betaFeedback

WorksMathematics

Energy-Dissipative, High-Order BDF Schemes for Simulating Geometric Flows via Energy Increments

Abstract

We propose energy-dissipative backward differentiation formula (BDF) methods for geometric flows in the Barrett–Garcke–Nürnberg (BGN) parametric finite element framework. Starting from a Lagrange-multiplier formulation, we replace the multistep approximation of the energy derivative by an energy-increment equation while retaining the BDF geometric equations. The Simpson quadrature and two interleaved time grids provide third- and fourth-order constructions. The positive quadrature weights ensure stepwise decay of the original discrete geometric energy for every nondegenerate solution of the nonlinear system. We establish an explicit conditional consistency bound for the scalar energy equation and a regularity criterion for the Newton system of curve-shortening flows. Then, the proposed schemes extend to simulating curve diffusion, mean curvature flow, and surface diffusion. Finally, numerical experiments demonstrate geometric accuracy, energy decay, good mesh distribution, and high performance of the proposed schemes.

Introduction

Curvature-driven flows provide models for interface motion in materials science, biomembrane dynamics, and image processing [8, 9, 12]. Many of these models are gradient flows of a geometric energy WW and satisfy

ddtW(Γ(t))=−D(Γ(t)),D(Γ(t))≥0,(1)\frac{d}{dt}W(\Gamma(t))=-\mathcal{D}(\Gamma(t)),\qquad\mathcal{D}(\Gamma(t))\ge0, \tag*{(1)}

where Γ(t)\Gamma(t) is the evolving curve or surface and D\mathcal{D} is the dissipation rate. Curve shortening and curve diffusion flows, for example, decrease perimeter with dissipation rates ∫Γκ2 ds\int_{\Gamma}\kappa^2\,\mathrm{d}s and ∫Γ∣∂sκ∣2 ds\int_{\Gamma}|\partial_s\kappa|^2\,\mathrm{d}s, respectively; analogous laws hold for mean curvature flow and surface diffusion [7]. A numerical method should reproduce this energy decrease while resolving the interface motion accurately. Combining higher-order time integration with discrete energy dissipation and a well-distributed moving mesh is the problem addressed here.

Parametric finite element methods (PFEMs) approximate the interface directly by a polygonal curve or a triangulated surface. Their performance depends on how the mesh evolves along with the geometry. The Barrett–Garcke–Nürnberg (BGN) formulation [4, 5, 6] couples the normal-velocity equation to a weak curvature relation. Higher-order BGN methods must approximate both the interface velocity and the geometry on which the discrete equations are evaluated. Jiang, Su, and Zhang developed a second-order method with interlaced perimeter decay [20] and backward differentiation formula (BDF) methods of high orders from the second to fourth [21]. In the latter, lower-order BGN solves supply the predicted integration geometry. A related predictor–corrector construction was developed for curve and surface diffusion [22]. These developments motivate retaining the BGN spatial structure while improving the treatment of its energy dissipation law.

Lagrange multipliers provide a direct way to impose geometric balance laws as additional equations. Garcke, Jiang, Su, and Zhang [16] used this approach to enforce energy dissipation and area or volume conservation in PFEMs for surface diffusion. Their BDF formulation applies the multistep difference operator to the energy itself. For BDF2, a nonincreasing starting step allows an induction argument for subsequent energy decay. At orders three and four, however, a nonpositive BDF approximation of the energy derivative does not by itself imply Wm+1≤WmW^{m+1} \le W^m. The issue is the sign of the individual energy increments in the multistep recurrence, rather than the order of approximation of the derivative. Moreover, replacing this equation by a backward-Euler energy update would generally reduce the consistency order of the coupled method.

Several other approaches achieve high-order energy stability for phase-field gradient flow models. On fixed domains, scalar auxiliary variable (SAV) methods [19, 25, 26] and energy-diminishing Runge–Kutta methods [13, 14] provide important examples. For evolving interfaces, Guo and Li [17] combine BDF integration with minimal-deformation, SAV, and Lagrange-multiplier formulations for anisotropic surface diffusion. The dissipation theorem for their SAV variants concerns an auxiliary energy, which is distinct from monotonicity of the original discrete geometric energy. Duan, Li, and Zhang [11] obtained high-order energy-diminishing methods for curvature-dependent surface energies by averaged vector-field collocation. More recently, Ma and Rao [24] constructed algebraically stable implicit Runge–Kutta extensions of BGN-type methods for curves and genus-zero surfaces. Zhang, Andrews, and Farrell [27] developed continuous Petrov–Galerkin time discretizations of arbitrary order with auxiliary variables that preserve energy dissipation and, for surface diffusion, enclosed volume of the surface. These methods establish that high-order geometric energy dissipation is achievable. In this paper, our objective is its realization within the BDF3 and BDF4 Lagrange-multiplier formulations, with their geometric evolution and curvature equations retained.

We replace the multistep energy-derivative equation by an approximation of the integrated balance in (1). Writing Wm=W(Γm)W^m = W(\Gamma^m) and τ=tm+1−tm\tau= t_{m+1} - t_m, the scalar update takes the form

Wm+1−Wm=−τQkm+1,(2)W^{m+1} - W^m = -\tau Q_k^{m+1}, \tag*{(2)}

where Qkm+1Q_k^{m+1} approximates the time-averaged dissipation rate by a quadrature rule with positive weights. For k=3,4k = 3, 4, Simpson’s rule combines endpoint and midpoint dissipation values. An auxiliary half-integer time grid supplies the midpoint states, and the two grids are advanced alternately. The energy in (2) is, respectively, the length for curve evolution, area for surface evolution, or anisotropic free energy for anisotropic geometric flows. Thus, every nondegenerate solution satisfying the scalar equation inherits the corresponding original energy decrease rate, separately on each time level. The construction only changes the scalar balance rather than the BDF approximation of the geometric evolution.

The analysis identifies the conditions needed to interpret this construction. We establish stepwise energy dissipation and quantify the scalar consistency defect in terms of the quadrature error and the errors in the nodal dissipation data. We also derive the Newton system for curve shortening flow (CSF), give a Schur-complement criterion for its local regularity, and relate the nonlinear energy residual to the computed energy balance. These results make explicit the roles of predicted geometry, midpoint accuracy, initialization, and nonlinear solution accuracy.

The rest of the paper is organized as follows. Section 2 introduces the energy-increment BDF schemes and their scalar analysis. Section 3 presents the interleaved implementation, and Section 4 develops the Newton solver. Extensions to other geometric flows are given in Section 5, followed by numerical experiments in Section 6 and conclusions in Section 7.

Energy-increment BDF discretization

We use curve-shortening flow to introduce the construction. After fixing the geometric conventions and spatial discretization, we examine the energy constraint in derivative-based BDF schemes and replace it by an integrated balance. The resulting scalar equation admits separate consistency and dissipation arguments.

Curve shortening flow and spatial semidiscretization

Let Γ(t)⊂R2\Gamma(t) \subset\mathbb{R}^{2} be a smooth, simple closed curve with initial configuration Γ(0)=Γ0\Gamma(0) = \Gamma_0 and regular parametrization

X(ρ,t):=(x1(ρ,t),x2(ρ,t)):I×[0,T]→R2,I=[0,1] with endpoints identified.X(\rho,t):=(x_1(\rho,t),x_2(\rho,t)):\mathbb{I}\times[0,T]\to\mathbb{R}^{2},\qquad\mathbb{I}=[0,1]\text{ with endpoints identified}.

We orient the curve counterclockwise and choose the outward unit normal n\mathbf{n}. The convention κn=−∂ssX\kappa\mathbf{n}=-\partial_{ss}X then assigns curvature κ=1/R\kappa=1/R to a circle of radius RR. All continuous identities are understood while the curve remains smooth and regular. The arc-length variable is s(ρ,t):=∫0ρ∣∂qX∣ dqs(\rho,t):=\int_0^\rho|\partial_qX|\,dq, with derivative and measure ∂s:=1∣∂ρX∣∂ρ\partial_s:=\frac{1}{|\partial_\rho X|}\partial_\rho, ds=∣∂ρX∣ dρds=|\partial_\rho X|\,d\rho.

Write W(t)=L(Γ(t))W(t)=L(\Gamma(t)) for the perimeter and V=∂tX⋅nV=\partial_tX\cdot\mathbf{n} for the normal velocity. The perimeter variation underlying the Lagrange-multiplier approach [16] is

ddtL(Γ(t))=∫Γ(t)κV ds.\frac{d}{dt}L(\Gamma(t))=\int_{\Gamma(t)}\kappa V\,ds.

Introducing a scalar multiplier λ(t)\lambda(t) in the curvature direction and imposing the perimeter dissipation law gives

∂tX(⋅,t)⋅n=−κ+λ(t)κ,(3)\partial_tX(\cdot,t)\cdot\mathbf{n}=-\kappa+\lambda(t)\kappa, \tag*{(3)}
κn=−∂ssX,(4)\kappa\mathbf{n}=-\partial_{ss}X, \tag*{(4)}
dLdt=−∫Γ(t)κ2 ds.(5)\frac{dL}{dt}=-\int_{\Gamma(t)}\kappa^2\,ds. \tag*{(5)}

Combining the perimeter variation with (3) and (5) yields λ(t)∫Γ(t)κ2 ds=0\lambda(t)\int_{\Gamma(t)}\kappa^2\,ds=0. Thus λ(t)=0\lambda(t)=0 whenever κ≢0\kappa\not\equiv0, and (2.1) recovers classical curve-shortening flow.

For the spatial discretization, partition I\mathbb{I} uniformly into

I=[0,1]=⋃j=1NIj,Ij=[ρj−1,ρj],ρj=jh,h=1N.\mathbb{I}=[0,1]=\bigcup_{j=1}^{N}I_j,\qquad I_j=[\rho_{j-1},\rho_j],\qquad\rho_j=jh,\qquad h=\frac{1}{N}.

Define the finite element space

Vh:={u∈C(I):u∣Ij∈P1, ∀j=1,…,N, u(ρ0)=u(ρN)}⊆H1(I),V^h := \{u \in C(\mathbb{I}) : u|_{I_j} \in\mathbb{P}_1,\ \forall j=1,\ldots,N,\ u(\rho_0)=u(\rho_N)\} \subseteq H^1(\mathbb{I}),

where P1\mathbb{P}_1 denotes the set of polynomials of degree at most one. For Xh∈[Vh]2X^h \in[V^h]^2, let Γh=Xh(I)\Gamma^h=X^h(\mathbb{I}) be the associated polygon, with edges hj=Xh(ρj)−Xh(ρj−1)\mathbf{h}_j=X^h(\rho_j)-X^h(\rho_{j-1}). All polygons used to evaluate inner products or arc-length derivatives are assumed to have positive edge lengths. The discrete perimeter is

Wh(t):=L(Γh(t)):=∫Γh(t)1 ds=∑j=1N∣hj(t)∣.(6)W^h(t):=L(\Gamma^h(t)):=\int_{\Gamma^h(t)}1\,\mathrm{d}s=\sum_{j=1}^{N}|\mathbf{h}_j(t)|. \tag*{(6)}

On IjI_j, the discrete arc-length derivative is ∂s=(hj/∣hj∣)∂ρ\partial_s=(\mathbf{h}_j/|\mathbf{h}_j|)\partial_\rho. We use the standard inner product (u,v)Γh:=∫Γhu⋅v ds(u,v)_{\Gamma^h}:=\int_{\Gamma^h}u\cdot v\,\mathrm{d}s and its mass-lumped counterpart

(u,v)Γhh:=12∑j=1N∣hj∣[(u⋅v)(ρj−)+(u⋅v)(ρj−1+)],(7)(u,v)_{\Gamma^h}^h:=\frac{1}{2}\sum_{j=1}^{N}|\mathbf{h}_j|\left[(u\cdot v)(\rho_j^-)+(u\cdot v)(\rho_{j-1}^+)\right], \tag*{(7)}

for scalar- or vector-valued piecewise continuous functions. The values at ρj±\rho_j^\pm denote one-sided traces, with periodic extension at ρ0\rho_0 and ρN\rho_N.

Given compatible initial data, seek (Xh,κh,λh)∈[Vh]2×Vh×R(X^h,\kappa^h,\lambda^h)\in[V^h]^2\times V^h\times\mathbb{R} such that, for all (φh,ωh)∈Vh×[Vh]2(\varphi^h,\omega^h)\in V^h\times[V^h]^2,

(∂tXh,φh)Γhh+(1−λh)(κh,φh)Γhh=0,(8)(\partial_tX^h,\varphi^h)_{\Gamma^h}^h+(1-\lambda^h)(\kappa^h,\varphi^h)_{\Gamma^h}^h=0, \tag*{(8)}
(κhnh,ωh)Γhh−(∂sXh,∂sωh)Γh=0,(9)(\kappa^h\mathbf{n}^h,\omega^h)_{\Gamma^h}^h-(\partial_sX^h,\partial_s\omega^h)_{\Gamma^h}=0, \tag*{(9)}
dWhdt+(κh,κh)Γhh=0.(10)\frac{\mathrm{d}W^h}{\mathrm{d}t}+(\kappa^h,\kappa^h)_{\Gamma^h}^h=0. \tag*{(10)}

Here Γh(t)=Xh(I,t)\Gamma^h(t)=X^h(\mathbb{I},t), and the elementwise outward unit normal is

nh∣Ij=−hj⊥∣hj∣,j=1,…,N,\mathbf{n}^h|_{I_j}=-\frac{\mathbf{h}_j^\perp}{|\mathbf{h}_j|},\qquad j=1,\ldots,N,

with (v1,v2)⊥=(−v2,v1)(v_1,v_2)^\perp=(-v_2,v_1) denoting counterclockwise rotation by π/2\pi/2. This sign is consistent with the counterclockwise orientation.

For a differentiable polygonal evolution, dWh/dt=(∂sXh,∂s∂tXh)Γh\mathrm{d}W^h/\mathrm{d}t=(\partial_sX^h,\partial_s\partial_tX^h)_{\Gamma^h}. Testing (8) with κh\kappa^h and (9) with ∂tXh\partial_tX^h therefore gives dWh/dt=−(1−λh)(κh,κh)Γhh\mathrm{d}W^h/\mathrm{d}t=-(1-\lambda^h)(\kappa^h,\kappa^h)_{\Gamma^h}^h. Equation (10) then implies λh=0\lambda^h=0 when κh≢0\kappa^h\not\equiv0. The scalar constraint is thus compatible with the semidiscrete BGN equations; its role becomes essential after time discretization.

Remark 2.1 (Compatibility of the initial polygon). Let ℓj=∣hj∣\ell_j=|\mathbf{h}_j|, tj=hj/ℓj\mathbf{t}_j=\mathbf{h}_j/\ell_j, and bj=−(hj+hj+1)⊥/2\mathbf{b}_j=-(\mathbf{h}_j+\mathbf{h}_{j+1})^\perp/2, with cyclic indices. Testing (9) with the vector-valued nodal basis functions at vertex jj gives κjbj=tj−tj+1\kappa_j\mathbf{b}_j=\mathbf{t}_j-\mathbf{t}_{j+1}. Taking its dot product with hj+hj+1\mathbf{h}_j+\mathbf{h}_{j+1} yields

(ℓj−ℓj+1)(1−tj⋅tj+1)=0.(11)(\ell_j-\ell_{j+1})(1-\mathbf{t}_j\cdot\mathbf{t}_{j+1})=0. \tag*{(11)}

Thus adjacent edges with different unit tangents must have equal lengths. When bj≠0\mathbf{b}_j\ne0, the Euclidean least-squares approximation to the nodal curvature equation is

κjLS=bj⋅(tj−tj+1)∣bj∣2.(12)\kappa_j^{\mathrm{LS}}=\frac{\mathbf{b}_j\cdot(\mathbf{t}_j-\mathbf{t}_{j+1})}{|\mathbf{b}_j|^2}. \tag*{(12)}

This value can be used to initialize the fully discrete curvature, but it does not enforce (9) on an incompatible polygon. Accordingly, the consistency analysis below assumes the existence of a smooth, compatible semidiscrete reference solution.

BDF discretization with an energy-derivative constraint

Following the BGN/BDF construction of [21] and the Lagrange-multiplier approach of [16], we first apply the BDF operator to both the position and the energy. Let tm=mτt_m=m\tau, where τ=T/M>0\tau=T/M>0 and M∈NM\in\mathbb{N}. The predicted polygon Γ~m+1\widetilde{\Gamma}^{m+1} is constructed before assembling the equations and has the same nodal connectivity as the computed polygons. Functions on different polygons are identified by their coefficients on the common reference mesh. The normal and arc-length derivatives in the following equations are evaluated on Γ~m+1\widetilde{\Gamma}^{m+1}. Given the required history, for m≥k−1m\ge k-1 seek (Xm+1,κm+1,λm+1)∈[Vh]2×Vh×R(\mathbf{X}^{m+1},\kappa^{m+1},\lambda^{m+1})\in[V^h]^2\times V^h\times\mathbb{R} such that, for all (φh,ωh)∈Vh×[Vh]2(\varphi^h,\omega^h)\in V^h\times[V^h]^2,

(DtkXm+1τ,φhn~m+1)Γ~m+1h+(κm+1,φh)Γ~m+1h−λm+1(κm+1,φh)Γ~m+1h=0(13)\left(\frac{D_t^k\mathbf{X}^{m+1}}{\tau},\varphi^h\widetilde{\mathbf{n}}^{m+1}\right)_{\widetilde{\Gamma}^{m+1}}^h+\left(\kappa^{m+1},\varphi^h\right)_{\widetilde{\Gamma}^{m+1}}^h-\lambda^{m+1}\left(\kappa^{m+1},\varphi^h\right)_{\widetilde{\Gamma}^{m+1}}^h=0 \tag*{(13)}
(κm+1n~m+1,ωh)Γ~m+1h−(∂sXm+1,∂sωh)Γ~m+1h=0(14)\left(\kappa^{m+1}\widetilde{\mathbf{n}}^{m+1},\omega^h\right)_{\widetilde{\Gamma}^{m+1}}^h-\left(\partial_s\mathbf{X}^{m+1},\partial_s\omega^h\right)_{\widetilde{\Gamma}^{m+1}}^h=0 \tag*{(14)}
DtkWm+1τ=−(κm+1,κm+1)Γ~m+1h.(15)\frac{D_t^kW^{m+1}}{\tau}=-\left(\kappa^{m+1},\kappa^{m+1}\right)_{\widetilde{\Gamma}^{m+1}}^h. \tag*{(15)}

The dimensionless BDF operators are

Dt1Xm+1=Xm+1−Xm,D_t^1\mathbf{X}^{m+1}=\mathbf{X}^{m+1}-\mathbf{X}^m,
Dt2Xm+1=32Xm+1−2Xm+12Xm−1,D_t^2\mathbf{X}^{m+1}=\frac{3}{2}\mathbf{X}^{m+1}-2\mathbf{X}^m+\frac{1}{2}\mathbf{X}^{m-1},
Dt3Xm+1=116Xm+1−3Xm+32Xm−1−13Xm−2,D_t^3\mathbf{X}^{m+1}=\frac{11}{6}\mathbf{X}^{m+1}-3\mathbf{X}^m+\frac{3}{2}\mathbf{X}^{m-1}-\frac{1}{3}\mathbf{X}^{m-2},
Dt4Xm+1=2512Xm+1−4Xm+3Xm−1−43Xm−2+14Xm−3.D_t^4\mathbf{X}^{m+1}=\frac{25}{12}\mathbf{X}^{m+1}-4\mathbf{X}^m+3\mathbf{X}^{m-1}-\frac{4}{3}\mathbf{X}^{m-2}+\frac{1}{4}\mathbf{X}^{m-3}.

The same coefficients define DtkWm+1D_t^kW^{m+1}, with Wm:=W(Xm):=L(Γm)W^m:=W(\mathbf{X}^m):=L(\Gamma^m). Thus Dtk/τD_t^k/\tau approximates the time derivative. We call these schemes BGN-LM-BDFkk, k=1,2,3,4k=1,2,3,4. For k=1k=1, set Γ~m+1=Γm\widetilde{\Gamma}^{m+1}=\Gamma^m; for higher orders, use lower-order solves to predict the geometry. Section 3 specifies the predictors and start-up data.

Remark 2.2 (Derivative-based energy updates). Write ΔWm+1=Wm+1−Wm\Delta W^{m+1}=W^{m+1}-W^m and let Dm+1≥0\mathcal{D}^{m+1}\ge0 denote the endpoint dissipation. The BDF2 energy equation reads

3ΔWm+1=ΔWm−2τDm+1.3\Delta W^{m+1}=\Delta W^m-2\tau\mathcal{D}^{m+1}.

Consequently, stepwise decay follows by induction if W1≤W0W^1\le W^0. For BDF3 and BDF4, respectively,

11ΔWm+1=7ΔWm−2ΔWm−1−6τDm+1,11\Delta W^{m+1}=7\Delta W^m-2\Delta W^{m-1}-6\tau\mathcal{D}^{m+1},
25ΔWm+1=23ΔWm−13ΔWm−1+3ΔWm−2−12τDm+1.25\Delta W^{m+1}=23\Delta W^m-13\Delta W^{m-1}+3\Delta W^{m-2}-12\tau\mathcal{D}^{m+1}.

Nonpositive history increments do not ensure a nonpositive right-hand side. For BDF3, W0=2W^0=2, W1=W2=1W^1=W^2=1, and D3=0\mathcal{D}^3=0 give W3=13/11>W2W^3=13/11>W^2. For BDF4, W0=W1=2W^0=W^1=2, W2=W3=1W^2=W^3=1, and D4=0\mathcal{D}^4=0 give W4=38/25>W3W^4=38/25>W^3. These examples concern only the scalar recurrences and illustrate the limitation discussed in [16]; they are not counterexamples to a complete geometric scheme.

The Crank–Nicolson formulation in [16] already uses an energy increment, with midpoint dissipation. We instead couple an integrated energy balance to the BDF approximation of the geometric evolution, using quadrature consistent with the intended order.

Energy-increment formulation and scalar analysis

The proposed modification discretizes the scalar energy law in integral form. Integrating (5) over [tm,tm+1][t_m,t_{m+1}] gives

W(tm+1)−W(tm)+∫tmtm+1∫Γ(t)κ2 ds dt=0.W(t_{m+1})-W(t_m)+\int_{t_m}^{t_{m+1}}\int_{\Gamma(t)}\kappa^2\,\mathrm{d}s\,\mathrm{d}t=0.

We retain the geometric equations (2.7a)–(2.7b) and replace (15) by

Wm+1−Wm+τQkm+1=0,(16)W^{m+1}-W^m+\tau Q_k^{m+1}=0, \tag*{(16)}

where Qkm+1Q_k^{m+1} approximates the time-averaged dissipation rate. For k=1k=1, the right-endpoint rule gives

Q1m+1=(κm+1,κm+1)Γm+1h.(17)Q_1^{m+1}=(\kappa^{m+1},\kappa^{m+1})_{\Gamma_{m+1}^h}. \tag*{(17)}

For k=2k=2, we use the trapezoidal rule,

Q2m+1=12[(κm+1,κm+1)Γm+1h+(κm,κm)Γmh].(18)Q_2^{m+1}=\frac{1}{2}\left[(\kappa^{m+1},\kappa^{m+1})_{\Gamma_{m+1}^h}+(\kappa^m,\kappa^m)_{\Gamma_m^h}\right]. \tag*{(18)}

For k=3,4k=3,4, we use Simpson’s rule,

Qkm+1=16[(κm+1,κm+1)Γm+1h+4(κm+12,κm+12)Γm+12h+(κm,κm)Γmh].(19)Q_k^{m+1}=\frac{1}{6}\left[(\kappa^{m+1},\kappa^{m+1})_{\Gamma_{m+1}^h}+4(\kappa^{m+\frac{1}{2}},\kappa^{m+\frac{1}{2}})_{\Gamma_{m+\frac{1}{2}}^h}+(\kappa^m,\kappa^m)_{\Gamma_m^h}\right]. \tag*{(19)}

The resulting schemes are denoted by ES-BDFkk. ES-BDF3 and ES-BDF4 share this quadrature rule but differ in their BDF operators, predictors, and accuracy requirements for the intermediate data. The endpoint dissipation is evaluated on the predicted geometry, whereas the history and midpoint terms use their computed geometries. We now examine consistency and energy dissipation for this scalar update.

For the consistency analysis, assume that (2.4) admits a sufficiently smooth, compatible reference solution and write

Dh(t):=(κh(t),κh(t))Γh(t).\mathcal{D}^h(t):=(\kappa^h(t),\kappa^h(t))_{\Gamma_h(t)}.

The semidiscrete energy law (10) gives

Wh(tm+1)−Wh(tm)τ+1τ∫tmtm+1Dh(t) dt=0.(20)\frac{W^h(t_{m+1})-W^h(t_m)}{\tau}+\frac{1}{\tau}\int_{t_m}^{t_{m+1}}\mathcal{D}^h(t)\,\mathrm{d}t=0. \tag*{(20)}

Denote by Dm,θ\mathcal{D}_{m,\theta} the computed dissipation at quadrature node tm+θt_{m+\theta} for the interval [tm,tm+1][t_m,t_{m+1}]. In the present curve-shortening formulation,

Dm,1=(κm+1,κm+1)Γm+1h,Dm,0=(κm,κm)Γmh,\mathcal{D}_{m,1}=(\kappa^{m+1},\kappa^{m+1})_{\Gamma_{m+1}^h},\qquad\mathcal{D}_{m,0}=(\kappa^m,\kappa^m)_{\Gamma_m^h},

and, when Simpson’s rule is used,

Dm,1/2=(κm+12,κm+12)Γm+12h.\mathcal{D}_{m,1/2}=(\kappa^{m+\frac{1}{2}},\kappa^{m+\frac{1}{2}})_{\Gamma_{m+\frac{1}{2}}^h}.

The required accuracy of these data is expressed by

Dm,θ=Dh(tm+θ)+O(τk)(21)D_{m,\theta}=D^h(t_{m+\theta})+O(\tau^k) \tag*{(21)}

at all nodes of the selected rule, uniformly in time for fixed hh. This is an assumption on both curvature and geometry, since the latter determines the mass-lumped weights. In particular, Dm,0D_{m,0} need not equal Dm−1,1D_{m-1,1}: the same curvature coefficients are evaluated on Γm\Gamma^m when used as history, but on Γ~m\widetilde{\Gamma}^m when used as an endpoint. For k=3,4k=3,4, the assumption also concerns the midpoint states supplied by the staggered construction in § 2.

Proposition 2.3 (Conditional scalar consistency). Let a compatible semidiscrete solution exist on Im=[tm,tm+1]\mathcal{I}_m=[t_m,t_{m+1}], with Dh∈Cq(Im)D^h\in C^q(\mathcal{I}_m). For the right-endpoint, trapezoidal, and Simpson rules, respectively, set

(q,Cq)=(1,1/2),(2,1/12),(4,1/2880).(q,C_q)=(1,1/2),\quad(2,1/12),\quad(4,1/2880).

Define em=max⁡θ∣Dm,θ−Dh(tm+θ)∣e_m=\max_\theta|D_{m,\theta}-D^h(t_{m+\theta})|, with the maximum taken over the rule’s nodes. Then the defect in the energy-increment equation satisfies

∣Wh(tm+1)−Wh(tm)+τQkm+1∣≤τem+Cqτq+1∥(Dh)(q)∥L∞(Im).(22)\left|W^h(t_{m+1})-W^h(t_m)+\tau Q_k^{m+1}\right| \le\tau e_m+C_q\tau^{q+1}\left\|(D^h)^{(q)}\right\|_{L^\infty(\mathcal{I}_m)}. \tag*{(22)}

Under (2.13), this defect is O(τk+1)O(\tau^{k+1}) for k=1,2,3,4k=1,2,3,4, or O(τk)O(\tau^k) after division by τ\tau, with constants at fixed hh.

Proof. By (2.12), the defect is the difference between τQkm+1\tau Q_k^{m+1} and the exact dissipation integral. Add and subtract the quadrature rule applied to the exact dissipation values. If ωθ\omega_\theta are the normalized weights, the data error is bounded by

τ∑θωθ∣Dm,θ−Dh(tm+θ)∣≤τem,\tau\sum_\theta\omega_\theta|D_{m,\theta}-D^h(t_{m+\theta})|\le\tau e_m,

since ωθ≥0\omega_\theta\ge0 and ∑θωθ=1\sum_\theta\omega_\theta=1. The standard quadrature remainder gives the second term in (2.14). Finally, q≥kq\ge k for each selected rule, which gives the stated orders when em=O(τk)e_m=O(\tau^k).

Proposition 2.3 concerns the scalar constraint with exact reference energies at the interval endpoints. A convergence result for the coupled algorithm would also require control of the geometry, curvature, initialization, predictors, and staggered histories; the diffusion extensions additionally involve discrete-gradient errors. The estimate is at fixed hh, with no assertion of uniform constants under spatial refinement or mesh degeneration.

Proposition 2.4 (Stepwise energy dissipation). Let τ>0\tau>0, let all quadrature geometries be nondegenerate, and define Qkm+1Q_k^{m+1} by (2.9)–(2.11). Every solution satisfying the scalar equation (2.8) exactly obeys

Wm+1≤Wm.W^{m+1}\le W^m.

For curve shortening flows, the inequality is strict if (2.7b) also holds and the updated polygon has at least one nonzero edge.

Proof. Each dissipation term is nonnegative, and each quadrature weight is positive. Hence Qkm+1≥0Q_k^{m+1}\ge0, and (2.8) gives

Wm+1−Wpm=−τQkm+1≤0.W^{m+1}-W^{p_m}=-\tau Q_k^{m+1}\le0.

Equality would force the endpoint dissipation to vanish. For curve shortening, positivity of the mass-lumped weights on Γ~m+1\widetilde{\Gamma}^{m+1} then gives κm+1=0\kappa^{m+1}=0. Taking ωh=Xm\omega^h=X^m in (2.7b) would then give ∥∂sXm+1∥L2(Γ~m+1)2=0\|\partial_sX^{m+1}\|_{L^2(\widetilde{\Gamma}^{m+1})}^2=0, so every edge of the updated polygon would vanish, a contradiction.

The proposition applies to every solved step, but does not establish solvability for arbitrary τ\tau. In particular, the Simpson update and Wm+1>0W^{m+1}>0 require

τ6(4Dm,1/2+Dm,0)<Wm.(23)\frac{\tau}{6}\left(4D_{m,1/2}+D_{m,0}\right)<W^m. \tag*{(23)}

Thus the prescribed history and midpoint data can preclude a positive-energy solution; (23) is only a necessary condition. For successive solved steps on a single time grid, the scalar equations telescope to

WM+τ∑m=m0M−1Qkm+1=Wm0.W^M+\tau\sum_{m=m_0}^{M-1}Q_k^{m+1}=W^{m_0}.

This identity holds separately on the two grids used by the staggered implementation. For inexact nonlinear solves, the corresponding residual correction is discussed in § 3. The principal role of the increment constraint is at BDF orders three and four: BDF2 with a nonincreasing start, Crank–Nicolson, and the trapezoidal increment update already admit energy-decay arguments.

Boole’s rule has positive weights (7,32,12,32,7)/90(7,32,12,32,7)/90, but a higher-order extension would also require accurate intermediate states and a stable multistep coupling. Not all closed Newton–Cotes rules have positive weights. We focus here on design orders k=2,3,4k=2,3,4, with ES-BDF1 as a first-order reference. The remaining implementation task is to supply the midpoint geometry and curvature required by Simpson’s rule. § 2 describes the interleaved time grids and the associated accuracy requirements.

Staggered implementation

We now describe the implementation of the ES-BDFkk schemes for curve-shortening flow. For ES-BDF3 and ES-BDF4, Simpson’s rule requires geometric data at the midpoint of each time interval. We obtain these data by evolving an auxiliary sequence on the half-integer time grid.

Two ingredients follow the BGN/BDFkk methodology of [21]. First, the usual start-up target for an order-kk multistep method is an O(τk)O(\tau^k) error in the position history. A first-order start-up integrator with substep τ~\widetilde{\tau} over an interval of length O(τ)O(\tau) has an accumulated error of order O(ττ~)O(\tau\widetilde{\tau}), provided the local error and stability constants remain bounded. This motivates τ~=O(τk−1)\widetilde{\tau}=O(\tau^{k-1}). Choose τ~=τk−1/(2L)\widetilde{\tau}=\tau^{k-1}/(2L) with an integer LL so that every half-grid start-up time is reached exactly. This scaling does not by itself control the algebraic curvature variable or verify (21). A first-order scheme may have an O(τ~)O(\widetilde{\tau}) curvature error even when its accumulated position error is O(ττ~)O(\tau\widetilde{\tau}). The effects of these distinct errors on the coupled start-up require separate analysis.

Second, each target geometry is predicted before the BGN equations are assembled. We use BGN1 for ES-BDF2 and an ES-BDF2 predictor for ES-BDF3. Here BGN1 means (13) and (14) with k=1k=1, λ=0\lambda=0, and no additional scalar energy equation; it is distinct from ES-BDF1. Each nested ES-BDF2 prediction uses its own BGN1 geometry. ES-BDF2 uses only endpoint data and requires one time grid. For the Simpson updates, Algorithm 3.1 advances integer and half-integer sequences, each with time step τ\tau.

Algorithm 3.1 advances the integer and half-integer sequences alternately. The integer-level update at tm+1t_{m+1} uses the previously computed midpoint data at tm+12t_{m+\frac{1}{2}}. The subsequent half-integer update at tm+32t_{m+\frac{3}{2}} then uses the newly computed state at tm+1t_{m+1} as its midpoint data. The corresponding scalar energy equation on the half-

Algorithm 3.1 ES-BDF3 staggered update

Require: Initial data X0\mathbf{X}^{0}, time step τ\tau, start-up step τ~=τ2/(2L)=O(τ2)\widetilde{\tau}=\tau^{2}/(2L)=O(\tau^{2}), terminal time T=MτT=M\tau with integer M≥3M\geq3.

Ensure: Integer-level states (Xm,κm)(\mathbf{X}^{m},\kappa^{m}) and half-integer-level states (Xm+12,κm+12)(\mathbf{X}^{m+\frac{1}{2}},\kappa^{m+\frac{1}{2}}).

  1. Initialization.

  2. Compute κ0\kappa^{0} by (2.6).

  3. Compute (X12,κ12)(\mathbf{X}^{\frac{1}{2}},\kappa^{\frac{1}{2}}), (X1,κ1)(\mathbf{X}^{1},\kappa^{1}), and (X32,κ32)(\mathbf{X}^{\frac{3}{2}},\kappa^{\frac{3}{2}}) by successive BGN1 start-up solves with step τ~\widetilde{\tau}.

  4. Compute (X2,κ2)(\mathbf{X}^{2},\kappa^{2}) on the integer grid by ES-BDF2.

  5. Compute (X52,κ52)(\mathbf{X}^{\frac{5}{2}},\kappa^{\frac{5}{2}}) on the half-integer grid by ES-BDF2.

  6. Time stepping.

  7. for m=2,…,M−1m=2,\ldots,M-1 do

  8. Construct the predicted geometry Γ~m+1\widetilde{\Gamma}^{m+1} by an ES-BDF2 predictor on the integer grid.

  9. Solve the ES-BDF3 system for (Xm+1,κm+1,λm+1)(\mathbf{X}^{m+1},\kappa^{m+1},\lambda^{m+1}).

  10. if m<M−1m<M-1 then

  11. Construct the predicted geometry Γ~m+32\widetilde{\Gamma}^{m+\frac{3}{2}} by an ES-BDF2 predictor on the half-integer grid.

  12. Solve the ES-BDF3 system for (Xm+32,κm+32,λm+32)(\mathbf{X}^{m+\frac{3}{2}},\kappa^{m+\frac{3}{2}},\lambda^{m+\frac{3}{2}}).

  13. end if

  14. end for

integer grid is

Wm+32−Wm+12=−τQ3m+32,(24)W^{m+\frac{3}{2}}-W^{m+\frac{1}{2}}=-\tau Q_{3}^{m+\frac{3}{2}}, \tag*{(24)}

where

Q3m+32=16[(κm+32,κm+32)Γ~m+32h+4(κm+1,κm+1)Γm+1h+(κm+12,κm+12)Γm+12h].(25)Q_{3}^{m+\frac{3}{2}}=\frac{1}{6}\left[(\kappa^{m+\frac{3}{2}},\kappa^{m+\frac{3}{2}})_{\widetilde{\Gamma}^{m+\frac{3}{2}}}^{h}+4(\kappa^{m+1},\kappa^{m+1})_{\Gamma^{m+1}}^{h}+(\kappa^{m+\frac{1}{2}},\kappa^{m+\frac{1}{2}})_{\Gamma^{m+\frac{1}{2}}}^{h}\right]. \tag*{(25)}

After initialization, both sequences use BDF3. This matches their design orders but leaves (2.13) to be verified for the coupled evolution. Proposition 2.4 applies separately to the two sequences: it gives Wm+1≤WmW^{m+1}\leq W^{m} and Wm+32≤Wm+12W^{m+\frac{3}{2}}\leq W^{m+\frac{1}{2}}, without ordering Wm+12W^{m+\frac{1}{2}} and WmW^{m}.

The dependency structure of one staggered update is illustrated in Figure 1.

Figure 1. Dependency structure of one staggered ES-BDF3 update. Black arrows indicate data used for the integer-level update at tm+1t_{m+1}, and blue arrows indicate data used for the subsequent half-integer-level update at tm+32t_{m+\frac{3}{2}}. The arrows include both the BDF3 history and the midpoint data required by Simpson’s rule.

For ES-BDF4, take M≥4M\geq4 and retain four history values on each grid. The integer sequence is initialized through t3t_{3} and the half-integer sequence through t72t_{\frac{7}{2}} using lower-order solves refined sufficiently to target O(τ4)O(\tau^{4}) start-up errors. An unrefined ES-BDF2 macro step generally contributes an O(τ3)O(\tau^{3}) local error and is not by itself sufficient for this purpose. The ES-BDF3 predictor uses only the already available history and midpoint data and is discarded after construction of the predicted geometry. The histories are listed in Table 1. As for ES-BDF3, the integer solve uses the available half-integer midpoint, and the following half-integer solve uses the newly computed integer state. Each ES-BDF3 or ES-BDF4 macro step normally requires two nonlinear solves, together with its predictor solves. Newton iteration counts alone therefore do not measure total computational cost.

ES-BDF3ES-BDF4
Integer-level historym,m−1,m−2m,m-1,m-2m,m−1,m−2,m−3m,m-1,m-2,m-3
Half-integer-level historym+12,m−12,m−32m+\frac{1}{2},m-\frac{1}{2},m-\frac{3}{2}m+12,m−12,m−32,m−52m+\frac{1}{2},m-\frac{1}{2},m-\frac{3}{2},m-\frac{5}{2}
PredictorES-BDF2ES-BDF3
Integer startupthrough t2t_{2}through t3t_{3}
Half-integer startupthrough t5/2t_{5/2}through t7/2t_{7/2}

Table 1. Staggered implementation data for ES-BDF3 and ES-BDF4.

Newton solution of the ES-BDF3 system

At the target time tm+1t_{m+1}, Newton’s method solves for (Xm+1,κm+1,λm+1)∈[Vh]2×Vh×R(X^{m+1},\kappa^{m+1},\lambda^{m+1})\in[V^{h}]^{2}\times V^{h}\times\mathbb{R}. The predicted geometry, its normal and metric, and the midpoint data are fixed throughout the iteration. Initial curvature is obtained from (2.6), following [20 Remark 2.1]; this requires bj0≠0b_{j}^{0}\ne0, which is stronger than nonzero individual edge lengths. The position and curvature histories are supplied by § 2.

For a polygonal curve Γ=X(I)\Gamma=X(I) and a perturbation Y(ε)=X+εXδY(\varepsilon)=X+\varepsilon X^{\delta}, the first variation of the discrete perimeter is

δW(X)(Xδ)=ddεW(Y(ε))∣ε=0=(∂sX,∂sXδ)Γ.(26)\delta W(X)(X^{\delta})=\left.\frac{d}{d\varepsilon}W(Y(\varepsilon))\right|_{\varepsilon=0}=(\partial_{s}X,\partial_{s}X^{\delta})_{\Gamma}. \tag*{(26)}

Equivalently, δW(X)(Xδ)=∑j(hj/∣hj∣)⋅(Xδ(ρj)−Xδ(ρj−1))\delta W(X)(X^{\delta})=\sum_{j}(\mathbf{h}_{j}/|\mathbf{h}_{j}|)\cdot(X^{\delta}(\rho_{j})-X^{\delta}(\rho_{j-1})). This derivative is evaluated on the current iterate Γm+1,i\Gamma^{m+1,i}, whereas the metric in the endpoint dissipation remains fixed on Γ~m+1\widetilde{\Gamma}^{m+1}.

A natural initial guess is

Xm+1,0=X~m+1,κm+1,0=κm,λm+1,0=0.X^{m+1,0}=\widetilde{X}^{m+1},\qquad\kappa^{m+1,0}=\kappa^{m},\qquad\lambda^{m+1,0}=0.

Given (Xm+1,i,κm+1,i,λm+1,i)(X^{m+1,i},\kappa^{m+1,i},\lambda^{m+1,i}), the Newton correction (Xδ,κδ,λδ)∈[Vh]2×Vh×R(X^{\delta},\kappa^{\delta},\lambda^{\delta})\in[V^{h}]^{2}\times V^{h}\times\mathbb{R} is determined by the following linearized system. Let R1i(φh)R_{1}^{i}(\varphi^{h}) and R2i(ωh)R_{2}^{i}(\omega^{h}) be the left-hand sides of (2.7a) and (2.7b), evaluated at the current iterate with k=3k=3, and let REi=W(Xm+1,i)−Wm+τQ3m+1,iR_{E}^{i}=W(X^{m+1,i})-W^{m}+\tau Q_{3}^{m+1,i}. Then

(116τXδ,φhn~m+1)Γ~m+1+(1−λm+1,i)(κδ,φh)Γ~m+1−λδ(κm+1,i,φh)Γ~m+1=−R1i(φh),(27)\left(\frac{11}{6\tau}X^{\delta},\varphi^{h}\widetilde{\mathbf{n}}^{m+1}\right)_{\widetilde{\Gamma}^{m+1}}+(1-\lambda^{m+1,i})(\kappa^{\delta},\varphi^{h})_{\widetilde{\Gamma}^{m+1}}-\lambda^{\delta}(\kappa^{m+1,i},\varphi^{h})_{\widetilde{\Gamma}^{m+1}}=-R_{1}^{i}(\varphi^{h}), \tag*{(27)}
(κδn~m+1,ωh)Γ~m+1−(∂sXδ,∂sωh)Γ~m+1=−R2i(ωh),(28)(\kappa^{\delta}\widetilde{\mathbf{n}}^{m+1},\omega^{h})_{\widetilde{\Gamma}^{m+1}}-(\partial_{s}X^{\delta},\partial_{s}\omega^{h})_{\widetilde{\Gamma}^{m+1}}=-R_{2}^{i}(\omega^{h}), \tag*{(28)}
(∂sXm+1,i,∂sXδ)Γm+1,i+τ3(κm+1,i,κδ)Γ~m+1=−REi.(29)(\partial_{s}X^{m+1,i},\partial_{s}X^{\delta})_{\Gamma^{m+1,i}}+\frac{\tau}{3}(\kappa^{m+1,i},\kappa^{\delta})_{\widetilde{\Gamma}^{m+1}}=-R_{E}^{i}. \tag*{(29)}

The factor τ/3\tau/3 in (29) is twice the endpoint Simpson weight τ/6\tau/6; no geometry derivative enters this quadratic term because the predicted metric is fixed.

Proposition 4.1 (Regularity of the CSF Newton system). At an iterate with a nondegenerate current polygon and λm+1,i<1\lambda^{m+1,i}<1, suppose that the predicted polygon has positive edge lengths and its nodal vectors bj=−(h^j+h^j+1)⊥/2\mathbf{b}_j=-(\widehat{\mathbf{h}}_j+\widehat{\mathbf{h}}_{j+1})^\perp/2 span R2\mathbb{R}^2. Let KiK_i be the coefficient matrix of {(xref}(4.2)–(28)} for z=(Xδ,κδ)\mathbf{z}=(\mathbf{X}^{\delta},\kappa^\delta), with λδ=0\lambda^\delta=0. Then KiK_i is invertible. Writing the full Jacobian as

Ji=(KibiciT0),si=−ciTKi−1bi,(30)J_i=\begin{pmatrix}K_i & b_i\\ c_i^{\mathsf T} & 0\end{pmatrix},\qquad s_i=-c_i^{\mathsf T}K_i^{-1}b_i, \tag*{(30)}

where bib_i is the multiplier column and ciTc_i^{\mathsf T} is the energy row, JiJ_i is invertible if and only if si≠0s_i\ne0.

Proof. For a homogeneous solution of the first two equations, test with φh=κδ\varphi^h=\kappa^\delta and ωh=Xδ\omega^h=\mathbf{X}^\delta. Eliminating their mixed term gives

116τ∥∂sXδ∥L2(Γ~m+1)2+(1−λm+1,i)∥κδ∥h,Γ~m+12=0,\frac{11}{6\tau}\|\partial_s\mathbf{X}^\delta\|_{L^2(\widetilde{\Gamma}^{m+1})}^2+(1-\lambda^{m+1,i})\|\kappa^\delta\|_{h,\widetilde{\Gamma}^{m+1}}^2=0,

where ∥v∥h,Γ2=(v,v)Γh\|v\|_{h,\Gamma}^2=(v,v)^h_\Gamma. Hence κδ=0\kappa^\delta=0 and Xδ\mathbf{X}^\delta is constant. The first equation and the spanning condition force this constant to vanish. The square matrix KiK_i thus has trivial kernel. Block elimination then gives the scalar Schur complement sis_i in (30), proving the second assertion. ▫

At a nondegenerate root satisfying these conditions, the inverse function theorem gives local uniqueness of that root and standard Newton theory gives local quadratic convergence for sufficiently close initial guesses. This conditional statement proves neither existence of a root nor that sis_i stays away from zero along the computed evolution. The condition λm+1,i<1\lambda^{m+1,i}<1 is sufficient for the block argument, not asserted for every iterate. These conclusions concern CSF, not the anisotropic matrix.

After solving {(xref}(4.2)–(29)}, the iterate is updated by

Xm+1,i+1=Xm+1,i+Xδ,κm+1,i+1=κm+1,i+κδ,λm+1,i+1=λm+1,i+λδ.\mathbf{X}^{m+1,i+1}=\mathbf{X}^{m+1,i}+\mathbf{X}^{\delta},\qquad \kappa^{m+1,i+1}=\kappa^{m+1,i}+\kappa^\delta,\qquad \lambda^{m+1,i+1}=\lambda^{m+1,i}+\lambda^\delta.

The increment-based stopping criterion used to describe the iteration is

max⁡{∥Xδ∥L∞,∥κδ∥L∞,∣λδ∣}≤tol.(31)\max\left\{\|\mathbf{X}^{\delta}\|_{L^\infty},\|\kappa^\delta\|_{L^\infty},|\lambda^\delta|\right\}\le\mathrm{tol}. \tag*{(31)}

A small correction alone does not certify a small nonlinear residual. In particular, if the accepted state has scalar energy residual rEm+1:=Wm+1−Wm+τQkm+1r_E^{m+1}:=W^{m+1}-W^m+\tau Q_k^{m+1}, then

Wm+1−Wm=−τQkm+1+rEm+1.(32)W^{m+1}-W^m=-\tau Q_k^{m+1}+r_E^{m+1}. \tag*{(32)}

Consequently, a sufficient acceptance condition for dissipation is rEm+1≤τQkm+1r_E^{m+1}\le\tau Q_k^{m+1}. For example, ∣rEm+1∣≤ητQkm+1|r_E^{m+1}|\le\eta\tau Q_k^{m+1} with 0≤η<10\le\eta<1 gives Wm+1−Wm≤−(1−η)τQkm+1W^{m+1}-W^m\le-(1-\eta)\tau Q_k^{m+1}. When Qkm+1=0Q_k^{m+1}=0, a positive absolute tolerance alone cannot guarantee nonincrease. More generally,

∣WM−Wm0+τ∑m=m0M−1Qkm+1∣≤∑m=m0M−1∣rEm+1∣.\left|W^M-W^{m_0}+\tau\sum_{m=m_0}^{M-1}Q_k^{m+1}\right| \le\sum_{m=m_0}^{M-1}|r_E^{m+1}|.

Thus a per-step energy residual O(τk+1)O(\tau^{k+1}) contributes at most O(τk)O(\tau^k) to the accumulated balance over a fixed time interval. This controls the balance error, not the geometric solution error. These estimates motivate monitoring the energy residual alongside the correction size, with residual scaling, tolerances, and iteration limits specified for each implementation. Control of the geometric solution error additionally requires bounds on the other residuals.

Extensions to other geometric flows

The construction extends by changing the geometric constitutive relation and the dissipation functional. To display the common structure, for a nonnegative functional dΓ(u)d_{\Gamma}(u) define

Qm[d,u]=16[dΓ~m+1(um+1)+4dΓm+1/2(um+1/2)+dΓm(um)].(33)\mathcal{Q}_m[d,u]=\frac{1}{6}\left[d_{\widetilde{\Gamma}_{m+1}}(u^{m+1})+4d_{\Gamma_{m+1/2}}(u^{m+1/2})+d_{\Gamma_m}(u^m)\right]. \tag*{(33)}

The endpoint uses the predicted metric, whereas the history and midpoint use their computed geometries. The following ES-BDF3 systems share this quadrature and the staggered construction of § 2. Nondegenerate evaluation geometries are assumed.

Curve diffusion

We first consider curve diffusion, the H−1H^{-1}-gradient flow of the perimeter W(t)=L(Γ(t))W(t)=L(\Gamma(t)). Its normal velocity is

V=∂ssκ,V=\partial_{ss}\kappa,

where κ\kappa denotes the curvature. The corresponding Lagrange-multiplier formulation is

∂tX⋅n=∂ssκ+λ(t)κ,κn=−∂ssX,dWdt=−∫Γ(t)∣∂sκ∣2 ds.(34)\begin{aligned} \partial_t\mathbf{X}\cdot\mathbf{n}&=\partial_{ss}\kappa+\lambda(t)\kappa,\\ \kappa\mathbf{n}&=-\partial_{ss}\mathbf{X},\\ \frac{\mathrm{d}W}{\mathrm{d}t}&=-\int_{\Gamma(t)}|\partial_s\kappa|^2\,\mathrm{d}s. \tag*{(34)} \end{aligned}

Using the perimeter variation dW/dt=(κ,V)Γ(t)\mathrm{d}W/\mathrm{d}t=(\kappa,V)_{\Gamma(t)} and the first equation in (34) gives

λ(t)∫Γ(t)κ2 ds=0.\lambda(t)\int_{\Gamma(t)}\kappa^2\,\mathrm{d}s=0.

Hence λ(t)=0\lambda(t)=0 whenever κ≢0\kappa\not\equiv0, and the augmented formulation is equivalent to the classical curve-diffusion equation at the continuous level. Define dC,Γ(u)=∥∂su∥L2(Γ)2d_{C,\Gamma}(u)=\lVert\partial_su\rVert^2_{L^2(\Gamma)}. Given Γ~m+1\widetilde{\Gamma}^{m+1}, the ES-BDF3 scheme seeks

(Xm+1,κm+1,λm+1)∈[Vh]2×Vh×R(\mathbf{X}^{m+1},\kappa^{m+1},\lambda^{m+1})\in[V^h]^2\times V^h\times\mathbb{R}

such that, for all (φh,ωh)∈Vh×[Vh]2(\varphi^h,\omega^h)\in V^h\times[V^h]^2,

(Dt3Xm+1τ,φhnm+1)Γ~m+1h+(∂sκm+1,∂sφh)Γ~m+1−λm+1(κm+1,φh)Γ~m+1h=0(35)\left(\frac{D_t^3\mathbf{X}^{m+1}}{\tau},\varphi^h\mathbf{n}^{m+1}\right)_{\widetilde{\Gamma}_{m+1}}^h+\left(\partial_s\kappa^{m+1},\partial_s\varphi^h\right)_{\widetilde{\Gamma}_{m+1}}-\lambda^{m+1}\left(\kappa^{m+1},\varphi^h\right)_{\widetilde{\Gamma}_{m+1}}^h=0 \tag*{(35)}
(κm+1nm+1,ωh)Γ~m+1h−(∂sXm+1,∂sωh)Γ~m+1h=0(36)\left(\kappa^{m+1}\mathbf{n}^{m+1},\omega^h\right)_{\widetilde{\Gamma}_{m+1}}^h-\left(\partial_s\mathbf{X}^{m+1},\partial_s\omega^h\right)_{\widetilde{\Gamma}_{m+1}}^h=0 \tag*{(36)}
Wm+1−Wm+τQm[dC,κ]=0.(37)W^{m+1}-W^m+\tau\mathcal{Q}_m[d_C,\kappa]=0. \tag*{(37)}

Since Qm[dC,κ]≥0\mathcal{Q}_m[d_C,\kappa]\geq0, the last equation gives Wm+1≤WmW^{m+1}\leq W^m. Testing the first equation with φh=1\varphi^h=1 yields

(Dt3Xm+1τ,nm+1)Γ~m+1h=λm+1(κm+1,1)Γ~m+1h,\left(\frac{D_t^3\mathbf{X}^{m+1}}{\tau},\mathbf{n}^{m+1}\right)_{\widetilde{\Gamma}_{m+1}}^h=\lambda^{m+1}\left(\kappa^{m+1},1\right)_{\widetilde{\Gamma}_{m+1}}^h,

which is not an exact discrete enclosed-area conservation law. An additional area constraint, as in [16], is needed if exact conservation is required.

Anisotropic curve diffusion

We next consider anisotropic curve diffusion driven by the anisotropic perimeter

Wγ(Γ)=∫Γγ(θ) ds,W_\gamma(\Gamma)=\int_{\Gamma}\gamma(\theta)\,\mathrm{d}s,

where θ\theta is the angle of the outward normal measured counterclockwise from the positive first coordinate axis. We assume a positive, 2π2\pi-periodic energy density γ∈C2\gamma\in C^2. The unregularized smooth-flow setting below uses positive surface stiffness γ+γ′′>0\gamma+\gamma^{\prime\prime}>0. The normal velocity and anisotropic chemical potential are

V=∂ssμ,μ=(γ(θ)+γ′′(θ))κ.V=\partial_{ss}\mu,\qquad\mu=(\gamma(\theta)+\gamma^{\prime\prime}(\theta))\kappa.

For γ≡1\gamma\equiv1, one has μ=κ\mu=\kappa, and the formulation reduces to isotropic curve diffusion. With the sign convention used here, the first variation satisfies

ddtWγ(Γ(t))=(μ,V)Γ(t).\frac{\mathrm{d}}{\mathrm{d}t}W_\gamma(\Gamma(t))=(\mu,V)_{\Gamma(t)}.

For a polygonal curve Γm\Gamma^m, we define the corresponding discrete anisotropic energy by

Wγm:=Wγ(Γm)=∑j=1Nγ(θjm)∣hjm∣,W_\gamma^m:=W_\gamma(\Gamma^m)=\sum_{j=1}^{N}\gamma(\theta_j^m)\lvert\mathbf{h}_j^m\rvert,

where θjm\theta_j^m denotes the orientation angle of the outward unit normal on the jjth edge of Γm\Gamma^m.

For Newton iteration, the energy row must differentiate this anisotropic energy. At the current polygon,

δWγ(X)[Y]=∑j(γ(θj)tj−γ′(θj)nj)⋅(Yj−Yj−1).(38)\delta W_\gamma(\mathbf{X})[\mathbf{Y}]=\sum_j\left(\gamma(\theta_j)\mathbf{t}_j-\gamma^{\prime}(\theta_j)\mathbf{n}_j\right)\cdot(\mathbf{Y}_j-\mathbf{Y}_{j-1}). \tag*{(38)}

Indeed, δℓj=tj⋅δhj\delta\ell_j=\mathbf{t}_j\cdot\delta\mathbf{h}_j and δθj=−nj⋅δhj/ℓj\delta\theta_j=-\mathbf{n}_j\cdot\delta\mathbf{h}_j/\ell_j. The energy derivative includes both edge-length and angle variations; the coefficients in the predicted-geometry equations remain fixed. Following [23], we use the conservative representation

μn=−∂s(G(θ)∂sX),G(θ)=(γ(θ)−γ′(θ)γ′(θ)γ(θ)).\mu\mathbf{n}=-\partial_s(G(\theta)\partial_s\mathbf{X}),\qquad G(\theta)= \begin{pmatrix} \gamma(\theta) & -\gamma^{\prime}(\theta)\\ \gamma^{\prime}(\theta) & \gamma(\theta) \end{pmatrix}.

The augmented continuous formulation is then

∂tX⋅n=∂ssμ+λ(t)μ,μn=−∂s(G(θ)∂sX),dWγdt=−∫Γ(t)∣∂sμ∣2 ds.(39)\begin{aligned} \partial_t\mathbf{X}\cdot\mathbf{n}&=\partial_{ss}\mu+\lambda(t)\mu,\\ \mu\mathbf{n}&=-\partial_s(G(\theta)\partial_s\mathbf{X}),\\ \frac{\mathrm{d}W_\gamma}{\mathrm{d}t}&=-\int_{\Gamma(t)}\lvert\partial_s\mu\rvert^2\,\mathrm{d}s. \tag*{(39)} \end{aligned}

Combining the first-variation identity with (5.5) gives

λ(t)∫Γ(t)μ2 ds=0.\lambda(t)\int_{\Gamma(t)}\mu^2\,\mathrm{d}s=0.

and therefore λ(t)=0\lambda(t)=0 whenever μ≠0\mu\ne0.

Using dC,Γ(μ)=∥∂sμ∥L2(Γ)2d_{C,\Gamma}(\mu)=\lVert\partial_s\mu\rVert_{L^2(\Gamma)}^2 and the predicted geometry, the ES-BDF3 equations are, for every (ψh,wh)∈Vh×[Vh]2(\psi^h,w^h)\in V^h\times[V^h]^2,

(Dt3Xm+1τ,n~m+1ψh)Γ~m+1h+(∂sμm+1,∂sψh)Γ~m+1h−λm+1(μm+1,ψh)Γ~m+1h=0,(40)\left(\frac{D_t^3\mathbf{X}^{m+1}}{\tau},\widetilde{\mathbf{n}}^{m+1}\psi^h\right)_{\widetilde{\Gamma}_{m+1}}^h+\left(\partial_s\mu^{m+1},\partial_s\psi^h\right)_{\widetilde{\Gamma}_{m+1}}^h-\lambda^{m+1}\left(\mu^{m+1},\psi^h\right)_{\widetilde{\Gamma}_{m+1}}^h=0, \tag*{(40)}
(μm+1n~m+1,wh)Γ~m+1h−(G(θ~m+1)∂sXm+1,∂swh)Γ~m+1h=0,(41)\left(\mu^{m+1}\widetilde{\mathbf{n}}^{m+1},w^h\right)_{\widetilde{\Gamma}_{m+1}}^h-\left(G(\widetilde{\theta}^{m+1})\partial_s\mathbf{X}^{m+1},\partial_sw^h\right)_{\widetilde{\Gamma}_{m+1}}^h=0, \tag*{(41)}
Wγm+1−Wγm+τQm[dC,μ]=0.(42)W_\gamma^{m+1}-W_\gamma^m+\tau\mathcal{Q}_m[d_C,\mu]=0. \tag*{(42)}

Again, the positivity of the Simpson weights gives Wγm+1≤WγmW_\gamma^{m+1}\le W_\gamma^m for any solution of (5.6).

Remark 5.1 (Strong anisotropic cases). The constitutive relation can also be expressed using the symmetrized matrix of [1, 22]. To distinguish the stabilization parameter from the BDF order, we write this matrix as Zσ(n~m+1)Z_\sigma(\widetilde{\mathbf{n}}^{m+1}) in place of G(θ~m+1)G(\widetilde{\theta}^{m+1}) in (5.6b). The scalar energy-increment argument still applies to any solution with nonnegative dissipation. If γ+γ′′\gamma+\gamma^{\prime\prime} becomes negative in some orientations, however, algebraic energy decay and matrix symmetrization alone do not establish well-posedness of the unregularized continuum flow or convergence to it. Such a regime requires a specified regularization or solution concept and is not discussed here.

Initialization. We extend the least-squares construction of [20 Remark 2.1] to the anisotropic chemical potential μ\mu. For the initial polygon Γ0\Gamma^0, the resulting normal equations are

(N0TN0)μ0=N0TA0X0,(N_0^{\mathsf{T}}N_0)\mu^0=N_0^{\mathsf{T}}A_0\mathbf{X}^0,

where

N0=((φi,(n0)[1]φj)Γ0h(φi,(n0)[2]φj)Γ0h),X0=(x0y0),N_0=\begin{pmatrix}(\varphi_i,(\mathbf{n}^0)^{[1]}\varphi_j)_{\Gamma^0}^h\\(\varphi_i,(\mathbf{n}^0)^{[2]}\varphi_j)_{\Gamma^0}^h\end{pmatrix},\qquad\mathbf{X}^0=\begin{pmatrix}\mathbf{x}^0\\\mathbf{y}^0\end{pmatrix},

and

A0=(Aγ−Aγ′Aγ′Aγ),(Aη)ij=(η(θ0)∂sφj,∂sφi)Γ0.A_0=\begin{pmatrix}A_\gamma&-A_{\gamma'}\\A_{\gamma'}&A_\gamma\end{pmatrix},\qquad(A_\eta)_{ij}=(\eta(\theta^0)\partial_s\varphi_j,\partial_s\varphi_i)_{\Gamma^0}.

Here {φi}i=1N\{\varphi_i\}_{i=1}^N is the standard Lagrange basis on II, a[1]\mathbf{a}^{[1]} and a[2]\mathbf{a}^{[2]} denote the two components of a∈R2\mathbf{a}\in\mathbb{R}^2, and xj0=(X0)[1](ρj)x_j^0=(\mathbf{X}^0)^{[1]}(\rho_j) and yj0=(X0)[2](ρj)y_j^0=(\mathbf{X}^0)^{[2]}(\rho_j). Equivalently, μ0\mu^0 minimizes ∥N0μ−A0X0∥22\lVert N_0\mu-A_0\mathbf{X}^0\rVert_2^2. Uniqueness requires N0N_0 to have full column rank, or equivalently nonzero mass-weighted nodal normals. If the system is ill conditioned, a QR or singular-value decomposition avoids explicitly forming the normal equations. This supplies the initial chemical potential.

Surface evolution in three dimensions

We next consider closed surfaces in R3\mathbb{R}^3. The curvature relation is expressed in terms of the mean curvature H\mathcal{H}, and surface gradients replace arc-length derivatives. We first introduce the discrete surface notation used below.

At time tm=mτt_m=m\tau, the surface Γ(tm)\Gamma(t_m) is approximated by a polyhedral surface

Γm=⋃j=1Jσjm,σjm=△{qj1m,qj2m,qj3m},\Gamma^m=\bigcup_{j=1}^J\sigma_j^m,\qquad\sigma_j^m=\triangle\{\mathbf{q}_{j1}^m,\mathbf{q}_{j2}^m,\mathbf{q}_{j3}^m\},

consisting of nondegenerate triangles without hanging vertices, with the vertices of each triangle ordered consistently so that their cross product points outward. The continuous piecewise linear finite element space on Γm\Gamma^m is

Whm:={u∈C(Γm):u∣σjm is affine for all j=1,…,J}⊂H1(Γm).W_h^m := \left\{u \in C(\Gamma^m): u|_{\sigma_j^m}\text{ is affine for all }j=1,\ldots,J\right\} \subset H^1(\Gamma^m).

For scalar/vector piecewise continuous functions u,vu,v defined on Γm\Gamma^m, we denote

(u,v)Γm:=∫Γmu⋅v dA,(u,v)Γmh:=13∑j=1J∣σjm∣∑k=13(u⋅v)((qjkm)−).(u,v)_{\Gamma^m}:=\int_{\Gamma^m}u\cdot v\,dA,\qquad(u,v)_{\Gamma^m}^h:=\frac{1}{3}\sum_{j=1}^{J}|\sigma_j^m|\sum_{k=1}^{3}(u\cdot v)((q_{jk}^m)^-).

where (qjkm)−(q_{jk}^m)^- denotes the limit taken from within the triangle σjm\sigma_j^m. The outer unit normal on each triangle is given by

nm∣σjm=njm:=(qj2m−qj1m)×(qj3m−qj1m)∣(qj2m−qj1m)×(qj3m−qj1m)∣.\left.\mathbf{n}^m\right|_{\sigma_j^m}=\mathbf{n}_j^m:=\frac{(q_{j2}^m-q_{j1}^m)\times(q_{j3}^m-q_{j1}^m)}{\left|(q_{j2}^m-q_{j1}^m)\times(q_{j3}^m-q_{j1}^m)\right|}.

and the discrete surface gradient ∇Γmf\nabla_{\Gamma^m}f of f∈Whmf\in W_h^m is constant on each triangle and defined by

(∇Γmf)∣σjm=njm×(f(qj1m)(qj3m−qj2m)+f(qj2m)(qj1m−qj3m)+f(qj3m)(qj2m−qj1m))2∣σjm∣.\left.(\nabla_{\Gamma^m}f)\right|_{\sigma_j^m}=\frac{\mathbf{n}_j^m\times\left(f(q_{j1}^m)(q_{j3}^m-q_{j2}^m)+f(q_{j2}^m)(q_{j1}^m-q_{j3}^m)+f(q_{j3}^m)(q_{j2}^m-q_{j1}^m)\right)}{2|\sigma_j^m|}.

The discrete surface area of Γm\Gamma^m is denoted by

Wm:=Sm:=∫Γm1 dA.W^m:=S^m:=\int_{\Gamma^m}1\,dA.

For a nondegenerate current polyhedral surface Γ\Gamma with vertex vector X\mathbf{X}, its area variation in direction Y\mathbf{Y} is

δS(X)[Y]=∫Γ∇ΓId⁡:∇ΓY dA.(43)\delta S(\mathbf{X})[\mathbf{Y}]=\int_{\Gamma}\nabla_{\Gamma}\operatorname{Id}:\nabla_{\Gamma}\mathbf{Y}\,dA. \tag*{(43)}

Here Y\mathbf{Y} is identified with its field on the current surface. This is the surface analogue of (4.1) for the Newton energy row; the gradients in the dissipation term remain on the frozen predictor.

Mean curvature flow and surface diffusion. Let H\mathcal{H} denote the sum of the two principal curvatures (rather than their average), with the sign convention

Hn=−ΔΓId⁡.\mathcal{H}\mathbf{n}=-\Delta_{\Gamma}\operatorname{Id}.

For mean curvature flow, V=−HV=-\mathcal{H}, and the surface area W(t)=S(Γ(t))W(t)=S(\Gamma(t)) satisfies

dWdt=−∫Γ(t)∣H∣2 dA.\frac{dW}{dt}=-\int_{\Gamma(t)}|\mathcal{H}|^2\,dA.

The augmented velocity is V=−(1−λ)HV=-(1-\lambda)\mathcal{H}, with the same curvature relation and area-dissipation law. The first variation dW/dt=(H,V)ΓdW/dt=(\mathcal{H},V)_{\Gamma} then gives λ∥H∥L2(Γ)2=0\lambda\|\mathcal{H}\|_{L^2(\Gamma)}^2=0, so the augmentation reduces to mean curvature flow whenever H≢0\mathcal{H}\not\equiv0. For mean curvature flow, set dM,Γ(u)=(u,u)Γhd_{M,\Gamma}(u)=(u,u)_{\Gamma}^h in (5.1). All surfaces share a fixed mesh connectivity. The map Xm+1\mathbf{X}^{m+1} and historical position fields are represented on the predicted surface by nodal identification; scalar test functions belong to its continuous piecewise affine space, and vector test functions to its three-component counterpart.

(Dt3Xm+1τ,φh⋅n~m+1)Γ~m+1h+(1−λm+1)(Hm+1,φh)Γ~m+1h=0,(44)\left(\frac{D_t^3\mathbf{X}^{m+1}}{\tau},\boldsymbol{\varphi}^h\cdot\widetilde{\mathbf{n}}^{m+1}\right)_{\widetilde{\Gamma}^{m+1}}^h+(1-\lambda^{m+1})\left(\mathcal{H}^{m+1},\boldsymbol{\varphi}^h\right)_{\widetilde{\Gamma}^{m+1}}^h=0, \tag*{(44)}
(Hm+1,n~m+1⋅ωh)Γ~m+1h−(∇Γ~m+1Xm+1,∇Γ~m+1ωh)Γ~m+1h=0,(45)\left(\mathcal{H}^{m+1},\widetilde{\mathbf{n}}^{m+1}\cdot\boldsymbol{\omega}^h\right)_{\widetilde{\Gamma}^{m+1}}^h-\left(\nabla_{\widetilde{\Gamma}^{m+1}}\mathbf{X}^{m+1},\nabla_{\widetilde{\Gamma}^{m+1}}\boldsymbol{\omega}^h\right)_{\widetilde{\Gamma}^{m+1}}^h=0, \tag*{(45)}
Wm+1−Wm+τQm[dM,H]=0.(46)W^{m+1}-W^m+\tau\mathcal{Q}_m[d_M,\mathcal{H}]=0. \tag*{(46)}

The nonnegativity of Qm[dM,H]\mathcal{Q}_m[d_M,\mathcal{H}] again gives Wm+1≤WmW^{m+1}\leq W^m.

For surface diffusion, the normal velocity and surface-area dissipation are

V=ΔΓH,dWdt=−∫Γ(t)∣∇ΓH∣2 dA.V=\Delta_{\Gamma}\mathcal{H},\qquad\frac{\mathrm{d}W}{\mathrm{d}t}=-\int_{\Gamma(t)}\lvert\nabla_{\Gamma}\mathcal{H}\rvert^2\,\mathrm{d}A.

The corresponding augmented velocity equation is

∂tX⋅n=ΔΓH+λ(t)H.\partial_t\mathbf{X}\cdot\mathbf{n}=\Delta_{\Gamma}\mathcal{H}+\lambda(t)\mathcal{H}.

Set dS,Γ(u)=∥∇Γu∥L2(Γ)2d_{S,\Gamma}(u)=\lVert\nabla_{\Gamma}u\rVert_{L^2(\Gamma)}^2. The ES-BDF3 surface-diffusion system follows from (5.3) by replacing κ\kappa with H\mathcal{H}, ∂s\partial_s with ∇Γ\nabla_{\Gamma}, and dCd_C with dSd_S, using the surface finite element spaces above.

Numerical experiments

We examine geometric accuracy, energy decay, mesh distribution, and nonlinear iteration counts for the numerical schemes presented in Sections 3 and 5. Shrinking circles and spheres provide exact smooth benchmarks.

Curve evolution

We consider CSF, curve diffusion (CDF), and anisotropic curve diffusion. Since tangential motion may lead to different parametrizations of the same geometric curve, we measure geometric differences using the manifold distance. For two simple closed curves Γ1\Gamma_1 and Γ2\Gamma_2 enclosing bounded regions Ω1\Omega_1 and Ω2\Omega_2, respectively, the manifold distance is defined by [20]

M(Γ1,Γ2):=∣(Ω1∖Ω2)∪(Ω2∖Ω1)∣=∣Ω1∣+∣Ω2∣−2∣Ω1∩Ω2∣.\mathcal{M}(\Gamma_1,\Gamma_2):=\lvert(\Omega_1\setminus\Omega_2)\cup(\Omega_2\setminus\Omega_1)\rvert=\lvert\Omega_1\rvert+\lvert\Omega_2\rvert-2\lvert\Omega_1\cap\Omega_2\rvert.

When no exact solution is available, we use the Cauchy difference

EM,i:=EM(T,τi,τi+1)=M(ΓT/τi,ΓT/τi+1).\mathcal{E}_{\mathcal{M},i}:=\mathcal{E}_{\mathcal{M}}(T,\tau_i,\tau_{i+1})=\mathcal{M}\left(\Gamma^{T/\tau_i},\Gamma^{T/\tau_{i+1}}\right).

Under an asymptotic expansion in which the successive geometric differences are proportional to ∣τipi−τi+1pi∣\lvert\tau_i^{p_i}-\tau_{i+1}^{p_i}\rvert, an observed order pi>0p_i>0 can be estimated, when a positive solution exists, from

EM,iEM,i+1=τipi−τi+1piτi+1pi−τi+2pi.(47)\frac{\mathcal{E}_{\mathcal{M},i}}{\mathcal{E}_{\mathcal{M},i+1}}=\frac{\tau_i^{p_i}-\tau_{i+1}^{p_i}}{\tau_{i+1}^{p_i}-\tau_{i+2}^{p_i}}. \tag*{(47)}

For positive differences and τi>τi+1>τi+2>0\tau_i>\tau_{i+1}>\tau_{i+2}>0, a positive root exists uniquely precisely when

EM,iEM,i+1>log⁡(τi/τi+1)log⁡(τi+1/τi+2).\frac{\mathcal{E}_{\mathcal{M},i}}{\mathcal{E}_{\mathcal{M},i+1}}>\frac{\log(\tau_i/\tau_{i+1})}{\log(\tau_{i+1}/\tau_{i+2})}.

The right-hand side of (47) is strictly increasing in p>0p>0, with the displayed threshold as its p↓0p\downarrow0 limit. Zero differences or failure of this condition make the positive-order estimate undefined. For a constant refinement ratio r=τi/τi+1>1r=\tau_i/\tau_{i+1}>1, this reduces to

pi=log⁡(EM,i/EM,i+1)log⁡(τi/τi+1).p_i=\frac{\log(\mathcal{E}_{\mathcal{M},i}/\mathcal{E}_{\mathcal{M},i+1})}{\log(\tau_i/\tau_{i+1})}.

In the experiments below, the spatial and temporal resolutions are generally refined simultaneously. The resulting values of pip_i should therefore be interpreted as observed orders along the specified coupled space–time refinement paths, rather than as estimates obtained with a fixed spatial discretization. In particular, along τ=Ch2/k\tau=Ch^{2/k}, a second-order spatial error is also O(τk)O(\tau^k); a slope close to kk on this path does not isolate the temporal order. The symmetric-difference area measures geometric agreement, not an error in a particular parametrization.

At the continuous level, the Lagrange multiplier vanishes for the flows considered here under the nondegeneracy conditions stated in § 4. We therefore use

Eλ(T,τi)=∣λT/τi∣,pλ,i=log⁡(Eλ(T,τi)/Eλ(T,τi+1))log⁡(τi/τi+1).\mathcal{E}_{\lambda}(T,\tau_i)=\left|\lambda^{T/\tau_i}\right|,\qquad p_{\lambda,i}=\frac{\log\left(\mathcal{E}_{\lambda}(T,\tau_i)/\mathcal{E}_{\lambda}(T,\tau_{i+1})\right)}{\log(\tau_i/\tau_{i+1})}.

as a diagnostic of the discrete multiplier and its decay under refinement.

We also monitor the normalized energy W(t)/W(0)W(t)/W(0), the number of Newton iterations, and the mesh ratio

Ψm=max⁡1≤j≤N∣hjm∣min⁡1≤j≤N∣hjm∣.\Psi^m=\frac{\max_{1\le j\le N}|h_j^m|}{\min_{1\le j\le N}|h_j^m|}.

These quantities respectively describe the discrete energy evolution, nonlinear iteration behavior, and mesh distribution.

Example 1 (Curve-shortening flow, CSF). The initial polygon has equally spaced vertices on the unit circle. The exact solution is

Xex(ρ,t)=1−2t(cos⁡(2πρ),sin⁡(2πρ)),κex=11−2t,ρ∈I,t∈[0,0.5).\mathbf{X}_{\mathrm{ex}}(\rho,t)=\sqrt{1-2t}\bigl(\cos(2\pi\rho),\sin(2\pi\rho)\bigr),\qquad\kappa_{\mathrm{ex}}=\frac{1}{\sqrt{1-2t}},\qquad\rho\in I,\quad t\in[0,0.5).

For this example, define the geometric and curvature errors by

EM(T,τ)=M(ΓT/τ,Γex(T)),\mathcal{E}_{\mathcal{M}}(T,\tau)=\mathcal{M}(\Gamma^{T/\tau},\Gamma_{\mathrm{ex}}(T)),
Eκ(T,τ)=∥κT/τ−κex(T)∥∞.\mathcal{E}_{\kappa}(T,\tau)=\left\|\kappa^{T/\tau}-\kappa_{\mathrm{ex}}(T)\right\|_{\infty}.

For either error E\mathcal{E}, the observed order between two resolutions is

ord⁡(τ1,τ2)=log⁡(E(T,τ1)/E(T,τ2))log⁡(τ1/τ2).\operatorname{ord}(\tau_1,\tau_2)=\frac{\log(\mathcal{E}(T,\tau_1)/\mathcal{E}(T,\tau_2))}{\log(\tau_1/\tau_2)}.

Figure 2 shows the convergence results for the shrinking-circle benchmark. Under the refinements used in this experiment, the manifold-distance and curvature errors display the expected second-, third-, and fourth-order trends for ES-BDF2, ES-BDF3, and ES-BDF4, respectively. The multiplier magnitude Eλ=∣λT/τ∣\mathcal{E}_{\lambda}=|\lambda^{T/\tau}| also decreases at a high-order rate under refinement. The second row of Figure 2 gives the numerical performance of the scheme ES-BDF3. As shown in this figure, the perimeter decreases monotonically, consistent with Proposition 2.4; the multiplier remains close to zero, and the Newton iteration count always stays 3 during the evolution.

Convergence results and representative diagnostics for CSF with a unit-circle initial curve

Figure 2. Convergence results and representative diagnostics for CSF with a unit-circle initial curve. (a)–(c): log-log plots of EME_M, EκE_\kappa, and EλE_\lambda at T=0.25T = 0.25 for the ES-BDFkk schemes; (d)–(f): evolution of the normalized perimeter, Lagrange multiplier, and Newton iteration count for ES-BDF3 with N=1280N = 1280 and τ=1/160\tau= 1/160.

Example 2 (Curve diffusion flow, CDF). We next consider CDF starting from a 4×14 \times1 ellipse. Since no exact solution is available, convergence is assessed by Cauchy differences in the manifold distance along coupled space–time refinement paths. We set T=0.25T = 0.25 and use τ=0.05h\tau= 0.05h, τ=0.05h2/3\tau= 0.05h^{2/3}, and τ=0.05h1/2\tau= 0.05h^{1/2} for the second-, third-, and fourth-order schemes, respectively.

The first row of Figure 3 compares the observed convergence behavior. ES-BDF2 and BGN-LM-BDF2 exhibit comparable second-order trends. For the third-order schemes, ES-BDF3 and BGN/BDF3 display rates close to three along the reported refinement path. BGN-LM-BDF3 shows a lower observed rate on the coarser computations and approaches third-order behavior after the time-step coefficient is reduced from 0.05 to 0.01.

Convergence results and diagnostics for CDF starting from a $4 \times 1$ ellipse

Figure 3. Convergence results and diagnostics for CDF starting from a 4×14 \times1 ellipse. First row: Cauchy-type convergence tests for the reported BDFkk schemes at T=0.25T = 0.25; Second row: ES-BDF3 diagnostics with N=125N = 125 and τ=1/500\tau= 1/500 up to T=1T = 1: (d) normalized perimeter; (e) Lagrange multiplier; (f) Newton iteration count and mesh ratio Ψ(t)\Psi(t).

The same experiment also compares two treatments of the midpoint data in ES-BDF3. Replacing the staggered half-integer solve by a first-order half-step predictor produces larger errors over the tested refinement range. This numerical comparison supports the use of a higher-order midpoint reconstruction in the Simpson energy update. ES-BDF4 displays an approximately fourth-order trend along its reported refinement path. The second row of Figure 3 shows representative ES-BDF3 diagnostics. The normalized perimeter decreases monotonically, the multiplier remains close to zero, and the Newton iteration count and mesh ratio remain bounded over the reported time interval.

Example 3 (Anisotropic curve diffusion). We next consider anisotropic curve diffusion using the formulation introduced in § 4. The initial curve is an ellipse with semi-major axis 2 and semi-minor axis 1, and the anisotropic energy density is

γ(θ)=1+βcos⁡(4θ).\gamma(\theta) = 1 + \beta\cos(4\theta).

Since no exact solution is available, convergence is assessed by the generalized Cauchy procedure in (47). We take β=0.05\beta= 0.05 and use the coupled refinement paths τ=0.05h\tau= 0.05h, τ=0.05h2/3\tau= 0.05h^{2/3}, and τ=0.05h1/2\tau= 0.05h^{1/2} for ES-BDF2, ES-BDF3, and ES-BDF4, respectively. The convergence test is terminated at T=0.1T = 0.1. Here γ+γ′′=1−15βcos⁡(4θ)≥0.25>0\gamma+ \gamma^{\prime\prime} = 1 - 15\beta\cos(4\theta) \ge0.25 > 0, so the test lies in the weakly anisotropic regime.

Table 2 shows observed orders close to two and three for ES-BDF2 and ES-BDF3 over most of the reported refinement range. The ES-BDF4 results exhibit observed orders between approximately 3.5 and 3.7 over the tested range. The data therefore indicate convergence at a rate above three, but do not yet resolve a fourth-order asymptotic regime.

ES-BDF2ES-BDF3ES-BDF4
τ\tauEM\mathcal{E}_Morderτ\tauEM\mathcal{E}_Morderτ\tauEM\mathcal{E}_Morder
1/3201.555e−1–1/1805.390e−2–1/2003.071e−3–
1/6403.621e−22.101/3208.684e−32.981/2401.382e−33.72
1/12808.575e−32.081/5002.083e−32.961/2806.994e−43.71
1/25601.947e−32.141/7206.282e−43.021/3203.920e−43.59
1/51203.930e−42.311/9802.465e−42.731/3602.364e−43.52

Table 2. Convergence results for anisotropic curve diffusion with γ(θ)=1+βcos⁡(4θ)\gamma(\theta) = 1 + \beta\cos(4\theta), β=0.05\beta= 0.05, and T=0.1T = 0.1. The coupled space–time refinement paths are τ=0.05h\tau= 0.05h for ES-BDF2, τ=0.05h2/3\tau= 0.05h^{2/3} for ES-BDF3, and τ=0.05h1/2\tau= 0.05h^{1/2} for ES-BDF4.

For the longer simulation up to T=0.5T = 0.5, we use N=125N = 125 and τ=0.05h2/3\tau= 0.05h^{2/3}. Figure 4 shows the evolution of the curve under the anisotropic flow. Figure 5 reports the normalized anisotropic energy and the Lagrange multiplier. The discrete anisotropic energy decreases monotonically, while the multiplier remains close to zero over the reported interval, consistent with the continuous relation.

Evolution of the curve under anisotropic curve diffusion at $t=0$, $0.1$, $0.25$, and $0.5$

Figure 4. Evolution of the curve under anisotropic curve diffusion at t=0t=0, 0.10.1, 0.250.25, and 0.50.5.

Evolution of the normalized anisotropic energy and Lagrange multiplier

Figure 5. Evolution of the normalized anisotropic energy (left) and Lagrange multiplier (right) up to T=0.5T=0.5.

Surface evolution

To monitor the evolving triangulation, we use the two mesh-distribution indicators introduced in [21]. At t=tmt=t_m, they are defined by

rh(tm):=max⁡1≤j≤Jmax⁡{∥qj1m−qj2m∥,∥qj2m−qj3m∥,∥qj3m−qj1m∥}min⁡1≤j≤Jmin⁡{∥qj1m−qj2m∥,∥qj2m−qj3m∥,∥qj3m−qj1m∥},(48)r_h(t_m):=\frac{\max_{1\le j\le J}\max\left\{\lVert\mathbf{q}_{j1}^{m}-\mathbf{q}_{j2}^{m}\rVert,\lVert\mathbf{q}_{j2}^{m}-\mathbf{q}_{j3}^{m}\rVert,\lVert\mathbf{q}_{j3}^{m}-\mathbf{q}_{j1}^{m}\rVert\right\}}{\min_{1\le j\le J}\min\left\{\lVert\mathbf{q}_{j1}^{m}-\mathbf{q}_{j2}^{m}\rVert,\lVert\mathbf{q}_{j2}^{m}-\mathbf{q}_{j3}^{m}\rVert,\lVert\mathbf{q}_{j3}^{m}-\mathbf{q}_{j1}^{m}\rVert\right\}}, \tag*{(48)}

and

ra(tm):=max⁡1≤j≤J∣σjm∣min⁡1≤j≤J∣σjm∣.(49)r_a(t_m):=\frac{\max_{1\le j\le J}\left|\sigma_j^m\right|}{\min_{1\le j\le J}\left|\sigma_j^m\right|}. \tag*{(49)}

Thus rhr_h is the ratio of the largest to the smallest edge length, while rar_a is the corresponding ratio of triangle areas. These quantities provide simple diagnostics of edge-length and element-area variation during the evolution.

Example 4 (Mean curvature flow, MCF). We first consider MCF starting from the unit sphere. The exact solution remains spherical with radius

R(t)=1−4t,t∈[0,1/4).R(t)=\sqrt{1-4t},\qquad t\in[0,1/4).

We measure the geometric error by the maximal radial deviation

Erad(T,τ):=max⁡q∈ΓT/τ∣∥q∥2−R(T)∣,Order=log⁡(Erad(T,τ1)/Erad(T,τ2))log⁡(τ1/τ2).\mathcal{E}_{\mathrm{rad}}(T,\tau):=\max_{\mathbf{q}\in\Gamma_T/\tau}\left|\lVert\mathbf{q}\rVert_2-R(T)\right|,\qquad\mathrm{Order}=\frac{\log\left(\mathcal{E}_{\mathrm{rad}}(T,\tau_1)/\mathcal{E}_{\mathrm{rad}}(T,\tau_2)\right)}{\log(\tau_1/\tau_2)}.

Following [21], the initial triangulation contains J=93608J=93608 triangles and K=46806K=46806 vertices. The fine spatial resolution is used to reduce the contribution of the spatial discretization error in this temporal-refinement test. Observed temporal slopes are informative only over the range where the spatial error and solver error are smaller than the time-discretization error.

For the three methods tested in Figure 6(a), ES-BDF1, ES-BDF2, and ES-BDF3 show approximately first-, second-, and third-order trends, respectively. The surface area decreases monotonically, and the mesh indicators rh(t)r_h(t) and ra(t)r_a(t) remain bounded over the reported evolution.

Results for MCF starting from the unit sphere

Figure 6. Results for MCF starting from the unit sphere, for ES-BDF1, ES-BDF2, and ES-BDF3. (a) Maximal radial deviation EradE_{\mathrm{rad}}; (b) Triangle-area ratio ra(t)r_a(t); (c) Edge-length ratio rh(t)r_h(t); (d) Surface area W(t)W(t); (e) Lagrange multiplier; (f) Newton iteration count.

Figure 6(d)–(Figure 6f) compare the surface-area, multiplier, and Newton-iteration histories. The surface-area curves produced by the three schemes are close. The Lagrange multipliers remain small, with smaller magnitudes observed for the higher-order schemes in this test. The Newton iteration counts are also of comparable size for the three discretizations.

We next consider the torus

(x12+x22−R)2+x32=r2,(R,r)=(1,0.5),\left(\sqrt{x_1^2+x_2^2}-R\right)^2+x_3^2=r^2,\qquad(R,r)=(1,0.5),

using (J,K)=(2528,1264)(J,K)=(2528,1264) and τ=10−3\tau=10^{-3}. Figure 7 shows the numerical evolution toward the singular regime. The final pinch-off times for ES-BDF1, ES-BDF2, and ES-BDF3 are 0.145, 0.137, and 0.135, respectively, which are consistent with the numerical value 0.138 reported in [5].

Numerical evolution of the torus under MCF

Figure 7. Numerical evolution of the torus under MCF with (J,K)=(2528,1264)(J,K)=(2528,1264) and τ=10−3\tau=10^{-3}. First row: ES-BDF1 at t=0t=0, 0.10.1, 0.130.13, and 0.1450.145; second row: ES-BDF2 at t=0.1t=0.1, 0.130.13, and 0.1370.137; third row: ES-BDF3 at t=0.1t=0.1, 0.130.13, and 0.1350.135.

Example 5 (Surface diffusion flow, SDF). Finally, we consider SDF of an 8×1×18 \times1 \times1 cigar-like surface obtained by capping a cylinder of diameter 1 and length 7 with two hemispheres. The initial mesh contains (J,K)=(2640,1322)(J,K)=(2640,1322) elements and vertices, respectively, and we choose τ=10−3\tau=10^{-3}.

Figure 8 compares the evolutions produced by ES-BDF1, ES-BDF2, and ES-BDF3. The surface develops a progressively narrower neck and becomes nonconvex, and finally pinches off. The final pinch-off times are 0.257, 0.245, and 0.243 for ES-BDF1, ES-BDF2, and ES-BDF3, respectively. Compared to the pinch-off resolution, the

Numerical evolution of the 8 × 1 × 1 cigar-like surface under surface diffusion

Figure 8. Numerical evolution of the 8×1×18 \times1 \times1 cigar-like surface under surface diffusion with (J,K)=(2640,1322)(J,K)=(2640,1322) and τ=10−3\tau=10^{-3}. First row: ES-BDF1 at t=0t=0, 0.20.2, 0.240.24, and 0.2570.257; second row: ES-BDF2 at t=0.2t=0.2, 0.240.24, and 0.2450.245; third row: ES-BDF3 at t=0.2t=0.2, 0.240.24, and 0.2430.243.

values produced by ES-BDF2 and ES-BDF3 are much closer than to the value by ES-BDF1.

Figure 9 shows monotone area decay for all three schemes. The ES-BDF2 and ES-BDF3 energy curves remain close, while their separation from ES-BDF1 grows near the pinch-off time. The multiplier evolution shows a similar pattern, and the Newton iteration counts are comparable. Since the continuous Lagrange multiplier vanishes, its discrete magnitude keeps very close to 0 during the evolution, but exhibit abruptly increasing when the evolution approaches the pinch-off time, shown in Figure 9(b). The pinch-off time produced by ES-BDF3, 0.243, is consistent with the numerical value 0.2431 reported in the literature [5].

Numerical diagnostics for ES-BDFk applied to surface diffusion of the cigar-like surface

Figure 9. Numerical diagnostics for ES-BDFkk, k=1,2,3k=1,2,3, applied to surface diffusion of the cigar-like surface: (a) Normalized surface area; (b) Lagrange multiplier; (c) Newton iteration count.

Conclusions

We have developed energy-increment BDF methods for geometric flows within the BGN parametric finite element framework. The formulation replaces the multistep energy-derivative equation by a quadrature approximation of the integrated energy balance while retaining the BDF geometric equations. For BDF3 and BDF4, positive Simpson weights ensure decay of the original discrete geometric energy separately on each interleaved time grid for every nondegenerate solution satisfying the scalar constraint exactly. The analysis establishes conditional scalar consistency, a regularity criterion for the ES-BDF3 Newton system in curve shortening flow, and residual-based bounds on energy-balance errors. The smooth benchmarks exhibit accuracy trends consistent with the design orders under the reported refinements. The curve-diffusion comparison supports the importance of accurate midpoint data, while representative curve and surface tests document energy decay and mesh evolution. A complete solvability and convergence analysis of the coupled scheme remains to be established, requiring joint control of initialization, predicted geometry, midpoint data, and nonlinear solution errors.

References

  1. [1]W. Bao, W. Jiang, and Y. Li, A symmetrized parametric finite element method for anisotropic surface diffusion of closed curves, SIAM J. Numer. Anal., 61 (2023), pp. 617–641, https://doi.org/10.1137/22M1500575.
  2. [4]J. W. Barrett, H. Garcke, and R. Nürnberg, A parametric finite element method for fourth order geometric evolution equations, J. Comput. Phys., 222 (2007), pp. 441–467, https://doi.org/10.1016/j.jcp.2006.07.026.
  3. [5]J. W. Barrett, H. Garcke, and R. Nürnberg, On the parametric finite element approximation of evolving hypersurfaces in R³, J. Comput. Phys., 227 (2008), pp. 4281–4307, https://doi.org/10.1016/j.jcp.2007.11.023.
  4. [6]J. W. Barrett, H. Garcke, and R. Nürnberg, Parametric approximation of Willmore flow and related geometric evolution equations, SIAM J. Sci. Comput., 31 (2008), pp. 225–253, https://doi.org/10.1137/070700231.
  5. [7]J. W. Barrett, H. Garcke, and R. Nürnberg, Parametric finite element approximations of curvature-driven interface evolutions, in Handbook of numerical analysis, vol. 21, Elsevier, 2020, pp. 275–423, https://doi.org/10.1016/bs.hna.2019.05.002.
  6. [8]F. Cao, Geometric curve evolution and image processing, Springer Science & Business Media, 2003.DOI
  7. [9]F. Davi and M. E. Gurtin, On the motion of a phase interface by surface diffusion, Z. Angew. Math. Phys., 41 (1990), pp. 782–811, https://doi.org/10.1007/BF00945835.
  8. [11]B. DUAN, B. LI, AND Z. ZHANG, High-order fully discrete energy diminishing evolving surface finite element methods for a class of geometric curvature flows, Ann. Appl. Math., 37 (2021), pp. 405–436, https://doi.org/10.4208/aam.OA-2021-0007.
  9. [12]C. M. ELLIOTT AND B. STINNER, Modeling and computation of two phase geometric biomembranes using surface finite elements, J. Comput. Phys., 229 (2010), pp. 6585–6612.DOI
  10. [13]Z. FU, J. SHEN, AND J. YANG, Higher-order energy-decreasing exponential time differencing Runge–Kutta methods for gradient flows, Sci. China Math., 68 (2025), pp. 1727–1746, https://doi.org/10.1007/s11425-024-2337-3.
  11. [14]Z. FU, T. TANG, AND J. YANG, Energy diminishing implicit-explicit Runge–Kutta methods for gradient flows, Math. Comp., 93 (2024), pp. 2745–2767, https://doi.org/10.1090/mcom/4021.
  12. [16]H. GARCKE, W. JIANG, C. SU, AND G. ZHANG, Structure-preserving parametric finite element method for surface diffusion based on Lagrange multiplier approaches, SIAM J. Sci. Comput., 47 (2025), pp. A1983–A2011, https://doi.org/10.1137/24M1687546.
  13. [17]Y. GUO AND M. LI, Structure-preserving parametric finite element methods for anisotropic surface diffusion flow with minimal deformation formulation, Comput. Phys. Commun., 313 (2025), p. 109620, https://doi.org/10.1016/j.cpc.2025.109620.
  14. [19]F. HUANG AND J. SHEN, A new class of implicit-explicit BDFk SAV schemes for general dissipative systems and their error analysis, Comput. Methods Appl. Mech. Engrg., 392 (2022), p. 114718, https://doi.org/10.1016/j.cma.2022.114718.
  15. [20]W. JIANG, C. SU, AND G. ZHANG, A second-order in time, BGN-based parametric finite element method for geometric flows of curves, J. Comput. Phys., 514 (2024), p. 113220, https://doi.org/10.1016/j.jcp.2024.113220.
  16. [21]W. JIANG, C. SU, AND G. ZHANG, Stable backward differentiation formula time discretization of BGN-based parametric finite element methods for geometric flows, SIAM J. Sci. Comput., 46 (2024), pp. A2874–A2898, https://doi.org/10.1137/23M1625597.
  17. [22]W. JIANG, C. SU, G. ZHANG, AND L. ZHANG, Predictor-corrector, BGN-based parametric finite element methods for surface diffusion, J. Comput. Phys., 530 (2025), p. 113901, https://doi.org/10.1016/j.jcp.2025.113901.
  18. [23]Y. LI AND W. BAO, An energy-stable parametric finite element method for anisotropic surface diffusion, J. Comput. Phys., 446 (2021), p. 110658, https://doi.org/10.1016/j.jcp.2021.110658.
  19. [24]S. MA AND Q. RAO, High-order energy-stable BGN parametric finite element methods for geometric flows, arXiv preprint arXiv:2608.29877, (2026), https://doi.org/10.48550/arXiv.2608.29877.
  20. [25]J. SHEN, J. XU, AND J. YANG, The scalar auxiliary variable (SAV) approach for gradient flows, J. Comput. Phys., 353 (2018), pp. 407–416, https://doi.org/10.1016/j.jcp.2017.10.021.
  21. [26]J. SHEN, J. XU, AND J. YANG, A new class of efficient and robust energy stable schemes for gradient flows, SIAM Rev., 61 (2019), pp. 474–506, https://doi.org/10.1137/17M1150153.
  22. [27]G. ZHANG, B. D. ANDREWS, AND P. E. FARRELL, Arbitrary-order structure-preserving discretizations for geometric curvature flows, arXiv:2605.20371, 2026, https://doi.org/10.48550/arXiv.2605.20371.

Paper details

vlnc:W_a6xkz744
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