INTRODUCTION

Partial differential equations with time-fractional derivatives arise in mathematical modeling of anomalous diffusion and dynamic processes in materials with memory. Numerous articles are devoted to the approximation of boundary value problems for linear problems with time-fractional derivatives. In recent years, the numerical methods has been actively developed for equations with time-fractional derivatives and the elliptic parts containing the non-linearities. Thus, implicit mesh schemes for the equations with Caputo time-fractional derivatives and with right-hand sides depending on the solution are studied in [1–3]. In [4] an implicit scheme is investigated for the nonlinear diffusion equation with a coefficient depending on the solution and with Caputo time-fractional derivative. In [5–10], analysis of ADI schemes was performed for linear time-fractional equations and for a class of semilinear time-fractional equations.

The article [11] proposes a new mathematical model for a viscoelastic-plastic process by formulating a temporary fractional equation containing an elliptic operator, which depends on the gradient of the solution. In [12], the authors construct and study backward Euler and locally one dimensional schemes approximating Dirichlet problem with Caputo time-fractional derivative and a quasilinear elliptic part without mixed derivatives.

In this article, we consider a quasilinear parabolic problem with a time fractional derivative of the Caputo type and mixed boundary conditions. The quasilinear elliptic part of the problem defines a uniformly monotone and Lipschitz operator, which is energy equivalent to the Laplace operator. For this problem, we construct and investigate a linear regularized semidiscrete scheme in which the regularization term contains the Laplace operator and time-fractional derivative is approximated using well-known \(L1\)-approximation. Stability estimates for the semidiscrete problem by the variational method are obtained. Accuracy estimates are given provided that the initial data and the solution to the differential problem are sufficiently smooth. For discrete schemes obtained as a result of mesh approximations of the elliptic part of the semidiscrete problem, the stability estimates remain valid. They are also generalized for factorized schemes.

1 FORMULATION OF THE PROBLEM

Let \(\Omega\subset\mathbb{R}^{2}\) be a polygon with the boundary \(\partial\Omega=\Gamma_{D}\cup\Gamma_{N},\) \({\textrm{meas }}\Gamma_{D}>0\), and \(Q=\Omega\times(0,T]\), \(\Sigma_{D}=\Gamma_{D}\times(0,T]\), \(\Sigma_{N}=\Gamma_{N}\times(0,T]\). Define a Caputo-type time-fractional derivative:

$$\displaystyle\mathcal{D}_{t}y(t)=\int\limits_{0}^{t}G(t-s)\dfrac{\partial y}{\partial s}(s)ds$$
(1)

with a kernel \(G(t)\) satisfying the assumptions:

$$G(t):(0,+\infty)\to{\mathbb{R}^{+}}\text{ is continuous and strictly decreasing function,}\quad\displaystyle\int\limits_{0}^{+\infty}G(t)dt<\infty.$$
(2)

The conditions (2) are satisfied for

  • the generalized Caputo fractional derivative with \({G(t)}=\dfrac{r(t)}{\Gamma(1-\alpha)t^{\alpha}}\), \(0<\alpha<1,\) \(\Gamma(x)\) is gamma-function, a weighting function \(r(t)\in C^{2}[0,T]\), \(r(t)>0\) and \(r^{\prime}(t)\leqslant 0\) for all \(t\in[0,T]\) (\(r(t)\equiv 1\) corresponds to the classical Caputo fractional derivative);

  • the multi-term fractional derivative with \({G(t)}=\sum_{k=1}^{s}\dfrac{c_{k}}{\Gamma(1-\alpha_{k})t^{\alpha_{k}}},\quad 0<\alpha_{1}<\ldots<\alpha_{s}<1,\) \(c_{k}>0\).

See the definitions of the corresponding derivatives, e.g., in [13–15]).

Let us define a second order quasilinear elliptic differential operator \(L\) and the co-normal derivative \(\frac{\partial u}{\partial\nu_{L}}\) by the equalities

$$\displaystyle Lu=-\sum\limits_{i=1}^{2}\frac{\partial}{\partial x_{i}}g_{i}\big{(}x,t,u,\nabla u\big{)}+g_{0}\big{(}x,t,u,\nabla u\big{)},\quad\displaystyle\frac{\partial u}{\partial\nu_{L}}=\sum\limits_{i=1}^{2}g_{i}\big{(}x,t,u,\nabla u\big{)}\cos(n,x_{i}),$$
(3)

where \(n\) is the unit vector of the outward normal to \(\Sigma_{N}\). Suppose that the nonlinear coefficients \(g_{i}(t,p)=g_{i}(x,t,p_{0},p_{1},p_{2})\) of the equation satisfy the following assumptions for all \((x,t)\in\bar{Q}\) and \(p,q,r\in{\mathbb{R}^{3}}\):

$$\displaystyle g_{i}(t,p),\quad\text{for}\quad i,k=0,\ldots,2\quad\text{are continuous},$$
$$\displaystyle\sum\limits_{i=0}^{2}(g_{i}(t,p)-g_{i}(t,q))(p_{i}-q_{i})\geqslant c_{0}\sum\limits_{i=1}^{2}(p_{i}-q_{i})^{2},\quad c_{0}>0,$$
$$\displaystyle\sum\limits_{i=0}^{2}(g_{i}(t,p)-g_{i}(t,q))r_{i}\leqslant c_{1}\left(\sum\limits_{i=0}^{2}(g_{i}(t,p)-g_{i}(t,q))(p_{i}-q_{i})\right)^{1/2}\left(\sum\limits_{i=0}^{2}r_{i}^{2}\right)^{1/2}.$$
(4)

Consider a quasilinear parabolic problem with the time-fractional derivative and mixed boundary conditions:

$$\displaystyle\mathcal{D}_{t}u+Lu=f\quad\textrm{in}\quad Q_{T},$$
$$\displaystyle u=0\quad\textrm{on}\quad\Sigma_{D},\quad\frac{\partial u}{\partial\nu_{L}}=q\quad\textrm{on}\quad\Sigma_{N},$$
$$u=0\quad\textrm{for}\quad t=0,\quad x\in\Omega.$$
(5)

We assume that given functions \(f=f(x,t)\) and \(q=q(x,t)\) are continuous in \(t\) in order to use the simplest mesh approximations and do not specify their properties in details, since we do not investigate the existence of a solution to problem (5).

Let us make only a few remarks about well-known results on the existence of weak solutions of parabolic equations with time-fractional derivatives. The existence of a regular weak solution to Dirichlet problem for the case of linear operator \(L\) and classical Caputo fractional derivative was proved in [16]. A priori estimate for the solution in the norm of \(H^{\alpha}(0,T;L^{2}(\Omega))\cap L^{2}(0,T;H_{0}^{1}(\Omega)\cap H^{2}(\Omega))\) through the \(L^{2}(Q)\)-norm of \(f\) was derived. The existence of a weak solution from Besov space \(B^{\alpha/2}(Q_{T})=H^{\alpha/2}(0,T;L^{2}(\Omega))\cap L^{2}(0,T;H^{1}(\Omega))\) is substantiated in [12] for a quasilinear Dirichlet boundary value problem with classical Caputo derivative and a particular case of differential operator (3). This result can be generalized for the case of boundary value problem with elliptic operator (3) and Caputo time derivative. In a general case the problem of the existence of a unique solution to (5) seems to be open. In the future, we will assume that problem (5) has a smooth solution, and will focus on studying the mesh approximation of this problem.

2 APPROXIMATION

Multiplying the differential equation (5) by a sufficiently smooth function \(v(x):v|_{\Gamma_{D}}=0\), we obtain a variational equation that can serve as a basis for the constructing its approximations:

$$\displaystyle\int\limits_{\Omega}\mathcal{D}_{t}uvdx+\int\limits_{\Omega}\left(\sum\limits_{i=1}^{2}g_{i}\big{(}x,t,u,\nabla u\big{)}\frac{\partial v}{\partial x_{i}}+g_{0}\big{(}x,t,u,\nabla u\big{)}v\right)dx=\int\limits_{\Omega}fvdx+\int\limits_{\Gamma_{N}}qvd\Gamma.$$

Let \(V=\{y\in H^{1}(\Omega);y=0\quad\textrm{a.e.}\quad x\in\Gamma_{D}\}\) with the norm \(||u||_{V}^{2}=\int\limits_{\Omega}\big{(}u^{2}+|\nabla u|^{2}\big{)}dx\), and let \(||u||_{0}^{2}=\int\limits_{\Omega}u^{2}dx,\) \(||u||_{1}^{2}=\int\limits_{\Omega}|\nabla u|^{2}dx\) and \(||u||_{\Gamma}^{2}=\int\limits_{\Gamma_{N}}u^{2}d\Gamma.\) It is well-known that

$$||u||_{0}\leqslant c_{\Omega}||u||_{1},\quad||v||_{\Gamma}\leqslant c_{\Gamma}||v||_{1}\quad\forall v\in V.$$
(6)

Due to the first inequality in (6) the norms \(||u||_{V}\) and \(||u||_{1}\) are equivalent. By virtue of this fact and the assumptions (4) the form

$$\displaystyle\displaystyle a(t,u,v)=\int\limits_{Q_{T}}\Big{(}\sum\limits_{i=1}^{2}g_{i}\big{(}x,t,u,\nabla u\big{)}\frac{\partial v}{\partial x_{i}}+g_{0}\big{(}x,t,u,\nabla u\big{)}v\Big{)}dxdt$$

is well-defined on \(V\times V\) for all \(t\) and has the following properties:

$$\displaystyle a(t,u,v-u)-a(t,v,v-u)\geqslant c_{0}||u-v||_{1}^{2}\quad\forall u,v\in V,$$
$$\displaystyle a(t,u-v,w)\leqslant c_{2}||u-v||_{1}||w||_{1}\quad\forall u,v,w\in V,\quad c_{2}=c_{2}(c_{1},c_{\Omega}).$$
(7)

Let \(\omega_{\tau}=\{t_{j}=j\tau,j=0,1,\ldots M;M\tau=T\}\) be a uniform mesh on the segment \([0,T]\) and let us use the notation \(y^{j}=y(t_{j})\) for a continuous function \(y(t)\). The \(L1\)-approximation of the time derivative \(\mathcal{D}_{t}y(t)=\int\limits_{0}^{t}G(t-s)\dfrac{\partial y}{\partial s}(s)ds\) of a continuous function \(y(t)\) at a point \(t_{k}\in\omega_{\tau}\) is given by the following relations:

$$\displaystyle\mathcal{D}_{t}y(t_{k})\approx\partial_{t}\bar{y}^{k}=d_{1}y^{k}+\sum\limits_{j=1}^{k-1}(d_{j+1}-d_{j})y^{k-j}-d_{k}y^{0},$$
$$\displaystyle d_{j}=\dfrac{1}{\tau}\int\limits_{t_{k-j}}^{t_{k-j+1}}G(t_{k}-s)ds=\dfrac{1}{\tau}\int\limits_{(j-1)\tau}^{j\tau}G(u)du,\quad\bar{y}^{k}=(y^{0},y^{1},\ldots,y^{k}).$$

The properties (2) imply that the coefficients satisfy the inequalities

$$d_{1}>d_{2}>\cdots>d_{M}>0.$$
(8)

First, we consider the semidiscrete implicit scheme approximating problem (5): find \((y(x,t_{1}),y(x,t_{2}),\ldots,y(x,t_{M}))\) with \(y(x,t_{k})\in V\) and \(y(x,0)=0\), such that for all \(k\)

$$\int\limits_{\Omega}\partial_{t}\bar{y}^{k}(x)v(x)dx+a(t_{k},y^{k},v)=\int\limits_{\Omega}f(x,t_{k})v(x)dx+\int\limits_{\Gamma_{N}}q(x,t_{k})vd\Gamma\quad\forall v\in V.$$
(9)

Since \(\partial_{t}\bar{y}^{k}(x)=d_{1}y^{k}(x)+\sum\limits_{j=1}^{k-1}(d_{j+1}-d_{j})y^{k-j}(x),\) then a \(k\)-th equation of system (9) is a variational equation

$$\int\limits_{\Omega}d_{1}y^{k}(x)v(x)dx+a(t_{k},y^{k},v)=F_{k}(v)$$
(10)

with right-hand side

$$\displaystyle F_{k}(v)=\int\limits_{\Omega}f(x,t_{k})v(x)dx+\int\limits_{\Gamma_{N}}q(x,t_{k})vd\Gamma-\sum\limits_{j=1}^{k-1}(d_{j+1}-d_{j})\int\limits_{\Omega}\nabla y^{k-j}(x)\cdot\nabla v(x)dx.$$

Due to the properties (7) the left side of (10) defines a continuous and uniformly monotone operator from \(V\) to its dual \(V^{*}\). Using the inequalities (6), it is easy to see that linear functional \(F_{k}\in V^{*}\) for \(f(x,t_{k})\in L^{2}(\Omega)\) and \(q(x,t_{k})\in L^{2}(\Gamma_{N})\). Thus, the existence of a unique solution to problem (10) follows from [17], Chapter 2, section 2.1.

Below, instead of the nonlinear problem (9), a linear regularized problem is constructed. To shorten writing, we will use the notations

$$\displaystyle(u,v)=\int\limits_{\Omega}uvdx,\quad a_{k}(u,v)=a(t_{k},u,v)\quad(f^{k},v)=\displaystyle\int\limits_{\Omega}f(x,t_{k})v(x)dx+\int\limits_{\Gamma_{N}}q(x,t_{k})vd\Gamma.$$

Let us introduce a bilinear bounded form \(r(u,v)\) on the space \(V\times V\), which satisfies the inequality

$$r(u,u)\geqslant||u||_{1}^{2}\quad\forall u\in V.$$
(11)

A particular case of this bilinear form is \(r(u,v)=\int\limits_{\Omega}\nabla u\cdot\nabla vdx\) for \(u,v\in V.\)

The regularized scheme approximating (5) is as follows: for a given \(y^{0}=0\) find \(y^{k}\) for \(k\geqslant 1\) from the system of linear equations

$$(\partial_{t}\bar{y}^{k},v)+\sigma r(y^{k}-y^{k-1},v)+a(t_{k-1},y^{k-1},v)=(f^{k},v)\quad\forall v\in V,\quad\sigma>0.$$
(12)

For a fixed \(k\) the mesh scheme (12) can be written as

$$\displaystyle d_{1}(y^{k},v)+\sigma r(y^{k}-y^{k-1},v)+\sum\limits_{j=1}^{k-1}(d_{j+1}-d_{j})(y^{k-j},v)+a_{k-1}(y^{k-1},v)=(f^{k},v),$$
(13)

The linear variational problem (13) has a unique solution for any \(\sigma\geqslant 0\) by Lax–Milgram theorem.

The essential point used in proving stability estimates is the property of the \(L1\)-approximation of the time-fractional derivative, which is given by the following lemma:

Lemma 1 ([12]). Let

$$D=\begin{pmatrix}d_{2}&0&0&0&\cdots&0&0&0\\ 0&d_{2}&0&0&\cdots&0&0&0\\ d_{3}-d_{2}&0&d_{2}&0&\cdots&0&0&0\\ ...&...&...&...&\cdots&...&...\\ d_{M}-d_{M-1}&d_{M-1}-d_{M-2}&d_{M-2}-d_{M-3}&...&\cdots&d_{3}-d_{2}&0&d_{2}\end{pmatrix}.$$

Then the symmetric matrix \(1/2(D+D^{T})\) is positive definite.

Using the matrix \(D\), the approximation of time derivative at a point \(t_{k}=k\tau\in\omega_{\tau}\) can be written as

$$\partial_{t}y(t_{k})=(d_{1}-d_{2})(y^{k}-y^{k-1})+\big{(}D_{k}\bar{y}^{k}\big{)}^{k},$$

where \(D_{k}\) is \(k\times k\) leading submatrix of \(B\), and \(1/2(D_{k}+D_{k}^{T})\) is positive definite along with \(1/2(D+D^{T})\):

$$\sum\limits_{k=1}^{m}\big{(}D_{k}\bar{y}^{k}\big{)}^{k}y^{k}\geqslant 0\quad\forall\bar{y}^{k}=(y^{1},y^{2},\ldots,y^{k}).$$
(14)

Introduce also the bilinear, symmetric and positive definite form on \(V\times V\):

$$b(u,v)=(d_{1}-d_{2})(u,v)+\sigma r(u,v).$$

Now the equation (13) can be rewritten as

$$\displaystyle b(y^{k}-y^{k-1},v)+(\big{(}D_{k}\bar{y}^{k}\big{)}^{k},v)+a_{k-1}(y^{k-1},v)=(f^{k},v).$$
(15)

For generality, we derive an a priori estimate for the case of a nonzero initial value. For \(i=1,2\) we denote by \(y_{i}\) the solutions of the equation (13) with the right-hand sides \(f_{i}\) and the initial values \(y^{0}_{i}\). Let also \(y=y_{1}-y_{2}\), \(y^{0}=y_{1}^{0}-y_{2}^{0}\) and \(\psi=f_{1}-f_{2}\).

Lemma 2. Let the conditions (7) be satisfied and the regularization parameter satisfies the condition

$$\sigma=\frac{c_{2}^{2}}{2c_{0}}+\delta,\delta>0.$$
(16)

Then there is a constant \(\mu=\mu(\delta)\) such that for all \(m=1,2,\ldots M\) the following a priori estimate for the regularized mesh scheme (13) holds:

$$\displaystyle(d_{1}-d_{2})||y^{m}||_{0}^{2}+\mu\sum\limits_{k=1}^{m}||y^{k}||^{2}_{1}\leqslant||y^{0}||_{0}^{2}+r(y^{0},y^{0})+2\sum\limits_{k=1}^{m}|\big{(}\psi^{k},y^{k})\big{|}.$$
(17)

Proof. From the equations (24) written for \(y^{k}_{1}\) and \(y^{k}_{2}\) with test function \(v=2(y^{k}_{1}-y^{k}_{2})=2y^{k}\) we obtain:

$$2b(y^{k}-y^{k-1},y^{k})+2(\big{(}D_{k}\bar{y}^{k}\big{)}^{k},y^{k})+2a_{k-1}(y_{1}^{k-1},v)-2a_{k-1}(y_{1}^{k-1},y^{k})=2(\psi^{k},y^{k}).$$
(18)

In what follows, we use the following estimates:

$$2b(y^{k}-y^{k-1},y^{k})=||y^{k}||_{b}^{2}-||y^{k-1}||_{b}^{2}+||y^{k}-y^{k-1}||_{b}^{2};$$
$$2a_{k-1}(y_{1}^{k-1},y^{k})-2a_{k-1}(y_{2}^{k-1},y^{k})$$
$${}=2a_{k-1}(y_{1}^{k-1},y^{k-1}+y^{k}-y^{k-1})-2a_{k-1}(y_{2}^{k-1},y^{k-1}+y^{k}-y^{k-1})$$
$${}\geqslant 2c_{0}||y^{k-1}||_{1}^{2}-2c_{2}||y^{k-1}||_{1}||y^{k}-y^{k-1}||_{1}\geqslant(2c_{0}-\delta_{0})||y^{k-1}||_{1}^{2}-\frac{c_{2}^{2}}{\delta_{0}}||y^{k}-y^{k-1}||_{1}^{2}$$

with a constant \(\delta_{0}\in(0,2c_{0})\) in the second estimate.

Using the derived estimates and the inequality (14), after summing the equation (18) over \(k\) from 1 to \(m\) we obtain

$$||y^{m}||_{b}^{2}-||y^{0}||_{b}^{2}+\sum\limits_{k=1}^{m}\Big{(}||y^{k}-y^{k-1}||_{b}^{2}-\frac{c_{2}^{2}}{\delta_{0}}||y^{k}-y^{k-1}||_{1}^{2}\Big{)}$$
$${}+(2c_{0}-\delta_{0})\sum\limits_{k=1}^{m}||y^{k-1}||_{1}^{2}\leqslant 2\sum\limits_{k=1}^{m}|(\psi^{k},y^{k})|.$$
(19)

Now, due to the definition of the bilinear form \(b(.,.)\) and the inequalities (11) and (16) we have the inequality

$$||y^{k}-y^{k-1}||_{b}^{2}\geqslant\sigma||y^{k}-y^{k-1}||_{1}^{2}=\left(\frac{c_{2}^{2}}{2c_{0}}+\delta\right)||y^{k}-y^{k-1}||_{1}^{2}.$$

With an appropriate choice of the constant \(\delta_{0}=\delta_{0}(\delta)\in(0,2c_{0}),\) we obtain

$$||y^{k}-y^{k-1}||_{b}^{2}-\frac{c_{2}^{2}}{\delta_{0}}||y^{k}-y^{k-1}||_{1}^{2}\geqslant\left(\frac{c_{2}^{2}}{2c_{0}}+\delta-\frac{c_{2}^{2}}{\delta_{0}(\delta)}\right)||y^{k}-y^{k-1}||_{1}^{2}\geqslant 0,$$

whence

$$||y^{m}||_{b}^{2}-||y^{0}||_{b}^{2}+(2c_{0}-\delta_{0}(\delta))\sum\limits_{k=1}^{m}||y^{k-1}||_{1}^{2}\leqslant 2\sum\limits_{k=1}^{m}|(\psi^{k},y^{k})|.$$

Let \(\mu(\delta)=\min\{\sigma,2c_{0}-\delta_{0}(\delta)\}\). Since \(||y^{m}||_{b}^{2}=(d_{1}-d_{2})||y^{m}||^{2}+\sigma||y^{m}||_{1}^{2}\), then the last inequality implies (17) \(\Box\)

The a priori estimate (17) can be used to obtain various stability estimates for the regularized mesh scheme (13). We give one such estimate in the case of zero initial condition. Let

$$\displaystyle||f||_{-1}=\sup\limits_{u\in V\setminus\{0\}}\frac{(f,u)}{||u||_{1}}=\sup\limits_{u\in V\setminus\{0\}}\frac{\int_{\Omega}fudx}{(\int_{\Omega}|\nabla u|^{2}dx)^{1/2}}$$

and let the following norms be defined for the functions from \(\omega_{\tau}\) to \(V\) or to its subspace \(W\):

$$||y||^{2}_{L^{2}(\omega_{\tau};H^{1})}=\sum\limits_{k=1}^{M}\tau||y^{k}||_{1}^{2},\quad||f||^{2}_{L^{2}(\omega_{\tau};H^{-1})}=\sum\limits_{k=1}^{M}\tau||u^{k}||_{-1}^{2},\quad||q||^{2}_{L^{2}(\omega_{\tau};L^{2}(\Gamma))}=\sum\limits_{k=1}^{M}\tau||u^{k}||_{\Gamma}^{2}.$$

Theorem 1. Let the conditions (7) be satisfied, \(y_{i}\) be the solutions of the equation (13) with the right-hand sides \(f_{i},\) homogeneous initial conditions \(y_{i}^{0}=0\) and the right-hand sides in Neuman boundary conditions \(q_{i}\) . If the regularization parameter satisfies the condition (16):

$$\sigma=\frac{c_{2}^{2}}{2c_{0}}+\delta,\delta>0,$$

then the following stability estimate holds:

$$||y_{1}-y_{2}||_{L^{2}(\omega_{\tau};H^{1})}\leqslant C\Big{(}||f_{1}-f_{2}||_{L^{2}(\omega_{\tau};H^{-1})}+||q_{1}-q_{2}||_{L^{2}(\omega_{\tau};L^{2}(\Gamma))}\Big{)},\quad C=C(\delta).$$
(20)

Proof. The inequality (17) implies

$$\mu\sum\limits_{k=1}^{m}\tau||y_{1}^{k}-y_{2}^{k}||^{2}_{1}\leqslant 2\sum\limits_{k=1}^{m}\tau\big{|}(\psi^{k},y_{1}^{k}-y_{2}^{k})\big{|}$$
(21)

with

$$(\psi^{k},v)=\int\limits_{\Omega}(f_{1}(x,t_{k})-f_{2}(x,t_{k}))v(x)dx+\int\limits_{\Gamma_{N}}(q_{1}(x,t_{k})-q_{2}(x,t_{k})v(x)d\Gamma.$$

For all \(k\) and \(v\in V\)

$$\left|\int\limits_{\Omega}(f_{1}(x,t_{k})-f_{2}(x,t_{k}))v(x)dx\right|\leqslant||f_{1}^{k}-f_{2}^{k}||_{-1}||v||_{1},$$

and due to (6)

$$\left|\int\limits_{\Gamma_{N}}(q_{1}(x,t_{k})-q_{2}(x,t_{k})v(x)d\Gamma\right|\leqslant||q_{1}^{k}-q_{2}^{k}||_{\Gamma}||v||_{\Gamma}\leqslant c_{\Gamma}||q_{1}^{k}-q_{2}^{k}||_{\Gamma}||v||_{1}.$$

Now the estimate (20) immediately follows from (21) and the last two inequalities. \(\Box\)

Using the proven stability estimate, we can obtain an accuracy estimate under the assumption that the input data and the solution to the differential problem are sufficiently smooth. Let \(y^{k},k=1,2,\ldots,M,\) be the solution of a semidiscrete problem (12) and \(u^{k}=u(x,t_{k})\) be the exact solution \(u(x,t)\) of the differential problem (5). Then the estimate (18) applied to the difference \(y-u\) contains in the right-hand side the function \(\psi^{k}\) which is defined by the equality

$$(\psi^{k},v)=(\partial_{t}\bar{u}^{k}-\mathcal{D}_{t}u(t_{k}),v)+\sigma r(u^{k}-u^{k-1},v).$$

Let us consider the case of classical Caputo derivative. The following estimate is well-known ([19, 20]): \(\partial^{\alpha}_{t}\bar{u}^{k}=\mathcal{D}^{\alpha}_{t}u(t_{k})+O(\tau^{2-\alpha})\). It is easy to see that \(r(u^{k}-u^{k-1},v)=\tau r\left(\frac{u^{k}-u^{k-1}}{\tau},v\right)\leqslant O(\tau)||v||_{1}\). Thus, the most significant term in the approximation error is associated with the introduction of the regularizing term, and the accuracy estimate is

$$||y-u||_{L^{2}(\omega_{\tau};H^{1})}=O(\tau).$$

Note that for the implicit scheme (backward Euler scheme) and locally one-dimensional scheme (see [12]), the terms in the accuracy estimate related to the approximation in time have the following orders respectively:

$$||y-u||_{L^{2}(\omega_{\tau};H^{1})}=O(\tau^{2-\alpha}),\quad||y-u||_{L^{2}(\omega_{\tau};L^{2})}=O(\tau^{\alpha/2}).$$

Both of these schemes are nonlinear and require the use of iterative methods when implemented at each time level.

Remark 1. A fully discrete approximation of problem (5) can be constructed using an approximation with respect to the spatial variables of the semidiscrete problem (12). All the results on the existence of a unique solution and the proved stability estimates remain valid for mesh problems constructed using the finite element method with finite elements of the first order, as well as for finite element schemes with quadrature formulas.

Remark 2. The explicit (forward Euler) mesh scheme is a particular case of fully discrete regularized scheme with \(\sigma=0\). A sufficient condition of its stability follows from the estimate (19):

$$(d_{1}-d_{2})||u||^{2}\geqslant\frac{c_{2}^{2}}{\delta_{0}}||u||^{2}_{1}\quad\forall u\in V.$$
(22)

The last inequality is equivalent to the following estimate for the maximum eigenvalue \(\lambda_{\textrm{max}}\) of the mesh Laplace operator:

$$d_{1}-d_{2}\geqslant\lambda_{\textrm{max}}.$$

As is known, \(\lambda_{\textrm{max}}=O(h^{-2})\), where \(h\) characterizes the mesh step in the spatial variable. Since, for example, for the \(L1\)-approximation of the Caputo derivative, \(d_{1}-d_{2}=O(\tau^{-\alpha})\), then (22) leads to a very restrictive condition for the smallness of the time step:

$$\tau<ch^{2/\alpha}$$

with a constant \(c\), which depends on the problem data.

Remark 3. Let \(r_{1}(u,v)\) be a bounded bilinear form in \(V\times V\) and \(r_{1}(u,u)\geqslant 0\). Then the following scheme satisfies all the proven results on unique solvability and stability:

$$(\partial_{t}\bar{y}^{k},v)+\sigma r(y^{k}-y^{k-1},v)+r_{1}(y^{k},v)+a(t_{k-1},y^{k-1},v)=(f^{k},v)\quad\forall v\in V,\quad\sigma>0.$$
(23)

This fact allows to apply these results to the factorized mesh schemes. More precisely, let for a fully discrete scheme constructed using finite element or finite difference scheme \(r(u,v)=(Ru,v)\) and the mesh Laplace operator \(R=-\Delta_{h}\) with corresponding boundary conditions can be splitted into the sum \(R=R_{1}+R_{2}\) such that \(R_{1}R_{2}\) is a non-negative mesh operator: \((R_{1}R_{2}u,u)\geqslant 0\). In this case we set \(r_{1}(u,v)=(R_{1}R_{2}u,v)\) and take the regularization term in the form

$$\sigma r(y^{k}-y^{k-1},v)+\sigma^{2}d_{1}^{-1}r_{1}(y^{k},v).$$

The corresponding mesh scheme becomes the mesh scheme with a factorized operator on each time level:

$$\displaystyle d_{1}\big{(}(I+\sigma d_{1}^{-1}R_{1})(I+\sigma d_{1}^{-1}R_{2})y^{k},v\big{)}-\sigma(Ry^{k-1},v)+\sum\limits_{j=1}^{k-1}(d_{j+1}-d_{j})(y^{k-j},v)$$
$${}+a_{k-1}(y^{k-1},v)=(f^{k},v),\quad I\text{ is the identity operator}.$$
(24)

The implementation of (24) is reduced to sequential inversion of the mesh operators \(I+\sigma d_{1}^{-1}R_{1}\) and \(I+\sigma d_{1}^{-1}R_{2}\) and is very easy if these operators are locally one-dimensional (that corresponds to alternating direction method) or triangle (that corresponds to alternating triangular method). The term \(\sigma^{2}d_{1}^{-1}r_{1}(u^{k},v)\) is added to the approximation error. In case of classical Caputo derivative it is asymptotically \(O(\tau^{\alpha})\). So, the accuracy estimate is

$$||y-u||_{L^{2}(\omega_{\tau};H^{1})}=O(\tau^{\alpha}).$$