betaFeedback

WorksMathematics

Unconditionally Positive Nonstandard Finite Difference Schemes for pp-Laplacian Reaction–Diffusion Equations

Abstract

We propose a conservative, semi-implicit nonstandard finite difference scheme for the reaction–diffusion equation ut=λ(∣ux∣p−2ux)x+f(u)u_t=λ(|u_x|^{p-2}u_x)_x+f(u), p>1p>1, where, as is customary, the pp-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 [0,K][0,K] 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 1<p≤21<p\le 2, decreases a discrete energy. An example shows that the energy may increase for p>2p>2. The truncation error is O(k+h2)O(k+h^2) 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 pp-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 uimu_i^m in the update is 1−λk(ci+1/2+ci−1/2)/h21 - \lambda k (c_{i+1/2} + c_{i-1/2})/h^2, where the conductance c≈∣ux∣p−2c \approx|u_x|^{p-2} 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 pp grows (§ 4).

Nonstandard finite difference (NSFD) methods address such failures with two devices: a denominator function φ(k)=k+O(k2)\varphi(k) = k + O(k^2) that replaces kk 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 pp-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 MM-matrix. Unconditional positivity, an invariant region, exact equilibria, and an energy inequality then follow from short arguments, for every p>1p>1 and for a whole class of reactions. We also show where the argument stops: for p>2p>2 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 um>0u^m>0, the conductance is not regularized for p<2p<2, and the ℓ∞\ell^\infty 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 MM-matrix. At the cost of one tridiagonal solve per step, this yields properties that the explicit scheme does not provide: invariance of [0,K][0,K] uniformly in kk, hh and mm, mass conservation and a discrete maximum principle, the exact logistic flow in the reaction part (Remark 3.3), and a discrete energy inequality for 1<p≤21<p\leq2 together with an example showing that it fails for p>2p>2. We also give numerical evidence, including a comparison with iterating the lagged conductance to convergence (Table 4).

ppkkPicard solves per step∥u−uPic∥∞\lVert u-u_{\mathrm{Pic}}\rVert_\infty at TT
1.510−410^{-4}5.41.5×10−41.5\times10^{-4}
1.510−310^{-3}10.11.4×10−31.4\times10^{-3}
1.510−210^{-2}17.21.5×10−21.5\times10^{-2}
310−410^{-4}4.61.9×10−41.9\times10^{-4}
310−310^{-3}≥28.9\ge28.96.8×10−46.8\times10^{-4}
310−210^{-2}≥168\ge1681.2×10−21.2\times10^{-2}

Table 4. Scheme (3) (solution uu) versus the converged Picard iteration (solution uPicu_{\mathrm{Pic}}) within each step, stopped when ∥vj+1−vj∥∞<10−8\lVert v^{j+1}-v^j\rVert_\infty<10^{-8} or after 200 iterations; data as in Table 3. Entries marked ≥\ge are lower bounds, since the cap was reached in some steps.

The Scheme

For x∈(0,1)x\in(0,1) and t>0t>0 we consider a reaction–diffusion equation of the form

ut=λ(Fε(ux))x+f(u),Fε(s)=(s2+ε2)p−22s,(1)u_t=\lambda\left(F_\varepsilon(u_x)\right)_x+f(u),\qquad F_\varepsilon(s)=\left(s^2+\varepsilon^2\right)^{\frac{p-2}{2}}s, \tag*{(1)}

with zero-flux boundary conditions Fε(ux)=0F_\varepsilon(u_x)=0 at x=0,1x=0,1, initial data u(⋅,0)=u0≥0u(\cdot,0)=u_0\geq0, λ>0\lambda>0, p>1p>1, and a fixed regularization parameter ε>0\varepsilon>0, as in finite element treatments of the pp-Laplacian [3]. For p=2p=2, (1) reduces to the semilinear heat equation for every ε\varepsilon. The reaction is split as

f(s)=a(s)−s b(s),a,b∈C[0,∞),a,b≥0.(2)f(s)=a(s)-s\,b(s),\qquad a,b\in C[0,\infty),\qquad a,b\geq0. \tag*{(2)}

For the logistic term f(s)=s(1−s)f(s)=s(1-s) one takes a(s)=b(s)=sa(s)=b(s)=s; for f∈C1[0,∞)f\in C^1[0,\infty) with f(0)≥0f(0)\geq0 one may take a=f+a=f^+ and b(s)=f−(s)/sb(s)=f^-(s)/s, extended continuously to s=0s=0.

Let h=1/Nh=1/N, xi=ihx_i=ih, Div=(vi+1−vi)/hD_i v=(v_{i+1}-v_i)/h, and let w0=wN=h/2w_0=w_N=h/2 and wi=hw_i=h for 0<i<N0<i<N. For a grid function uu we define the conductances and the discrete diffusion operator

ci+1/2[u]=((Diu)2+ε2)p−22>0,(L[u]v)i=1wi(ci+1/2Div−ci−1/2Di−1v),c_{i+1/2}[u]=\left((D_i u)^2+\varepsilon^2\right)^{\frac{p-2}{2}}>0,\qquad(L[u]v)_i=\frac{1}{w_i}\left(c_{i+1/2}D_i v-c_{i-1/2}D_{i-1}v\right),

with the zero-flux convention c−1/2D−1v=cN+1/2DNv=0c_{-1/2}D_{-1}v=c_{N+1/2}D_Nv=0. With the denominator function φ(k)=(eLk−1)/L\varphi(k)=(e^{\mathcal{L}k}-1)/\mathcal{L}, L>0\mathcal{L}>0, which satisfies φ(k)=k+O(k2)\varphi(k)=k+O(k^2) [6], the NSFD scheme reads

uim+1−uimφ(k)=λ(L[um]um+1)i+a(uim)−b(uim)uim+1,i=0,…,N.(3)\frac{u_i^{m+1}-u_i^m}{\varphi(k)} =\lambda\left(L[u^m]u^{m+1}\right)_i+a(u_i^m)-b(u_i^m)u_i^{m+1},\qquad i=0,\ldots,N. \tag*{(3)}

The conductance is evaluated at the old level, the solution enters linearly at the new level, and the loss term sb(s)s b(s) is replaced by b(um)um+1b(u^m)u^{m+1}. Writing φ=φ(k)\varphi=\varphi(k) and θi=φλ/(wih)\theta_i=\varphi\lambda/(w_i h), we can rewrite (3) as Bmum+1=um+φa(um)B^m u^{m+1}=u^m+\varphi a(u^m) with the tridiagonal matrix

Biim=1+φb(uim)+θi(ci+1/2+ci−1/2),Bi,i+1m=−θici+1/2,Bi,i−1m=−θici−1/2,B_{ii}^m=1+\varphi b(u_i^m)+\theta_i\left(c_{i+1/2}+c_{i-1/2}\right),\qquad B_{i,i+1}^m=-\theta_i c_{i+1/2},\qquad B_{i,i-1}^m=-\theta_i c_{i-1/2},

where c−1/2=cN+1/2=0c_{-1/2}=c_{N+1/2}=0. Hence, one step costs O(N)O(N) 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 p>1p>1 and ε,λ,k,h>0\varepsilon,\lambda,k,h>0 be arbitrary, let the splitting of the reaction (2) hold, and let um≥0u^m\geq0 be a grid function.

(a) BmB^m is a strictly diagonally dominant MM-matrix with ∥(Bm)−1∥∞≤1\lVert(B^m)^{-1}\rVert_\infty\leq1. Hence (3) has a unique solution um+1u^{m+1}.

(b) um+1≥0u^{m+1}\geq0, and um+1>0u^{m+1}>0 if um>0u^m>0.

(c) If a(s)≤Kb(s)a(s)\leq Kb(s) on [0,K][0,K] for some K>0K>0 and 0≤u0≤K0\leq u^0\leq K, then 0≤um≤K0\leq u^m\leq K for all mm.

(d) If f(u∗)=0f(u^*)=0 for some u∗≥0u^*\geq0 and um≡u∗u^m\equiv u^*, then um+1≡u∗u^{m+1}\equiv u^*.

(e) If f≡0f\equiv0, then ∑iwiuim\sum_i w_i u_i^m is independent of mm, and min⁡iui0≤uim≤max⁡iui0\min_i u_i^0\leq u_i^m\leq\max_i u_i^0.

Proof. (a) BmB^m has a positive diagonal, non-positive off-diagonal entries, and Biim−∑j≠i∣Bijm∣=1+φb(uim)≥1B^m_{ii}-\sum_{j\ne i}|B^m_{ij}|=1+\varphi b(u_i^m)\geq1. For any vector vv, choose ii with ∣vi∣=∥v∥∞|v_i|=\lVert v\rVert_\infty. Then ∣(Bmv)i∣≥Biim∣vi∣−∑j≠i∣Bijm∣∣vj∣≥(Biim−∑j≠i∣Bijm∣)∥v∥∞≥∥v∥∞|(B^m v)_i|\geq B^m_{ii}|v_i|-\sum_{j\ne i}|B^m_{ij}||v_j|\geq(B^m_{ii}-\sum_{j\ne i}|B^m_{ij}|)\lVert v\rVert_\infty\geq\lVert v\rVert_\infty, so ∥Bmv∥∞≥∥v∥∞\lVert B^m v\rVert_\infty\geq\lVert v\rVert_\infty. Hence BmB^m is nonsingular and ∥(Bm)−1∥∞≤1\lVert(B^m)^{-1}\rVert_\infty\leq1. Writing Bm=D−RB^m=D-R with DD diagonal and R≥0R\geq0, diagonal dominance gives ∥D−1R∥∞<1\lVert D^{-1}R\rVert_\infty<1, so (Bm)−1=∑j≥0(D−1R)jD−1≥0(B^m)^{-1}=\sum_{j\geq0}(D^{-1}R)^jD^{-1}\geq0 entry-wise, with diagonal entries at least 1/Dii1/D_{ii}.

(b) Since a≥0a\geq0 and φ>0\varphi>0, the right-hand side satisfies um+φa(um)≥um≥0u^m+\varphi a(u^m)\geq u^m\geq0. As (Bm)−1≥0(B^m)^{-1}\geq0, we get um+1≥0u^{m+1}\geq0. If um>0u^m>0, then uim+1≥(uim+φa(uim))/Dii>0u_i^{m+1}\geq(u_i^m+\varphi a(u_i^m))/D_{ii}>0.

(c) We argue by induction. To do so, assume 0≤um≤K0\leq u^m\leq K. By (b), um+1≥0u^{m+1}\geq0. Let M:=max⁡iuim+1=ui∗m+1M:=\max_i u_i^{m+1}=u_{i^*}^{m+1} and β:=b(ui∗m+1)\beta:=b(u_{i^*}^{m+1}). In row i∗i^* of (3) the diffusion part θi∗∑j=i∗±1(ci∗+j)/2(ui∗m+1−ujm+1)\theta_{i^*}\sum_{j=i^*\pm1}(c_{i^*+j})/2(u_{i^*}^{m+1}-u_j^{m+1}) is nonnegative, so (1+φβ)M≤ui∗m+φa(ui∗m)≤K+φKβ(1+\varphi\beta)M\leq u_{i^*}^m+\varphi a(u_{i^*}^m)\leq K+\varphi K\beta, where we used ui∗m≤Ku_{i^*}^m\leq K and a≤Kba\leq Kb on [0,K][0,K]. Hence M≤KM\leq K.

(d) Constants lie in the kernel of the diffusion part, and a(u∗)=u∗b(u∗)a(u^*)=u^*b(u^*). Hence Bmu∗=(1+φb(u∗))u∗=u∗+φa(u∗)B^m u^*=(1+\varphi b(u^*))u^*=u^*+\varphi a(u^*), i.e., u∗u^* solves (3).

(e) With a=b=0a=b=0, multiplying (3) by wiw_i and summing over ii, the diffusion terms telescope to zero by the zero-flux convention. At a maximizer i∗i^* of um+1u^{m+1}, row i∗i^* gives ui∗m+1≤ui∗m≤max⁡iuimu_{i^*}^{m+1}\leq u_{i^*}^m\leq\max_i u_i^m. The minimum is treated in the same way.

Remark 3.2. The proof of Theorem 3.1 uses only the positivity of the conductances ci+1/2c_{i+1/2}, not the way they are computed. Parts (a)–(e) therefore hold for any positive conductance in place of ci+1/2[um]c_{i+1/2}[u^m], 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 c=c[um]c=c[u^m].

Remark 3.3. For the logistic reaction with a(s)=b(s)=sa(s)=b(s)=s and L=1L=1, the reaction part of (3) is the map u↦u(1+φ)/(1+φu)=uek/(1+u(ek−1))u\mapsto u(1+\varphi)/(1+\varphi u)=ue^k/(1+u(e^k-1)), which is the exact flow of u′=u(1−u)u'=u(1-u). Part (c) holds with K=1K=1.

For the energy estimate, let G(s)=(s2+ε2)p/2/pG(s)=(s^2+\varepsilon^2)^{p/2}/p, so that G′=FεG'=F_\varepsilon, and define

Eh(v)=h∑i=0N−1G(Div),∥v∥W2=∑i=0Nwivi2.E_h(v)=h\sum_{i=0}^{N-1}G(D_i v),\qquad\lVert v\rVert_W^2=\sum_{i=0}^{N}w_i v_i^2.

Theorem 3.4. Let f≡0f\equiv0 and 1<p≤21<p\leq2. Then, for every k,h,ε,λ>0k,h,\varepsilon,\lambda>0,

Eh(um+1)+1λφ(k)∥um+1−um∥W2≤Eh(um).E_h(u^{m+1})+\frac{1}{\lambda\varphi(k)}\lVert u^{m+1}-u^m\rVert_W^2\leq E_h(u^m).

Proof. Write u+=um+1u^{+}=u^{m+1}, u−=umu^{-}=u^{m}, z=u+−u−z=u^{+}-u^{-}, and ci=ci+1/2[u−]c_i=c_{i+1/2}[u^{-}]. Summation by parts with the zero-flux convention gives ∑iwizi(L[u−]u+)i=−h∑iciDizDiu+\sum_i w_i z_i(L[u^{-}]u^{+})_i=-h\sum_i c_iD_i zD_i u^{+}. Multiplying (3) by wiziw_i z_i and summing yields

h∑iciDizDiu+=−1λφ∥z∥W2.(4)h\sum_i c_iD_i zD_i u^{+}=-\frac{1}{\lambda\varphi}\|z\|_W^2. \tag*{(4)}

Let g(t)=(t+ε2)p/2/pg(t)=(t+\varepsilon^2)^{p/2}/p, so that G(s)=g(s2)G(s)=g(s^2). For p≤2p\le2 the function gg is concave, hence G(s′)≤G(s)+g′(s2)(s′2−s2)G(s')\le G(s)+g'(s^2)(s'^2-s^2) with g′(s2)=12(s2+ε2)(p−2)/2g'(s^2)=\frac{1}{2}(s^2+\varepsilon^2)^{(p-2)/2}. Applying this with s=Diu−s=D_i u^{-} and s′=Diu+s'=D_i u^{+} gives Eh(u+)≤Eh(u−)+h2∑ici((Diu+)2−(Diu−)2)E_h(u^{+})\le E_h(u^{-})+\frac{h}{2}\sum_i c_i((D_i u^{+})^2-(D_i u^{-})^2). Since (Diu+)2−(Diu−)2=2DizDiu+−(Diz)2(D_i u^{+})^2-(D_i u^{-})^2=2D_i zD_i u^{+}-(D_i z)^2, we obtain Eh(u+)≤Eh(u−)+h∑iciDizDiu+−h2∑ici(Diz)2E_h(u^{+})\le E_h(u^{-})+h\sum_i c_iD_i zD_i u^{+}-\frac{h}{2}\sum_i c_i(D_i z)^2, and the claim follows from (4) after dropping the last (non-positive) term.

Remark 3.5. For p>2p>2 the function gg is convex, and not only the argument but also the conclusion fails. Take p=3p=3, N=40N=40, ε=0.01\varepsilon=0.01, λ=1\lambda=1, k=0.1k=0.1, and u0(x)=1+12cos⁡7πx+12e−60(x−1/2)2u_0(x)=1+\frac{1}{2}\cos7\pi x+\frac{1}{2}e^{-60(x-1/2)^2}. In the third step, EhE_h rises by about 22%, from about 1.7⋅10−41.7\cdot10^{-4} to 2.1⋅10−42.1\cdot10^{-4} (Figure 1, right).

Discrete energy curves for p=1.5 and p=3

Figure 1. Discrete energy Eh(um)E_h(u^m) for f=0f=0, N=40N=40, ε=0.01\varepsilon=0.01, λ=1\lambda=1, and u0(x)=1+12cos⁡7πx+12e−60(x−1/2)2u_0(x)=1+\frac{1}{2}\cos7\pi x+\frac{1}{2}e^{-60(x-1/2)^2}. Left: p=1.5p=1.5 (monotone for all kk). Right: p=3p=3 (an increase occurs for k=0.1k=0.1).

Proposition 3.6. Let aa and bb be Lipschitz continuous, and let uu be a solution of (1) that is C4C^4 in xx and C2C^2 in tt. Inserting uu into (3) leaves a residual τim=O(k+h2)\tau_i^m=O(k+h^2) at interior nodes and τim=O(k+h)\tau_i^m=O(k+h) for i∈{0,N}i\in\{0,N\}, with constants depending on pp, ε\varepsilon, λ\lambda, and norms of uu.

Proof. Since φ(k)=k+O(k2)\varphi(k)=k+O(k^2), the time quotient equals ut+O(k)u_t+O(k). Because ci+1/2≤εp−2c_{i+1/2}\le\varepsilon^{p-2} for p<2p<2 and ci+1/2c_{i+1/2} is bounded on bounded gradients for p≥2p\ge2, replacing u(tm+1)u(t_{m+1}) by u(tm)u(t_m) in the diffusion and reaction terms costs O(k)O(k). For the spatial part, Diu=ux(xi+1/2)+h224uxxx(xi+1/2)+O(h3)D_i u=u_x(x_{i+1/2})+\frac{h^2}{24}u_{xxx}(x_{i+1/2})+O(h^3) and FεF_\varepsilon is smooth, so ci+1/2Diu=Fε(ux)(xi+1/2)+h2ψ(xi+1/2)+O(h3)c_{i+1/2}D_i u=F_\varepsilon(u_x)(x_{i+1/2})+h^2\psi(x_{i+1/2})+O(h^3) with ψ∈C1\psi\in C^1. The difference quotient of this flux is (Fε(ux))x+O(h2)(F_\varepsilon(u_x))_x+O(h^2) at interior nodes. At i=0i=0 the half cell gives (2/h)Fε(ux)(h/2)+O(h)=(Fε(ux))x(0)+O(h)(2/h)F_\varepsilon(u_x)(h/2)+O(h)=(F_\varepsilon(u_x))_x(0)+O(h), using Fε(ux(0))=0F_\varepsilon(u_x(0))=0. The case i=Ni=N is analogous.

Numerical Experiments

Unless stated otherwise, all runs use L=1\mathcal{L}=1, λ=1\lambda=1, and the logistic reaction with a(s)=b(s)=sa(s)=b(s)=s. We use the manufactured solution u(x,t)=1+12e−tcos⁡πxu(x,t)=1+\frac{1}{2}e^{-t}\cos\pi x and add to (1) the source term gg that makes it exact for the regularized operator, g+g^{+} is included in aa and g−/ug^{-}/u in bb. Table 1 and Table 2 show first order in time and second order in space for p=1.5p=1.5 and p=3p=3. The last spatial rates fall below two because the first-order time error is still visible: for p=3p=3 and N=320N=320, halving kk changes the error from 2.85⋅10−62.85\cdot10^{-6} to 2.55⋅10−62.55\cdot10^{-6}.

Table 1: Time refinement with N=320N=320: maximum-norm errors and observed orders (ε=0.2\varepsilon=0.2, T=0.2T=0.2).

kkp=1.5p=1.5p=3p=3
ErrorOrderErrorOrder
T/21T/2^15.10×10−25.10\times10^{-2}–3.88×10−23.88\times10^{-2}–
T/22T/2^23.25×10−23.25\times10^{-2}0.652.02×10−22.02\times10^{-2}0.94
T/23T/2^31.86×10−21.86\times10^{-2}0.801.05×10−21.05\times10^{-2}0.94
T/24T/2^41.00×10−21.00\times10^{-2}0.905.37×10−35.37\times10^{-3}0.97
T/25T/2^55.19×10−35.19\times10^{-3}0.952.72×10−32.72\times10^{-3}0.98
T/26T/2^62.64×10−32.64\times10^{-3}0.971.36×10−31.36\times10^{-3}0.99
T/27T/2^71.33×10−31.33\times10^{-3}0.986.84×10−46.84\times10^{-4}1.00

Table 1.

Table 2: Space refinement with k=T/217k=T/2^{17}: maximum-norm errors and observed orders (ε=0.2\varepsilon=0.2, T=0.2T=0.2).

NNp=1.5p=1.5p=3p=3
ErrorOrderErrorOrder
103.95×10−23.95\times10^{-2}–3.69×10−33.69\times10^{-3}–
203.18×10−33.18\times10^{-3}3.635.57×10−45.57\times10^{-4}2.73
402.87×10−42.87\times10^{-4}3.471.45×10−41.45\times10^{-4}1.94
806.99×10−56.99\times10^{-5}2.043.67×10−53.67\times10^{-5}1.98
1601.84×10−51.84\times10^{-5}1.929.63×10−69.63\times10^{-6}1.93
3205.59×10−65.59\times10^{-6}1.722.85×10−62.85\times10^{-6}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 h=0.01h=0.01. For p=3p=3 the explicit method blows up even for k=10−4k=10^{-4}, since the conductance grows with the gradient. For p=1.5p=1.5 it produces negative values for every kk we tried. For k=10−4k=10^{-4} the violation is small (min⁡u=−3.2⋅10−5\min u=-3.2\cdot10^{-5} at intermediate times, and the solution is positive again at TT), but it becomes large for larger kk. Scheme (3) stays in [0,1][0,1] in all cases, as predicted by Theorem 3.1. Note that for large kk we claim only positivity and boundedness, the accuracy is still O(k)O(k).

ppkkExplicit FDMScheme (3): [min⁡u,max⁡u][\min u,\max u] at TT
1.510−410^{-4}negative values (min⁡u=−3.2×10−5\min u=-3.2\times10^{-5})[0.1388,0.1582][0.1388, 0.1582]
1.510−310^{-3}negative values (min⁡u=−2.1\min u=-2.1)[0.1369,0.1603][0.1369, 0.1603]
1.510−210^{-2}negative values (min⁡u=−311.9\min u=-311.9)[0.1261,0.1832][0.1261, 0.1832]
310−410^{-4}blow-up at t=0.0003t=0.0003[0.0764,0.2211][0.0764, 0.2211]
310−310^{-3}blow-up at t=0.0030t=0.0030[0.0753,0.2222][0.0753, 0.2222]
310−210^{-2}blow-up at t=0.0200t=0.0200[0.0544,0.2432][0.0544, 0.2432]

Table 3. Explicit FDM versus scheme (3) for u0(x)=0.02+0.98e−200(x−1/2)2u_0(x)=0.02+0.98e^{-200(x-1/2)^2}, ε=0.05\varepsilon=0.05, h=0.01h=0.01, T=0.05T=0.05; “blow-up” means ∥u∥∞>104\lVert u\rVert_\infty>10^4.

The lag can be removed by a Picard iteration within each step: starting from v0=umv^0=u^m, the iterate vj+1v^{j+1} solves (3) with c[um]c[u^m] replaced by c[vj]c[v^j], 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 ∥vj+1−vj∥∞<10−8\lVert v^{j+1}-v^j\rVert_\infty<10^{-8} for the data of Table 3. For p=1.5p=1.5 it ranges from about 5 to 17: for p=3p=3 it grows quickly with kk, 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 kk, 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 K=1K=1 and a small diffusion coefficient, this variant produces negative values for k=0.5k=0.5 (min⁡u=−0.20\min u=-0.20) and reaches −181-181 for k=1k=1. The nonlocal treatment keeps u≥1u\ge1 throughout, with final state in [1,1.0313][1,1.0313] for all kk tested. This lower bound follows from the minimum version of the argument in Theorem 3.1(c), since a=ba=b.

kkExplicit reactionNonlocal reaction (3)
0.11.00001.0000
0.251.00001.0000
0.5−0.2017-0.20171.0000
1−181.1501-181.15011.0000

Table 5. Smallest value of uu over the whole run for data above K=1K=1: u0(x)=1+2e−50(x−1/2)2u_0(x)=1+2e^{-50(x-1/2)^2}, p=3p=3, λ=10−3\lambda=10^{-3}, ε=0.05\varepsilon=0.05, N=100N=100, T=3T=3. Both columns use the same φ(k)\varphi(k) and the same lagged diffusion.

Figure 1 shows the discrete energy Eh(um)E_h(u^m) for f=0f=0 and the initial datum of Remark 3.5. For p=1.5p=1.5 the energy decreases monotonically for k=0.01,0.1,1k=0.01,0.1,1, as guaranteed by Theorem 3.4.

For p=3p=3 the same data give a nonmonotone history for k=0.1k=0.1 (Remark 3.5). In randomized tests with 3000 steps for each pp (random data, k∈[10−4,10]k \in[10^{-4},10], ε∈[10−3,10−0.5]\varepsilon\in[10^{-3},10^{-0.5}]), the energy never increased for p∈{1.5,2,2.5}p \in\{1.5,2,2.5\}, but it increased in 95 steps for p=3p=3 and in 250 steps for p=5p=5.

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 1<p≤21<p\le2, 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 h−1h^{-1}. The dependence of the error constants on ε\varepsilon and the limit ε→0\varepsilon\to0 are not analyzed. The scheme is only first-order accurate in time, and since φ(k)\varphi(k) grows exponentially in Lk\mathcal{L}k, L\mathcal{L} should be kept moderate. In two space dimensions the MM-matrix structure persists, but a sparse linear system has to be solved in each step. Finally, an energy-stable variant for p>2p>2 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. [1]E. DiBenedetto, Degenerate Parabolic Equations, Universitext, Springer New York, 1993, eBook published 2012. doi:10.1007/978-1-4612-0895-2.
  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. [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. [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. [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. [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. [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. [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. [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. [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. [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.

Paper details

vlnc:W_7b4z4j4p
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents