A Priori Error Analysis for NCVEM Discretization of Elliptic Optimal Control Problem ()
1. Introduction
The main purpose of this paper is to discuss the prior error analysis of the NCVEM discretization for the elliptic optimal control problem. Consider the following optimal control problems with state constrained:
(1.1)
subject to
(1.2)
where
is the objective functional, q is the state variable,
is the desired state and
is a positive constant parameter. The aim of the control is to make the state variable q as close as to the desired state
. Ω is a bounded polygon on R2. The admissible set of the control is given by
The above state constrained optimal control models perform an increasingly important role in many science and engineering fields. For this reason,the research on optimal control problems becomes meaningful. Different types of optimal control problems are solved by finite elements method (FEM) [1] [2] , discontinuous Galerkin method [3] , spectral method [4] so on. In [5] error estimates of approximate local optimal control for semi-linear elliptic equation with finite many state constraints were given. In [6] A control vector was used instead of a control function to establish a high order error estimate for similar Settings. In [7] , these results were generalized to a less regular setting for the states and the convergence of FEM approximations for semilinear distribution and boundary control problems was obtained. In [8] , a discretization concept was investigated, which used the relationship between adjoint state and control to discretized control variables. They discretized the equation of state using linear FEM and proved the convergence of the equation of state on L2 norm. On the other hand, for the optimal control problems of elliptic equations, Stokes equations and parabolic equations, the corresponding posteriori error estimates of the conforming FEM were given in [9] [10] [11] .
The basic driving force of the VEM comes from the processing of arbitrarily shaped polygons [12] . The traditional meaning of finite element or finite difference requires discretization of physical entities with significant geometric features to solve, which to some extent loses the “macro” description of entity geometric information. However, in practical engineering requirements, more and more calculations involve dealing with specific geometric structures, such as the deformation of non convex polygons and the contact of complex structures. Compared to the FEM, the VEM also requires discretization of the geometric space, approximating the actual problem by forming and solving a system of linear equations. The difference lies in: 1) The approximation function used in the finite element method is an explicit polynomial function; In the VEM, in addition to polynomial functions, approximate functions also have functions that are continuous polynomials at the boundary of the element and satisfy certain conditions inside the element (such as polynomials after Laplace operation). These functions are not explicitly expressed in the element domain; 2) In the FEM, the degrees of freedom are all values of approximate functions. The VEM avoids calculating the values of approximate functions inside the element when forming a stiffness matrix by defining reasonable degrees of freedom; 3) The stiffness matrix of the finite element method only contains one term, while the stiffness matrix of the virtual element method includes the coordination matrix and the stability matrix to ensure the convergence of the calculation.
Due to these advantages of VEM, the VEM for solving various partial differential equations has been proposed and developed, such as linear elasticity [13] , Stokes or Navier-Stokes equations [14] [15] , Cahn-Hilliard equations [16] , and so on. Back to the NCVEM proposed in this paper. The NCVEM was first introduced for elliptic problems in [17] . In the past years, it has been successfully used to solve different models [18] - [23] . But so far, NCVEM has not been used to solve the elliptic optimal control problem. In this paper, NCVEM is introduced to approach the elliptic optimal control problem with pointwise control constraints. Based on NCVEM approximation (through VEM projection operator), a VEM discrete scheme is constructed. Then we obtain the corresponding priori estimate for three variables in H1 and L2 norms.
Throughout this paper, for an open bounded domain
in Ω, standard notation
and
denote seminorm and norm, respectively, in the Sobolev space
. when
,
and
denote the inner product and the norm of
. When
is the whole domain Ω, the subscript can be omitted. Let
denote the space of polynomials of degree at most
on E. Usually,
.
The aim of this paper is to construct a NCVEM for constrained optimal control problems. The rest of this paper is as follows. In the next section, we introduce some preliminaries knowledge about VEM. Then, the continuous first-order optimality system of elliptic optimal control problem are introduced. In Section 3, we derive the VEM discrete scheme, discrete first-order optimality condition. Then the priori error estimates are derived both for three variables in H1 and L2 norm. In Section 4, we show a numerical example to verify our theoretical analysis. Finally, we make some summaries in Section 5.
2. Preliminaries Knowledge
In this section, we mainly introduce local projection operators and the definitions of virtual element space.
Fristly, suppose
is a family of decompositions of the domain Ω divided into star-shaped polygons E. For any
,
and
The set of edges s of
is denoted by
, which is subdivided into the set of boundary edges
and the set of internal edges
. Finally,
denote the length of the edge s.
Before introducing the virtual element space, we first make the following assumptions about the grid.
Assumption 2.1 (See [24] ) (mesh regularity) We assume that there exists a real number
such that, for every
satisfies the following two assumptions.
1) Every element E is star-shaped with respect to a circle with a radius
;
2) Every edge s of E has length
;
Next we give the definition of the projection operator.
Definition 2.1 (See [24] ) Define the L2 projection operator
as follows:
(2.1)
Definition 2.2 (See [24] ) Define the H1 projection operator
as follows:
plus
A finite dimensional function space
, where
(2.2)
(not necessarily a subspace of
We denote by
the set of scaled polynomials
The global virtual element space in each case is constructed from [18] as a subspace of an infinite dimensional space W, defined differently for the VEM and NCVEM. For the VEM we simply take
. For the NCVEM, we introduce the subspace
of the nonconforming broken Sobolev space
defined in (3.1), by imposing certain weak inter-element continuity requirements such that
The jump operator
across a mesh interface
is defined as follows for
.
If
,
there exist
and
such that
. Denote by
the trace of
on s from within
and by
the unit outward normal on s from
.
If
,
, w representing the trace of v from within the element E, having s as an interface and
is the unit outward normal on s from E.
The modified virtual element space from is defined as
where
We can see that the space
is not a subspace of
, so we need to define a discrete norm
Similar to reference [18] , we define the degrees of freedom as
(2.3)
(2.4)
Using the above degrees of freedom (2.3) and (2.4), projection
and
are exactly computable.
Next we introduce the continuous first-order optimality system
Theorem 2.1 Let
be the the solution of (1.1) and (1.2). Then the following first-order optimality system holds
and
where p is called the adjoint state variable and
(2.5)
Let
(2.6)
denotes the pointwise projection onto the admissible set
. Similar to the discussion in Becker et al. [25] , (2.6) is equivalent to
Let
. Then the bilinear form of the continuous first-order system reads:
(2.7)
From classical Scott-Dupont theory [26] we first introduce a local estimate.
Lemma 2.1 For any
, there is a piecewise polynomial
satisfying
(2.8)
where
For any function
, using degrees of freedom (2.3) and (2.4) we can define the interpolation
. We denote by
the ith degrees of freedom given by (2.3) and (2.4), where
and
is the number of degrees of freedom (2.3) and (2.4). Then we have
From [23] the following lemma provides the interpolation error estimation of the nonconforming virtual element
Lemma 2.2 For any
and
, we have
(2.9)
3. Virtual Element Approximation
3.1. Virtual Element Discrete Scheme for State Equation
The bilinear form of the state equation is as follows
The corresponding virtual element discrete scheme of (1.2) can be defined by
where
There are many choices for
, and following [12] we take the simple choice
such that it satisfies the following property:
Here
is the number of degrees of freedom on the element E and denotes the
value of the yth local degree of freedom defining
in
.
We define consistency and stability as follows:
Definition 3.1 see ( [23] )
- Consistency: For all
and for all
(3.1)
- Stability: There exist positive constants
and
independent of h and the mesh element E such that
(3.2)
Lemma 3.1 See ( [23] ) The Discrete bilinear form
satisfies the polynomial consistency property and the stability property. Then, we obtain
is coercive.
The virtual element approximation of control problem (1.1)-(1.2) is to find
such that
(3.3)
subject to
(3.4)
In [27] , it also showed that (3.3) and (3.4) has a unique solution
and that
is the solution of (3.3) and (3.4) if and only if there is a co-state
such that
satisfies the following discrete first-order optimality conditions:
(3.5)
3.2. A Priori Error Estimate
Lemma 3.2 see ( [28] )) There exists a positive constant C such that, for all
and all smooth enough functions w defined on E, it holds:
(3.6)
To derive a priori error estimate we need to introduce the following auxiliary problems:
(3.7)
We make the following data assumption:
Assumption 3.1 (Data assumption) We assume the solution
of the optimal control problem and
satisfy:
Then we have the following estimates.
Note that
, from integration by parts, we get
The consistency error is
(3.8)
Similar to the theoretical analysis in Xiao et al. [23] , we have
(3.9)
Theorem 3.1 Suppose that
is the solution of (2.7), and
is the solution of auxiliary problem (3.7), under Assumpiton 2.1 and assumption 3.1 we have
Proof:
(3.10)
where
and
From the coercivity of
(3.11)
Using formula (2.8), interpolation estimates (2.9) and stability property of
, we get
(3.12)
By the consistency and the stability of
, we obtain
(3.13)
Then, by the minkowski inequality and the property of the L2 projection, we get
(3.14)
By integration by parts, we write
Furthermore, from (3.9) we have
(3.15)
By substituting (3.12), (3.13), (3.14) and (3.15) into (3.11), we have
(3.16)
From Lemma 2.2 (2.9) we get
(3.17)
By combining (3.16) and (3.17), we have
(3.18)
By repeating the proof procedure of (3.18), we can get
,
Similar to the discussion in Theorem 3.1 and theoretical analysis in Andrea et al. [18] we have the L2 error estimates between solution of (2.7) and the solution of auxiliary problem (3.7)
Lemma 3.3 see ( [18] ) Suppose that
is the solution of (2.7), and
is the solution of auxiliary problem (3.7), under Assumpiton 2.1 and assumption 3.1 we have
Theorem 3.2 (A priori error estimate) Let
and
are the solutions of (2.7) and (3.5) respectively. Under the Assumptions 2.1 and 3.1 we derive
and
Proof: We decompose the errors
and
into
and
From the discrete first-order optimality system of the optimal control problem (3.5) and auxiliary problems (3.7) we get
Let
, and from (3.2) we have
(3.19)
then
By the Lemma 3.2 and Theorem 3.1, we have
Combining these inequalities, we get
and
By the Lemma 3.2 and Theorem 3.1, we have
Let
. From (3.2) we derive
(3.20)
Further we have
By triangle inequality we derive
(3.21)
similar to (3.20) we obtain
Through the triangle inequality, we can infer
and
Because both the estimation of the state and the adjoint state depend on the estimation of the control variables, now we need estimate
Define
Then, we can prove that
(3.22)
Using (3.7) and the property of the projection of L2 we can deduce
Then from (3.22) we have
This shows
Note that
Then by Lemma 3.1 and (3.21) we have
Inserting the
into the
and
, we obtain the result.
,
Remark 1 (comparison with conforming VEM). The global virtual element space defined differently for the VEM and NCVEM, For the VEM simply take
. For NCVEM, we introduce the nonconforming broken Sobolev space
by imposing certain weak inter-element continuity requirements. In contrast to conforming VEM, since
is not a subset of
in general, the substitution of discrete function
in the weak formulation leads to a nonconformity error such as (3.15).
4. Numerical Experiments
In this section, we present three different sequences of meshes to validate the performance of our error analysis presented in this paper. Through decomposing the domain into multiple squares, we obtain the first sequence of meshes (labeled square). The second meshes (labeled Lloyd) is given through Voronoi mesh generator [29] . The third sequence (labeled distorted) is to divide the Lloyd meshes into multiple distorted Lloyd meshes. These three sequences of meshes are respectively shown in Figures 1(a)-(c).
We will confirm the priori error on the three grids by showing the relative errors in L2 and H1 norm between
and the solution
given by the NCVEM. We use
to denote the relative errors in the L2 norm between
and
. Similarly, we respectively use
to denote the relative errors in the H1 norm between
and
.
Example: The optimal control problem (1.1)-(1.2) is restricted to the unit square
. Let
,
,
. We chose the following exact solution
f and
can be determined from the exact solutions
.
In Table 1, there different meshes data of size parameter (mesh diameter), number of elements and vertices are shown.
In Figures 2-4, we present the convergence rate curves of the state, adjoint
Table 1. Mesh data for three grid meshes. h represents the mesh size parameter, and
and
represent the number of elements and vertices of the mesh.
Figure 2. Relative errors of state variables in L2 and H1 norm.
Figure 3. Relative errors of adjoint state variables in L2 and H1 norm.
Figure 4. Relative errors of control variables in L2 and H1 norm.
state and control variables in L2 and H1 norm in Tables 2-7, the numerical results about the relative errors and convergence are shown on three different meshes. In Figures 5-7, we present the figure of the solution of three variables in three meshes.
Table 2. Relative errors and convergence rates of state, adjoint state, control variables in L2 norm on Square mesh of Example 4.1.
Table 3. Relative errors and convergence rates of state, adjoint state variables in H1 norm on Square mesh of Example 4.1.
Table 4. Relative errors and convergence rates of state, adjoint state, control variables in L2 norm on Lloyd mesh of Example 4.1.
Table 5. Relative errors and convergence rates of state, adjoint state variables in H1 norm on Lloyd mesh of Example 4.1.
Table 6. Relative errors and convergence rates of state, adjoint state, control variables in L2 norm on Distorted mesh of Example 4.1.
Table 7. Relative errors and convergence rates of state, adjoint state variables in H1 norm on Distorted mesh of Example 4.1.
Figure 5. The solution of state variables in Lloyd meshes.
Figure 6. The solution of adjoint state variables in Lloyd meshes.
Figure 7. The solution of control variables in Lloyd meshes.
5. Conclusions
In this paper, NCVEM is applied to approximate elliptic optimal control problems with pointwise control constraints. The priori error estimates are derived, numerical examples verifies the theoretical results.
In our future work, we will increase the complexity of the elliptic problem, generalize the problem to linear indefinite elliptic problems. And because of the flexibility of the VEM, we will derive a posteriori error estimate for the problem, and derive an adaptive grid algorithm to guide mesh refinement.