- Open Access
Discretisation of abstract linear evolution equations of parabolic type
© Gonçalves et al; licensee Springer. 2012
- Received: 6 October 2011
- Accepted: 16 February 2012
- Published: 16 February 2012
We investigate the discretisation of the linear parabolic equation du/dt = A(t)u + f(t) in abstract spaces, making use of both the implicit and the explicit finite-difference schemes. The stability of the explicit scheme is obtained, and the schemes' rates of convergence are estimated. Additionally, we study the special cases where A and f are approximated by integral averages and also by weighted arithmetic averages.
MSC 2010: 65J10.
- parabolic evolution equations
- finite-difference methods
- financial mathematics
where, for every t ∈ [0, T] with T ∈ (0, ∞), A(t) is a linear operator from a reflexive separable Banach space V to its dual V*, u: [0, T] → V is an unknown function, f: [0, T] → V*, g belongs to a Hilbert space H, with f and g given, and V is continuously and densely embedded into H. We assume that operator A(t) is continuous and impose a coercivity condition.
Our motivation lies in the numerical approximation of multidimensional PDE problems arising in European financial option pricing. Let us consider the stochastic modeling of a multi-asset financial option of European type under the framework of a general version of Black-Scholes model, where the vector of asset appreciation rates and the volatility matrix are taken time and space-dependent. Owing to a Feynman-Kač type formula, pricing this option can be reduced to solving the Cauchy problem (with terminal condition) for a second-order linear parabolic PDE of nondivergent type, with null term and unbounded coefficients, degenerating in the space variables (see, e.g., ).
with real coefficients, f and g are given real-valued functions (the free term f is included to further improve generality), and T ∈ (0, ∞) is a constant. For each t ∈ [0, T] the operator - L is degenerate elliptic, and the growth in the spatial variables of the coefficients a, b, and of the free data f, g is allowed. One possible approach for the numerical approximation of the PDE problem (2) is to proceed to a two-stage discretisation. First, the problem is semi-discretised in space, and both the possible equation degeneracy and coefficient unboundedness are dealt with (see, e.g., [2, 3], where the spatial approximation is pursued in a variational framework, under the strong assumption that the PDE does not degenerate, and ). Subsequently, a time discretisation takes place.
For the time discretisation, the topic of the present article, it can be tackled by approximating the linear evolution equation problem (1) which the PDE problem (2) can be cast into. This simpler general approach, which we follow, is powerful enough to obtain the desired results. On the other hand, it covers a variety of problems, namely initial-value and initial boundary-value problems for linear parabolic PDEs of any order m ≥ 2.
Several studies dealing with the discretisation of parabolic evolution problems in abstract spaces can be found in the literature. Most of them are concerned with the discretisation of problems with constant operator A (see, e.g., [5–9]). Other studies (see, e.g., [10–13]), study the general case where the operator A is time-dependent, under Hölder or Lipschitz-continuity assumptions. Also, in some of the above mentioned studies and in others, as in , the discretisation is pursued by considering a particular discretisation of the datum f (namely, by using integral averages).
In the present study, we study the discretisation in time of problem (1) with time-dependent operator A in a general setting. We use both the implicit and the explicit finite-difference schemes. To further improve generality, we proceed to the study leaving the discretised versions of A and f nonspecified. Also, in order to obtain the convergence of the schemes, we need to assume that the solution of (1) satisfies a smoothness condition but weaker than the usual Hölder-continuity.
It is well known that, to guarantee the explicit scheme stability, an additional assumption has to be made, usually involving an inverse inequality between V and H (see, e.g., ). In our study, the explicit discretisation is investigated by assuming instead a not usual inverse inequality between H and V*.
In addition, we illustrate our study by exploring examples where different choices are made for the discretised versions of A and f.
First, we consider the approximation of A and f by integral averages. We show that the standard smoothness and coercivity assumptions for problem (1) induce correspondent properties for the discretised problem, so that stability results can be proved. Moreover, the rate of convergence we obtain is optimal. Then, we study the alternative approximation of A and f by weighted arithmetic averages of their respective values at consecutive time-grid points. In this case, stronger smoothness assumptions are needed in order to obtain the scheme convergence.
We emphasize that none of the above mentioned choices is artificial: there are applications where the available information regards the values of A and f at the time-grid points and others the integral averages, but usually not both.
The article is organized as follows. In Section 2, we set an abstract framework for a linear parabolic evolution equation and present a solvability classical result. In the following two sections, we study the discretisation of the evolution equation with the use of the Euler's implicit scheme (Section 3) and the Euler's explicit scheme (Section 4). In Sections 5 and 6, we discuss some examples, respectively, for the implicit and the explicit discretisation schemes and, finally, in Section 7, we present some computational results.
We establish some facts on the solvability of linear evolution equations of parabolic type.
with T ∈ (0, ∞), where A(t) is a linear operator from V to V* for every t ∈ [0, T] and A(·)v : [0, T] → V* is measurable for fixed v ∈ V, u : [0, T] → V is an unknown differentiable function, f : [0, T] → V* is a measurable given function, d/dt is the standard derivative with respect to the time variable t, and g ∈ H is given.
We assume that the operator A(t) is continuous and impose a coercivity condition, as well as some regularity on the free data f and g.
Assumption 1. Suppose that there exist constants λ > 0, K, M, and N such that
1., ∀v ∈ V and ∀t ∈ [0,T];
2. ∥A(t)v∥V*≤ M∥v∥ V , ∀v ∈ V and ∀t ∈ [0, T];
3.and ∥g∥ H ≤ N.
We define the generalized solution of problem (3).
Definition 1. We say that u ∈ C([0,T]; H) is a generalized solution of (3) on [0,T] if
1. u ∈ L2([0,T];V);
The following well-known result states the existence and uniqueness of the generalized solution of problem (3) (see, e.g., ).
where N is a constant.
We will now study the time discretisation of problem (3) making use of an implicit finite-difference scheme. We begin by constructing an appropriate discrete framework.
where k := T/n. Denote t j = jk for j = 0,1,..., n.
Let A k , f k be some time-discrete versions of A and f, respectively, i.e., A k (t j ) is a linear operator from V to V* for every j = 0,1,..., n and f k : T n → V* a function. For all z ∈ V, denote Ak,j+1z = A k (tj+1) z , fk,j+1= f k (tj+1), j = 0, 1,..., n - 1.
Problem (5) is a time-discrete version of problem (3).
Assumption 2. Suppose that
1., ∀v ∈ V, j = 0,1,..., n - 1,
2. ∥Ak,j+1v∥V*≤ M∥v∥ V , ∀v ∈ V, j = 0,1,..., n - 1,
3.and ∥g∥ H ≤ N,
where λ, K, M, and N are the constants in Assumption 1.
Remark 1. Note that as problem (5) is a time-discrete version of problem (3) and g denotes the same function in both problems, under Assumption 1 we have that g ∈ H and ∥g∥ H ≤ N.
Under the above assumption, we establish the existence and uniqueness of the solution of problem (5).
Theorem 2. Let Assumption 2 be satisfied and the constant K be such that Kk ≤ 1. Then for all n ∈ ℕ there exists a unique vector v0, v1, ... ,v n in V satisfying (5).
Lemma 1 (Lax-Milgram). Let B : V → V* be a bounded linear operator. Assume there exists λ > 0 such that, for all v ∈ V. Then Bv = v* has a unique solution v ∈ V for every given v* ∈ V*.
Proof. (Theorem 2)
From (5), we have that (I - kAk,1)v1 = g + fk,1k and
(I - kAk,i+1)vi+1= v i + fk,i+1k, for i = 0,1,..., n - 1, with I the identity operator on V.
Then, as Kk ≤ 1, we have that and the hypotheses of Lemma 1 are satisfied.
For v1, we have that (I - kAk 1)v1 = g + fk,1k. This equation has a unique solution by Lemma 1. Suppose now that equation (I - kAk,i)v i = vi-1+ fk,ik has a unique solution. Then equation (I - kAk,i+1)vi+1= v i + fk,i+1k has also a unique solution, again by Lemma 1. The result is obtained by induction.
Next, we prove an auxiliary result and then obtain a version of the discrete Gronwall's lemma convenient for our purposes.
Lemma 2. Letbe a finite sequence of numbers for every integer n ≥ 1 such that, for all j = 1, 2,..., n, where C is a positive constant and c0 ≥ 0 is some real number. Then, for all j = 1, 2,..., n.
and the result is proved.
for all integers n ≥ 1 and j = 1, 2,..., n, where K q := -K ln(1 - q)/q.
the result is proved.
We are now able to prove that the scheme (5) is stable, that is, the solution of the discrete problem remains bounded independently of k.
Theorem 3. Let Assumption 2 be satisfied and assume further that constant K satisfies: 2Kk < 1. Denote vk,j, with j = 0, 1, ..., n, the unique solution of problem (5) in Theorem 2. Then there exists a constant N independent of k such that
Remark 3. Under Assumption 2, with K satisfying 2Kk < 1, Theorem 2 obviously holds so that problem (5) has a unique solution.
Proof. (Theorem 3)
Estimate (2) follows.
We will now study the convergence properties of the scheme we have constructed. We impose stronger regularity on the solution u = u(t) of problem (3):
for all i = 0, 1, ..., n - 1.
Remark 4. Assume that u satisfies the following condition: "There exist a fixed number δ ∈ (0,1] and a constant C such that ∥u(t) - u(s)∥ V ≤ C|t - s| δ , for all s, t ∈ [0,T]". Then Assumption 3 obviously holds.
By assuming this stronger regularity of the solution u of (3), we can prove the convergence of the solution of problem (5) to the solution of problem (3) and determine the convergence rate. The accuracy we obtain is of order δ.
Theorem 4. Let Assumptions 1 and 2 be satisfied and assume further that constant K satisfies: 2Kk < 1. Denote u(t) the unique solution of (3) in Theorem 1 and vk,j, j = 0, 1, ..., n, the unique solution of (5) in Theorem 2. Let also Assumption 3 be satisfied. Then there exists a constant N independent of k such that
where φ(ti+1) := fk,i+1k - u(ti+1) + u(t i ) + Ak,i+1u(ti+1)k.
Let us estimate separately each one of the three terms in (13).
with λ > 0.
with λ > 0, using Cauchy's inequality.
with N a constant. Following the same steps as in the proof of Theorem 3, estimates (1) and (2) follow.
Next result is an immediate consequence of Theorem 4.
with N be a constant independent of k.
We now approach the time-discretisation with the use of an explicit finite-difference scheme. As in the previous section, we begin by setting a suitable discrete framework and then investigate the stability and convergence properties of the scheme.
with A h (t), f h (t), and g h "space-discrete versions" of A(t), f(t), and g, and h ∈ (0,1] a constant. We will use the notation (·,·) h for the inner product in H h and 〈·,·〉 h for the duality between and V h .
with j = 0,1,..., n - 1.
with v j = v(t j ), j = 0,1,..., n, in V h .
We make some assumptions.
Assumption 4. Suppose that
1., ∀v ∈ V h , j = 0,1,..., n - 1,
2., ∀v ∈ V h , j = 0,1,..., n - 1,
where λ, K, M, and N are the constants in Assumption 1.
Remark 5. We refer to Remark 1 and note that, under Assumption 1, g h ∈ H h and.
The following version of the discrete Gronwall's inequality is an immediate consequence of Lemma 3.
for all integers n ≥ 1 and j = 0, 1,..., n, where K q := -K ln(1 - q)/q.
for j = 1, 2,..., n. The result follows.
In order to obtain stability for the scheme (19) we make an additional assumption, involving an inverse inequality between H h and . We note that, for the case of the implicit scheme, there was no such need: the implicit scheme's stability was met unconditionally.
with the last inequality above due to (22).
Remark 7. Assumption 5 is not void. For example, when the solvability of a multidimensional linear PDE of parabolic type is considered in Sobolev spaces, and its discretised version solvability in discrete counterparts of those spaces (see), (21) is satisfied with C h such that, with C a constant independent of h.
Theorem 5. Let Assumptions 4 and 5 be satisfied and λ, K, M, and C h the constants defined in the Assumptions. Denote by v hk,j , with j = 0,1, ...,n, the unique solution of problem (19). Assume that constant K is such that 2Kk < 1. If there exists a number p such thatthen there exists a constant N, independent of k and h, such that
Remark 8. Remark 2 applies to the above theorem with the obvious adaptations.
Proof. (Theorem 5)
with λ > 0.
with μ > 0.
where L := (μM2 + λ(1 + μ)p)/λμM2.
where K q is the constant defined in Lemma 4. (1) follows.
and (2) follows.
Finally, we prove the convergence of the scheme and determine the convergence rate. The accuracy obtained is of order δ, with δ given by Assumption 3.
Theorem 6. Let Assumptions 1, 4, and 5 be satisfied and λ, K, M, and C h the constants defined in the Assumptions. Denote by u h (t) the unique solution of problem (18) in Theorem 1 and by vhk,j, with j = 0, 1,..., n, the unique solution of problem (19). Assume that constant K is such that 2Kk < 1 and that Assumption 3 is satisfied. If there exists a number p such thatthen there exists a constant N, independent of k and h, such that
where φ(t i ) := fhk,ik - u h (ti+1) + u h (t i ) + Ahk,iu h (t i )k.
where L := ((3M2 + λp+ νλp)μν + (1 + μ+ ν + ν 2 )λp)/μνλM2. Estimates (1) and (2) are obtained following the same steps as in Theorem 5.
Next result follows immediately from Theorem 6.
with N a constant independent of k.
for all z ∈ V, j = 0,1,..., n - 1.
We prove that, under Assumption 1, Ā k and satisfy Assumption 2.
Proposition 1. Under Assumption 1, operator Ā k and functionsatisfy
1. , ∀v ∈ V, j = 0,1,..., n - 1,
2. , ∀v ∈ V, j = 0,1,..., n - 1,
where λ, K, M, and N are the constants in Assumption 1.
with j = 0,1,..., n - 1, and (1) is proved.