Backward Stochastic Differential Equations

and Applications

ETIENNE PARDOUX

Institut Universitaire de France

and
Labo. d'Analyse, Topologie, Probabilités
Centre de Mathématiques et d'Informatique
Université de Provence
Rue Joliot Curie, 13453 Marseille Cedex 13
France

Introduction
A new type of stochastic differential equation, called the backward stochastic dif-
ferentil equation (BSDE), where the value of the solution is prescribed at the final
(rather than the initial) point of the time interval, but the solution is nevertheless
required to be at each time a function of the past of the underlying Brownian mo-
tion, has been introduced recently, independently by Peng and the author in [16],
and by Dufne and Epstein in [7]. This class of equations is a natural nonlinear ex-
tension of linear equations that appear both as the equation for the adjoint process
in the maximum principle for optimal stochastic control (see [2]), and as a basic
model for asset pricing in financial mathematics. It was soon after discovered (see
[22], [17]) that those BSDEs provide probabilistic formulas for solutions of certain
semilinear partial differential equations (PDEs), which generalize the well-known
Feynmann-Kac formula for second order linear PDEs. This provides a new addi-
tional tool for analyzing solutions of certain PDEs, for instance reaction-diffusion
equations.

1 Backward stochastic differential equations

Let {Bt; 0 < t < T} denote a d-dimensional Brownian motion defined on a
probability space (Q,^, P). For 0 < t < T, we denote by Tt the a-algebra
o{Bs\ 0 < s < t}, augmented with the P-null sets of T. We are given two objects:
a final condition £ G L2(Q, TT, P: Rk);
a coefficient f, which is a mapping from Q x [0, T] x Rk x Rkxd into Rk, and
is such that there exists K > 0 with:

f(-,y,z) is progressively measurable, 1 My G Rk,z G Rkxd: (i)

1) That is for each t, the restriction to [0, t] x f2 of the mapping (S,CJ) —• f(s,uj,y,,z) is
13(0, t) 0 Tt measurable.

Proceedings of the International Congress

of Mathematicians, Zürich, Switzerland 1994
© Birkhäuser Verlag, Basel, Switzerland 1995
 Backward Stochastic Differential Equations 1503

< j l/(*,0,0)|2 dt < oo ; (ii)

Jo
\f(t,y,z) - f(t,y',z')\ < K(\y - y'\ + \z- z'\), Vt,y,y',z,z'. (iii)
A solution of the BSDE(£, / ) is a pair {(Yt,Zt), 0 < t < T} of progressively
measurable processes with values in Rk x Rkxd such that

E f |Zt|2d*<oo (j)

Y
t =t +J f(s,Ys,Zs)ds- J ZsdBs,0<t<T. (jj)

We have the following.

THEOREM 1.1. [16] Under the above conditions, in particular (i), (ii), and (iii),
the BSDEfë, f) has a unique solution (Yt,Zt), 0 <t <T.

REMARK 1.2. The constraint that the solution be progressively measurable (i.e.
adapted to the past of B at each time t) is in a sense rather unnatural for the
solution of a backward equation. This is the reason why we need to have the freedom
of choosing Z independently ofY.

REMARK 1.3. What makes the solution random is the randomness of Ç and f.
The stochastic integral is there to make Y progressively measurable, which in fact
means "reducing the randomness of Y ". In particular, if for some stopping time
r < T, Ç and f are TT measurable, then on the interval [r, T], Z = 0 and Y is
given by the solution of the ODE

dYt/dt = -f(t,Yt,0), YT = £.

REMARK 1.4. Using Ito calculus, it follows from the square integrability of'£, (ii),
(iii), (j), and (jj) that if(Y,Z) solves the BSDE(£,f), E ^ s u p ^ ^ \Yt\2] < oo. It is
easy to check that a solution of the BSDE(£,0) is given as follows. Yt = E[£f!Ft\,
and Z is given by Itô's representation theorem of functionals of Brownian motion,
which says that £ = E[Ç] +JQ Zt dBt, for a certain progressively measurable process
Z satisfying (j). Note that if we require only JQ \Zt\2 dt < oo, then the uniqueness
of Z is not guaranteed.

2 Applications
Before giving some indication for the proof of Theorem 1.1, let us motivate that
notion by presenting several applications. Later we shall present our main appli-
cation, which is to semilinear PDEs.
2.1 Application in financial mathematics. Consider a typical model for continuous
time asset pricing. Let Vt denote the total wealth of an agent at time t, which he
can invest in n + 1 different assets, one nonrisky asset, whose price per unit Pt° is
1504 Etienne Pardoux

governed by the linear ordinary differential equation (ODE) dP®/dt = P^rt, and
n risky assets, where the price process for one share of the zth stock is governed by
the linear stochastic differential equation (SDE) dP\ = P\\p\ dt + Y?j=i G%t dB{\.
The asset pricing problem is as follows. Given a contingent claim £, which is an
^-measurable random variable that we suppose to be square integrable, find an
initial wealth Vo and a portfolio (U.\, 0 < t < T, 1 < i < n) such that the wealth at
time T is exactly £. Hence, we need to solve the following linear BSDE (1 denotes
the d-dimensional vector whose coordinates are all equal to 1):

Vt=Ç- j rs\Vs-lTsl}ds- j ni[psds + osdBs}.

This linear BSDE is a very classical model in financial mathematics. It is

in particular the starting point of the celebrated Black-Scholes formula for option
pricing. No general theory is necessary to study such a linear equation. However,
there is at least one unreasonable assumption in our model: Vt — U^l represents
an amount of money that is deposited in the bank whenever it is positive, but it
represents an amount of money that is borrowed from the bank if it is negative.
As the interest rate for borrowing is in fact bigger than the bond rate, we should
rather write the above equation as a nonlinear BSDE, with some interest rate
process Rt > rt

Vt=i- j rs[Vs-Il*sl}+ds+ [ Rs[Vs-U;i]-ds- [ U;[psds + osdBs}.

Jt Jt Jo
Note that this last BSDE is of the type considered in Section 1, in the par-
ticular case k = 1. There are several other reasons for using nonlinear BSDEs as
models in financial mathematics, including taking into account technology con-
straints, as well as the notion of recursive utility (see [7], [8], [9], [10], [11] and the
bibliographies therein).
2.2 Application in stochastic control (see [11], [22]). Suppose now that k = 1, and
the coefficient / of our BSDE is concave in the variables y and z. We define the
following upolar" process:

F(t, 3,7) := sup [f(t, y, z) - ßy - 7 - 3 ] .

It follows from a measurable selection theorem that to each progressively measur-

able process (Yt,Zt), one can associate a progressively measurable pair (/?*,7t*)
such that
F(L ßlril) = f(t, Yt, Zt) - ß\Yt -il-Zt,0<t< T.
Let A denote the set of progressively measurable "control" processes (ßt,Jt) that
satisfy E JQ F(t, ßt,lt)2 dt < 00. Consider for each t > 0 the scalar forward linear
SDE
rf;; = 1 + / T^[ßrdr + 7r dB r \, s>t.
We then have the following.
 Backward Stochastic Differential Equations 1505

THEOREM 2.1. [11] Let (Yt,Zt) be the unique solution of the BSDE(£,f). Then
for each 0 <t <T, Yt is the value function of a stochastic control problem, in the
sense that
rT
Yt = sup E J T^F(s,ßsns)ds + T^
(Pn)eA

2.3 Application in stochastic geometry (see [6]). One can show that the construc-
tion of a gamma-martingale (which is a notion of martingale adapted to processes
with values in a manifold equipped with a connection T) with prescribed final
value £ can be achieved by solving a backward SDE where the coefficient / takes
the form

j,k,q

One can assume that T is bounded and Lipschitz, however / here is not Lipschitz
in z, hence Theorem 1.1 does not apply directly. However, combining BSDE and
gamma-martingale techniques, one can show existence and uniqueness of a solution
in this case.

3 Proof of Theorem 1.1

We now indicate a proof of our basic Theorem 1.1. The notation is as in Section
1. Let B2 denote the set of pairs {(Yt, Zt), 0 < t < T} of progressively measurable
processes with values in Rk x Rkxd satisfying (j) and E JQ \Yt\2 dt < co.
We define a mapping $ from B2 into itself as follows. Given (U, V) G B2, let
(Y, Z) = $(U, V) be defined by: Yt = E[Ç + j f f(s, US.VS) ds\Ft], where Z is the
process given by Itô's martingale representation theorem applied to the martingale
Mt = E[Ç + J0T f(s, US.VS) ds\Ft}. We then have that

Yt=t-r f f(s,Us,Vs)ds- f ZsdBs,0<t<T.

Define the following norm on B2, for r > 0:

\\(Y,Z)\\r=(EJ\rt[\Yt\'2 + \Zt\

One can show by Ito calculus that, if r is large enough, the mapping $ is a strict
contraction on B2 equipped with the norm || • || r , hence it has a unique fixed point,
which means that the BSDE(£, / ) has a unique solution.
REMARK 3.1. Note that one has a similar result to Theorem 1.1, if one assumes
£ G Lp(Q,J7,P;Rk) for some p > 1, instead of p = 2. There is apparently no
general theory for BSDEs with locally Lipschitz coefficient f. However, in addition
to the result of [6], one can find some results in that direction in [19].
1506 Etienne Pardoux

One important tool for analyzing BSDEs is the following comparison theorem.
THEOREM 3.2. Suppose k = 1, and let (£, / ) and (£', / ' ) be two pairs of data that
satisfy the assumptions of Section 1. Suppose in addition that
£ < £' a.s. and f(t, y, z) < f'(t, y, z) V(t, y, z) G [0, T] x R x Rd and a.s.,
then Yt <Yj,0<t< T, a.s.

4 BSDEs and viscosity solutions of second order semilinear PDEs

We now describe the relation between BSDEs and systems of second order semilin-
ear PDEs. It turns out that solutions of BSDEs are naturally related with viscosity
solutions of PDEs. This approach allows us to minimize the regularity requirements
on the coefficients, while we will have to restrict ourselves to the case where the
ith component of / depends on the ith line of the matrix Z only.
Before introducing the system of PDEs, we need to put the BSDE in a Marko-
vian framework, i.e. to let £ and / be functions of the Brownian motion B through
a Markov-diffusion process, solution of a forward SDE driven by B.
Let b : Rd -> Rd and o : Rd -> Rdxd be Lipschitz functions, and for any
(t,x) G [0,T] x Rd, let {Xl'x: t<s<T} denote the solution of the forward SDE

Xl-X =x-r f b(X^x)dr + f o(Xtrx)dBr, (4.1)

and consider the backward SDE

rt.X g(Xlix) + f /(r, x£ x , Yrtx, Z*x) dr - f Z^x dBr, t<s<T, (4.2)

where g and f map respectively Md and [0,T] xRd x Kfc x R fexd into Rk, g and
f(t,-,y,z) are jointly continuous and there exist constants K,p such that for all
t e [0,T], x € R d , y,y' e Mfc, z,z' e Rkxd :
\g(x)\ < K{1 + \x\"), \f(tx,y,z)\ < K{1 + \x\' + \y\ + \z\),
\f(t,x,y,z) - f{Lx,y',z')\ < K(\y -y'\ + \z - z'\).
We now associate to (4.1) and (4.2) the following system of parabolic second
order semilinear PDEs:

J -^(t.x) +Lu(t,x) -r f(tx.u(Ux),Vua(t,x)) =0. (tx) G [0,T] xRd. oX

< ot d
(4.3)
u(T,x) =g(x), xeR ,
where
' Lu\ ' 1 d2 b_d_
Lu= \ : J, L = -((T**)^l,J ^ ^ -r k%
2 dxidxj dxi
K Luk
I
(with the convention of summation over repeated indices).
Let us first recall the notion of viscosity solutions of the system of PDEs (4.3)
(see [5], [14]).
 Backward Stochastic Differential Equations 1507

DEFINITION 4.1. u G C([0, T] x Rd; Rk) is called a viscosity sub-solution of (4.3)

whenever ui(T,x) < gi(x), 1 < i < k, x G Rd, and for each 1 < i < k, (t,x) G
(0,T) x Rd, ip G Cli2((0,T) x Rd) such that (tx) is a local maximum of Ui - ip,

--^-(t. x) - L(p(t.x) - fi(Ux,u(t.x), V<po(tx)) < 0.

ot
u G C([0,T] x Rrf;3Rfc) is called a viscosity super-solution of (4.3) whenever
Ui(T,x) > Qi(x), 1 < i < k, x G Rd, and for each 1 < i < k, (tx) G [0,T] x Rd,
ip G C 1 , 2 ((0,T) x Rd) such that (t,x) is a local minimum of U{ — tp,

--^(tx) -Lip(t,x) - fi(t<x<u(t,x)^ip<j(t,x)) >0.

ot
u G C([0,T] x Ed;]Rfe) is called a viscosity solution of (4.3) if it is both a
viscosity sub- and super-solution of (4.3).

We can now state the main result of this section. (Cp([0,T] x Md;Rfc) denotes
the set of continuous functions from [0,T] x Rd into Rk, which grow at most
polynomialy at infinity).
THEOREM 4.2. u(t,x) := Y^x belongs to Cp([0,T] x R d ;R fc ) and is the unique
viscosity solution of (4.3).

Proof. Uniqueness is proved by methods from viscosity solutions, see [5]. In order
to prove that {Y*'x} is a sub-solution we assume that 1 < i < k, (t, x) G [0, T] x Rd,
if G C 1,2 ((0, T) x Rd) are such that Ui(t,x) = y(t,x) and Ui(s,y) < y(s,y),
(s,y) G [0,T] x Rd. We first note that from uniqueness of the solution of the
BSDE, Y^h = u(t + h,Xl,xh). Hence, deleting the superscripts t, x for notational
simplicity,
t+h çt+h
/
fi(Xs,Ys,Zl)ds- I Z\dBs.
Let (Ys, Zs, t < s < t + h) be the solution of the BSDE
t+h _ _ ft+h
/
fi(Xr,(Yr,YÏ),Zr)dr-
/ ZrdBr,
Js
where (Yr,Y*) denotes the vector whose ith component equals Yr, and the oth-
ers equal the corresponding components of the vector Yr. Prom the compar-
ison theorem, ui(t,x) < Yt. Applying Itô's formula to tp(s,Xs) and defining
Ys =YS — ip(s,Xs), Zs = Zs — Vipo(s,Xs), we have that
0<Yt=EJ [(^ + L<p)(s,Xs)
+ h{Xs,(Ys + <p(s,Xs),Yls),Zs + Vcp<r(s,Xs))] ds.

It remains essentially to divide by h and let h —> 0.

1508 Etienne Pardoux

REMARK 4.3. Suppose that fc = 1 and f(t,x,y,z) = c(t,x)y. Then by the varia-
tion of constants formula,

y/- x =g{XY) cxp[j c(s, Xi1) ds] - j cxp[J" c(r, X**) &]&<*, dB„)

=EYtUx

=E L(XÏx)Cxp[j\(S,Xl-xds}\ ,

which is the well-known Feynman-Kac formula.

Let us now indicate some results on reaction-diffusion equations that have

been obtained recently, with the help of the above stochastic representation, fol-
lowing [23]. Consider a reaction-diffusion equation of the type

du d u

where for instance / is of the "KPP type", f(r) = r(l — r). Suppose the initial
condition is of the form u 0 (x) = lx_(:r). The function us(t,x) = u(t/e,x/e) be-
haves for small e as v(x~cat), where v decreases from 1 to 0 on R. Moreover a,
the speed of the front, can be computed in terms of the derivative of / at 0+.
For proving this type of result, Freidlin [13] uses the Feynman-Kac formula (which
is implicit here, because of the nonlinearity of / ) and results from the theory of
large deviations. Using BSDEs, one can write an explicit probabilistic formula for
the solution of the PDE, even when the nonlinear term / depends on the first
derivative of u, and obtain asymptotic results on u£ in that more general set-up.

5 Extensions
We now indicate several results that have been obtained recently and generalize
the results presented above.
5.1 BSDEs with respect to Brownian motion and Poisson random measure. It is
possible to solve BSDEs with the Brownian motion replaced by a general martin-
gale, see [3]. One can also consider a diffusion process with jumps X. solution of a
forward SDE driven by both a Brownian motion and a Poisson random measure,
and consider a backward SDE whose final condition and coefficient are functions
of X. This provides a stochastic formula for a semilinear integro-partial differential
equation [1], [4]. One can also use these results for giving a probabilistic formula
for the solution of a system of parabolic PDEs, where the second order PDE op-
erator is different from one line to the other, i.e. with the notation of the previous
section, hu = (L\Ui,... .L^uu)', see [20].
5.2 Coupled forward-backward SDEs [15]. In the last section, the diffusion process
X was perturbing the coefficients of the BSDE for (Y,Z). Suppose now that we
can solve a pair of forward-backward SDEs, where (Y, Z) appear in the coefficients
 Backward Stochastic Differential Equations 1509

of the forward SDE for X, and X appears in the coefficients of the backward
SDE for (Y,Z). Then u(t,x) := Y*,x would solve a general type of quasilinear
PDE. However, it seems so far that in this case the pair of SDEs can be solved
directly only in very restricted cases, and that results on the corresponding PDE
are necessary to solve the pair of SDEs under rather general assumptions.
5.3 Equations with boundary conditions. So far we have considered only PDEs
in the whole space Rd. It is possible to consider PDEs with Dirichlet or Neumann
boundary conditions at the boundary of a domain D cRd, provided one replaces
the diffusion X by a diffusion cither stopped at the boundary of D, or reflected at
the same boundary.
5.4 Infinite-time horizon BSDEs, and elliptic PDEs. If one replaces the final time
T by -hoc, or by a stopping time T, it is possible to give probabilistic formulas for
quasilinear elliptic PDEs. Of course, more restrictive assumptions on the coefficient
/ are then required. For a result on infinite-time horizon BSDEs, see [10].
5.5 Reflected BSDEs. The following reflected BSDE has been studied in [12].

Yt=t+ I f(t,Xt,YuZt)dt + KT-Kt- j ZsdBs,0<t<T

Yt>St,0<t<T] [ (Yt - St)dKt = 0,

Jo

where £ and / are as in Section 1, and {St, 0 < t < T} is a continuous process
satisfying E(supt |St| 2 ) < oc and ST < £ a.s. The solution is a triple (Y,Z,K) of
progressively measurable processes, where K is continuous and increasing. Note
that, unlike the case of reflected forward SDEs, whenever S is a semimartingale,
the increasing process K is absolutely continuous.
In the case where the data (£, / , S) is a given function of a diffusion X, we
get the probabilistic interpretation of an obstacle problem for a quasilinear PDE.
5.6 Backward doubly stochastic differential equations and stochastic PDEs [18].
If we introduce another independent Brownian motion {Wt}, the equation

Yt=£-r [ f(s,Ys,Zs)ds-r i g(s,Ys,Zs)dWs- f ZsdBs

Jt Jt Jt

has a unique solution {(Yt, Zt); 0 < t < T} which is adapted at each time t to the
sup of the past of B and the future increments of W, provided both / and g are
Lipschitz. When put in a Markovian framework, we obtain a formula for a system
of quasilinear SPDEs driven by {W*}.
1510 Etienne Par doux

References
[1] G. Barles, R. Buckdahn, and E. Pardoux, BSDEs and integro-partial differential
equations, preprint.
[2] J. M. Bismut, Théorie probabiliste du contrôle des diffusions, Mem. Amer. Math.
Soc. 176 (1973).
[3] R. Buckdahn, Backward stochastic differential equations driven by a martingale,
LATP prépublication 93-5.
[4] R. Buckdahn and E. Pardoux, BSDEs with jumps and associated integro-partial
differential equations, Proc. Conf. Metz 1994, to appear.
[5] M. Crandall, H. Ishii, and P. L. Lions, User's guide to viscosity solutions of second
order partial differential equations, Bull. Amer. Math. Soc. (New Ser.) 27 (1992),
1-67.
[6] R. Darling, Constructing Gamma-martingales with prescribed limit, using backward
SDE, LATP prépublication 94-2.
[7] D. Duffie and L. Epstein, Stochastic differential utility, Econometrica 60 (1992),
353-394.
[8] D. Duffie and L. Epstein, Asset pricing with stochastic differential utility, Rev.
Financial Stud. 5 (1992), 411-436.
[9] D. Duffie and P. L. Lions, PDE solutions of stochastic differential utility, J. Math.
Econom. 21 (1992).
[10] D. Duffie, J. Ma, and J. Yong, Black's consol rate conjecture, IMA preprint series
# 1164 (1993).
[11] N. El Karoui, S. Peng, and M. C. Quenez, Backward stochastic differential equations
in finance, preprint.
[12] N. El Karoui, C. Kapoudjian, E. Pardoux, S. Peng, and M. C. Quenez, Reflecting
BSDE and associated obstacle problem for PDEs, preprint.
[13] M. Freidlin, Semi-linear PDEs and limit theorems for large deviations, in: Ecole
d'été de probabilités de St Flour 1990, Lecture Notes in Mathematics 1527 (1992),
1-109, Springer-Verlag, Berlin and New York.
[14] H. Ishii and S. Koike, Viscosity solutions for monotone systems of second order
elliptic PDEs, Comm. Partial Differential Equations 16 (1991), 1095-1128.
[15] J. Ma, P. Protter, and J. Yong, Solving forward-backward stochastic differential
equations explicitly — A four-step scheme, preprint.
[16] E. Pardoux and S. Peng, Adapted solutions of backward stochastic equations, Sys-
tems Control Lett. 14 (1990), 55-61.
[17] E. Pardoux and S. Peng, Backward stochastic differential equations and quasilinear
parabolic partial differential equations, in: Stochastic Partial Differential Equations
and Their Applications (B. L. Rozovskii and R. S. Sowers, eds.), LNCIS 176 (1992),
200-217, Springer-Verlag, Berlin and New York.
[18] E. Pardoux and S. Peng, Backward doubly stochastic differential equations and
systems of quasilinear SPDEs, Probab. Theory Related Fields 98 (1994), 209-227.
[19] E. Pardoux and S. Peng, Some backward SDEs with non-Lipschitz coefficients,
Proc. Conf. Metz 1994, to appear.
[20] E. Pardoux, F. Pradeilles, and Z. Rao, BSDEs and viscosity solutions of systems of
second order parabolic PDEs, preprint.
[21] S. Peng, Probabilistic interpretation for systems of quasilinear parabolic partial dif-
ferential equations, Stochastics and Stochastics Reports 37 (1991), 61-74.
[22] S. Peng, Backward stochastic differential equations and applications to optimal con-
trol, Appi. Math. Optim. 27 (1993), 125-144.
[23] F. Pradeilles, Equations différentielles stochastiques rétrogrades et équations de
réaction-diffusion, in preparation.