betaFeedback

WorksMathematics

Fully Explicit Stabilized Runge–Kutta Methods for Allen–Cahn and Cahn–Hilliard Dynamics

Abstract

Stiff phase-field dynamics are typically solved by semi-implicit or fully implicit time discretizations, since the interfacial scales in the Allen–Cahn and Cahn–Hilliard equations impose severe step-size restrictions on standard explicit methods. In this paper, we investigate a different strategy based on stabilized Runge–Kutta solvers, in particular the Runge–Kutta–Legendre (RKL) and Runge–Kutta–Gegenbauer (RKG) families. These schemes possess enlarged stability regions and therefore provide a natural explicit framework for diffusion-dominated gradient flows without requiring implicit solves. For the Allen–Cahn equation, combined with Strang splitting, we establish a maximum principle in 1-dimensional case, and prove energy dissipation in all the dimensions. For the modified energy dissipation analysis, we further introduce two new first-order stabilized schemes of RKL and RKG type and derive a novel eigenvalue estimate for the second-order RKG method. We also conduct extensive numerical experiments with both constant and adaptive time stepping, on regular domains as well as on an L-shaped irregular domain. The numerical results show that stabilized explicit Runge–Kutta methods provide an effective and flexible approach for the numerical simulation of Allen–Cahn and Cahn–Hilliard dynamics while preserving key structural properties. The Python implementation is available at https://github.com/ZT220501/SRK-for-Gradient-Flows.

Graphical Abstract

Highlights

Fully Explicit Stabilized Runge–Kutta Methods for Allen–Cahn and Cahn–Hilliard Dynamics

Zheng Tan, Tariq D. Aslam, Andrea L. Bertozzi

  • Stabilized Runge–Kutta methods for fully explicit, non-spectral solution approach of the Allen–Cahn and the Cahn–Hilliard Dynamics.

  • The Maximum principle is proved when applying to the Allen–Cahn equation.

  • Modified energy dissipation is proved for two Strang splitting orders.

  • The method applies to irregular domains without Fourier spectral methods.

Introduction

Phase-field models are prevalent in science and engineering, particularly in material science and fluid dynamics, see for example [1, 2, 3, 4, 5, 6, 7]. Therefore, it is important to efficiently and accurately solve the equations numerically. In this work, we consider two of the most well-known phase-field equations, namely the Allen–Cahn equation

∂u∂t=ε2∇2u−F′(u),x∈Ω,t∈[0,T],(1)\frac{\partial u}{\partial t} = \varepsilon^{2}\nabla^{2}u - F'(u), \qquad x \in\Omega, t \in[0,T], \tag*{(1)}

and the Cahn–Hilliard equation

∂u∂t=∇2(−ε2∇2u+F′(u)),x∈Ω,t∈[0,T],(2)\frac{\partial u}{\partial t} = \nabla^{2}\left(-\varepsilon^{2}\nabla^{2}u + F'(u)\right), \qquad x \in\Omega, t \in[0,T], \tag*{(2)}

with initial condition u(x,0)=u0(x)u(x,0)=u_{0}(x). Here Ω⊂Rn\Omega\subset\mathbb{R}^{n} is a bounded domain and T>0T>0 is a finite time. Throughout this paper, we use the classical double-well potential F(u):=14(u2−1)2F(u):=\frac{1}{4}(u^{2}-1)^{2}. We use either the periodic boundary conditions, or the homogeneous Neumann boundary condition

∂u∂n∣∂Ω=∂u∂n∣∂Ω=0,\left.\frac{\partial u}{\partial n}\right|_{\partial\Omega}=\left.\frac{\partial u}{\partial n}\right|_{\partial\Omega}=0,

*Corresponding author

Email addresses: [email protected] (Zheng Tan), [email protected] (Tariq D. Aslam), [email protected] (Andrea L. Bertozzi) where μ=−ε2∇2u+F′(u)\mu= -\varepsilon^2\nabla^2 u + F'(u) and nn is the outward normal of ∂Ω\partial\Omega.

The Allen–Cahn and the Cahn–Hilliard equations are well-known to be the gradient flows of the Ginzburg–Landau energy

E(u(t))=∫Ωε22∣∇u∣2+F(u) dx.(3)E(u(t)) = \int_{\Omega} \frac{\varepsilon^2}{2}|\nabla u|^2 + F(u)\,dx. \tag*{(3)}

More precisely, the L2L^2-gradient flow results in the Allen–Cahn equation, while the H−1H^{-1}-gradient flow results in the Cahn–Hilliard equation. Therefore, a successful numerical scheme should also preserve the energy dissipation property to obtain physically correct solutions; more precisely, the numerical solution should satisfy

E(u(t1))≤E(u(t0)),∀t1≥t0.E(u(t_1)) \le E(u(t_0)), \quad\forall t_1 \ge t_0.

We call a scheme energy stable if the above condition is satisfied. Another important property of the second-order parabolic Allen–Cahn equation is the maximum principle: if ∥u(x,0)∥∞≤1\lVert u(x,0)\rVert_{\infty} \le1, then ∥u(x,t)∥∞≤1\lVert u(x,t)\rVert_{\infty} \le1 for all t>0t > 0. Therefore, it is also essential for a numerical scheme to satisfy the maximum principle for the Allen–Cahn equation.

Preserving numerical energy stability is central to schemes for gradient flow simulations. An abundance of numerical evidence shows that spurious oscillations may occur when the energy stability condition is violated. Numerical effort has been devoted for energy stable schemes development, including [8, 9, 10, 11]. Since the spatial derivative is second or fourth order, the stability constraint limits the usage of classical explicit methods, leading to the need to solve an implicit system at each step. The recently deployed exponential time differencing Runge–Kutta schemes (ETDRK) [12, 13, 14], although they avoid implicit solves and monotonically dissipate energy, rely heavily on the Fourier spectral methods so that on non-periodic irregular domains, it degenerates into an implicit method because of the need for a stabilization parameter.

In this paper, we explore the potential of fully explicit stabilized Runge–Kutta methods called Runge–Kutta–Legendre (RKL) and Runge–Kutta–Gegenbauer (RKG) methods for solving the Allen–Cahn and Cahn–Hilliard equations. The RKL and RKG methods are super-time-stepping multistage methods that admit stability constraints to be relaxed quadratically with their number of stages. We prove that the RKL and RKG methods preserve the maximum principle for the Allen–Cahn equation when combined with the Strang splitting, under periodic or Neumann boundary conditions. Modified energy dissipations for the RKL and RKG methods are proved when they are applied to the Allen–Cahn equation, and the dissipated energy is shown to converge to the exact energy. We also conduct numerous experiments to demonstrate the behaviors when the RKL and RKG methods are applied to the gradient flows. Specifically, an irregular domain experiment is conducted to demonstrate the effectiveness of the methods for a case where the Fourier spectral method cannot be easily applied.

This paper is organized as follows. In Section 2 we review some of the well-known numerical schemes that efficiently and accurately solve the gradient flows while preserving the energy stability, with their advantages and limitations. Section 3 introduces the RKL and RKG methods and describes how we apply them to the gradient flows. Following that, in Section 4 we prove the maximum principle and the modified energy dissipation for RKL and RKG schemes. We conduct comprehensive experiments in Section 5 to demonstrate the behavior of RKL and RKG methods when applied to gradient flows. Concluding remarks and potential research directions are presented in Section 6.

Review of schemes for simulating gradient flows

Various schemes have been designed in order to satisfy the maximum principle of the Allen–Cahn equation and energy stability of the gradient flows. In this section, we present several of the most representative methods, in roughly chronological order.

Convex splitting

One of the pioneering works for energy dissipating Cahn–Hilliard simulation is the convex splitting technique [8]. Assume that the energy F(u)F(u) can be split into the difference of two convex functions; more precisely, F(u)=F1(u)−F2(u)F(u) = F_1(u) - F_2(u) where both F1,F2F_1, F_2 are convex. Then one treats the positive convex part to be implicit and negative convex part to be explicit, so that the first-order convex splitting method for Cahn–Hilliard reads

un+1−unτ=∇2μn+1;μn+1=(−ε2∇2un+1+F1′(un+1)−F2′(un)).(4)\frac{u^{n+1}-u^n}{\tau}=\nabla^2\mu^{n+1};\qquad\mu^{n+1}=\left(-\varepsilon^2\nabla^2u^{n+1}+F'_1(u^{n+1})-F'_2(u^n)\right). \tag*{(4)}

It is shown that [8]

E(un+1)−E(un)≤−τ∥∇μn+1∥2,E(u^{n+1})-E(u^n)\leq-\tau\lVert\nabla\mu^{n+1}\rVert^2,

so that the scheme dissipates the energy as desired. However, a nonlinear system needs to be solved at every step, leading to inefficient computation.

Implicit-Explicit (IMEX) methods

Another approach, to some extent similar to the convex splitting idea, is the implicit-explicit (IMEX) method [9]. Instead of splitting the energy F(u)F(u), the IMEX approach directly treats the linear part implicitly while treating the nonlinear part explicitly. Therefore, only a linear system needs to be solved at each step.

Several works have studied IMEX methods for parabolic equations and gradient flows [9, 15, 16, 17]. Although only linear system solving is required, the IMEX method may still only attain conditional energy dissipation, unless additional stabilization techniques [18] are used. Therefore, more unconditional energy dissipation techniques are developed, as will be introduced in the following subsections.

Lagrange multiplier approaches

One of the novel approaches for solving gradient flows utilizes the idea of Lagrange multiplier. X. Yang and his collaborators apply the Lagrange multiplier approach to the class of gradient flows with free energies bounded below [10, 19, 20, 21, 22, 23, 24, 25, 26] and develop the invariant energy quadratization (IEQ) method. To illustrate the approach, consider the Cahn–Hilliard equation written in the form of

∂u∂t=∇2μ;μ=−∇2u+F′(u).(5)\frac{\partial u}{\partial t}=\nabla^2\mu;\qquad\mu=-\nabla^2u+F'(u). \tag*{(5)}

With the free energy F(u)F(u) uniformly bounded below such that F(u)≥−C0∈RF(u)\geq-C_0\in\mathbb{R}, one introduces the Lagrange multiplier q(t,x;u)=F(u)+C0q(t,x;u)=\sqrt{F(u)+C_0}, and rewrites Equation (5) as

∂u∂t=∇2μ;μ=−∇2u+qF(u)+C0F′(u);∂q∂t=F′(u)2F(u)+C0∂u∂t.(6)\frac{\partial u}{\partial t}=\nabla^2\mu;\qquad\mu=-\nabla^2u+\frac{q}{\sqrt{F(u)+C_0}}F'(u);\qquad\frac{\partial q}{\partial t}=\frac{F'(u)}{2\sqrt{F(u)+C_0}}\frac{\partial u}{\partial t}. \tag*{(6)}

The IEQ method solves (6) with numerical methods such as Crank-Nicolson or BDF schemes. Taking the inner product of the above with μ\mu, utu_t, and 2q2q respectively, the IEQ method naturally satisfies the modified energy dissipation law

ddt(12∥∇u∥2+∫Ωq2 dV)=−∥∇μ∥2,\frac{d}{dt}\left(\frac{1}{2}\lVert\nabla u\rVert^2+\int_{\Omega}q^2\,dV\right)=-\lVert\nabla\mu\rVert^2,

so that the energy stability is automatically guaranteed.

Though unconditionally energy stable, the IEQ method requires solving a variable coefficient system of equations every step. This disadvantage is overcome by the scalar auxiliary variable (SAV) approach [11]. Instead of assuming the free energy F(u)F(u) is bounded below, the SAV approach assumes that the nonlinear part of energy E1(u):=∫ΩF(u)dVE_1(u):=\int_{\Omega}F(u)dV is bounded from below such that E1(u)≥−C0∈RE_1(u)\geq-C_0\in\mathbb{R}. By introducing the Lagrange multiplier r(t;u)=E1(u)+C0r(t;u)=\sqrt{E_1(u)+C_0}, the Cahn–Hilliard equation (5) can be rewritten as

∂u∂t=∇2μ;μ=−∇2u+rE1(u)+C0F′(u);∂r∂t=12E1(u)+C0∫ΩF′(u)∂u∂t dV.(7)\frac{\partial u}{\partial t}=\nabla^2\mu;\qquad\mu=-\nabla^2u+\frac{r}{\sqrt{E_1(u)+C_0}}F'(u);\qquad\frac{\partial r}{\partial t}=\frac{1}{2\sqrt{E_1(u)+C_0}}\int_{\Omega}F'(u)\frac{\partial u}{\partial t}\,dV. \tag*{(7)}

Using Crank-Nicolson or BDF methods, only scalar-coefficient systems are needed to be solved at each step for the SAV method. A similar modified energy dissipation law also holds for the SAV approach

ddt(12∥∇u∥2+r2(t))=−∥∇μ∥2.\frac{d}{dt}\left(\frac{1}{2}\lVert\nabla u\rVert^2+r^2(t)\right)=-\lVert\nabla\mu\rVert^2.

Extensive follow-up works of the SAV method [27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37] demonstrate its effectiveness across various gradient flow problems. However, the method still requires solving an implicit system of equations at each step since explicit solvers cannot be applied to (6) or (7) directly, leading to the need for solving two Poisson-type equations per step [11].

Exponential time differencing (ETD) methods

A popular and active line of research is the use of exponential time differencing (ETD) methods [38] for solving the phase-field equation. The phase-field equation has the semi-linear right hand side structure

∂u∂t=G(Lu+f(u)),\frac{\partial u}{\partial t}=G(Lu+f(u)),

where GG is a non-positive operator, LL is a positive-definite operator, and ff is a (nonlinear) function. Using Duhamel’s principle, the exact solution is

u(x,t)=eGL(t−t0)u(x0,t0)−eGL(t−t0)∫t0te−GL(s−t0)Gf(u(x,s)) ds.(8)u(x,t)=e^{GL(t-t_0)}u(x_0,t_0)-e^{GL(t-t_0)}\int_{t_0}^{t}e^{-GL(s-t_0)}Gf(u(x,s))\,ds. \tag*{(8)}

The ETD methods are then developed by approximating the integral. A major advantage of the ETD methods is that the linear stiff part is exactly integrated, so that larger timesteps can be chosen after the matrix exponentials are computed using the Fourier transform.

The most widely used ETD methods are the ETD Runge–Kutta (ETDRK) methods, which use RK schemes to approximate the integral in (8). Much of the existing literature has studied the ETDRK methods for semi-linear parabolic equations and gradient flows [13, 39, 40, 12, 14, 41]. The discrete maximum principle for first- and second-order ETDRK methods (ETDRK1, ETDRK2) are proved for the second-order parabolic equations [40]. The dissipation of the original energy for the Ginzburg–Landau gradient flow was proved for ETDRK1, ETDRK2 [12]; later a general sufficient criterion for ETDRK of arbitrary order is later proposed, along with a novel energy dissipating ETDRK3 method [14]. The existence of energy dissipating ETDRK4 or higher order ETDRK methods is still an open problem.

Although implicit equation solvers are not needed for the ETDRK methods, the success of ETDRK methods for gradient flows largely relies on the effectiveness in computing the matrix exponentials under the periodic boundary condition within a regular domain. Since the Laplacian operator is diagonalizable in the Fourier domain, the computational cost is only O(N2log⁡N)O(N^2 \log N) if there are NN grids for each direction in two dimension [13]. However, if a non-periodic boundary condition is applied, the Fourier spectral method cannot be directly applied [42]. Moreover, a stabilization parameter is needed in practice in order for the method to behave well and preserve the maximum principle [12], making the method essentially implicit with the need for tuning the stabilization parameter.

Stabilized Runge–Kutta methods

Due to the stiff nature of gradient flows, the existing methods in Section 2 either use (semi-)implicit approaches, or utilize the exponential time differencing approach along with the Fourier transform and stabilization technique. Although linear and energy stability can be guaranteed with mild or no constraints, the (semi-)implicit methods require solving a linear or nonlinear system at each step, while the exponential time differencing approach requires computing matrix exponential, leading to heavy computational overhead.

In this section, we utilize a class of stabilized Runge–Kutta (SRK) methods for solving the Allen–Cahn and Cahn–Hilliard equations, more specifically the Runge–Kutta–Legendre (RKL) and Runge–Kutta–Gegenbauer (RKG) methods. The SRK methods are fully explicit methods, and do not rely on any computations in the Fourier domain or stabilization techniques. Although the SRK methods are only conditionally stable due to their explicit nature, the region of absolute stability of the SRK methods grow proportionally to s2s^2 where ss is the number of sub-stages, leading to mild stability constraints in practice. Rigorous analysis shows that along with Strang splitting, the RKL and RKG methods preserve the maximum principle under the linear stability constraints. Following similar arguments as in [43, 44], modified energy dissipation for two different ways of Strang splittings are proved for the Allen–Cahn equation, with new first-order methods and an estimation for the second-order RKG (RKG2) method. The modified energy is shown to converge to the exact energy as τ→0\tau\to0.

We note that the existing literature has briefly explored the use of super-time-stepping method for Cahn–Hilliard equation simulation [45]. However, they utilize Runge–Kutta–Chebyshev method and the DUMKA3 methods and only simulated the fourth-order Cahn–Hilliard and molecular beam epitaxy (MBE) equations; such methods have been numerically shown to violate the discrete maximum principle [46], so that they cannot be extended to the Allen–Cahn equation simulation. Moreover, our method is independent of spatial discretization, so that it can be applied to more generally shaped domains in addition to rectangular periodic domains, as will be shown in § 7. Finally, we provide an energy dissipation theoretical guarantee for the Allen–Cahn equation simulation with the SRK, for different ways of Strang splittings.

Stabilized Runge–Kutta methods overview

The Runge–Kutta–Legendre methods [46] and Runge–Kutta–Gegenbauer methods [47, 48] are super-time-stepping methods originally designed to solve for an ordinary differential equation (ODE) system

dudt=Mu,(9)\frac{du}{dt}=Mu, \tag*{(9)}

where MM is a symmetric, constant coefficient matrix that represents the discretization of a parabolic operator; the operator MM is required to have non-positive, real eigenvalues to ensure convergence. It is also shown [46, 47, 48] that the RKL and RKG methods are effective for variable coefficient problems, and later [49, 50] apply this class of method for broader applications.

We present the derivation of the RKL and RKG methods. The RKL and RKG methods are based on the shifted Legendre and Gegenbauer polynomials, respectively. The general form of the RKL and RKG methods is

ut+τ=Rs(τM)ut,(10)u^{t+\tau}=R_s(\tau M)u^t, \tag*{(10)}

where Rs(τM)R_s(\tau M) is the “stability polynomial" of the numerical method. Numerical stability is ensured if ∣Rs(τλ)∣≤1\lvert R_s(\tau\lambda)\rvert\le1 for all eigenvalues λ\lambda of MM; therefore, boundedness of Legendre polynomials and Gegenbauer polynomials enables them to be the stability polynomial choices [46, 48]. By matching the leading order terms in Taylor expansion as in [46, 48], the RKL1, RKL2, RKG1, and RKG2 methods, respectively, take the form of

ut+τ=Ps(I+2τs2+sM)ut;(11)u^{t+\tau}=P_s\left(I+\frac{2\tau}{s^2+s}M\right)u^t; \tag*{(11)}
ut+τ=[(1−s2+s−22s(s+1))I+s2+s−22s(s+1)Ps(I+4τs2+s−2M)]ut;(12)u^{t+\tau}=\left[\left(1-\frac{s^2+s-2}{2s(s+1)}\right)I+\frac{s^2+s-2}{2s(s+1)}P_s\left(I+\frac{4\tau}{s^2+s-2}M\right)\right]u^t; \tag*{(12)}
ut+τ=2(s+1)(s+2)Cs(λ)(I+4τs2+3sM)ut;(13)u^{t+\tau}=\frac{2}{(s+1)(s+2)}C_s^{(\lambda)}\left(I+\frac{4\tau}{s^2+3s}M\right)u^t; \tag*{(13)}
ut+τ=[(1−2(s−1)(s+4)3s(s+3))I+4(s−1)(s+4)3s(s+1)(s+2)(s+3)Cs(λ)(I+6τ(s+4)(s−1)M)]ut,(14)u^{t+\tau}=\left[\left(1-\frac{2(s-1)(s+4)}{3s(s+3)}\right)I+\frac{4(s-1)(s+4)}{3s(s+1)(s+2)(s+3)}C_s^{(\lambda)}\left(I+\frac{6\tau}{(s+4)(s-1)}M\right)\right]u^t, \tag*{(14)}

where Ps(x)P_s(x) is the ss-th degree Legendre polynomial and Cs(λ)(x)C_s^{(\lambda)}(x) is the ss-th degree Gegenbauer polynomial with parameter λ\lambda defined in the standard literature. In all of our experiments, λ\lambda is chosen to be 32\frac{3}{2} in order to be consistent with [48].

To obtain a Runge–Kutta formulation of the RKL and RKG methods, [46, 48] use the recurrence relationship of the Legendre and Gegenbauer polynomials. For example, the RKL2 method uses the Legendre polynomial recurrence relationship

sPs(x)=(2s−1)xPs−1(x)−(s−1)Ps−2(x).(15)sP_s(x)=(2s-1)xP_{s-1}(x)-(s-1)P_{s-2}(x). \tag*{(15)}

Together with the initial Legendre polynomials, after rearranging terms, one superstep of the ss-stage RKL2 method can be written as

Y0=u(t0),Y1=Y0+μ~1τMY0,Yj=μjYj−1+νjYj−2+(1−μj−νj)Y0+μ~jτMYj−1+γ~jτMY0,2≤j≤s,u(t0+τ)=Ys,(16)\begin{aligned} Y_0 &= u(t_0), \quad Y_1 = Y_0 + \widetilde{\mu}_1 \tau M Y_0,\\ Y_j &= \mu_j Y_{j-1} + \nu_j Y_{j-2} + (1-\mu_j-\nu_j)Y_0 + \widetilde{\mu}_j \tau M Y_{j-1} + \widetilde{\gamma}_j \tau M Y_0,\quad2 \leq j \leq s,\\ u(t_0+\tau) &= Y_s, \tag*{(16)} \end{aligned}

where the parameters are specified in [46]. Similar derivations work for the RKL1 method [46] and the RKG methods [48] using the recurrence relation between the Legendre and Gegenbauer polynomials of different degrees.

Although the formulas appear complicated, the method is entirely explicit and can be easily implemented with only three stages of the solution storage required. Moreover, the stability constraint is much more relaxed than a usual explicit method: denote τexpl:=2λmax⁡\tau_{\mathrm{expl}} := \frac{2}{\lambda_{\max}} where λmax⁡\lambda_{\max} is the maximum eigenvalue of the operator MM; as shown in [46], the largest possible timestep is τmax⁡=τexpls2+s−24\tau_{\max} = \tau_{\mathrm{expl}}\frac{s^2+s-2}{4}. The restriction is relaxed quadratically with respect to ss, so that when we choose ss large enough, the restriction on τ\tau related to dxdx is almost negligible. Moreover, it is simple to control the stage number ss in the implementation. These advantages lead to the adaptive use of ss in our complete algorithm.

Despite the fact that RKL and RKG methods were primarily designed to ensure the discrete maximum principle under different boundary conditions, the theoretical proof of their discrete maximum principle remained unsolved for a long time. Our recent work [49] fully proved the discrete maximum principle of RKL1, RKL2, RKG1, and RKG2 methods when applied to the heat equation, under the periodic or homogeneous Neumann boundary condition for rectangular domains. The discrete maximum principle of the Allen–Cahn equation therefore follows, as shown in § 6.1.

Stabilized Runge–Kutta methods applied to gradient flows

With the above description, we now apply the stabilized Runge–Kutta methods to the Allen–Cahn and Cahn–Hilliard equations. For the Allen–Cahn equation (1), since the nonlinear part is a double-well potential, we utilize the Strang splitting [51] technique so that the nonlinear part is solved exactly, while the numerical monotonicity can be guaranteed easily as will be shown in Theorem 4.1. By writing

SL(τ)(un)=SRK(un),SN(τ)(un):=un(1−(un)2)e−2τ+(un)2,S_L(\tau)(u^n)=S_{RK}(u^n),\qquad S_N(\tau)(u^n):=\frac{u^n}{\sqrt{\left(1-(u^n)^2\right)e^{-2\tau}+(u^n)^2}},

our Strang splitting scheme reads

un+1=SN(τ/2)SL(τ)SN(τ/2)un,u^{n+1}=S_N(\tau/2)S_L(\tau)S_N(\tau/2)u^n,

with the discrete Laplacian used being the standard central difference operator.

For the Cahn–Hilliard equation, since in this paper we in general do not want to use the Fourier spectral method for spatial discretization, we choose to not use Strang splitting as in [45] but instead discretize the spatial domain with standard central difference, and apply the SRK method with MuMu corresponding to the discretized right-hand side of the Cahn–Hilliard equation.

Maximum Principle and Modified Energy Dissipation

In this section, we prove the maximum principle and the energy dissipation when applying RKL and RKG methods to the Allen–Cahn equation via Strang splitting.

Maximum principle for Allen–Cahn via Strang splitting

In this section, we present the proof of the discrete maximum principle for the RKL and RKG methods when applied to the Allen–Cahn equation. More precisely, we want to show that if the initial condition is uniformly bounded by 1, then ∥u(t)∥∞≤1\lVert u(t)\rVert_\infty\leq1 for any further time t>0t>0. The discrete analogue is therefore described and proved in the following theorem, utilizing the power of Strang splitting [51] and the monotone stability results proved in [49].

Theorem 4.1. Consider the 1D Allen–Cahn equation with periodic or homogeneous Neumann boundary conditions

∂u∂t=ε2Lu−F′(u),\frac{\partial u}{\partial t}=\varepsilon^2Lu-F'(u),

where F(u)=14(u2−1)2F(u)=\frac{1}{4}(u^2-1)^2, and LL is the central difference approximation to the second-order spatial derivative. Consider the Strang splitting along with any SRK method when solving the above equation un+1=SL(τ/2)SN(τ)SL(τ/2)unu^{n+1}=S_L(\tau/2)S_N(\tau)S_L(\tau/2)u^n or un+1=SN(τ/2)SL(τ)SN(τ/2)unu^{n+1}=S_N(\tau/2)S_L(\tau)S_N(\tau/2)u^n. Assume ∥u0∥∞≤1\lVert u^0\rVert_\infty\le1. Then ∥un∥∞≤1\lVert u^n\rVert_\infty\le1 under the linear stability constraint of the SRK method for all n∈Nn\in\mathbb{N}.

Proof. Notice that for each spatial grid ii, 1−(uin)2≥01-(u_i^n)^2\ge0 by assumption, therefore

∣SN(τ)uin∣=∣uin(1−(uin)2)e−2τ+(uin)2∣≤∣uin∣(uin)2=1,\left|S_N(\tau)u_i^n\right|=\left|\frac{u_i^n}{\sqrt{(1-(u_i^n)^2)e^{-2\tau}+(u_i^n)^2}}\right|\le\frac{|u_i^n|}{\sqrt{(u_i^n)^2}}=1,

and therefore the nonlinear update SN(τ)S_N(\tau) preserves the maximum principle for any τ>0\tau>0. By Theorem 5.7 and Corollary 5.12 in [49], we know that the linear update SL(τ)S_L(\tau) also preserves the maximum principle, therefore the maximum principle of the Strang splitting follows. □

Remark 4.2. The theorem only works for the most commonly used periodic and homogeneous Neumann boundary conditions. It is known that the RKL methods are not monotonically stable for the homogeneous Dirichlet boundary condition, and therefore cannot be guaranteed to preserve the maximum principle of the Allen–Cahn equation. For the RKG methods, it is conjectured in [48] that they are monotonically stable under the homogeneous Dirichlet boundary condition, while no theoretical proof has been established yet.

Remark 4.3. The above theorem only works for 1D because of the limitation of the theorem in [49]. Nevertheless, numerical evidence suggests that the maximum principle also holds a for 2D regular domain with periodic boundary condition and irregular domains under the Neumann boundary condition. However, direct computations reveal that RKL2 does not always preserve numerical monotonicity in 2D even with the CFL condition.s

Modified energy dissipation for the Allen–Cahn equation

We now prove the modified energy dissipation of the Allen–Cahn equation. For the linear-nonlinear-linear Strang splitting, we are inspired by the work [43, 44], with an inexact solution of the linear part. Since the computational cost of the linear part often dominates the cost of the nonlinear part, we propose a different modified energy for the more practical splitting un+1=SN(τ/2)SL(τ)SN(τ/2)unu^{n+1}=S_N(\tau/2)S_L(\tau)S_N(\tau/2)u^n and prove that it dissipates and converges to the exact energy. New first-order RKL and RKG methods are proposed in order to fit into the proof framework, along with an estimation of the RKG2 method spectrum. Moreover, throughout this section we only require the discrete finite difference Laplacian operator LL to be real, symmetric, and negative-definite; the domain is not required to be regular.

Linear-Nonlinear-Linear Strang splitting

We begin by considering the linear-nonlinear-linear splitting un+1=SL(τ/2)SN(τ)SL(τ/2)unu^{n+1}=S_L(\tau/2)S_N(\tau)S_L(\tau/2)u^n.

Theorem 4.4 (Modified Energy Dissipation, linear-nonlinear-linear). Define the modified energy as

E~(un):=12τ⟨(I−SL2(τ/2))un,un⟩+⟨g(SL(τ/2)un),1⟩,(17)\widetilde{E}(u^n):=\frac{1}{2\tau}\left\langle\left(I-S_L^2(\tau/2)\right)u^n,u^n\right\rangle+\left\langle g(S_L(\tau/2)u^n),1\right\rangle, \tag*{(17)}

where

g(x):=14+12τx2−eττ(e2τ−1)(1+(e2τ−1)x2−1).(18)g(x):=\frac{1}{4}+\frac{1}{2\tau}x^2-\frac{e^{\tau}}{\tau(e^{2\tau}-1)}\left(\sqrt{1+(e^{2\tau}-1)x^2}-1\right). \tag*{(18)}

Then under the linear-nonlinear-linear Strang splitting defined above, we have

  1. E~(un+1)≤E~(un)\widetilde{E}(u^{n+1})\le\widetilde{E}(u^n);

  1. For fixed grid size hh, E~(un)\widetilde{E}(u^n) converges to E(un)E(u^n) uniformly as τ→0\tau\to0; more precisely, for 0≤τ≤min⁡{1,CFL constraint}0 \le\tau\le\min\{1,\mathrm{CFL\ constraint}\}, assuming enough regularity of u0u^0, we have that

sup⁡n≥0∣E~(un)−E(un)∣≲τ.\sup_{n\ge0}\left|\widetilde{E}(u^n)-E(u^n)\right|\lesssim\tau.

Proof. We will first prove that E~(u~n+1)≤E~(u~n)\widetilde{E}(\widetilde{u}^{n+1})\le\widetilde{E}(\widetilde{u}^n). By construction, the spectrum of SL(τ/2)S_L(\tau/2) is within [−1,1][-1,1]; therefore, we have that I−SL2(τ/2)I-S_L^2(\tau/2) is positive semi-definite. The non-negativity of the energy and its dissipation result then directly follows from the proof of Theorem 2.1 in [44]. To show the uniform convergence, we follow the procedure as in [43]. Define u~n:=SL(τ/2)un\widetilde{u}^n:=S_L(\tau/2)u^n; taking the difference between u~n\widetilde{u}^n and unu^n and using Taylor expansion,

∥u~n−un∥=∥SL(τ/2)un−un∥=∥(I+τ2L+O(τ2))un−un∥≲O(τ).\left\|\widetilde{u}^n-u^n\right\|=\left\|S_L(\tau/2)u^n-u^n\right\|=\left\|\left(I+\frac{\tau}{2}L+O(\tau^2)\right)u^n-u^n\right\|\lesssim O(\tau).

Notice that ∥u~n∥∞≤∥SL(τ/2)∥∞∥un∥∞≤C∥SL(τ/2)∥2≤C\|\widetilde{u}^n\|_\infty\le\|S_L(\tau/2)\|_\infty\|u^n\|_\infty\le C\|S_L(\tau/2)\|_2\le C for some grid size dependent constant CC uniformly when the CFL condition is satisfied. By direct calculation, we have that

g(u~n)=14−12(u~n)2+14(u~n)4+O(τ),g(\widetilde{u}^n)=\frac{1}{4}-\frac{1}{2}(\widetilde{u}^n)^2+\frac{1}{4}(\widetilde{u}^n)^4+O(\tau),

we can bound the nonlinear part of the energy using the fact that

∣∫TdF~(u~)dx−∫Td14((un)2−1)2dx∣≲τ.\left|\int_{\mathbb{T}^d}\widetilde{F}(\widetilde{u})dx-\int_{\mathbb{T}^d}\frac{1}{4}\left((u^n)^2-1\right)^2dx\right|\lesssim\tau.

For the linear part, we can bound the difference by

∣12τ⟨(I−SL2(τ/2))un,un⟩−12⟨ε2(−Δ)un,un⟩∣≤∣12τ⟨(I−SL2(τ/2))un,un⟩−12τ⟨(I−eε2τL)un,un⟩∣+∣12τ⟨(I−eε2τL)un,un⟩−12⟨ε2(−L)un,un⟩∣+∣12⟨ε2(−L)un,un⟩−12⟨ε2(−Δ)un,un⟩∣≲O(τ).\begin{aligned} \left|\frac{1}{2\tau}\left\langle\left(I-S_L^2(\tau/2)\right)u^n,u^n\right\rangle-\frac{1}{2}\left\langle\varepsilon^2(-\Delta)u^n,u^n\right\rangle\right| &\le\left|\frac{1}{2\tau}\left\langle\left(I-S_L^2(\tau/2)\right)u^n,u^n\right\rangle-\frac{1}{2\tau}\left\langle\left(I-e^{\varepsilon^2\tau L}\right)u^n,u^n\right\rangle\right|\\ &\quad+\left|\frac{1}{2\tau}\left\langle\left(I-e^{\varepsilon^2\tau L}\right)u^n,u^n\right\rangle-\frac{1}{2}\left\langle\varepsilon^2(-L)u^n,u^n\right\rangle\right|\\ &\quad+\left|\frac{1}{2}\left\langle\varepsilon^2(-L)u^n,u^n\right\rangle-\frac{1}{2}\left\langle\varepsilon^2(-\Delta)u^n,u^n\right\rangle\right|\\ &\lesssim O(\tau). \end{aligned}

This completes our proof.

Nonlinear-Linear-Nonlinear Strang splitting

We now consider the more practical nonlinear-linear-nonlinear splitting un+1=SN(τ/2)SL(τ)SN(τ/2)unu^{n+1}=S_N(\tau/2)S_L(\tau)S_N(\tau/2)u^n. If the energy in the previous section is directly used, the proof fails. Therefore, a new non-negative energy is needed, which should ideally converge to the exact Ginzburg–Landau energy. Following the logical arguments in [43] we define a suitable u~\widetilde{u} as follows in the proof below.

Theorem 4.5 (Modified Energy Dissipation, nonlinear-linear-nonlinear). For any symmetric positive-definite SL(τ)S_L(\tau) with spectrum in (0,1](0,1], define the modified energy as

E~(un):=12τ⟨(I−SL(τ))SN(τ/2)un,SL(τ)SN(τ/2)un⟩+⟨g(SL(τ)SN(τ/2)un),1⟩,(19)\widetilde{E}(u^n):=\frac{1}{2\tau}\left\langle\left(I-S_L(\tau)\right)S_N(\tau/2)u^n,S_L(\tau)S_N(\tau/2)u^n\right\rangle+\left\langle g\left(S_L(\tau)S_N(\tau/2)u^n\right),1\right\rangle, \tag*{(19)}

where the function gg is defined as in Equation (18). Then under the nonlinear-linear-nonlinear Strang splitting defined above, we have

  1. E~(un+1)≤E~(un)\widetilde{E}(u^{n+1})\le\widetilde{E}(u^n),

  2. For fixed grid size hh, E~(un)\widetilde{E}(u^n) converges to E(un)E(u^n) uniformly as τ→0\tau\to0; more precisely, for 0≤τ≤min⁡{1,CFL constraint}0\le\tau\le\min\{1,\mathrm{CFL\ constraint}\},

assuming enough regularity of u0u^0 we have that

sup⁡n≥0∣E~(u~n)−E(un)∣≲τ.\sup_{n\geq0}|\widetilde{E}(\widetilde{u}^n)-E(u^n)|\lesssim\tau.

Remark 4.6. As will be shown in § 6.2.3 and § 6.2.4, the assumptions can be satisfied by modified versions of RKL1 and RKG1. It will also be naturally satisfied by RKL2, but requires some more careful estimations and restrictions for RKG2.

Proof. The non-negativity of E~(un)\widetilde{E}(u^n) can be easily deduced from the non-negativity of gg and the fact that (I−SL(τ))SL(τ)(I-S_L(\tau))S_L(\tau) is positive semi-definite. To show that the modified energy dissipates, set u~n:=SL(τ)SN(τ2)un\widetilde{u}^n:=S_L(\tau)S_N\left(\frac{\tau}{2}\right)u^n. We know that

u~n+1=SL(τ)SN(τ)u~n,\widetilde{u}^{n+1}=S_L(\tau)S_N(\tau)\widetilde{u}^n,

so that

SL(τ)−1u~n+1=eτu~n1+(e2τ−1)(u~n)2.S_L(\tau)^{-1}\widetilde{u}^{n+1}=\frac{e^{\tau}\widetilde{u}^n}{\sqrt{1+(e^{2\tau}-1)(\widetilde{u}^n)^2}}.

Rewriting the above and using direct computation, we find

1τ(SL(τ)−1−I)u~n+1+1τ(u~n+1−u~n)=−g′(u~n).\frac{1}{\tau}(S_L(\tau)^{-1}-I)\widetilde{u}^{n+1}+\frac{1}{\tau}(\widetilde{u}^{n+1}-\widetilde{u}^n)=-g'(\widetilde{u}^n).

Also by direct computation and self-adjointness of SL(τ)S_L(\tau) and the fact that all the eigenvalues of SL(τ)S_L(\tau) are non-negative, we have that

1τ⟨(SL(τ)−1−I)u~n+1,u~n+1−u~n⟩=1τ⟨(I−SL(τ))SN(τ2)un+1,SL(τ)SN(τ2)(un+1−un)⟩=1τ⟨(I−SL(τ))SL(τ)SN(τ2)un+1,SL(τ)SN(τ2)(un+1−un)⟩=12τ(⟨(I−SL(τ))SL(τ)SN(τ2)un+1,SL(τ)SN(τ2)un+1⟩−⟨(I−SL(τ))SL(τ)SN(τ2)un,SL(τ)SN(τ2)un⟩+⟨(I−SL(τ))SL(τ)SN(τ2)(un+1−un),SL(τ)SN(τ2)(un+1−un)⟩)=12τ(⟨(SL(τ)−1−I)u~n+1,u~n+1⟩−⟨(SL(τ)−1−I)u~n,u~n⟩+⟨(I−SL(τ))SL(τ)SN(τ2)(un+1−un),SL(τ)SN(τ2)(un+1−un)⟩).\begin{aligned} \frac{1}{\tau}\langle(S_L(\tau)^{-1}-I)\widetilde{u}^{n+1},\widetilde{u}^{n+1}-\widetilde{u}^n\rangle &=\frac{1}{\tau}\langle(I-S_L(\tau))S_N\left(\frac{\tau}{2}\right)u^{n+1},S_L(\tau)S_N\left(\frac{\tau}{2}\right)(u^{n+1}-u^n)\rangle\\ &=\frac{1}{\tau}\langle(I-S_L(\tau))\sqrt{S_L(\tau)}S_N\left(\frac{\tau}{2}\right)u^{n+1},\sqrt{S_L(\tau)}S_N\left(\frac{\tau}{2}\right)(u^{n+1}-u^n)\rangle\\ &=\frac{1}{2\tau}\bigl(\langle(I-S_L(\tau))\sqrt{S_L(\tau)}S_N\left(\frac{\tau}{2}\right)u^{n+1},\sqrt{S_L(\tau)}S_N\left(\frac{\tau}{2}\right)u^{n+1}\rangle\\ &\qquad-\langle(I-S_L(\tau))\sqrt{S_L(\tau)}S_N\left(\frac{\tau}{2}\right)u^n,\sqrt{S_L(\tau)}S_N\left(\frac{\tau}{2}\right)u^n\rangle\\ &\qquad+\langle(I-S_L(\tau))\sqrt{S_L(\tau)}S_N\left(\frac{\tau}{2}\right)(u^{n+1}-u^n),\sqrt{S_L(\tau)}S_N\left(\frac{\tau}{2}\right)(u^{n+1}-u^n)\rangle\bigr)\\ &=\frac{1}{2\tau}\bigl(\langle(S_L(\tau)^{-1}-I)\widetilde{u}^{n+1},\widetilde{u}^{n+1}\rangle-\langle(S_L(\tau)^{-1}-I)\widetilde{u}^n,\widetilde{u}^n\rangle\\ &\qquad+\langle(I-S_L(\tau))\sqrt{S_L(\tau)}S_N\left(\frac{\tau}{2}\right)(u^{n+1}-u^n),\sqrt{S_L(\tau)}S_N\left(\frac{\tau}{2}\right)(u^{n+1}-u^n)\rangle\bigr). \end{aligned}

Multiplying the equation and summing over the grid, and using Taylor expansion of gg up to second-order, we have that

E~(u~n+1)−E~(u~n)=−12τ⟨(I−SL(τ))SL(τ)SN(τ2)(un+1−un),SL(τ)SN(τ2)(un+1−un)⟩−⟨(1τ−12g′′(ξn))(u~n+1−u~n)2,1⟩.\begin{aligned} \widetilde{E}(\widetilde{u}^{n+1})-\widetilde{E}(\widetilde{u}^n) &=-\frac{1}{2\tau}\langle(I-S_L(\tau))\sqrt{S_L(\tau)}S_N\left(\frac{\tau}{2}\right)(u^{n+1}-u^n),\sqrt{S_L(\tau)}S_N\left(\frac{\tau}{2}\right)(u^{n+1}-u^n)\rangle\\ &\qquad-\langle\left(\frac{1}{\tau}-\frac{1}{2}g''(\xi^n)\right)(\widetilde{u}^{n+1}-\widetilde{u}^n)^2,1\rangle. \end{aligned}

By noticing that I−SL(τ)I-S_L(\tau) is symmetric positive-definite and noticing that 1τ−12g′′(ξn)≥0\frac{1}{\tau}-\frac{1}{2}g''(\xi^n)\geq0, we can conclude that E~(u~n+1)≤E~(u~n)\widetilde{E}(\widetilde{u}^{n+1})\leq\widetilde{E}(\widetilde{u}^n) as desired.

Now we show that the modified energy converges to the exact energy. We will first show that u~n\widetilde{u}^n and unu^n differ by O(τ)O(\tau). By definition we have that u~n:=SL(τ)SN(τ2)un\widetilde{u}^n:=S_L(\tau)S_N\left(\frac{\tau}{2}\right)u^n; taking the difference and using Taylor expansion, and by noticing that the norm of discrete Laplacian ∥L∥∞\lVert L\rVert_{\infty} is bounded for fixed grid size, we have that

∥u~n−un∥∞=∥SL(τ)SN(τ2)un−un∥∞=∥(I+τL+O(τ2))(eτ/2+O(τ))un−un∥∞=∥(I+τL+O(τ2))(I+O(τ))un−un∥∞≲O(τ).\begin{aligned} \lVert\tilde{u}^{n}-u^{n}\rVert_{\infty} &=\lVert S_L(\tau)S_N\left(\frac{\tau}{2}\right)u^n-u^n\rVert_{\infty}\\ &=\lVert(I+\tau L+O(\tau^2))(e^{\tau/2}+O(\tau))u^n-u^n\rVert_{\infty}\\ &=\lVert(I+\tau L+O(\tau^2))(I+O(\tau))u^n-u^n\rVert_{\infty}\\ &\lesssim O(\tau). \end{aligned}

Now notice that ∥u~n∥∞≤∥SL(τ/2)∥∞∥SN(τ/2)un∥∞≤C∥SL(τ/2)∥2≤C\lVert\tilde{u}^{n}\rVert_{\infty}\leq\lVert S_L(\tau/2)\rVert_{\infty}\lVert S_N(\tau/2)u^n\rVert_{\infty}\leq C\lVert S_L(\tau/2)\rVert_2\leq C for some grid size dependent constant CC uniformly when the CFL condition is satisfied. Then by the same argument as in equations (2.26) and (2.27) of [43], we are done with the nonlinear part convergence.

For the inner product part of the energy, by direct calculation we know that

∣12τ⟨(I−SL(τ))SL(τ)SN(τ2)un,SL(τ)SN(τ2)un⟩−12⟨ε2Lun,un⟩∣≤∣12τ⟨(I−SL(τ))SL(τ)SN(τ2)un,SL(τ)SN(τ2)un⟩−12⟨ε2(−L)SL(τ)SN(τ2)un,SL(τ)SN(τ2)un⟩∣+∣12⟨ε2(−L)SL(τ)SN(τ2)un,SL(τ)SN(τ2)un⟩−12⟨ε2(−L)un,un⟩∣+∣12⟨ε2(−L)un,un⟩−12⟨ε2(−Δ)un,un⟩∣≲O(τ).\begin{aligned} &\left|\frac{1}{2\tau}\left\langle(I-S_L(\tau))\sqrt{S_L(\tau)}S_N\left(\frac{\tau}{2}\right)u^n,\sqrt{S_L(\tau)}S_N\left(\frac{\tau}{2}\right)u^n\right\rangle-\frac{1}{2}\left\langle\varepsilon^2Lu^n,u^n\right\rangle\right|\\ &\leq\left|\frac{1}{2\tau}\left\langle(I-S_L(\tau))\sqrt{S_L(\tau)}S_N\left(\frac{\tau}{2}\right)u^n,\sqrt{S_L(\tau)}S_N\left(\frac{\tau}{2}\right)u^n\right\rangle-\frac{1}{2}\left\langle\varepsilon^2(-L)\sqrt{S_L(\tau)}S_N\left(\frac{\tau}{2}\right)u^n,\sqrt{S_L(\tau)}S_N\left(\frac{\tau}{2}\right)u^n\right\rangle\right|\\ &\quad+\left|\frac{1}{2}\left\langle\varepsilon^2(-L)\sqrt{S_L(\tau)}S_N\left(\frac{\tau}{2}\right)u^n,\sqrt{S_L(\tau)}S_N\left(\frac{\tau}{2}\right)u^n\right\rangle-\frac{1}{2}\left\langle\varepsilon^2(-L)u^n,u^n\right\rangle\right|\\ &\quad+\left|\frac{1}{2}\left\langle\varepsilon^2(-L)u^n,u^n\right\rangle-\frac{1}{2}\left\langle\varepsilon^2(-\Delta)u^n,u^n\right\rangle\right|\lesssim O(\tau). \end{aligned}

This completes our proof.

Remark 4.7. The linear-nonlinear-linear energy dissipation utilizes the power of Strang splitting instead of the power of super-time-stepping methods. In fact, any linear stable approximation to the heat equation can be fitted into the proof outline above. The nonlinear-linear-nonlinear Strang splitting energy dissipation, on the other hand, does rely on the fact that the stability polynomial of the second-order SRK methods are (almost) positive-definite.

It is easy to verify all the conditions required in the theorem above, except for the positive-definiteness of the SL(τ)S_L(\tau) operators for the RKL and RKG methods. We therefore proceed and check the condition in the following two sections, and show that the RKL and RKG methods are indeed able to be fitted into the proof above up to modifications and minor additional timestep constraint.

Positive-definite modified RKL1 and RKG1 methods

Since the nonlinear-linear-nonlinear energy definition and the proof rely on the fact that SL(τ)S_L(\tau) has spectrum within (0,1](0,1], we now determine the condition that such an assumption holds. For the RKL1 and RKG1 methods

ut+τ=Ps(I+2τs2+sM)ut,u^{t+\tau}=P_s\left(I+\frac{2\tau}{s^2+s}M\right)u^t,
ut+τ=2(s+1)(s+2)Cs(2)(I+4τs2+3sM)ut.u^{t+\tau}=\frac{2}{(s+1)(s+2)}C_s^{(2)}\left(I+\frac{4\tau}{s^2+3s}M\right)u^t.

According to the Legendre and Gegenbauer polynomial properties, ∣SL(τ)∣≤1\lvert S_L(\tau)\rvert\leq1 is the best possible estimation. Therefore, modifications are needed in order for the first-order methods to have a non-negative spectrum. We then follow the general derivation procedure as in § 5.1 to derive new modified first-order RKL and RKG methods.

For the modified RKL1 method, instead of using the Legendre polynomial, consider the shifted Legendre polynomial P~n(x)=1+Pn(x)2\widetilde{P}_n(x)=\frac{1+P_n(x)}{2}, where Pn(x)P_n(x) is the nn-th order Legendre polynomial. Since the Legendre polynomials satisfy Pn(x)∈[−1,1]P_n(x)\in[-1,1] for x∈[−1,1]x\in[-1,1] and ∀n∈N\forall n\in\mathbb{N}, we have that P~n(x)∈[0,1]\widetilde{P}_n(x)\in[0,1] for x∈[−1,1]x\in[-1,1] and ∀n∈N\forall n\in\mathbb{N}. The three-point recurrence relationship is then

jP~j(x)=(2j−1)xP~j−1(x)−(j−1)P~j−2(x)+2j−12(1−x),j≥2.j\widetilde{P}_j(x)=(2j-1)x\widetilde{P}_{j-1}(x)-(j-1)\widetilde{P}_{j-2}(x)+\frac{2j-1}{2}(1-x),\qquad j\geq2.

Now assume that the stability polynomial takes the form of

Rs(z)=as+bsP~n(w0+w1z).R_s(z)=a_s+b_s\widetilde{P}_n(w_0+w_1z).

For first-order, we can assume that as=0a_s=0 for all s∈Ns\in\mathbb{N} and choose w0=1w_0=1 as in [46], so that

Rs(z)=P~s(1+4s2+sz).R_s(z)=\widetilde{P}_s\left(1+\frac{4}{s^2+s}z\right).

Plugging the stability polynomial into the three-point recurrence relationship, we obtain that

P~j(1+4s2+sz)=2j−1j(1+4s2+sz)P~j−1(1+4s2+sz)−j−1jP~j−2(x)+2j−12j(−4s2+sz).\widetilde{P}_j\left(1+\frac{4}{s^2+s}z\right)=\frac{2j-1}{j}\left(1+\frac{4}{s^2+s}z\right)\widetilde{P}_{j-1}\left(1+\frac{4}{s^2+s}z\right)-\frac{j-1}{j}\widetilde{P}_{j-2}(x)+\frac{2j-1}{2j}\left(-\frac{4}{s^2+s}z\right).

The modified RKL1 method is then

Y0=u(t0),Y1=Y0+μ~1τMY0,Yj=μjYj−1+νjYj−2+μ~jτMYj−1+γ~jτMY0,u(t0+τ)=Ys.\begin{aligned} Y_0&=u(t_0),\\ Y_1&=Y_0+\widetilde{\mu}_1\tau MY_0,\\ Y_j&=\mu_jY_{j-1}+\nu_jY_{j-2}+\widetilde{\mu}_j\tau MY_{j-1}+\widetilde{\gamma}_j\tau MY_0,\\ u(t_0+\tau)&=Y_s. \end{aligned}

where

μj=2j−1j,νj=1−jj,μ~j=2j−1j4s2+s,γ~j=−4j−2js(s+1).\mu_j=\frac{2j-1}{j},\quad\nu_j=\frac{1-j}{j},\quad\widetilde{\mu}_j=\frac{2j-1}{j}\frac{4}{s^2+s},\quad\widetilde{\gamma}_j=-\frac{4j-2}{js(s+1)}.

Similarly, we can derive the modified RKG1 method by using the shifted Gegenbauer polynomial

C~n(λ)(x)=12(1+1(n+1)(n+2)Cn(λ)(x)).\widetilde{C}^{(\lambda)}_n(x)=\frac{1}{2}\left(1+\frac{1}{(n+1)(n+2)}C^{(\lambda)}_n(x)\right).

By using exactly the same procedure as above, the RKG1 method can be written as

Y0=u(t0),Y1=Y0+μ~1τMY0,Yj=μjYj−1+νjYj−2+μ~jτMYj−1+γ~jτMY0,u(t0+τ)=Ys.\begin{aligned} Y_0&=u(t_0),\\ Y_1&=Y_0+\widetilde{\mu}_1\tau MY_0,\\ Y_j&=\mu_jY_{j-1}+\nu_jY_{j-2}+\widetilde{\mu}_j\tau MY_{j-1}+\widetilde{\gamma}_j\tau MY_0,\\ u(t_0+\tau)&=Y_s. \end{aligned}

where

μj=2j+1j+2,νj=−j−1j+2,μ~j=2j+1j+28s2+3s,γ~j=−2j+12(j+2)8s2+3s,μ~1=4s2+3s.\mu_j=\frac{2j+1}{j+2},\quad\nu_j=-\frac{j-1}{j+2},\quad\widetilde{\mu}_j=\frac{2j+1}{j+2}\frac{8}{s^2+3s},\quad\widetilde{\gamma}_j=-\frac{2j+1}{2(j+2)}\frac{8}{s^2+3s},\quad\widetilde{\mu}_1=\frac{4}{s^2+3s}.

The advantage of these modified RKL1 and RKG1 is that they control the eigenvalues of the stability polynomial to be all non-negative, so that they can be directly fitted into the proof of Theorem 4.5. Moreover, those methods naturally satisfy the discrete maximum principle since the regular RKL1 and RKG1 methods are, under the corresponding boundary condition requirements of the RKL and RKG methods. The drawback of this modified RKL1 is that the maximum possible τ\tau is only half of the original RKL1 by following the same derivation as in [46], although it still follows the O(s2)O(s^2) order.

Positive-definiteness estimation for RKL2 and RKG2 methods

We now turn our attention to the second-order RKL and RKG. According to the derivations in [46] and [48], it is not difficult to see that the shifted RKL2 and RKG2 are the same as the usual RKL2 and RKG2, since there are exactly three equations and three unknowns. Therefore, we need to estimate the conditions for RKL2 and RKG2 to give us positive (semi)-definite stability polynomials.

Recall that the stability polynomial for RKL2 is

Rs(τM)=(1−s2+s−22s(s+1))I+s2+s−22s(s+1)Ps(I+4τs2+s−2M).R_s(\tau M)=\left(1-\frac{s^2+s-2}{2s(s+1)}\right)I+\frac{s^2+s-2}{2s(s+1)}P_s\left(I+\frac{4\tau}{s^2+s-2}M\right).

Notice that 0≤s2+s−22s(s+1)<120\leq\frac{s^2+s-2}{2s(s+1)}<\frac{1}{2} for s∈Z+s\in\mathbb{Z}_{+}; since the Legendre polynomial is bounded by 1 when the argument is within [−1,1][-1,1], we know that 0<Rs(τM)≤10<R_s(\tau M)\leq1 when τ\tau is within the stability constraint. This shows that RKL2 automatically gives us a positive-definite stability polynomial.

The RKG2 method does not automatically have a positive-definite stability polynomial. Recall that the RKG2 stability polynomial is

(1−2(s−1)(s+4)3s(s+3))I+4(s−1)(s+4)3s(s+1)(s+2)(s+3)Cs(λ)(I+6τ(s+4)(s−1)M).\left(1-\frac{2(s-1)(s+4)}{3s(s+3)}\right)I+\frac{4(s-1)(s+4)}{3s(s+1)(s+2)(s+3)}C_s^{(\lambda)}\left(I+\frac{6\tau}{(s+4)(s-1)}M\right).

Notice that 0≤2(s−1)(s+4)3s(s+3)<230\leq\frac{2(s-1)(s+4)}{3s(s+3)}<\frac{2}{3} for s∈Z+s\in\mathbb{Z}_{+}; since the Gegenbauer polynomial of degree ss and λ=32\lambda=\frac{3}{2} has value within [−(s+1)(s+2)2,(s+1)(s+2)2]\left[-\frac{(s+1)(s+2)}{2},\frac{(s+1)(s+2)}{2}\right], the spectrum of the RKG2 stability polynomial takes values within (−13,1)\left(-\frac{1}{3},1\right). Since the most negative values of the Gegenbauer polynomials correspond to argument near −1-1 with ss odd, we need to estimate the domain of the Gegenbauer polynomial so that the argument within the polynomial is away from −1-1. With more asymptotic analysis and the control of the positions for the zeros of the Gegenbauer polynomial, we manage to show that the characteristic polynomial of RKG2 is “almost positive-definite” in the sense that with a small sacrifice of the discrete timestep size away from the maximum possible discrete timestep size, the RKG2 operator is positive-definite. More precisely, it is shown in the theorem below.

Theorem 4.8 (RKG2 operator is “almost positive-definite”). Consider RKG2 in Equation (14) with parameter λ=32\lambda=\frac{3}{2}, and let τmax⁡\tau_{\max} denote the largest timestep satisfying the CFL constraint. Then for any 0<ε≪10<\varepsilon\ll1, there exists s0∈Ns_0\in\mathbb{N} such that for all 0<τ≤(1−ε)τmax⁡0<\tau\leq(1-\varepsilon)\tau_{\max} and s≥s0s\geq s_0, SL(τ)S_L(\tau) is positive-definite.

Proof. First recall that

Cs(λ)(x)=Γ(λ+12)Γ(2λ)Γ(s+2λ)Γ(s+λ+12)Ps(λ−12,λ−12)(x)C_s^{(\lambda)}(x)=\frac{\Gamma(\lambda+\frac{1}{2})}{\Gamma(2\lambda)}\frac{\Gamma(s+2\lambda)}{\Gamma(s+\lambda+\frac{1}{2})}P_s^{(\lambda-\frac{1}{2},\lambda-\frac{1}{2})}(x)

where Γ(x)\Gamma(x) is the standard gamma function [52]. Now arbitrarily fix δ>0\delta>0. Now fix c>0c>0 small enough such that cos⁡cs>xs\cos\frac{c}{s}>x_s where xsx_s is the right-most root of Cs(λ)(x)=Cs(3/2)(x)C_s^{(\lambda)}(x)=C_s^{(3/2)}(x). Notice that this is attainable via Theorem 4.1 of [53]: for ϑs,k(1,1)\vartheta_{s,k}(1,1) such that cos⁡(ϑs,k(1,1))\cos(\vartheta_{s,k}(1,1)) are the zeros of the Gegenbauer polynomial Ps(1,1)(x)=Cs(3/2)(x)P_s^{(1,1)}(x)=C_s^{(3/2)}(x), we have that specifically

ϑs,1(1,1)=j1,1ν+O(s−5),\vartheta_{s,1}(1,1)=\frac{j_{1,1}}{\nu}+O(s^{-5}),

where j1,1j_{1,1} denotes the left-most zero of J1(x)J_1(x) for x∈[0,∞)x\in[0,\infty), and ν=((s+32)2−14)1/2∼O(s)\nu=\left(\left(s+\frac{3}{2}\right)^2-\frac{1}{4}\right)^{1/2}\sim O(s); since cos⁡(ϑs,1(1,1))\cos(\vartheta_{s,1}(1,1)) corresponds to the right-most zero of Cs(3/2)(x)C_s^{(3/2)}(x), we can choose 0<c≲j1,10<c\lesssim j_{1,1} small enough to achieve the desired goal.

Notice that the positive-definiteness of the RKG2 operator, with the timestep restriction in the theorem statement, is equivalent to showing that

C~s(λ)(x):=Cs(λ)(x)Cs(λ)(1)>−12(20)\widetilde{C}_s^{(\lambda)}(x):=\frac{C_s^{(\lambda)}(x)}{C_s^{(\lambda)}(1)}>-\frac{1}{2} \tag*{(20)}

for x∈[−1,1]x\in[-1,1] that is not close to −1-1, and as s→∞s\to\infty. This is because the eigenvalues lie in the interval of the form

1−2(s−1)(s+4)3s(s+3)+4(s−1)(s+4)3s(s+1)(s+2)(s+3)Cs(λ)(x)=1−2(s−1)(s+4)3s(s+3)+2(s−1)(s+4)3s(s+3)C~s(λ)(x),1-\frac{2(s-1)(s+4)}{3s(s+3)}+\frac{4(s-1)(s+4)}{3s(s+1)(s+2)(s+3)}C_s^{(\lambda)}(x) =1-\frac{2(s-1)(s+4)}{3s(s+3)}+\frac{2(s-1)(s+4)}{3s(s+3)}\widetilde{C}_s^{(\lambda)}(x),

along with the fact that 2(s−1)(s+4)3s(s+3)⟶23\frac{2(s-1)(s+4)}{3s(s+3)}\longrightarrow\frac{2}{3}. Therefore, we will show that eq. (20) holds; more specifically, we will show that after the standard change of variable x:=cos⁡θx:=\cos\theta, it holds for 0<θ<π−δ0<\theta<\pi-\delta in order to complete our proof.

To conduct the estimation, we dissect the θ\theta region [0,π−δ][0,\pi-\delta] into three different regimes. We first choose d>0d>0 to be specified later.

Regime 1: 0≤θ<cs0 \le\theta< \frac{c}{s}. For the region for 0<θ<cs0 < \theta< \frac{c}{s}, by the way we construct cc, naturally Cs(λ)(x)≥0C_s^{(\lambda)}(x) \ge0 so that C~s(λ)(x)≥0\widetilde{C}_s^{(\lambda)}(x) \ge0 since Cs(λ)(1)>0C_s^{(\lambda)}(1) > 0 always holds.

Regime 2: cs≤θ≤ds\frac{c}{s} \le\theta\le\frac{d}{s}. In this regime, consider the Mehler–Heine formula [52]

lim⁡s→∞s−αPs(α,β)(cos⁡(zs))=(z2)−αJα(z).\lim_{s \to\infty} s^{-\alpha}P_s^{(\alpha,\beta)}\left(\cos\left(\frac{z}{s}\right)\right)=\left(\frac{z}{2}\right)^{-\alpha}J_\alpha(z).

Notice that

C~s(3/2)(x)=Cs(3/2)(x)Cs(3/2)(1)=s+22Ps(1,1)(x)2(s+1)(s+2)=1s+1Ps(1,1)(x).\widetilde{C}_s^{(3/2)}(x)=\frac{C_s^{(3/2)}(x)}{C_s^{(3/2)}(1)}=\frac{s+2}{2}P_s^{(1,1)}(x)\frac{2}{(s+1)(s+2)}=\frac{1}{s+1}P_s^{(1,1)}(x).

Therefore, we have that

lim⁡s→∞C~s(3/2)(cos⁡(zs))=2J1(z)z.\lim_{s \to\infty}\widetilde{C}_s^{(3/2)}\left(\cos\left(\frac{z}{s}\right)\right)=\frac{2J_1(z)}{z}.

By elementary estimation of J1(z)z\frac{J_1(z)}{z}, we have that for all sufficiently large s>0s>0, C~s(3/2)(cos⁡(zs))>−12\widetilde{C}_s^{(3/2)}\left(\cos\left(\frac{z}{s}\right)\right)>-\frac{1}{2} for all z∈[c,d]z\in[c,d], or equivalently θ∈[cs,ds]\theta\in\left[\frac{c}{s},\frac{d}{s}\right].

Regime 3: ds≤θ≤π−δ\frac{d}{s}\le\theta\le\pi-\delta. By theorem 8.21.12 in [52],

(sin⁡θ2)α(cos⁡θ2)βPs(α,β)(cos⁡θ)=N−αΓ(s+α+1)s!(θsin⁡θ)1/2Jα(Nθ)+θ1/2O(s−3/2),cs≤θ≤π−δ.\left(\sin\frac{\theta}{2}\right)^\alpha\left(\cos\frac{\theta}{2}\right)^\beta P_s^{(\alpha,\beta)}(\cos\theta)=N^{-\alpha}\frac{\Gamma(s+\alpha+1)}{s!}\left(\frac{\theta}{\sin\theta}\right)^{1/2}J_\alpha(N\theta)+\theta^{1/2}O(s^{-3/2}),\qquad\frac{c}{s}\le\theta\le\pi-\delta.

where N=s+α+β+12N=s+\frac{\alpha+\beta+1}{2}, Jα(x)J_\alpha(x) is the Bessel function of the first kind of order α\alpha [52]. In our case of λ=32\lambda=\frac{3}{2}, α=λ−12=1\alpha=\lambda-\frac{1}{2}=1, and the above formula is equivalent to

12sin⁡θPs(1,1)(cos⁡θ)=N−1Γ(s+2)s!(θsin⁡θ)1/2J1(Nθ)+θ1/2O(s−3/2),ds≤θ≤π−δ.\frac{1}{2}\sin\theta P_s^{(1,1)}(\cos\theta)=N^{-1}\frac{\Gamma(s+2)}{s!}\left(\frac{\theta}{\sin\theta}\right)^{1/2}J_1(N\theta)+\theta^{1/2}O(s^{-3/2}),\qquad\frac{d}{s}\le\theta\le\pi-\delta.

Using the relationship between C~s(3/2)(x)\widetilde{C}_s^{(3/2)}(x) and P(1,1)(x)P^{(1,1)}(x), we have that

C~s(3/2)(x)=2sin⁡θ(s+1)N−1Γ(s+2)s!(θsin⁡θ)1/2J1(Nθ)+θ1/2O(s−3/2),ds≤θ≤π−δ.\widetilde{C}_s^{(3/2)}(x)=\frac{2}{\sin\theta(s+1)}N^{-1}\frac{\Gamma(s+2)}{s!}\left(\frac{\theta}{\sin\theta}\right)^{1/2}J_1(N\theta)+\theta^{1/2}O(s^{-3/2}),\qquad\frac{d}{s}\le\theta\le\pi-\delta.

By using the fact that ∣J1(z)∣≤Cz1/2|J_1(z)|\le\frac{C}{z^{1/2}} for some C>0C>0, we have that

∣2sin⁡θ(s+1)N−1Γ(s+2)s!(θsin⁡θ)1/2J1(Nθ)∣≤A1sin⁡θ(s+1)(θsin⁡θ)1/21(sθ)1/2≤As3/2(sin⁡θ)3/2,\left|\frac{2}{\sin\theta(s+1)}N^{-1}\frac{\Gamma(s+2)}{s!}\left(\frac{\theta}{\sin\theta}\right)^{1/2}J_1(N\theta)\right|\le A\frac{1}{\sin\theta(s+1)}\left(\frac{\theta}{\sin\theta}\right)^{1/2}\frac{1}{(s\theta)^{1/2}}\le\frac{A}{s^{3/2}(\sin\theta)^{3/2}},

for some constant AA. Now choose d>cd>c such that Ad3/2<14\frac{A}{d^{3/2}}<\frac{1}{4}; notice this bound is not circular, as AA is independent of dd but depends only on the Bessel function bound uniform constant CC. The worst case scenario of the bound above occurs at ds\frac{d}{s} for all s≥s0s\ge s_0 with some s0∈Ns_0\in\mathbb{N} large enough, so that

∣2sin⁡θ(s+1)N−1Γ(s+2)s!(θsin⁡θ)1/2J1(Nθ)∣≤Ad3/2.\left|\frac{2}{\sin\theta(s+1)}N^{-1}\frac{\Gamma(s+2)}{s!}\left(\frac{\theta}{\sin\theta}\right)^{1/2}J_1(N\theta)\right|\le\frac{A}{d^{3/2}}.

Then for s≫1s\gg1 such that the error term is also less than 14\frac{1}{4} in magnitude and the leading term dominates, we get that

∣C~s(3/2)(x)∣<12.\left|\widetilde{C}_s^{(3/2)}(x)\right|<\frac{1}{2}.

This ends our three regime analysis.

Then since we need θ∈[0,π−δ]\theta\in[0,\pi-\delta], we have that x=cos⁡θ∈[−cos⁡δ,1]x=\cos\theta\in[-\cos\delta,1]. By the RKG2 argument definition, we have that

τ=0.25\tau= 0.25RKL2RKG2
L2L^2-normL∞L^\infty-normL2L^2-normL∞L^\infty-norm
ErrorRateErrorRateErrorRateErrorRate
τ\tau--------
τ/2\tau/21.31e-05-9.263e-08-1.304e-05-9.215e-08-
τ/4\tau/43.301e-061.9912.331e-081.9913.280e-061.9912.319e-081.990
τ/8\tau/88.264e-071.9985.837e-091.9988.214e-071.9985.808e-091.998
τ/16\tau/162.067e-071.9991.460e-091.9992.054e-071.9991.452e-091.999
τ/32\tau/325.168e-082.0003.650e-102.0005.136e-082.0003.632e-102.000

Table 1. Temporal rate of convergence estimation for the RKL2 and RKG2 methods applied to the Allen–Cahn equation, along with Strang splitting. The convergence rates for both methods are close to 2 as expected.

x(τ)j=1+6τ(s+4)(s−1)μj,x(\tau)_j = 1 + \frac{6\tau}{(s+4)(s-1)}\mu_j,

where μj\mu_j is an eigenvalue of the discrete Laplacian operator. Since we want x≥−cos⁡δx \ge-\cos\delta, we need that

1+6τ(s+4)(s−1)μmax⁡≥−cos⁡δ⟹τ≤(s+4)(s−1)6μmax⁡(1+cos⁡δ),1 + \frac{6\tau}{(s+4)(s-1)}\mu_{\max} \ge-\cos\delta\Longrightarrow\tau\le\frac{(s+4)(s-1)}{6\mu_{\max}}(1+\cos\delta),

where μmax⁡\mu_{\max} is the maximum eigenvalue in absolute value. Notice that the τmax⁡\tau_{\max} of the RKG2 method is achieved at τmax⁡=(s+4)(s−1)3μmax⁡\tau_{\max} = \frac{(s+4)(s-1)}{3\mu_{\max}} because we want to set x(τ)=−1x(\tau)=-1 in this case; therefore, we get that

τ≤1+cos⁡δ2τmax⁡=(1−ε)τmax⁡,\tau\le\frac{1+\cos\delta}{2}\tau_{\max} = (1-\varepsilon)\tau_{\max},

if we set ε:=1−cos⁡δ2>0\varepsilon:= \frac{1-\cos\delta}{2} > 0. As desired, since δ>0\delta> 0 is a small number, we have that ε>0\varepsilon> 0 is a small number. This concludes our proof.

Remark 4.9. The above theorem is only a rough estimate; nevertheless, it shows that RKG2 can be controlled away from its most negative value, so that RKG2 with a more restricted timestepping and large enough stage number ss can be fitted into the proof of Theorem 4.5.

Experimental results

In this section, we present numerical experiments of the RKL2 and RKG2 methods applied to the Allen–Cahn and Cahn–Hilliard equations. For all the Allen–Cahn simulations, we take the nonlinear-linear-nonlinear Strang splitting order written in § 6.2.2. We first verify the second-order convergence rate of the RKL2 and RKG2 methods when applied to the Allen–Cahn and Cahn–Hilliard equations. Following that, we present qualitative results for the Allen–Cahn and Cahn–Hilliard coarsening behaviors, by comparing them with known results in existing papers and random initial condition coarsening. Furthermore, we simulate the Allen–Cahn and the Cahn–Hilliard equations for an LL-shaped domain; finite difference approximation rather than the Fourier spectral method is used for spatial discretization.

Convergence rate test

We consider the two dimensional Allen–Cahn equation in Equation (1), with ε=0.1\varepsilon= 0.1. The simulation domain is [0,2π]2[0, 2\pi]^2 with periodic boundary condition. The spatial domain is divided into 256×256256 \times256 intervals, so the spatial discretization error is smaller than the time discretization error. The initial condition is

u(x,y,0)=0.05sin⁡(x)sin⁡(y).u(x,y,0)=0.05\sin(x)\sin(y).

We test the order of convergence for the second-order Runge–Kutta–Legendre (RKL2) and Runge–Kutta–Gegenbauer (RKG2) methods, when using Strang splitting. We choose the number of substeps s=100s = 100 and simulate until T=1T = 1. We adopt the standard step-doubling convergence rate estimation since there is no exact solution available;

convergence rate≈log⁡(∥ei∥/∥ei+1∥)log⁡2,\text{convergence rate} \approx\frac{\log(\lVert e_i\rVert/\lVert e_{i+1}\rVert)}{\log2},

where ei:=uτ/2i(T)−uτ/2i+1(T)e_i := u_{\tau/2^i}(T) - u_{\tau/2^{i+1}}(T) if we denote uτ(T)u_\tau(T) to be the simulated solution with step size τ\tau at time TT, and ∥⋅∥\lVert\cdot\rVert is either L2L^2-norm or L∞L^\infty-norm. As shown in Table 1, the rate of convergence for the RKL2 and RKG2 methods in both norms are close to the expected second-order. We test the order of convergence for RKL2 and RKG2 when applied to the Cahn–Hilliard equation, under the same conditions as above up to T=0.016T = 0.016. The temporal convergence rate is shown in Table 2, with the desired second-order convergence.

τ=1.6e−3\tau= 1.6\mathrm{e}{-3}RKL2RKG2
L2L^2-normL∞L^\infty-normL2L^2-normL∞L^\infty-norm
ErrorRateErrorRateErrorRateErrorRate
τ\tau--------
τ/2\tau/28.529e-08-5.288e-10-6.404e-08-3.971e-10-
τ/4\tau/42.135e-081.9981.324e-101.9981.603e-081.9999.942e-111.998
τ/8\tau/85.340e-091.9993.313e-111.9994.009e-091.9992.488e-111.999
τ/16\tau/161.335e-092.0008.289e-121.9991.002e-092.0006.222e-121.999
τ/32\tau/323.339e-102.0002.080e-121.9952.504e-102.0011.558e-121.998

Table 2. Temporal rate of convergence estimation for the RKL2 and RKG2 method applied to the Cahn–Hilliard equation. The convergence rates for both methods are close to 2 as expected.

Coarsening process and energy evolution in 2-dimensional space

In this section, we simulate the rescaled Allen–Cahn and Cahn–Hilliard equation

∂u∂t=ε2∇2u−F′(u),∂u∂t=∇2(−ε2∇2u+F′(u))\frac{\partial u}{\partial t}=\varepsilon^2\nabla^2u-F'(u),\quad\frac{\partial u}{\partial t}=\nabla^2\left(-\varepsilon^2\nabla^2u+F'(u)\right)

in the domain [0,2π]×[0,2π][0, 2\pi]\times[0, 2\pi] with periodic boundary condition. We use ε=0.1\varepsilon= 0.1 for all the tests. We first simulate the coarsening behavior to validate our method, by comparing the simulation results to the ones in the existing literature. For the Allen–Cahn equation, we use the initial condition

u(x,y,0)=0.05sin⁡(x)sin⁡(y).(21)u(x,y,0)=0.05\sin(x)\sin(y). \tag*{(21)}

The spatial grid is 512×512512 \times512, with τ=1e−1\tau= 1\mathrm{e}{-1}. We use s=10s=10 for RKL2 and s=12s=12 for the RKG2 methods so that their stability can be guaranteed, without an excessive number of substeps. We simulate until T=8T=8. As shown in Figure 1, the results are visually indistinguishable between methods and are consistent with Figure 1 in [12]. The maximum principle is also preserved for the simulation.

Allen–Cahn equation coarsening evolution at $t=2,4,6$ and $8$

Figure 1. Allen–Cahn equation coarsening evolution at t=2,4,6t = 2, 4, 6 and 88, with initial condition Equation (21).

For the Cahn–Hilliard equation, it is verified in [45] that the symmetric sinusoidal initial condition above leads to an unstable evolution. Therefore, we instead use the asymmetric initial condition in [45]

u(x,y,0)=0.05sin⁡(x)sin⁡(y)+0.001.(22)u(x,y,0) = 0.05\sin(x)\sin(y) + 0.001. \tag*{(22)}

The spatial grid is 256×256256 \times256. We use s=200s=200 for RKL2 and s=240s=240 for RKG2 to guarantee stability. The timestep is τ=1e−2\tau=1\mathrm{e}{-2}. We again simulate until T=8T=8. The results in Figure 3 agree with Figure 5.12 in [28], and using the fact that the energy and the solution do not change after about T=8T=8.

Cahn–Hilliard simulation using RKL2 at $t = 1$, $2$, $5$ and $8$

(a) Cahn–Hilliard simulation, RKL2

Cahn–Hilliard simulation using RKG2 at $t = 1$, $2$, $5$ and $8$

(b) Cahn–Hilliard simulation, RKG2

Figure 3. Cahn–Hilliard equation coarsening evolution at t=1t = 1, 22, 55 and 88, with initial condition Equation (22).

We also compute the energy for the numerical solutions; as shown in Figure 2 and Figure 4, both the Allen–Cahn simulation and the Cahn–Hilliard simulations obtain monotonically decreasing energy curves; the shape and the values of the curves are also similar to the ones in Figure 2 of [13]. For the Allen–Cahn equation, we also compute the modified energy in Equation (19) in addition to the standard energy. For visualization purpose, when plotting the modified energy, we use the exact energy at t=0t=0. As shown in Figure 2, there is no visible difference between the exact energy and the modified energy.

(a) Allen–Cahn equation energy, RKL2

(b) Allen–Cahn equation energy, RKG2

Figure 2. Original energy and modified energy dissipation of the Allen–Cahn equation simulations, with initial condition Equation (21). The modified energy is defined by Equation (19).

(a) Cahn–Hilliard equation energy, RKL2

(b) Cahn–Hilliard equation energy, RKG2

Figure 4. Energy dissipation of the Cahn–Hilliard equation simulations, with initial condition Equation (22).

Then we simulate Allen–Cahn and Cahn–Hilliard equations with random initial conditions. For Allen–Cahn, we use the initial condition

u(x,y,0)=0.5Rand⁡(x,y),(23)u(x,y,0) = 0.5\operatorname{Rand}(x,y), \tag*{(23)}

with the same settings as above. For Cahn–Hilliard, we again use an asymmetric random initial condition

u(x,y,0)=0.0025+0.004Rand⁡(x,y),(24)u(x,y,0) = 0.0025 + 0.004\operatorname{Rand}(x,y), \tag*{(24)}

where Rand⁡(x,y)\operatorname{Rand}(x,y) denotes random numbers sampled from [−1,1][-1,1] uniformly. For Allen–Cahn, the spatial discretization is chosen to be 512×512512 \times512, the timestep is chosen to be τ=1e−1\tau=1\mathrm{e}{-1}, and we simulate until T=8T=8. We still use s=10s=10 for the RKL2 method and s=12s=12 for the RKG2 method. For Cahn–Hilliard, the spatial discretization is chosen to be 256×256256 \times256, the timestep is chosen to be τ=1e−2\tau=1\mathrm{e}{-2}, and we simulate until T=2T=2. The coarsening parameter is chosen to be ε=0.1\varepsilon=0.1. We again use s=200s=200 for RKL2 and s=240s=240 for RKG2 as in the previous example. Periodic boundary condition is applied. The coarsening behaviors are shown in Figure 5 and Figure 7; as shown, both RKL2 and RKG2 produce visually the same coarsening behaviors. Notice that for consistency, we use the color bar with range [−1,1][-1,1] for the Cahn–Hilliard equation plots, although some of the simulation values might exceed this range with trivial amount due to the lack of maximum principle; visually the effect is negligible. The energy curves of the methods are shown

Allen–Cahn equation coarsening evolution at $t=2$, $4$, $6$ and $8$ for RKL2 and RKG2

Figure 5. Allen–Cahn equation coarsening evolution at t=2t=2, 44, 66 and 88, with initial condition Equation (23).

Cahn–Hilliard simulation snapshots using RKL2 and RKG2

Figure 7. Cahn–Hilliard equation coarsening evolution at t=0.5t = 0.5, 11, 1.51.5 and 22, with initial condition Equation (24).

in Figure 6 and Figure 8, both of which are monotonically decreasing. The Allen–Cahn simulations in addition have almost identical energy and modified energy as shown in Figure 6.

Figure 6. Original energy and modified energy dissipation of the Allen–Cahn equation solutions, with initial condition Equation (23). The modified energy is defined by Equation (19).

Figure 8. Energy dissipation of the Cahn–Hilliard equation simulations, with initial condition Equation (24).

Coarsening process and energy evolution in 2-dimensional irregular domain

As discussed in § 4, the exponential time differencing Runge–Kutta (ETDRK) methods in rectangular domains with periodic boundary condition can be solved extremely efficiently, using the Fourier spectral method for spatial discretization. However, it is also known that for irregular domains with non-periodic boundary conditions, Fourier spectral method cannot be applied, and other methods such as Chebyshev or Legendre spectral methods need to be utilized in general [42]. Therefore, RKL and RKG methods are attractive as explicit methods that do not rely on the Fourier spectral approximation. We hence conduct numerical simulations on an LL-shaped domain for Allen–Cahn and Cahn–Hilliard equations, using RKL and RKG methods; the domain is chosen to be [0,2π]∖[π,2π][0, 2\pi] \setminus[\pi, 2\pi]. We again choose the random initial condition Equation (23) and Equation (24), with the same parameters as in the case of rectangular domain. The boundary condition is chosen to be homogeneous Neumann boundary condition. As shown in Figure 9, Figure 11, Figure 10, and Figure 12, the simulations of both methods are visually the same and the energy curves are decreasing. More importantly, the modified energy and the original energy again perfectly match each other, which confirms our claim that the proof of Theorem 4.5 does not rely on the fact that the domain is rectangular.

Allen–Cahn simulations in an L-shaped irregular domain at $t=2$, $4$, $6$, and $8$ using RKL2 and RKG2

Figure 9. Allen–Cahn equation coarsening evolution at t=2t=2, 44, 66 and 88, within the LL-shaped irregular domain and initial condition Equation (23).

Energy and modified energy of Allen–Cahn equation simulations using RKL2 and RKG2

Figure 10. Energy dissipation of the Allen–Cahn equation simulations on LL-shaped domain, with initial condition Equation (23). The modified energy is defined by Equation (19).

Cahn–Hilliard simulation coarsening evolution on an L-shaped domain using RKL2 and RKG2

Figure 11. Cahn–Hilliard equation coarsening evolution at t=0.5t = 0.5, 11, 1.51.5 and 22, within the LL-shaped irregular domain and initial condition Equation (24). As shown, there is no visual difference between the RKL2 and RKG2 simulations.

Figure 12. Energy dissipation of the Cahn–Hilliard equation simulations on LL-shaped domain, with initial condition Equation (24).

Adaptive time stepping

As can be seen from the previous examples, the phase-field models behavior can drastically change quickly over short timescales while varying very slowly over longer timescales. Various existing methods [45, 12] have utilized adaptive time stepping for faster simulation; we therefore implement adaptive time stepping for the Cahn–Hilliard equation. We use the strategy motivated in in [54], which defines a dimensionless modified local truncation error and calculates the infinity norm of the local truncation error as a criterion. However, if their strategy were directly used in our simulation, division by zero would be (almost) inevitable and would lead to ineffective measurement. We therefore consider the further modified local truncation error

LTE=∣ek+1−ττoldek∣,whereek+1=∥uk+1−uk∥∞∥uk∥∞,ek=∥uk−uk−1∥∞∥uk−1∥∞,τ=tk+1−tk,τold=tk−tk−1.(25)\mathrm{LTE} = \left| e_{k+1} - \frac{\tau}{\tau_{\mathrm{old}}} e_k \right|, \qquad \text{where}\quad e_{k+1} = \frac{\left\|u^{k+1}-u^k\right\|_{\infty}}{\left\|u^k\right\|_{\infty}},\quad e_k = \frac{\left\|u^k-u^{k-1}\right\|_{\infty}}{\left\|u^{k-1}\right\|_{\infty}},\quad\tau= t^{k+1}-t^k,\quad\tau_{\mathrm{old}} = t^k-t^{k-1}. \tag*{(25)}

Compared to the original formulation in [54], our formulation first applies the infinity norm to the finite difference errors ek+1e_{k+1} and eke_k, so that the division by zero issue is eliminated.

Another noteworthy component of the adaptive time stepping strategy is the update of number of substeps ss for the

Algorithm 1Substep ss changing procedure
Require:Timestep τ\tau and spatial grid size hh; old substep number ss, method RKL or RKG; case max or min.
Step 1:Compute CFL number C=ετh4C = \frac{\varepsilon\tau}{h^4}.
Step 2:If method is RKL, set snew=ceil⁡(−1+9+512C2)s_{\mathrm{new}} = \operatorname{ceil}\left(\frac{-1+\sqrt{9+512C}}{2}\right). If method is RKG, set snew=ceil⁡(−3+25+768C2)s_{\mathrm{new}} = \operatorname{ceil}\left(\frac{-3+\sqrt{25+768C}}{2}\right).
Step 3:Relax the condition a little bit by setting snew=ceil⁡(1.02s)s_{\mathrm{new}} = \operatorname{ceil}(1.02s). If case is max, set s=max⁡{snew,s}s = \max\{s_{\mathrm{new}},s\}. If case is min, set s=min⁡{snew,s}s = \min\{s_{\mathrm{new}},s\}.

Table 1.

SRK methods. Since the implicit method used in [54] is unconditionally stable, their adaptive time stepping strategy does not need to account for the stability constraint. However, the stability constraints of our SRK methods are fully controlled by the number of substeps ss, therefore we need to change ss each time the timestep τ\tau increases or decreases to avoid redundant or insufficient number of substeps. Since the fourth-order term dominates in the Cahn–Hilliard, we choose our ss analytically according to the CFL condition given in [46, 48] with a small relaxation. The detailed method for changing ss is described in Algorithm 1.

We summarize the algorithm described above in Algorithm 2, and conduct Cahn–Hilliard equation simulations accordingly. We choose the adaptive timestep tolerance toltol to be 1e-3 and MaxCountMaxCount to be 3. We again choose the Algorithm 2 Adaptive time stepping procedure

Require: Numerical solutions of two steps uku^k, uk−1u^{k-1}; current and previous timesteps τ\tau, τold\tau_{\mathrm{old}}; adaptive time tolerance toltol; small error counting countcount; maximum number of count MaxCountMaxCount.

Step 1: Compute uk+1u^{k+1} using timestep τ\tau.

Step 2: Calculate LTE according to Equation (25).

Step 3: If the original energy is not monotone, set τold=τ\tau_{\mathrm{old}}=\tau and τ=0.8τ\tau=0.8\tau. Set ss with τ\tau and case min, then go back to step 1.

Step 4: Else, if LTE <tol< tol, count=count+1count=count+1. If countcount reaches MaxCountMaxCount, set τold=τ\tau_{\mathrm{old}}=\tau, τ=1.2τ\tau=1.2\tau, and count=0count=0.

Step 5: Set ss with τ\tau and case max, then go back to step 1.

Step 6: Endif.

initial condition to be the perturbed sinusoidal initial condition Equation (22), with the same initial parameters as in the constant timestep case. As shown in Figure 13 and Figure 14, the adaptive time stepping simulations and the energy curves match the constant timestep results in Figure 3 and Figure 4. Moreover, we plot the τ\tau used for each timestep for both methods in Figure 15; as shown in the figures, the τ\tau used significantly increases from T=0T=0 to T=8T=8. We also conduct a wall-clock comparison of the simulation times. For RKL2 the constant timestep version uses 896.02s, while the adaptive timestep version uses 344.95s. For RKG2, the constant timestep version uses 1185.29s, while the adaptive timestep version uses 502.32s. This shows the effectiveness of our adaptive timestepping strategy in the Cahn–Hilliard equation simulations.

Cahn–Hilliard simulation adaptive time stepping, RKL2 and RKG2

Figure 13. Cahn–Hilliard equation coarsening evolution at (approximately) t=1,2,5t = 1, 2, 5 and 88, with initial condition Equation (22) and adaptive time stepping. As shown, there is no visual difference between the RKL2 and RKG2 simulations, and the results are similar to the constant timestep simulations in Figure 3.

Figure 14. Energy dissipation of the Cahn–Hilliard equation solutions with adaptive time stepping.

Change of timestep size for the Cahn–Hilliard equation, RKL2 and RKG2

Figure 15. Evolution of τ\tau over the adaptive time stepping simulation of the Cahn–Hilliard equation, with sinusoidal initial condition Equation (22).

Conclusion

In this paper, we present simulations of Allen–Cahn and Cahn–Hilliard equations using the explicit stabilized Runge–Kutta methods. Theoretical analysis guaranties that the discrete maximum principle is preserved for the Allen–Cahn equation with Strang splitting within the stability constraint. We also prove that the modified discrete energy, under two different ways of splitting, dissipate with a minor tradeoff of maximum timestep allowed, by estimating the positive-definiteness of the stability polynomial. Numerical simulations also verify and validate the rate of convergence and the energy dissipation of both Allen–Cahn and Cahn–Hilliard equations, for both rectangular domains and irregular domains, under periodic and Neumann boundary conditions.

For future work, one can potentially develop higher-order stabilized Runge–Kutta methods specifically designed for gradient flows. Another direction can be to prove dissipation of the original energy for Allen–Cahn and the Cahn–Hilliard equations, or find out the limitation of the methods, since only empirical evidence is provided for 2D maximum principle and Cahn–Hilliard energy dissipation. Moreover, our methods can be applied to higher-order parabolic equations with variable coefficients easily, and gain more efficiency when compared with the implicit methods. Finally, theoretical guarantee of the maximum principle when applying RKG methods to the Allen–Cahn equation with Dirichlet boundary condition can also help us better understand the stabilized Runge–Kutta methods, and it would also be beneficial if the maximum principle for RKL and RKG methods can be proved for more general domain shapes.

Acknowledgment

The work of the second author was supported by the Advanced Simulation and Computing Program (ASC) at Los Alamos National Laboratory, operated by Triad National Security, LLC, for the National Nuclear Security Administration of the U.S. Department of Energy contract 89233218CNA000001. The work of the first and the third author was supported by ONR grant DOD N00014-23-1-2565.

The first author used AI-assisted tools, including ChatGPT and Claude, for support with function visualizations in theorem proving, English language polishing, LaTeX editing, and partial code checking. The third author used Claude for one stage of proofreading but not the final proofreading of the paper, which was checked by a human editor.

References

References

  1. [1]S. M. Allen, J. W. Cahn, A microscopic theory for antiphase boundary motion and its application to antiphase domain coarsening, Acta Metallurgica 27 (6) (1979) 1085–1095. doi:https://doi.org/10.1016/0001-6160(79)90196-2. URL https://www.sciencedirect.com/science/article/pii/0001616079901962
  2. [2]J. W. Cahn, J. E. Hilliard, Free energy of a nonuniform system. I. interfacial free energy, The Journal of Chemical Physics 28 (2) (1958) 258–267. doi:10.1063/1.1744102.DOI
  3. [3]L.-Q. Chen, Phase-field models for microstructure evolution, Annual Review of Materials Research 32 (Volume 32, 2002) (2002) 113–140. doi:https://doi.org/10.1146/annurev.matsci.32.112001.132041. URL https://www.annualreviews.org/content/journals/10.1146/annurev.matsci.32.112001.132041
  4. [4]L. Golubović, A. Levandovsky, D. Moldovan, Interface dynamics and far-from-equilibrium phase transitions in multilayer epitaxial growth and erosion on crystal surfaces: Continuum theory insights, East Asian Journal on Applied Mathematics 1 (4) (2011) 297–371. doi:10.4208/eajam.040411.030611a.DOI
  5. [5]M. E. Gurtin, D. Polignone, J. Viñals, Two-phase binary fluids and immiscible fluids described by an order parameter, Mathematical Models and Methods in Applied Sciences 06 (06) (1996) 815–831. arXiv:https://doi.org/10.1142/S0218202596000341, doi:10.1142/S0218202596000341. URL https://doi.org/10.1142/S0218202596000341
  6. [6]O. Penrose, P. C. Fife, Thermodynamically consistent models of phase-field type for the kinetic of phase transitions, Physica D: Nonlinear Phenomena 43 (1) (1990) 44–62. doi:https://doi.org/10.1016/0167-2789(90)90015-H. URL https://www.sciencedirect.com/science/article/pii/016727899090015H
  7. [7]H. Gomez, L. Cueto-Felgueroso, R. Juanes, Three-dimensional simulation of unstable gravity-driven infiltration of water into a porous medium, Journal of Computational Physics 238 (2013) 217–239. doi:https://doi.org/10.1016/j.jcp.2012.12.018. URL https://www.sciencedirect.com/science/article/pii/S0021999112007498
  8. [8]D. J. Eyre, Unconditionally gradient stable time marching the Cahn–Hilliard equation, in: Computational & Mathematical Models of Microstructural Evolution, Vol. 529 of MRS Online Proceedings Library, Materials Research Society, 1998, pp. 39–46. doi:10.1557/PROC-529–39.DOI
  9. [9]U. M. Ascher, S. J. Ruuth, B. T. R. Wetton, Implicit-explicit methods for time-dependent partial differential equations, SIAM Journal on Numerical Analysis 32 (3) (1995) 797–823. arXiv:https://doi.org/10.1137/0732037, doi:10.1137/0732037. URL https://doi.org/10.1137/0732037
  10. [10]Z. Xu, X. Yang, H. Zhang, Z. Xie, Efficient and linear schemes for anisotropic Cahn–Hilliard model using the stabilized-invariant energy quadratization (S-IEQ) approach, Computer Physics Communications 238 (2019) 36–49. doi:https://doi.org/10.1016/j.cpc.2018.12.019. URL https://www.sciencedirect.com/science/article/pii/S0010465518304417
  11. [11]J. Shen, J. Xu, J. Yang, The scalar auxiliary variable (SAV) approach for gradient flows, Journal of Computational Physics 353 (2018) 407–416. doi:https://doi.org/10.1016/j.jcp.2017.10.021. URL https://www.sciencedirect.com/science/article/pii/S002199911730774X
  12. [12]Z. Fu, J. Yang, Energy-decreasing exponential time differencing Runge–Kutta methods for phase-field models, Journal of Computational Physics 454 (2022) 110943. doi:https://doi.org/10.1016/j.jcp.2022.110943. URL https://www.sciencedirect.com/science/article/pii/S0021999122000055
  13. [13]Q. Du, L. Ju, X. Li, Z. Qiao, Maximum principle preserving exponential time differencing schemes for the nonlocal Allen–Cahn equation, SIAM Journal on Numerical Analysis 57 (2) (2019) 875–898. arXiv:https://doi.org/10.1137/18M118236X, doi:10.1137/18M118236X. URL https://doi.org/10.1137/18M118236X
  14. [14]Z. Fu, J. Shen, J. Yang, Higher-order energy-decreasing exponential time differencing Runge-Kutta methods for gradient flows, Science China Mathematics 68 (7) (2025) 1727–1746. doi:10.1007/s11425-024-2337-3.DOI
  15. [15]B. P. Vollmayr-Lee, A. D. Rutenberg, Fast and accurate coarsening simulation with an unconditionally stable time step, Physical Review E 68 (6) (2003) 066703. doi:10.1103/PhysRevE.68.066703.
  16. [16]N. Zheng, J. Shen, Stability and error analysis of linear imex schemes for sixth-order Cahn–Hilliard-type equations, Communications in Nonlinear Science and Numerical Simulation 145 (2025) 108724. doi:https://doi.org/10.1016/j.cnsns.2025.108724. URL https://www.sciencedirect.com/science/article/pii/S1007570425001352
  17. [17]Z. Fu, T. Tang, J. Yang, Energy diminishing implicit-explicit Runge–Kutta methods for gradient flows, Mathematics of Computation 93 (2024) 2745–2767. doi:10.1090/mcom/3509.
  18. [18]J. Shen, X. Yang, Numerical approximations of Allen-Cahn and Cahn-Hilliard equations (2010). doi:10.3934/dcds.2010.28.1669. URL https://www.aimsciences.org/article/id/b4cba61a-377b-449a-b226-e85e2c00cc6aDOI
  19. [19]X. Yang, Linear, first and second-order, unconditionally energy stable numerical schemes for the phase field model of homopolymer blends, Journal of Computational Physics 327 (2016) 294–316. doi:https://doi.org/10.1016/j.jcp.2016.09.029. URL https://www.sciencedirect.com/science/article/pii/S0021999116304508
  20. [20]X. Yang, Efficient, second-order in time, and energy stable scheme for a new hydrodynamically coupled three components volume-conserved Allen–Cahn phase-field model, Mathematical Models and Methods in Applied Sciences 31 (04) (2021) 753–787. arXiv:https://doi.org/10.1142/S0218202521500184, doi:10.1142/S0218202521500184. URL https://doi.org/10.1142/S0218202521500184
  21. [21]X. Yang, L. Ju, Efficient linear schemes with unconditional energy stability for the phase field elastic bending energy model, Computer Methods in Applied Mechanics and Engineering 315 (2017) 691–712. doi:https://doi.org/10.1016/j.cma.2016.10.041. URL https://www.sciencedirect.com/science/article/pii/S0045782516306016
  22. [22]X. Yang, L. Ju, Linear and unconditionally energy stable schemes for the binary fluid–surfactant phase field model, Computer Methods in Applied Mechanics and Engineering 318 (2017) 1005–1029. doi:https://doi.org/10.1016/j.cma.2017.02.011. URL https://www.sciencedirect.com/science/article/pii/S0045782516317856
  23. [23]X. Yang, J. Zhao, X. He, Linear, second order and unconditionally energy stable schemes for the viscous Cahn–Hilliard equation with hyperbolic relaxation using the invariant energy quadratization method, Journal of Computational and Applied Mathematics 343 (2018) 80–97. doi:https://doi.org/10.1016/j.cam.2018.04.027. URL https://www.sciencedirect.com/science/article/pii/S0377042718302085
  24. [24]X. Yang, J. Zhao, Q. Wang, Numerical approximations for the molecular beam epitaxial growth model based on the invariant energy quadratization method, Journal of Computational Physics 333 (2017) 104–127. doi:https://doi.org/10.1016/j.jcp.2016.12.025. URL https://www.sciencedirect.com/science/article/pii/S0021999116306799
  25. [25]X. Yang, J. Zhao, Q. Wang, J. Shen, Numerical approximations for a three-component Cahn–Hilliard phase-field model based on the invariant energy quadratization method, Mathematical Models and Methods in Applied Sciences 27 (11) (2017) 1993–2030.DOI
  26. [26]H. Yu, X. Yang, Numerical approximations for a phase-field moving contact line model with variable densities and viscosities, Journal of Computational Physics 334 (2017) 665–686. doi:https://doi.org/10.1016/j.jcp.2017.01.026. URL https://www.sciencedirect.com/science/article/pii/S0021999117300426
  27. [27]X. Bai, J. Sun, J. Shen, W. Yao, Z. Guo, A Ginzburg–Landau-$H^{-1}$ model and its SAV algorithm for image inpainting, Journal of Scientific Computing 96 (2) (2023) 40. doi:10.1007/s10915-023-02252-z. URL https://doi.org/10.1007/s10915-023-02252-z
  28. [28]Q. Cheng, C. Liu, J. Shen, A new Lagrange multiplier approach for gradient flows, Computer Methods in Applied Mechanics and Engineering 367 (2020) 113070. doi:https://doi.org/10.1016/j.cma.2020.113070. URL https://www.sciencedirect.com/science/article/pii/S0045782520302541
  29. [29]Q. Cheng, C. Liu, J. Shen, Generalized SAV approaches for gradient systems, Journal of Computational and Applied Mathematics 394 (2021) 113532. doi:https://doi.org/10.1016/j.cam.2021.113532. URL https://www.sciencedirect.com/science/article/pii/S0377042721001515
  30. [30]Q. Cheng, J. Shen, C. Wang, Unique solvability and error analysis of a scheme using the Lagrange multiplier approach for gradient flows, SIAM Journal on Numerical Analysis 63 (2) (2025) 772–799. arXiv:https://doi.org/10.1137/24M1659303,doi:10.1137/24M1659303. URL https://doi.org/10.1137/24M1659303
  31. [31]Y. Di, Y. Ma, J. Shen, J. Zhang, A variable time-step IMEX-BDF2 SAV scheme and its sharp error estimate for the Navier–Stokes equations, ESAIM: Mathematical Modelling and Numerical Analysis 57 (3) (2023) 1143–1170. doi:10.1051/m2an/2023007. URL https://doi.org/10.1051/m2an/2023007
  32. [32]X. Huang, J. Shen, Bound/positivity preserving SAV schemes for the Patlak-Keller-Segel-Navier-Stokes system, Journal of Computational Physics 480 (2023) 112034. doi:https://doi.org/10.1016/j.jcp.2023.112034. URL https://www.sciencedirect.com/science/article/pii/S0021999123001298
  33. [33]X. Li, J. Shen, Error estimate of a consistent splitting GSAV scheme for the Navier-Stokes equations, Applied Numerical Mathematics 188 (2023) 62–74. doi:https://doi.org/10.1016/j.apnum.2023.03.004. URL https://www.sciencedirect.com/science/article/pii/S0168927423000661
  34. [34]X. Liu, J. Shen, N. Zheng, Efficient energy stable numerical schemes for Cahn–Hilliard equations with dynamical boundary conditions, Journal of Computational Physics 509 (2024) 113037. doi:https://doi.org/10.1016/j.jcp.2024.113037. URL https://www.sciencedirect.com/science/article/pii/S0021999124002869
  35. [35]Z. Liu, Y. Zhang, X. Li, A novel energy-optimized technique of SAV-based (EOP-SAV) approaches for dissipative systems, Journal of Scientific Computing 101 (2) (2024) 38. doi:10.1007/s10915-024-02677-0. URL https://doi.org/10.1007/s10915-024-02677-0
  36. [36]J. Shen, J. Xu, J. Yang, A new class of efficient and robust energy stable schemes for gradient flows, SIAM Review 61 (3) (2019) 474–506. arXiv:https://doi.org/10.1137/17M1150153, doi:10.1137/17M1150153. URL https://doi.org/10.1137/17M1150153
  37. [37]Y. Zhang, J. Shen, A generalized SAV approach with relaxation for dissipative systems, Journal of Computational Physics 464 (2022) 111311. doi:https://doi.org/10.1016/j.jcp.2022.111311. URL https://www.sciencedirect.com/science/article/pii/S0021999122003734
  38. [38]M. Hochbruck, A. Ostermann, Exponential integrators, Acta Numerica 19 (2010) 209–286. doi:10.1017/S0962492910000048.DOI
  39. [39]S. Cox, P. Matthews, Exponential time differencing for stiff systems, Journal of Computational Physics 176 (2) (2002) 430–455. doi:https://doi.org/10.1006/jcph.2002.6995. URL https://www.sciencedirect.com/science/article/pii/S0021999102969950
  40. [40]Q. Du, L. Ju, X. Li, Z. Qiao, Maximum bound principles for a class of semilinear parabolic equations and exponential time-differencing schemes, SIAM Review 63 (2) (2021) 317–359. arXiv:https://doi.org/10.1137/19M1243750, doi:10.1137/19M1243750. URL https://doi.org/10.1137/19M1243750
  41. [41]Y. Liu, C. Quan, D. Wang, On the maximum bound principle and energy dissipation of exponential time differencing methods for the matrix-valued Allen–Cahn equation, IMA Journal of Numerical Analysis (2024) drae090arXiv:https://academic.oup.com/imajna/advance-article-pdf/doi/10.1093/imanum/drae090/61251649/drae090.pdf, doi:10.1093/imanum/drae090. URL https://doi.org/10.1093/imanum/drae090
  42. [42]J. Shen, T. Tang, L.-L. Wang, Spectral methods: algorithms, analysis and applications, Vol. 41, Springer Science & Business Media, 2011.DOI
  43. [43]D. Li, C. Quan, J. Xu, Stability and convergence of Strang splitting. Part I: Scalar Allen-Cahn equation, Journal of Computational Physics 458 (2022) 111087. doi:https://doi.org/10.1016/j.jcp.2022.111087. URL https://www.sciencedirect.com/science/article/pii/S0021999122001498
  44. [44]M. Gao, C. Quan, Z. Zhang, Convergence analysis of a novel Strang directional splitting method for the Allen–Cahn equations, Journal of Computational and Applied Mathematics 473 (2026) 116868. doi:https://doi.org/10.1016/j.cam.2025.116868. URL https://www.sciencedirect.com/science/article/pii/S0377042725003826
  45. [45]Y. Cheng, A. Kurganov, Z. Qu, T. Tang, Fast and stable explicit operator splitting methods for phase-field models, Journal of Computational Physics 303 (2015) 45–65. doi:https://doi.org/10.1016/j.jcp.2015.09.005. URL https://www.sciencedirect.com/science/article/pii/S0021999115005859
  46. [46]C. D. Meyer, D. S. Balsara, T. D. Aslam, A stabilized Runge–Kutta–Legendre method for explicit super-time-stepping of parabolic and mixed equations, Journal of Computational Physics 257 (2014) 594–626. doi:https://doi.org/10.1016/j.jcp.2013.08.021. URL https://www.sciencedirect.com/science/article/pii/S0021999113005597
  47. [47]S. O’Sullivan, Runge–Kutta–Gegenbauer explicit methods for advection-diffusion problems, Journal of Computational Physics 388 (2019) 209–223. doi:https://doi.org/10.1016/j.jcp.2019.03.001. URL https://www.sciencedirect.com/science/article/pii/S0021999119301706
  48. [48]T. Skaras, T. Saxton, C. Meyer, T. D. Aslam, Super-time-stepping schemes for parabolic equations with boundary conditions, Journal of Computational Physics 425 (2021) 109879. doi:https://doi.org/10.1016/j.jcp.2020.109879. URL https://www.sciencedirect.com/science/article/pii/S0021999120306537
  49. [49]Z. Tan, T. D. Aslam, A. L. Bertozzi, Explicit monotone stable super-time-stepping methods for finite time singularities, SIAM Journal on Applied Mathematics 86 (2) (2026) 457–481. arXiv:https://epubs.siam.org/doi/pdf/10.1137/25M1747646, doi:10.1137/25M1747646. URL https://epubs.siam.org/doi/abs/10.1137/25M1747646DOI
  50. [50]Z. Tan, W. Wang, A. L. Bertozzi, E. K. Ryu, STORK: Faster diffusion and flow matching sampling by resolving both stiffness and structure-dependence, in: The Fourteenth International Conference on Learning Representations, 2026. URL https://openreview.net/forum?id=Ce0IVXM14rarxiv.org/abs/2505.24210
  51. [51]G. Strang, On the construction and comparison of difference schemes, SIAM Journal on Numerical Analysis 5 (3) (1968) 506–517. arXiv:https://doi.org/10.1137/0705041, doi:10.1137/0705041. URL https://doi.org/10.1137/0705041
  52. [52]G. Szegő, Orthogonal Polynomials, American Math. Soc: Colloquium publ, American Mathematical Society, 1975. URL https://books.google.com/books?id=2phzwgEACAAJ
  53. [53]L. Gatteschi, On the zeros of Jacobi polynomials and Bessel functions, in: International conference on special functions: theory and computation (Turin, 1984). Rend. Sem. Mat. Univ. Politec. Torino (Special Issue), 1985, pp. 149–177.
  54. [54]T. Kostić, A. L. Bertozzi, Statistical density estimation using threshold dynamics for geometric motion, Journal of Scientific Computing 54 (2-3) (2013) 513–530. doi:10.1007/s10915-012-9615-6.DOI

Paper details

vlnc:W_b6cxtr9y
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

ChatGPT, Claude

View statement
Looking for related papers…
Contents