On a nonlinear partial integro-differential equation

Frdric Abergel, Rmi Tachet November 17, 2009 Ecole Centrale Paris, 92295 Chtenay-Malabry, France Electronic addresses: frederic.abergel@ecp.fr, remi.tachet@ecp.fr

Introduction
Financial modelling has been an area of extremely rapid growth in the past 30 years, and some extremely interesting mathematical challenges have emerged. One of the utmost importance for real-life applications to derivatives trading is that of calibration. Similar to common situations in many areas of physics and engineering, once a model has been suggested, its parameters have to be estimated using external data. In the case of derivative modelling, those data are the liquid (tradable) options, generally known as the vanilla products. It is well known since the pioneering work of Litzenberger and Breeden [1] and its celebrated extension by Bruno Dupire [3] that the knowledge of market data such as the prices of vanilla options across all strikes and maturities is equivalent to the knowledge of the risk-neutral marginals of the underlying stock distribution, and moreover, that there is a unique one-dimensional driftless diffusion which recovers exactly such marginals. However, it has also been well-known for almost as many years that the evolution in time of the so-called local volatility is not stable, thereby leading researchers and nancial engineers to look for a more robust, stochastic volatility type of modelling. In this paper, we consider the calibration problem for a generic stochastic volatility model: more precisely, we address the issue of calibrating to market data a generic model with a stochastic component and a local component for the volatility process. Such models are very useful in practice, since they offer both the exibility and realistic dynamics of stochastic volatility models, and the exact calibration properties of local volatility models. In mathematical terms, the problem we consider is a non linear partial integrodifferential equation for which we are able to prove short-time existence of classical solutions under suitable assumptions. The paper is organized as follows: Section 1 is devoted to the mathematical formulation of the problem. Section 2, to notations and statement of the main result. In Section 3, we recall some important technical results stemming from the general theory of parabolic PDEs. Section 4 contains the proof of the main result. Finally, Section 5 is a short conclusion.

The Local and Stochastic Volatility model and its calibration

The LSV model is an extension of the Dupire local volatility model. In the simplest situation - the two-dimensional case - the dynamics of the model are given by the following system of SDEs dSt = a(t , St )b(Yt )dBt1 + t dt St dYt = (t , Yt )dBt2 + t dt
1

Electronic copy available at: http://ssrn.com/abstract=1508490

Here, (St , t 0) is the stock price process and (Yt , t 0) the stochastic component of the volatility. The function b simply transforms that factor into a proper volatility. a is the local volatility part of the model, choosing its value properly will enable us to calibrate the vanillas of the model. is the volatility of the volatility factor and and are drift terms that may depend on the state variables and on time. B1 and B2 are standard brownian motions with correlation . In order to t the vanillas of this model, we write the Kolmogorov forward equation on the joint density p(t , S, y) of the couple (St , Yt ) 2 1 2 2 2 2 2 1 2 p 2 ( a b S p) (abSp) 2 ( p)) + ( p) + (rSp) + rp = 0 t S 2 S y y 2 y S p(S, y, 0) = (S = S0 , y = y0 ) with (S0 , y0 ) the initial conditions. Taking q = pdy the marginal of S, we get the equation q 2 1 2 ( a2 S2 ( b2 pdy)) + (rSq) + rq = 0 t S 2 S Using Dupires results from [3], we know that q has to solve the following equation in order to t perfectly the vanillas of the market 2 1 q 2 2 ( 2 (rSq) + rq = 0 D S q) + t S 2 S where D is Dupires local volatility and contains the information about the vanillas we want to reproduce. We identify the terms in this last formula. This gives us the value of a2 (t , S) = pdy q 2 2 D (t , S) b2 pdy = D (t , S) b2 pdy . Eventually, the joint density that calibrates the smile of our model is solution of the nonlinear partial integro-differential equation 2 1 2 2 1 2 2 2 pdy 2 p pdy 1 2 ( b S p) (D bS( 2 ) p) 2 ( p)) t S2 2 D b2 pdy S y b pdy y 2 + ( p) + (rSp) + rp = 0 y S The rest of this paper is devoted to the study of a more general n-dimensional version of this equation.

Generalized equation and notations

Throughout this article, we denote by 0 < t T the time-variable and by x = (x1 , x2 , ..., xn ) Rn the n-dimensional space variable where is an open subset with a sufciently smooth boundary (we will precise this notion later). When we consider the equation from a nancial point of view, the rst-variable x1 stands for the spot and the last n 1 for the volatility. Hence, we write S = x1 , y = (x2 , ..., xn ). We also denote by DT =]0, T [ the domain of denition and by B = {0} , BT = {T } and CT =]0, T [ the different parts of the boundary. Given the particular part played by the spot, we consider S = {S R/y Rn1 , (S, y) } and n1 /(S, y) }. We are interested in the following equation: S S , S y = {y R
n p 2 2 2 2 (11 1 I ( p) p) (1i 1 i O( p) := t S i=2 Sxi n

I ( p) p) (1)

n 2 (i j i j p) + (i p) + p = 0 on DT BT i, j=2 xi x j i=1 xi

Electronic copy available at: http://ssrn.com/abstract=1508490

where (i j )1i, jn is a correlation matrix ie is symetric positive denite and veries: ii = 1 2 for all 1 1 T i and 2 < i j < 2 for i = j. We add the boundary condition p = on B C with constant on CT . We also let p0 denote the function p0 (t , S, y) = (S, y). The complexity of this equation stems from the following integral term: I ( p)(t , S) =
S y

p(t , S, x2 , ..., xn )dx2 ...dxn

b S y

2 (x , ..., x ) p(t , S, x , ..., x )dx ...dx n n n 2 2 2

p(t , S, y)dy S y b2 (y) p(t , S, y)dy S y

(2)

(1) belongs to the class of nonlinear, parabolic and nonlocal equations. An interesting reference concerning that kind of equations is [7]. However, our case doesnt fall under the scope of that paper: the operator I (.) is not dened on Cb (DT ). Let us now make a few remarks about our particular equation: 1. in the case of an equation with no I ( p) term, it becomes a classic linear equation of parabolic type. That kind of equation has been properly solved for quite some time now, see [4] or [5]. 2. when b is constant, the problem is reduced to the previous remark. This observation is the key to our resolution method. First, we suppose that b does not vary too much and approximate the nonlocal term I ( p) with a suitable constant. We then isolate the error made during this process in the second term and use a xed point method to solve the new equation. 3. in order to use this method and the results from [4], one has to assume that the coefcients of the equation belong to Hlder spaces H k,h,h/2 (we shall dene them in the preliminaries). 4. the question whether I is properly dened is natural. To answer it, we have to prove that b2 (y) p(t , S, y)dy is bounded away from 0. To do it, we assume that b is non-negative and S y that the initial condition is strictly positive. By restricting ourselves to short times, we are sure that p is not too far from its initial condition and thus is strictly positive. 5. from a nancial viewpoint, it is natural to consider a domain cylindrical with respect to the spot. However, since it may become very challenging to study a PDE on a domain with corners, we reduce our study to domains with a S-section depending on S. The theorem we will prove requires the following assumptions on b and . (H1) b C1 (Rn1 ), (1 , 2 ) R2 , 0 < 1 b 2 on Rn1
b (H2) 2 i n, | xi | b on y where b is a constant we will choose later
2

(H3) is strictly positive and in H 2,h,h/2 . This gives us two results on p0 . First, p0 belongs to H 2,h,h/2 (DT ) and second 0 < p0 = in f p0 sup p0 = p0 (H4) O() = 0 on B in a sense described in the preliminaries Under the previous assumptions, we have the following result: Theorem 1. If the i belong to H 2,h,h/2 (DT ), are positive and bounded away from 0 by a stricly positive constant e, if the i are in H 1,h,h/2 (DT ) and if belongs to H 0,h,h/2 (DT ), then, for b small enough, there exists 0 < T T and a solution of the equation (1) on DT BT . The rest of the paper will be devoted to the proof of Theorem 1 above.

Preliminaries

In this section, we write xi and xi without distinction. Let us start as in [4] and [5] with the following notion of distance d (P, Q) = [|x x |2 + |t t |]1/2 where P = (t , x) and Q = (t , x ) belong to DT and |x| is the norm of the n-dimensionnal vector x. Given such a metric d, we can dene the concept of Hlder continuity. For a function u, we write: |u|D 0 = sup|u|
DT
T T

D Hh (u) = sup

|u(P) u(Q)| d (P, Q)h P,QDT

D D |u|D h = |u|0 + Hh (u)

D (u) < if and only if u is uniformly hlder (exponent h) in DT . We denote by H 0,h,h/2 (DT ) Hh T the set of all functions u for which |u|D h < . Now, if all the derivatives used in the equation exist, we write for k 2: D D k D D |u|D k+h = |u|h + |x u|h + ... + |x u|h + |t u|h
T T T T T

(3)

where the sums are taken over all the partial derivatives of the indicated order. We denote by T H k,h,h/2 (DT ) the set of all functions u for which |u|D k+h < . It is a Banach space and an algebra with the norm given by denition 3. Indeed, for all u,v in H k,h,h/2 (DT ), we have:
D D |uv|D k+h |u|k+h |v|k+h
T T T

(4)

We can now make the assumptions about DT more precise: for every point Q of CT , there exists an (n+1)-dimensional neighborhood V such that V CT can be represented, for some i (1 i n), in the form xi = r(t , x1 , ..., xi1 , xi+1 , ..., xn ) with r, x r, 2 xx r , t r Hlder continuous (exponent h) and 2 r simply continous. 2 r , xt tt We also have to consider functions dened on B CT . Such a function is said to belong to H k,h,h/2 if there exists a in H k,h,h/2 (DT ) such that = on B CT . We then dene ||k+h = inf
k,h,h/2 (DT ) which coincide with on ||D k+h where the inf is taken with respect to all the s in H B CT . This process denes a norm on H k,h,h/2 . The following results will be useful in the proof of our result. They concern the PDE:
T

Lu :=

n n u 2 u u ai j (x, t ) + bi (x, t ) + c(x, t )u = f (x, t ) on DT BT t i, j=1 xi x j i=1 xi

(5)

u = on B CT We shall need the assumptions: the coefcients of the operator L belong to H 0,h,h/2 (DT ), let K1 be a bound on their norm
n

for all (x, t ) in DT and for all Rn , H 2,h,h/2 and | f |D h <

T

i, j=1

ai j (x, t )i j K2 | |2 (K2 > 0)

In addition, given the assumption about DT , if we consider a function H 2,h,h/2 , for any extension of , t is uniquely dened (by continuity) on the boundary B of B, and the denition is independent of . We denote this function (on B) by t . The other terms of L are also uniquely dened (by continuity) on B. Thus, the quantity L is well-dened on B.

Theorem 2. Under the previous assumptions and if L = f on B, there exists a unique solution of the equation 5, this solution belongs to H 2,h,h/2 (DT ) and we have the Schauder inequality (with KH 2 depending only on K1 , on K2 , on h and on DT )
D |u|D 2+h KH 2 (||2+h + | f |h )
t t

(6)

Furthermore, if = 0, we can write a bound containing the time on the supremum of the solution |u|D 0
t

tK0 | f |D 0

(7)

where K0 only depends on K1 , on K2 , on h and on . Proof: the rst part of the result is classic, its proof can be found in [4]. As to the result with = 0, which is more original, one needs a result from [5] about volume potentials and representation of solutions of parabolic equations. It is the theorem (16.2) of section IV.16 we shall use. One reads that the solution of the equation 5 with = 0 can be written as
t

u(x, t ) =
0

G(x, z, t , ) f (z, )dz

where G is the Greens function for the operator L and veries |G(x, y, t , )| K (t ) 2 exp(K
n

|x y|2 ) t

(8)

with K and K two constants depending on the data of the problem. Using both these results, we get, for all t t and x |u(x, t )| t | f |D 0
t

d
0

K (t ) 2 1 exp(K

t |x y|2 )dz t | f |D 0 K0 t

where K0 depends on K1 , on K2 , on h and on DT .

Proof of Theorem 1
n 2 2 p 2 (11 1 I ( p) p) (1i 1 i t S2 i=2 Sxi n

We are interested in the equation I ( p) p)

n 2 (i j i j p) + (i p) + p = 0 on DT BT i, j=2 xi x j i=1 xi

and want to prove the Theorem. If the i belong to H 2,h,h/2 (DT ), are positive and bounded away from 0 by a stricly positive constant e, if the i are in H 1,h,h/2 (DT ) and if belongs to H 0,h,h/2 (DT ), then, for b small enough, there exists 0 < T T and a solution of the equation (1) on DT BT . Proof: The assumption (H2) gives us some control over the variations of b. Let us denote by b = b(y0 ) a strictly positive value taken by b (with y0 S y for some arbitrary S S ). We use the assumption on b to approximate the integral term I ( p) with 1/b2 , the gap between those two quantities is quantied with the

Lemma 1. There exists a constant Kb (depending only on h, n, 1 , 2 , p0 and ) and a polynomial 2,h,h/2 (DT ) veriying p p, function P strictly positive and increasing on R 0 + such that p H we have |I ( p) 1 Dt 1 Dt Dt | + | I ( p ) | b K P ( | p | b 2 + h 2 + h 2+h ). b b2

Remark. As a consequence of this lemma, we see that p H 2,h,h/2 (DT ) veriying p0 p, I ( p) belongs to H 2,h,h/2 (DT ). We then write the equation as
n n n 1 2 2 p 2 1 2 (11 2 p ) ( p ) ( p ) + (i p) + p ij i j 1i 1 i 1 2 t S b b i=2 Sxi i, j=2 xi x j i=1 xi

n 2 1 2 1 2 ( ( I ( p ) (1i 1 i ( I ( p) ) p) ) p ) + 11 1 2 2 S b b i=2 Sxi

To solve this equation, we apply a xed point method and use the lemma 1 to get an upper bound on the second term. T t We take a real number x | p0 |D 2+h and t R+ and let Xx denote the set
t Xx = { p H 2,h,h/2 (Dt ), | p|D 2+h x,
t

p0 p0 p p0 + , p = on B CT } 2 2

t clearly contains the function p . We then consider the application M which takes a The set Xx 0 function u X and sends it on v H 2,h,h/2 (DT ) solution of the equation

O v :=

n n n v 2 1 1 2 2 2 (11 2 ( v ) ( v ) + (i v) + v v ) i i j i j 1 i 1 1 2 t S b b i, j=2 xi x j i=1 xi i=2 Sxi

n 2 1 2 1 2 ( ( I ( u ) ( ( ) u ) + I ( u ) )u) (9) i 11 1 i 1 1 S 2 b b2 i=2 Sxi

with the boundary condition v = on B CT . The existence of v is given by Theorem 2. Indeed, the coefcients of this equation belong to the appropriate spaces and because of (H4) the necessary condition O=
n 2 1 2 1 2 ( ( I ( ) ) ) + (1i 1 i ( I () )) 11 1 2 2 S b b i=2 Sxi

on B is veried. It remains to prove that this operator is elliptic: let (i )1in be n real numbers 1 and (t , S, y) DT . We write f1 = b and f i = i for i 2, we have
n i, j=1 n n

i j fi(t , S, y) f j (t , S, y)i j

fi2 (t , S, y)2 i

K e

i=1

i=1

2 i

where the existence of K is a consequence of being a positive denite matrix. This proves the ellipticity of the operator, v exists and belongs to H 2,h,h/2 (DT ). We now want to show that for t ie that suitable x and t , v belongs to Xx |v|D 2+h x
t

p0 p0 v p0 + 2 2

For the rst inequality, we apply 6

D |v|D 2+h KH 2 (||2+h + |
T t t

n t 2 1 2 1 2 ( ( I ( u ) ) u ) + (1i 1 i ( I (u) )u)|D 11 1 h ) 2 2 S b b i=2 Sxi n t 1 Dt 1 ) u | + |1i 1 i ( I (u) )u|D 2+h 2+h ) 2 b b i=2 n
T

2 KH 2 (||D 2+h + |11 1 (I (u)

T T

D 2 D KH 2 (||D 2+h + (|11 1 |2+h + |1i 1 i |2+h )(|I (u)

where K

T KH 2 (||D 2+h + b K

i=2 t Dt P(|u|2+h )|u|D 2+h )

T

1 Dt 1 t Dt | + | I (u) |D 2+h )|u|2+h ) 2 2+h b b (10)

DT = (|11 2 1 |2+h +

n i=2

|1i1i|D 2+h )Kb (we apply lemma 1 for the last line).
t

t , thus |u|D x and then We remember that u belongs to Xx 2+h D |v|D 2+h KH 2 (||2+h + b K P(x)x) D Taking x = max(KH 2 (||D 2+h + 1), | p0 |2+h ) and b
T T t T

It remains to prove that Ov = O p0 +

p0 2

v p0 +

p0 2.

1 K P(x )x ,

10 gives us |v|D 2+h x .

Let us write v = v p0 . It is clear that v veries

n 2 1 1 2 2 ( ( I ( u ) ) u ) + ( ( I ( u ) )u) i 11 1 i 1 1 S 2 S x b b2 i i=2

on DT BT with v = 0 on B CT (here we use the fact that is constant on CT ). We now apply the second part of Theorem 2, the inequality 7, to this function v
t |v |D 0

n t 2 1 1 2 2 tK0 |O p0 + 2 (11 1 (I (u) 2 )u) + (1i 1 i ( I (u) )u)|D 0 S b b i=2 Sxi

tK0 (|O

T p0 |D 0 T

n t 2 2 1 1 2 + | 2 (11 1 (I (u) 2 )u) + (1i 1 i ( I (u) )u)|D h ) S b b i=2 Sxi

tK0 (|O p0 |D 0 + 1)
T

(11)
t

0 0 Taking T K0 (|O p0 |D |D , the last inequality 0 + 1) = 2 , we get |v 0 2 . Eventually, since v = p0 + v T T is proved and v belongs to Xx . The application M maps Xx into itself. T Using this statement, we construct a bounded sequence ( pn )nN of functions belonging to Xx

p0 has been previously dened by induction, we write pn+1 = M ( pn )

By construction, we have n N, | pn |D 2+h x . Repeated applications of the Ascoli-Arzel theorem give us a function p C2 (DT ) limit in C2 (DT ) of a subsequence of pn . Since
T

| sup{

2 pn 2 pn ( x , t ) xi x j xi x j (x , t ) | ; (x, t ), (x (| x x |2 + | t t |)/2

, t ) DT } x

We have | sup{
2 p 2 p xi x j (x, t ) xi x j (x , t ) | , (x, t ); (x (| x x |2 + | t t |)/2

, t ) DT } x

And this computation being true for all the derivatives appearing in the norm H 2,h,h/2 , we nd that p H 2,h,h/2 . The last step of the proof is to take the limit in 9. The only result needed is I ( pn ) I ( p). Since pn T , the denominator is bounded away from 0. Two applications of the dominated convergence Xx theorem give us the convergence we need. Thus, it is possible to take the limit in 9 which gives us
n n n p 2 2 1 2 2 1 2 (11 1 2 p) (1i 1 i p) (i j i j p) + (i p) + p t S b b i=2 Sxi i=1 xi i, j=2 xi x j

n 2 1 2 1 2 ( I ( p ) ( (1i 1 i ( I ( p) ) p) ) p ) + 11 1 2 2 S b b i=2 Sxi

That concludes the proof of the theorem, p is solution of our equation. Let us now prove lemma 1. By denition, we have: |I ( p) 1 DT 1 DT pdy p(b2 b2 )dy DT = | = | | | |2+h 2+h b2 pdy b2 2+h b2 b2 b2 pdy T T 1 1 p(b2 b2 )dy|D |D 2| 2 2+h | 2+h b pdy b

Let us compute one after another the terms appearing in this norm (we remember that those functions only depend on t and S). Let (t , S) belong to ]0, T [ S . We have |
S y

p(t , S, y)(b2 b2 (y))dy|

S y

p(t , S, y)|b2 b2 (y)|dy | p|D 0

T

S y

|b2 b2 (y)|dy |y0 y|dy| p|D 0

T

| p|D 0

S y
T

|b2 (y0 ) b2 (y)|dy b (n 2)

S y

b K | p|D 0

here and in the rest of the proof, K stands for some constant depending only on the data of the problem (, n, 1 ...) but not on p nor on b . We get the last line from the following computation where y = (x2 , ., xn ) and y0 = (x2 , ., xn ):
n

|b2 (y0 ) b2 (y)| = |b2 (x2 , ., xn ) b2 (x2 , ., xn )|

i=3 n i=3

|b2(x2, ., xi, xi+1, ., xn) b2(x2, ., xi1, xi, ., xn)|

(b2 ) Dt | |x xi | (n 2)b |y y| xi 0 i

Now, let (t , S) and (t , S ) belong to ]0, T [ S . We compute |

S y

p(t , S, y)(b2 (y0 ) b2 (y))dy

S y

p(t , S , y)(b2 (y0 ) b2 (y))dy| p(t , S, y)|b2 (y0 ) b2 (y)|dy p(t , S, y)|b2 (y0 ) b2 (y)|dy |b2 (y0 ) b2 (y)|dy |b2 (y0 ) b2 (y)|dy |y0 y|dy)

S S y y

| p(t , S, y) p(t , S , y)||b2 (y0 ) b2 (y)|dy + +

S S y \y

S S y \y T

D Hh ( p)D((t , S), (t , S ))

S S y y

|b2 (y0 ) b2 (y)|dy + | p|D 0 +| p|D 0

S S y \y T

S y

\S y

b K | p|D h (D((t , S), (t , S )) +

S S y \y

|y0 y|dy +

S S y \y

By assumption on the boundary of our domain, it is possible to nd a constant K depending only on such as S, S S , S \S |y0 y|dy KD(S, S ). This gives us
y y

|
T

D p(b2 b2 )dy|D h b K | p|h

T

2 2 D And since | p(b2 b2 )dy|D 2+h = | p(b b )dy|h + |

T 2 p 2 (b b2 )dy|D h , S 2

p 2 2 DT t (b b )dy|h

+|

p 2 2 DT S (b b )dy|h

we get from the previous computation |

D p(b2 b2 )dy|D 2+h b K | p|2+h
T 1 |D . b2 pdy 2+h T t

We now have to nd a bound on |

2 . 2 1 p0V ()

T , we have | Since p belongs to Xx

T 1 |D b2 pdy 0

Now, let (t , S) and (t , S ) belong to ]0, T [ S . We write | 1 | b2 (y) p(S , t , y)dy

S y

1 | 2 b (y) p(S, t , y)dy S y

b2 (y) p(S , t , y)dy

t

b S y

2 (y) p(S, t , y)dy|

S y

b2 (y) p(S, t , y)dy S S y y

b2 (y) p(S , t , y)dy

D ( p)D((t , S), (t , S )) KHh

We used the same kind of arguments than earlier, K denotes here another constant depending on 1 , T 1 DT 2 , p0 and . This gives us | b21pdy |D h K (1 + | p|h ). As for derivatives of b2 pdy , for instance with respect to S, we have
p b2 1 DT 2 DT DT S dy DT | ( 2 )|h = | | K ( 1 + | p | ) | p | h h 1+h S b pdy ( b2 pdy)2

The same kind of computation is true for the derivative of second order
p p p 2 dy DT b2 b2 2( b2 T 3 DT DT 2 DT 2 DT S dy S dy) S2 | ( )|h = | |h K [(1 + | p|D h ) (| p|1+h ) + (1 + | p|h ) | p|2+h ] 2 2 2 3 2 2 S ( b pdy) ( b pdy) ( b pdy)
2

Eventually, we get |
T 1 |D 2+h 2 b pdy

D 2 D 3 D 4 D 5 K (1 + | p|D 2+h + (| p|2+h ) + (| p|2+h ) + (| p|2+h ) + (| p|2+h ) )

Combining this result with the previous computations, we nd |I ( p)

t 1 DT | b KP(| p|D 2+h ) 2 2+h b

with P a polynomial function of degree 6, strictly positive on R + . Now, writing

I ( p)
1 b2 I ( p)+ 1 b

1 I ( p) b =

, we nd the same kind of results for the second term involved in the lemma. This con-

cludes the proof.

Conclusion

In this paper, we have shown that the equation driving the calibration problem for local and stochastic volatility models is well-posed in the case of suitably regularized initial conditions. It is however clear that the solution of the full Kolmogorov equation with Dirac initial condition does not obtain as a consequence of Theorem 1 : possible extensions of our results towards this direction are currently being explored. Let us also mention that a generalization of Theorem 1 to multidimensional correlation calibration have already been investigated and will be presented in [8]

References
[1] D. Breeden, R. Litzenberger, State Contingent Prices Implicit in Option Prices. Journal of Business 51, 621-651, 1978 [2] H. Berestycki, J. Busca, I. Florent, Computing the implied volatility in stochastic volatility models. Comm. Pure Appl. Math., 57(10):1352-1373, 2004 [3] B. Dupire, Pricing and Hedging with Smiles. Proc AFFI Conf, La Baule, 1993 [4] A. Friedman, Partial differential equations of parabolic type, Prentice-Hall Inc., Englewood Cliffs, N.J., 1964. [5] O.A. Ladyzhenskaya, V.A. Solonnikov, N.N. Uralceva, Linear and quasilinear equations of parabolic type, Trans. Math. Monographs, vol. 23, Amer. Math. Soc., Providence, RI, 1968. [6] A. Pazy, Semigroups of linear operators and applications to partial differential equations, Springer-Verlag, New-York, 1983 [7] N. Alibaud, Existence, uniqueness and regularity for nonlinear parabolic equations with nonlocal terms equations with nonlocal terms, NoDEA Nonlinear Differential Equations Appl., 14 (3-4):259289, 2007 [8] F. Abergel, R. Tachet, work in progress