Abstract
The main goal of the paper is to establish time semidiscrete and space-time fully discrete maximal parabolic regularity for the time discontinuous Galerkin solution of linear parabolic equations. Such estimates have many applications. They are essential, for example, in establishing optimal a priori error estimates in non-Hilbertian norms without unnatural coupling of spatial mesh sizes with time steps.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
Let \(\Omega \) be a Lipschitz domain in \(\mathbb {R}^d\), \(d=2,3\) and \(I=(0,T)\). We consider the heat equation as a model of a parabolic second order partial differential equation,
with a right-hand side \(f \in L^s(I;L^p(\Omega ))\) for some \(1\le p,s\le \infty \) and \(u_0\in L^p(\Omega )\), \(1\le p\le \infty \). The maximal parabolic regularity for \(u_0\equiv 0\) says that there exists a constant C such that,
(see, e.g., [1–3]). The maximal parabolic regularity is an important analytical tool and has a number of applications, especially to nonlinear problems and/or optimal control problems when sharp regularity results are required (cf. [4–7]). Our aim in this paper is to establish similar maximal parabolic regularity results for time discrete discontinuous Galerkin solutions as well as for the fully discrete Galerkin approximations. Such results are very useful, for example, in fully discrete a priori error estimates and are essential in order to keep the spatial mesh size h and the time steps k independent of each other (cf. [8]). In [9] we apply the results of this paper to establish pointwise best approximation estimates for fully discrete Galerkin solutions.
Maximal parabolic regularity with applications to semidiscrete finite element Galerkin solutions in space were analyzed for smooth domains in [10, 11] and for convex polyhedra in [12]. Time discrete results are much less known in the finite element community. Explicit methods are treated in [13–15]. Implicit Euler methods with pointwise norms in time are considered in [16, 17]. A more systematic investigation of discrete maximal parabolic regularity for various time schemes was carried out by Sobolevskiĭ and Ashyralyev and summarized in the book [18].
In this paper, we investigate maximal parabolic regularity for a family of time discontinuous Galerkin (dG) methods, which were first deeply analyzed for linear second order parabolic problems in [19]. There is a number of important properties that make the dG schemes attractive for temporal discretization of parabolic problems. For example, such schemes allow for a priori error estimates of optimal order with respect to discretization parameters, such as the size of time steps and the mesh width, as well as with respect to the regularity requirements for the solution (see, e.g., [20, 21]). Different systematic approaches for a posteriori error estimation and adaptivity developed for finite element discretizations can be adapted for dG temporal discretization of parabolic equations, (see, e.g., [22, 23]). Since the trial space allows for discontinuities at the time nodes, the use of different spatial discretizations for each time step can be directly incorporated into the discrete formulation, (see, e.g., [22]). Compared to the continuous Galerkin methods, dG schemes are not only A-stable but also strongly A-stable, (see, e.g., [24]). An efficient and easy to implement approach that avoids complex coefficients, which arise in the equations obtained by a direct decoupling for high order dG schemes, was developed in [25]. For the treatment of optimal control problems, Galerkin methods are particularly suitable since they expose an important property that the two approaches optimize-then-discretize, i.e., the discretization of the optimality system built up on the continuous level, and discretize-then-optimize, i.e., discretization of the state equation and subsequent construction of the optimality system on the discrete level, lead to the same discretization scheme, (see, e.g., [26]). Compared to continuous Petrov–Galerkin time-stepping schemes (see [27] for details), dG schemes also have the advantage that the adjoint state can use the same discretization as the state variable. This allows for unified numerical treatment and simplifies a priori and a posteriori error analysis, (see, e.g., [28–31]).
The main results of this paper for the time semidiscrete discontinuous Galerkin \(u_k\) solution consist roughly of two parts. First, for the homogeneous problem (i.e. \(f=0\)) with \(u_0 \in L^p(\Omega )\), \(1\le p\le \infty \) we show
for \(m=1,2,\dots ,M\). Then, using this smoothing result, we also establish discrete maximal parabolic regularity for the inhomogeneous problem when \(u_0=0\). We show,
for \(1\le s\le \infty \) and \(1\le p\le \infty \), with obvious notation changes in the case of \(s=\infty .\) In the case of the lowest order piecewise constant method, i.e., \(q=0\), the first terms on the left-hand side of the above estimates vanish. In contrast to the continuous case, the limiting cases \(s,p \in \{1,\infty \}\) are allowed, which explains the logarithmic factor in (3). We also provide the fully discrete analog of (2) and (3).
The rest of the paper is organized as follows. In the next section we introduce the discretization method and the resolvent estimates, which build the main analytical tool of the paper. For better communication of the ideas we first analyze the dG(0) method, which is technically much simpler, and in the following section we analyze the general dG(q) case. That is done in Sects. 3 and 4, respectively. At the end of Sect. 4 we provide an example of how such maximal parabolic regularity results can rather easily lead to optimal order error estimates. Finally, Sect. 5 is devoted to fully discrete Galerkin solutions. In Sect. 6 we provide an extension of our results to the case of a general norm fulfilling a resolvent estimate. This generalization, being of an independent interest, is used, for example, in [9] for derivation of pointwise interior (local) error estimates of fully discrete Galerkin solutions.
2 Preliminaries
To introduce the time discontinuous Galerkin discretization for the problem, we partition \(I =(0,T)\) into subintervals \(I_m = (t_{m-1}, t_m]\) of length \(k_m = t_m-t_{m-1}\), where \(0 = t_0< t_1<\cdots< t_{M-1} < t_M =T\). The maximal and minimal time steps are denoted by \(k =\max _{m} k_m\) and \(k_{\min }=\min _{m} k_m\), respectively. We impose the following conditions on the time mesh (as in [32]):
-
(i)
There are constants \(c,\beta >0\) independent on k such that
$$\begin{aligned} k_{\min }\ge ck^\beta . \end{aligned}$$ -
(ii)
There is a constant \(\kappa >0\) independent on k such that for all \(m=1,2,\ldots ,M-1\)
$$\begin{aligned} \kappa ^{-1}\le \frac{k_m}{k_{m+1}}\le \kappa . \end{aligned}$$ -
(iii)
It holds \(k\le \frac{1}{4}T\).
The semidiscrete space \(X_k^q\) of piecewise polynomial functions in time is defined by
where \(\mathcal {P}_{q}(V)\) is the space of polynomial functions of degree q in time with values in a Banach space V. We will employ the following notation for functions in \(X_k^q\)
Next we define the following bilinear form
where \((\cdot ,\cdot )_{\Omega }\) and \((\cdot ,\cdot )_{I_m \times \Omega }\) are the usual \(L^2\) space and space-time inner-products, \(\langle \cdot ,\cdot \rangle _{I_m \times \Omega }\) is the duality product between \( L^2(I_m;H^{-1}(\Omega ))\) and \( L^2(I_m;H^{1}_0(\Omega ))\). We note, that the first sum vanishes for \(u \in X^0_k\). The dG(q) semidiscrete (in time) approximation \(u_k\in X_k^q\) of (1) is defined as
Rearranging the terms in (5), we obtain an equivalent (dual) expression of B:
The analysis of such schemes in non-Hilbertian setting is usually done by using a semigroup approach that represents time stepping formulas as a Dunford–Taylor integral in the complex plane [33, Ch. 9]. This approach requires certain resolvent estimates. For Lipschitz domains and a given \(\gamma \in (0,\pi /2)\), the resolvent estimate (see [34]) guarantees the existence of a constant C such that for all \(u\in L^p(\Omega )\), \(1\le p\le \infty \), and any \(z \in \mathbb {C}{\setminus } \Sigma _{\gamma }\) the following estimate holds:
where the Laplace operator \(-\Delta \) is supplemented with homogeneous Dirichlet boundary conditions, and
Using the identity \(\Delta (z+\Delta )^{-1}={\text {Id}}-z(z+\Delta )^{-1}\), one immediately obtains,
We note, that all our results for semidiscrete solutions hold if we replace the Laplace operator \(-\Delta \) with a more general self-adjoint second order elliptic operator A provided it satisfies (8).
3 Estimates for dG(0)
For the ease of the presentation, we first establish the results for the lowest order piecewise constant discretization dG(0). In this case, we use the following notation,
First, we establish results for the homogeneous problem. In this case the dG(0) method is equivalent to the Backward Euler method.
3.1 Results for the homogeneous problem
Let \(f=0\), \(u_0\in L^p(\Omega )\) and let \(u_k\in X_k^0\) be the semidiscrete approximation of (1) defined by
i.e., the dG(0) solution \(u_k\) satisfies
The first result shows that the solution can not grow from one time step to the next one.
Lemma 1
Let \(u_k\) be the solution of (12). Then, for \(u_0 \in L^p(\Omega )\), \(1\le p\le \infty \) there holds
Proof
First, we assume \(u_0 \in L^\infty (\Omega )\) and establish
It is sufficient to consider only a single time step,
We want to show that \(\Vert u_{k,1}\Vert _{L^\infty (\Omega )}\le \Vert u_{0}\Vert _{L^\infty (\Omega )}\). Assume it is false. Let \(x_0\in \Omega \) be a point where \(u_{k,1}\) attains a maximum. By [35, Theorem 3.3], we know that \(u_{k,1}\in C_0(\Omega )\), hence, there exists an open ball \(B_\delta (x_0)\) of radius \(\delta >0\) centered at \(x_0\) with \(\overline{B_\delta (x_0)} \subset \Omega \) such that
Hence,
By the maximum principle, from
we obtain a contradiction to the assumption that \(u_{k,1}\) has a maximum at the interior point \(x_0\). This contradiction establishes (14). Next, using a duality argument, we will show
Consider the problem, to find \(z_{k,1}\in H^1_0(\Omega )\) that satisfies,
The solution \(z_{k,1}\) can be thought of as a single step of the dG(0) method to a parabolic problem with initial condition \({\text {sgn}}u_{k,1}\). Thus,
where we have used (14) for \(z_k\) and the fact that \(\Vert z_{0}\Vert _{L^\infty (\Omega )}=||{\text {sgn}}{u}_{k,1}||_{L^\infty (\Omega )}=1\). This establishes (16). Interpolating, we obtain the lemma for \(1\le p\le \infty \). Next we will establish a smoothing result. \(\square \)
Theorem 1
(Homogeneous smoothing estimate) Let \(u_k\in X_k^0\) be the solution of (12) with \(u_0 \in L^p(\Omega )\), \( 1\le p\le \infty \). Then there exists a constant C independent of k such that
Proof
The proof is given on page 1321 in [36] for the \(L^2(\Omega )\) norm, but the proof is valid for the \(L^p(\Omega )\) norm as well by using the resolvent estimate (8) with respect to the \(L^p(\Omega )\) norm. \(\square \)
Remark 1
Let \(u_k\in X_k^0\) be the solution of (12) with \(u_0 \in L^p(\Omega )\), \( 1\le p\le \infty \). Then there exists a constant C independent of k such that
From (13), we immediately obtain the following result.
Corollary 1
Let \(u_k\in X_k^0\) be the solution of (12) with \(u_0 \in L^p(\Omega )\), \( 1\le p\le \infty \). Then there exists a constant C independent of k such that
3.2 Results for the inhomogeneous problem
Now we consider \(u_{k}\in X_k^0\) to be the dG(0) solution to the parabolic equation with \(u_0=0\), i.e., \(u_k\) satisfies,
Thus, the dG(0) solution satisfies
where
Since \(f_m\) is the \(L^2\) projection of f onto the piecewise constant functions on each subinterval \(I_m\), we have
We now state our main result for the dG(0) approximations.
Theorem 2
(Maximal parabolic regularity) Let \(1\le s,p\le \infty \) and \(u_0=0\). Then, there exists a constant C independent of k such that for every \(f\in L^s(I;L^p(\Omega ))\) and \(u_k\) satisfying (17), the following estimate holds:
Proof
Using (18), we can write the dG(0) solution as
where \(r(z)=(1+z)^{-1}.\) Then,
Hence
From Remark 1, since each term in the sum on the right-hand side can be thought of as a homogeneous solution with initial condition \(f_l\) at \(t=t_{l-1}\), we have
Thus, we obtain
For \(s=\infty \), we obtain from the above estimate and using (19),
where in the last step we used that
by using the assumption \(k_{\min }\ge Ck^\beta \) and \(k\le \frac{T}{4}\). For \(s=1\), we have
Changing the order of summation and using (19), we obtain
where we used again that
Interpolating between \(s=1\) and \(s=\infty \), we obtain the result for any \(1\le s\le \infty \). \(\square \)
Remark 2
The appearance of the logarithmic term is natural for the critical values \(s=1\), \(p=1\), \(s=\infty \), or \(p=\infty \), since the corresponding maximal parabolic regularity results for the continuous problem hold only for \(1<s,p<\infty \). For \(s=2\) or \(p=2\), the power of the logarithm can be lowered. Thus, for \(p=2\), from [30] we know,
and from (20), we have
Interpolating between \(s=2\) and \(s=\infty \) and between \(s=2\) and \(s=1\), we obtain
Similarly, we can obtain,
Corollary 2
(Maximal parabolic regularity for jumps) Let \(1\le s,p\le \infty \) and \(u_0=0\). Then, there exists a constant C independent of k such that for every \(f\in L^s(I;L^p(\Omega ))\) and \(u_k\) satisfying (17), the following estimate holds,
where the jump term \([u_k]_0\) at \(t = 0\) is defined as \(u_{k,1}\).
Proof
Since by (18) on each time subinterval \(I_m\) we have
\(\square \)
by using Theorem 2, we have
Similarly, using Theorem 2, for \(1\le s<\infty \) we have
where the constant \(C_s\) depends on s. By taking the s-root we obtain the corollary.
4 Estimates for dG(q)
In this section we will establish the dG(q) version of the results from the previous section. It is convenient to introduce some additional notation. Let \(q\ge 1\) and \(\psi _l(t)\in P_q([0,1])\), \(l=0,1,\dots ,q\) be the standard Lagrange basis functions on the interval [0, 1], i.e., \(\psi _l\left( \frac{j}{q}\right) =\delta _{lj}\), where \(\delta _{lj}\) is the Kronecker symbol. Then for any \(u_k\in X_k^q\) on the time interval \(I_m=(t_{m-1},t_m]\) we have
with \(U^m_l\in H^1_0(\Omega )\) independent of t. In this notation, we have
4.1 Results for the homogeneous problem
Let \(u_{k}\in X^q_k\) be the semidiscrete in time solution to the parabolic equation with \(f\equiv 0\), i.e.,
Alternatively, on a single interval \(I_m\), we have
where the rational functions \(r_{l,0}\) are of the form,
with \(\hat{p}\) being a polynomial of degree \(q+1\) with no roots on the right-half complex plane and \(p_{l,0}\), \(l=0,1,\dots ,q\) being polynomials of degree q (cf. [36], page 1322). Since \(r_{q,0}(\lambda )\) is a subdiagonal Padé approximation of \(e^{-\lambda }\), we also have (cf. [37])
as \(\lambda \rightarrow 0\). The rational functions \(r_{l,0}\) satisfy the following properties, which we will often use
where \(\tilde{p}_l(\lambda )\) are some polynomials of degree q. The first property follows, for example, by considering the homogeneous Neumann problem with initial condition \(u_0=1\). Then the exact solution u and the dG(q) solution \(u_k\) are the same and equal to 1, i.e., \(u=u_k=1\). Hence, all nodal values \(U^m_l=1\) for all \(m=1,2,\dots ,M\) and \(l=0,1,\dots ,q\). For example for \(m=1\), we have
and as a result \(r_{l,0}(0)=1\). The second property in (27) is just a consequence of the first one.
Remark 3
The dG(1) solution \(u_k\) on each subinterval \(I_m\) is of the form
and the rational functions are \(\hat{p}(\lambda )=1+\frac{2}{3}\lambda +\frac{\lambda ^2}{6}\), \(r_{0,0}(\lambda )=1+\frac{2}{3}\lambda \), and \(r_{1,0}(\lambda )=1-\frac{\lambda }{3}\).
For later proof we require two supplementary results.
Lemma 2
Let the rational function r(z) be of the form \(r(z)= \frac{p(z)}{\hat{p}(z)},\) where \(\hat{p}(z)\) is a polynomial of degree \(q+1\) with no roots on the right half complex plane and p(z) is a polynomial of degree q, for some \(q\ge 0\). Then, there exists a constant C independent of \(k>0\), such that for any \(g\in L^p(\Omega )\)
Proof
For simplicity we assume that the roots \(z_1,z_2,\dots ,z_q\) of \(\hat{p}\) are pairwise distinct. If it is not the case, the argument can be slightly modified. For \(q=0\) we have \(r(z) = \frac{c_0}{z-z_0}\) and the desired estimate follows directly by the resolvent estimate (8), since
and therefore by (8)
For \(q>0\) we use the partial fraction decomposition
with some \(c_i \in \mathbb {C}\). Applying the estimate for \(q_0\) to each summand we obtain
which completes the proof. \(\square \)
Lemma 3
Let the rational function r(z) be of the form \(r(z)= \frac{zp(z)}{\hat{p}(z)}\), where \(\hat{p}(z)\) is a polynomial of degree \(q+1\) with no roots on the right-half complex plane and p(z) is a polynomial of degree q, for some \(q\ge 0\). Then for any \(g\in L^p(\Omega )\) with \(\Delta g\in L^p(\Omega )\), \(1\le p\le \infty \), there exists a constant C independent of k such that
Proof
This lemma is just a consequence of the previous one. We set \(\tilde{r}(z)=\frac{p(z)}{\hat{p}(z)}\) and obtain:
The the result follows by Lemma 2. \(\square \)
Lemma 4
Let the rational function r(z) be of the form \(r(z)= \frac{zp(z)}{\hat{p}(z)},\) where \(\hat{p}(z)\) is a polynomial of degree \(q+1\) with no roots on the right half complex plane and p(z) is a polynomial of degree q, for some \(q\ge 1\). Then, there exists a constant C independent of k, such that for any \(g\in L^p(\Omega )\)
Proof
We set \(\tilde{r}(z)=\frac{p(z)}{\hat{p}(z)}\) and obtain:
The estimate
is provided on the top of page 1322 in [36] using a decomposition \(r(z)=r_1(z)+r_2(z)\), where \(r_1(z)=\frac{c}{z-z_0}\), with \(z_0\) being a root of \(\hat{p}(z)\) and c such that the degree of the polynomial in the numerator of \(r_2(z)\) is less or equal \(q-1\). Then the estimate for \(\Delta \tilde{r}_1(-k\Delta ) g\) follows directly by applying a dG(0) type argument and the term \(\Delta \tilde{r}_2(-k\Delta ) g\) is estimated using the Dunford–Taylor formula. \(\square \)
Next we provide some properties of the dG(q) solutions of the homogeneous problem.
Lemma 5
Let \(u_k\) be the solution of (23) with \(u_0 \in L^p(\Omega )\), \(1\le p\le \infty \). Then,
Proof
The proof is given in [36, Thm. 5.1] for the \(L^2(\Omega )\) norm, but the proof is valid for the \(L^p(\Omega )\) norm as well by using the resolvent estimate (8) with respect to the \(L^p(\Omega )\) norm.
Theorem 3
(Homogeneous smoothing estimate) Let \(u_k\) be the solution of (23) with \(u_0 \in L^p(\Omega )\), \(1\le p\le \infty \). Then there exists a constant C independent of k such that
Proof
Again the proof is given in [36, Thm. 5.1] for the \(L^2(\Omega )\) norm, but the proof is valid for the \(L^p(\Omega )\) norm as well by using the resolvent estimate (8) with respect to the \(L^p(\Omega )\) norm. \(\square \)
Remark 4
Notice that the statement of Theorem 3 is equivalent to
which we will use in the following proofs.
Remark 5
Let \(u_k\) be the solution of (23). Then there exists a constant C independent of k such that
or in terms of nodal values
Theorem 4
(Homogeneous smoothing estimate for jumps) Let \(u_k\) be the solution of (23) with \(u_0 \in L^p(\Omega )\), \(1\le p\le \infty \). Then there exists a constant C independent of k such that
where \([u_k]_{0}=U_0^1-u_0\).
Proof
For \(m>1\), using (24), we have
Using (27) and Lemma 3, we obtain
Now by Remark 4 and the assumption on the time mesh (ii), we obtain
That finishes the proof for this case.
For \(m=1\), by Lemma 5 we have,
\(\square \)
Similarly, we can obtain the corresponding result for the time derivative.
Theorem 5
(Homogeneous smoothing estimate for time derivatives) Let \(u_k\) be the solution of (23) with \(u_0 \in L^p(\Omega )\), \(1\le p\le \infty \). Then there exists a constant C independent of k such that
Proof
For \(m>1\), using (22) and (24), we have
By the fact that \(\sum _{l=0}^q \psi _l\left( \frac{t-t_{m-1}}{k_m}\right) =1\) we have \(\sum _{l=0}^q \psi _l'\left( \frac{t-t_{m-1}}{k_m}\right) =0\). Using (27), i.e., \(r_{l,0}(0)=1\) we obtain
where \(\hat{p}(z)\) is the same polynomial as in (25) and \(\tilde{p}_t(z)\) is some polynomial of degree \(q-1\) whose coefficients are time dependent, but uniformly bounded on \(I_m\). Thus again by Lemma 3, we obtain
Remark 4 and the assumption on the time mesh (ii), finishes the proof for \(m>1\).
For \(m=1\), by Lemma 5 we have,
\(\square \)
4.2 Results for the inhomogeneous problem
In this section we establish properties of the dG(q) solution \(u_{k}\in X^q_k\) to the inhomogeneous parabolic equation with \(u_0=0\), that satisfies,
Alternatively, on a single time interval \(I_m\), we have
where
and the rational functions
are as in the homogenous case with \(\hat{p}\) being a polynomial of degree \(q+1\) with no roots on the right half complex plane and \(p_{l,j}\), \(l,j=0,1,\dots ,q\) being polynomials of degree q (cf. [36], page 1322).
Notice that for \(m=1,2,\dots ,M\),
Theorem 6
(Maximal parabolic regularity) Let \(u_k\) satisfy (32) with \(f \in L^s(I;L^p(\Omega ))\) for \(1\le s,p\le \infty \). There exists a constant C independent of k and f such that
Proof
Using (33), we have the following representation
where
with the usual convention that \(\prod _{j=1}^0\) is an empty product. The proof now follows along the lines of Theorem 2. Taking the Laplacian of both sides we obtain
and as a result
By Lemma 4, we have
and by Lemma 2 we also have
On the other hand by Remark 5 for any \(l=0,1,\dots ,q\), since each term in the sum on the right-hand side can be thought of as a homogeneous solution with initial condition \(G_q^n\) at \(t=t_{n-1}\), we have
To establish the result for \(s=\infty \), we observe
where in the last step we used (21). Using (35) we can conclude that for \(s=\infty \)
Similarly, for \(s=1\), we have
Changing the order of summation and using (21) we obtain,
Thus, by using (35), we have
Interpolating between \(s=1\) and \(s=\infty \) we obtain the result for any \(1\le s\le \infty \). \(\square \)
Remark 6
As in the case of dG(0) the appearance of a logarithmic term is natural, since in contrast to the continuous case the choices \(s,p \in \{1,\infty \}\) are allowed. The power of the logarithm can be improved for \(p=2\) or \(s=2\). In fact, we can obtain the following estimates (cf. Remark 2),
and
Theorem 7
(Maximal parabolic regularity for jumps) Let \(u_k\) satisfy (32) with \(f \in L^s(I;L^p(\Omega ))\) for \(1\le s,p\le \infty \). Then there exists a constant C independent of k and f such that
Proof
Using (33) and (36), we have the following representation for the jump terms
Using that \(r_{0,0}-1\) satisfies (27) and using Lemmas 2, 3, and proceeding similarly to the proof of Theorem 6, we have
Now, the proof of the cases \(s=1\) and \(s=\infty \) is identical to the one of the previous Theorem 6 and we have
For \(1<s<\infty \) using the Hölder inequality with \(\frac{1}{s}+\frac{1}{s'}=1\), we obtain,
Hence
Changing the order of summation, we obtain
Taking the s-root we finish the proof. \(\square \)
Theorem 8
Let \(u_k\) satisfy (32). Then there exists a constant C independent of k and f such that
Proof
Similarly to the proof of Theorem 4, using (22) and (33), we have
Using (27) and \(\sum _{l=0}^q \psi _l'\left( \frac{t-t_{m-1}}{k_m}\right) =0\), we can conclude that
where \(\hat{p}(z)\) is the same polynomial as in (25) and \(\tilde{p}_t(z)\) is some polynomial of degree q whose coefficients are time dependent, but uniformly bounded on \(I_m\). Thus again by Lemma 3 and Lemma 4, we obtain
The rest of the proof is identical to the proof of the previous theorem. \(\square \)
4.3 Application to optimal order error estimates
As an application of the maximal parabolic regularity, we show optimal convergence rates for the dG(q) solution. First, we establish that the error is bounded by a certain projection error. A similar result was obtained for the \({L^2(I;L^2(\Omega ))}\) norm in [29]. First, we define a projection \(\pi _k\) for \(u \in C(I,L^2(\Omega ))\) with \(\pi _k u|_{I_m} \in P_q(L^2(\Omega ))\) for \(m=1,2,\dots ,M\) on each subinterval \(I_m\) by
In the case \(q = 0\), \(\pi _ku\) is defined solely by the second condition.
Theorem 9
Let u be the solution to (1) with \(u \in C(\bar{I}; L^p(\Omega ))\) and \(u_k\) be its dG(q) approximation (6), for \(q\ge 0\). Then there exists a constant C independent of k such that
where the projection \(\pi _k\) is defined above in (41).
Proof
Put \(e:=u-u_k=(u-\pi _k u)+(\pi _ku-u_k):=\eta _k+\xi _k\). For \(1\le s,p<\infty \), we have
For each such \(\psi \), we consider a dual problem for \(z_k \in X_k^q\) satisfying
Thus, we have
Using the Hölder inequality, we find
On the other hand using that \(B(u-u_k,\chi _k)=0\) for any \(\chi _k\in X_k^q\), we obtain
where we used that the first sum vanishes due to (41a) and the sum involving jumps due to (41b). Integrating by parts in space, using the Hölder inequality and Theorem 6, we obtain
Combining the estimates for \(J_1\) and \(J_2\) we obtain the result. \(\square \)
If the exact solution is sufficiently smooth then the above result easily leads to an optimal convergence rate, modulo a logarithmic term.
Corollary 3
Let \(u\in W^{q+1,s}(I;L^p(\Omega ))\) be the solution to (1) and \(u_k\) be its dG(q) approximation for \(q\ge 0\). Then there exists a constant C independent of k such that
Remark 7
The above result can be extended to the case of non-homogeneous Dirichlet boundary conditions. Let \(g \in C(I;L^2(\Omega ))\cap L^2(I;H^1(\Omega ))\) be given and consider the equation
It turns out, that it is convenient to use \(\pi _k g\) as boundary conditions for the semidiscrete solution, i.e.
Then following the lines of the proof of Theorem 9 and using that \(\xi _k = \pi _k u - u_k\) has homogeneous boundary conditions, i.e., \(\xi _k \in X_k^q\), we obtain
Under an additional assumption on \(\Omega \) that for any \(v \in H^1_0(\Omega )\) with \(\Delta v \in L^{p'}(\Omega )\) the estimate
holds, we obtain
The above assumption is fulfilled, for example, if on \(\Omega \) the \(W^{2,p'}\) elliptic regularity holds.
5 Fully discrete solutions
In this section, we consider the fully discrete approximation of the equation (1). From now on we assume that the domain \(\Omega \) is a polygonal/polyhedral convex domain. For \(h \in (0, h_0]\); \(h_0 > 0\), let \(\mathcal {T}\) denote a quasi-uniform triangulation of \(\Omega \) with mesh size h, i.e., \(\mathcal {T} = \{\tau \}\) is a partition of \(\Omega \) into cells (triangles or tetrahedrons) \(\tau \) of diameter \(h_\tau \) such that for \(h=\max _{\tau } h_\tau \),
hold. Let \(V_h\) be the set of all functions in \(H^1_0(\Omega )\) that are polynomials of degree r on each \(\tau \), i.e., \(V_h\) is the usual space of conforming finite elements. To obtain the fully discrete approximation we consider the space-time finite element space
We define a fully discrete analog \(u_{kh} \in X^{q,r}_{k,h}\) of \(u_k\) introduced in (6) by
Moreover, we introduce the discrete Laplace operator \(\Delta _h :V_h \rightarrow V_h\) by
The semidiscrete results from the first part of the paper translate almost immediately to the fully discrete setting provided we have the corresponding resolvent estimate,
with some constant C independent of h. Such a result was established in [38] for smooth domains. Later it was extended to convex polyhedral domains in [39] (for some \(\gamma >0\)) via stability and smoothing properties of the semigroup \(E_h(t)=e^{-\Delta _ht}\) and directly for an arbitrary \(\gamma >0\) but with logarithmic dependence of the constant C on h in [40].
5.1 Result for the homogeneous problem
Let \(u_{kh}\in X^{q,r}_{k,h}\) be the fully discrete dG(q)cG(r) solution to the parabolic equation with \(f\equiv 0\), i.e.
Theorem 10
(Fully discrete homogeneous smoothing estimate) Let \(u_{kh}\) be a solution of (45) with \(u_0 \in L^p(\Omega )\), \(1\le p\le \infty \). Then there exists a constant C independent of k and h such that
for \(m=1,2,\dots ,M\).
5.2 Results for the inhomogeneous problem
Let \(u_{kh}\in X^{q,r}_{k,h}\) be the dG(q)cG(r) solution to the inhomogeneous parabolic equation with \(u_0=0\), i.e.
Theorem 11
(Fully discrete maximal parabolic regularity) Let \(u_{kh}\) satisfy (46) with \(f \in L^s(I;L^p(\Omega ))\), \(1\le s,p\le \infty \). Then there exists a constant C independent of k and h such that
with obvious notation changes in the case of \(s=\infty \).
5.3 Application to optimal order error estimates
Similarly to the semidiscrete case, as an application of the maximal parabolic regularity, we show optimal convergence rates for the dG(q)cG(r) solution.
Theorem 12
Let u be the solution to (1) with \(u \in C(\bar{I}; L^p(\Omega ))\) and \(u_{kh}\) be the dG(q)cG(r) solution for \(q\ge 0\) and \(r\ge 1\). Then there exists a constant C independent of k and h such that for \(1\le s,p<\infty \),
where the projection \(\pi _k\) is defined in (41), \(P_h :L^2(\Omega )\rightarrow V_h\) is the orthogonal \(L^2\) projection and \(R_h :H^1_0(\Omega )\rightarrow V_h\) is the Ritz projection.
Proof
Put \(e:=u-u_{kh}=(u-P_h\pi _k u)+(P_h\pi _ku-u_{kh}):=\eta _{kh}+\xi _{kh}\). For \(1\le s,p<\infty \), we have
For each such \(\psi \), consider a dual problem
Thus, we have
Using the Hölder inequality, the triangle inequality, the stability of the \(L^2\) projection \(P_h\) in \(L^p(\Omega )\) and the approximation properties of \(\pi _k\) and \(P_h\), we find
On the other hand, using that \(B(u-u_{kh},\chi _{kh})=0\) for any \(\chi _{kh}\in X^{q,r}_{k,h}\), and the properties of the \(L^2\) projection and the properties of \(\pi _k\), we obtain
where we used that the first sum vanishes due to (41a) and the sum involving jumps due to (41b). Using the properties of the Ritz projection, integrating by parts in space, and using the Hölder inequality and Theorem 6, we obtain
Combining the estimates for \(J_1\) and \(J_2\) we obtain the result. \(\square \)
Corollary 4
If the solution u to (1) satisfies \(u\!\in \! W^{q+1,s}(I;L^p(\Omega ))\cap L^{s}(I;W^{r+1,p}(\Omega ))\) and \(\Omega \) such that elliptic \(W^{2,p'}\)- regularity holds, then there exists a constant C independent of k and h such that
6 Fully discrete results in general norms
For the future references we provide discrete maximal parabolic regularity results in general norms. For example, we use these results to establish pointwise best approximation estimates in [9] for fully discrete Galerkin solutions.
Let \(\Omega \) be a Lipschitz domain and let \(\mathcal {T} = \{\tau \}\) be an arbitrary partition of \(\Omega \) into cells \(\tau \) (triangles, tetrahedrons, quads, or hexahedrons, not necessary quasi-uniform). Let \(V_h\) be the set of all functions in \(H^1_0(\Omega )\) that belong to a certain polynomial space (i.e., \(P_r\) or \(Q_r\)) on each \(\tau \). As before, we define a fully discrete solution \(u_{kh} \in X^{q,r}_{k,h}\) by
where
As in the previous section, we introduce the discrete Laplace operator \(\Delta _h :V_h \rightarrow V_h\) by
and the orthogonal \(L^2\) projection \(P_h :L^2(\Omega ) \rightarrow V_h\) by
Let be a norm on \(V_h\) such that for some \(\gamma \in (0,\frac{\pi }{2})\) the following resolvent estimate holds,
for all \(\chi \in V_h\), where \(\Sigma _{\gamma }\) is defined in (9) and the constant \(M_h\) is independent of z.
For quasi-uniform meshes, this assumption is fulfilled for with a constant \(M_h \le C\) independent of h, see [39], as discussed and exploited above. For a weighted norm with the weight \(\sigma _{x_0}(x)=\sqrt{|x-x_0|^2+h^2}\) and \(M_h \le C |\ln {h}|\) we established this estimate in [9], and used the corresponding result to obtain interior (local) pointwise estimates. Moreover, the resolvent estimate (49) is known also to hold in \({L^p(\Omega )}\) norms on a class of non quasi-uniform meshes as well, see [41].
6.1 Smoothing estimates for the homogeneous problem in general norms
For the homogeneous heat equation (1), i.e. \(f=0\) and its discrete approximation \(u_{kh} \in X^{q,r}_{k,h}\) defined by
we have the following smoothing result.
Theorem 13
(Fully discrete smoothing estimate in general norms) Let be a norm on \(V_h\) fulfilling the resolvent estimate (49). Let \(u_{kh}\) be the solution of (50). Then, there exists a constant C independent of k and h such that
for \(m=1,2,\dots ,M\), where \(P_h :L^2(\Omega )\rightarrow V_h\) is the orthogonal \(L^2\) projection. For \(m=1\) the jump term is understood as \([u_{kh}]_0 = u_{kh,0}^+-P_h u_0\).
6.2 Discrete maximal parabolic estimates for the inhomogeneous problem in general norms
Now, we consider the inhomogeneous heat equation (1), with \(u_0=0\) and its discrete approximation \(u_{kh} \in X^{q,r}_{k,h}\) defined by
Theorem 14
(Discrete maximal parabolic regularity in general norms) Let be a norm on \(V_h\) fulfilling the resolvent estimate (49) and let \(1 \le s \le \infty \). Let \(u_{kh}\) be a solution of (51). Then, there exists a constant C independent of k and h such that
where \(P_h :L^2(\Omega )\rightarrow V_h\) is the orthogonal \(L^2\) projection and with obvious notation change in the case of \(s=\infty \). For \(m=1\) the jump term is understood as \([u_{kh}]_0=u_{kh,0}^+\).
The proofs of the above two results are identical to the proofs of the corresponding time discrete results from Sect. 4, provided the resolvent estimate (49) holds.
References
Coulhon, T., Duong, X.T.: Maximal regularity and kernel bounds: observations on a theorem by Hieber and Prüss. Adv. Differ. Equ. 5(1–3), 343–368 (2000)
Haller-Dintelmann, R., Rehberg, J.: Maximal parabolic regularity for divergence operators including mixed boundary conditions. J. Differ. Equ. 247(5), 1354–1396 (2009). doi:10.1016/j.jde.2009.06.001
Hieber, M., Prüss, J.: Heat kernels and maximal \(L^p\)-\(L^q\) estimates for parabolic evolution equations. Commun. Partial Differ. Equ. 22(9–10), 1647–1669 (1997). doi:10.1080/03605309708821314
Hömberg, D., Meyer, C., Rehberg, J., Ring, W.: Optimal control for the thermistor problem. SIAM J. Control Optim. 48(5), 3449–3481 (2009/10). doi:10.1137/080736259
Krumbiegel, K., Rehberg, J.: Second order sufficient optimality conditions for parabolic optimal control problems with pointwise state constraints. SIAM J. Control Optim. 51(1), 304–331 (2013). doi:10.1137/120871687
Kunisch, K., Pieper, K., Vexler, B.: Measure valued directional sparsity for parabolic optimal control problems. SIAM J. Control Optim. 52(5), 3078–3108 (2014)
Leykekhman, D., Vexler, B.: Optimal a priori error estimates of parabolic optimal control problems with pointwise control. SIAM J. Numer. Anal. 51(5), 2797–2821 (2013). doi:10.1137/120885772
Leykekhman, D., Vexler, B.: A priori error estimates for three dimensional parabolic optimal control problems with pointwise control. SIAM J. Control Optim. (2016) (Accepted)
Leykekhman, D., Vexler, B.: Pointwise best approximation results for Galerkin finite element solutions of parabolic problems. SIAM J. Numer. Anal. 54(3), 1365–1384 (2016). doi:10.1137/15M103412X
Geissert, M.: Discrete maximal \(L_p\) regularity for finite element operators. SIAM J. Numer. Anal. 44(2), 677–698 (electronic) (2006). doi:10.1137/040616553
Geissert, M.: Applications of discrete maximal \(L_p\) regularity for finite element operators. Numer. Math. 108(1), 121–149 (2007). doi:10.1007/s00211-007-0110-1
Li, B.: Maximum-norm stability and maximal \(L^p\) regularity of FEMs for parabolic equations with Lipschitz continuous coefficients. Numer. Math. 131(3), 489–516 (2015)
Blunck, S.: Analyticity and discrete maximal regularity on \(L_p\)-spaces. J. Funct. Anal. 183(1), 211–230 (2001). doi:10.1006/jfan.2001.3740
Blunck, S.: Maximal regularity of discrete and continuous time evolution equations. Stud. Math. 146(2), 157–176 (2001). doi:10.4064/sm146-2-3
Portal, P.: Maximal regularity of evolution equations on discrete time scales. J. Math. Anal. Appl. 304(1), 1–12 (2005). doi:10.1016/j.jmaa.2004.09.003
Guidetti, D.: Backward Euler scheme, singular Hölder norms, and maximal regularity for parabolic difference equations. Numer. Funct. Anal. Optim. 28(3–4), 307–337 (2007). doi:10.1080/01630560701285594
Guidetti, D.: Maximal regularity for parabolic equations and implicit Euler scheme. In: Mathematical analysis seminar, University of Bologna Department of Mathematics: Academic Year 2003/2004. Academic Year 2004/2005 (Italian), pp. 253–264. Tecnoprint, Bologna (2007)
Ashyralyev, A., Sobolevskiĭ, P.E.: Well-posedness of parabolic difference equations. In: Operator Theory: Advances and Applications, vol. 69. Birkhäuser Verlag, Basel (1994). doi:10.1007/978-3-0348-8518-8 (Translated from the Russian by A. Iacob)
Eriksson, K., Johnson, C., Thomée, V.: Time discretization of parabolic problems by the discontinuous Galerkin method. RAIRO Modél. Math. Anal. Numér. 19(4), 611–643 (1985)
Eriksson, K., Johnson, C.: Adaptive finite element methods for parabolic problems. I. A linear model problem. SIAM J. Numer. Anal. 28(1), 43–77 (1991). doi:10.1137/0728003
Eriksson, K., Johnson, C.: Adaptive finite element methods for parabolic problems. II. Optimal error estimates in \(L_\infty L_2\) and \(L_\infty L_\infty \). SIAM J. Numer. Anal. 32(3), 706–740 (1995). doi:10.1137/0732033
Schmich, M., Vexler, B.: Adaptivity with dynamic meshes for space-time finite element discretizations of parabolic equations. SIAM J. Sci. Comput. 30(1), 369–393 (2007/08). doi:10.1137/060670468
Schötzau, D., Wihler, T.P.: A posteriori error estimation for \(hp\)-version time-stepping methods for parabolic partial differential equations. Numer. Math. 115(3), 475–509 (2010). doi:10.1007/s00211-009-0285-8
Lasaint, P., Raviart, P.A.: On a finite element method for solving the neutron transport equation. In: Mathematical Aspects of Finite Elements in Partial Differential Equations (Proc. Sympos., Math. Res. Center, Univ. Wisconsin, Madison, Wis., 1974), pp. 89–123. Publication No. 33. Math. Res. Center, Univ. of Wisconsin-Madison, Academic Press, New York (1974)
Richter, T., Springer, A., Vexler, B.: Efficient numerical realization of discontinuous Galerkin methods for temporal discretization of parabolic problems. Numer. Math. 124(1), 151–182 (2013). doi:10.1007/s00211-012-0511-7
Becker, R., Meidner, D., Vexler, B.: Efficient numerical solution of parabolic optimization problems by finite element methods. Optim. Methods Softw. 22(5), 813–833 (2007). doi:10.1080/10556780701228532
Meidner, D., Vexler, B.: A priori error analysis of the Petrov–Galerkin Crank–Nicolson scheme for parabolic optimal control problems. SIAM J. Control Optim. 49(5), 2183–2211 (2011). doi:10.1137/100809611
Chrysafinos, K.: Discontinuous Galerkin approximations for distributed optimal control problems constrained by parabolic PDE’s. Int. J. Numer. Anal. Model. 4(3–4), 690–712 (2007)
Meidner, D., Vexler, B.: Adaptive space-time finite element methods for parabolic optimization problems. SIAM J. Control Optim. 46(1), 116–142 (electronic) (2007). doi:10.1137/060648994
Meidner, D., Vexler, B.: A priori error estimates for space-time finite element approximation of parabolic optimal control problems. Part I: problems without control constraints. SIAM J. Control Optim. 47(3), 1150–1177 (2008)
Meidner, D., Vexler, B.: A priori error estimates for space-time finite element approximation of parabolic optimal control problems. Part II: problems with control constraints. SIAM J. Control Optim. 47(3), 1301–1329 (2008)
Meidner, D., Rannacher, R., Vexler, B.: A priori error estimates for finite element discretizations of parabolic optimization problems with pointwise state constraints in time. SIAM J. Control Optim. 49(5), 1961–1997 (2011). doi:10.1137/100793888
Thomée, V.: Galerkin finite element methods for parabolic problems. Springer Series in Computational Mathematics, vol. 25, 2nd edn. Springer, Berlin (2006)
Shen, Z.W.: Resolvent estimates in \(L^p\) for elliptic systems in Lipschitz domains. J. Funct. Anal. 133(1), 224–251 (1995). doi:10.1006/jfan.1995.1124
Haller-Dintelmann, R., Meyer, C., Rehberg, J., Schiela, A.: Hölder continuity and optimal control for nonsmooth elliptic problems. Appl. Math. Optim. 60(3), 397–428 (2009). doi:10.1007/s00245-009-9077-x
Eriksson, K., Johnson, C., Larsson, S.: Adaptive finite element methods for parabolic problems. VI. Analytic semigroups. SIAM J. Numer. Anal. 35(4), 1315–1325 (electronic) (1998). doi:10.1137/S0036142996310216
Egert, M., Rozendaal, J.: Convergence of subdiagonal Padé approximations of \(C_0\)-semigroups. J. Evol. Equ. 13(4), 875–895 (2013). doi:10.1007/s00028-013-0207-1
Bakaev, N.Y., Thomée, V., Wahlbin, L.B.: Maximum-norm estimates for resolvents of elliptic finite element operators. Math. Comput. 72(244), 1597–1610 (electronic) (2003). doi:10.1090/S0025-5718-02-01488-6
Li, B., Sun, W.: Maximal \(l^p\) analysis of finite element solutions for parabolic equations with nonsmooth coefficients in convex polyhedra. (Preprint) arXiv:1501.07345 (2015)
Leykekhman, D., Vexler, B.: Finite element pointwise results on convex polyhedral domains. SIAM J. Numer. Anal. 54(2), 561–587 (2016). doi:10.1137/15M1013912
Bakaev, N.Y., Crouzeix, M., Thomée, V.: Maximum-norm resolvent estimates for elliptic finite element operators on nonquasiuniform triangulations. M2AN Math. Model. Numer. Anal. 40(5), 923–937 (2006). doi:10.1051/m2an:2006040 (2007)
Acknowledgments
The authors would like to thank Dominik Meidner and Konstantin Pieper for the careful reading of the manuscript and providing valuable suggestions that help to improve the presentation of the paper.
Author information
Authors and Affiliations
Corresponding author
Additional information
The author was partially supported by NSF Grant DMS-1522555.
Rights and permissions
About this article
Cite this article
Leykekhman, D., Vexler, B. Discrete maximal parabolic regularity for Galerkin finite element methods. Numer. Math. 135, 923–952 (2017). https://doi.org/10.1007/s00211-016-0821-2
Received:
Revised:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00211-016-0821-2