ESAIM: M2AN ESAIM: Mathematical Modelling and Numerical Analysis

Vol. 41, No 3, 2007, pp. 485–511 www.edpsciences.org/m2an

DOI: 10.1051/m2an:2007029

A POSTERIORI ERROR ANALYSIS FOR PARABOLIC

VARIATIONAL INEQUALITIES ∗

Kyoung-Sook Moon 1 , Ricardo H. Nochetto 2 , Tobias von Petersdorff 3

and Chen-song Zhang 3

Abstract. Motivated by the pricing of American options for baskets we consider a parabolic varia-
tional inequality in a bounded polyhedral domain Ω ⊂ Rd with a continuous piecewise smooth obstacle.
We formulate a fully discrete method by using piecewise linear finite elements in space and the back-
ward Euler method in time. We define an a posteriori error estimator and show that it gives an upper
bound for the error in L2 (0, T ; H 1 (Ω)). The error estimator is localized in the sense that the size of
the elliptic residual is only relevant in the approximate non-contact region, and the approximability of
the obstacle is only relevant in the approximate contact region. We also obtain lower bound results
for the space error indicators in the non-contact region, and for the time error estimator. Numerical
results for d = 1, 2 show that the error estimator decays with the same rate as the actual error when
the space meshsize h and the time step τ tend to zero. Also, the error indicators capture the correct
behavior of the errors in both the contact and the non-contact regions.

Mathematics Subject Classification. 58E35, 65N15, 65N30.

Received January 20, 2006.

1. Introduction
The pricing of options is of considerable importance in finance [28]. We consider the setting of the classical
Black-Scholes model [3]: assume that the price S(t) of the underlying risky asset (e.g., a stock) is described by
geometric Brownian motion with volatility α > 0 and interest rate r > 0. An American put option with strike
price K and expiration date T gives the holder the right to sell one asset at any time t before the expiration
date at price K. At the time t when the option is exercised, its value is given by H(S(t)) with the payoff
function H(S) = (K − S)+ = max{K − S, 0}. The problem we want to solve is the following: If at time t we
have an asset price S(t), what is the fair price V (S, t) of the option? And when is the optimal time to exercise
the option?

Keywords and phrases. A posteriori error analysis, finite element method, variational inequality, American option pricing.
∗ Partially supported by NSF Grants DMS-0204670 and DMS-0505454.
1 Department of Mathematics and Information, Kyungwon University, Bokjeong-dong, Sujeong-gu, Seongnam-si, Gyeonggi-do,
461-701, Korea. [email protected]
2 Department of Mathematics and Institute for Physical Science and Technology, University of Maryland, College Park,

MD 20742, USA. [email protected]

3 Department of Mathematics, University of Maryland, College Park, MD 20742, USA. [email protected];

[email protected]
c EDP Sciences, SMAI 2007

Article published by EDP Sciences and available at http://www.esaim-m2an.org or http://dx.doi.org/10.1051/m2an:2007029

486 A POSTERIORI ERROR ANALYSIS FOR PARABOLIC VARIATIONAL INEQUALITIES

Under the standard assumption of a frictionless market without arbitrage one can formulate this problem as an
optimal stopping problem and find that the solution V (S, t) satisfies a parabolic variational inequality problem.
Using the time to maturity t̃ = T − t and x = log S as independent variables, the function u(x, t̃) := V (ex , T − t̃)
satisfies the following differential inequality (we will write t instead of t̃ from now on):


∂u ∂u α2 ∂ 2 u α2 ∂u
+ Lu = − + −r + ru ≥ 0 for x ∈ R and 0 < t < T (1.1)
∂t ∂t 2 ∂x2 2 ∂x

with the obstacle constraint

u(x, t) ≥ χ(x) for x ∈ R and 0 < t < T (1.2)
and the initial condition

u(x, 0) = u0 (x) = H(ex ) = max(K − ex , 0) for x ∈ R (1.3)

where χ(x) = u0 (x) is the payoff function in the log of the asset price. Moreover, for each point (x, t) ∈ R×(0, T )
either (1.1) or (1.2) is an equality, which is the so-called complementarity condition. The set of points C :=
{(x, t) ∈ R × (0, T ) : u(x, t) = χ(x)} is called the contact set, and its complement N the non-contact set. The
boundary F between the two sets is called the free boundary. The optimal stopping time t∗ (i.e. the optimal
time to exercise the option) is given by
 
t∗ = inf t ∈ [0, T ] : (log S(t), t) ∈ C . (1.4)

That means t∗ is the first time when the graph of log S(t) crosses the free boundary F . The solution u(x, t) has
a singular behavior in both time and space close to t = 0 and x = log K (i.e., time close to maturity and price
close to strike price).
This problem has to be solved on the whole real line. For practical computations one uses a bounded interval
Ω := (−R, R) and boundary conditions u(−R, t) = χ(−R), u(R, t) = 0. This causes an error which decreases
exponentially in R, and is therefore negligible for R large enough. By subtracting a suitable function from u(x)
we can obtain a problem with zero Dirichlet conditions and a right-hand side function f . The resulting problem
is formulated variationally in Ω × [0, T ] in terms of the bilinear form
  
α2   2
a(v, w) := 2 v (x)w (x) + ( α2 − r)v  (x)w(x) + rv(x)w(x) dx ∀v, w ∈ H01 (Ω). (1.5)
Ω

However, in finance one also needs to consider baskets of assets with prices S1 , . . . , Sd which lead to a parabolic
variational inequality for u(x, t) with x ∈ Rd and bilinear form a(·, ·) and operator L given by

a(v, w) := Lv, w , Lv := −div(µ2 ∇v) + b · ∇v + cv ∀v, w ∈ H01 (Ω), (1.6)

where ·, · denotes either the L2 scalar product or the duality pairing between H −1 (Ω) and H01 (Ω). Note that
the bilinear form a(·, ·) is nonsymmetric: for d = 1 it is possible to symmetrize the problem but this is not
always the case for d > 1. Therefore we are interested in variational inequalities in Ω ⊂ Rd for a nonsymmetric
operator L. We allow variable coefficients µ, b, c, so that we can have volatilities depending on the underlying
asset prices. A weak solution of the corresponding parabolic variational inequality is a function u(x, t) in
C([0, T ]; L2 (Ω)) ∩ L2 (0, T ; H 1(Ω)) such that

∂t u, u − v + a(u, u − v) ≤ f, u − v ∀v ∈ K(t) := {v ∈ H01 (Ω) : v ≥ χ(t)} a.e. t ∈ [0, T ] (1.7)

and u(t) ∈ K(t) a.e. t ∈ [0, T ]. Here instead of restricting ourselves to the problem (1.1), we consider a more
general right-hand side function f ∈ L2 (0, T ; L2(Ω)).
 K.-S. MOON ET AL. 487

Several numerical methods have been proposed to approximate this problem for d = 1; among them are
lattice methods, Monte Carlo methods, finite difference methods and finite element methods (see [6, 15, 28] and
the references therein). In the special case of the classical American option for one asset there exist various
analytic results and specialized numerical methods [13]. For d > 1, variational inequality (1.7) is usually solved
by a finite element (or finite difference) discretization in space, and a backward Euler discretization in time.
We refer to the monograph by Glowinski [12] and the references therein for a survey of numerical methods for
parabolic variational inequalities.
There are a number of a priori error estimates available for the parabolic variational inequality (1.7), such
as Johnson [14], Fetter [10], and Vuik [27], but they all assume a certain amount of regularity not valid in
2
our setting. For instance, Johnson assumes that u0 ∈ W∞ (Ω), and χ = 0, and obtains for L symmetric (with
some additional regularity assumptions) an error estimate O((log τ −1 )1/4 τ 3/4 + h) for the L2 (0, T ; H 1 (Ω)) error.
In option pricing problems, however, the obstacle χ in (1.2) and the initial condition u0 in (1.3) are merely
Lipschitz, thereby giving rise to a solution u singular close to maturity time and points in space where the
payoff function χ is nonsmooth; in fact, ∂t u and ∂xx u blow up in L2 (Ω) at t goes down toward 0 (see numerical
example 5.4). Therefore, the a priori error analysis does not provide a satisfactory description of (1.7) for
realistic financial applications. This also includes variable time steps and graded meshes, which are relevant to
resolve space-time singularities of u but are not discussed in [10, 14, 27] where uniform time and space mesh
refinement is assumed.
The a posteriori error analysis is much more recent and rather intricate. To gain some insight on the
difficulties involved, we let A(u) := Lu + σ(u) be the nonlinear operator of (1.7), which consists of the linear
nonsymmetric operator L and the nonlinear part σ that accounts for the unilateral constraint u ≥ χ. The latter
is the so-called Lagrange multiplier, and satisfies

σ(u) = 0 in N = {u > χ}, σ(u) = f − ut − Lu ≤ 0 in C = {u = χ}; (1.8)

hence σ(u) restores the equality in (1.7), namely,

ut + Lu + σ(u) = f. (1.9)

A posteriori error estimates of residual type are obtained by plugging the discrete solution U into the PDE.
Roughly speaking, we get the defect measure

G = Ut + LU + σ(U ) − f, (1.10)

which is called Galerkin functional in this nonlinear context; precise definitions are given in Section 2.5. This is a
replacement for the usual residual in linear theory and shows that, to obtain sharp a posteriori error estimators,
we must be able to provide a discrete multiplier σ(U ) with properties similar to (1.8). In fact, the linear part R
of G, that is R := Ut + LU − f , does not give correct information in the contact set N , where the solution
adheres to the obstacle regardless of the size of R.
Our analysis hinges on the following two key ideas, both about dealing with σ(u):
• Time discretization: The nonlinear operator A(u) = Lu + σ(u) of (1.7) is angle-bounded (see Sect. 2.2). This
allows for optimal order-regularity error estimates for the implicit Euler method with variable time steps,
as shown by Nochetto et al. [17, 18]. Notice the importance of exploiting the structure of σ(u) because its
dependence on u is not smooth.
• Space discretization: A suitable definition of discrete contact and non-contact sets (see Sect. 2.3), as well as
a discrete Lagrange multiplier σ(U ) (see Sect. 2.4), inspired by those of Fierro and Veeser [11]. This yields a
space error estimator which is localized in that the estimator vanishes in the discrete contact set, except for
obstacle resolution, whereas it reduces to the linear residual R in the discrete non-contact set where obstacle
resolution is immaterial.
488 A POSTERIORI ERROR ANALYSIS FOR PARABOLIC VARIATIONAL INEQUALITIES

We point out that failure to recognize the importance of σ(u) leads to a global upper bound of the error but not
to a global lower bound [8]; overestimation is thus possible. This issue was first addressed for elliptic variational
inequalities by Veeser [23] and further improved by Fierro and Veeser [11] in H 1 (Ω). Nochetto et al. extended
these estimates to L∞ (Ω) and derived barrier set estimates [19, 20]. If the variational inequality becomes an
equality, our energy estimates reduce to those in [2,21,25]. The duality approach, reported in [1], is not suitable
in this setting because of the singular character of σ(u).
In this paper we construct, analyze, and test a computable a posteriori error estimator of residual type for
a full discretization of (1.7), consisting of continuous piecewise linear finite elements in space and the implicit
Euler method in time, under realistic regularity assumptions. This undertaking is critical in finance but seems
to be missing in the literature. Our main result is an a posteriori upper bound (up to a multiplicative constant)
for the error in the L2 (0, T ; H 1 (Ω))-norm. We also obtain lower bounds for the error in terms of the leading
term of the space error estimator in the non-contact region and the time error estimator. This means that the
space error estimator as well as the time error estimator must decay with the correct rate. Moreover, the space
error estimator is a sum of local terms which are bounded from below by the local actual error. Therefore these
local terms are suitable as indicators for refinement in an adaptive method. Similarly, the time estimator is a
sum of terms for each time step which can be used as indicators for refinement of the time step. We do not
pursue this endeavor here.
In our numerical experiments of Section 5, we measure the error in L2 (0, T ; H 1 (Ω)). Our results indicate that
the time and space components of the error as well as the error estimator indeed converge with the expected
rates. For smooth initial data, and uniform meshes in space and time with meshsizes τ and h, respectively, we
obtain an optimal convergence rate of O(τ + h); see Sections 5.2 and 5.6. In the case of nonsmooth initial data,
the convergence rate is suboptimal but the ratio between total error estimator and error is close to a moderate
constant. For the American option pricing problem with initial condition u0 given by (1.3), thus almost in
3
H 2 (Ω), we obtain a suboptimal rate O(τ 3/4 ) for both the total error estimator and the time error estimator.
Because of the expected asymptotic solution behavior at t → 0, we can compensate for this with an (a priori)
graded time partition, and thereby obtain an error estimator which converges with optimal rate O(τ + h); see
Section 5.4. The same strategy can be used in several space dimensions d > 1 for a continuous, piecewise smooth
payoff function χ(x). In the case when the space mesh does not resolve the obstacle χ in the contact region, we
observe that both the obstacle consistency error estimator and true error decay with the same suboptimal rate.
Using appropriate local refinement, we can recover the optimal rates for both true error and error estimator;
see Section 5.3.
The outline of the paper is as follows. In Section 2 we introduce the parabolic variational inequality and the
corresponding discrete problem. Since the underlying operator is angle-bounded, we review the definition and
properties as well. We also define the discrete non-contact set and full-contact set, and the discrete Lagrange
multipliers which play a key role in the construction of the error estimator. In Section 3 we present the error
estimator for a time independent piecewise linear obstacle χ and prove that it gives an upper bound of the
actual error in L2 (0, T ; H 1 (Ω)). We also prove that the space and time error estimator terms are lower bounds
of the actual error. In Section 4 we consider obstacles χ that are not contained in the finite element space and
give an upper bound with additional terms resulting from the obstacle approximation, the so-called obstacle
consistency terms. In Section 5 we present several numerical results for d = 1 and d = 2 including American
option problem, and compare the behavior of error and estimators. We finally summarize our conclusions in
Section 6.

2. Variational formulation
Throughout this paper we use the following notation. For ω ⊂ Ω, we denote by ·, ·ω the inner product in
1
L2 (ω) or the duality pairing between H −1 (ω) and H01 (ω), and by · ω := ·, ·ω2 the corresponding norm. We
also employ the energy norm |||·|||ω induced by the operator L as well as its dual norm |||·|||∗,ω ; see (2.3). We omit
 K.-S. MOON ET AL. 489

the subscript ω provided ω = Ω and no confusion arises. Finally, the symbol A
B means A ≤ CB with a
constant C independent of space meshsize h and time step τ , and A ≈ B abbreviates A
B
A.

2.1. Continuous and discrete problems

Let Ω be an open bounded polyhedron domain in Rd and Q := (0, T ) × Ω be the parabolic cylinder with
lateral boundary Γ := (0, T ) × ∂Ω. Consider an obstacle χ ∈ H 1 (Q) such that χ ≤ 0 on Γ and nonempty convex
sets
K(t) := {v ∈ H01 (Ω) : v ≥ χ(t)} a.e. t ∈ [0, T ], (2.1)
1
where H0 (Ω) is the usual Sobolev space of function with square integrable weak derivatives and vanishing trace.
If H −1 (Ω) denotes the dual space of H01 (Ω), we consider the linear operator L : H01 (Ω) → H −1 (Ω) given by

Lv := −div(µ2 ∇v) + b · ∇v + cv (2.2)

with coefficients µ2 ∈ L∞ (Ω), b ∈ [W 1,∞ (Ω)]d , c ∈ L∞ (Ω) satisfying

1
µ2 (x) ≥ µ20 > 0, b(x) ≤ b0 , ρ(x) := − div b(x) + c(x) ≥ ρ20 ≥ 0 a.e. x ∈ Ω,
2
for some nonnegative constants µ0 , b0 , ρ0 . Such an operator induces the energy norm

  12
1
2 2 2
|||v|||ω := Lv, vω =
2
µ |∇v| + ρ|v| dx ∀v ∈ H01 (ω), (2.3)
ω

as well as the dual norm |||·|||∗,ω := supv∈H01 (Ω) ·, vω / ||v|||ω . It is easy to check that |||·|||ω and |||·|||∗,ω are equivalent
to the norms in H01 (ω) and H −1 (ω), respectively. The operator L also gives rise to the continuous and coercive
bilinear form a(·, ·) : [H01 (Ω)]2 → R defined by

a(v, w) := µ2 ∇v · ∇w + (b · ∇v)w + cvw ∀v, w ∈ H01 (Ω).
Ω

We then consider the following continuous parabolic variational inequality:

Problem 2.1. Let K(t) be the convex sets defined in (2.1) and u0 ∈ K(0). Given data f ∈ L2 (0, T ; L2(Ω)),
find u ∈ C([0, T ]; L2 (Ω)) ∩ L2 (0, T ; H 1 (Ω)) such that

∂t u, u − v + a(u, u − v) ≤ f, u − v, ∀v ∈ K(t) a.e. t ∈ [0, T ] (2.4)

and
u(t) ∈ K(t) a.e. t ∈ [0, T ]
with initial condition u(0, x) = u0 (x) in Ω and boundary condition u(t, x) = 0 on Γ for a.e. t ∈ [0, T ] (see [4],
Chap. III).
For the numerical treatment of this problem, we discretize the spatial domain Ω into simplexes S ∈ Th , and
partition the time domain [0, T ] into N subintervals, i.e. 0 = t0 < t1 < · · · < tN = T and let τn := tn − tn−1 .
Let hS be the diameter of S ∈ Th and h(x) be the local meshsize, that is the piecewise constant function with
h|S := hS for all S ∈ Th . Given a node z, we set hz := max{diam(S) : S ∈ Th and z ∈ S}.
Let Vh be the usual conforming piecewise linear finite element subspace of H01 (Ω) over the mesh Th . In
this paper, we assume that meshes do not change in time. Mesh adaptation as well as the design of adaptive
algorithms will be considered in [16]. Consider the corresponding discrete convex set

Khn := {v ∈ Vh : v ≥ χnh } (2.5)

490 A POSTERIORI ERROR ANALYSIS FOR PARABOLIC VARIATIONAL INEQUALITIES

where the sequence χnh ∈ Vh is a piecewise linear approximation of the obstacle χ(tn ) for 0 ≤ n ≤ N . For
example, when the obstacle χ is continuous, we could take χnh to be the piecewise linear Lagrange interpolant
of χ(tn ). Now we formulate the following fully discrete numerical approximation of Problem 2.1 by using linear
finite elements in space and backward Euler method in time:
Problem 2.2. Given the approximation Fhn ∈ L2 (Ω) of f at time tn for 1 ≤ n ≤ N , and initial guess Uh0 ∈ Kh0 ,
find an approximate solution Uhn ∈ Khn for 1 ≤ n ≤ N such that

1
U n − Uhn−1 , Uhn − vh  + a(Uhn , Uhn − vh ) ≤ Fhn , Uhn − vh , ∀vh ∈ Khn . (2.6)
τn h

This problem admits a unique solution [12]. Moreover, let {φzi }Ii=1 be the set of nodal basis functions
(here {zi }Ii=1 are interior nodes), and let A := (φzi , φzj  + τn a(φzi , φzj ))Ii,j=1 be the resulting matrix of (2.6).
I
If U = (Ui ), X = (Xi ) ∈ RI are the vector of nodal values of Uhn and χnh , namely Uhn = i=1 Ui φzi and

Xhn = Ii=1 Xi φzi , and F = (Uhn−1 + τn Fhn , φzi )Ii=1 = (Fi ), then U satisfies the complementarity system:
T
AU ≥ F , U ≥ X, AU − F U − X = 0. (2.7)

Therefore, the equality (AU )i = Fi holds at every node zi where Ui > Xi . This leads to the notion of non-contact
node defined in Section 2.3.

2.2. Angle-bounded operators

A linear monotone operator L : H01 (Ω) → H −1 (Ω) is called sectorial if it satisfies the strong sector condition
2 2 2
Lv, w ≤ 4γ 2 |||v||| |||w||| ∀v, w ∈ H01 (Ω). (2.8)

This is equivalent to the following inequality for the skew-symmetric part of L [5], Proposition 11:

Lv, w − Lw, v ≤ 2λ |||v||| |||w||| ∀v, w ∈ H01 (Ω), (2.9)

with a positive constant λ satisfying γ 2 = (λ2 +1)/4. We observe that (2.8) implies that L is Lipschitz continuous
and |||Lv|||∗ = supw∈H01 (Ω) Lv, w/ |||w||| satisfies

1 2 2 2
|||Lv|||∗ ≤ |||v||| ≤ |||Lv|||∗ ∀v ∈ H01 (Ω).
4γ 2

We are now ready to recall the notion of angle-bounded operator. This was introduced by Brézis and Browder [5]
as a nonlinear generalization of sectorial operators, and more recently revisited by Caffarelli in the context of
regularity theory [7].
Definition 2.3. Let H be a Hilbert space, and let D(A) ⊂ H be the domain of an operator A : H → 2H . Then
A is said to be γ 2 -angle-bounded if there exists a positive constant γ such that

A(v) − A(w), w − z ≤ γ 2 A(v) − A(z), v − z ∀v, w, z ∈ D(A). (2.10)

Lemma 2.4 (equivalence). The conditions (2.8) and (2.10) are equivalent for L linear.
Proof. We simply set ṽ = v − z and w̃ = w − z in (2.10) to get the equivalent formulation (we omit the tildes)

Lv, w ≤ γ 2 Lv, v + Lw, w ∀v, w ∈ D(L). (2.11)

Then replace v by αv with α ∈ R and argue with the resulting quadratic inequality in α to realize that (2.8)
and (2.11) are equivalent. 
 K.-S. MOON ET AL. 491

Lemma 2.5 (L is sectorial). The operator L defined in (2.2) is sectorial with constant γ given by


1 b2
γ2 = 1+ 2 2 0 2 2 , (2.12)
4 µ0 (ρ0 + µ0 P0 )

where P0 is the constant in the Poincaré inequality P0 v ≤ ∇v for all v ∈ H01 (Ω).
Proof. This proof hinges on (2.9). Since
2 2
|||v||| ≥ µ20 ∇v 2 , |||v||| ≥ (ρ20 + µ20 P02 ) v 2
∀v ∈ H01 (Ω),

the skew-symmetric part of L satisfies for all v, w ∈ H01 (Ω)

2b0
Lv, w − Lw, v ≤ b0 ∇v w + b0 ∇w v ≤ 1 |||v||| |||w||| ,
µ0 (ρ20 + µ20 P02 ) 2
1
whence λ = b0 µ−1 2 2 2 −2
0 (ρ0 + µ0 P0 ) . The assertion (2.12) follows from γ 2 = (1 + λ2 )/4 [5], Proposition 11. 
If K = K(t) is the convex set defined in (2.1), we let A : K → H −1 (Ω) be the multivalue operator associated
with the variational inequality in K, i.e.

v∗ ∈ A(v) ⇔ a(v, v − w) ≤ v∗ , v − w ∀w ∈ K. (2.13)

If we further define the multivalue operator σ(v) := A(v) − Lv with D(σ) = K, we see that σ(v) ≤ 0 in Ω and
σ(v) = 0 in N = {v > χ(t)} (simply argue with w = v + ϕ). It turns out that σ is the subdifferential ∂IK of the
indicator function IK of K, which is defined as zero in K and ∞ in the complement. Such a σ can be viewed as
a Lagrange multiplier of the constraint v ≥ χ(t), an interpretation that we exploit in Section 2.4.
Lemma 2.6 (A is angle-bounded). The operator A = L + σ is γ02 -angle-bounded with constant γ0 = max(1, γ),
γ being defined in (2.12). Moreover, A satisfies for all v, w, z ∈ K

A(v) − A(w), w − z ≤ γ 2 Lv − Lz, v − z + σ(v), v − z

(2.14)
≤ γ 2 Lv − Lz, v − z + σ(v) − σ(z), v − z.

Proof. Since A(v) = Lv + σ(v), in view of Lemmas 2.4 and 2.5 we only need to deal with σ(v). We resort to
the fact that σ(v) = ∂IK (v), which translates into the property

σ(v), w − v ≤ 0 ∀v, w ∈ K.

In fact, if v > χ(t) then σ(v) = 0 whereas if v = χ(t) ≤ w then σ(v) ≤ 0. Consequently

σ(v) − σ(w), w − z = σ(v), v − z + σ(v), w − v + σ(w), z − w ≤ σ(v), v − z ≤ σ(v) − σ(z), v − z,

whence we deduce (2.14)

A(v) − A(w), w − z ≤ γ 2 Lv − Lw, w − z + σ(v) − σ(z), v − z ≤ γ02 A(v) − A(z), v − z.

The last inequality implies that A is γ0 -angle-bounded, as asserted. 

We conclude this section with the coercivity Lemma 4.3 of [18], which will be crucial in Section 3.1.
Lemma 2.7 (coercivity). Let γ be defined in (2.12). The linear operator L defined in (2.2) satisfies

2 1 2 2
Lv − Lw, w − z ≤ 2γ 2 |||v − z||| − |||v − w||| + |||z − w||| ∀ v, w, z ∈ K. (2.15)
4
492 A POSTERIORI ERROR ANALYSIS FOR PARABOLIC VARIATIONAL INEQUALITIES

Proof. In view of the definition (2.3) of |||·|||, Lemma 2.5, and Cauchy-Schwarz inequality we get

Lv − Lw, w − z = Lv − Lw, w − v + Lv − Lw, v − z

≤ − |||v − w|||2 + 2γ |||v − w||| |||v − z|||
1
≤ − |||v − w|||2 + 2γ 2 |||v − z|||2 .
2
Similarly, we get
Lv − Lw, w − z = Lz − Lw, w − z + Lv − Lz, w − z
2
≤ − ||z − w||| + 2γ |||v − z||| |||w − z|||
1 2 2
≤ − |||z − w||| + 2γ 2 |||v − z||| .
2
Adding the last two inequalities gives (2.15). 

2.3. Residuals and discrete sets

For any sequence {W n }N
n=1 , we define the piecewise constant interpolant W and the piecewise linear inter-
polant W as

W (t) := W n , W (t) := l(t)W n−1 + (1 − l(t))W n ∀t ∈ (tn−1 , tn ], 1 ≤ n ≤ N, (2.16)

where the linear function l(t) is defined by

tn − t
l(t) := ∀t ∈ (tn−1 , tn ]. (2.17)
τn

We also denote by {δW n }N

n=1 the discrete derivative of the sequence {W }n=1
n N

W n − W n−1
δW n := ∀1 ≤ n ≤ N. (2.18)
τn

For a function w continuous in time, we let wn (·) := w(tn , ·) be its semi-discrete approximation and w be defined
as in (2.16).
We now introduce spatial quantities for 1 ≤ n ≤ N fixed. We define the jump residual Jhn on the side S1 ∩ S2
to be
Jhn := −µ2 (∇Uhn |S1 · ν1 + ∇Uhn |S2 · ν2 ) (2.19)
where νi is the unit outer normal vector to the element Si ∈ Th for i = 1, 2. We also define the interior residuals
associated with the linear operator L for each element of Th

hn := Fhn − δUhn − b · ∇Uhn − cUhn ,

R hn + ∇(µ2 ) · ∇Uhn .
Rhn := Fhn − δUhn − LUhn = R (2.20)

We observe that Rhn is the usual full residual, including the second order term div(µ2 ∇Uhn ) = ∇(µ2 ) · ∇Uhn , and
that Rn = Rn provided that µ is constant.
h h
Our next task is to define discrete sets that mimic the contact and non-contact sets, C and N , of (1.8). This
is a rather tricky endeavor because our choice is related to the definition of discrete multiplier and localization.
We adopt the definition of Fierro and Veeser [11], which requires dealing with stars ωz . They are the support
of the basis functions {φz }z∈Ph , where Ph is the set of all nodes of Th , including the boundary nodes. We let
γz be the skeleton of ωz , namely the set of all interior sides of ωz which contain z; for d = 1 we have γz = {z}
and so v γz reduces to the absolute value of v(z). We then split Ph into three disjoint sets

Ph = Nhn ∪ Chn ∪ Fhn

 K.-S. MOON ET AL. 493

with the non-contact nodes Nhn , full-contact nodes Chn , and free boundary nodes Fhn defined as follows [11]:

Nhn := {z ∈ Ph : Uhn > χnh in int ωz }, (2.21a)

Chn := {z ∈ Ph : Uhn = χnh and Rhn ≤ 0 in ωz , Jhn ≤ 0 on γz }, (2.21b)
Fhn := Ph \ (Nhn ∪ Chn ). (2.21c)

The definition of Nhn is clear since z ∈ Nh if and only if Uhn (z) > χnh (z). In contrast, the definition of Chn
deserves an explanation. It is motivated by localization, that is the fact that modifying Uhn locally in ωz should
not improve the resolution, provided χnh = χ(tn ) in ωz . In fact, if Kzn := {v + Uhn : 0 ≤ v ∈ H01 (ωz )} is the
convex set of local admissible functions, then we consider the variational inequality problem
1
Find V ∈ Kzn : V − Uhn−1 , V − W  − a(V, V − W ) ≤ Fhn , V − W  ∀W ∈ Kzn .
τn
We claim that the unique solution of this local problem is V = Uhn and thus local improvement of Uhn is no
longer possible. If v = 0, then the above variational inequality becomes
1
Fhn , w − U n − Uhn−1 , w + a(Uhn , w) ≤ 0 ∀0 ≤ w ∈ H01 (ωz ),
τn h
or, after integration by parts,

Rhn , wωz + Jhn , wγz ≤ 0 ∀0 ≤ w ∈ H01 (ωz ).

This is equivalent to the requirements Rhn ≤ 0 in ωz and Jhn ≤ 0 on γz in (2.21b).

2.4. Lagrange multipliers

We first recall the definition of Lagrange multiplier σ(u) of (1.9) and (2.13)

σ(u), ϕ := f − ∂t u − b · ∇u − cu, ϕ − µ2 ∇u, ∇ϕ ∀ϕ ∈ H01 (Ω), (2.22)

as well as the properties (1.8), namely,

σ(u) = 0 in N = {u > χ}, σ(u) = f − ut − Lu ≤ 0 in C = {u = χ}.

We note that in the non-contact region N , where the constraint is inactive, we have σ(u) = 0, whereas in the
contact region C we have σ(u) ≤ 0. We could thus interpret σ(u) as a reaction to the obstacle.
It is crucial for us to mimic these continuous
 properties at the discrete level. A first attempt would be to
define a piecewise linear function σhn = z∈Ph snz φz in such a way that the nodal values snz are weighted means
on stars ωz :    n 
 1  φz − µ2 ∇U n ∇φz , z ∈ Ph ∩ Ω
R
n φ ω h h
sz := ωz z z
(2.23)
0, z ∈ Ph ∩ ∂Ω;
note that σ is zero on ∂Ω ∩ N , which motivates us to define snz = 0 on ∂Ω. This implies
 
snz φz = Rhn φz − µ2 ∇Uhn ∇φz .
ωz ωz

This definition yields snz ≤ 0 and snz = 0 for z ∈ Nhn , and it is thus quite appropriate for Nhn but not necessarily
for z ∈ Chn . In fact, to achieve localization of the error estimator σhn must equal the linear residual in ωz for
z ∈ Chn , thereby leading to
σhn = Fhn − δUhn − LUhn
494 A POSTERIORI ERROR ANALYSIS FOR PARABOLIC VARIATIONAL INEQUALITIES

in the sense of distributions in ωz . We can blend the two competing alternatives via the partition of unity
{φz }z∈Ph as proposed by Fierro and Veeser [11] and Nochette et al. [20]

σhn , ϕ := σhn , ϕφz  ∀ ϕ ∈ H01 (Ω), (2.24)
z∈Ph

where
 
n 2
ωz Rnh ϕφz − µ ∇Uh ∇(ϕφz ) z ∈ Chn
n
σhn , ϕφz  := (2.25)
ωz sz ϕφz z ∈ Ph \ Chn .

We stress that σhn is not a discrete object, but its definition (2.24) guarantees that σhn ≤ 0 in Ω. This is a
consequence of snz ≤ 0 and the sign conditions in (2.21b).

2.5. Galerkin functionals

For a linear parabolic problem ∂t u + Lu = f it is natural to use the distribution f − δU − LU as a defect
measure of the error. This so-called residual fails to be the correct quantity for variational inequalities because it
provides wrong information in the contact region. The Lagrange multiplier σ = σ(u) of (2.22) restores equality

∂t u + Lu + σ − f, ϕ = 0 ∀ϕ ∈ H01 (Ω),

and leads to the notion (1.10) of Galerkin functional Gh , which is a nonlinear defect measure; see [11, 19, 20, 23]
for elliptic variational inequalities. The precise definition of Gh is as follows: Gh ∈ L2 (0, T ; H −1(Ω)) is the
piecewise constant functional in time given by

Gh , ϕ := ∂t Uh + LU h + σ h − F h , ϕ
= ∂t (Uh − u) + L(U h − u) + (σ h − σ), ϕ − F h − f, ϕ ∀ϕ ∈ H01 (Ω). (2.26)

Note that for 1 ≤ n ≤ N , the Galerkin functional Ghn satisfies

n , ϕφz  + µ2 ∇U n , ∇(ϕφz )

Ghn , ϕφz  = σhn − R ∀z ∈ Ph , (2.27)
h h

whence Ghn , ϕφz  = 0 for all z ∈ Chn . This crucial property is responsible for localization and will be instrumental
in Section 3.

3. A POSTERIORI error estimate: piecewise linear time-independent obstacles

In this section we derive a posteriori upper and lower bounds for the error in the L2 (0, T ; H 1 (Ω))-norm. To
simplify the presentation we assume in this section that (i) the obstacle χ is time independent, and (ii) that the
obstacle is a piecewise linear function and χ = χh . In this case we let χnh := χh and have

χ(tn ) = χnh = χh ⇒ Khn ⊂ K ∀n = 0, . . . , N. (3.1)

3.1. Upper bound

Upon choosing the function ϕ = Uh − u in (2.26), we get the error equation

1 d 2
Uh − u + L(U h − u), Uh − u + σ h − σ, Uh − u = Gh , Uh − u + F h − f, Uh − u.
2 dt
To estimate the second and third terms on the left-hand side we would like to resort to the angle-bounded
structure of A = L + σ, namely Lemma 2.6. However, this is not possible because σ h does not coincide with
 K.-S. MOON ET AL. 495

the subdifferential ∂IK (Uh ); this is the effect of both space and time discretization. The crucial monotonicity
property σ h − σ, Uh − u ≥ 0 is thus no longer valid. Therefore, we split the analysis into the linear part
involving L and the nonlinear part involving the multipliers, but still aim at an inequality similar to (2.14).
Applying Lemma 2.7, namely taking v = U h , w = u and z = Uh in (2.15), we arrive at

1 d 2 1 2 2
Uh − u + U h − u + |||Uh − u|||
2 dt 4
2
≤ 2γ 2 U h − Uh − σ h − σ, Uh − u + Gh , Uh − u + F h − f, Uh − u.

Invoking the definition of dual norm |||·|||∗ , a straightforward calculation gives

2 2 1 2
Gh , Uh − u + F h − f, Uh − u ≤ 4 |||Gh |||∗ + 4 F h − f ∗
+ |||Uh − u||| ,
8
whence
d 2 1 2 1 2
Uh − u + U h − u + |||Uh − u|||
dt 2 4 (3.2)
2 2 2
≤ 4γ 2 U h − Uh + 8 |||Gh |||∗ + 8 F h − f ∗
− 2σ h − σ, Uh − u.
On the other hand, using the strong sector condition (2.8) and the triangle inequality in (2.26) yields

2 2 2 2
|||∂t (Uh − u) + (σ h − σ)|||∗ ≤ 12γ 2 U h − u + 3 ||Gh |||∗ + 3 F h − f ∗
. (3.3)

Combining the last two inequalities, (3.2) and (3.3), and integrating on [0, T ] gives
 T
2 1 2 2
(Uh − u)(T ) + Uh − u + |||Uh − u||| dt
4 0
 T
1  2
|||∂t (Uh − u) + (σ h − σ)|||∗ dt ≤ Uh0 − u0 
2
+ (3.4)
48γ 2 0
 T
2 2
+ 4γ 2 U h − Uh − 2σ h − σ, Uh − u + λ |||Gh |||2∗ + F h − f ∗
dt,
0

1
with λ := 8 + 16γ 2 and

 T 
N
2 2
|||Gh |||∗ dt = τn |||Ghn |||∗
0 n=1
 T N 
 tn
σ h − σ, Uh − udt = σhn − σ, Uh − udt.
0 n=1 tn−1

2
Hence, to get the desired global upper bound, we need to find an appropriate upper bound of |||Ghn |||∗ and a lower
 tn
bound of tn−1 σhn − σ, Uh − u dt for 1 ≤ n ≤ N . This is achieved in the next two lemmas.

Lemma 3.1 (upper bound of Galerkin functional). Let the sequence {Uhn }N n=1 be the discrete solution of the
variational inequality (2.6) and the discrete Galerkin functional Ghn be defined as in Section 2.5. Then there is
a constant CG ≥ 0, independent of the meshsize h, such that

|||Ghn |||2∗ ≤ CG (ηzn )2 ∀n = 1, . . . , N. (3.5)
z∈Ph \Ch
n
496 A POSTERIORI ERROR ANALYSIS FOR PARABOLIC VARIATIONAL INEQUALITIES

Here the space discretization error indicator ηzn is defined by

    12
2  n n 2
ηzn := hz Jhn L2 (γz ) + h2z Rh − Rz  2 , (3.6)
L (ωz )

with Jhn and Rhn defined in (2.19) and (2.20), respectively, and Rzn given by
 1 
|ωz | ωz Rh , z ∈ Ph ∩ Ω
n
 n
Rz := (3.7)
0, z ∈ Ph ∩ ∂Ω.

Proof. We employ the partition of unity to write Ghn , ϕ = z∈Ph Ghn , ϕφz  for any function ϕ ∈ H01 (Ω). We
next recall (2.27), in particular Ghn , ϕφz  = 0 for z ∈ Chn , to deduce
  
Ghn , ϕ = µ2 ∇Uhn , ∇(ϕφz )ωz − Rhn − σhn , ϕφz ωz .
z∈Ph \Ch
n

In view of definitions (2.25) and (2.23) of σhn and snz , we see that
 
Gh , ϕφz  =
n 2 n 
µ ∇Uh · ∇(ϕφz ) − (Rh − σh )ϕφz =
n n
µ2 ∇Uhn · ∇((ϕ − ϕ  n (ϕ − ϕ
z )φz ) − R z )φz
h
ωz ωz

z is the constant
where ϕ  
ϕφz
ωz , if z ∈ Ph ∩ Ω
z :=
ϕ ω z
φz (3.8)
0, if z ∈ Ph ∩ ∂Ω.
We now recall the definitions (2.20) of R  n and Rn , and integrate by parts to infer that
h h

Ghn , ϕφz  = µ2 ∇Uhn · ∇((ϕ − ϕ z )φz ) + ∇(µ2 ) · ∇Uhn (ϕ − ϕ z )φz − Rhn (ϕ − ϕ
z )φz
ωz
  (3.9)
=− Jhn (ϕ − ϕ
z )φz − (Rhn − R n )(ϕ − ϕ
z )φz ,
z
γz ωz

because of the orthogonality relation ωz R  n (ϕ − ϕ
z )φz = 0 which, in light of (3.8), holds for arbitrary constants
z
 
Rz for z ∈ Ph ∩ Ω and Rz = 0 for z ∈ Ph ∩ ∂Ω.
n n

Standard interpolation arguments based on a trace theorem and the Bramble-Hilbert lemma yield [11, 24]
 
1
 zn 
|Ghn , ϕφz |
hz2 Jhn γz + hz Rhn − R  ∇ϕ ωz .
ωz

n , adding over
Finally, the asserted estimate (3.5) follows at once upon choosing the optimal value (3.7) for Rz
z ∈ Ph \ Ch , and using the finite overlapping property of the stars ωz .
n

 tn
The second step to prove the upper bound is to show that the term tn−1 σhn − σ, Uh − u dt admits a lower
bound. The conformity assumption Khn ⊂ K, never used before, will play an essential role now.
Lemma 3.2 (lower bound of Lagrange multipliers). Let {Uhn }N n=1 solve the discrete problem (2.6) and let the
continuous and discrete Lagrange multipliers σ and σhn be defined in (2.22) and (2.24), respectively. Then the
following inequality holds
 tn  τn 
σhn − σ, Uh − udt ≥ − σhn , (Uhn − Uhn−1 )φz  + τn snz dnz , (3.10)
tn−1 2
n n z∈Ch ∪Fh n z∈Fh
 K.-S. MOON ET AL. 497

where the constants snz are defined in (2.23) and

 
dnz := (Uhn − χnh )φz = (Uhn − χh )φz ≥ 0. (3.11)
ωz ωz

Proof. In view of (2.22) and (2.4), we have σ, u − v ≤ 0 for all v ∈ K. Since Uh (t) ∈ K a.e. t ∈ [0, T ] according
to (3.1), we obtain σ, u − Uh  ≥ 0, whence

σhn − σ, Uh − u ≥ σhn , Uh − u = σhn , (Uh − u)φz ωz .
z∈Ph

The crucial property σhn ≤ 0, in conjunction with the consequence u ≥ χ = χnh of (3.1), implies that

σhn , (Uh − u)φz  = σhn , (Uh − χh + χh − u)φz ωz ≥ σhn , (Uh − χh )φz ωz .

Using definition Uh = l(t)Uhn−1 + (1 − l(t))Uhn , with l(t) given in (2.17), and integrating in time yields
 tn
τn n
σhn , (Uh − χh )φz dt = σ , (Uhn−1 + Uhn − 2χh )φz 
tn−1 2 h
τn
= σhn , (Uhn−1 − Uhn )φz  + τn σhn , (Uhn − χh )φz .
2

We finally observe that σhn , ϕφz  = 0 for z ∈ Nhn , because snz = 0, and Uhn = χh in ωz for z ∈ Chn . Therefore
σhn , (Uhn −χh )φz  = snz dnz for z ∈ Fhn and zero otherwise, whence the desired estimate (3.10) follows immediately.


Replacing the estimates of Lemmata 3.1 and 3.2 into (3.4) yields the main result of this section.
Theorem 3.3 (upper bound: piecewise linear time-independent obstacles). Let u be the continuous solution
2 2
of (2.4) and {Uhn }Nn=1 be the discrete solution of (2.6). Let data f and u0 satisfy f ∈ L (0, T ; L (Ω)) and
u0 ∈ L2 (Ω). If σhn denotes the discrete Lagrange multiplier of (2.25), and snz , dnz indicate the constants defined
in (2.23) and (3.11), then the following a posteriori upper bound holds
 
2 1 T 2 2 1 T
2
(Uh − u)(T ) + Uh − u + |||Uh − u||| dt + |||∂t (Uh − u) + (σ h − σ)|||∗ dt
4 0 48γ 2 0
 2
≤ U 0 − u0 
h initial error
N
4 2 2
+ γ τn Uhn − Uhn−1 time error
n=1
3

N    
+ τn λCG (ηzn )2 − 2snz dnz space error
n=1 z∈Ph \Ch
n z∈Fhn


N 
+ τn σhn , (Uhn − Uhn−1 )φz  mixed error
n=1 n ∪F n
z∈Ch h
 T
2
+λ Fh − f ∗
dt data consistency
0

1
with constants γ 2 , CG given in (2.12), (3.5), λ := 8 + 16γ 2 , and space error estimator ηzn given in (3.6).
498 A POSTERIORI ERROR ANALYSIS FOR PARABOLIC VARIATIONAL INEQUALITIES

Remark 3.4 (oscillation term). The spatial data oscillation term is hidden in the second part hz Rhn − R  n ωz
z
of ηz . In the case of the Poisson equation, we have Rh = Fh if z ∈ Ph ∩ Ω, and such a term reduces to the
n n n

usual data oscillation hz Fhn − Fzn ωz . This quantity is generically of higher order than the interior residual
hz Fhn ωz , as can be seen in the numerical experiments of Section 5; see also [11, 20].
Remark 3.5 (localization). The space error indicator ηzn is fully localized, i.e., there is no contribution from
z ∈ Chn , the discrete contact set. Note that this is consistent with the absence of obstacle approximation error
because χh = χ. Likewise, the term −snz dnz ≥ 0 contributes only when z ∈ Fhn . This extends the localized
elliptic estimates of [11, 20] to the parabolic setting. One may also wonder whether the sets of full-contact
nodes Chn and free boundary nodes Fhn are good approximations of the actual contact region and free boundary,
respectively. We will explore this point further via numerical experiments in Section 5.
Remark 3.6 (mixed error term). The mixed error estimator mimics the nonlinear term in (2.14), and combines
effects of both time and space discretization. For z ∈ Chn , this term contributes solely where Uhn = Uhn−1 , that
is close to the discrete free boundary Fhn because χh (t) ≡ χh , for t ∈ [0, T ]. The numerical results in Section 5
show that this term is relatively small and decays with higher order than other terms of the proposed space
and time error estimators.

3.2. Lower bound

We first derive the lower bound of the local error based on our space error indicator ηzn . Since localization
entails ηzn = 0 for z ∈ Chn , the contact set, we just have to consider z ∈ Ph \ Chn .
Lemma 3.7 (lower bound for space error). Under the same assumptions as in Theorem 3.3, we have the
following local lower bound

h1/2 Jhn γz
γ |||Uhn − u|||ωz + ||∂t (Uh − u) + (σhn − σ)|||∗,ωz+ xn ωx + ||Fhn − f |||
hx Rhn − R
z ∗,ωz (3.12)
x∈(Ph \Ch
n )∩ω
z

for any z ∈ Ph \ Chn and n = 1, . . . , N .

Proof. Let z ∈ Ph \ Chn and ϕ ∈ H01 (ωz ). In view of (3.9) we can write

  
Ghn , ϕ = − Jhn (ϕ − ϕ
x )φx + xn )(ϕ − ϕ
(Rhn − R x )φx .
x∈(Ph \Ch
n )∩ω γx ωx
z

At the same time, the definition (2.26) of Galerkin functional together with (2.8) gives

|||Ghn |||∗,ωz ≤ 2γ |||Uhn − u|||ωz + |||∂t (Uh − u) + (σhn − σ)|||∗,ωz + |||Fhn − f |||∗,ωz .

We now proceed as Fierro and Veeser [11] to construct an explicit test function ϕ to plug into Ghn , ϕ; see also
Verfürth [24]. Let E ⊂ γz be a generic edge and let ωE denote the union of the elements S ∈ Th sharing E.
Consider the bubble functions  
ψE := φx , ψS := φx ; (3.13)
x∈Ph ∩E x∈Ph ∩S
note that ωE := supp(ψE ). Let ζ be defined by

ζ = ψE − αS,x ψS φx
x∈Ph ∩S, S⊂ωE

with coefficients αS,x ∈ R given by the linear system

  
αS,x ψS φx φy = ψE φy ∀ y ∈ Ph ∩ S.
x∈Ph ∩S S S
 K.-S. MOON ET AL. 499

It is easy to see that supp ζ = ωE , and that ωy
ζφy = 0 for all y ∈ Ph . Moreover, ζ satisfies

∇ζ ωE + h−1
z ζ ωE ≤ h−1/2
z ψE E,

because ζ|E = ψE . We finally set ϕ := Jhn |E ζ and observe that [24], Lemma 3.3, implies
   
Jhn 2E
Jhn Jhn ψE φx = Jhn ϕφx ,
x∈Ph \Ch
n E x∈Ph \Ch
n E

whence
 
Jhn 2
E
−Ghn , ϕωE − xn )(ϕ − ϕ
(Rhn − R x )φx
x∈(Ph \Ch
n )∩ω ωx
E


|||Ghn |||∗,ωE |||ϕ|||ωE + xn
hx Rhn − R ωx ∇ϕ ωE
x∈(Ph \Ch
n )∩ω
E


|||Ghn |||∗,ωE + zn
hx Rhn − R ωx h−1/2
z Jhn E.
x∈ (Ph \Ch
n )∩ω
E

Consequently, we get the desired result. 

Remark 3.8 (local implies global). Since the local lower bound (3.12) contains local negative norms, and
negative norms are nonlocal, it is not clear that squaring and adding (3.12) would lead to a global lower bound.
This is indeed the case as the following argument shows. Let Λ ∈ H −1 (Ω), and

ϕz ∈ H01 (ωz ) : ∇ϕz ωz = Λ ∗,ωz , Λ 2

∗,ωz = Λ, ϕz  ∀z ∈ Ph .
  
Then ϕ = z∈Ph ϕz ∈ H01 (Ω) and ∇ϕ 2Ω
z∈Ph ∇ϕz 2
ωz = z∈Ph Λ 2
∗,ωz because of the finite overlap-
ping property of the stars {ωz }z∈Ph . Hence
   1/2
2 2
Λ ∗,ωz = Λ, ϕz  = Λ, ϕ ≤ Λ ∗,Ω ∇ϕ Ω ≤ Λ ∗,Ω Λ ∗,ωz ;
z∈Ph z∈Ph z∈Ph

compare with [2], Proposition 3.7, for the heat equation. The reversed inequality is in general not valid.
The following lower estimate is rather elementary due to the structure of the energy error in Theorem 3.3,
which involves both U h and Uh . This is to be compared with [2], Proposition 3.6, for the heat equation.
Lemma 3.9 (lower bound for time error). The following local lower bound holds


N  T
2 2
τn Uhn − Uhn−1 ≤6 Uh − u + |||Uh − u|||2 dt.
n=1 0

2  tn 2  tn 2
Proof. We have τn Uhn − Uhn−1 =3 tn−1
U h − Uh dt ≤ 6 tn−1
Uh − u + |||Uh − u|||2 dt because of
definition (2.16) of U h and Uh , and the triangle inequality. 

4. A POSTERIORI error estimate: general obstacles

For option pricing problems, we need to consider obstacles which are neither piecewise constant nor time-
independent, as in the previous section. Although one can always use the transformation v = u − χ and solve
the parabolic variational inequality for v, this does not avoid dealing with the obstacle χ especially in defining
500 A POSTERIORI ERROR ANALYSIS FOR PARABOLIC VARIATIONAL INEQUALITIES

tn−1 tn

χ
Uhn−1 χh
Uhn Uh

χn
h

χn−1
h

Figure 1. Obstacle consistency: If the obstacle χ and its space-time piecewise linear approx-
imation χh do not coincide in ωz × (tn−1 , tn ) for nodes z ∈ Ph \ Nhn , then the quantities
σhn , (χ − Uh )+ φz  and σhn , (χh − χ)+ φz  measure the local lack of conformity. Note that these
quantities vanish for z ∈ Nhn , that is for the non-contact nodes.

the contact region. In this section, we consider the case of general obstacles χ ∈ H 1 (Q) directly. This extension
affects only Lemma 3.2, which must be revisited within the present
 tn context.
Therefore, in what follows we derive a lower bound for tn−1 σhn − σ, Uh − u dt. To this end, we let
χh = l(t)χn−1
h + (1 − l(t))χnh ∈ C([0, T ]; Vh (Ω)) be a space-time piecewise linear approximation of χ, and
observe that in general χh (t)
χ(t) for 0 ≤ t ≤ T . To handle this lack of consistency, we follow Veeser [23] and
introduce the auxiliary function Uh∗ := max(Uh , χ) ∈ K. Since σ, u − Uh∗  ≥ 0, we have that

σhn − σ, Uh − u ≥ σhn , Uh∗ − u + σ − σhn , Uh∗ − Uh . (4.1)

For the first term on the right-hand-side of (4.1), we invoke σhn ≤ 0 and σhn , (χ − u)φz  ≥ 0 to obtain
 
σhn , Uh∗ − u = σhn , (Uh∗ − u)φz  ≥ σhn , (Uh∗ − χ)φz 
z∈Ph z∈Ph
 (4.2)
= σhn , (Uh − χh )φz  + σhn , (Uh∗ − Uh )φz  + σhn , (χh − χ)φz .
z∈Ph

Arguing as in the proof of Lemma 3.2 with the first term on the right-hand side, we deduce

  tn  τn 
σhn , (Uh − χh )φz dt = − σ n , (Uhn − Uhn−1 ) − (χnh − χn−1 ) φz  + τn snz dnz .
tn−1 2 h h
z∈Ph z∈Ph z∈Ph \Ch
n

This step corresponds to that of Lemma 3.2; in fact the first term on the right-hand side is the most general
form of the mixed error in Theorem 3.3. However, we now have two additional terms in (4.2) that account for
the obstacle inconsistent approximation, as illustrated in Figure 1. To bound them we utilize the definition of
Uh∗ , which results in Uh∗ − Uh = (χ − Uh )+ , as well as σhn ≤ 0, and end up with

σhn , (Uh∗ − Uh )φz  ≥ σhn , (χ − Uh )+ φz , σhn , (χh − χ)φz  ≥ σhn , (χh − χ)+ φz .

We can also rewrite the second term on the right-hand side of (4.1) as follows:

σ − σhn , Uh∗ − Uh  = ∂t (u − Uh ) + (σ − σhn ), (χ − Uh )+  − ∂t (u − Uh ), (χ − Uh )+ .

 K.-S. MOON ET AL. 501

The second term on the right-hand side is most problematic. We handle it via integration by parts in time:
 T T
 T
+ +
− ∂t (u − Uh ), (χ − Uh )  = −u − Uh , (χ − Uh )  + u − Uh , ∂t (χ − Uh )+ dt.
0 0 0

Note that we can eliminate the first term on the right-hand side at t = 0 because if χ0 (x) > Uh0 (x) then
u0 (x) ≥ χ0 (x) > Uh0 (x) whence u0 − Uh0 , (χ0 − Uh0 )+  ≥ 0.
After applying the Cauchy-Schwarz inequality three times, we arrive at
 T 
N  τn 
σ h − σ, Uh − udt ≥ − σhn , (Uhn − Uhn−1 ) − (χnh − χn−1
h ) φz  − τn snz dnz
0 n=1 z∈Ph
2 n
z∈Fh
  T
+ σ h , (χ − Uh )+ + (χh − χ)+ φz dt
z∈Ph 0
 T  T
ε1 1 2
− |||∂t (u − Uh ) + (σ − σ h )|||2∗ dt − (χ − Uh )+ dt
2 0 2ε1 0
ε2 1
− (u − Uh )(T ) 2 − (χ − Uh )+ (T ) 2
2 2ε2
 T
ε3 2 1 2
− |||u − Uh ||| + ∂t (χ − Uh )+ ∗ dt,
0 2 2ε 3

1 1
with ε1 , ε2 , ε3 > 0 arbitrary. We finally choose ε1 = 96γ 2 , ε2 = 2 and ε3 = 18 , and insert the above estimate
into (3.4) to obtain the following upper bound.
Theorem 4.1 (upper bound: general obstacles). With the same notation as in Theorem 3.3, but general
obstacle χ ∈ H 1 (Q), we have the following upper a posteriori bound
 T  T
2 1 2 1 2 1 2
(Uh − u)(T ) + Uh − u + |||Uh − u||| dt + |||∂t (Uh − u) + (σ h − σ)|||∗ dt
0 2 4 48γ 2 0
 2
≤ 2 Uh0 − u0  initial error
8γ 2 
N
2
+ τn Uhn − Uhn−1 time error
3 n=1

N  
N 
+ 2λCG τn (ηzn )2 −4 τn snz dnz space error
n=1 z∈Ph \Ch
n n=1 z∈Fhn


N 
+2 τn σhn , (Uhn − Uhn−1 ) − (χnh − χn−1
h ) φz  mixed error
n=1 n ∪F n
z∈Ch h
 T
+ 2 2 2
+ 4 (χ − Uh ) (T ) + 192 γ 2 (χ − Uh )+ + 16 ∂t (χ − Uh )+ ∗
dt obstacle consistency 1
0

N   tn
−4 σhn , {(χ − Uh )+ + (χh − χ)+ }φz  dt obstacle consistency 2
n=1 n ∪F n
z∈Ch tn−1
h
 T
2
+ 2λ Fh − f ∗
dt. data consistency
0
502 A POSTERIORI ERROR ANALYSIS FOR PARABOLIC VARIATIONAL INEQUALITIES

Remark 4.2 (obstacle consistency 1). Terms involving (χ − Uh )+ are only active away from the non-contact
set, a crucial localization property, and account for the lack of constraint consistency Uh < χ in both space and
time; see Figure 1.
Remark 4.3 (obstacle consistency 2). The space-time situation χh > χ, depicted in Figure 1, is only detected
by the term σhn , (χh − χ)+ φz . In particular, if z ∈ Chn is a full-contact node, then this is the only nonzero
local indicator. Besides justifying its presence, this argument shows that such a term can be regarded as a
complement to the notion of full contact nodes which hinges on the condition χnh = χn in ωz ; see Section 2.3.
For a kink or cusp pointing downwards the relation χh > χ is not only to be expected but it might suggest that
one needs strong local refinement. This is not true because asymptotically the discrete solution detaches from
the obstacle and so σhn , (χh − χ)+ φz  = 0; see [19] for a full discussion.
Remark 4.4 (inactive constraint). For the non-contact nodes Nhn , the variational inequality becomes an
equality. This is reflected on the vanishing of all terms that account for the unilateral constraint. The resulting
estimator reduces to an energy-type estimator for a linear diffusion equation. It is however different from earlier
versions [2,21,25] in that we develop star-based error indicators, the interior residual is of higher order, and the
linear sectorial operator L is more general than the Laplacian.
Remark 4.5 (lower bounds). The lower bounds proved in Lemmas 3.7 and 3.9 are still valid in this context.
The additional terms accounting for non-conformity can all be justified, as explained in Remarks 4.2 and 4.3,
except for |||∂t (χ − Uh )+ |||∗ .

5. Numerical experiments
In this section we present several numerical examples computed with the adaptive finite element toolbox
ALBERTA of Schmidt and Siebert [22]. The experiments are meant to demonstrate the localization property,
as well as to check the asymptotic convergence rates, of the proposed error estimators with respect to both
time and space discretization. We thus consider mostly uniform partitions in time and space, the latter made of
triangular elements. We do not attempt here to use our estimators to drive an adaptive algorithm, which is the
subject of current research. We provide instead computational evidence that our estimators possess essential
information for adaptivity.

5.1. Implementation
In all our numerical experiments, obstacles χ are piecewise smooth functions. We take χh and Uh0 to be
piecewise linear interpolants of χ and u0 , respectively, whereas Fhn is an approximate L2 projection of f (·, tn )
into the space of piecewise linears. We note that the free boundary is generally unknown, and so free boundary
points are not mesh nodes usually.
In each time step, we use the projective SOR method (PSOR) [9, 12] to solve the resulting finite dimensional
linear complementary problem (2.7). The termination condition for the PSOR method is to make the relative
residual less than 10−8 .
We denote by E the total error estimator
1
E := Eh2 + Eτ2 + Eτ2h + Eχ2 2
,

where Eh , Eτ , Eτ h , and Eχ are the space, time, mixed, and obstacle consistency error estimators
1
2 2 2
Eh := Eh,1 + Eh,2 + Eh,3 2

 N  12
 2
Eτ := τn Uhn − Uhn−1
n=1
 K.-S. MOON ET AL. 503
⎛ ⎞ 12

N 
Eτ h := ⎝ τn σhn , (Uhn − Uhn−1 ) − (χnh − χn−1
h ) φz ⎠
n=1 n ∪F n
z∈Ch h
⎛ ⎞ 12
 T 
N   tn
2
Eχ := ⎝ (χ − Uh )+ dt − σhn , {(χ − Uh )+ + (χh − χ)+ }φz  dt⎠ ,
0 n ∪F n
n=1 z∈Ch tn−1
h

N n 1/2
with Eh,i := n=1 τn Eh,i for i = 1, 2, 3, and

   2 
2  zn 
n
Eh,1 := hz Jhn γz
n
Eh,2 := h2z Rhn − R  n
Eh,3 := − snz dnz .
ωz
z∈Ph \Ch
n z∈Ph \Ch
n z∈Fhn

In the experiments below we examine the relative sizes of these four quantities. We also take all constants in
the proposed error estimators to be 1; hence no ad-hoc scaling is used. We employ the notation:

e = energy error u − U L2 (0,T ;H 1 (Ω))

EOC = experimental order of convergence (based on last two experiments)
N = number of time steps
DOF = number of degrees of freedom in space.

Remark 5.1 (nonhomogeneous Dirichlet conditions). Since u = g = 0 on Γ = ∂Ω × (0, T ) in our experiments,

we take U (z) = g(z) at boundary nodes z ∈ Ph ∩ ∂Ω. For d = 1 and d = 2 with piecewise linear g, this is
equivalent to the problem with zero boundary data. Otherwise, in addition to the above estimators, there is an
interpolation error which is neglected because its order is higher than O(h) for smooth g.
Remark 5.2 (negative norm estimators). We implement all the estimators in Theorem 4.1, except for
|||∂t (χ − Uh )+ |||∗ and F h − f ∗ . We would expect the first one to be of at least the same order as |||(χ − Uh )+ |||
(see example 5.5 for numerical evidence) and the second term to be of higher order than O(h). In fact, since f
is always Lipschitz continuous across the free boundary and smooth elsewhere, we expect its L2 -projection F h
with (7th order) quadrature into the space of piecewise linears to superconverge in H −1 (Ω).
Remark 5.3 (nodal-based space error indicator). The space error estimator Eh can be written as Eh =
N 
z∈Ph τn Ez , where the nodal-based space error indicator Ez is given by
n n
n=1

⎧  2 1
⎪
⎪ 2  n  2
⎪ hz Jhn + h2z Rhn − Rz − snz dnz z ∈ Fhn
⎨ γz
 2 12
ωz
Enz := 2  zn 
⎪
⎪ hz Jhn + h2z Rhn − R  z ∈ Nhn
⎪
⎩
γz
ωz
0 z ∈ Chn .

We see that Enz vanishes at full-contact nodes z ∈ Chn , the crucial localization property alluded to in Remark 3.5.
On the other hand, this estimator disregards the case χh = χ, which is taken care of by the obstacle consistency
estimator Eχ .

5.2. 1d tent obstacle: case χh = χ

2
We take L := − ∂x
∂
2 , the domain Ω := [−1.0, 1.0], the time interval [0.5, 1.0], and the noncontact and contact

sets to be N := {|x| > t/6} and C := {|x| ≤ t/6}. If the obstacle is χ(x) = 1 − 3|x|, then the exact solution u
504 A POSTERIORI ERROR ANALYSIS FOR PARABOLIC VARIATIONAL INEQUALITIES

Figure 2. 1d tent obstacle: Exact solution u(·, t) at times t = 0.5, 0.75, 1.0. The obstacle χ is
piecewise linear with a kink at x = 0, belonging to all partitions. This implies χh = χ.

Table 1. 1d tent obstacle problem (χh = χ): The space and time estimators Eh , Eτ , decrease
with optimal order 1, but the mixed estimator Eτ h is of higher order. The ratio between total
estimator E and energy error e is quite stable and of moderate size.

N DOF Eh Eτ Eτ h E e E/e
64 127 2.256e+0 2.121e+0 2.731e–2 3.097e+0 7.347e–1 4.219
128 255 1.138e+0 1.059e+0 9.686e–3 1.555e+0 3.700e–1 4.202
256 511 5.716e–1 5.294e–1 3.338e–3 7.791e–1 1.857e–1 4.202
512 1023 2.864e–1 2.646e–1 1.181e–3 3.900e–1 9.301e–2 4.184
1024 2047 1.434e–1 1.323e–1 4.148e–4 1.951e–1 4.655e–2 4.184
EOC 0.998 1.000 1.510 0.999 0.999 –

and forcing function f are given by


36t−2 x2 − (3 + 12t−1 )|x| + 2 in N
u(x, t) =
1 − 3|x| in C,

−12t−2 (6t−1 x2 − |x| + 6) in N
f (x, t) =
−72t−2 in C.

Function u is depicted in Figure 2 at times t = 0.5, 0.75, and 1.0.

To test the asymptotic convergence rates of both the proposed error estimator E and exact error e, we halve
time step τ and space meshsize h in each experiment and report the results in Table 1 and Figure 3. To
investigate the decay of each component Eh,i of the space estimator Eh we fix the time-step to be τ = 2.5 × 10−4 ,
so small that the error is dominated by the space discretization. Table 2 displays their behavior under uniform
space refinement: the nodal-based estimator Eh,1 exhibits optimal order 1 and dominates the other two terms.
We display in Figure 4 the nodal-based space error estimator Enz at different stages tn = 0.6, 0.8, 1.0 of the
evolution. We see that Enz vanishes at full-contact nodes z ∈ Chn , as predicted by theory, and that the exact
free-boundary is captured within one element. This is further documented in Table 3 which shows exact and
approximate free boundary locations at times tn = 0.6, 0.8, 1.0.

5.3. 1d tent obstacle: case χh = χ

In general, we cannot expect the underlying mesh to match the singular behavior of the obstacle, as in
Example 5.2, even for piecewise linear obstacles. This happens, for instance, when the obstacles change in time.
The question thus arises whether or not the proposed error estimator E is able to capture the correct behavior
of the solution when a singularity is not resolved by the mesh.
 K.-S. MOON ET AL. 505

Table 2. Decay of each component Eh,i of the space estimator Eh for a fixed time-step τ =
2.5 × 10−4 so small that the time estimator Eτ is insignificant. Left: 1d tent obstacle problem;
Right: 2d oscillating moving obstacle problem. In both cases the nodal-based estimator Eh,1
exhibits the expected order 1 whereas the other two superconverge.

1d tent obstacle example (χh = χ) 2d oscillating moving circle example

DOF Eh,1 Eh,2 Eh,3 DOF Eh,1 Eh,2 Eh,3
129 2.282 3.034e–1 4.929e–2 145 1.094 1.323 1.194e–2
257 1.144 1.073e–1 1.823e–2 545 5.660e–1 4.974e–1 3.936e–3
513 5.729e–1 3.792e–2 6.295e–3 2113 2.880e–1 1.817e–1 1.368e–3
1025 2.866e–1 1.341e–2 2.250e–3 8321 1.453e–1 6.532e–2 4.652e–4
2049 1.434e–1 4.740e–3 7.903e–4 33 025 7.295e–2 2.329e–2 1.617e–4
EOC 0.999 1.500 1.509 EOC 0.994 1.488 1.525

error estimator
real error
optimal convergence rate

0
10
−1
10
error & estimator

error & estimator

−1
10 −2
10

4 6 6 8
10 10 10 10
N × DOF N × DOF

Figure 3. Error estimator E and energy error e vs. total number of degrees of freedom
(N · DOF) for 1d tent obstacle example with χh = χ (left) and 2d oscillating moving circle
problem (right). Since N · DOF  τ h1 d  hd+1 1
, provided τ  h, the optimal error decay is
1
− d+1
O(h) = O((N · DOF) ) and is indicated by the dotted lines with slopes –1/2 (left) for d = 1
and –1/3 (right) for d = 2. This shows optimal decay of both E and e.

Table 3. 1d tent obstacle problem (χh = χ): Since the meshsize is h ≈ 7.8 × 10−3 the FEM
captures the exact free boundary within one element.

Time Exact free boundary Approx free boundary

0.6 ±1.0000 × 10−1 ±1.0156 × 10−1
0.8 ±1.3333 × 10−1 ±1.3328 × 10−1
1.0 ±1.6667 × 10−1 ±1.6406 × 10−1

To answer this question, we modify Example 5.2 by the shift v(x − 13 , t) for v = u, χ, f but keep the same
meshes and time steps as before. In this case, the kink at x = 1/3 is never a mesh point and χh = χ. Since χ is
almost in H 3/2 we expect a rate of convergence 0.5 in H 1 . This is confirmed by the results of Table 4, which also
shows that the only estimator that detects this reduced order is Eχ , the obstacle consistency error estimator.
We observe that Eh and Eτ dominate at the beginning and it takes quite awhile to reach the asymptotic regime.
506 A POSTERIORI ERROR ANALYSIS FOR PARABOLIC VARIATIONAL INEQUALITIES

error estimator
exact free boundary

0.1 0.1 0.1

space error estimator

0.05 0.05 0.05

0 0 0
−0.5 0 0.5 −0.5 0 0.5 −0.5 0 0.5
t = 0.60000 t = 0.80000 t = 1.00000

Figure 4. 1d tent obstacle example: Nodal-based space error estimator Enz at times tn =
0.6, 0.8, 1.0 for DOF = 255 and τ = 2.5 × 10−4 . The localization property that Enz vanishes
at the full-contact nodes z ∈ Chn is clearly visible, along with the fact that free-boundary
approximation takes place within one element (see Tab. 3).

Table 4. 1d tent obstacle problem (χh = χ): The kink is not resolved by the underlying
meshes with uniform mesh refinement. The only estimator that detects the reduced order 0.5
is Eχ . The total estimator is dominated by Eh and Eτ at the beginning but eventually Eχ takes
over. This combined effect is reflected in the behavior of the total estimator E.

N DOF Eh Eτ Eτ h Eχ E e E/e
1024 2047 1.434e–1 1.548e–1 4.154e–4 9.882e–2 2.330e–1 8.175e–2 2.850
2048 4095 7.172e–2 7.741e–2 1.466e–4 6.988e–2 1.266e–1 5.050e–2 2.507
4096 8191 3.587e–2 3.871e–2 5.181e–5 4.941e–2 7.229e–2 3.282e–2 2.203
8192 16 383 1.794e–2 1.935e–2 1.831e–5 3.494e–2 4.378e–2 2.213e–2 1.978
16 384 32 767 8.970e–3 9.676e–3 6.471e–6 2.471e–2 2.801e–2 1.527e–2 1.834
EOC 1.000 1.000 1.501 0.500 0.644 0.535 –

We wonder whether making a suitable local mesh refinement near the kink may restore the optimal linear
rate. We conduct an experiment consisting of locally refined meshes only at the kink location, where the
meshsize is h2 , whereas it remains uniform and equal to h elsewhere. The interpolation error in H 1 becomes
1
now proportional to h, both at the kink location and elsewhere, because the error in W∞ is O(1) and O(h),
respectively. This heuristic argument is corroborated by the results of Table 5, which illustrates the potentials
of mesh refinement to achieve optimal complexity along with the importance of Eχ .

5.4. 1d American option

In American option pricing problems, we start from an initial condition u0 , as in (1.3), which is in the Sobolev
3
space H 2 − for any  > 0 but not in any smoother regularity class. The a priori error estimates of [18] imply
a rate of convergence O(τ 1/2 ) for u0 ∈ H 1 and O(τ ) for u0 ∈ H 2 . Given the fractional regularity right in
the middle between H 1 and H 2 , we expect, from interpolation theory, that the convergence rate with uniform
time-step would be about O(τ 3/4 ). Our experiments confirm this expectation.
 K.-S. MOON ET AL. 507

Table 5. 1d tent obstacle problem (χh = χ): The underlying partition is locally refined at
the kink location, where the meshsize is h2 , but is otherwise uniform with meshsize h. This
restores the optimal linear rate for both Eχ and e, as well as the total estimator E (compared
with the reduced rate reported in Tab. 4 for uniform meshes).

N DOF Eh Eτ Eτ h Eχ E e E/e
32 45 8.710 4.915 1.787e–1 1.398e–1 1.000e+1 2.802 3.570
64 93 4.477 2.467 5.648e–2 9.882e–2 5.113 1.431 3.573
128 191 2.267 1.236 1.947e–2 3.494-e-2 2.583 7.230e–1 3.572
256 382 1.141 6.186e–1 6.631e–3 1.747e–2 1.298 3.634e–1 3.572
512 767 5.723e–1 3.095e–1 2.326e–3 8.735e–3 6.507e–1 1.822e–1 3.572
EOC 0.995 0.999 1.511 1.000 0.996 0.996 –

Table 6. 1d American put option problem: Uniform time and space partitions yield subopti-
mal rates for Eτ and Eχ due to the fractional regularity of the initial condition, which is about
H 3/2 . This explains the order of about 0.75 of Eτ , that accounts for the initial transient regime,
but not quite the suboptimal order of Eχ .

N DOF Eh Eτ Eτ h Eχ E
128 511 4.353e–2 1.149e–1 3.240e–3 3.843e–1 4.035e–1
256 1023 2.172e–2 7.023e–2 1.147e–3 2.434e–1 2.543e–1
512 2047 1.091e–2 5.026e–2 4.035e–4 1.195e–1 1.301e–1
1024 4095 5.461e–3 2.940e–2 1.416e–4 7.581e–2 8.150e–2
2048 8191 2.736e–3 1.751e–2 4.980e–5 4.931e–2 5.240e–2
EOC 0.997 0.748 1.505 0.620 0.637

We take an American put option problem on a single stock with strike price K = 100, maturity time
T = 0.5 year, volatility α = 0.4, interest rate r = 6%, and forcing f = 0. We choose the space domain to be
Ω = (−1, 7). Table 6 displays all four estimators and Eτ has indeed the expected rate of about 0.75.
We now explore the effect of refining the time partition to restore the optimal convergence rate. We design
an algebraically graded time grid
n β
tn = ∀1 ≤ n ≤ N,
N
with β > 0 to be determined so that the time error estimator Eτ ≈ O(N −1 ). The time-step τn reads

n β n−1 β β n β−1 β 1−1/β

τn = − ≈ ⇒ τn ≈ t .
N N N N N n
We recall the regularizing effect for linear parabolic problems, namely,

∂t u(·, t) H1 ≈ u(·, t) H3
t−3/4

provided the initial condition u0 ∈ H 3/2 . We proceed heuristically and assume the same asymptotic
 tn behavior
to be valid for parabolic variational inequalities. We next formally replace Uhn − Uhn−1
tn−1 |||∂t u(·, t)||| dt
in the definition of Eτ to get
N 
 tn  T
2 β
Eτ2 ≈ |||∂t u(·, t)||| τn2 dt ≈ t−3/2+2(1−1/β) dt ≈ O(N −1 ),
n=1 tn−1 N 0

provided β > 4/3. This argument can be made rigorous for linear parabolic equations upon using Theorem 4.5
of [26] and carefully approximating the solution on the first time interval. To test this heuristic argument for
508 A POSTERIORI ERROR ANALYSIS FOR PARABOLIC VARIATIONAL INEQUALITIES

n 3/2
Table 7. 1d American put option problem: Algebraically graded time partition tn = N
and uniform space mesh. This grading restores the optimal linear convergence rate of both Eτ
and Eχ (compared with Tab. 6).

N DOF Eh Eτ Eτ h Eχ E
80 1023 2.386e–2 8.152e–2 1.945e–3 1.833e–1 2.021e–1
160 2047 1.159e–2 4.397e–2 6.693e–4 8.679e–2 9.798e–2
320 4095 5.657e–3 2.235e–2 2.313e–4 4.385e–2 4.954e–2
640 8191 2.793e–3 1.137e–2 8.030e–5 2.238e–2 2.526e–2
1280 16 383 1.388e–3 5.899e–3 2.787e–5 1.162e–2 1.310e–2
EOC 1.018 0.975 1.526 0.970 0.972

parabolic variational inequalities, we take β = 1.5 and report the results in Table 7. We see that this properly
chosen time partition restores the optimal convergence rate not only for Eτ but also for Eχ . Moreover, this
argument explains why uniform time stepping, i.e. β = 1, yields a suboptimal convergence rate for the time
estimator Eτ (see Tab. 6).

5.5. 1d American option with moving obstacle

To test the asymptotic behavior of the obstacle consistency term |||∂t (χ − Uh )+ |||∗ , which we omitted in E, we
modify the previous American option problem in the following way: from time t = 0 to 0.5, we still have the
same American option pricing problem as in Section 5.4. From time t = 0.5 to 1.0, we raise the obstacle at a
constant rate ξ ∈ R+ . In other words, the obstacle in the previous example has been replaced by:

+  
χ(x, t) := (K − ex ) 1 + ξ(t − 0.5)+ x ∈ (−1, 7), t ∈ [0, T ].

In this way, we exclude the initial transient region from our consideration and the singular point log(K) is
always a mesh point. Also we choose the speed ξ moderate to prevent the free boundary point to recede to
log(K). Similar to the analysis for the last example, the uniform space mesh and algebraically graded time
partition should give optimal convergence rate as before. We report in Table 8 the mixed error estimator terms
Eτ h , Eχ and
  12
T
Eχo := ∂t (χ − Uh )+ 2L2 (Ω) dt .
0

Since it is difficult to compute the dual norm in the term

  12
T
2
Eχ∗ := ∂t (χ − Uh ) +
∗
dt ,
0

we compute the term Eχo with the L2 -norm instead.

From Table 8 we see that the experimental convergence rate of Eχo is 1.0. Since Eχ∗
Eχo , the numerical results
show evidence that Eχ∗ is of at least the same order as the obstacle consistency term Eχ and this justifies the
comments in Remark 5.2. On the other hand, we see that the convergence rate of the mixed error term Eτ h
is greater than 1.0 and becomes closer to 1.0 as the obstacle moves faster and faster. Notice that we omit in
Table 8 the space and time error estimators Eh and Eτ which also converge at the optimal rate 1.0.
 K.-S. MOON ET AL. 509

Table 8. Modified American put option problem: Algebraically graded time partition tn =
n 3/2
N and uniform space mesh.

ξ 0.01 0.1 1.0

N DOF Eτ h Eχ Eχo Eτ h Eχ Eχo Eτ h Eχ Eχo
40 511 1.024e–2 3.116e–1 1.250e–1 1.128e–2 3.321e–1 1.250e–1 8.120e–2 5.144e–1 1.280e–1
80 1023 3.623e–3 1.559e–1 6.311e–2 4.015e–3 1.660e–1 6.310e–2 3.613e–3 2.565e–1 6.421e–2
160 2047 1.255e–3 7.801e–2 3.157e–2 1.523e–3 8.304e–2 3.157e–2 1.895e–2 1.280e–1 3.231e–2
320 4095 4.472e–4 3.902e–2 1.584e–2 5.554e–4 4.153e–2 1.584e–2 8.698e–3 6.398e–2 1.611e–2
EOC 1.489 0.999 0.995 1.455 1.000 0.995 1.123 1.000 1.005

Table 9. 2d oscillating moving circle problem: The space and time estimators Eh , Eτ , decrease
with optimal order 1, but the mixed estimator Eτ h is of higher order. The ratio between total
estimator E and energy error e is quite stable and of moderate size.

N DOF Eh Eτ Eτ h E e E/e
64 1985 3.432e–1 8.110e–2 2.219e–3 3.527e–1 8.328e–2 4.237
128 8065 1.597e–1 4.008e–2 8.087e–4 1.646e–1 4.204e–2 3.922
256 32 513 7.664e–2 1.996e–2 2.899e–4 7.920e–1 2.111e–2 3.745
512 130 561 3.749e–2 9.965e–3 1.037e–4 3.879e–2 1.058e–2 3.663
1024 523 265 1.853e–2 4.980e–3 3.691e–5 1.919e–2 5.297e–3 3.623
EOC 1.017 1.001 1.490 1.015 0.998 –

5.6. 2d oscillating moving circle

Let the operator be L := −∆, the domain be Ω = [−1, 1]2 , the time interval be [0, 0.25], and the noncontact
and contact sets be N := {|x − c(t)|2 > r0 (t)} and C := {|x − c(t)|2 ≤ r0 (t)} with

T
r0 (t) = 1/3 + 0.3 sin(4 ωπt), c(t) = r1 cos(ωπt), sin(ωπt) ,

and r1 = 1/3, ω = 4.0. The obstacle is χ ≡ 0, and the exact solution u and forcing function f are
 2
1 2 2
u(x, t) = 2 |x − c(t)|2 − r0 (t) in N
0 in C,
⎧
⎨ 4 r02 (t) − 2|x − c(t)|22 − 1
|x − c(t)|22 − r02 (t) (x − c(t)) · c (t) + r0 (t)r0 (t) in N
2
f (x, t) =
⎩ −4r02 (t) 1 − |x − c(t)|22 + r02 (t) in C.

The free boundary is an oscillating circle with radius r0 (t) and center c(t) moving counterclockwise along the
circle of radius r1 centered at the origin. The initial and boundary conditions are given by u.
We halve both time-step τ and space meshsize h in each experiment and report the results in Table 9 and
Figure 3; we observe optimal linear convergence rate. We also investigate in Table 2 the decay of the space
estimators alone. We fix the time-step τ = 2.5 × 10−4 and halve the meshsize size in each experiment. We
observe optimal linear decay of Eh,1 but higher order of convergence for Eh,2 , Eh,3 .
In Figure 5, we show the nodal-based space error indicator Enz on the cross section x2 = 0 at different stages
of the evolution tn = 0.02, 0.05, 0.18. For the same times and cross section, we also compare the exact and
approximate free boundaries in Table 10. Their difference is well within one meshsize.
510 A POSTERIORI ERROR ANALYSIS FOR PARABOLIC VARIATIONAL INEQUALITIES

−3 −3 −3
x 10 x 10 x 10
7 7 7

error estimator
exact free boundary
6 6 6

5 5 5
space error estimator

4 4 4

3 3 3

2 2 2

1 1 1

0 0 0
−1 0 1 −1 0 1 −1 0 1
t = 0.02000 t = 0.05000 t = 0.18000

Figure 5. 2d oscillating moving circle problem: nodal-based error estimator Enz in the cross
section x2 = 0 for DOF = 8065, τ = 2.5 × 10−4 and tn = 0.02, 0.05, 018. Note the vanishing of
Enz for full-contact nodes and the monotone behavior for the rest.

Table 10. 2d oscillating moving circle: Exact and approximate free boundaries on the cross
section x2 = 0. Their differences are less than one meshsize, which is about 2.2 × 10−2 .

Time Exact free boundaries Approx. free boundaries

0.02 {−2.5788 × 10−1 , 9.0361 × 10−1 } {−2.5000 × 10−1 , 9.0625 × 10−1 }
0.05 {−2.0083 × 10−1 , 7.4017 × 10−1 } {−1.8750 × 10−1 , 7.1875 × 10−1 }
0.18 {−5.7430 × 10−1 , 1.4942 × 10−1 } {−5.6250 × 10−1 , 1.5625 × 10−1 }

6. Conclusions
We have developed a novel a posteriori error analysis for parabolic variational inequalities, including local-
ization features to the non-contact region, and illustrated it with several numerical experiments, some relevant
in finance. Upon comparing theory and practice we have the following concluding remarks:
• Error decay: For problems with smooth data, the energy error in L2 (0, T ; H 1 (Ω)) decays linearly, namely
O(h + τ ). Although the experimental evidence is consistent, there seems to be no theoretical justification yet.
If the obstacle χ exhibits a singularity not resolved by the mesh, as in Section 5.3, or the initial condition is
rough, as in Section 5.4, the actual error decays with a suboptimal rate. Suitable mesh refinement in either
space or time appears to cure this problem; see again Sections 5.3 and 5.4.
• Estimator decay: The numerical experiments corroborate that the proposed fully localized error estimator E
decays with the same rate as the actual error e. We have demonstrated experimentally that the components
Eh , Eτ , Eχ of E provided valuable a posteriori information of the solution. We are confident that they can
be utilized to drive an adaptive algorithm for problems with unilateral constraints. This is a topic of future
research [16].
• Dominant estimators: The dominant parts of the error estimator E for smooth data are Eh,1 and Eτ , which
are of order O(h) and O(τ ), respectively. Our lower bounds of Section 3.2 are in fact derived for these two
terms. The other terms Eh,2 , Eh,3 , Eτ h are of higher order in all our test examples. The obstacle consistency
term Eχ is relevant when χh = χ, in which case it may be dominant, and Tables 4, 6 and 7 show that it
provides correct information for obstacle approximation.
• Localization of space estimator: Figures 4 and 5 show that the nodal-based space estimator Enz vanishes
at full-contact nodes z ∈ Chn . Its contribution comes only from the non-contact region where the solution
 K.-S. MOON ET AL. 511

behaves like the solution of a linear parabolic equation. This estimator yields an upper bound also for globally
linear parabolic problems and seems to be new in the literature of parabolic PDE.
• Exercise boundary approximation: accurate approximation of the free (exercise) boundary is an impor-
tant problem in option pricing. Numerical results in Sections 5.2 and 5.6, particularly Figures 4 and 5 as well
as Tables 3 and 10, suggest an excellent agreement between approximate and exact free boundaries. This
observation could be made rigorous, upon extending the idea in [20], provided pointwise a posteriori error
estimates were available. This is under further investigation.

References
[1] W. Bangerth and R. Rannacher, Adaptive Finite Element Methods for Differential Equations. Lectures in Mathematics ETH
Zürich, Birkhäuser Verlag (2003).
[2] A. Bergam, C. Bernardi and Z. Mghazli, A posteriori analysis of the finite element discretization of some parabolic equations.
Math. Comp. 74 (2005) 1117–1138 (electronic).
[3] F. Black and M. Scholes, The pricing of options and corporate liabilities. J. Polit. Econ. 81 (1973) 637–659.
[4] H. Brézis, Opérateurs maximaux monotones et semi-groupes de contraction dans les espaces de Hilbert. North Holland (1973).
[5] H. Brézis and F.E. Browder, Nonlinear integral equations and systems of Hammerstein type. Adv. Math. 18 (1975) 115–147.
[6] M. Broadie and J. Detemple, Recent advances in numerical methods for pricing derivative securities, in Numerical Methods in
Finance, L.C.G. Rogers and D. Talay Eds., Cambridge University Press (1997) 43–66.
[7] L.A. Caffarelli, The regularity of monotone maps of finite compression. Comm. Pure Appl. Math. 50 (1997) 563–591.
[8] Z. Chen and R.H. Nochetto, Residual type a posteriori error estimates for elliptic obstacle problems. Numer. Math. 84 (2000)
527–548.
[9] C.W. Cryer, Successive overrelaxation methods for solving linear complementarity problems arising from free boundary prob-
lems, Free boundary problems I, Ist. Naz. Alta Mat. Francesco Severi (1980) 109–131.
[10] A. Fetter, L∞ -error estimate for an approximation of a parabolic variational inequality. Numer. Math. 50 (1987) 57–565.
[11] F. Fierro and A. Veeser, A posteriori error estimators for regularized total variation of characteristic functions. SIAM J.
Numer. Anal. 41 (2003) 2032–2055.
[12] R. Glowinski, Numerical methods for nonlinear variational problems. Springer series in computational physics, Springer-Verlag
(1984).
[13] P. Jaillet, D. Lamberton and B. Lapeyre, Variational inequalities and the pricing of American options. Acta Appl. Math. 21
(1990) 263–289.
[14] C. Johnson, Convergence estimate for an approximation of a parabolic variational inequatlity. SIAM J. Numer. Anal. 13
(1976) 599–606.
[15] D. Lamberton and B. Lapeyre, Introduction to stochastic calculus applied to finance. Springer (1996).
[16] R.H. Nochetto and C.-S. Zhang, Adaptive mesh refinement for evolution obstacle problems (in preparation).
[17] R.H. Nochetto, G. Savaré and C. Verdi, Error control for nonlinear evolution equations. C.R. Acad. Sci. Paris Ser. I 326
(1998) 1437–1442.
[18] R.H. Nochetto, G. Savaré and C. Verdi, A posteriori error estimates for variable time-step discretizations of nonlinear evolution
equations. Comm. Pure Appl. Math. 53 (2000) 525–589.
[19] R.H. Nochetto, K.G. Siebert and A. Veeser, Pointwise a posteriori error control for elliptic obstacle problems. Numer. Math.
95 (2003) 163–195.
[20] R.H. Nochetto, K.G. Siebert and A. Veeser, Fully localized a posteriori error estimators and barrier sets for contact problems.
SIAM J. Numer. Anal. 42 (2005) 2118–2135.
[21] M. Picasso, Adaptive finite elements for a linear parabolic problem. Comput. Methods Appl. Mech. Engrg. 167 (1998) 223–237.
[22] A. Schmidt and K.G. Siebert, Design of adaptive finite element software: the finite element toolbox ALBERTA. Lecture Notes
in Computational Science and Engineering, Springer (2005).
[23] A. Veeser, Efficient and reliable a posteriori error estimators for elliptic obstacle problems. SIAM J. Numer. Anal. 39 (2001)
146–167.
[24] R. Verfürth, A review of a posteriori error estimation and adaptive mesh-refinement techniques. Wiley Teubner (1996).
[25] R. Verfürth, A posteriori error estimates for finite element discretizations of the heat equation. Calcolo 40 (2003) 195–212.
[26] T. von Petersdorff and C. Schwab, Numerical solution of parabolic equations in high dimensions. ESAIM: M2AN 38 (2004)
93–127.
[27] C. Vuik, An L2 -error estimate for an approximation of the solution of a parabolic variational inequality. Numer. Math. 57
(1990) 453–471.
[28] P. Wilmott, J. Dewynne, and S. Howison, Option Pricing: Mathematical Models and Computation. Oxford Financial Press,
Oxford, UK (1993).