 Research
 Open Access
 Published:
An efficient highorder compact finite difference scheme based on proper orthogonal decomposition for the multidimensional parabolic equation
Advances in Difference Equations volume 2019, Article number: 341 (2019)
Abstract
In this paper, the combination of efficient sixthorder compact finite difference scheme (ECFDS6) based proper orthogonal decomposition and Strang splitting method (ECFDS6SSM) is constructed for the numerical solution of the multidimensional parabolic equation (MDPE). For this purpose, we first develop the CFDS6 to attain a high accuracy for the onedimensional parabolic equation (ODPE). Then, by the Strang splitting method, we have converted the MDPE into a series of onedimensional ODPEs successfully, which is easier to implement and program with CFDS6 than an alternating direction implicit method. Finally, we employ proper orthogonal decomposition techniques to improve the computational efficiency of CFDS6 and build the ECFDS6SSM with fewer unknowns and sufficiently high accuracy. Six numerical examples are presented to demonstrate that the ECFDS6SSM not only can largely alleviate the computational load but also hold a high accuracy and simplify the process of the program for the numerical solution of MDPE.
Introduction
Many physical phenomena are simulated with parabolic equations such as proliferation of gas, the penetration of liquids, heat conduction and spread of impurities in semiconductor materials. However, due to the complexity of practical problems or the lack of rules for initial values, their exact solutions for practical engineering problems are not generally sought out so that we have to rely on numerical solutions. Currently, extensive numerical methods including the finite difference method, finite element method, finite volume method and spectral method have been developed for the numerical solution of parabolic equation. In these methods, the FDM has proved to be the most popular and efficient method for finding the numerical solution of a parabolic equation because of its simplicity, wide application in applied fields of sciences and easy programming. Although it is feasible to solve these equations by means of some traditional FDM such as the Euler method and the central difference method, these schemes, which converge very slowly, may largely deviate from the exact solution. Therefore, it is imperative for us to construct a scheme that can guarantee satisfying numerical solution and reflect the properties of equations.
In the existing papers, many scholars focus their eyes on the compact finite difference scheme (CFDS), which has been widely used in numerical solution of various types of partial differential equations. The outstanding advantage of CFDS is that it possesses a faster convergence rate than the corresponding explicit schemes, not significantly increasing the points in each coordinate direction of the grid. As one of the most effective numerical implementations, there have been many numerical research reports concerning with CFDS. For example, Hammad [1] constructed CFDS for Burgers–Huxley and Burgers–Fisher equations. Wang [2] developed CFDS for Poisson equation. Mohebbi [3] combined CFDS with a radial basis functions meshless approach to solve the 2D Rayleigh–Stokes problem. Düring [4] applied CFDS to stochastic volatility models on nonuniform grids. Chen [5] provided highorder CFDS to solve parabolic equation. Especially, some attempts have been made whose main idea is to combine fourth Runge–Kutta in time and a sixthorder compact finite difference in space (CFDS6) by the researchers [6,7,8]. However, the CFDS6 for parabolic equation, especially the case of desirable accuracy in high dimension, they usually need small spatial discretization or extended finite difference stencils and a small time step which brings about heavy computational loads. Therefore, an important problem for CFDS6 is how to build a scheme which not only saves the computational time in the practical problems but also holds a sufficiently accurate numerical solution.
A large number of numerical examples have proved that the proper orthogonal decomposition (POD) is a powerful technique offering the adequate approximation for numerical models with fewer unknowns, which means that the models based on POD can alleviate the computational loads and guarantee sufficiently accurate solution [9,10,11]. The POD is also known as Karhunen–Loève expansions in signal analysis or principal component analysis in statistics, which has been widely used in reallife applications. Especially, it also has been combined with some numerical methods such as finite element methods [12], meshless methods [13], finite difference methods [14] and finite volume methods [15] successfully. However, to the best of our knowledge, there are no published results in papers concerning efficient CFDS6 (ECFDS6) for parabolic equations. Thus, the first task in this paper is to build the ECFDS6 based on the POD method for the parabolic equations.
The MDPE is widely used to describe many physical phenomena, such as engineering problems, typically involving many complex physical phenomena in nature or for irregular computing areas. The study of MDPE has seen a growing interest, which plays an extremely important role in the kinds of fields of physics. Until now, a large amount of work has been done to study FDM with better properties for the numerical solution of MDPE. Among these methods, the alternating direction implicit (ADI) metthod is a popular and wellknown method. The main idea of the ADI [16] method is that the collection of onedimensional problems can be obtained by the discretization of multidimensional problem in space and time, which is feasible since each onedimensional problem in each time level can be solved by tridiagonal matrices. Nevertheless, the accuracy of ADI only is of second order, which results in the accumulation of truncated errors in the process of computation. Especially, ADI is more complex to execute as a computer program and it requires a large amount of computational effort, which may produce considerable dissipation. Thus, a crucial issue is how to build more accurate or convenient scheme to solve the multidimensional problems. The authors of [17] introduced a splitting technique whose main idea is to split the Burgers’ equation into two subequations and solve them by finite difference schemes. In [18], the authors have split the full problem into hyperbolic, nonlinear and linear problems, solved by different numerical methods. Sun [19] had applied a splitting method for solving the radiative transfer problem. Therefore, the Strang splitting method (SSM) is an effective numerical method for solving MDPE by converting MDPE to successions of ODPE. Then we just need to solve the sequence of a linear tridiagonal system, and the whole process of programming is also simplified. SSM has been applied successfully in a diffusion–reaction problem [20]. To the best of our knowledge, no highorder CFDS combined with POD and SSM (ECFDS6SSM) aimed to solve MDPE efficiently has been developed so far. Hence, the second task in this paper is to develop the ECFDS6SSM based on the POD and SSM methods to attain a highly accurate numerical solution of MDPE, which only contains very few unknowns and simplifies the whole process of computation.
The outline of this paper is organized as follows. In Sect. 2, a brief background is given on the theoretical foundations of highorder compact finite difference scheme and POD technique. Then the formulation of the CFDS6 is given for ODPE. Besides, the efficient CFDS6 based on POD for solving ODPE is presented. In Sect. 3, the Strang splitting method is described and the ECFDS6 are extended to MDPE. In Sect. 4, the efficiency, simplicity and capabilities of ECFDS6SSM are verified by six numerical examples. The conclusions are drawn in Sect. 5.
Some highorder difference schemes for ODPE
In this section, we will give a brief description of the CFDS6 and POD techniques, then the construction of ECFDS6 based on POD for solving ODPE is derived.
The construction of sixthorder compact finite difference scheme for ODPE
Firstly, consider the following initial and boundary value problem:
where \(g_{1}(t)\), \(g_{2}(t)\) and \(\varphi(x)\) are given enough smooth functions. Let h be the spatial step increment in the xdirection and τ be the time step increment, and then write \(x_{j} = ( j  1 )h\) \((j = 1,2,3, \ldots,J)\), \(t_{n} = n\tau \) \((n = 0,1,2, \ldots,N1)\), \(u_{j}^{n} \approx u(x_{j},j_{n})\).
The CFDS can be summarized into two broad categories. The main idea of the first methods is to apply the central difference to the governing partial differential equation and then constantly replace the higherorder derivatives in the truncation error with loworder derivatives of the partial differential equation, which is called the traditional explicit finite difference method. The basic idea in the second methods is that all the spatial derivatives in the governing PDEs can be obtained through solving a system of linear equations [21,22,23]. In this paper, we choose the second way to build a highorder compact finite difference scheme for a parabolic equation.
Because the parabolic equation (1) containing the secondorder derivatives, we only give the compact finite difference scheme for secondorder derivatives. Next, we derive the CFDS6 for an ODPE. For the secondorder derivatives at interior nodes \(u_{j}\), the sixthorder scheme formula can be written as follows:
At the most left boundary point \(x_{1}\), a sixthorder formula can be given as follows:
At the second left boundary point \({x_{2}}\), the sixthorder formula is given as follows:
According to the symmetry, at the second right boundary point \({x_{J  1}}\), the sixthorder formula is
Similarly, at the most right boundary point \({x_{J}}\), a sixthorder formula is
Note that the scheme of Eqs. (2)–(6) can be written as
where
As mentioned above, the parabolic equation in (1) has been converted into a system of initial value problem of ordinary differential equations (ODEs) by the compact scheme (2)–(6). Then the fourthorder Runge–Kutta (RK4) scheme is applied to integrate the timedependent governing ODEs,
where R denotes a spatial differential operator. Assuming that the value of \(\mathbf {U}^{n}\) at \({t_{n}}\) is given, then the numerical solution \(\mathbf {U}^{n + 1}\) at \({t_{n + 1}} = {t_{n}} + \tau\) is obtained through the following operations:
Using the sixthorder compact difference scheme listed in Eq. (7), the second derivative related to the operator \(R(\mathbf {U})\) at each time level is obtained. Then we can get the numerical solution at \({t_{n + 1}}\) by the RK4 method. Thus, if the initial value is known, we can calculate the value at any time steps through many iterations.
The establishment of ECFDS6 based on POD technique
In this section, we use the POD technique to build the ECFDS6. For more details, not described here, please refer to [15, 24,25,26]. Meanwhile, the POD method has a variety of interpretations, refer to [9,10,11] to find more interpretations. As described in the Introduction, the main goal of POD is to seek a set of orthogonal matrices generated by applying a singular value decomposition (SVD) into sample space, which is called an optimal basis function. Then, by using the first M sequences of the optimal basis function, the samples can be expressed optimally. In this method, POD will be used to calculate the optimal basis function. To this aim, we need a set of snapshots and use SVD to construct the optimal basis.
We suppose that there are d samples (also usually called snapshots) \({\mathbf{{s}}^{{{1}}}},{\mathbf{{s}}^{{{2}}}}, \ldots , {\mathbf{{s}}^{{{d}}}}\) which can be written as a matrix \(\mathbf{{S}} = ({\mathbf{{s}}^{1}},{\mathbf{{s}}^{2}}, \ldots ,{\mathbf{{s}} ^{d}})\), where \(\mathbf{{s}}^{i} \in{R^{J}}\) \((i = 1,2, \ldots,d)\). \(\mathbf{{S}} \in{R^{J \times d}}\), and \(\mathbf{{S}}{\mathbf{{S}} ^{T}} \in{R^{J \times J}}\) is an \(J \times J\) semidefinite matrix. Applying the SVD on matrix S:
the matrix \(\mathbf {G} = (\boldsymbol{\alpha}_{1},\boldsymbol{\alpha}_{2}, \ldots, \boldsymbol{\alpha}_{J})\) with J rows and J columns. The G and \({\mathbf {V}_{d \times d}}\) are both orthogonal matrices, \(\mathbf {D}_{r} = \operatorname{diag}({\lambda_{1}},{\lambda_{2}}, \ldots,{\lambda_{r}})\). The orthogonal eigenvectors of \(\mathbf {S}{\mathbf {S} ^{T}}\) are contained in the matrix \(\mathbf {G} = (\boldsymbol{\alpha}_{1}, \boldsymbol{\alpha}_{2}, \ldots,\boldsymbol{\alpha}_{J})\). The singular values \({\lambda_{i}}\) \((i = 1,2, \ldots,r)\) satisfy \({\lambda_{1}} \ge {\lambda_{2}} \ge\cdots\ge{\lambda_{r}} > 0\).
Denote d columns of S by \({\boldsymbol{\beta}^{l}} = {(s_{1}^{l},s _{2}^{l}, \ldots,s_{J}^{l})^{T}}\) \((l = 1,2, \ldots,d)\), the projection \({P_{k}}\) is defined as follows:
the \(0 < M \le d\) and \(( \cdot, \cdot)\) represents the inner product of vectors, then we can obtain the following result [24]:
where \(\Vert \cdot\Vert_{{2}}\) is standard norm of vector. Hence, \({\boldsymbol{\alpha}_{1}},{\boldsymbol{\alpha}_{2}}, \ldots,\boldsymbol{\alpha}_{M}\) are a group of the optimal POD basis, which from basis matrix \(\boldsymbol{\alpha} = (\boldsymbol{\alpha}_{1},\boldsymbol{\alpha}_{2}, \ldots, \boldsymbol{\alpha}_{M})\). It should be pointed out that the basis matrix fulfills the orthogonality condition, i.e., \({{\boldsymbol{\alpha}}^{T}} {\boldsymbol{\alpha}} = \mathbf {I}\) (I is unit matrix with M dimension).
In the following, the procedure of establishing ECFDS6 for parabolic equation is listed by the POD basis.
If U of Eq. (7) is substituted for
we have
and noting that \({{\boldsymbol{\alpha}}^{T}}{\boldsymbol{\alpha}} = \mathbf {I}\), let \({\mathbf {V}_{\mathbf{{0}}}}= {\boldsymbol{\alpha} ^{T}}{\mathbf{{U}}^{n}}\) then the RK4 for the reduced solution is given as follows:
We can obtain the global solution \({\mathbf {U}^{n + 1}} = {\boldsymbol{\alpha}} {\mathbf {V}^{n + 1}}\) when the reduced solution \({\mathbf {V}^{n + 1}}\) has been obtained from Eq. (15). Here, the procedure of ECFDS6 for parabolic equation is listed as follows:

step 1
Snapshot S is generated from experiments or numerical simulations.

step 2
Formulate the optimal POD basis matrix α by the SVD method.

step 3
Apply Eq. (14) to work out the reduced secondorder derivative \(\mathbf {V}''\).

step 4
Obtain the reduced solution by solving Eq. (15).

step 5
Having applied \({\mathbf {U}^{n + 1}} = {\boldsymbol{\alpha}}{ \mathbf {V}^{n + 1}}\), then the reduced solution is expanded.
It is easy to see from that above algorithm that ECFDS6 needs to solve only \(M \times M\) equations (Eq. (14)) at each iteration, but CFDS6 includes \(J \times J\) equations (Eq. (7)) to solve at each iteration. In general, M is much smaller than J, which means that ECFDS6 needs less computational time than CFDS6. Applying that whole procedure, we may complete the entire calculation from \({t^{n}}\) to \({t^{n + 1}}\). Moreover, due to the use of a sixthorder compact scheme for discretizing the space variables, it is not difficult to find that our algorithm is of sixthorder accuracy.
Multidimensional case
The traditional numerical method for solving MDPE is the ADI method, which replaces complex multidimensional problems with a number of onedimensional problems. It is a classical algorithm. However, with the use of the Crank–Nicolson method in time and the center difference method in space, ADI is shown to only have secondorder accuracy in both time and space. Besides, this method produces a very complex set of equations in MDPE, which is very expensive to solve. The benefits of the ADI method are that the equation needed to be solved in each step is relatively simple, and the tridiagonal matrix algorithm can be used to solve the equation successfully. The ADI method is unconditionally stable. Compared with ADI, SSM is an effective numerical method that will lead to no loss of accuracy. SSM is extremely efficient for solving MDPE by converting MDPE to a product of ODPE and programming is very simple. It also can be use to accelerate that calculation of problems related to operators of different time scales. In this section, instead of using ADI, we applied ECFDS6SSM to solve MDPE. First, we apply SSM to decompose MDPE into a product of ODPEs. Then we solve each ODPE by ECFDS6.
Solutions to twodimensional parabolic equation
Considering the following twodimensional parabolic equation:
We rewrite Eq. (16) as follows:
and
where \({\mathbf {H}_{x}}\) and \({\mathbf {H}_{y}}\) are different operators solved by Eq. (7) in the xdirection and the ydirection. The initial value of 2D equation \({\mathbf {U}^{n }}\) at \({t^{n }}\) is split, respectively, into \(\mathbf {U}_{1}^{n }\) and \(\mathbf {U}_{2}^{n }\) in the xdirection and the ydirection by Eq. (17). Then we use ECFDS6 to compute \(\mathbf {U}_{1}^{n + 1}\) and \(\mathbf {U}_{2}^{n + 1}\) by Eq. (15). Finally, we obtain the following equations:
where \({\mathbf {U}^{n + 1}}\) represents the numerical result of 2D equation at \({t^{n+1}}\). Thus, we split the twodimensional problem into two onedimensional problems, which indicates that we only solve each ODPE by ECFDS6 introduced in Sect. 2, rather than solve a set of complicated equations.
Solutions to threedimensional parabolic equation
For the threedimensional parabolic equation, the above ECFDS6SSM scheme can be extended directly to the threedimensional case. Similarly, consider the following threedimensional parabolic equation:
We also rewrite Eq. (20) as follows:
and
where \({\mathbf {H}_{x}}\), \({\mathbf {H}_{y}}\), \({\mathbf {H}_{z}}\) are different operators obtained by Eq. (7) in the xdirection, ydirection, and zdirection. It is necessary to note that the way of acquiring \({\mathbf {U}^{n + 1}}\) is similar to Eq. (19), which will not be listed again.
The \(\mathbf {U}_{1}^{n + 1}\), \(\mathbf {U}_{2}^{n + 1}\) and \(\mathbf {U}_{3}^{n + 1}\) represent the numerical results of each ODPE obtained by Eq. (15) in the xdirection, ydirection and zdirection. Compared with ADI, it is so simple that we do not solve a very complicated set of equations in MDPE in the way mentioned for ADI, and we just need to compute three ODPEs, respectively. It should be pointed out that the scheme has the same accuracy as the onedimensional cases.
Formulate POD basis for each ODPE split by MDPE
In this subsection, we will illustrate our methods and formulate the POD basis for each ODPE. As described above, we have split 2D or 3D parabolic equations into a series of ODPE. By solving the formulation of Eq. (15), we can get the approximate solutions \(\{ {u_{j}^{n}} \}_{n = 1}^{N}\) \(( j = 2, 3, \ldots,J  1 )\). Then we may select \(\{ {u_{j}^{{n_{i}}}} \}_{i = 1}^{d}\) \(( j = 2, 3, \ldots,J  1, 1 \le{n_{1}} \le {n_{2}} < \cdots< n_{{{d}}} \le N )\) from the CFDS6 solutions \(\{ {u_{j}^{n}} \}_{n = 1}^{N}\) \(( j = 2, 3, \ldots,J  1 )\) of Eq. (15), which is called the method of ‘snapshots’. Finally, we can construct the POD basis for each ODPE in the way mentioned in Eqs. (10)–(13).
Although the snapshot is obtained through the approximate solution of CFDS6 in this paper, in fact, we can get the collection of snapshots through experiments and interpolation when calculating the actual problem. If the development and change of a large number of future natural phenomena (for example, weather change, biology anagenesis, and so on) are closely related to previous results, or if the physical system of the natural phenomena performs well, that is, the past dynamics is representative and inclusive of the future dynamics, then the previous or existing experimental data can be used to construct a snapshot. Then the POD basis is obtained by using the POD method in Eqs. (10)–(13) and we can derive an efficient scheme. Therefore, one can effectively simulate and predict the development and change of some future natural phenomena, which is of great significance for practical applications.
Numerical examples
In order to see whether the present method is capable of getting an accurate solution, in this section, the ECFDS6SSM will be evaluated for six examples of the MDPE given below. In the case of the different number of nodes, we have some tests of the accuracy and efficiency for the method described in this article. We performed our computations using Matlab 2018a software on a Ryzen 7 1800X, 3.6 GHz CPU machine with 16 GB of memory. The convergence order of the method presented in this article is calculated with this formula:
Example 1
Consider the following 2D parabolic equation (\(\mathrm{SP} _{1}\)):
where \(\varOmega= \{ (x,y);0 \le x \le2,0 \le y \le2\}\), ∂Ω denotes the boundary of Ω. The exact solution is \(u(x,y,t) = {10^{9}}{e^{  2{\pi^{2}}t}}\sin\pi x\sin\pi y\).
The obtained solutions and pointwise error of CFDS6SSM with \(41\times 41\) uniformly distributed points are shown in Fig. 1 while the corresponding numerical results of ECFDS6SSM are shown in Fig. 2. We choose h as 0.05 and \(\tau=0.0001\). It is not difficult to see that the results of ECFDS6SSM are in very good agreement with those of CFDS6SSM, which means that the ECFDS6SSM method possesses a high computational accuracy, as well as CFDS6SSM. We compare the error of ECFDS6SSM and CFDS6SSM with the error of the D’Yakonov alternating direction implicit (DADI) method with \({h_{x}} = {h_{y}} = 0.01\) and \(\tau= 0.01\) in Table 1, which indicates that our algorithm significantly improves the accuracy. In order to compare the accuracy and efficiency of two approximate methods, we list the error and computational time of ECFDS6SSM and CFDS6SSM in Table 1 and Table 2. The error of CFDS6SSM and the error of ECFDS6SSM are also drawn in Figs. 1 and 2. It also can be clearly seen that the ECFDS6SSM is almost as accurate as CFDS6SSM and the computational times of ECFDS6SSM are less than those of CFDS6SSM under the same number of nodes. In Fig. 3, the error between CFDS6SSM and ECFDS6SSM is no more than \(3 \times 10^{  7}\). It also can be found that the accuracy of ECFDS6SSM is almost identical to that of CFDS6SSM. Besides, we also found that the order of convergence obtained by ECFDS6SSM and CFDS6SSM is almost the same in Table 2. In addition, it should be noted that our algorithm is easier to execute than the classical ADI algorithm.
Example 2
Consider the following 2D parabolic equation (\(\mathrm{SP} _{2}\)):
where \(\varOmega= \{ (x,y);0 \le x \le2\pi,0 \le y \le2\pi\}\), ∂Ω denotes the boundary of Ω. The exact solution is \(u(x,y,t) = {e^{  t}}\sin x\sin\frac{y}{2}\).
We design this equation to let the solution u change much faster in the xdirection than in ydirection. We give the figures of CFDS6SSM and ECFDS6SSM in Figs. 4 and 5. Obviously, they are very similar for the same nodes. Meanwhile, we also plot the error of ECFDS6SSM and CFDS6SSM in the righthand side of Figs. 4 and 5 for comparison. The error of CFDS6SSM, the error of ECFDS6SSM and the computational time are also shown in Table 3 and Table 4. It is not difficult to see that the results of ECFDS6SSM are in very good agreement with those of CFDS6SSM, which means that the ECFDS6SSM not only possesses a high accuracy as well as CFDS6SSM, but also alleviates computational load. In addition, we compare the error of ECFDS6SSM and CFDS6SSM with the error of DADI with \({h_{x}} = {h_{y}} = 0.01\pi\) and \(\tau= 0.01\) in Table 3, it also manifested that our algorithm significantly improves the accuracy. In Fig. 6, the error between CFDS6SSM and ECFDS6SSM is no more than \(1.5 \times10^{  9}\). It also indicated that the accuracy of ECFDS6SSM is almost identical to that of CFDS6SSM under the same nodes and time step. Besides, we also give the order of convergence obtained by ECFDS6SSM and CFDS6SSM in Table 4, by which it can be seen clearly that the numerical results confirm the convergence with the rate \(O({h^{6}})\) for this equation.
Example 3
Consider the following 2D parabolic equation (\(\mathrm{SP} _{3}\)):
where \(\varOmega= \{ (x,y);0 \le x \le\pi,0 \le y \le\pi\}\), ∂Ω denotes the boundary of Ω. The exact solution is \(u(x,y,t) = {e^{  t}}\sin x\sin y\).
Figure 7 and Fig. 8 plot the numerical solutions and pointwise absolute errors of CFDS6SSM and ECFDS6SSM, respectively. For comparison, Table 5 lists the error of ECFDS6SSM and DADI with \({h_{x}} = {h_{y}} = 0.005\pi\) and \(\tau= 0.01\), which indicates that our methods remarkably improve the accuracy. Similarly, we also list the absolute error, computational time and convergence order in Table 6. It can be clearly seen that the ECFDS6SSM greatly saved the time consumption and vastly improved the computational efficiency compared with CFDS6SSM. Besides, the convergence order in Table 6 confirms the convergence with the rate \(O({h^{6}})\) for this equation. The error between CFDS6SSM and ECFDS6SSM is also drawn in Fig. 9. It also can be clearly seen that the CFDS6SSM is almost accurate with the ECFDS6SSM.
Example 4
Consider the following 3D parabolic equation (\(\mathrm{SP} _{4}\)):
where \(\varOmega= \{ (x,y,z);0 \le x \le2\pi,0 \le y \le2\pi,0 \le z \le2\pi\}\), ∂Ω denotes the boundary of Ω. The exact solution is \(u(x,y,t) = {e^{  t}}\sin x\sin y \sin z\).
In this example, we also let the solution u change differently in each direction. Compared with the algorithm mentioned in [14], our algorithm is easier to carry out and understand. In Fig. 10, the numerical solutions of CFDS6SSM and ECFDS6SSM are shown, which indicates that they are similar. Next, we let \(z=0.4\pi\), the figure becomes the threedimensional figure in Fig. 11. we report the absolute error and computational time of ECFDS6SSM and CFDS6SSM in Table 7. It is easy to see that the results of ECFDS6SSM are in very good agreement with those of CFDS6SSM, which means that the ECFDS6SSM holds accuracy. It also shows that the ECFDS6SSM can greatly save the time consumption and vastly improve the computational time. Then we plot pointwise the error of CFDS6SSM and ECFDS6SSM in Figs. 12 and 13. It can be seen that the ECFDS6SSM is similar to CFDS6SSM. Besides, in Table 7, in order to evaluate the overall convergence rate of the ECFDS6SSM, we have reported the order of convergence obtained by ECFDS6SSM and CFDS6SSM, which shows that the ECFDS6SSM still can produce very accurate solution.
Example 5
Consider the following 3D parabolic equation (\(\mathrm{SP} _{5}\)):
where \(\varOmega= \{ (x,y,z);  1 \le x \le1,  1 \le y \le1,  1 \le z \le1\}\), ∂Ω denotes the boundary of Ω. The exact solution is \(u(x,y,z,t) = {e^{  t}}\sin\pi x\sin\pi y \sin\pi z\).
In this example, we observe the threedimensional parabolic problem by the slice figures of the fourdimensional images in Fig. 14. In Fig. 14, the solutions of CFDS6SSM and ECFDS6SSM are almost similar. In order to make the image more aestheticcally pleasing, we let \(z=0.4\), the figure becomes the threedimensional figure in Fig. 15. We also list the absolute error of ECFDS6SSM and CFDS6SSM in Table 8. It is easy to see that the results of ECFDS6SSM are in very good agreement with those of CFDS6SSM, which means that the ECFDS6SSM without missing significant loss in accuracy. Then the corresponding error of CFDS6SSM and ECFDS6SSM are drawn in Fig. 16. It can be seen that the CFDS6SSM is slightly less accurate than ECFDS6SSM. In Fig. 17, the error between ECFDS6SSM and CFDS6SSM is given. Besides, in Table 8, we also gave the computational time and order of convergence obtained by ECFDS6SSM and CFDS6SSM, which indicates that the ECFDS6SSM is more efficient than the CFDS6SSM for solving the parabolic equation and ECFDS6SSM holds same accuracy.
Example 6
Consider the following 3D parabolic equation (\(\mathrm{SP} _{6}\)):
where \(\varOmega= \{ (x,y,z);0 \le x \le2\pi,0 \le y \le2\pi,0 \le z \le2\pi\}\), ∂Ω denotes the boundary of Ω. The exact solution is \(u(x,y,t) = {e^{  t}}\sin x\sin y \sin z\).
In this example, in Fig. 18, the numerical solutions of CFDS6SSM and ECFDS6SSM are almost identical. Similarly, we let \(z=0.4\pi\), the figure becomes the threedimensional figure in Fig. 19. We list the absolute error and computational time of ECFDS6SSM and CFDS6SSM in Table 9. It is easy to see that the results of ECFDS6SSM are in very good agreement with those of CFDS6SSM, which means that the ECFDS6SSM does not have a significant loss in accuracy. It also shows that the ECFDS6SSM can greatly save the time consumption and vastly improve the computational time. Then we plot the numerical error and pointwise absolute errors of CFDS6SSM and ECFDS6SSM in Figs. 20 and 21. It can be seen that the ECFDS6SSM is similar to CFDS6SSM. Besides, in Table 9, we also give the order of convergence obtained by ECFDS6SSM and CFDS6SSM, which indicates that the ECFDS6SSM is not only more efficient than the CFDS6SSM for solving the parabolic equation, but it holds the same accuracy.
Conclusions
In this article, we have established the ECFDS6SSM for the numerical solution of MDPE. For this purpose, firstly, the snapshot is formed by means of the initial few solutions such as the numerical simulation results or the collection of experiment data and the POD basis is formulated by the SVD. Secondly, we have established ECFDS6 based on the POD technique, which can be used to solve ODPE efficiently. Finally, we split MDPE into a product of ODPEs. Coupled with SSM, we have obtained the ECFDS6SSM for MDPE. As listed in the table of Sect. 4, we have compared our method with ADI, which shows that the ECFDS6SSM improves the accuracy greatly. In addition, the computational time of ECFDS6SSM is greatly less than CFDS6SSM and the whole process of implementation of ECFDS6SSM is simpler than CFDS6SSM. We test our algorithm by six numerical experiments, which implies that the ECFDS6SSM has high efficiency and is reliable for solving the MDPE.
References
Hammad, D.A., ElAzab, M.S.: 2N order compact finite difference scheme with collocation method for solving the generalized Burgers–Huxley and Burgers–Fisher equations. Appl. Math. Comput. 258, 296–311 (2015)
Wang, H., Zhang, Y., Ma, X., Qiu, J., Liang, Y.: An efficient implementation of fourthorder compact finite difference scheme for Poisson equation with Dirichlet boundary conditions. Comput. Math. Appl. 71, 1843–1860 (2016)
Mohebbi, A., Abbaszadeh, M., Dehghan, M.: Compact finite difference scheme and RBF meshless approach for solving 2D Rayleigh–Stokes problem for a heated generalized second grade fluid with fractional derivatives. Comput. Methods Appl. Mech. Eng. 264, 163–177 (2013)
Düring, B., Fournié, M., Heuer, C.: Highorder compact finite difference schemes for option pricing in stochastic volatility models on nonuniform grids. J. Comput. Appl. Math. 271, 247–266 (2014)
Chen, J., Ge, Y.: High order locally onedimensional methods for solving twodimensional parabolic equations. Adv. Differ. Equ. 2018, 361 (2018)
Bhatt, H.P., Khaliq, A.Q.M.: Fourthorder compact schemes for the numerical simulation of coupled Burgers’ equation. Comput. Phys. Commun. 200, 117–138 (2016)
Sari, M., Gürarslan, G.: A sixthorder compact finite difference scheme to the numerical solutions of Burgers’ equation. Appl. Math. Comput. 208, 475–483 (2009)
Zhang, X., Zhang, P.: A reduced highorder compact finite difference scheme based on proper orthogonal decomposition technique for KdV equation. Appl. Math. Comput. 339, 535–545 (2018)
Liang, Y.C., Lee, H.P., Lim, S.P., Lin, W.Z., Lee, K.H., Wu, C.G.: Proper orthogonal decomposition and its applications—Part I: theory. J. Sound Vib. 252, 527–544 (2002)
Rathinam, M., Petzold, L.R.: A new look at proper orthogonal decomposition. SIAM J. Numer. Anal. 41, 1893–1925 (2003)
Kerschen, G., Golinval, J.C., Vakakis, A.F., Bergman, L.A.: The method of proper orthogonal decomposition for dynamical characterization and order reduction of mechanical systems: an overview. Nonlinear Dyn. 41, 147–169 (2005)
Ullmann, S., Rotkvic, M., Lang, J.: PODGalerkin reducedorder modeling with adaptive finite element snapshots. J. Comput. Phys. 325, 244–258 (2016)
Dehghan, M., Abbaszadeh, M.: A combination of proper orthogonal decomposition–discrete empirical interpolation method (PODDEIM) and meshless local RBFDQ approach for prevention of groundwater contamination. Comput. Math. Appl. 75, 1390–1412 (2018)
An, J., Luo, Z., Li, H., Sun, P.: Reducedorder extrapolation spectralfinite difference scheme based on POD method and error estimation for threedimensional parabolic equation. Front. Math. China 10, 1025–1040 (2015)
Luo, Z., Li, H., Sun, P.: A reducedorder Crank–Nicolson finite volume element formulation based on POD method for parabolic equations. Appl. Math. Comput. 219, 5887–5900 (2013)
Peaceman, D.W., Rachford, H.H. Jr.: The numerical solution of parabolic and elliptic differential equations. J. Soc. Ind. Appl. Math. 3, 28–41 (1955)
Seydaoğlu, M.: An accurate approximation algorithm for Burgers’ equation in the presence of small viscosity. J. Comput. Appl. Math. 344, 473–481 (2018)
Gidey, H.H., Reddy, B.D.: Operatorsplitting methods for the 2D convective Cahn–Hilliard equation. Comput. Math. Appl. 77, 3128–3153 (2019)
Sun, J., Eichholz, J.A.: Splitting methods for differential approximations of the radiative transfer equation. Appl. Math. Comput. 322, 140–150 (2018)
Einkemmer, L., Moccaldi, M., Ostermann, A.: Efficient boundary corrected Strang splitting. Appl. Math. Comput. 332, 76–89 (2018)
Li, J., Chen, Y., Liu, G.: Highorder compact ADI methods for parabolic equations. Comput. Math. Appl. 52, 1343–1356 (2006)
Li, J., Visbal, M.R.: Highorder compact schemes for nonlinear dispersive waves. J. Sci. Comput. 26, 1–23 (2006)
Li, J.: Highorder finite difference schemes for differential equations containing higher derivatives. Appl. Math. Comput. 171, 1157–1176 (2005)
Sun, P., Luo, Z., Zhou, Y.: Some reduced finite difference schemes based on a proper orthogonal decomposition technique for parabolic equations. Appl. Numer. Math. 60, 154–164 (2010)
Luo, Z., Yang, X., Zhou, Y.: A reduced finite difference scheme based on singular value decomposition and proper orthogonal decomposition for Burgers equation. J. Comput. Appl. Math. 229, 97–107 (2009)
Luo, Z., Li, H., Sun, P., Gao, J.: A reducedorder finite difference extrapolation algorithm based on POD technique for the nonstationary Navier–Stokes equations. Appl. Math. Model. 37, 5464–5473 (2013)
Acknowledgements
The authors express their sincere thanks to the anonymous reviews for their valuable suggestions and corrections for improving the quality of this paper.
Availability of data and materials
Not applicable.
Funding
This work is financially supported by the Academic Mainstay Foundation of Hubei Province of China (No. D20171202), and the National Natural Science Foundation of China (Grant No. 11826208).
Author information
Authors and Affiliations
Contributions
All authors contributed equally and significantly in writing this article. All authors read and approved the final manuscript.
Corresponding author
Ethics declarations
Competing interests
The authors declare that they have no competing interests.
Additional information
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendix: The detailed implementation of ADI for 2D parabolic equation
Appendix: The detailed implementation of ADI for 2D parabolic equation
For the 2D problem
We let \({x_{i}} = i\cdot\Delta x\) \((0 \le i \le a/\Delta x), {y_{j}} = j\cdot\Delta y\) \((0 \le j \le b/\Delta y)\) and \({t_{k}} = k\cdot\Delta t\) \((0 \le k \le T/\Delta t)\).
Firstly, define the following central difference operators:
We can rewrite Eq. (23) by these difference operators
Then we have the formula
where \({r_{1}} = \frac{{\Delta t}}{{\Delta{x^{2}}}}\) and \({r_{2}} = \frac{{\Delta t}}{{\Delta{y^{2}}}}\).
In order to obtain a collection of onedimensional problems that can be solved by the tridiagonal matrix, we add \(\frac {{{r_{1}}{r_{2}}}}{4}\delta_{x}^{2}\delta_{y}^{2}u_{i,j}^{k + 1}\) in the lefthand side of Eq. (25) and \(\frac {{{r_{1}}{r_{2}}}}{4}\delta_{x}^{2}\delta_{y}^{2}u_{i,j}^{k}\) in the righthand side of Eq. (25). Then Eq. (25) can be written as
Then Eq. (26) also can be rewritten as
By introducing the variable \({V_{i,j}}\), the wellknown Peaceman–Rachford scheme can be obtained as follows:
It should be noticed that the boundary condition should be given in the first scheme of Eq. (28).
We assume the \(u_{i,j}^{k}\) is given. In the first scheme of Eq. (28), we first fix j \((1 \le j \le n1)\), then the system consisting of \(m1\) equations with \(m1\) unknowns can be solved to get the V. Similarly, we first fix i \((1 \le i \le m1)\), then, by the V, the system consisting of \(n1\) equations with \(n1\) unknowns can be solved to get the \(u_{i,j}^{k+1}\).
Rights and permissions
Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.
About this article
Cite this article
Xu, B., Zhang, X. An efficient highorder compact finite difference scheme based on proper orthogonal decomposition for the multidimensional parabolic equation. Adv Differ Equ 2019, 341 (2019). https://doi.org/10.1186/s1366201922733
Received:
Accepted:
Published:
DOI: https://doi.org/10.1186/s1366201922733
Keywords
 Compact finite difference scheme
 Proper orthogonal decomposition
 Multidimensional parabolic equation
 Strang splitting method