) = (σΦ0 ) ◦ Φ−1 (y),

and

Φ0 b + 21 (Φ0 )0 X
µ(dy) = ◦ Φ−1 (y)dy + βxi δΦ(xi ) (dy),
(Φ0 )2
xi ∈I∪J

where,
Φ0 (x+)(1 + cx ) − Φ0 (x−)(1 − cx )
βx = ,
Φ0 (x+)(1 + cx ) + Φ0 (x−)(1 − cx )
with cx = 0 if x ∈ J \ I .

Remark 3.1 Note that γ obviously belongs to Coeff(λ00 , Λ00 ) for some 0 < λ00 < Λ00 < ∞, and that µ is
in M, so it makes sense to speak of Sde(γ, µ).
We can say that the class of SDE of type Sde(σ, ν) is stable by transformation by a bijection belonging
to T(λ, Λ) for some 0 < λ < Λ < ∞.

Proof of Proposition 3.1. We prove i) ⇒ ii). The converse can be proven in the same manner quite
being technically more cumbersome.
By the symmetric Itô-Tanaka formula we rst get:
Rt Rt
Yt = Φ(Xt ) = Φ(X0 ) + 0
(σΦ0 )(Xs )dWs + 0
(σ 2 bΦ0 )(Xs )ds

Φ0 (xi +)+Φ0 (xi −)

cxi Lxt i (X)
P
+ xi ∈I 2

Rt Φ0 (xi +)−Φ0 (xi −) xi

+ 12 [σ 2 (Φ0 )0 ](Xs )ds +
P
0 xi ∈J 2 Lt (X)
Rt
= Φ(X0 ) + 0
(σΦ0 ) ◦ Φ−1 (Ys )dWs
Rt
[σ 2 (Φ0 b + 12 (Φ0 )0 )] ◦ Φ−1 (Ys )ds + Kxi Lxt i (X),
P
+ 0 xi ∈I∪J

with Kx = cx (Φ0 (x+) + Φ0 (x−))/2 + (Φ0 (x+) − Φ0 (x−))/2.

We have then to express Lxt (X) in function of LΦ(x)
t (Y ) for x ∈ I ∪ J .

255
Using Corollary VI.1.9 of [RY91] it can be shown that
Φ(x)±
Lt (Y ) = Φ0 (x±)Lx±
t (X). (3.1)
Besides theorem VI.1.7 of [RY91] leads to (Lx+
t (X) − Lt (X))/2 = cx Lt (X) and combining with
x− x

(Lt (X) + Lt (X))/2 = Lt (X) we get

x+ x− x

Lx+ x
t (X) = (1 + cx )Lt (X). (3.2)

In a similar manner we have (LΦ(x)+

t
Φ(x)−
(Y ) − Lt (Y ))/2 = Kx Lxt (X) and we can get

Φ(x) Φ(x)+
Kx Lt (X) + Lt (Y ) = Lt (Y ).

Then using (3.1) and (3.2) we get

Φ(x)
Lt (Y ) = (Φ0 (x+)(1 + cx ) − Kx )Lxt (X),

and the formula is proved.

To prove the proposition below, Le Gall used in [Leg85] a space bijection that enters in the general setting
of Proposition 3.1.

Proposition 3.2 (Le Gall 1985) Let σ be in Coeff(λ, Λ) for some 0 < λ < Λ < ∞ and ν be in M.
There is a unique strong solution to Sde(σ, ν).
We need two lemmas.

Lemma 3.1 (Le Gall 1985) Let σ be in Coeff(λ, Λ) for some 0 < λ < Λ < ∞. There is a unique
strong solution to Sde(σ, 0).
Proof. See [Leg85].
The next lemma will play a great role for calculations in the sequel.

Lemma 3.2 Let ν be in M. There exists a function fν in Coeff(λ, Λ) (for some 0 < λ < Λ < ∞),
unique up to a multiplicative constant, such that:

fν0 (dx) + (fν (x+) + fν (x−))ν(dx) = 0. (3.3)

Proof of Proposition 3.2. It suces to set

Z x
Φν (x) = fν (y)dy.
0

By Lemma 3.2 Φν obviously belongs to T(λ, Λ) for some 0 < λ < Λ < ∞. By Proposition 3.1
and Lemma
3.2 we get that X solves Sde(σ, ν) if and only if Y := Φν (X) solves Sde (σfν ) ◦ Φ−1
ν , 0 . By Lemma
3.1 the proof is completed.

256
3.2 Convergence results

Le Gall proved the next consistency result for equations of the type Sde(σ, ν).

Theorem 3.1 (Le Gall 1985) Let be two sequences (σn ) and (ν n ) for which there exist 0 < λ < Λ < ∞,
0 < M < ∞ and δ > 0 such that

(H1) σ n ∈ Coeff(λ, Λ), ∀n ∈ N,

(H2) |ν n ({x})| ≤ 1 − δ, ∀n ∈ N, ∀x ∈ G.

(H3) |ν n |(G) ≤ M, ∀n ∈ N,

so that each ν n is in M. Assume that there exist two functions σ and f in Coeff(λ0 , Λ0 ) (for some
0 < λ0 < Λ0 < ∞) such that

L1 (R) L1 (R)
σn −−loc
−−−→ σ
n→∞
and fν n −−loc
−−−→ f,
n→∞

and set:
f 0 (dx)
ν(dx) = − . (3.4)
f (x+) + f (x−)

Let (Ω, F, (Ft )t≥0 , Px ) be a ltered probability space carrying an adapted Brownian motion W . On this
space, for each n ∈ N let be X n the strong solution of Sde(σn , ν n ), and let be X the strong solution of
Sde(σ, ν). Then:

E [ sup |Xsn − Xs | ] −−−−→ 0

n→∞
and L
(Xtn , t ≥ 0) −−−−→ (Xt , t ≥ 0).
n→∞
0≤s≤t

Remark 3.2 In this theorem ν n approaches ν in the sense that fν tends to fν for the L1loc convergence.
n

Note we can have ν * ν1 , but f → fν2 for the

n
νn convergence, with ν1 6= ν2 (See [Leg85] p 65 for
L1loc
an example). The theorem asserts that X n tends to X that solves Sde(σ, ν2 ) and not Sde(σ, ν1 ) !

Le Gall also proved a Donsker theorem for solution to SDEs of the type Sde(σ, ν) for σ ≡ 1. Let be µ
in M and Y be the solution to Sde(1, µ).
We dene some coecients βkn for all k ∈ Z, and all n ∈ N∗ , by:

1 − βkn k k+1
 1 − µ({y})  fµ ( k+1

n )
(3.5)
Y
n = exp − 2µc (] , ]) = .
1 + βk n n k k+1
1 + µ({y}) fµ ( nk )
n <y≤ n

We dene a sequence (µn ) of measures in M by

(3.6)
X
µn = βkn δ k ,
n

and a sequence (Y n ) of processes such that each Y n solves Sde(1, µn ).

Finally we dene for all n ∈ N∗ a sequence (τpn )p∈N of stopping times by,

257
 τ0n = 0
(3.7)
n 1
τp+1 = inf{t > τpn : |Ytn − Yτnpn | = n }.

We have the next theorem.

Theorem 3.2 (Le Gall 1985) In the previous context Spn := nYτn denes a sequence of random walks
n
p
on the integers such that:
i)
S0n = 0, ∀n ∈ N∗ ,

n
P[Sp+1 = k + 1|Spn = k] = 21 (1 + βkn ), ∀n, p ∈ N∗ , ∀k ∈ Z,

n
P[Sp+1 = k − 1|Spn = k] = 21 (1 − βkn ), ∀n, p ∈ N∗ , ∀k ∈ Z.

ii)
The sequence of processes dened by Yetn := (1/n)S[n
n
t] , where 2 b.c stands for the integer part of a non
negative real number, veries for all 0 < T < ∞:
L
E|Yetn − Yt | −−−−→ 0, ∀t ∈ [0, T ] and (Yetn , t ≥ 0) −−−−→ (Yt , t ≥ 0).
n→∞ n→∞

4 Link between DFO and SDELT

This link is stated by the following proposition.

Proposition 4.1 Let a and ρ be in Coeff(λ, Λ) for some 0 < λ < Λ < ∞. Let us denote by I the set of
the points
√
of discontinuity of a. Then L(a, ρ) is the innitesimal generator of the unique strong solution
of Sde( aρ, ν) with,
 a0 (x)  X a(xi +) − a(xi −)
ν(dx) = dx + δx (dx). (4.1)
2a(x) a(xi +) + a(xi −) i
xi ∈I

In [LM06] the authors proved the proposition above by the use of Dirichlet forms and Revuz measures.
We give here a more simple proof, based on smoothing the coecients and using the consistency theorems
of the two preceeding sections.

Proof of Proposition 4.1. As a and ρ are in Coeff(λ, Λ) the function √ρa is in Coeff(λ, Λ). Besides,
as |a − b|/|a + b| < 1 for any√a,b in R∗+ , the measure ν dened by (4.1) is in M. The existence of a unique
strong solution X to Sde( aρ, ν) follows from Proposition 3.2.
We then identify the innitesimal generator of X . We can build two sequences (an ) and (ρn ) of functions
in Coeff(λ, Λ) ∩ C ∞ (G), such that

an −−−−→ a a.e. and ρn −−−−→ ρ a.e.

n→∞ n→∞

For any n in N we denote by X n the process generated by L(an , ρn ).

On one hand, by dominated convergence the hypotheses of Theorem 2.2 are fullled, and we have,

258
 L
(Xtn , t ≥ 0) −−−−→ (X
et , t ≥ 0), (4.2)
n→∞

where the process Xe is generated by L(a, ρ).

On the other hand we will show by Theorem 3.1 that
L
(Xtn , t ≥ 0) −−−−→ (Xt , t ≥ 0). (4.3)
n→∞

Thus taking in care (4.2) and (4.3) we will conclude that the innitesimal generator of X is L(a, ρ).
As an and ρn are C ∞ , (Ln , D(Ln )) = L(an , ρn ) can be written,

ρn h n 0 d d2 i
Ln = a + an 2 ,
2 dx dx
so it is standard to say that X n solves
Z tp t 0
ρn (Xsn )an (Xsn )
Z
Xtn = x + ρn (Xsn )an (Xsn )dWs + ds. (4.4)
0 0 2

As dhX n is = ρn (Xsn )an (Xsn )ds, by the occupation time density formula we can rewrite (4.4) and assert
that X n solves, Z t p Z
Xtn = x + ρn (Xsn )an (Xsn )dWs + νn (dx)Lxt (X n ),
0 R

where νn (dx) = (a (x)/2a (x))λ(dx).

n0 n

Then elementary calculations show that the function fνn associated to νn by Lemma 3.2 is of the form
fνn (x) = K/an (x) with K a real number. This obviously tends to K/a(x) =: f (x) for the L1loc (R)-
convergence. We then determine the measure ν associated to f by (3.4). First we check that ν c (dx) =
(a0 (x)/2(a(x))λ(dx). Second, the set {x ∈ G : ν({x}) 6= 0} is equal to I , and we have for all x ∈ I ,

f (x+) − f (x−) a(x+) − a(x−)

ν({x}) = − = .
f (x+) + f (x−) a(x+) + a(x−)

So the measure ν is equal to the one dened by (4.1). As it is obvious that

√ L1 (R) √
ρn an −−loc
−−−→ ρa,
n→∞

and that the hypotheses (H1) − (H3) of Theorem 3.1 are fullled, we can say that (4.3) holds. The proof
is completed.

5 Random walk approximation

5.1 Monte Carlo Approximation

From now the horizon 0 < T < ∞ is xed. For any a, ρ ∈ Coeff(λ, Λ) and any initial condition f we
denote by (P)(a, ρ, f ) the parabolic PDE

259
 ∂u(t,x)
for (t, x) ∈ [0, T ] × G,


 ∂t = Lu(t, x),



(P)(a, ρ, f )

u(t, l) = u(t, r) = 0 for t ∈ [0, T ],


for x ∈ G,

u(0, x) = f (x)


with (L, D(L)) = L(a, ρ).

Let be 0 < λ < Λ < ∞. From now till the end of this paper we assume that a and ρ are in Coeff(λ, Λ).
We denote by I = {xi }i∈I the set of the points of discontinuity of a (I = {0 ≤ i ≤ k1 } ⊂ Z is nite). We
set X to be the process generated by L(a, ρ).
We seek for a probabilistic numerical method to approximate the solution of (P)(a, ρ, f ). By Theorem
2.1 and some standard PDEs renements we know that for all f ∈ L2 (G), (P)(a, ρ, f ) has a unique weak
solution u(t, x) in C([0, T ], L2 (G, mρ )) ∩ L2 (0, T ; H10 (G)). We know that Ex [f (Xt )] is a continuous version
of u(t, x).
Our goal is to build a process Xb n easy to simulate and such that
L
btn , t ≥ 0) −−−−→ (Xt , t ≥ 0).
(X (5.1)
n→∞

Thus Ex [f (Xbtn )] → Ex [f (Xt )] for any t ∈ [0, T ], and the strong law of large numbers asserts that,
N
1 X b n,(i) n→∞
f (Xt ) −−−−→ u(t, x), (5.2)
N i=1 N →∞

where for each i, Xb n,(i) is a realisation of the random variable Xb n .

5.2 Skew Brownian Motion

The Skew Brownian Motion (SBM) of parameter β ∈ (−1, 1) starting from y , which we denote by Y β,y
is known to solve:

Ytβ,y = y + Wt + βLyt (Y β,y ), (5.3)

i.e. Y β,y solves Sde(1, βδ0 ) (see [HS81]).
It was rst constructed by Itô and McKean in [IM74] (Problem 1 p115) by ipping the excursions of a
reected Brownian motion with probability α = (β + 1)/2. On SBM see also [Wal78] . It behaves like a
Brownian motion except in y where its behaviour is pertubated, so that

P(Ytβ,y > y) = α, ∀t > 0. (5.4)

We denote by T (∆) the law of the stopping time τ = inf{t ≥ 0, |Wt | = ∆} where W is a standard
Brownian motion starting at zero. For the SBM Y β,0 and ∆ in R∗+ we dene the stopping time τ∆ =
inf{t ≥ 0, Ytβ,0 ∈ {∆, −∆}}. Our approach relies on the following lemma.

Lemma 5.1 Let y and x be in R, ∆ in R∗+ and β in (−1, 1). Set α = (β + 1)/2. Then
i) Y β,y + x ∼
= Y β,y+x , and ii) (τ∆ , Yτβ,0 ) ∼
= T (∆) ⊗ Ber(α).
∆

Proof. This follows from the construction of the SBM by Itô and McKean in [IM74].

260
5.3 Some possible approaches

Recently two methods have been proposed to build Xb n satisfying (5.1).

In [Mar04]
√
M. Martinez proposed to use an Euler scheme. We know by Proposition 4.1 that X solves
Sde( aρ, ν) with ν dened by (4.1). We have seen in the proof of Proposition 3.2 that if we dene
Φ(x) = 0 fν (y) then Y = Φ(X) solves Sde(γ, 0) with γ in some Coeff(m, M ). Thus an Euler scheme
Rx

approximation Yb n of Y can be built and by setting Xb n = Φ−1 (Yb n ) we get an approximation of X .

Because the coecient γ is not Lipschitz if a and ρ are not, evaluating the speed of convergence of such
a scheme is not easy.
In [LM06], A. Lejay and M. Martinez proposed to use the SBM. They rst build a piecewise
constant
√ napproximation (an , ρn ) of (a, ρ) in order that the process X n generated by L(an , ρn ) solves
Sde( a ρ , ν ) with ν satisfying (ν n )c = 0. Second
n n n
P by a proper bijection Φn ∈ T(m, M ), they get
that Y = Φ (X ) solves Sde(1, µ ) with µ = βk δyk , i.e. Y behaves locally like a SBM. Third
n n n n n n

they proposed a scheme Yb n for Y n based on Lemma 5.1 and simulations of exit times of the SBM and
they nally set Xb n = (Φn )−1 (Yb n ).
Our method can be seen as a variation of this last approach because it also deeply relies on getting such
a Y n and using Lemma 5.1. But we then use random walks instead of the scheme proposed in [LM06].

5.4 The basic idea of our approach

We focus on weak convergence and propose a three-step approximation scheme diering slightly from the
one proposed by Theorem 3.2.

We x n ∈ N∗ , and 1/n will be the spatial discretization step size.

STEP 1. We build (an , ρn ) in Coeff(λ, Λ) × Coeff(λ, Λ) such that:
i) The functions an and ρn are piecewise constant. The points of discontinuity of either an and ρn
are included in some set I n . We assume I n = {xnk }k∈I n for I n = {0 ≤ k ≤ k1n } ⊂ Z nite and
xnk < xnk+1 , ∀k ∈ I n .

ii) For each xnk ∈ I n we have an (xnk ) = a(xnk ) and ρn (xnk ) = ρ(xnk ).
iii) Consider the function
kn,x −1
xn − xnk x − xnkn,x
(5.5)
X
Φn (x) = p k+1n + q ,
k=0
a(xk )ρ(xnk ) a(xnkn,x )ρ(xnkn,x )

where the integer kn,x veries xnkn,x ≤ x ≤ xnkn,x +1 .

The set I n satises Φn (I n ) = {k/n, k ∈ Z} ∩ Φn (G). From now we assume xnk is the point of I n such
that Φn (xnk ) = k/n.

Remark 5.1 In fact the rst thing to do is to construct the grid I n satisfying iii). It is very easy and
only requires to know the coecients a and ρ (see point 1 of the algorithm in Subsection 5.5). Then an
and ρn can be constructed.

Remark 5.2 The sets I and I n may have no common points.

We take X n to be the process generated by L(an , ρn ).

261
Remark 5.3 It can be shown by Theorem 2.2 that X n converges in law to X .
√
STEP 2. By Proposition 4.1 the process X n solves Sde( an ρn , ν n ) with

X an (xn +) − an (xn −)
νn = k k
δxn .
n n
an (xnk +) + an (xnk −) k
xk ∈I

The function0 Φn dened by (5.5) belongs to T(1/Λ, 1/λ). The points of discontinuity of Φn are those in
0

I n , and (Φn )0 = 0, so by Proposition 3.1 the process Y n = Φn (X n ) solves

k/n
(5.6)
X
Ytn = Y0n + Wt + βkn Lt (Y n ),
xn
k ∈I
n

where
p q
a(xnk )/ρ(xnk ) −
a(xnk−1 )/ρ(xnk−1 )
βkn = p q . (5.7)
a(xnk )/ρ(xnk ) + a(xnk−1 )/ρ(xnk−1 )

To write these coecients we have used the fact that an and ρn are r.c.l.l. and that for instance an (xnk +) =
a(xnk ) and an (xnk −) = a(xnk−1 ).

Remark 5.4 We have got Y n that solves Sde(1, βkn δk/n ) in a dierent way than the one used by Le
P

Gall in Theorem 3.2. We now use his method to get Ye n that veries

E|Yetn − Ytn | −−−−→ 0, ∀t ∈ [0, T ]. (5.8)

n→∞

STEP 3. Like in (3.7) we dene a sequence (τpn )p∈N of stopping times by,

τ0n = 0, and n
τp+1 = inf{t > τpn : |Ytn − Yτnpn | = 1/n}.

Thanks to the uniformity of the grid {k/n, k ∈ I n } we have the following lemma.

Lemma 5.2 i) For all k ∈ Z and all p ∈ N, (Yτn +u − Yτn , 0 ≤ u ≤ τp+1

n
p
n
p
n
− τpn ) knowing that {Yτn n
p
= k/n}
n
has the same law as (Yt , 0 ≤ t ≤ τ1/n ).
β ,0
k

ii)
∀p ∈ N, σpn := n2 τpn − τp−1
n ∼


= T (1),

and the σpn 's are independent.

Proof. The statement i) follows simply from point i) of Lemma 5.1 and the comparison between (5.3) and
(5.6). From i), the strong Markov property, and point ii) of Lemma 5.1, we get that (τpn −τp−1
n
= T (1/n2 ),
)∼
and the statement ii) follows by scaling. Using again the Markov property we get the independence of
the σpn 's.

Le Gall used this lemma in his proof of Theorem 3.2. Indeed it is obvious by the i) of Lemma 5.2 and
the ii) of Lemma 5.1 that Spn := nYτnpn satises

262
 S0n = 0,

n
P[Sp+1 = k + 1|Spn = k] = 12 (1 + βkn ) =: αkn , ∀p ∈ N∗ , ∀k ∈ I n , (5.9)
n
P[Sp+1 = k − 1|Spn = k] = 21 (1 − βkn ) = 1 − αkn , ∀p ∈ N∗ , ∀k ∈ I n .

Moreover the ii) of Lemma 5.2 allows to show that Yetn := (1/n)S[n
n n
2 t] = Yτ n
2
satises (5.8).
[n t]

Thus the idea is to take

btn := (Φn )−1 1 Sbn 2 , (5.10)

X
n [n t]

where Sbn is a random walk on the integers dened by (5.9). The process Xb n is a random walk on the
grid I n . In fact this grid is made in order that Xb n spends the same average time in each of its cells.
Combining remark 5.3 and theorem 3.2 we should have (5.1). To sum up this section we write our scheme
in the algorithm form. In the next section we will estimate the approximation error of our scheme.

5.5 The algorithm

Note that by construction (Φn )−1 k/n = xnk for all k ∈ I n .

We dene a function ALGO in the next manner:

INPUT DATA: the coecients a and ρ, the starting point x, the precision order n and the nal time t.
OUTPUT DATA: an approximation in law Xb n of X at time t.
1. Set xn0 ← l.
while xnk ≤ r
set xnk ← a(xnk )ρ(xnk )(1/n) + xnk and k ← k + 1.
p

endwhile
2. Compute the αkn = (1 + βkn )/2 with βkn dened by (5.7).
3. Set y ← Φn (x).
if (n y − bn yc) < 0.5
set s0 ← bn yc.
else
set s0 ← bn yc + 1.
endif
4. for i = 0 to i = bn2 tc =: N
if xns
i
∈ R \ (l, r)
Return xnsi .
endif
We have si = k for some k ∈ I k . Simulate a realization B of Ber(αkn ).
Then set si ← si + B .

263
endfor
5. Return xnsN .

6 Speed of convergence
In this section we will prove the following theorem.
Theorem 6.1 Assume that a, ρ ∈ Coeff(λ, Λ) for some 0 < λ ≤ Λ < ∞. Let be 0 < T < ∞ and X the
process generated by L(a, ρ). For n ∈ N consider the process Xb n starting from x dened by,

∀t ∈ [0, T ], btn = ALGO(a, ρ, x, n, t).

X
For all f ∈ W01,∞ (G) ∩ C0 (G), all ε > 0, and all γ ∈ (0, 1/2) there exists a constant C depending on ε, γ ,
T , λ, Λ, G, ka0 k∞ , kρ0 k∞ kf k∞ , kdf /dxk2 , kdf /dxk∞ , supi∈I 1/(xi+1 − xi ), and the two rst moments
of T (1) such that, for n large enough,

sup btn ) ≤ Cn−γ .

Ex f (Xt ) − Ex f (X
(t,x)∈[ε,T ]×Ḡ

We have,

|Ex f (Xt ) − Ex f (X
btn )| ≤ |Ex f (Xt ) − Ex f (Xtn )| + |Ex f (Xtn ) − Ex f (X
btn )|
(6.1)
=: e1 (t, x, n) + e2 (t, x, n).
We will estimate e1 (t, x, n) by PDEs techniques and e2 (t, x, n) by very simple probabilistic techniques.

6.1 Estimate of a weak error

In this subsection we prove the following proposition.

Proposition 6.1 Assume f belongs to H10 (G)∩C0 (G). Let be u(t, x) and un (t, x) respectively the solutions
of (P)(a, ρ, f ) and (P)(an , ρn , f ), with an and ρn like in Subsection 5.4, Step 1. Then for all ε > 0 there
is a constant C1 depending on ε, T , λ, Λ, G, kf k∞ , kdf /dxk2 , ka0 k∞ , kρ0 k∞ , and supi∈I 1/(xi+1 − xi )
such that for n large enough,
1
sup |u(t, x) − un (t, x)| ≤ C1 √ .
(t,x)∈[ε,T ]×Ḡ n

As we will see in Proposition 6.2, if we had I ⊂ I n we could obtain an upper bound for |||u − un |||∞,∞
of the form K(ka − an k2∞ + kρ − ρn k∞ ). But this is not necessary the case (see Remark 5.2). However
it is possible to modify a and ρ in order to rend us in a situation close to this one, and we will do that
to prove Proposition 6.1.
Proposition 6.2 Let be f ∈ H10 (G)∩C0 (G). Let be a1 , ρ1 , a2 , ρ2 ∈ Coeff(λ, Λ), and I1 and I2 respectively
the set of points of discontinuity of a1 and ρ1 and a2 and ρ2 . Assume I1 ⊂ I2 . Let be u1 (t, x) the weak
solution of (P)(a1 , ρ1 , f ) and u2 (t, x) the weak solution of (P)(a2 , ρ2 , f ). There exists a constant Ce1
depending on T , λ, Λ, G, kf k∞ , and kdf /dxk2 , such that,
 
e1 ka1 − a2 k2 + kρ1 − ρ2 k
|||u1 − u2 |||∞,∞ ≤ C ∞ ∞ .

264
We need a lemma asserting some standard estimates.
Lemma 6.1 i) Let be f ∈ H10 (G). Let be u(t, x) the weak solution of (P)(a, ρ, f ). Then ∂t u is in
L2 (0, T ; L2 (G)) and more precisely,

Λ df
|||∂t u|||2,2 ≤ . (6.2)
2 dx 2

ii) Let be f ∈ L2 (G). Let be u(t, x) the weak solution of (P)(a, ρ, f ). Then du
dx is in L2 (0, T ; L2 (G)) and
more precisely,
du 1
≤ kf k2 . (6.3)
dx 2,2 λ

Proof. i)Step 1. Assume rst that a and ρ are C ∞ (G) and that f is Cc∞ (G) so that u(t, x) is itself
C ((0, T ) × G). As u(t, x) is a weak solution of (P)(a, ρ, f ), and u, Lu and ∂t u are C ∞ , using ∂t u as a
∞

test function and integrating by parts with respect to x, we get

Z T Z Z T Z Z T Z
1 du d(∂t u)
|∂t u|2 mρ (dx)dt = ∂t u Lu mρ (dx)dt = − a dx dt.
0 G 0 G 2 0 G dx dx
Then using Fubini's theorem, interverting the partial derivatives, integrating by parts with respect to t
and then again with respect to x, we get
Z T Z Z 2 Z 2
2 1 du(0, x) 1 du(T, x)
2 |∂t u| mρ (dx)dt = a dx − a dx,
0 G 2 G dx 2 G dx
which leads to (6.2).
Step 2. In the general case, with a and ρ in Coeff(λ, Λ), and f in H10 (G), we use a regularization
argument, Theorem 2.2, Step 1, a compactness argument and an integration by parts with respect to t,
to exhibit a function w ∈ L2 (0, T ; L2 (G)) satifying:
Z T Z Z T Z
Λ df
∀ϕ ∈ Cc∞ ((0, T ) × G), wϕ = − u∂t ϕ, and |||w|||2,2 ≤ .
0 G 0 G 2 dx 2

That is ∂t u is in L2 (0, T ; L2 (G)) and veries (6.2).

ii) Thanks to point i) and because Lu(t, .) ∈ L2 (G) we can write
Z T Z T
h∂t u, uiL2 (G,mρ ) dt = hLu, uiL2 (G,mρ ) dt, (6.4)
δ δ

for all δ > 0. As u is in C 1 ([δ, T ], L2 (G, mρ ))dt and we have (see [Bre83]),
d
2 h∂t u, uiL2 (G,mρ ) =
2
kukL2 (G,mρ ) , ∀t ∈ [δ, T ], (6.5)
dt
using an integration by part with respect to x in the right hand side of (6.4), and making δ tend to 0 we
get,
2
λ du 1 2 2

≤ ku(0, .)kL2 (G,mρ ) − ku(T, .)kL2 (G,mρ ) ,
2 dx 2,2 2

which leads to (6.3).

265
Proof of Proposition 6.2. Step 1. We introduce the following norm on
C(0, T ; L2 (G)) ∩ L2 (0, T ; H10 (G)):

2 1/2

2 dv
|v|G,T := sup kv(t, .)k2 + .
t∈[0,T ] dx 2,2

We have the following estimate:

|||v|||∞,∞ ≤ K|v|G,T , (6.6)

where the constant K depends only on T and G (see [LSU68], II.Ÿ.3 inequality (3) p 74 with r = ∞ and
q = ∞).
We set v := u2 − u1 . Our goal is now to estimate |v|G,T .
Step 2. Set (L2 , D(L2 )) = L(a2 , ρ2 ). Elementary computations show that, v(t, x) is a weak solution to
ρ2 (x) d du1 (t, x)
ρ2 (x)
∂t v(t, x) = L2 v(t, x) + (a2 (x) − a1 (x)) +( − 1)∂t u1 (t, x),
2 dx dx ρ1 (x)

that is to say for all ϕ ∈ Cc∞ ((0, T ) × G) we have,

RT R RT R dv dϕ
− 0 G
v∂t ϕmρ2 (dx)dt = − 0 G
a2 dx dx dx dt

R T R  ρ2
(6.7)
ρ2
d
− a1 ) du

+ 2 dx ((a2 dx ) + ( ρ1 − 1)∂t u1 ϕmρ2 (dx)dt.
1
0 G

We take v as a test function in (6.7). Using (6.5) with v , integration by parts, ab ≤ (λ/2)a2 + (2/λ)b2
and a1 , ρ2 ∈ Coeff(λ, Λ) we nally get,
T T
κ0 
Z Z  du 2 Z Z 
1 1 
|v|G,T ≤ (a1 − a2 ) 2 1
dxdt + − ∂t u1 vdxdt , (6.8)
κ 0 G dx 0 G ρ1 ρ2

where κ = min(1/Λ, λ/4) and κ0 = max(1/λ, 1).

Step 3. As f ∈ C0 (G) and the semigroup (St1 )t≥0 and (St2 )t≥0 generated respectively by (L1 , D(L1 )) =
L(a1 , ρ1 ) and (L2 , D(L2 )) are Feller, we can consider that

∀t ∈ [0, T ], kv(t, .)k∞ ≤ ku2 (t, .)k∞ + ku1 (t, .)k∞ ≤ 2 kf k∞ .

Thus |||v|||∞,∞ ≤ 2 kf k∞ ; besides f ∈ H10 (G) thus using point i) of Lemma 6.1 and Hölder inequality we
get,
RT R
0
(1
G ρ1
− 1
ρ2 )∂t u1 vdxdt ≤ 2 1
ρ1 − 1
ρ2 kf k∞ |||∂t u1 |||1,1
∞

(6.9)
2
p
≤ λ2 T |G| kρ1 − ρ2 k∞ kf k∞ |||∂t u1 |||2,2 .

df
Λ
p
≤ λ2 T |G| kf k∞ dx kρ1 − ρ2 k∞ .
2

To nish we have,
Z T Z 2
du1 2 2 du1
(a1 − a2 )2 ( ) dxdt ≤ ka1 − a2 k∞ ,
0 G dx dx 2,2

266
and point ii) of Lemma 6.1 completes the proof because f ∈ L2 (G).
Note that the fact that I1 ⊂ I2 allows to consider the quantities kρ1 − ρ2 k∞ and ka1 − a2 k∞ .

We are then ready to prove Proposition 6.1.

Proof of Proposition 6.1. The idea is to build a bijection φn such that the points of discontinuity of
a ◦ φn and ρ ◦ φn are included in I n .
Step 1. We build a piecewise linear bijection φn such that φ−1
n (I) ⊂ I in the following manner:
n

We rst dene a projection πn : G → I n by,

 n
 xkn,x if (x − xnkn,x )/(xnkn,x +1 − xnkn,x ) < 1/2,
πn (x) =
xnkn,x +1 if (x − xnkn,x )/(xnkn,x +1 − xnkn,x ) ≥ 1/2.


Then we set
x0 −l


 πn (x0 )−l (x − l) + l if x ∈ [l, πn (x0 )[,




xi+1 −xi
− πn (xi )) + xi if x ∈ [πn (xi ), πn (xi+1 )[,

πn (xi+1 )−πn (xi ) (x

φn (x) =

 for 0 < i < k1 − 1,




 r−xk1

r−πn (xk1 ) (x − πn (xk1 )) + xk1 if x ∈ [πn (xk1 ), r].

Note that we then have for all i ∈ I , φn (πn (xi )) = xi .

Step 2. We set ũn (t, x) := u(t, φn (x)). If a and ρ are smooth simple computations show that ũn (t, x)
solves (P)(ãn , ρ̃n , f ◦ φn ) with,

a ◦ φn (x)
ãn (x) = , and ρ̃n (x) = ρ ◦ φn (x).
φ0n (x)

It can be shown that this is still the case for a and ρ in Coeff(λ, Λ), using a regularization argument and
again Theorem 2.2.
We also dene ūn (t, x) to be the solution of (P)(ãn , ρ̃n , f ).
Step 3. For all (t, x) ∈ [ε, T ] × Ḡ, we have,

|u(t, x) − un (t, x)| ≤ |u(t, x) − u(t, φn (x))| + |||ũn − ūn |||∞,∞ + |||ūn − un |||∞,∞ .

The points of discontinuity of ãn and ρ̃n belong to I n . So by Proposition 6.2 there is a constant Ce1 not
depending on n such that
 
e1 kãn − an k2 + kρ̃n − ρn k
|||ūn − un |||∞,∞ ≤ C ∞ ∞ . (6.10)

Besides, as for each n the semigroup (Setn )t≥0 generated by L(ãn , ρ̃n ) is Feller, we have

df
q
|||ũn − ūn |||∞,∞ ≤ kf − f ◦ φn k∞ ≤ kid − φn k∞ . (6.11)
dx 2

267
Finally, we know that u(t, x) is continuous on [0, T ] × Ḡ and of class C 1 on each [ε, T ] × (xi , xi+1 ) (see in
[LRU68] Theorems 6 and 7). So if x and φn (x) belong to the same interval (xi , xi+1 ) we have,

du
|u(t, x) − u(t, φn (x))| ≤ sup | (t, x)| · |x − φn (x)|.
(t,x)∈[ε,T ]×(xi ,xi+1 ) dx

Let us set

du
M = sup sup | (t, x)|.
i∈I (t,x)∈[ε,T ]×(xi ,xi+1 ) dx

If for instance x ∈ (xi−1 , xi ) and φn (x) ∈ (xi , xi+1 ) we have,

|u(t, x) − u(t, φn (x))| ≤ |u(t, x) − u(t, xi )| + |u(t, xi ) − u(t, φn (x))| ≤ 2M kid − φn k∞ . (6.12)
We will see below that kid − φn k∞ → 0 as n → ∞, so for n large enough we are always at least in the
last situation.
Step 4. By construction (see point 1 of the algorithm) the grid I n satises |xnk+1 − xnk | ≤ Λ/n, for all
k ∈ I n.
So elementary computations show that

3Λ 1 1 1
kid − φn k∞ ≤ and 1− ≤ 2Λ sup . .
n φ0n ∞ i∈I xi+1 − xi n

As we have said above, for n large enough (6.12) is valid and we then have,
1
|u(t, x) − u(t, φn (x))| ≤ 6M Λ . (6.13)
n
It remains to evaluate kãn − an k∞ and kρ̃n − ρn k∞ . On each (xi , xi+1 ), a is of class C 1 and a0 is r.c.l.l.,
so it makes sense to speak of ka0 k∞ . Moreover each φn ([xnk , xnk+1 )) is included in some [xi , xi+1 ] that
contains xnk , so we have,

kãn − an k∞ ≤ supk∈I n supx∈[xnk ,xnk+1 ) |a(φn (x))/φ0n (x) − an (x)|

≤ ka0 k∞ supx∈[xnk ,xnk+1 ) |φn (x) − xnk | + Λ supx∈[xnk ,xnk+1 ) |1 − 1

φ0n (x) |.

In addition |φn (x) − xnk | ≤ |φn (x) − x| + |x − xnk | ≤ 4Λ/n, for all x ∈ [xnk , xnk+1 ), and we can get a similar
bound for |1 − φ0n1(x) | so nally there exists K1 such that
1
kãn − an k∞ ≤ K1 . (6.14)
n
In a similar manner we get K2 such that,
1
kρ̃n − ρn k∞ ≤ K2 . (6.15)
n
Thus, combining (6.10), (6.11), (6.13), (6.14) and (6.15), we complete the proof.

268
6.2 Estimate of a strong error

Proposition 6.3 In the context of Subsection 5.4, for all γ ∈ (0, 1/2) there exist a constant C2 depending
on T , γ and the two rst moments of T (1) such that,

∀n ∈ N∗ , ∀t ∈ [0, T ], E|Yτnn 2 − Ytn | ≤ C2 n−γ .

[n t]

Note that in all this subsection we drop any reference to the starting point x in the notation of the
expectation. Indeed the calculations we make are uniform with respect to this variable.
The proof of Proposition 6.3 will follow from two simple lemmas.
For all γ ∈ (0, 1/2), all T > 0, and for any process Y let us introduce the notation

|Yt − Ys |
MTγ (Y ) := sup γ
.
s6=t, s,t∈[0,T ] |t − s|

Lemma 6.2 Let (µn ) be a sequence in M and let be (Y n ) the sequence of processes such that each Y n
solves Sde(1, µn ). Assume there exist two positive constants m and M such that,

∀n ∈ N, fµn ∈ Coeff(m, M ). (6.16)

Then for all γ ∈ (0, 1/2), all k ∈ N∗ such that 0 < γ < 1/2 − 1/(2k), and all T > 0, there exists a positive
constant Ckγ,T , not depending on n, which veries,
 1/2k
∀n ∈ N, E[(MTγ (Y n ))2k ] ≤ Ckγ,T < ∞. (6.17)

Proof. We have to use the Kolmogorov-ƒentsov theorem. Let be γ ∈ (0, 1/2) and k ∈ N∗ such that
0 < γ < 1/2 − 1/(2k). For all n ∈ N let us dene
Z x
Fn (x) := fµn (y)dy,
0

and

hn := fµn ◦ Fn−1

Let us evaluate E|Ytn − Ysn |2k for n ∈ N and t, s > 0. We have

E|Ytn − Ysn |2k = E|Fn−1 (Ztn ) − Fn−1 (Zsn )|2k , (6.18)

where Z n := Fn (Y n ) is solution of Sde(hn , 0) by the virtue of Proposition 3.1.
By (6.16) we have that ||(Fn−1 )0 ||∞ ≤ 1/m, so a simple use of the mean value theorem leads from (6.18)
to,
Z t
1 1
E|Ytn − Ysn |2k ≤ E|Ztn − Zsn |2k = 2k E| hn (Yrn )dWr |2k .
m2k m s

Using now the Burkholder-Davis-Gundy inequality and the majoration part of (6.16) we get

1 Z t k  M 2k
E|Ytn − Ysn |2k ≤ 2k Ck E h2n (Yrn )ds = Ck (t − s)1+(k−1) ,
m s m

269
  2k
where Ck is a constant not depending on n. The constant M m Ck does not depend on n and we have
2k > 1. Thus, having a look at the proof of the Kolmogorov-ƒentsov theorem (see [RY91] for instance),
and identifying each Y n with its γ -Hölder modication we can say that there is a positive constant Ckγ,T ,
such that (6.17) holds.

Lemma 6.3 There exists a constant K depending on T and the two rst moments of T (1) such that:
1
∀n ∈ N∗ , ∀t ∈ [0, T ], n
E|τ[n 2
2 t] − t| ≤ K . (6.19)
n2

Proof. First we notice that if t < 1/n2 then (6.19) holds with K = 1. We then assume now that
t ≥ 1/n . For all t ∈ [0, T ] and with the σpn dened in Lemma 5.2 we set σpn,t := tσpn . Let T1 ∼
2
= T (1).
Using,

(v1 + . . . + vk )2 ≤ k(v12 + . . . + vk2 ), (6.20)

we get
2 2
[n t] [n t]
1 X n,t 1 X n,t
(6.21)
n 2 n 2 2
E|τ[n 2 t] − t| ≤ 2E τ[n 2 t] − 2 σ + 2E 2 σ −t .
[n t] p=1 p [n t] p=1 p

We have (τpn − τp−1

n
= T (1/n2 ) so we get,
)∼

P[n2 t] P[n2 t] 2
n 1 2 [n2 t]−n2 t

E τ[n 2 t] − [n2 t] p=1 σpn,t =E p=1 (τpn − τp−1
n
) [n2 t]

≤ 1
P[n2 t]
E(τpn − τp−1
n
)2 (6.22)
[n2 t] p=1

≤ E(T12 ) n14 .

For the second term of (6.21), as ET1 = 1 (see [Bre68] for instance) and the σpn 's are independent, we
have

1
P[n2 t] 2 t2
P[n2 t] 2
E [n2 t] p=1 σpn,t − t = [n2 t]2 E p=1 (σpn − 1)

= t2
P[n2 t]
V(σpn ) (6.23)
[n2 t]2 p=1

≤ 2T V(T1 ) n12 .

Taking in account (6.22), (6.23) and (6.21) we complete the proof.

Proof of Proposition 6.3. Let be γ ∈ (0, 1/2), k ∈ N∗ such that 0 < γ < 1/2 − 1/(2k), and n ∈ N∗ .
By the Hölder inequality we have for p and q conjugate

(6.24)

p
∀t ∈ [0, T ], E|Yτnn 2 − Ytn | ≤ [E MTγ (Y n ) ]1/p .[E|τ[n
n
2 t] − t|
γq 1/q
] .
[n t]

270
Let us x q = 2/γ . We have 1 < p < 4/3. Each Y n solves Sde(1, µn ) with µn = k∈I n βkn δk/n with the
P
βkn dened by (5.7). The function fµn is unique up to a multiplicative constant. If we impose fµn (l) = 1
then simple calculations show that
q
Y 1− βkn ρ(xnkn,x )
fµn (x) = =q .
k n
1+ βkn a(xnkn,x )
n ≤x,k∈I

Thus each fµn is in Coeff( Λ/λ, λ/Λ) and by Lemma 6.2, there exists Ckγ,T verifying (6.17).
p p

As p < 2k by Jensen inequality there exists Kγ uniform in n such that

∀n ∈ N∗ , [E(MTγ (Y n )p )]1/p < Kγ < ∞. (6.25)

By Lemma 6.3 there exists K such that

∀n ∈ N∗ , ∀t ∈ [0, T ], [E|τ[n
n
2 t] − t|
γq 1/q
] n
= [E|τ[n 2 1/q
2 t] − t| ] ≤ K n−2/q = K n−γ . (6.26)
Combining (6.24), (6.25) and (6.26) we can complete the proof.

6.3 Proof of Theorem 6.1

Combining the two preceeding subsections we can nally prove Theorem 6.1.

Proof of Theorem 6.1. We have e1 (t, x, n) = |u(t, x) − un (t, x)| where u(t, x) and un (t, x) are those of
Proposition 6.1.
Moreover, by construction, Xb n of Theorem 6.1 is distributed as Xe n := (Φn )−1 (Ye n ) with the Ye n of
Subsection 5.4 and Proposition 6.3, that lives in the same probality space as Y n . So, as ∇(Φn )−1 ∞ ≤
1/Λ, we have,

df
x etn ) df dx
∞
e2 (t, x, n) = |E [f (X − f (Xtn )]| ≤ E x etn
|X − Xtn | ≤ Ex |Yetn − Ytn |.
dx ∞ Λ

Thus using Propositions 6.1 and 6.3 we complete the proof.

7 Numerical experiments
Example 1. We take a and ρ to be
 
1 if x < 0, 1 if x < 0,
a(x) = and ρ(x) =
5 if x ≥ 0, 1/5 if x ≥ 0.

Then by Proposition 4.1, the process X generated by L(a, ρ) solves

Sde(1, 2/3 δ0 ), i.e. X is distributed as the simple SBM Y x,β of parameter β = 2/3.
We know the exact density p(t, x, y) of the transition probability of Y x,β (see [Wal78]). For x = 0 we
have:

271
 0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
−5 −4 −3 −2 −1 0 1 2 3 4 5

Figure 1: Approximated density p(t, x, y) for X with t = 1 and x = 0 (with 50000 particles), together
with the exact density (represented by the ∗-line).

if y ≥ 0,

√  (β + 1) exp{−y 2 /2t}
2πt p(t, 0, y) =

(1 − β) exp{−y 2 /2t} if y < 0.
We simulate N = 50000 random variables Xbtn = ALGO(a, ρ, x, n, t) with x = 0, t = 1 and the precision
order n = 20. We plot on the same graph (Figure 1) the histogram we get and the exact p(1, 0, y).

Example 2. We take ρ ≡ 1 and the coecient a represented by Figure 2. We plot a histogram

approximating p(t, x, y) for x = 0 at three successive times, t = 1, t = 2 and t = 3.5. We used N = 10000
particules and took n = 20 (Figure 3).

Example 3. We take the same n, a and ρ as in example 2. For f ∈ W01,∞ (G) ∩ C0 (G), we know

200

180

160

140

120

100

80

60
−60 −40 −20 0 20 40 60

Figure 2: Graph of a.

272
 0.04 0.04

0.035 0.035

0.03 0.03

0.025 0.025

0.02 0.02

0.015 0.015

0.01 0.01

0.005 0.005

0 0
−100 −80 −60 −40 −20 0 20 40 60 80 100 −100 −80 −60 −40 −20 0 20 40 60 80 100

0.04

0.035

0.03

0.025

0.02

0.015

0.01

0.005

0
−100 −80 −60 −40 −20 0 20 40 60 80 100

Figure 3: Approximation of p(t, x, y) for x = 0 at times t = 1, t = 2 and t = 3.5 (with N=10000

particules).

(Subsection 5.1) that if each Xb n,(i) is a realisation of Xb n starting at x, the quantity (1/N ) N b n,(i) ] P
i=1 f [X
approaches u(t, x), the solution of (P)(a, ρ, f ).
We then consider that G = (−100, 100) and take f (x) = sin(π (x + 100)/200). We take x = −10. We
compute usto (t, x) := (1/N ) N b n,(i) ], with N = 10000, for t belonging to a time grid that is a
P
i=1 f [Xt
discretisation of [0, 4]. We use a deterministic algorithm to compute an approximation udet (t, x) of u(t, x)
for t ∈ [0, 4]. We plot usto (t, x) and udet (t, x) on the same graph (Figure 4).

Example 4. We wish to compare our scheme with the one proposed by Lejay and Martinez in [LM06].
In order to do that we take ρ ≡ 1 and a dened by

if x < 0,

2 + sin(x)
a(x) =
5 + sin(x + π) if x ≥ 0.

We take t = 1, x = 0.5, n = 10 and N = 10000. We plot a histogram of the values of Xb1n approximating
p(1, 0.5, y) on Figure 5. We plot on the same gure the histogram obtained in [LM06] for the same
parameters. Note that in [LM06] the authors also compared their histogram with the one obtained by
the Euler Scheme of M. Martinez in [Mar04].

Acknowledgement: The author would like to thank Antoine Lejay for his encouragements and
constructive remarks.

273
 0.99

0.98

0.97

0.96

0.95

0.94

0.93
0 0.5 1 1.5 2 2.5 3 3.5 4

Figure 4: Graphs of usto (t, x) and udet (t, x) (represented by the ∗-line) for x = −10 and t ∈ [0, 4].

0.35

0.3

0.25

0.2

0.15

0.1

0.05

0
−10 −8 −6 −4 −2 0 2 4 6 8 10

Figure 5: Approximation of p(t, x, y) in example 4 by our algorithm together with the one by Lejay and
Martinez (represented by the dashed line).

274
References
[BC05] R.F. Bass Z.-Q. Chen One-dimensional Stochastic Dierential Equations with Singular and
Degenerate Coecients. Sankhya 67 (2005), 19-45.
[Bre68] L. Breiman. Probability. Addison-Wesley Series in Statistics, 1968. MR229267
[Bre83] H. Brezis. Analyse fonctionnelle. Masson, 1983. MR697382
[DDG05] M. Decamps, A. De Schepper and M. Goovaerts. Applications of δ -function pertubation to the
pricing of derivative securities. Physica A 342:3-4 (2004), 677-692. MR2105796
[EK86] S.N. Ethier and T.G. Kurz Markov processes. Characterization and convergence. Wiley Series
in Probability and Mathematical Statistics, 1986. MR838085
[HS81] J.M. Harrison and L.A. Shepp. On skew Brownian Motion, The Annals of Probability 9 (1981),
309-313. MR606993
[IM74] K. Itô and H.P. McKean. Diusion Processes and Their Sample Paths. Springer, Berlin, 1974.
[KP92] P.E. Kloeden and E. Platen. Numerical Solution of Stochastic Dierential Equations. Springer,
Berlin, 1992. MR1214374
[LSU68] O.A. Ladyzenskaja, V.A. Solonnikov and N.N. Ural'ceva. Linear and Quasilinear Equations of
Parabolic Type. American Mathematical Society, 1968. MR241822
[LRU68] O.A. Ladyzenskaja, V.J. Rivkind and N.N. Ural'ceva. The classical solvability of diraction
problems. Trudy Mat. Inst. Steklov 92 (1966), 116-146. MR211050
[Lej00] A. Lejay. Méthodes probabilistes pour l'homogénéisation des opérateurs sous forme divergence:
cas linéaires et semi-linéaires. PhD thesis, Université de Provence, Marseille, France, 2000.
[LM06] A. Lejay and M. Martinez. A scheme for simulating one-dimensional diusion processes with
discontinuous coecients. To appear in The Annals of Applied Probability.
[Leg85] J.F. Le Gall. One-dimensional stochastic dierential equations involving the local time of the
unknown process, Stochastic Analysis and Applications, vol. 1095 of Lect. Notes in Math. 51-82,
Springer 1985. MR777514
[Mar04] M. Martinez. Interprétations probabilistes d'opérateurs sous forme divergence et analyse de
méthodes numériques probabilistes associées. PhD thesis, Université de Provence, Marseille,
France, 2004.
[MT06] M. Martinez and D. Talay. Discrétisation d'équations diérentielles stochastiques
unidimensionnelles à générateur sous forme divergence avec coecient discontinu. C.R. Acad.
Sci. Paris 342:1 (2006), 51-56.
[RY91] D. Revuz and M. Yor. Continuous Martingale and Brownian Motion. Springer, Heidelberg,
1991. MR1083357
[Str82] D. W. Stroock. Aronson's Estimate for Elliptic Operators in Divergence Form. Lecture Notes
in Mathematics, Seminaire de Probabilités XXII (1321), Springer 1988, 316-347.
[Wal78] J.B. Walsh. A Diusion with a Discontinuous Local Time. Société Mathématique de France,
Astérisque 52-53 (1978) 37-45.

275