WorksMathematics
Unconditionally Positive Nonstandard Finite Difference Schemes for -Laplacian Reaction–Diffusion Equations
Abstract
We propose a conservative, semi-implicit nonstandard finite difference scheme for the reaction–diffusion equation , , where, as is customary, the -Laplacian is regularized by a parameter \eps>0. The scheme combines a denominator function, a lagged diffusion coefficient, and a nonlocal treatment of the loss part of the reaction. For every time step and mesh size we prove that the scheme is well posed and positive, that it preserves an invariant region whenever the reaction admits one, and that it reproduces every constant equilibrium exactly. Without reaction, it conserves mass, satisfies a discrete maximum principle, and, for , decreases a discrete energy. An example shows that the energy may increase for . The truncation error is at interior nodes. Numerical tests confirm first order in time and second order in space, and show that the scheme remains positive and bounded for step sizes at which the explicit method breaks down.
Introduction
Reaction–diffusion equations with -Laplacian-type diffusion arise in models of non-Newtonian flow, filtration, and nonlinear heat conduction [1, 2, 3, 4]. Since the unknown is usually a density or a concentration, a numerical method should return nonnegative values and, if the reaction has a carrying capacity, respect it. Explicit finite differences cannot guarantee this. When the explicit scheme is written in conservative form, the coefficient of in the update is , where the conductance is evaluated at the old time level. Positivity can therefore be guaranteed only under a step size restriction that depends on the gradients, which are not known in advance, and this restriction becomes severe as grows (§ 4).
Nonstandard finite difference (NSFD) methods address such failures with two devices: a denominator function that replaces in the difference quotient, and nonlocal
Email addresses: [email protected] (Achraf Zinihi), [email protected] (Matthias Ehrhardt), [email protected] (Moulay Rchid Sidi Ammi) (two-level) approximations of nonlinear terms [5, 6, 7, 8]. Evaluating the diffusion coefficient at the old level (“lagged diffusivity”) is a classical device in image processing and in the iterative solution of -Poisson problems [9, 10]. Here we combine these ingredients: lagging the conductance and treating the loss part of the reaction nonlocally yields a scheme whose system matrix is a tridiagonal -matrix. Unconditional positivity, an invariant region, exact equilibria, and an energy inequality then follow from short arguments, for every and for a whole class of reactions. We also show where the argument stops: for the discrete energy can increase.
We do not claim that the individual devices are new. In the previous work [7] we proposed an explicit NSFD scheme for the same class of problems, in which only the central node of each discrete flux is taken at the new time level. That scheme is positive for every step size, but its spatial operator is not in conservative form, it requires , the conductance is not regularized for , and the bound proved there grows with the number of time steps. The present scheme evaluates the conductance at cell midpoints in conservative form and lags it in time, so that the entire new time level enters through a tridiagonal -matrix. At the cost of one tridiagonal solve per step, this yields properties that the explicit scheme does not provide: invariance of uniformly in , and , mass conservation and a discrete maximum principle, the exact logistic flow in the reaction part (Remark 3.3), and a discrete energy inequality for together with an example showing that it fails for . We also give numerical evidence, including a comparison with iterating the lagged conductance to convergence (Table 4).
| Picard solves per step | at | ||
| 1.5 | 5.4 | ||
| 1.5 | 10.1 | ||
| 1.5 | 17.2 | ||
| 3 | 4.6 | ||
| 3 | |||
| 3 |
Table 4. Scheme (3) (solution ) versus the converged Picard iteration (solution ) within each step, stopped when or after 200 iterations; data as in Table 3. Entries marked are lower bounds, since the cap was reached in some steps.
The Scheme
For and we consider a reaction–diffusion equation of the form
with zero-flux boundary conditions at , initial data , , , and a fixed regularization parameter , as in finite element treatments of the -Laplacian [3]. For , (1) reduces to the semilinear heat equation for every . The reaction is split as
For the logistic term one takes ; for with one may take and , extended continuously to .
Let , , , and let and for . For a grid function we define the conductances and the discrete diffusion operator
with the zero-flux convention . With the denominator function , , which satisfies [6], the NSFD scheme reads
The conductance is evaluated at the old level, the solution enters linearly at the new level, and the loss term is replaced by . Writing and , we can rewrite (3) as with the tridiagonal matrix
where . Hence, one step costs operations with the Thomas algorithm [11] the same order as an explicit step.
Properties of the Scheme
We now collect the main properties of scheme (3) that hold for every time step and mesh size.
Theorem 3.1. Let and be arbitrary, let the splitting of the reaction (2) hold, and let be a grid function.
(a) is a strictly diagonally dominant -matrix with . Hence (3) has a unique solution .
(b) , and if .
(c) If on for some and , then for all .
(d) If for some and , then .
(e) If , then is independent of , and .
Proof. (a) has a positive diagonal, non-positive off-diagonal entries, and . For any vector , choose with . Then , so . Hence is nonsingular and . Writing with diagonal and , diagonal dominance gives , so entry-wise, with diagonal entries at least .
(b) Since and , the right-hand side satisfies . As , we get . If , then .
(c) We argue by induction. To do so, assume . By (b), . Let and . In row of (3) the diffusion part is nonnegative, so , where we used and on . Hence .
(d) Constants lie in the kernel of the diffusion part, and . Hence , i.e., solves (3).
(e) With , multiplying (3) by and summing over , the diffusion terms telescope to zero by the zero-flux convention. At a maximizer of , row gives . The minimum is treated in the same way.
Remark 3.2. The proof of Theorem 3.1 uses only the positivity of the conductances , not the way they are computed. Parts (a)–(e) therefore hold for any positive conductance in place of , in particular for every iterate of a Picard iteration that updates the conductance within a time step. Theorem 3.4 below, in contrast, does use .
Remark 3.3. For the logistic reaction with and , the reaction part of (3) is the map , which is the exact flow of . Part (c) holds with .
For the energy estimate, let , so that , and define
Theorem 3.4. Let and . Then, for every ,
Proof. Write , , , and . Summation by parts with the zero-flux convention gives . Multiplying (3) by and summing yields
Let , so that . For the function is concave, hence with . Applying this with and gives . Since , we obtain , and the claim follows from (4) after dropping the last (non-positive) term.
Remark 3.5. For the function is convex, and not only the argument but also the conclusion fails. Take , , , , , and . In the third step, rises by about 22%, from about to (Figure 1, right).

Figure 1. Discrete energy for , , , , and . Left: (monotone for all ). Right: (an increase occurs for ).
Proposition 3.6. Let and be Lipschitz continuous, and let be a solution of (1) that is in and in . Inserting into (3) leaves a residual at interior nodes and for , with constants depending on , , , and norms of .
Proof. Since , the time quotient equals . Because for and is bounded on bounded gradients for , replacing by in the diffusion and reaction terms costs . For the spatial part, and is smooth, so with . The difference quotient of this flux is at interior nodes. At the half cell gives , using . The case is analogous.
Numerical Experiments
Unless stated otherwise, all runs use , , and the logistic reaction with . We use the manufactured solution and add to (1) the source term that makes it exact for the regularized operator, is included in and in . Table 1 and Table 2 show first order in time and second order in space for and . The last spatial rates fall below two because the first-order time error is still visible: for and , halving changes the error from to .
Table 1: Time refinement with : maximum-norm errors and observed orders (, ).
| Error | Order | Error | Order | |||
| – | – | |||||
| 0.65 | 0.94 | |||||
| 0.80 | 0.94 | |||||
| 0.90 | 0.97 | |||||
| 0.95 | 0.98 | |||||
| 0.97 | 0.99 | |||||
| 0.98 | 1.00 |
Table 1.
Table 2: Space refinement with : maximum-norm errors and observed orders (, ).
| Error | Order | Error | Order | |||
| 10 | – | – | ||||
| 20 | 3.63 | 2.73 | ||||
| 40 | 3.47 | 1.94 | ||||
| 80 | 2.04 | 1.98 | ||||
| 160 | 1.92 | 1.93 | ||||
| 320 | 1.72 | 1.76 |
Table 2.
Table 3 compares scheme (3) with the explicit method based on the same conservative spatial discretization. The initial datum is a narrow bump on a nearly vanishing background, and . For the explicit method blows up even for , since the conductance grows with the gradient. For it produces negative values for every we tried. For the violation is small ( at intermediate times, and the solution is positive again at ), but it becomes large for larger . Scheme (3) stays in in all cases, as predicted by Theorem 3.1. Note that for large we claim only positivity and boundedness, the accuracy is still .
| Explicit FDM | Scheme (3): at | ||
| 1.5 | negative values () | ||
| 1.5 | negative values () | ||
| 1.5 | negative values () | ||
| 3 | blow-up at | ||
| 3 | blow-up at | ||
| 3 | blow-up at |
Table 3. Explicit FDM versus scheme (3) for , , , ; “blow-up” means .
The lag can be removed by a Picard iteration within each step: starting from , the iterate solves (3) with replaced by , until convergence. The reaction is treated as in (3). By Remark 3.2 every iterate is positive, so positivity does not distinguish the two approaches – the difference lies in the cost. Table 4 reports the mean number of tridiagonal solves per step needed to reach for the data of Table 3. For it ranges from about 5 to 17: for it grows quickly with , and the cap of 200 iterations was reached in some steps. The solution of (3) differs from the converged one by an amount of order , as expected from lagging the conductance. In this sense, (3) is the one-iteration member of this family.
Table 5 isolates the effect of the reaction term. We keep the denominator function and the lagged implicit diffusion of (3) but treat the reaction explicitly. For data above and a small diffusion coefficient, this variant produces negative values for () and reaches for . The nonlocal treatment keeps throughout, with final state in for all tested. This lower bound follows from the minimum version of the argument in Theorem 3.1(c), since .
| Explicit reaction | Nonlocal reaction (3) | |
| 0.1 | 1.0000 | 1.0000 |
| 0.25 | 1.0000 | 1.0000 |
| 0.5 | 1.0000 | |
| 1 | 1.0000 |
Table 5. Smallest value of over the whole run for data above : , , , , , . Both columns use the same and the same lagged diffusion.
Figure 1 shows the discrete energy for and the initial datum of Remark 3.5. For the energy decreases monotonically for , as guaranteed by Theorem 3.4.
For the same data give a nonmonotone history for (Remark 3.5). In randomized tests with 3000 steps for each (random data, , ), the energy never increased for , but it increased in 95 steps for and in 250 steps for .
Conclusion
Scheme (3) guarantees positivity, an invariant region, exact constant equilibria, mass conservation, and a maximum principle for every step size, as well as an energy inequality for , at the cost of one tridiagonal solve per step. Several questions remain open. Convergence of (3) to the solution of the regularized problem is supported by Proposition 3.6 and Table 1–Table 2, but is not proved here. The standard error analysis is hampered by the dependence of the conductance on the discrete gradient, which introduces a factor . The dependence of the error constants on and the limit are not analyzed. The scheme is only first-order accurate in time, and since grows exponentially in , should be kept moderate. In two space dimensions the -matrix structure persists, but a sparse linear system has to be solved in each step. Finally, an energy-stable variant for remains open. Remark 3.5 shows that it requires a modification of (3).
Declarations
Data availability
No data was used for the research described in this letter.
Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this letter.
References
- [1]E. DiBenedetto, Degenerate Parabolic Equations, Universitext, Springer New York, 1993, eBook published 2012. doi:10.1007/978-1-4612-0895-2.
- [2]J. L. Vázquez, The Porous Medium Equation: Mathematical Theory, Oxford Mathematical Monographs, Oxford University Press, Oxford, 2006. doi:10.1093/acprof:oso/9780198569039.001.0001.DOI
- [3]J. W. Barrett, W. B. Liu, Finite element approximation of the p-Laplacian, Math. Comput. 61 (204) (1993) 523–537. doi:10.1090/S0025-5718-1993-1192966-4.DOI
- [4]A. Zinihi, M. R. Sidi Ammi, D. F. M. Torres, Fractional differential equations of a reaction–diffusion SIR model involving the Caputo-fractional time-derivative and a nonlinear diffusion operator, Evol. Eqs. Contr. Theo. 14 (2025) 944–967. doi:10.3934/eect.2025018.DOI
- [5]R. E. Mickens, Nonstandard finite difference schemes for reaction–diffusion equations, Numer. Meth. Part. Diff. Eqs. 15 (2) (1999) 201–214. doi:10.1002/(SICI)1098-2426(199903)15:2<201::AID-NUM5>3.0.CO;2-H.doi.org/10.1002/(sici)1098-2426(199903)15:2<201::aid-num5>3.0.co;2-h
- [6]R. E. Mickens, Calculation of denominator functions for nonstandard finite difference schemes for differential equations satisfying a positivity condition, Numer. Meth. Part. Diff. Eqs. 23 (3) (2007) 672–691. doi:10.1002/num.20198.DOI
- [7]A. Zinihi, M. Ehrhardt, M. R. Sidi Ammi, A nonstandard finite difference scheme for a nonlinear parabolic equation with p-Laplacian-type diffusion, preprint, arXiv:2607.00489 (2026). doi:10.48550/arXiv.2607.00489.DOI
- [8]M. Ehrhardt, R. E. Mickens, A nonstandard finite difference scheme for convection–diffusion equations having constant coefficients, Appl. Math. Comput. 219 (12) (2013) 6591–6604. doi:10.1016/j.amc.2012.12.068.DOI
- [9]T. F. Chan, P. Mulet, On the convergence of the lagged diffusivity fixed point method in total variation image restoration, SIAM J. Numer. Anal. 36 (2) (1999) 354–367. doi:10.1137/S0036142997327075.DOI
- [10]L. Diening, M. Fornasier, R. Tomasi, M. Wank, A relaxed Kačanov iteration for the p-Poisson problem, Numer. Math. 145 (1) (2020) 1–34. doi:10.1007/s00211-020-01107-1.DOI
- [11]L. H. Thomas, Elliptic problems in linear difference equations over a network, Watson Sci. Comput. Lab. Rept., Columbia University, New York 1 (71) (1949) 71–72.