You are on page 1of 36

Setting

Zs
s = (5.13)
(s, Xs )

we have that Yt = VtY0 , and then VTY0 , g(XT ). Thus, Y0 p0 .

Now let v s.t. 9 , VTv, g(XT ) = YT . We observe that V v, has the same dynamics as a BSDE and thus we can use

the comparison theorem to conclude that v Y0 . Taking the infinimum on v leads to p0 Y0 .

179
5.2 Main properties in the Markov setting

See [67]

• One can show that Yt = u(t, Xt ) for some Lipschitz function u.

• u is solution (in the viscosity sense) of

u(0) (t, x) f x, u(t, x), @x u(t, x) (t, x) = 0 (t, x) 2 [0, T ) ⇥ Rd (5.14)

and u(T, x) = g(x) . (5.15)

• Assuming that u is a smooth solution to the above non-linear pde, one can prove conversely

(Yt = u(t, Xt ), Zt = @x u(t, Xt ) (t, Xt )). (5.16)

180
Proof. Apply Ito’s Formula. 2

Notations (in the one-dimensional setting) For a function ' : [0, T ] ⇥ R ! R, we define

'(0) (t, x) = L(0) '(t, x) := @t ' + LX '(t, x) (5.17)

'(1) (t, x) = L(1) '(t, x) := (t, x)@x '(t, x) (5.18)

where LX is given in (2.7).

For a multi-index, [n 1 {0, 1}n with = (j1 , . . . , jk ) for some k 1, we denote by := (j2 , . . . , jk ) and

' = L ' = L(j1 ) [L '] .

181
5.3 Numerical analysis of backward Methods

• we seek to approximate (Y, Z) in order to approximate u.

• to alleviate the presentation, we assume that f := f (y, z) only

• We work under two different sets of assumptions

1. (HL): the coefficient b, , f , g are L-Lipschitz continuous.

2. (Hr): the function u, b, are Cb2,4 (up to T ).

182
Backward Algorithm: [17, 77] on a discrete grid ⇡

- at tn = T :

(Yn⇡ , Zn⇡ ) = (g(XT ), 0) (5.19)

- i < n, compute
8
>
> h i
>
> ⇡ Wti+1 Wti
<Zi = Eti Yi+1

( ti+1 ti )
(5.20)
>
> ⇥ ⇡ ⇤
>
>
:Yi⇡ = Eti Yi+1 + (ti+1 ti )f (Yi⇡ , Zi⇡ )

where the grid ⇡ satisfies n|⇡|  C where |⇡| := maxi (ti+1 ti ).

Remark 5.2. .
1
(i) This is an implicit Euler scheme and the error is O(|⇡| 2 ) in the Lipschitz case and assuming smoothness of the

183
coefficient O(|⇡|) at most, generally (see below).

(ii) Explicit version of the scheme: in the smooth case Zn⇡ := (XT )g 0 (XT ).

(iii) One has to estimate the conditional expectations!

heuristics.

184
5.3.1 L2 -stability

• Perturbation approach: Consider

h i
Ỹi = Eti Ỹi+1 + (ti+1 ti )f (Ỹi , Z̃i ) + ⇣i (5.21)

Wt W ti
Z̃i = Eti Ỹi+1 ( i+1 ) (5.22)
ti+1 ti

where ⇣i 2 L2 (Fti ).

• Let us set Y := Y ⇡ Ỹ , Z := Z ⇡ Z̃, we introduce

Definition 5.1 (L2 -stability). The scheme given in (5.20) is L2 -stable if there exists a constant C > 0 s.t.
" n 1
#
⇥ ⇤ X
max E | Yi |2  CE | Yn |2 + n⇣i2 (5.23)
i
i=0

185
for |⇡| small enough, for all perturbation ⇣.

Proposition 5.2. If f is Lipschitz continous, the scheme (5.20) is L2 -stable.

Proof. Let us observe that the scheme can be rewritten

Yi⇡ = Yi+1

+ hi f (Yi⇡ , Zi⇡ ) hi Zi⇡ Hi Mi (5.24)

where Hi = h .
Wi
Note that (5.26) defines Mi , moreover it satisfies

⇥ ⇤
Eti[ Mi ] = Eti[ Mi Hi ] = 0 and E | Mi |2 < 1 (5.25)

These properties are obtained by using the definition of the scheme given in (5.20).

For the perturbed scheme, we have similarly:

Ỹi = Ỹi+1 + hi f (Ỹi , Z̃i ) + ⇣i hi Z̃i Hi M̃i (5.26)

186
Denoting fi = f (Yi⇡ , Zi⇡ ) f (Ỹi , Z̃i ) and Mi = Mi M̃i , we observe that

Y i + h i Z i Hi + Mi = Yi+1 + hi fi + ⇣i . (5.27)

Squaring both sides and taking conditional expectation, we compute, using Young’s inequality,

h i C 2
2
2 2
| Yi | + hi | Zi |  (1 + Ch)Eti ( Yi+1 + hi fi ) + |⇣i | . (5.28)
h

Note that

( Yi+1 + hi fi )2  (| Yi+1 | + Chi | Yi | + Chi | Zi |)2 (5.29)

h ✏
 (1 + )(| Yi+1 | + Chi | Yi |)2 + C(1 + )h2i | Zi |2 (5.30)
✏ h

187
Choosing h and ✏ such that C(h + ✏)  12 , we obtain

1
( Yi+1 + hi fi )2  (1 + Ch)| Yi+1 |2 + Ch| Yi |2 + h2i | Zi |2 (5.31)
2

Inserting the previous inequality in (5.28), we get

⇥ ⇤ C
| Yi |2  (1 + Ch)Eti | Yi+1 |2 + |⇣i |2 (5.32)
h
⇥ ⇤ C
 eCh Eti | Yi+1 |2 + |⇣i |2 (5.33)
h

Taking expectation on both sides and iterating over i, we obtain (5.23).

188
5.4 Convergence analysis assuming no error on X

5.4.1 Truncation error

• Let us introduce

Wi
Ẑti = Eti Yti+1 . (5.34)
hi

• We define the local truncation error as


Z ti+1
⇣ˆi := Eti [f (Yt , Zt ) f (Yti , Ẑti )]dt . (5.35)
ti

• It measures how well the true solution satisfies the scheme.

Indeed, we observe that:

189
• The global truncation error is then defined as

X h i
T (⇡) := ˆ 2
E n|⇣i | . (5.36)
i

• Assume at this point that there is no error made on the forward process, thus Yn := 0 and we have

⇥ ⇤
max E |Yti Yi⇡ |2  CT (⇡) . (5.37)
i

ˆ
Proof. This comes directly from the L2 -stability of the scheme with the perturbation given by ⇣. 2

• we now study the order of the truncation error under (HL) or (Hr).

190
5.4.2 Order of convergence in the smooth case

Theorem 5.3. Under (Hr), we have that

T (|⇡|)  C|⇡|2 (5.38)

and thus the scheme is of order 1.

Proof. (proof in the one dimensional case) 1. We observe that


Z ti+1
|⇣ˆi |  Eti [f (Yt , Zt ) f (Yti , Zti )]dt + Chi |Zti Ẑti |. (5.39)
ti

191
Using the PDE satisfied by u, we get

h i
|Eti[f (Yt , Zt ) (0) (0)
f (Yti , Zti )] | = |Eti u (t, Xt ) u (ti , Xti ) | (5.40)
Z ti+1
= |Eti u(0,0) (t, Xt )dt | (5.41)
ti

 C|⇡| (5.42)

where we used Ito’s formula for the last equality. Now we compute, setting Hi := Wi
hi

 Z ti+1 Z ti+1
⇥ ⇤ 1
Ẑti = Eti u(ti+1 , Xti+1 )Hi = Eti Hi (0)
u (t, Xt )dt + u(1) (t, Xt )dt (5.43)
ti hi ti

192
Observe that
 Z ti+1  Z ti+1
|Eti Hi (0)
u (t, Xt )dt | = |Eti Hi {u(0) (t, Xt ) u(0) (ti , Xti )}dt | (5.44)
ti ti

 C|⇡| (5.45)

and that

h i  Z t
(1) (1)
Eti u (t, Xt ) = Eti u (ti , Xti ) + u(0,1) (s, Xs )ds (5.46)
ti

We thus get
 Z ti+1
1
|Eti u(1) (t, Xt )dt u(1) (ti , Xti )|  C|⇡| (5.47)
hi ti

193
leading to |Zti Ẑti |2  C|⇡|. We then obtain that

|⇣ˆi |  C|⇡|2 . (5.48)

And, by summing this local error estimate, we conclude that T (⇡)  C|⇡|2 . 2

194
5.4.3 Order of convergence in the Lipschitz case

• The Lipschitz case is more involved.

Proposition 5.3. Under (HL), we have that

n 1Z
!
⇥ ⇤ X ti+1 ⇥ ⇤
T (|⇡|)  C |⇡| + max E |Yt Yti | 2
+ E |Zt 2
Z̄ti | dt (5.49)
i ti
i=0

where
Z ti+1
1
Z̄ti = Eti Zt dt .
hi ti

Proof. We first observe that


Z ti+1 Z ti+1
Ẑti = Eti Zs dWs Hi + Hi f (Yt , Zt )dt (5.50)
ti ti

195
and then compute
Z ti+1
Z̄ti = Eti Zs dWs Hi (5.51)
ti
 Z ti+1 Z ti+1
|Eti Hi 2
f (Yt , Zt )dt |  Eti |f (Yt , Zt )|2 dt (5.52)
ti ti

leading to
Z ti+1
|Z̄ti 2
Ẑti |  CEti |f (Yt , Zt )|2 dt (5.53)
ti

196
Then
Z ti+1  Z ti+1
|Eti [f (Yt , Zt ) f (Yti , Ẑti )]dt |2  hi Eti C [|Yt Yti |2 + |Zt Z̄ti |2 + |Z̄ti Ẑti |2 ]dt (5.54)
ti ti
Z ti+1
 hi CEti [|Yt Yti |2 + |Zt Z̄ti |2 ]dt (5.55)
ti
Z ti+1
+ Ch2i Eti |f (Yt , Zt )|2 dt (5.56)
ti

The proof is concluding by summing on i and recalling that


Z T
E |f (Yt , Zt )|2 dt  C. (5.57)
0

197
• We give without proof the following result
Z ti+1 X Z ti+1
max E |Zt |2 dt + E |Zt Z̄ti |2 dt  C|⇡| (5.58)
i ti ti
i

due to Zhang [77].

• It is based on a representation of Z obtained by means of Malliavin calculus and require some heavy computations.

Theorem 5.4. Under (HL), we have that

T (|⇡|)  C|⇡| (5.59)

and thus the scheme is of order 1/2.

198
Proof. Observe that
" # Z Z
ti+1 ti+1
E sup |Yt Yti | 2
 CE 2
|f (Yt , Zt )| dt + |Zt |2 dt (5.60)
t2[ti ,ti+1 ] ti ti

 C|⇡| (5.61)

where we used (5.58). Then, the proof is concluded by combining the previous estimate and (5.58) with Proposition

5.3. 2

199
5.5 Full discrete-time error analysis

• We replace X by its Euler scheme X ⇡ .

• The scheme for (Y, Z) is the same but the terminal condition (5.19) is now

(Yn , Zn ) = (g(XT⇡ ), 0) (5.62)

Theorem 5.5. Under (Hr), the following holds

|Y0 u(0, X0 )|  C|⇡| . (5.63)

200
5.5.1 Truncation error

• Define Ỹi = u(ti , Xt⇡i ) and



Wi
Z̃i = Eti Ỹi+1 .
ti+1 ti

Proposition 5.4. With the above definition,

h i
Ỹi = Eti Ỹi+1 + (ti+1 ti )f (Ỹi , Z̃i ) + ⇣i (5.64)

where

h i
⇣i = Eti ⇣ie + ⇣if + ⇣iz (5.65)

and ⇣ie , ⇣if and ⇣iz are defined in (5.67), (5.69) and (5.70) respectively.

201
Proof. (b = 0) 1. Applying Ito’s formula, we have, setting Vsi = (Xt⇡i )@x u(s, Xs⇡ )
Z ti+1 Z ti+1
1
u(ti+1 , Xt⇡i+1 ) = u(ti , Xt⇡i ) + {@t u(s, Xs⇡ ) + 2
(Xt⇡i )@xx
2
u(s, Xs⇡ )}ds + Vsi dWs (5.66)
ti 2 ti

Then, introducing
Z ti+1
1
⇣ie := 2
(Xt⇡i ) 2
(Xs⇡ ) @xx
2
u(s, Xs⇡ )ds (5.67)
2 ti

the above equality rewrites, using the PDE satisfied by u,


Z ti+1 ⇣ ⌘ Z ti+1
Ỹi+1 = Ỹi hf (Ỹi , Vtii ) + u (0)
(s, Xs⇡ ) u (0)
(ti , Xt⇡i ) ds + ⇣ie + Vsi dWs (5.68)
ti ti

202
Now we define
Z ti+1 ⇣ ⌘
⇣if = u (0)
(s, Xs⇡ ) u (0)
(ti , Xt⇡i ) ds (5.69)
ti

⇣iz = h{f (Ỹi , Z̃i ) f (Ỹi , Vtii )} (5.70)

and observe that

h i
Ỹi = Eti Ỹi+1 + hf (Ỹi , Z̃tii ) + ⇣ie + ⇣if + ⇣iz . (5.71)

5.5.2 Convergence analysis

• We prove below Theorem 5.5: this relies (as usual) on

203
1. the stability result for the BTZ scheme

2. a control of the truncation error coming from the regularity of u.

Lemma 5.1. Under (Hr), the following holds

|Eti[⇣ie ] |2 = O(h4 ) (5.72)


h i
|Eti ⇣if |2 = O(h4 ) (5.73)

|Eti[⇣iz ] |2 = O(h4 ) (5.74)

Proof. See Exercise ?? 2

Proof of Theorem 5.5

We simply observe that (Ỹi , Z̃i )i defines a pertubed scheme as in Section 5.3.1. We can then combine Proposition 5.2

204
with Lemma 5.1 to conclude the proof, recall also (5.37).

205
5.6 Numerical illustration and further consideration

5.6.1 Monte Carlo based methods

• Regression methods, as the one used for US options, can be easily adapted to the BSDEs setting, introduced in

[37, 57] and extensively studied in [39, 40, 38, 34]

• Any methods that permits to estimate conditional expectation can be used, e.g. Malliavin method [29, 17, 16]

5.6.2 Tree based methods

• We illustrate Section 5.5 with Example 5.1.

• A more systematic approach is based on Cubature methods (it allows in particular to work with degenerate

206
diffusion), see e.g. [27, 28, 26].

Example 5.1. In this example [22], we work with d = 3 and X = W , We consider the following BSDE
Z 1 Z 1
!(1, W1 ) 5
Yt = + (Zs1 + Zs2 + Zs3 )( Ys )ds Zs dWs . (5.75)
1 + !(1, W1 ) t 6 t

where !(t, x) = exp (x1 + x2 + x3 + t). Applying Itô formula, we verify that the solution is given by

!(t, Wt ) !(t, Wt )
Yt = and Ztl = , l 2 {1, 2, 3} , t 2 [0, 1] . (5.76)
1 + !(t, Wt ) (1 + !(t, Wt ))2

• In this very smooth setting, we can introduce higher order scheme too. We consider the Crank-Nicholson scheme

(CN) and a linear multi-step scheme (AMB2) on the graph below.

• The number of points used to approximate W depends on the theoretical rate of convergence of the discrete-time

error and the number of moment to match, recall Section 2.4.

207
• Since X = W , we can use easily a recombination procedure.

Figure 5: Empirical convergence, Example 5.1

208
5.6.3 (Markovian) quantisation

• We illustrate here the Markovian quantisation with Example [23] in the case of quadratic BSDEs.

• "Classical" quantization can be used for BSDEs as well, see e.g. [65].

Example 5.2. [23] We consider the following quadratic Markovian BSDE:


8
>
> R
< Xt` = X0` + 0t ⌫Xs` dWs` , ` 2 {1, 2, 3}
, 0 6 t 6 1, (5.77)
>
> R1 a R 1
: Yt = g(X1 ) + 2
t 2 kZs k ds t Zs dWs

where a, ⌫, and (X0` )`2{1,2,3} are given real positive parameters and g : Rd ! R is a bounded Lipschitz function.

Applying Ito’s formula, one can show that the solution is given by

1 ⇣ ⇥ ⇤⌘
Yt = log Et exp ag(X1 ) , t61. (5.78)
a

209
For any given g, ⌫ and a, it is possible to estimate the solution Y0 at time 0 using an approximation of the Gaussian
⌫2 `
distribution at time T = 1, since X1` = X0` e 2 +⌫W1 .

• For our numerical illustration, d = 2 and g is given by

2
X
g : x 7! 3 sin2 (x` ) , (5.79)
`=1

• We use a markovian quantisation scheme as introduced in Section 4.5.3.

• The non-Lipschitz setting may cause instability and we use in the graph below a scheme taming this quadratic

growth also (called ’adaptive truncation’)

210
Figure 6: Comparison of schemes’ convergence

211
5.7 An example of forward method

• We present here a machine learning method to approximate BSDEs.

• Consider, for y 2 R and Z 2 H2


Z t Z t
Yty,Z =y f (Ysy,Z , Zs )ds + Zs dWs (5.80)
0 0

• and the following optimisation problem

h i
V := min E |g(XT ) YTy,Z |2 (5.81)
(y,Z)2R⇥H2

• In the Lipschitz setting, the solution of the above optimisation is the BSDEs (5.1) with V = 0!

• Main idea: solve numerically the optimisation problem (5.81) to get an approximation of the BSDE.

212
5.7.1 Discretization of the optimisation problem (5.81)

• The dynamics (5.80) of Y y,Z is discretised using a classical Euler scheme on ⇡:

Yt⇡n+1 = Yt⇡n + hf (Yt⇡n , Ztn ) + Ztn (Wtn+1 Wtn ) and Y0⇡ = y . (5.82)

• The random variable Z must be discretised also in some sense: parametric specification!

1. Non-linear specification: Z = Z ⇥ for some ⇥ 2 RK where ⇥ stands for the coefficient of a Neural Network

'NN and Zt⇥n = 'NN (tn , Xtn )

2. Linear specification: Zt✓n = 'L (tn , Xtn ) where

K
X
'L (tn , ·) = ✓k k (tn , ·) , ✓ 2 RdK
k=1

213
where ( k )1kK are some basis functions.

• Often, in practice, X has to be approximated too, by an Euler scheme for example, recall (2.14).

• The discrete optimisation problem is now given by

⇥ ⇤
V ⇡ := min E |g(XT⇡ ) YT⇡, |2 (5.83)
2R1+K̄

where = (y, ⇥0 , . . . , ⇥N 1 ) and K̄ = d + (N 1)K for the non-linear specification, or = (y, ✓0 , . . . , ✓N 1 )

and K̄ = d + (N 1)dK for the linear specification.

• To solve the discrete optimisation problem one uses generally a (stochastic) gradient descent algorithm.

214

You might also like