Abstract
We present and analyze a semi-discrete finite element scheme for a system consisting of a geometric evolution equation for a curve and a parabolic equation on the evolving curve. More precisely, curve shortening flow with a forcing term that depends on a field defined on the curve is coupled with a diffusion equation for that field. The scheme is based on ideas of Dziuk (SIAM J Numer Anal 36(6):1808–1830, 1999) for the curve shortening flow and Dziuk and Elliott (IMA J Numer Anal 27(2), 262–292, 2007) for the parabolic equation on the moving curve. Additional estimates are required in order to show convergence, most notably with respect to the length element: While in Dziuk (SIAM J Numer Anal 36(6):1808–1830, 1999) an estimate of its error was sufficient we here also need to estimate the time derivative of the error which arises from the diffusion equation. Numerical simulation results support the theoretical findings.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
We aim for approximating the following problem: Given a closed initial curve \(\Gamma _{0}\) and a function \(c_{0} : \Gamma _{0} \rightarrow \mathbb {R}\) find a moving closed curve \(\{ \Gamma (t) \}_{t \in [0,T]} \subset \mathbb {R}^2\) and a family of fields \(c(t) : \Gamma (t) \rightarrow \mathbb {R}\), \(t \in [0,T]\), such that
Here, s is an arc-length parameter of the actual curve \(\Gamma (t)\), v is the (scalar) velocity in the direction of a unit normal field \(\nu \), \(\kappa \) is the (scalar) curvature, \(f : \mathbb {R}\rightarrow \mathbb {R}\) is a coupling function, and \({\partial }^{\bullet }_{t}\) is the material derivative (\({\partial }^{\bullet }_{t}c = \partial _t c + v \partial _\nu c\) if c is smoothly extended away from \(\Gamma \)).
The system consisting of (1.1), (1.2) can fairly be regarded as the simplest system coupling a geometric evolution equation to an equation for a conserved field on the evolving manifold. We do not have any specific application in mind for (1.1), (1.2). But more sophisticated geometric evolution equations and parabolic PDEs on the moving manifold feature, for instance, in cell biology as an effective approach to cell motility [18, 23]. Problems in soft matter physics such as the relaxation dynamics of two-phase biomembranes can also be modeled by such type of systems [16, 17]. From a mathematical point of view, the evolution of pattern forming PDE systems on deforming surfaces is of general interest, for instance, see [25].
Working in a parametric setting we assume that the curves can be parametrized by a family of functions \(u(t) : S^1 \rightarrow \mathbb {R}^2\), i.e., \(\Gamma (t) = u(S^1,t)\). For the initial curve we write \(\Gamma _0 = u_0(S^1)\). By \((\cdot )^\perp \) we denote the counter-clockwise rotation by 90\(^{\circ }\) in \(\mathbb {R}^2\). We write \(\tau = u_{x}/|u_{x}|\) for a unit tangent field and assume that the orientation is such that \(\nu = u_{x}^\perp /|u_{x}|\). For convenience, the field c on the evolving curve will be denoted by c again after transformation to the parameter space. A strong formulation of the geometric equation (with no tangential velocity) in the parametric setting then is
while for the PDE on the evolving curve we obtain
In order to approximate the solution let \(Y_h\) denote a finite element space (details will be provided later on in Sect. 3) and let \(X_h = Y_h^2\). Then consider the problem of finding functions \(u_h(\cdot ,t) \in X_h\) and \(c_h(\cdot ,t) \in Y_h\), \(t\in [0,T]\), such that \(u_h(\cdot ,0) = u_{h0} := I_h(u_{0})\), \(c_h(\cdot ,0) = c_{h0} := I_h(c_0)\), and such that for all \(\varphi _{h} \in X_h\) and \(\zeta _{h} \in Y_h\) at almost all times \(t \in [0,T]\)
Here, \(I_h\) stands for the interpolation operator for both scalar and vector valued functions.
With regards to the Eq. (1.5) for c, the approximation by (1.7) is inspired by [13]. The resulting scheme is intrinsic in the sense that it does not require any knowledge about the parametrization but only the positions of the vertices that are given in terms of \(u_h\) (see Algorithm 6.1 below). However, for the numerical analysis we cannot resort to the methods in [13] because the moving curve \(\Gamma (t)\) is not explicitly given but by the solution u of the geometric equation (1.4). Its approximation by (1.6) is based on [10] where a scheme for two-dimensional surfaces is presented. The one-dimensional semi-discrete case but with anisotropic surface energy has been analyzed in [12] (evolution in a plane) and in [26] (higher co-dimension), see also [11] for the isotropic case. In addition, there is the forcing term \(f(c) u_{x}^\perp /|u_{x}|\) which is of lower order but, because of the c dependence, requires a coupling of the error estimates for u to those for c.
Regarding the estimate for c, the main difficulty arises from the term \(c |u_x|_t/|u_x|\) in (1.5). The error of the length element \(|u_x| - |u_{hx}|\) already had to be estimated in the \(L^\infty ([0,T],L^2(S^1))\) norm when proving convergence of the approximation to curve shortening flow in [12]. However, here we need an estimate for the time derivative of the length element \(|u_x|_t - |u_{hx}|_t\). The key observation is that \(|u_x|_t\) can be estimated in terms of the squared velocity and the length element, see (2.7) in Lemma 2.4 below. Mimicking these calculations for the error \(|u_x|_t - |u_{hx}|_t\) is the content of the novel Lemma 4.1 which subsequently proves sufficient to obtain suitable estimates for \(c - c_h\). Our results are summarized by:
Theorem 1.1
Under Assumption 2.2 there exists \(h_0>0\) such that for all \(0 < h \le h_0 \) there exists a unique solution \((u_h,c_h)\) of (1.6), (1.7), and the error between the smooth solution and the discrete solutions can be estimated as follows:
with a constant \(C>0\). The constant depends on the final time T, on the bounds \(\Vert f\Vert _{L^{\infty }(\mathbb {R})}\) and \(\Vert f'\Vert _{L^{\infty }(\mathbb {R})}\) of the coupling function, on the bounds \(\Vert u \Vert _{W^{1,\infty }([0,T],H^2(S^1))}\), \(\Vert c \Vert _{W^{1,\infty }([0,T],H^1(S^1))}\), and \(\Vert c \Vert _{L^\infty ([0,T],H^2(S^1))}\) of the solution (which includes the bounds \(\Vert u_0 \Vert _{H^2(S^1)}\) and \(\Vert c_0 \Vert _{H^1(S^1)}\) of the initial values), on the bound \(C^{**}\) from below of the length element, see (2.5) in Assumption 2.2, and on the constant \(\bar{C}\) ruling the grid regularity [cf. (3.1)].
Our proof follows the lines of [12] on anisotropic curve shortening flow though we should mention that for the isotropic curve shortening flow other ideas and techniques have also been used, for instance, see [8]. From a practical point of view, mesh degeneration is an important problem for long-time simulations. We will not address this issue here but for ideas to move vertices in tangential direction as appropriate we refer to [2, 3, 15, 22]. In [1] an additional forcing term is accounted for, see also [6] for analytical results on such a problem. Also with regards to PDEs on evolving surfaces there are other methods. For instance, in [24] a surface reconstruction is used which is based on a fixed bulk mesh and in [19] a grid based particle method. Of course, there are also other approaches to surfaces PDEs and geometric PDEs which are not based on any parametrization but on level sets, phase field, or other ideas. We here only refer to the overviews [9, 14].
We start with specifying the assumptions on the solution to the continuous problem and showing some properties in Sect. 2. After, we carefully describe the finite element approach and, proceeding analogously to the continuous case, show some properties of the semi-discrete solution. Section 4 then contains the technical estimates required for convergence which is stated in the section after. In the final section we report on numerical simulation results which support the findings.
2 The continuous problem
Here and in the following sections, constants which, in general, will vary from line to line in the various computations will be denoted by capital C. Moreover we occasionally use the abbreviation
The finite element approximation consisting of (1.6) and (1.7) emerges from the following weak formulation of the system (1.4) and (1.5):
Problem 2.1
(Weak problem) Find functions \(u : S^1 \times [0,T] \rightarrow \mathbb {R}^2\) and \(c : S^1 \times [0,T] \rightarrow \mathbb {R}\) such that \(u(\cdot ,0) = u_0\), \(c(\cdot ,0) = c_0\), and such that for all test functions \(\varphi : S^1 \rightarrow \mathbb {R}^2\) and \(\zeta : S^1 \rightarrow \mathbb {R}\) and almost all times \(t \in [0,T]\)
Note that if \(\zeta : S^1 \times [0,T] \rightarrow \mathbb {R}\) is a time dependent test function then (2.3) becomes
Clearly, we can not expect the flow to be eternal, since the flow might exhibit singularities in finite time (like the curve shortening flow). We thus make the following assumptions regarding existence, uniqueness, and regularity of the weak solution:
Assumption 2.2
Both f and its derivative \(f'\) are bounded,
There is a unique solution (u, c) of (2.2), (2.3) on the time interval [0, T] with initial values \(u(\cdot ,0)=u_0(\cdot ) \in H^2(S^1)\), \(c(\cdot ,0)=c_0(\cdot ) \in H^1(S^1)\) which satisfies
Moreover, there is a constant \(C^{**}>0\) such that
Remark 2.3
There is a huge literature on the curve shortening flow (and more generally on the mean curvature flow), see for instance [7, 21]. There are also results for curve shortening flow with a forcing term. For instance, in [5] it is shown that if f is smooth and the initial curve \(u_{0}\) is embedded then the maximal existence time of a smooth solution is bounded from below by a quantity that depends on the initial data and \(\Vert f\Vert _{L^{\infty }(\mathbb {R})}\). There do not seem to exist any results on short time well-posedness, regularity, and long-time behavior for our specific type of problem. However we count upon the standard methods for proving short-time well-posedness for parabolic systems to work thanks to the relatively nice elliptic second order structure of the spatial part of the differential operator. We leave these analytical questions for future studies and here focus on approximating the solution as it is postulated in the above Assumption 2.2.
From now on (u, c) will always denote the solution as specified above. Note that direct consequence of Assumption 2.2 is that
with a constant \(C>0\). Obviously then also \(\Vert c\Vert _{L^2([0,T],H^1(S^1))} \le C\) and (by embedding theory) \(\Vert c\Vert _{C([0,T],L^\infty (S^1))} \le C\) hold.
Although the bounds derived in the next lemma are implied by the regularity assumptions imposed on the continuous solution, the derived equations and methods of proof will be important to derive discrete analogues later on.
Lemma 2.4
-
1.
For the length element we have that
$$\begin{aligned} |u_x|_t&=-|u_t|^2 \,|u_x| + u_t \cdot r \,|u_x|. \end{aligned}$$(2.7) -
2.
Furthermore,
$$\begin{aligned} |u_x| \le C^{*} \end{aligned}$$(2.8)with a constant \(C^{*}>0\).
-
3.
For \({{\bar{t}}} \in [0,T]\) we have that
$$\begin{aligned} \int _0^{{\bar{t}}} |u_t|^2 dt \le C \quad \text{ and } \quad \int _0^{{\bar{t}}} |u_t - r|^2 dt \le C \quad \text{ on } S^1\,. \end{aligned}$$(2.9)
Proof
We have that
by (1.4), (2.1), and the fact that \(u_t\) is a normal vector. The second claim follows from the boundedness of f and a Gronwall argument applied to
Finally observe that from (2.7) we know that \( |u_t|^2 \le -\frac{|u_x|_t}{|u_x|} + |u_t| \, |r| \le - \frac{|u_x|_t}{|u_x|} + \frac{1}{2} |r|^2 + \frac{1}{2} |u_{t}|^2\,,\) whence
Integration, (2.5), and (2.8) gives the third claim. \(\square \)
3 Spatial discretization
Let \(S^1= \bigcup _{j=1}^N S_j\) be a decomposition of \(S^1\) into segments given by the nodes \(x_j\). We think of \(S_j\) as the interval \([x_{j-1}, x_{j}] \subset [0,2\pi ]\) for \(j=1, \ldots , N\). Here and in the following, indices related to the grid have to be considered modulo N. For instance, we identify \(x_0=x_N\). Let \(h_j=|S_j|\) and \(h=\max _{j=1,\ldots , N}h_j\) be the maximal diameter of a grid element. We assume that for some constant \(\bar{C}>0\) we have
Clearly the first inequality yields \(\bar{C}h_{j+1} \le h_j \le \frac{h_{j+1}}{\bar{C}} \). For a discretization of (2.2) we introduce the discrete finite dimensional spaces
of continuous periodic piecewise affine functions on the grid. The scalar nodal basis functions of \(Y_h\) are denoted by \(\varphi _j\), \(j=1, \ldots , N\), and defined by \(\varphi _j(x_i)=\delta _{ij}\).
For a continuous function \(v \in C^0 (S^1, \mathbb {R})\) let \(I_h v \in Y_h\) be the linear interpolant uniquely defined by \(I_h v(x_i)= v(x_i)\) for all \(i=1, \ldots ,N\). For convenience we also denote the interpolation onto \(X_h\) by \(I_h\). We shall use the standard interpolation estimates (both for scalar and vector valued functions):
Recall also the inverse estimates for any \(w_{h} \in Y_{h}\) and \(j=1, \ldots , N\):
Problem 3.1
(Semi-discrete scheme) Find functions \(u_h(\cdot ,t) \in X_h\) and \(c_h(\cdot ,t) \in Y_h\), \(t\in [0,T]\), of the form
with \(u_j(t) \in \mathbb {R}^2\) and \(c_j(t) \in \mathbb {R}\), such that \(u_h(\cdot ,0) = u_{h0} := I_h(u_0)\), \(c_h(\cdot ,0) = c_{h0} := I_h(c_0)\), and such that for all \(\varphi _{h} \in X_h\) and \(\zeta _{h} \in Y_h\) at almost all times \(t \in [0,T]\) (1.6) and (1.7) are satisfied.
Note that we may want to use a time dependent test function in the equation for \(c_h\) of the form
In analogy to (2.4) Eq. (1.7) then becomes
Recalling that indices referring to the grid always are understood modulo N, let
If we insert \(\varphi _j\), \(j=1, \ldots ,N\), separately for each component of \(\varphi _h\) in (1.6) then we get the following \(2 \times N\) ordinary differential equations:
and the initial values are given by \(u_j(0) = u_0(x_j)\), \(j=1,\ldots ,N\). With
and
we can rewrite the system (3.8) with the initial condition as
Define the piecewise constant function
A short calculation shows that another equivalent formulation to (1.6) is
Next we aim at giving the discrete equivalents of the results in Lemma 2.4.
Lemma 3.2
Let \({{\bar{t}}} \in (0,T]\) and assume that \((u_h,c_h)\) is a solution of (1.6), (1.7) for \(t \in [0,{{\bar{t}}}]\) such that \(q_j(t) >0\) for all \(j=1, \ldots , N\) and all \(t \in [0,{{\bar{t}}}]\).
-
1.
For \(j=1,\ldots ,N\) we have that
$$\begin{aligned} \dot{q}_j&= \tau _j \cdot (r_j -r_{j-1})-\frac{|\tau _{j+1}-\tau _j|^2 }{(q_j + q_{j+1})} -\frac{|\tau _{j-1}-\tau _j|^2 }{(q_j + q_{j-1})} \end{aligned}$$(3.13)$$\begin{aligned}&= \tau _j \cdot (r_j -r_{j-1})-\frac{(q_j + q_{j+1})}{4}|\dot{u}_{j}-r_j|^2 - \frac{(q_j + q_{j-1})}{4}|\dot{u}_{j-1}-r_{j-1}|^2 \,. \end{aligned}$$(3.14) -
2.
Furthermore with a constant \(C>0\)
$$\begin{aligned} \begin{array}{ll} \max _{1 \le j \le N} q_{j}(t) &{}\le C \max _{1 \le j \le N} q_{j}(0), \\ \Vert u_{hx}(\cdot ,t) \Vert _{L^\infty (S^1)} &{}\le C \Vert u_{h0x}(\cdot ) \Vert _{L^\infty (S^1)}. \end{array} \end{aligned}$$(3.15) -
3.
Moreover, there is a \(C>0\) such that
$$\begin{aligned} \int _0^{{\bar{t}}} (q_j + q_{j+1}) |\dot{u}_j-r_j|^2 dt \le C h, \quad \int _0^{{\bar{t}}} \frac{|\tau _{j\pm 1}-\tau _{j}|^{2}}{(q_{j}+q_{j\pm 1})} dt \le C h. \end{aligned}$$(3.16)
Proof
From the definition of \(q_j\) we obtain by differentiating in time
From the system (3.11) together with \(\tau _j \cdot (\tau _{j+1}-\tau _j)=-\frac{1}{2}|\tau _{j+1}-\tau _j|^2\) we infer that
Arguing similarly for the term \(\tau _j \cdot \dot{u}_{j-1}\) one obtains Eq. (3.13). Using (3.11) we can write \(\tau _{j+1}-\tau _j= \frac{q_j+q_{j+1}}{2} (\dot{u}_j-r_j)\) and (3.14) follows which proves the first assertion.
For the second assertion we set \(f_j := f(c_j)\) for simplicity. Note that by (3.9)
Since \(\frac{q_{j \pm 1}}{q_{j}+q_{j \pm 1}} \le 1\) we get that
and, similarly,
Therefore
Equations (3.13) and (3.17) with \(\epsilon =1\) yield that
Integrating with respect to t we infer that
Applying a Gronwall argument we obtain the first estimate of (3.15). The second one is a direct consequence of the first one thanks to (3.1).
From (3.14) and (3.17) we infer that
where we have used (3.11) in the last equality. Choosing \(\epsilon \) appropriately, integrating with respect to time, and using that \(q_{j}(t) = h_{j}|u_{hx}|_{|_{S_{j}}} \le Ch\) thanks to (3.15), we obtain the estimates (3.16). \(\square \)
4 Error estimates
In this section we prove some estimates that will enable us to show convergence of the semi-discrete solutions \((u_h, c_h)\) of (1.6), (1.7) to the solution (u, c) of the continuous problem as specified in Assumption 2.2. For this purpose let us assume that for \(h>0\) there is a unique solution \((u_h, c_h)\) for \(t \in [0,{{\bar{t}}}]\) with some \({{\bar{t}}} \in (0,T]\) (this question will be addressed at the beginning of the proof of Theorem 1.1 in Sect. 5).
We commence with some calculations for the error of the length element \(| u_{x} | - | u_{hx} |\) and show some preliminary estimates in Lemma 4.1. These are used to obtain an estimate of \(c - c_h\) in suitable norms, see Lemma 4.2. An estimate of \(u - u_h\) in suitable norms (see Lemma 4.3) follows the lines of [12] and involves an integral term of the error of the length element which we estimate last in Lemma 4.4.
For the convenience of the reader we recall the abbreviations
for the solution to the continuous problem while for the semi-discrete case we recall
By (2.7) and (3.14) we can write for each grid element \(S_j=[x_{j-1},x_j]\) the following equation:
Using (3.9) we can write
Observe that
by (3.11), so we can write
Similarly one can show that \(\frac{2}{q_{j}+q_{j-1}} \tau _{j} \cdot \nu _{j-1} = -\nu _{j} \cdot (r_{j-1} -\dot{u}_{j-1})\) whence
Let us also set
Lemma 4.1
Assume that \({{\bar{t}}} \in (0,T]\) is such that
Then there exists a constant C such that for any time \(t \in (0,{{\bar{t}}}]\) we have:
-
1.
On each \(S_{j}\) we can write
$$\begin{aligned} ||u_{x}|_{t}- |u_{hx}|_{t}| \le C \Big (|u_{t}-\dot{u}_{j}|^{2} + |u_{t}-\dot{u}_{j-1}|^{2} \Big ) + C L_{j}, \end{aligned}$$where
$$\begin{aligned} L_{j}&:=|c-c_j| + |c-c_{j-1}| + |\tau -\tau _j| + |\tau -\tau _{j-1}| + |\tau - \tau _{j+1}| \\&\qquad + \left| q-\frac{q_{j}}{h_{j}}\right| + \left| q-\frac{q_{j+1}}{h_{j+1}}\right| + \left| q-\frac{q_{j-1}}{h_{j-1}}\right| + \left| u_t - \dot{u}_j\right| + \left| u_t - \dot{u}_{j-1}\right| + h . \end{aligned}$$ -
2.
Moreover
$$\begin{aligned} \sum _{j=1}^{N}\int _{S_{j}}|u_{t}-\dot{u}_{j}|^{2} + |u_{t}-\dot{u}_{j-1}|^{2} dx \le C h^{2} + C \int _{S^1} |u_t -u_{ht}|^2 dx \end{aligned}$$(4.5)and
$$\begin{aligned} \sum _{j=1}^{N}\int _{S_{j}}|L_{j}|^{2} dx\le & {} C \int _{S^1}|c-c_h|^2 dx + C \int _{S^1} |\tau -\tau _h|^2 dx \nonumber \\&+ C \int _{S^1} (|u_{hx}|-|u_x|)^2 dx + C \int _{S^1} |u_t -u_{ht}|^2 dx + C h^2.\nonumber \\ \end{aligned}$$(4.6)
Proof
As we have assumed that \(2C^* \ge q_h \ge C^{**} /2\), the discrete length elements are comparable, in other words
Note that \(|f(c)-f(c_j)| \le \Vert f'\Vert _{L^\infty (\mathbb {R})}|c-c_j|\) and
(which follows by (3.1)). Thus, using (2.8), (4.7), and the bound \(|u_{hx}| \le 2C^*\) we obtain from (4.2), (4.3) for some \(C>0\) that
Note that thanks to Assumption 2.2 and (2.1) \(\Vert u_{t}-r\Vert _{L^\infty ([0,T],L^\infty (S^1))} < C\). Observe that on \(S_j\)
and similarly for \(|r - r_{j-1}|\). Hence we get
Note that \(B_{3}^{+}\) defined in (4.4) can be written as
Using the \(L^{\infty }\)-bounds for \(u_{t}\), r and \(r_{j}\) [recall (2.1), (3.10), and \(|\bar{\nu }_{j}| \le 1\)], (4.7), the bound \(|u_{hx}|\le 2 C^{*}\), embedding theory, and arguments similar to those employed in (4.8), and (4.9), we infer that
Arguing similarly for \(B^-\), and putting all estimates together we finally obtain from (4.1) that
which shows the first claim.
As \(u_{ht_{|_{S_j}}}=\dot{u}_{j-1} + (\dot{u}_j - \dot{u}_{j-1})\frac{(x-x_{j-1})}{h_j}\), we have that \(u_{ht}(x_j)=\dot{u}_j\). On the other hand \(I_h u_t (x_j)= u_t(x_j)\). Therefore for \(x\in S_j\) we can write
For \(w_h (x): = I_h u_t (x) - u_{ht} (x)\) we can use the inverse estimate (3.6). Therefore
by (3.1) and (3.2). Arguing similarly for the term \(|u_t - \dot{u}_{j-1}|\), integrating, and summing up over the grid intervals we obtain (4.5).
Regarding the last estimate, observe that for any \(y\in S_{j+1}\) and \(x\in S_j\) we can write
Thanks to the continuity of q we can choose \(y \in S_{j+1}\) such that
Using this fact and (3.1) yields that
With similar arguments for \(q - q_{j}{/}h_{j}\) and \(q - q_{j-1}/h_{j-1}\) we obtain that
The terms \(|c-c_j|\) and \(|c-c_{j-1}|\) can be estimated similarly as \(|u_{t} - \dot{u}_{j}|\) and \(|u_{t} - \dot{u}_{j-1}|\) whence
We can use the boundedness of \(|u_x|\) from below (2.5) to get for any x, \(y \in S_j \cup S_{j+1}\) (suppose \(y\le x\) or change the order of integration otherwise)
Choosing \(y \in S_{j+1}\) such that \(h_{j+1}|\tau (y)-\tau _{j+1}|^2 \le \int _{S_{j+1}} (\tau -\tau _h)^2 dx\) we can write
Repeating the same sort of argument for \(|\tau - \tau _{j-1}|\) and integrating over \(S_j\) we get
Putting all estimates together and summing up over the grid intervals (4.6) follows. \(\square \)
Lemma 4.2
Assume that \({{\bar{t}}} \in (0,T]\) is such that
where \({\hat{C}}(S^1)\) is a constant for the embedding \(H^1(S^1) \hookrightarrow L^\infty (S^1)\). Then the following estimate holds with some constant \(C>0\):
Proof
The difference between the continuous (2.4) and the discrete version (3.7) reads
for all test functions \(\zeta _h(x,t)\) of the form \(\zeta _h = \sum _j \zeta _j(t) \varphi _j(x)\). Choosing
a short calculation yields that
Using Lemma 4.1 we can write
Together with (2.6), the assumptions (4.13), and (4.6) we obtain that
Similarly for \(K_{2}\), using again Lemma 4.1, (2.6), embedding theory and the assumptions (4.13) to estimate \(\Vert c - c_h \Vert _{L^\infty (S^1)}\) we can write
For \(K_{3}\) we note that by (2.6), (4.13), and (3.2)
with \(\hat{\varepsilon }>0\) that will be picked later on. We will refer to this estimate later on when integrating (4.15) with respect to time. For the term \(K_{4}\) we infer from (3.2) and (4.13) that
By the interpolation estimates (3.2), (3.3), (4.13), and embedding theory we have the following estimates for the terms involving spatial gradients (for \(\epsilon > 0\) arbitrarily small):
Summarizing all these estimates we obtain from (4.15) that we arrive at
Integrating with respect to time from 0 to \({{\bar{t}}}\), using (4.16), (4.13), and embedding theory we get for \(\epsilon \) small enough that
Note that
and, similarly with some arguments as used to estimate \(K_3\)
Choosing \(\hat{\varepsilon }\) small enough and using the above estimates for the initial data yields the claimed estimate (4.14). \(\square \)
Lemma 4.3
Assume that \({{\bar{t}}} \in (0,T]\) is such that
Then the following estimate holds for some \(C>0\):
Proof
The proof of this lemma follows the lines of the analogous Lemma 5.1 in [12]. However, some additional terms concerning the dependence on c have to be estimated. More precisely, while the terms \(I_1, \ldots ,I_5\) as defined below in (4.19) have been treated in [12] already, the terms \(J_1, \ldots J_6\) depend on c or \(c_{h}\) and are new. They can be dealt with using similar arguments, though.
Let us first write down the difference between the continuous geometric equation (2.2) and its discrete version (3.12):
for all \(\varphi _h \in X_h\). As a test function we choose
Observing that
some straightforward calculations show that
An evaluation of the integrals \(I_1\), \(I_2\) and \(I_3\) is given in [12, Lemma 5.1], therefore we can assert
with \(\epsilon >0\) to be chosen later. Note that, for \(I_{2}\), one uses Young’s inequality \(ab \le a^{2}+ b^{2}/4\) and (3.4) to obtain
Next we use interpolation (3.3) and (4.17) to obtain that
Noting that
by (4.17) we can infer that
The integral \(J_1\) can be estimated exactly as \(I_1\) because of its similar structure. Using (4.17)
with \(\epsilon >0\) to be chosen later. For \(J_2\) we note the following using (4.17), (3.4), (3.5), (3.2), and the boundedness of \(f'\):
Using Young’s inequality, noting that \(|\nu - \nu _h| = |\tau - \tau _h|\), and using interpolation estimates we also infer that
The second last term \(J_5\) can be estimated using (4.17), (3.4) and the \(L^\infty \)-bounds for f and \(f'\) as follows:
Similarly,
Collecting all the estimates and by embedding theory we obtain from (4.19) that (for \(h \le 1\))
Choosing \(\epsilon \) small enough, integrating with respect to time from 0 to \({{\bar{t}}}\) and using (4.17) we obtain that
Note that by
and by interpolation theory (3.3) we have that
Thus, (4.20) yields the claimed estimate. \(\square \)
Lemma 4.4
Assuming that
there exists a constant \(C>0\) such that for all \(t \in [0,{{\bar{t}}}]\)
Proof
Note that thanks to the assumption that \(2C^* \ge q_h \ge C^{**} /2\) the discrete length elements are comparable, that is \(C^{-1} q_{j+1} \le q_j \le C q_{j+1}\,\).
Integrating (4.1) with respect to t we obtain
Clearly
Using (4.2), (4.3), and (2.8) we get (in \(S_{j}\))
Using (2.9), the fact that \(q_k \le 2 C^* h_k\) for all k, \(|f(c) -f(c_{k})| \le \Vert f'\Vert _{L^{\infty }(\mathbb {R})} |c-c_k|\) and the boundedness of f, \(f'\), \(u_{t}\) and r we obtain that
Integrating (4.4) with respect to t yields
Thanks to (2.9), (3.16), the bounds for \(u_{t}\), r, and the fact that \(q_k \le 2 C^* h_k\) for all k we obtain
Repeating the same arguments for \(B^-\) and putting all estimates together we infer from (4.24) and recalling (4.9) that
Squaring the above expression, integrating with respect to space over \(S_{j}\), using and (4.11), (4.12), and (4.10) leads to
where \(M_j := S_j \cup S_{j+1} \cup S_{j-1}\). Summing up over all grid elements and using that
a Gronwall argument yields the claimed estimated (4.23). \(\square \)
5 Proof of the convergence Theorem 1.1
Thanks to the estimates in the previous Sect. 4 we are ready to prove the main result. We follow the lines of [12, Theorem 5.3] but need to also derive the estimates for \(c - c_h\) and repeat some arguments for the convenience of the reader.
Proof
First of all note that from standard ODE theory we have local existence and uniqueness of a discrete solution \((u_h,c_h)\) of (1.6), (1.7). Assume that \(T^* \in (0,T)\) is the maximal time for which we have that
where \({\hat{C}}(S^1)\) is a constant for the embedding \(H^1(S^1) \hookrightarrow L^\infty (S^1)\). Inserting equation (4.23) into (4.18) [note that (4.17) and (4.22) are satisfied thanks to (5.1)] gives for \({{\bar{t}}} \in [0,T^*]\) that
where, for the last inequality, we have used the monotonicity of the integrands. For instance,
A Gronwall argument yields that
Inserting this estimate into (4.23) gives for \(t \in [0, T^*]\) (again using the monotonicity of the integrands)
Next, we plug (5.2) and (5.3) into (4.14) [note that (4.13) is satisfied thanks to (5.1)] to obtain for \({{\bar{t}}} \in [0,T^*]\) that
Applying Gronwall again yields
Inserting this into (5.2) and (5.3) we obtain that
The constants appearing so far do not depend on \(T^*\). Since \(u_{hx}\) is constant on each grid interval, the above estimate together with classical embedding theory (see for example [4, Theorem 2.2]) implies
for all \(h \le h_0\) with \(h_0 \in (0,1)\) sufficiently small independently of \(T^*\). Similarly, after eventually decreasing \(h_0\) (recall also (2.8)), \(|u_{hx}| \le \frac{3}{2} C^*\) for all \(h \le h_0\) independently of \(T^*\).
Next observe that using (5.4), (3.6), and embedding theory we can write for \(t \in [0,T^*]\)
for all \(h \le h_0\) independently of \(T^*\) (after decreasing \(h_0\) if required). Using (5.4) we can easily derive that
for all \(h \le h_0\) independently of \(T^*\) (after decreasing \(h_0\) again if required). Continuity of the solution \((u_h,c_h)\) with respect to time yields a contradiction to the maximality of \(T^*\). It follows that \(T^*=T\) and the theorem is proved. \(\square \)
Corollary 5.1
Under the assumptions of Theorem 1.1 we have that
Proof
From Theorem 1.1, (4.21), and the fact that \(|u_x|\) and \(|u_{hx}|\) are bounded we obtain immediately the bound for the semi-norm \(|u-u_h|_{H^1 (S^1)}\). To prove the \(L^2\)-bound note that \(u(x,t)-u_h (x,t)= u(x,0)-u_h (x,0) + \int _0^t u_t (x,t')- u_{ht} (x,t') dt'\) and use Theorem 1.1 again with the interpolation result (3.2) for the initial values. \(\square \)
6 Numerical simulations
6.1 Sources and reaction terms
We now aim for assessing the results in Theorem 1.1. Exact solutions to the PDE system (1.4), (1.5) are difficult to obtain whence we prescribe functions (u, c) and account for source terms to ensure that they are solutions, i.e., we consider
with functions \(s_u : S^1 \times [0,T] \rightarrow \mathbb {R}^2\) and \(s_c: S^1 \times [0,T] \rightarrow \mathbb {R}\).
The required extension of the weak formulation (2.2), (2.3) is straightforward. With respect to the spatial discretization of the source terms we apply the interpolation \(I_h\) as follows: Instead of the equations (1.6), (1.7) we have
6.2 Time discretization
We apply a semi-implicit scheme which reads as follows:
Problem 6.1
(Fully Discrete Scheme) Given a time step \(\delta > 0\), let \(M = T/ \delta \) and find functions \(u_{\delta h}^{(m)}(\cdot ) \in X_h\) and \(c_{\delta h}^{(m)}(\cdot ) \in Y_h\), \(m \in \{ 0, \ldots , M \}\), of the form
with \(u_j^{(m)} \in \mathbb {R}^2\) and \(c_j^{(m)} \in \mathbb {R}\), such that \(u_{\delta h}^{(0)}(\cdot ) = u_{h0}\), \(c_{\delta h}^{(0)}(\cdot ) = c_{h0}\), and such that for all \(\varphi _{h} \in X_h\), \(\zeta _{h} \in Y_h\) and all \(m \in \{ 0, \ldots , M-1 \}\)
For a more sophisticated time discretization of PDEs on evolving surfaces we refer, for instance, to [20]. We solve the above fully discrete problem using the following algorithm:
Algorithm 6.1
Given data: N (number of nodes), \(\delta \) (time step), \((u_0, c_0)\) (initial data), M (number of time steps), tol (abort if any segment length becomes smaller).
-
1.
Set \(m=0\).
Initialize \(u_{h}^{(0)} = u_{h0} = I_h u_{0}\) and \(c_{h}^{(0)} = c_{h0} = I_h c_{0}\) by computing the values \(u_{i}^{(0)} = u_{0}(x_{i})\) and \(c_{i}^{(0)} = c_{0}(x_{i})\), \(i = 1, \ldots , N\).
Also, compute \(q_{i}^{(0)} = | u_{i}^{(0)} - u_{i-1}^{(0)} |\), \(i = 1, \ldots , N\).
Abort if \(\min _j q_j^{(0)} < tol\).
-
2.
Compute the vertex positions at time \(t^{(m+1)} = (m+1) \delta \) from
$$\begin{aligned}&\tfrac{1}{2\delta } (q_{i+1}^{(m)} + q_{i}^{(m)}) u_{i}^{(m+1)} - \tfrac{1}{q_{i}^{(m)}} u_{i-1}^{(m+1)} + \left( \tfrac{1}{q_{i+1}^{(m)}} + \tfrac{1}{q_{i}^{(m)}} \right) u_{i}^{(m+1)} - \tfrac{1}{q_{i+1}^{(m)}} u_{i+1}^{(m+1)} \\&\quad = \tfrac{1}{2} (q_{i+1}^{(m)} + q_{i}^{(m)}) \left( \tfrac{1}{\delta } u_{i}^{(m)} + s_{u}(x_{i})^{(m+1)} \right) \\&\qquad + \tfrac{1}{2} f(c_{i}^{(m)}) \left( u_{i+1}^{(m)} - u_{i-1}^{(m)} \right) ^\perp , \quad i=1, \ldots , N, \end{aligned}$$and compute \(q_{i}^{(m+1)} = | u_{i}^{(m+1)} - u_{i-1}^{(m+1)} |\), \(i = 1, \ldots , N\).
-
3.
Compute the surface field values at time \(t^{(m+1)} = (m+1) \delta \) from
$$\begin{aligned}&\left( \tfrac{1}{3\delta } (q_{j+1}^{(m+1)} + q_{j}^{(m+1)}) + \left( \tfrac{1}{q_{j+1}^{(m+1)}} + \tfrac{1}{q_{j}^{(m+1)}} \right) \right) c_{j}^{(m+1)} \\&\quad \left( \tfrac{1}{6\delta } q_{j+1}^{(m+1)} - \tfrac{1}{q_{j+1}^{(m+1)}} \right) c_{j+1}^{(m+1)} + \left( \tfrac{1}{6\delta } q_{j}^{(m+1)} - \tfrac{1}{q_{j}^{(m+1)}} \right) c_{j-1}^{(m+1)} \\&\quad \quad \quad = \tfrac{1}{3\delta } (q_{j+1}^{(m)} + q_{j}^{(m)}) \left( c_{j}^{(m)} + \delta s_{c}(x_{j})^{(m+1)} \right) \\&\quad \quad \quad \quad + \tfrac{1}{6\delta } q_{j+1}^{(m)} \left( c_{j+1}^{(m)} + \delta s_{c}(x_{j+1})^{(m+1)} \right) \\&\qquad \qquad + \tfrac{1}{6\delta } q_{j}^{(m)} \left( c_{j-1}^{(m)} + \delta s_{c}(x_{j-1})^{(m+1)} \right) , \quad j = 1, \ldots , N. \end{aligned}$$ -
4.
If \(\min _j q_j^{(m+1)} \ge tol\) and \(m+1<M\) then increase m by one and go to step 2.
Observe that the parametrization does not feature any more in the algorithm. The identities in steps two and three are straightforward to compute. For instance, step two is easily obtained from the continuous version (3.8).
6.3 A radially symmetric solution
Consider a radially symmetric setting and denote by R(t) and B(t) the radius of the evolving circle and the constant (in space) value of c along the circle, respectively. We pick \(\nu \) to be the outer unit normal of the enclosed ball. Then \(v = R'(t)\) and \(\kappa = -1/R(t)\). The system (1.1), (1.2) becomes
We consider the forcing function
Note that this function is not bounded and thus does not satisfy the assumptions of Theorem 1.1. However, the values of B in the subsequent simulations are bounded. We may therefore think of cutting off f at suitable high and low values which are outside of the computed values and locally smooth it sufficiently. This does not alter the computational results but the Theorem then applies.
The constant functions \((R(t),B(t)) = (1,1)\) are a stationary and stable solution to (6.7). The solution for initial values \(R(0) = 1.25\) and \(B(t) = 0.8\) converges back to this stable point and has been approximated with a standard MATLAB routine for the comparison in Fig. 1.
Now let \(h = 1/N\) with \(N \in \mathbb {N}\) and define the initial position of the curve approximation by
in which we set \(c_j^{(0)} = B(0)\). Furthermore, we set \(\delta = h^2\). We then perform numerical simulations with the scheme described in Algorithm 6.1. In order to be able to compare with the solution to the ODE system (6.7) we use the length of the computed polygon divided by \(2\pi \) and the average of the values of \(c_h\) in the nodes,
Figure 1 gives a nice impression of the convergence as the computational effort is increased. Note that the errors essentially are due to the spatial discretization. We checked that changing the time step only has a marginal impact on the graphs.
6.4 An oscillating solution
Consider now the functions
and
for \(x \in [0,1]\) and \(t \in [0,T]\) with \(T = 1\). Let \(f(c) = 2c\) (with regards to the lack of bound the remark in the previous section applies again). Then (u, c) is a solution to (6.1), (6.2) if the source terms are given by (writing \(s_u = (s_{u1},s_{u2})\))
For the numerical simulations with Algorithm 6.1 we monitored the following errors:
where we used sufficiently accurate quadrature rules on each interval \(S_j\) for the spatial integration.
We first picked several values for N as displayed in Table 1 and time steps of the size \(\delta = h^2\) where \(h = 1/N\). We checked that by this choice of the time step the spatial discretization error is dominating. The EOCs of \(\mathcal {E}_2\), \(\mathcal {E}_3\), and \(\mathcal {E}_5\) are close to two which is what Theorem 1.1 asserts. The error in \(c_x\) is relatively high but this is not surprising in view of the spatial oscillations of the surface quantity c which are at a higher frequency than those of the position field u. In turn, the EOCs of \(\mathcal {E}_1\) and \(\mathcal {E}_4\) are close to four and thus better than Theorem 1.1 predicts, a behavior which may be expected for \(\mathcal {E}_1\).
We also assessed the discretization error with respect to the time stepping. For the results in Table 2 we fixed a very fine spatial mesh with \(N=2001\) nodes and varied the time step. Note that our semi-implicit time discretization is of consistency order one. In accordance with this the EOCs of all errors are close to two for all fields. The drops of the EOCs of some errors for small time steps (from about \(m=5\) in Table 2) are due to the spatial discretization error becoming more significant.
References
Balažovjech, M., Mikula, K.: A higher order scheme for a tangentially stabilized plane curve shortening flow with a driving force. SIAM J. Sci. Comput. 33(5), 2277–2294 (2011)
Barrett, J.W., Garcke, H., Nürnberg, R.: Numerical approximation of gradient flows for closed curves in \({\mathbb{R}}^d\). IMA J. Numer. Anal. 30(1), 4–60 (2010)
Barrett, J.W., Garcke, H., Nürnberg, R.: The approximation of planar curve evolutions by stable fully implicit finite element schemes that equidistribute. Numer. Methods Partial Differ. Equ. 27(1), 1–30 (2011)
Buttazzo, G., Giaquinta, M., Hildebrandt, S.: One-dimensional variational problems: an introduction, Oxford Lecture Series in Mathematics and its Applications, vol. 15. The Clarendon Press Oxford University Press, New York (1998)
Cesaroni, A., Novaga, M., Valdinoci, E.: Curve shortening flow in heterogeneous media. Interfaces Free Bound 13(4), 485–505 (2011)
Chambolle, A., Novaga, M.: Implicit time discretization of the mean curvature flow with a discontinuous forcing term. Interfaces Free Bound 10(3), 283–300 (2008)
Chou, K.-S., Zhu, X.-P.: The curve shortening problem. Chapman & Hall/CRC, Boca Raton (2001)
Deckelnick, K., Dziuk, G.: On the approximation of the curve shortening flow. In: Calculus of variations, applications and computations (Pont-à-Mousson, 1994), Pitman Res. Notes Math. Ser. Longman Sci. Tech., vol. 326, pp. 100–108. Harlow (1995)
Deckelnick, K., Dziuk, G., Elliott, C.M.: Computation of geometric partial differential equations and mean curvature flow. Acta Numer. 14, 139–232 (2005)
Dziuk, G.: An algorithm for evolutionary surfaces. Numer. Math. 58(6), 603–611 (1991)
Dziuk, G.: Convergence of a semi-discrete scheme for the curve shortening flow. Math. Models Methods Appl. Sci. 4(4), 589–606 (1994)
Dziuk, G.: Discrete anisotropic curve shortening flow. SIAM J. Numer. Anal. 36(6), 1808–1830 (1999)
Dziuk, G., Elliott, C.M.: Finite elements on evolving surfaces. IMA J. Numer. Anal. 27(2), 262–292 (2007)
Dziuk, G., Elliott, C.M.: Finite element methods for surface PDEs. Acta Numer. 22, 289–396 (2013)
Elliott, C.M., Fritz, H.: On approximations of the curve shortening flow and of the mean curvature flow based on the DeTurck trick. IMA. J. Numer. Anal. doi:10.1093/imanum/drw020
Elliott, C.M., Stinner, B.: Modeling and computation of two phase geometric biomembranes using surface finite elements. J. Comput. Phys. 229(18), 6585–6612 (2010)
Elliott, C.M., Stinner, B.: Computation of two-phase biomembranes with phase dependent material parameters using surface finite elements. Commun. Comput. Phys. 13(2), 325–360 (2013)
Elliott, C.M., Stinner, B., Venkataraman, C.: Modelling cell motility and chemotaxis with evolving surface finite elements. J. R. Soc. Interface 9(76), 3027–3044 (2012)
Leung, S., Lowengrub, J., Zhao, H.: A grid based particle method for solving partial differential equations on evolving surfaces and modeling high order geometrical motion. J. Comput. Phys. 230(7), 2540–2561 (2011)
Lubich, C., Mansour, D., Venkataraman, C.: Backward difference time discretization of parabolic differential equations on evolving surfaces. IMA J. Numer. Anal. 33(4), 1365–1385 (2013)
Mantegazza, C.: Lecture notes on mean curvature flow, Progress in Mathematics, vol. 290. Birkhäuser/Springer Basel AG, Basel (2011)
Mikula, K., Ševčovič, D., Balažovjech, M.: A simple, fast and stabilized flowing finite volume method for solving general curve evolution equations. Commun. Comput. Phys. 7(1), 195–211 (2010)
Neilson, M.P., Mackenzie, J.A., Webb, S.D., Insall, R.H.: Modeling cell movement and chemotaxis using pseudopod-based feedback. SIAM J. Sci. Comput. 33(3), 1035–1057 (2011)
Olshanskii, M.A., Reusken, A.: Error analysis of a space-time finite element method for solving PDEs on evolving surfaces. SIAM J. Numer. Anal. 52(4), 2092–2120 (2014)
Plaza, R.G., Sánchez-Garduño, F., Padilla, P., Barrio, R.A., Maini, P.K.: The effect of growth and curvature on pattern formation. J. Dyn. Differ. Equ. 16(4), 1093–1121 (2004)
Pozzi, P.: Anisotropic curve shortening flow in higher codimension. Math. Methods Appl. Sci. 30(11), 1243–1281 (2007)
Acknowledgments
The authors would like to thank the Isaac Newton Institute for Mathematical Sciences for its hospitality during the programme Coupling Geometric PDEs with Physics for Cell Morphology, Motility and Pattern Formation supported by EPSRC Grant Number EP/K032208/1.
Author information
Authors and Affiliations
Corresponding author
Rights and permissions
About this article
Cite this article
Pozzi, P., Stinner, B. Curve shortening flow coupled to lateral diffusion. Numer. Math. 135, 1171–1205 (2017). https://doi.org/10.1007/s00211-016-0828-8
Received:
Revised:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00211-016-0828-8