Abstract
A convergent numerical method for \(\alpha \)-dissipative solutions of the Hunter–Saxton equation is derived. The method is based on applying a tailor-made projection operator to the initial data, and then solving exactly using the generalized method of characteristics. The projection step is the only step that introduces any approximation error. It is therefore crucial that its design ensures not only a good approximation of the initial data, but also that errors due to the energy dissipation at later times remain small. Furthermore, it is shown that the main quantity of interest, the wave profile, converges in \(L^{\infty }\) for all \(t \ge 0\), while a subsequence of the energy density converges weakly for almost every time.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
In this article, we present a numerical algorithm for \(\alpha \)-dissipative solutions of the Cauchy problem for the Hunter–Saxton (HS) equation
The equation was derived as an asymptotic model of the director field of a nematic liquid crystal [19], and possesses a rich mathematical structure. We mention a few of its properties here: it is bi-Hamiltonian and admits a Lax pair [20], it can be interpreted as a geodesic flow [23, 24], and it admits numerous extensions and generalizations [22, 26, 31].
A lot of the interest for (1.1) is generated by the fact that weak solutions in general will develop singularities in finite time, and, consequently, they are not unique, see [9, 13, 19, 21]. This phenomenon is known as wave breaking. In particular, \(u_{x} \rightarrow -\infty \) pointwise while the energy \(\Vert u_x (t, \cdot )\Vert _2\) remains uniformly bounded, and the solution u, itself, Hölder continuous. Furthermore, at wave breaking, energy concentrates on a set of measure zero. Thus the energy density is in general not absolutely continuous. To overcome this problem it is common to augment the solution with a finite, positive Radon measure \(\mu \) describing the energy density, see [9, 13, 18]. In particular, \(\mu \) coincides with the usual kinetic energy density \(u_x^2\) in regions where there is no wave breaking, thus, \(d\mu _{\text {ac}}=u_x^2dx\). The energy is then described by \(F(x) = \mu ((-\infty , x))\).
Weak solutions can be extended past wave breaking in various ways, see for instance [3, 9, 13, 21]. The two most prominent notions in the literature are that of a dissipative solution, where one removes all the concentrated energy from the system, and that of a conservative solution, where one reinserts the concentrated energy. In this work, we consider the concept of \(\alpha \)-dissipative solutions, first introduced in [12] for the related Camassa–Holm equation, and in [13] for (1.1). Instead of removing all the concentrated energy or none of it, an \(\alpha \)-fraction, where \(\alpha \in [0, 1]\), is removed. This way, the notion of \(\alpha \)-dissipative solutions acts as a continuous interpolant between the two extreme cases, \(\alpha =0\) corresponding to conservative solutions, and \(\alpha =1\) corresponding to dissipative solutions. Thus, the notion of \(\alpha \)-dissipative solutions allows for a uniform treatment of weak solutions with nonincreasing energy. The existence of \(\alpha \)-dissipative solutions was established in [13] for the two-component Hunter–Saxton system which generalizes (1.1), and in the more general setting where \(\alpha \in W^{1, \infty }(\mathbb {R}, [0, 1)) \cup \{1 \}\).
A common approach for solving the Cauchy problem (1.1) is to use a generalized method of characteristics, see for instance [9, 13]. This approach is followed here. In particular, given an \(\alpha \)-dissipative solution, the corresponding Lagrangian coordinates are governed by a linear system of differential equations, whose right hand side admits discontinuities at fixed times. These times can be computed a priori based on the initial data. Whence, if the initial data in Lagrangian coordinates is piecewise linear, the Lagrangian coordinates will remain piecewise linear at all later times, an important property which we take advantage of in this work.
Despite receiving a considerable amount of attention from a theoretical perspective, only a few number of numerical schemes have been proposed for the HS equation. In [17] several finite difference schemes for dissipative solutions were proved to converge. Also for dissipative solutions, [32] introduces a convergent, discontinuous Galerkin method. Furthermore, in [30] a finite difference scheme on a periodic domain for a modified HS equation was derived and proven to converge towards conservative solutions. More recently, a Godunov-inspired scheme [25, Chp. 12.1] for conservative solutions based on tracking the solution along characteristics, was introduced in [14]. The scheme was proved to converge and a convergence rate prior to wave breaking was derived.
We contribute to this line of research by introducing a numerical algorithm well-suited for \(\alpha \)-dissipative solutions. This algorithm is based on applying a tailor-made projection operator to the initial data, and then solving exactly along characteristics. Thus, the projection step is the only step that introduces any approximation error and it is therefore of particular importance that it not only yields a good approximation of the initial data, but also ensures that additional prospective errors, due to the energy dissipation, remain small and hence do not prevent convergence.
To highlight the importance of the correct choice of the projection operator, we compare our projection operator with the one introduced in [14]. Motivated by the fact that a piecewise linear structure of the initial data is preserved at all later times, see [1, 20], the numerical scheme in [14] uses a standard piecewise linear projection operator in Eulerian coordinates at every timestep. In between timesteps the numerical solution is evolved exactly along characteristics. The projection operator used treats u and F completely independently, although u and \(\mu \) are strongly connected through the absolutely continuous part. Consequently, a deviation is introduced in the sense that \(d\mu _{\Delta x, \textrm{ac}} \ne u_{\Delta x, x}^2dx\), where \(\Delta x\) denotes the spatial discretization parameter. Thus, as pointed out in [14], one is no longer dealing with the Hunter–Saxton equation, but rather a reformulated version of the two-component Hunter–Saxton system [28], which is accompanied by a density \(\rho \). Or, in other words, the projection operator maps into the Eulerian set for the two-component Hunter-Saxton equation, see [28, Def. 2.2], for which it has been established that \(\rho \) has a regularizing effect. In particular, if \(\rho > 0\) in some interval \([x_1, x_2]\), then the solution will never experience wave breaking in the interval spanned by the characteristics emanating from \([x_1, x_2]\), see [11, Thm 7.1]. While this works out neatly in the setting of conservative solutions, since no energy is lost, for \(\alpha \)-dissipative solutions we are highly dependent on wave breaking actually occurring in order to remove concentrated energy. Thus, more care is needed to derive a suitable projection operator. That is why we introduce a piecewise linear projection operator \(P_{\Delta x}\), which in contrast to the the one in [14], ensures that the projected initial data satisfies \(d\mu _{\Delta x, \textrm{ac}} = u_{\Delta x, x}^2dx\).
We accompany the numerical algorithm with a convergence analysis, which shows that the limit is a weak \(\alpha \)-dissipative solution. In particular, we obtain that \(u_{\Delta x} \rightarrow u\) in \(L^{\infty }(\mathbb {R})\) for all \(t \ge 0\), which is the main quantity of interest in practice, while a subsequence of the energy measure \(\mu _{\Delta x}(t)\) converges weakly for almost every time.
The outline of the paper is as follows. In Sect. 2 we set the stage for the Cauchy problem of (1.1) by defining the Eulerian set \(\mathcal {D}\), the Lagrangian set \(\mathcal {F}\), the mappings L and M between them and the notion of an \(\alpha \)-dissipative solution. Then in Sect. 3 we focus on deriving and motivating our choice of the projection operator \(P_{\Delta x}\), before we discuss the practical implementation of our numerical algorithm. In Sect. 4 we conduct the convergence analysis. In the last section, Sect. 5, we provide a few numerical experiments to illustrate the theoretical results and to go beyond the theory established here by investigating the convergence rate.
2 Preliminaries
In this section, we set the stage for the numerical algorithm by defining the sets we are working in, as well as recalling the construction of \(\alpha \)-dissipative solutions by using a generalized method of characteristics, as introduced in [13] and used in [15].
Assume that \(\alpha \in [0,1]\) and denote by \(\mathcal {M}^+(\mathbb {R})\) the space of positive, finite Radon measures on \(\mathbb {R}\). To define the set of Eulerian coordinates, we first need to recall some important spaces from [13] and [15].
Introduce
which is a Banach space when equipped with the norm
Furthermore, define
and introduce a partition of unity \(\chi ^+\) and \(\chi ^-\) on \(\mathbb {R}\), i.e., a pair of functions \(\chi ^+, \chi ^- \in C^{\infty }(\mathbb {R})\) such that
-
\(\chi ^+ + \chi ^- = 1\),
-
\(0 \le \chi ^{\pm } \le 1\),
-
\(\text {supp}(\chi ^+) \subset (-1, \infty )\) and \(\text {supp}(\chi ^-) \subset (-\infty , 1)\).
Then, we can define the following mappings
which are linear, continuous, and injective, see [4]. Based on those, we introduce the Banach spaces \(E_1\) and \(E_2\) as the images of \(H_1^1(\mathbb {R})\) and \(H_2^1(\mathbb {R})\), respectively, that is,
and endow them with the norms
By construction the spaces \(E_1\) and \(E_2\) do not rely on the particular choice of \(\chi ^+\) and \(\chi ^-\), cf. [10]. Furthermore, observe that the mapping \(R_1\) is also well-defined for functions in \(L_1^2(\mathbb {R}) = L^2(\mathbb {R}) \times \mathbb {R}\) and therefore, we set
and equip this space with the norm
At last, we can define the set of Eulerian coordinates, \(\mathcal {D}\), in which we also seek numerical solutions.
Definition 2.1
The space \(\mathcal {D}\) consists of all triplets \((u, \mu , \nu )\) such that
-
(i)
\(u \in E_2\),
-
(ii)
\(\mu \le \nu \in \mathcal {M}^{+}(\mathbb {R})\),
-
(iii)
\(\mu _{\textrm{ac}}\le \nu _{\textrm{ac}}\),
-
(iv)
\(d\mu _{\textrm{ac}} = u_x^2dx\),
-
(v)
\(\mu \left( (-\infty , \cdot ) \right) \in E_1^0\),
-
(vi)
\(\nu \left( (-\infty , \cdot ) \right) \in E_1^0\),
-
(vii)
If \(\alpha = 1\), then \(d\nu _{\textrm{ac}}=d\mu = u_x^2dx\),
-
(viii)
If \(\alpha \in [0, 1)\), then \(\frac{d\mu }{d\nu }(x) \in \{1 - \alpha , 1 \}\), and \(\frac{d\mu _{\textrm{ac}}}{d\nu _{\textrm{ac}}}(x) = 1\) if \(u_x(x) < 0\).
As \(\mu \), \(\nu \in \mathcal {M}^+(\mathbb {R})\), we can define the primitive functions \(F(x) = \mu ((-\infty , x))\) and \(G(x) = \nu ((- \infty , x))\). These are bounded, increasing, left-continuous and satisfy
We will interchangeably use the notation (u, F, G) and \((u, \mu , \nu )\) to refer to the same triplet in \(\mathcal {D}\), since by [7, Thm. 1.16], there is a one-to-one correspondence between (F, G) and \((\mu , \nu )\).
Moreover, for practical purposes, we will often restrict the initial data to belong to
Let \(B=E_2 \times E_2 \times E_1 \times E_1\) endowed with the norm
then the set of Lagrangian coordinates, \(\mathcal {F}\), is defined as follows.
Definition 2.2
The set \(\mathcal {F}\) consists of all quadruplets \(X = (y, U, V, H)\) with \((y-\text { {id}}, U, V, H) \in B\) such that
-
(i)
\((y-\text { {id}}, U, V, H) \in \big [W^{1, \infty }(\mathbb {R}) \big ]^4\),
-
(ii)
\( y_{\xi }, H_{\xi } \ge 0\) and there exists \(c > 0\) such that \(y_{\xi } + H_{\xi } \ge c \) holds a.e.,
-
(iii)
\(y_{\xi }V_{\xi }= U_{\xi }^2 \) a.e.,
-
(iv)
\(0 \le V_{\xi } \le H_{\xi }\) a.e.,
-
(v)
If \(\alpha = 1\), then \(y_{\xi }(\xi )=0\) implies that \(V_{\xi }(\xi )=0\), and \(y_{\xi }(\xi ) > 0\) implies that \(V_{\xi }(\xi ) = H_{\xi }(\xi )\) a.e.,
-
(vi)
If \(\alpha \in [0, 1)\), then there exists a function \(\kappa : \mathbb {R}\rightarrow \{1 - \alpha , 1 \}\) such that \(V_{\xi }(\xi ) = \kappa (\xi )H_{\xi }(\xi )\) a.e., and \(\kappa (\xi )=1\) whenever \(U_{\xi }(\xi ) < 0\).
In the convergence analysis, the following subsets of \(\mathcal {F}\) will play an important role
and
To construct the \(\alpha \)-dissipative solution using a generalized method of characteristics means to study the time evolution in Lagrangian rather than in Eulerian coordinates, and therefore the mappings between \(\mathcal {D}\) and \(\mathcal {F}\) are an essential part.
Definition 2.3
Let \(L: \mathcal {D}\rightarrow \mathcal {F}_0\) be defined by \(L \left( (u, \mu , \nu ) \right) = (y, U, V, H)\), where
Next we introduce the mapping taking us from Lagrangian to Eulerian coordinates. To this end, recall that the pushforward of a Borel measure \(\lambda \) by a measurable function f is the measure \(f_{\#}\lambda \) defined for all Borel sets \(A\subset \mathbb {R}\) by
Definition 2.4
Define \(M: \mathcal {F}\rightarrow \mathcal {D}\) by \(M((y, U, V, H)) = (u, \mu , \nu )\), where
For proofs that these mappings are well-defined we refer to [29, Prop. 2.1.5 and 2.1.7]. Furthermore, it should be noted that the triplets \((u,\mu ,\nu )\) are mapped to quadruplets (y, U, V, H) and hence there cannot be a one-to-one correspondence between Eulerian and Lagrangian coordinates. However, as pointed out in [29], one can identify equivalence classes in Lagrangian coordinates, so that each equivalence class corresponds to exactly one element in Eulerian coordinates.
Moreover, it should be pointed out that all the important information in Eulerian coordinates is encoded in the pair \((u,\mu )\), and hence contained in the triplet (y, U, V) in Lagrangian coordinates. In contrast, the mapping L relies heavily on \(\nu \) and hence changing \(\nu \) changes not only H, but also (y, U, V).
Finally, we can turn our attention to the time evolution. The \(\alpha \)-dissipative solution in Lagrangian coordinates, \(X(t) = (y, U, V, H)(t)\), with initial data \(X(0)=X_0\in \mathcal {F}\), is the unique solution to the following system of differential equations
Here \(\tau :\mathbb {R}\rightarrow [0, \infty ]\) is the wave breaking function given by
and \(V_{\infty }(t) = \lim _{\xi \rightarrow \infty }V(t, \xi )\) denotes the total Lagrangian energy at time t.
For a proof of the uniqueness of the solution to (2.2), we refer to [13, Lem. 2.3]. Furthermore, it should be pointed out that the solution operator respects equivalence classes in the following sense: If \(X_{A,0}\) and \(X_{B,0}\) belong to the same equivalence class, then also \(X_A(t)\) and \(X_B(t)\) belong to the same equivalence class for all \(t\ge 0\), see [13, Prop. 3.7].
Based on (2.2) we now define the solution operator \(S_t\) in Lagrangian coordinates.
Definition 2.5
For any \(t\ge 0\) and \(X_0\in \mathcal {F}\) define \(S_t(X_0)=X(t)\), where X(t) denotes the unique \(\alpha \)-dissipative solution to (2.2) with initial data \(X(0)=X_0\).
To finally obtain the \(\alpha \)-dissipative solution in Eulerian coordinates, we combine the solution operator \(S_t\) with the mappings L and M as follows.
Definition 2.6
For any \(t\ge 0\) and \((u_0,\mu _0,\nu _0)\in \mathcal {D}\) the \(\alpha \)-dissipative solution at time t is given by
As we mentioned earlier \(L((u,\mu ,\nu ))\) is heavily influenced by the choice of \(\nu \) in the triplet \((u,\mu ,\nu )\). Nevertheless, it has been shown in [16, Lem. 2.13], that the choice of \(\nu \) has no influence on the solution, in the following sense. Given any two triplets of initial data \((u_{A,0}, \mu _{A,0}, \nu _{A,0})\) and \((u_{B,0}, \mu _{B,0}, \nu _{B,0})\) in \(\mathcal {D}\) such that
then
As consequence of this result, we will restrict ourselves from now on to consider only initial data \((u_0,\mu _0,\nu _0)\in \mathcal {D}_0\), which means in particular that \(\mu _0=\nu _0\).
Furthermore, in the case of conservative and dissipative solutions the uniqueness of weak solutions has been established in [9] and [6], respectively, by showing that if a weak solution of the Hunter–Saxton equation satisfies certain properties, which are heavily dependent on the class of solutions one is interested in, then it can be computed using Definition 2.6. For the remaining values of \(\alpha \), i.e., \(\alpha \in (0,1)\), this is yet an open question.
3 The Numerical Algorithm
This section is devoted to presenting our numerical algorithm, which combines a projection operator \(P_{\Delta x}\) with the solution operator \(T_t\) as follows.
Definition 3.1
We define the numerical solution \((u_{\Delta x}, F_{\Delta x}, G_{\Delta x})\) for \(t \in [0,T]\) by
for any \((u_0, F_0, G_0) \in \mathcal {D}_0\).
As the projection operator \(P_{\Delta x}\) is the only part that actually introduces any error, its construction is the crucial step. In particular, \(P_{\Delta x}\) must preserve key properties of \(\alpha \)-dissipative solutions such as the total energy \(\mu ({\mathbb {R}})\) and Definition 2.1 (iv). Beside the construction of \(P_{\Delta x}\), we will, at the end of this section, discuss the implementation of our algorithm.
3.1 The Projection Operator
Any \(\mu \in \mathcal {M}^+(\mathbb {R})\) can be split into an absolutely continuous part \(\mu _{\text {ac}}\), and a singular part \(\mu _{\text {sing}}\) with respect to the Lebesgue measure, see [27, Prop. 9.8], i.e.,
Thus \(F(x)=\mu ((-\infty ,x))\) can be written as
where \(F_{\textrm{ac}}(x) = \mu _{\textrm{ac}}((-\infty , x))\) and \(F_{\textrm{sing}}(x) = \mu _{\textrm{sing}}((- \infty , x))\). Thus a projection operator acting on F(x) can be a combination of two projections, one for \(F_{\textrm{ac}}(x)\) and another one for \(F_{\textrm{sing}}(x)\). As we will see, this is how we define \(P_{\Delta x}\).
To derive \(P_{\Delta x}\), let \(\{x_j \}_{j \in {\mathbb {Z}}}\) be a uniform discretization of \(\mathbb {R}\), where \(x_j = j \Delta x\) for \(j \in {\mathbb {Z}}\) and \(\Delta x > 0\) is fixed. Furthermore, set \(P_{\Delta x}\left( (u, F, G)\right) = (u_{\Delta x}, F_{\Delta x}, G_{\Delta x})\). To ensure that \((u_{\Delta x}, F_{\Delta x}, G_{\Delta x})\) satisfies Definition 2.1 (iv), the projection operator is defined over two grid cells, i.e., over \([x_{2j}, x_{2j+1}] \cup [ x_{2j+1}, x_{2j+2}]\). Furthermore, to preserve the continuity of \((u, F_{\textrm{ac}})\) and the total energy, we require that \(\left( u_{\Delta x}, F_{\Delta x, \text {ac}} \right) \) coincides with \((u, F_{\textrm{ac}})\) in every other gridpoint, i.e.,
To interpolate \(u_{\Delta x}\) and \(F_{\Delta x}\) between the gridpoints \(\{x_{2j} \}_{j \in {\mathbb {Z}}}\), we fit two lines, \(p_1\) and \(p_2\), such that the resulting wave profile
is continuous, \((u_{\Delta x}, F_{\Delta x, \textrm{ac}})\) satisfies (3.2), and
For an arbitrary \(j \in {\mathbb {Z}}\) these constraints then read
Introducing the operators
for any sequence \(\{f_j \}_{j \in {\mathbb {Z}}}\) and solving (3.3a)–(3.3d), we end up with
where
Note that (3.4) is well-defined, as we have by the Cauchy–Schwarz inequality
For the singular part, we set
where the last equality follows from (3.1). Hence, all the discontinuities of \(F_{\Delta x, \text {sing}}(x)\) are located within the countable set \(\{x_{2j} \}_{j \in {\mathbb {Z}}}\). Finally, introduce
which is left-continuous. Then we can associate to \(F_{\Delta x}\) a positive and finite Radon measure \(\mu _{\Delta x}\) through
and, by construction,
To summarize, the projection operator \(P_{\Delta x}\) is defined as follows.
Definition 3.2
(Projection operator) We define the projection operator \(P_{\Delta x}:\mathcal {D}_0\rightarrow \mathcal {D}_0\) by \(P_{\Delta x} \left( (u,F,G) \right) =(u_{\Delta x},F_{\Delta x}, G_{\Delta x})\), where
and
for all \(x \in \mathbb {R}\). The absolutely continuous part of \(F_{\Delta x}\) is given by
and the singular part is defined by
Finally, let \(\mu _{\Delta x} ((- \infty , x))=F_{\Delta x}(x)\) and \(\nu _{\Delta x}((- \infty , x)) = G_{\Delta x}(x)\) be the unique, finite and positive Radon measures associated with \(F_{\Delta x}\) and \(G_{\Delta x}\), respectively.
Remark 3.3
The projection operator \(P_{\Delta x}\) in Definition 3.2 is only defined for \((u,\mu ,\nu )\in \mathcal {D}_0\) due to Definition 3.1. In this case \(\mu =\nu \) and by construction \(\mu _{\Delta x}= \nu _{\Delta x}\). Nevertheless, \(P_{\Delta x}\) can be extended to a projection operator \({\tilde{P}}_{\Delta x}\), which is well-defined for any \((u,\mu ,\nu ) \in \mathcal {D}\) as follows. We keep \(u_{\Delta x}\) and \(\mu _{\Delta x}((-\infty ,x))=F_{\Delta x}(x) \) given by (3.5)–(3.7), but \(\nu _{\Delta x}((-\infty ,x))=G_{\Delta x}(x)\) needs to be adjusted:
Following the same idea as for \(F_{\Delta x, \textrm{sing}}\), we set
Since we now have \(d\mu _{\textrm{ac}} = u_{x}^2dx\) and \(\mu _{\textrm{ac}} \le \nu _{\textrm{ac}}\), which implies that \(d\nu _{\textrm{ac}} = f dx \) for some \(f \ge u_x^2\), we approximate the deviation \((f - u_x^2)\) with the integral average over two grid cells. Following these lines, we end up with
for \(x\in (x_{2j}, x_{2j+1})\cup (x_{2j+1}, x_{2j+2})\). Computing \(G_{\Delta x, \textrm{ac}}(x)=\nu _{\Delta x, \textrm{ac}}((-\infty ,x))\), yields \(G_{\Delta x, \textrm{ac}}(x_{2j}) = G_{\textrm{ac}}(x_{2j})\), and
for \(x \in (x_{2j}, x_{2j+2}]\). Finally, we define \(\nu _{\Delta x}\) implicitly by
Last but not least, note that there is one drawback with \(\tilde{P}_{\Delta x}\). It preserves all properties in Definition 2.1 when \(d\mu _{\textrm{ac}} = d\nu _{\textrm{ac}}\). However, in the case \(d\mu _{\textrm{ac}} \ne d\nu _{\textrm{ac}}\), the Radon–Nikodym derivative \(\frac{d\mu _{\Delta x}}{d\nu _{\Delta x}}\) belongs to the interval \([1-\alpha , 1]\) rather than the set \(\{1-\alpha , 1 \}\), and hence the relation \(\frac{d\mu _{\textrm{ac}}}{d\nu _{\textrm{ac}}}(x) =1\) whenever \(u_x(x) < 0\) in Definition 2.1 (vi) will not be obeyed in general. Therefore, \({\tilde{P}}_{\Delta x}\) maps into a larger space \(\hat{\mathcal {D}} \supset \mathcal {D}\). Nevertheless, \({\tilde{P}}_{\Delta x}\) is relevant for numerical algorithms which are based on applying the projection operator after each time step \(\Delta t\) for \(\alpha \not =0\).
A closer look at (3.5) and (3.6) reveals that there are two possible sign choices on each interval \([x_{2j}, x_{2j+2}]\). Especially on a coarse grid, it is vital to make the right choice, as the following example shows.
Example 3.4
Consider the tuple (u, F, G), where
This models a scenario where wave breaking takes place at \(x=0\).
We discretize the domain \([-\frac{1}{2}, \frac{5}{2}]\) with \(\Delta x = \frac{1}{2}\), but for convenience we use \(x_j = - \frac{1}{2} + j \Delta x\) as our gridpoints. Using (3.5) and (3.6) we find
There are several possible ways to choose the signs in (3.8)–(3.9) and some of those are shown in Fig. 1: \(u_{\Delta x, 1}\) is based on using minus sign over \([x_{2j}, x_{2j+1}]\) and plus over \([x_{2j+1}, x_{2j+2}]\), while \(u_{\Delta x, 2}\) is based on using plus over \([x_{2j}, x_{2j+1}]\) and minus over \([x_{2j+1}, x_{2j+2}]\). By carefully choosing the signs, \(u_{\Delta x}\) fully overlaps with the exact solution. This choice which we denote as \(u_{\Delta x, \text {opt}}\), is given by
and equals \((u, F_{\textrm{ac}})\). The singular parts are approximated by
As the above example shows, choosing the best sign on each of the two grid cells in \([x_{2j}, x_{2j+2}]\) is vital, since it significantly affects the accuracy in the projection step and hence the whole algorithm, especially for a coarse grid. Instead of trial and error, which is never a good choice except for illustrative purposes, usually a selection criterion is imposed. In the implementation we decided to choose the sign over \([x_{2j}, x_{2j+1}]\) that minimizes the distance between \(u_{\Delta x}(x_{2j+1})\) and \(u(x_{2j+1})\), which can be formalized as follows. Introduce \(m \in \{0, 1\}\) such that for \(x \in [x_{2j}, x_{2j+1}]\) we may write (3.5) as
Subtracting \(u_{\Delta x}(x_{2j+1})\) from \(u(x_{2j+1})\), and finding the m that minimizes the distance can then be expressed as
Note that the sign over \([x_{2j+1}, x_{2j+2}]\) will be \((-1)^{k_{2j}+1}\).
3.2 Numerical Implementation of \(T_t\)
Since \(T_t\) associates to each piecewise linear initial data, the corresponding solution at time t, which again is piecewise linear, our numerical implementation of \(T_t\) will yield the exact solution with the projected initial data.
Fix a discretization parameter \(\Delta x > 0\) and an initial datum \((u, F, G)(0) \in \mathcal {D}_0\). Following Definition 3.1, the numerical Lagrangian initial data is given by
and we therefore first focus on the numerical implementation of L. We will observe that each component of \(X_{\Delta x}(0)\) is again a piecewise linear function. Moreover, the associated Lagrangian grid is non-uniform and has possibly a larger number of gridpoints than the original Eulerian grid.
3.2.1 Implementation of L
To avoid any ambiguity between breakpoints of a function and points of wave breaking, we denote the former as nodes in the following.
By construction, \(x+G_{\Delta x}(0, x)\) is an increasing, piecewise linear function with nodes situated at the points \(\{x_j\}_{j\in {\mathbb {Z}}}\), and hence \(y_{\Delta x}(0,\xi )\) is again an increasing and piecewise linear function due to Definition 2.3. Furthermore, \(y_{\Delta x}(0,\xi )\) is continuous and its nodes can be identified by finding all \(\xi \) which satisfy \(y_{\Delta x}(0,\xi )=x_j\) for some \(j\in {\mathbb {Z}}\) as follows. By Definition 2.3
and, due to \(X_{\Delta x}(0)\in \mathcal {F}_0\), we have
Since \(G_{\Delta x}(0, x)\) is continuous except possibly at the points \(\{x_{2j} \}_{j \in {\mathbb {Z}}}\), to every \(x_{2j+1}\) with \(j\in {\mathbb {Z}}\), there exists a unique \(\xi _{2j+1}\) such that \(y_{\Delta x}(0, \xi _{2j+1})=x_{2j+1}\) and, using (3.11), \(\xi _{2j+1}\) is given by
At the points \(\{x_{2j}\}_{j\in {\mathbb {Z}}}\), the function \(G_{\Delta x}(0, x)\) might have a jump. Therefore, there exists a maximal interval \(I_{2j}=[\xi _{2j}^{l}, \xi _{2j}^{r}]\) such that \(y_{\Delta x}(0,\xi )=x_{2j}\) for all \(\xi \in I_{2j}\), and using once more (3.11), \(\xi _{2j}^{l}\) and \(\xi _{2j}^{r}\) are given by
Note that \(\xi _{2j}^{l}=\xi _{2j}^{r}\) if and only if \(G_{\Delta x}(0, x)\) has no jump at \(x=x_{2j}\).
Set
Then the nodes of \(y_{\Delta x} (0,\xi )\) are situated at the points \(\{{\hat{\xi }}_{j}\}_{j\in {\mathbb {Z}}}\), as shown in Fig. 2. As mentioned earlier \(y_{\Delta x}(0)\) is piecewise linear and continuous, and hence also \(U_{\Delta x}(0)\), \(V_{\Delta x}(0)\), and \(H_{\Delta x}(0)\) are piecewise linear and continuous. Moreover, the nodes of \(X_{\Delta x}(0)\) are located at \(\{{\hat{\xi }}_{j}\}_{j\in {\mathbb {Z}}}\), since the nodes of \((u_{\Delta x}, F_{\Delta x}, G_{\Delta x})(0)\) are situated at \(x=x_{j}\), \(j\in {\mathbb {Z}}\). Thus, once the nodes are identified and the values of \(X_{\Delta x}(0)\) at these nodes are determined using Definition 2.3, the value of \(X_{\Delta x}(0,\xi )\) can be computed at any point \(\xi \).
Furthermore, each two grid cells \([x_{2j}, x_{2j+1}],[x_{2j+1}, x_{2j+2}]\) are mapped to 3 grid cells \([{\hat{\xi }}_{3j}, {\hat{\xi }}_{3j+1}], [{\hat{\xi }}_{3j+1}, {\hat{\xi }}_{3j+2}], [{\hat{\xi }}_{3j+2}, {\hat{\xi }}_{3j+3}]\) by L. In addition, while the Eulerian discretization is uniform with size \(\Delta x\), we now have
and hence the Lagrangian discretization is non-uniform.
3.2.2 The Wave Breaking Function
The next step is to compute the numerical wave breaking function \(\tau _{\Delta x}: \mathbb {R}\rightarrow [0, \infty ]\), using (2.3). For \(\xi \in [{\hat{\xi }}_{3j}, {\hat{\xi }}_{3j+1}]\), we have \(y_{\Delta x, \xi }(0,\xi ) = U_{\Delta x, \xi } (0,\xi )= 0\) and hence \(\tau _{\Delta x}(\xi )=0\) by (2.3). For \(\xi \in ({\hat{\xi }}_{3j+1}, {\hat{\xi }}_{3j+2})\), note that \(U_{\Delta x,\xi }(0,\xi )\) has the same sign as \(u_{\Delta x,x}(0, x)= Du_{2j}\mp q_{2j}\) in \((x_{2j}, x_{2j+1})\). Likewise, for \(\xi \in ({\hat{\xi }}_{3j+2}, {\hat{\xi }}_{3j+3})\), \(U_{\Delta x,\xi }(0,\xi )\) has the same sign as \(u_{\Delta x,x}(0, x)= Du_{2j}\pm q_{2j}\) in \((x_{2j+1}, x_{2j+2})\). Furthermore, \(U_{\Delta x,\xi }(0,\xi )=u_{\Delta x,x}(0, y_{\Delta x}(0,\xi ))y_{\Delta x,\xi }(0,\xi )\) for all \(\xi \in ({\hat{\xi }}_{3j+1}, {\hat{\xi }}_{3j+3})\backslash \{ {\hat{\xi }}_{3j+2}\}\). Therefore, introducing \(\tau _{3j+ \frac{3}{2}}\) and \(\tau _{3j+ \frac{5}{2}}\) as
the wave breaking function \(\tau _{\Delta x}\) for \(\xi \in [{\hat{\xi }}_{3j}, {\hat{\xi }}_{3j+3})\) is given by
Figure 3 illustrates the relation between the slopes of \(U_{\delta {x}}(0)\) and the value attained by \(\tau _{\delta {x}}\). Note that an interval where \(U_{\delta {x}}(0)\) is strictly increasing leads to an interval where \(\tau _{\delta {x}}\) is unbounded.
3.2.3 Implementation of \(S_t\)
We proceed by considering \(X_{\Delta x}(t)=S_t(X_{\Delta x}(0))\). Introduce the function
which we prefer to work with as \(\zeta _{\Delta x}\in L^{\infty }(\mathbb {R})\) in contrast to \(y_{\Delta x}\), see Definition 2.2. Furthermore, let
It then turns out to be advantageous to compute the time evolution of \({\hat{X}}_{\Delta x,\xi }\) rather than the one of \({\hat{X}}_{\Delta x}\), due to possible drops in \(V_{\Delta x,\xi }\), cf. (2.2c). However, this forces us to slightly change our point of view, since by differentiating (3.10) we find that \({\hat{X}}_{\Delta x, \xi }(0)\) is a piecewise constant function, whose discontinuities are situated at the nodes \({\hat{\xi }}_j\), where \(j\in {\mathbb {Z}}\). Therefore, we associate to each \(j \in {\mathbb {Z}}\),
This sequence is evolved numerically according to,
The above system is obtained by differentiating (2.2) with respect to \(\xi \). Since computing \({\hat{X}}_{\Delta x, \xi }(t,\cdot )\) exactly, yields a piecewise constant function whose discontinuities are again located at the nodes \(\{{\hat{\xi }}_j\}_{j \in {\mathbb {Z}}}\), considering the above sequence \(\{{\hat{X}}_{j, \xi }(t)\}_{j \in {\mathbb {Z}}}\) does not yield any additional error. Furthermore, the exact \({\hat{X}}_{\Delta x, \xi }(t,\cdot )\) can be read off from \(\{{\hat{X}}_{j, \xi }(t)\}_{j \in {\mathbb {Z}}}\).
Finally, we can exactly recover \({\hat{X}}_{\Delta x}(t)\), which is continuous with respect to \(\xi \), from \(\{{\hat{X}}_{j,\xi }(t)\}_{j \in {\mathbb {Z}}}\). Since the asymptotic behavior of \({\hat{X}}_{\Delta x}(t,\xi )\) as \(\xi \rightarrow \pm \infty \) changes in accordance with (2.2), this must also be taken into account by our algorithm. The fact that the initial data is in the space \({\mathcal {D}}\) combined with (2.2c) implies
where the abbreviation
has been introduced to ease the notation. Hence, for all \(j\in {\mathbb {Z}}\),
From (2.2d) it follows that
For \(U_{\Delta x}(t,\xi )\), we have, combining (2.2b) and Fubini’s theorem,
and for all \(j\in {\mathbb {Z}}\)
For \(\zeta _{\Delta x}(t,\xi )\), we find, using (2.2a) and \(\zeta _{\Delta x,-\infty }(0) = 0\),
and for all \(j\in {\mathbb {Z}}\),
Recalling that \({{\hat{X}}}_{\Delta x}(t)\) is piecewise linear and continuous, with nodes situated at \({\hat{\xi }}_j\) with \(j\in {\mathbb {Z}}\), we can now recover \({{\hat{X}}}_{\Delta x}(t,\xi )\) for all \(\xi \in {\mathbb {R}}\), and, using that \(y_{\Delta x}(t)= \zeta _{\Delta x}(t)+ \text { {id}}\), we end up with \(X_{\Delta x}(t,\xi )\) for all \(\xi \in {\mathbb {R}}\).
In practice, we have to limit the numerical approximation of the Cauchy problem (1.1) to initial data where \(u_{x}(0)\) and \(\mu _{\textrm{sing}}(0)\) have compact support. For such initial data there exists \(N \in {\mathbb {N}}\) such that
and by (3.14) we have
3.2.4 Implementation of M
Finally, to recover the solution in Eulerian coordinates we apply the mapping M to \(X_{\Delta x}(t)\), i.e.,
Here it is important to note that \(X_{\Delta x}(t)\) is piecewise linear and continuous, thus \(u_{\Delta x}(t,\cdot )\) is also piecewise linear and continuous, while \(F_{\Delta x}(t, \cdot )\) and \(G_{\Delta x}(t, \cdot )\) are piecewise linear, increasing and in general only left-continuous. Furthermore, their nodes are situated at the points \(\{y_{\Delta x}(t, {\hat{\xi }}_j)\}_{j \in {\mathbb {Z}}}\). Numerically, we therefore apply a piecewise linear reconstruction.
Given \(x \in \mathbb {R}\), there exists \(j \in {\mathbb {Z}}\) such that \(x \in (y_{\Delta x}(t, \hat{\xi }_j), y_{\Delta x}(t, \hat{\xi }_{j+1})]\).
If \(x=y_{\Delta x}(t, \hat{\xi }_{j+1})\), we have, by the definition of M,
This covers also the case where \(\hat{\xi }_j = \hat{\xi }_{j+1}\), and the case where wave breaking occurs, i.e., \(y_{\Delta x}(t, \hat{\xi }_j) = y_{\Delta x}(t, \hat{\xi }_{j+1})\).
If \(x\in (y_{\Delta x}(t, \hat{\xi }_j), y_{\Delta x}(t, \hat{\xi }_{j+1}))\), observe that
and
Therefore, for \(x\in (y_{\Delta x}(t, \hat{\xi }_j), y_{\Delta x}(t, \hat{\xi }_{j+1}))\) the linear interpolation, which coincides with \(M\left( X_{\Delta x}(t) \right) \), is given by
where \(D_+^{\xi }y_{\Delta x}(t, \hat{\xi }_j) = y_{\Delta x}(t, \hat{\xi }_{j +1 }) - y_{\Delta x}(t, \hat{\xi }_{j})\). Finally, note that the Eulerian grid changes significantly with respect to time. First of all the number of grid cells is changing over time and secondly, the grid does not remain uniform.
4 Convergence of the Numerical Method
In this section we prove that our family of approximations \(\{ \left( u_{\Delta x}, F_{\Delta x}, G_{\Delta x} \right) \}_{\Delta x > 0}\) converges to the \(\alpha \)-dissipative solution of (1.1). We start by proving convergence of the projected initial data in Eulerian coordinates. Thereafter, we show that this induces convergence, initially and at later times, towards the unique \(\alpha \)-dissipative solution in Lagrangian coordinates. Finally, we investigate in what sense convergence in Lagrangian coordinates carries over to Eulerian coordinates.
4.1 Convergence of the Initial Data in Eulerian Coordinates
Let \(P_{\Delta x}\) be the projection operator given by Definition 3.2. We have the following result.
Proposition 4.1
For \((u,\mu , \nu ) \in \mathcal {D}_0\), let \((u_{\Delta x}, F_{\Delta x}, G_{\Delta x})=P_{\Delta x} \left( (u,F, G) \right) \). Then
Proof
Let \(x \in [x_{2j}, x_{2j+1}]\). By (3.5), we have
where \(q_{2j}\) is defined by (3.4). The second term is bounded by
The first term can be estimated as follows:
where we applied the Cauchy–Schwarz inequality. Combining (4.2) and (4.3) yields
By a similar argument for \(x \in [x_{2j+1}, x_{2j+2}]\), (4.1a) is established.
Using (4.4) we get
which yields (4.1b).
Next, we show (4.1c). For the \(L^1(\mathbb {R})\)-estimate, we have
and for the \(L^2(\mathbb {R})\)-estimate,
To prove (4.1e), it suffices to show that \(\mu _{\Delta x} \rightarrow \mu \) vaguely as \(\Delta x \rightarrow 0\), see [7, Prop. 7.19]. That is,
for all \(\phi \in C_0(\mathbb {R})\), where \(C_0(\mathbb {R})\) denotes the space of continuous functions vanishing at infinity. Furthermore, since \(C_c^{\infty }(\mathbb {R})\) is dense in \(C_0(\mathbb {R})\), cf. [7, Prop 8.17], it suffices to show that (4.5) holds for all \(\phi \in C_c^{\infty }(\mathbb {R})\).
Let \(\phi \in C_c^{\infty }(\mathbb {R})\). Combining integration by parts and (4.1c) with \(p=1\), we find
and consequently \(\mu _{\Delta x} \rightarrow \mu \) vaguely as \(\Delta x\rightarrow 0\).
Since \(F=G\) and \(F_{\Delta x}=G_{\Delta x}\) by assumption, (4.1d) and (4.1f) hold. \(\square \)
Next we establish convergence of the spatial derivative \(u_{\Delta x, x}\).
Lemma 4.2
Let \((u,F,G)\in \mathcal {D}_0\) and \((u_{\Delta x}, F_{\Delta x}, G_{\Delta x})=P_{\Delta x} \left( (u,F, G) \right) \), then
Proof
We apply the Radon–Riesz theorem. Thus, we have to show that \(\Vert u_{\Delta x, x} \Vert _2 \rightarrow \Vert u_x \Vert _2\) and \(u_{\Delta x, x} \rightharpoonup u_x\) in \(L^2(\mathbb {R})\). A direct calculation, using (3.5), yields
For the weak convergence, it suffices to consider test functions \(\phi \in C_c^\infty ({\mathbb {R}})\), as \(C_c^{\infty }(\mathbb {R})\) is dense in \(L^2(\mathbb {R})\). Let \(\phi \in C_c^{\infty }(\mathbb {R})\). Integration by parts combined with the Cauchy–Schwarz inequality and (4.1b) yields
Thus \(u_{\Delta x, x} \rightharpoonup u_{x}\) in \(L^2(\mathbb {R})\) as \(\Delta x\rightarrow 0\). \(\square \)
The auxiliary function f(Z) will be an essential part in the upcoming convergence analysis, when we want to relate convergence in Eulerian coordinates to convergence in Lagrangian coordinates. It is defined as
where
and
Recalling that \(d\mu _{\textrm{ac}}=u_x^2dx\) by Definition 2.1 (iv), and \(\mu _{\textrm{ac}}=\nu _{\textrm{ac}}\) for \((u, \mu , \nu ) \in \mathcal {D}_0\), we establish the following \(L^1(\mathbb {R})\)-estimate.
Lemma 4.3
Let \((u,F,G)\in \mathcal {D}_0\) and \((u_{\Delta x}, F_{\Delta x}, G_{\Delta x})=P_{\Delta x} \left( (u,F, G) \right) \). Then f(Z), \(f(Z_{\Delta x}) \in L^1({\mathbb {R}})\), and
Proof
We only show that \(f(Z)\in L^1(\mathbb {R})\), since the argument for \(f(Z_{\Delta x})\) is exactly the same. By construction \(f(Z)(x)\le u_x^2(x)\) for all \(x\in \mathbb {R}\) and, since \(u_x\in L^2(\mathbb {R})\), it follows that
and hence \(f(Z)\in L^1(\mathbb {R})\).
For (4.7), we use a splitting based on \({\tilde{\Omega }}_d(Z)\) and \({\tilde{\Omega }}_c(X)\). If \(x \in \tilde{\Omega }_{d}(Z)\cap {\tilde{\Omega }}_{d}(Z_{\Delta x})\), then
and, by the Cauchy–Schwarz inequality and (4.8), we obtain
We can proceed similarly for \(x \in {\tilde{\Omega }}_c(Z) \cap \tilde{\Omega }_c(Z_{\Delta x})\).
Finally, consider \(x \in \left( {\tilde{\Omega }}_c(Z)\cap \tilde{\Omega }_d(Z_{\Delta x})\right) \cup \left( {\tilde{\Omega }}_d(Z)\cap {\tilde{\Omega }}_c(Z_{\Delta x})\right) \). By symmetry, it is sufficient to consider \(x \in {\tilde{\Omega }}_c(Z)\cap {\tilde{\Omega }}_d(Z_{\Delta x})\), for which
since \(u_{x}(x) \ge 0\) and \(u_{\Delta x, x}(x) <0\). Therefore, by the Cauchy–Schwarz inequality,
Combining (4.9) and (4.10) with analogous estimates for the other cases yields (4.7). \(\square \)
4.2 Convergence in Lagrangian Coordinates
We start by showing convergence of the initial data in Lagrangian coordinates.
Lemma 4.4
Given \((u, \mu , \nu ) \in \mathcal {D}_0\), let \(X = \left( y, U, V, H \right) = L \left( (u, \mu , \nu ) \right) \) and \(X_{\Delta x}= (y_{\Delta x}, U_{\Delta x}, V_{\Delta x}, H_{\Delta x}) = L \circ P_{\Delta x} \left( (u, \mu , \nu ) \right) \), then
Proof
Given \(\xi \in \mathbb {R}\), there exists \(j \in {\mathbb {Z}}\) such that \(\xi \in [\hat{\xi }_{3j}, \hat{\xi }_{3j+3})\). Moreover, by construction, y and \(y_{\Delta x}\) are continuous, increasing and satisfy
as discussed in Sect. 3.2. Thus
and, as X and \(X_{\Delta x}\) belong to \(\mathcal {F}_0\), we have shown both (4.11a) and (4.11c).
To prove (4.11b), note that by (2.1a), to any \(x \in {\mathbb {R}}\) there exist characteristic variables \(\xi _{\Delta x}\) and \(\xi \), such that \(y_{\Delta x}( \xi _{\Delta x}) = x=y(\xi )\), and hence
By (2.1b), we have, \(u(y(\xi )) = U(\xi )\) and \(u_{\Delta x}(y_{\Delta x}(\xi )) = U_{\Delta x}(\xi )\). Therefore, by rearranging (4.13), we obtain
The first term is bounded by (4.1a), while for the second term we apply the Cauchy–Schwarz inequality, yielding
Thus,
\(\square \)
For the initial derivatives we have the following convergence result.
Lemma 4.5
Given \((u, \mu , \nu ) \in \mathcal {D}_0\), let \(X= (y, U, V, H) = L \left( (u, \mu , \nu ) \right) \) and \(X_{\Delta x} = \left( y_{\Delta x}, U_{\Delta x}, V_{\Delta x}, H_{\Delta x} \right) = L \circ P_{\Delta x} \left( (u, \mu , \nu ) \right) \), then as \({\Delta x}\rightarrow 0\),
Proof
For a proof of \(H_{\Delta x, \xi } \rightarrow H_{\xi }\) and \(U_{\Delta x, \xi } \rightarrow U_{\xi }\) in \(L^2(\mathbb {R})\) we refer to [8, Sec. 5]. As \(X\in \mathcal {F}_0^0\) and, in particular, \(U_{\xi }^2 = y_{\xi }H_{\xi }\), we have
Since (4.15) also holds for \(X_{\Delta x}\), we obtain, using the Cauchy–Schwarz inequality,
and \(H_{\Delta x, \xi } \rightarrow H_\xi \) in \(L^1(\mathbb {R})\). Lastly, (4.14a) follows from (4.14c), as \(X_{\Delta x}, X \in \mathcal {F}_0^0\). \(\square \)
We proceed by introducing the stability function g(X), which is the key element when showing convergence at later times in Lagrangian coordinates. In particular, it describes the loss of energy at wave breaking in a continuous way, in contrast to the actual energy density \(V_{\xi }\), which drops abruptly at wave breaking.
Let
where
Note that for each \(\xi \in \Omega _c\) no wave breaking occurs in the future and hence \(V_\xi (t,\xi )\) is continuous forward in time, while for \(\xi \in \Omega _d\) wave breaking occurs in the future and hence \(V_\xi (t,\xi )\) might be discontinuous forward in time.
The function g(X) is then defined as
Note that the Eulerian counterpart of g(X) is the previously defined function f(Z), cf. (4.6). The relation between the two functions is clarified in the following remark.
Remark 4.6
If \((u,\mu ,\nu )\in \mathcal {D}_0\), \(\nu \) is purely absolutely continuous, and \(X=L((u,\mu ,\nu ))\), the functions (4.17) and (4.6) are related through
The following result is the final one concerning convergence of the initial data in Lagrangian coordinates.
Proposition 4.7
Given \((u, \mu , \nu ) \in \mathcal {D}_0\), let \(X= (y, U, V, H) = L \left( (u, \mu , \nu ) \right) \) and \(X_{\Delta x} = \left( y_{\Delta x}, U_{\Delta x}, V_{\Delta x}, H_{\Delta x} \right) = L \circ P_{\Delta x} \left( (u, \mu , \nu ) \right) \), then as \({\Delta x}\rightarrow 0\),
Proof
To begin with observe that for any \(X\in \mathcal {F}_0^0\)
since \(0\le V_\xi (\xi )=H_\xi (\xi )\le 1\) for all \(\xi \in \mathbb {R}\), which follows from (2.1c) and the fact that both y and H are increasing functions. We can therefore apply the Radon–Riesz theorem to establish \(g(X_{\Delta x})\rightarrow g(X)\) in \(L^2(\mathbb {R})\). Accordingly, we split the proof of (4.18) into three parts, i) \(L^2\)-norm convergence, ii) weak \(L^2\)-convergence, and iii) \(L^1\)-convergence.
i) Verification of \(\Vert g(X_{\Delta x}) \Vert _2 \rightarrow \Vert g(X) \Vert _{2}\). Note that since \(X_{\Delta x}\) and X belong to \(\mathcal {F}_0^0\), we have
where we used (4.15) in the last step.
To estimate (4.19a) and (4.19b), introduce
and \(B=y({\mathcal {S}})\). Then it has been shown in the proof of [12, Thm. 27] that
where \(\mu \vert _B\) denotes the restriction of \(\mu \) to B, that is \(\mu \vert _B(E)=\mu (E\cap B)\) for any Borel set E. Thus
Along the same lines, by defining
we have
Since \(\Omega _d(X)=\Omega _d(X)\cap {\mathcal {S}}^c\) and \(\Omega _d(X_{\Delta x})=\Omega _d(X_{\Delta x})\cap {\mathcal {S}}^c_{\Delta x}\), due to Definition 2.2 (iii), (4.21) implies that
Furthermore, \(U_{\xi }(\xi ) = u_x (y(\xi ))y_{\xi }(\xi )\) for almost every \(\xi \in \Omega _d(X)\) and likewise for \(U_{\Delta x,\xi }\). Hence, we can replace in the above integrals \(y\left( \Omega _d(X) \right) \) and \(y_{\Delta x}\left( \Omega _d(X_{\Delta x}) \right) \), by \({\tilde{\Omega }}_d(Z)\) and \({\tilde{\Omega }}_{d}(Z_{\Delta x})\), respectively, and end up with
The terms on the right hand side can be estimated using the same approach as in the proof of Lemma 4.3, which yields
For (4.19b), we get
The two terms in (4.26a) have a similar structure and we therefore only consider the first one. We have
since \(U_{\xi }^2 = y_{\xi }V_{\xi }\) and \(0 \le y_{\xi } \le 1\). Estimating (4.26b) in much the same way, yields
Finally, combining (4.19), (4.25) and (4.27), we have shown that
As \(H_{\Delta x, \xi } \rightarrow H_{\xi }\) and \(U_{\Delta x, \xi } \rightarrow U_{\xi }\) in \(L^2(\mathbb {R})\) by Lemma 4.5 and \(u_{\Delta x, x} \rightarrow u_{x}\) in \(L^2(\mathbb {R})\) by Lemma 4.2, it holds that \(\Vert g(X_{\Delta x})\Vert _2 \rightarrow \Vert g(X) \Vert _2\).
ii) We show that \(g(X_{\Delta x}) \rightharpoonup g(X)\) in \(L^2(\mathbb {R})\). To that end, we interpret \(g(X_{\Delta x})\) and g(X) as positive Radon measures with the associated functions
If we show that \({\hat{G}}_{\Delta x} \rightarrow {\hat{G}}\) pointwise, then [7, Prop. 7.19] implies that \(g(X_{\Delta x}) \rightarrow g(X)\) vaguely, and as a consequence, \(g(X_{\Delta x}) \rightharpoonup g(X)\) in \(L^2(\mathbb {R})\).
Let \(\xi \in [\hat{\xi }_{3j}, \hat{\xi }_{3j+3})\), and note that due to Definition 3.2, Definition 2.3, (3.12)–(3.13), and (4.21),
Thus,
For \(I_1\) note that using the same argument as the one leading to (4.24), we obtain
and due to Lemma 4.3,
The term \(I_2\) can be estimated as follows,
The first term is bounded from above by \(\Vert H - H_{\Delta x} \Vert _{\infty } \le 2\Delta x\), due to (4.11c). For the second term, let \(x= y(\xi )\) and \(x_{\Delta x}=y_{\Delta x}(\xi )\), which both belong to \([x_{2j}, x_{2j+2}]\). Using once more the same argument as the one leading to (4.24), \(\nu _{\text {ac}}=u_x^2\), and \(\nu _{\Delta x, \text {ac}}=u_{\Delta x, x}^2\), we end up with
and therefore
Thus,
By Lemma 4.2, \(u_{\Delta x, x} \rightarrow u_x\) in \(L^2(\mathbb {R})\), and, since \(F_{\text {ac}}\) is continuous, also the third term tends to zero as \(\Delta x \rightarrow 0\). Therefore, (4.28) implies that \({\hat{G}}_{\Delta x} \rightarrow {\hat{G}}\) pointwise, which gives, see [7, Prop. 7.19], that
Since g(X) and \(g(X_{\Delta x})\) belong to \(L^2(\mathbb {R})\) and \(C_c^\infty (\mathbb {R})\) is a dense subset of \(L^2(\mathbb {R})\), (4.29) holds for all \(\phi \in L^2(\mathbb {R})\) and \(g(X_{\Delta x})\rightharpoonup g(X)\) in \(L^2(\mathbb {R})\).
Consequently, combining i) and ii), the conditions of the Radon–Riesz theorem are met and \(g(X_{\Delta x})\rightarrow g(X)\) in \(L^2(\mathbb {R})\).
iii) It is left to show that \(g(X_{\Delta x}) \rightarrow g(X)\) in \(L^1(\mathbb {R})\). Again we will use a splitting based on the sets \({\mathcal {S}}\) and \({\mathcal {S}}_{\Delta x}\) defined in (4.20) and (4.22), respectively.
Since \(g(X_{\Delta x})(\xi ) = g(X) (\xi )= 1\) for all \(\xi \in {\mathcal {S}}\cap {\mathcal {S}}_{\Delta x}\), we have
By the Cauchy–Schwarz inequality and (4.23), (4.30a) satisfies
and we treat the first term in (4.30) the same way.
The term (4.30b), on the other hand, requires a bit more work. Introduce the set \(A = \{\xi : H_{\xi }(\xi ) \ge \frac{1}{2}\}\), which satisfies, due to Chebyshev’s inequality with \(p=1\),
Furthermore, let \(E_{\Delta x} = {\mathcal {S}}^c \cap {\mathcal {S}}_{\Delta x}^c\), so that (4.30b) can be written as
For (4.33), the Cauchy–Schwarz inequality and (4.32) imply
Combining (4.30), (4.31) (4.33), and (4.34) and recalling that \(g(X_{\Delta x})\rightarrow g(X) \) in \(L^2(\mathbb {R})\), it remains to show that (4.33) tends to zero as \(\Delta x\rightarrow 0\). We follow the proof of [11, Lem. 7.3] closely. For almost every \(\xi \in E_{\Delta x}\), we have
and
Thus combining (4.35) and (4.36),
Concerning the integral of (4.37a), note that \(f(Z) \in L^1(\mathbb {R})\) by Lemma 4.3. Thus, given \(\epsilon > 0\) there exists \(\psi \in C_c(\mathbb {R})\) such that \(\Vert f(Z) - \psi \Vert _{L^1(\mathbb {R})} \le \epsilon \), since \(C_c(\mathbb {R})\) is dense in \(L^1(\mathbb {R})\), and
since \(0\le y_\xi \), \(y_{\Delta x,\xi }\le 1\). As \(y_{\Delta x} \rightarrow y\) in \(L^{\infty }(\mathbb {R})\) by (4.11a), the support of \(\psi (y_{\Delta x})\) is contained inside some compact set which can be chosen independently of \(\Delta x\). Therefore, by the Lebesgue dominated convergence theorem, \(\psi (y_{\Delta x}) \rightarrow \psi (y)\) in \(L^1(\mathbb {R})\). Consequently, the left hand side of (4.38) vanishes as \(\epsilon \rightarrow 0\) and \(\Delta x \rightarrow 0\).
For the term (4.37b) we use the change of variables \(x=y_{\Delta x}(\xi )\), \(0 \le y_{\xi } \le 1\), and Lemma 4.3 to deduce that
For (4.37c) note that
which implies
Since \(u_x^2 \in L^1(\mathbb {R})\), an argument similar to the one for the integrals of (4.37a) and (4.37b) shows that the term on the left hand side tends to zero as \(\Delta x\rightarrow 0\).
Finally, for (4.37d), we observe that
and hence (4.37d) is bounded from above by
which is a combination of the integrals of (4.37a) and (4.37b) which tend to zero as \({\Delta x}\rightarrow 0\). This finishes the proof of \(g(X_{\Delta x}) \rightarrow g(X)\) in \(L^1(\mathbb {R})\). \(\square \)
To establish convergence at later times in Lagrangian coordinates, we equip the set \(\mathcal {F}\) with the following metric, which has been introduced in [13, Def. 4.6].
Definition 4.8
Let \(d: \mathcal {F}\times \mathcal {F}\rightarrow [0, \infty )\) be defined by
This metric has one major drawback: it separates solutions in Lagrangian coordinates belonging to the same equivalence class. Nevertheless, it is well suited for us, since we are only interested in comparing \(X_{\Delta x}(t) = S_t (X_{\Delta x}(0))\) with \(X(t) = S_t (X(0))\) and not their respective equivalence classes.
Theorem 4.9
Given \((u_0, \mu _0, \nu _0) \in \mathcal {D}_0\) and \(t\ge 0\), let \(X(t) = S_t \circ L \left( (u_0, \mu _0, \nu _0 )\right) \) and \(X_{\Delta x}(t) = S_{t} \circ L \circ P_{\Delta x} \left( (u_0, \mu _0, \nu _0) \right) \), then as \({\Delta x}\rightarrow 0\),
Furthermore,
Proof
Let \(X(0)=L((u_0, \mu _0,\nu _0))\) and \(X_{\Delta x}(0)= L\circ P_{\Delta x}((u_0,\mu _0,\nu _0))\). Combining Lemma 4.4, Lemma 4.5, and Proposition 4.7 yields \(d(X(0), X_{\Delta x}(0)) \rightarrow 0\). Furthermore, see [13, Thm. 4.18],
and hence (4.39a)–(4.39e) hold.
Since \(g(X)(t,\cdot )-g(X_{\Delta x})(t,\cdot )\) and \(H(t,\cdot )-H_{\Delta x}(t,\cdot )\) are time-independent, (4.39f) and (4.40a) follow immediately from Proposition 4.7 and (4.11c).
It remains to show (4.40b). Recalling (4.16), introduce
Using Fubini’s theorem, we write
Since no wave breaking occurs for \(\xi \in \Omega _{c,c}(0)\),
The terms (4.41a) and (4.41b) can be treated similarly, so we only consider (4.41a). Since any \(\xi \in \Omega _d(X_{\Delta x}(0))\) might enter \(\Omega _c\) within the time interval [0, t], we have
where we used in the last step that X(0) and \(X_{\Delta x}(0)\) belong to \(\mathcal {F}_0^0\).
For (4.41c) we use a similar decomposition, yielding
For the first term we have
For (4.44a), recalling (4.17) and that X(0) and \(X_{\Delta x}(0)\) belong to \(\mathcal {F}_0^0\), we have
For the last term in the above inequality, observe that \(U_\xi (t,\xi )\) and \(U_{\Delta x, \xi }(t,\xi )\) are increasing functions, cf. (2.2b), which equal zero at \(t= \tau (\xi )\) and \(t= \tau _{\Delta x}(\xi )\), respectively, for \(\xi \in \Omega _{d, d}(0)\). Furthermore, (2.2b) implies
and thus
The term (4.44b) is bounded similarly.
To estimate (4.44c), we split \(\Omega _{d, d}(0) \cap \Omega _{c, c}(t)\) into two sets,
where
We only present the details for the integral over A(t), since the other one can be treated similarly. Following closely the argument leading to (4.46), we get
Combining (4.42)–(4.43), (4.45), (4.46)–(4.47), and analogous estimates for the remaining cases, yields
where
Furthermore, \(\textrm{meas} \! \left( \{\xi \in \mathbb {R}: \tau (\xi ) \le t \}\right) \le (1 + \frac{1}{4}t^2)H_{\infty }(0)\), and likewise for the set \(\{\xi : \tau _{\Delta x}(\xi ) \le t \}\), by [13, Cor. 2.4]. Hence
Thus, by applying the Cauchy–Schwarz inequality to the last term in (4.48) and inserting (4.49), we obtain (4.40b). \(\square \)
Corollary 4.10
Let \((u_0, \mu _0, \nu _0) \in \mathcal {D}_0\) and \(t \ge 0\). Set \(X(t) = S_t \circ L \left( (u_0, \mu _0, \nu _0 ) \right) \) and \(X_{\Delta x}(t) = S_{t} \circ L \circ P_{\Delta x} \left( (u_0, \mu _0, \nu _0 )\right) \). Then there exists a subsequence \(\{\Delta x_m \}_{ m \in {\mathbb {N}}}\) with \(\Delta x_m \rightarrow 0\) as \(m \rightarrow \infty \), such that for a.e. \(t \in [0, \infty )\) we have
Proof
By (4.40b) and [7, Cor. 2.32], there exists a subsequence \(\{V_{\Delta x_m, \xi }\}_{m \in {\mathbb {N}}}\) such that \(V_{\Delta x_m, \xi }(t,\cdot ) \rightarrow V_{\xi }(t,\cdot )\) in \(L^1(\mathbb {R})\) for a.e. \(t \in [0, \infty )\). Let \(N \subset [0, \infty )\) be the null set of times for which the convergence does not hold. Thus, for \(t \in N^c\) we have
\(\square \)
Observe that the times for which the convergence in Corollary 4.10 fails depend on the particular chosen subsequence. Therefore, there is no natural way to extend this convergence to the whole sequence.
4.3 Convergence of the \(\alpha \)-Dissipative Solution in Eulerian Coordinates
Finally, we can examine in what sense convergence in Lagrangian coordinates, given by Theorem 4.9 and Corollary 4.10, carries over to Eulerian coordinates.
Lemma 4.11
Given \((u_0, \mu _0, \nu _0) \in \mathcal {D}_0\), let \((u, \mu , \nu )(t)= T_t \left( (u_0, \mu _0, \nu _0 )\right) \) and \((u_{\Delta x}, \mu _{\Delta x}, \nu _{\Delta x})(t) = T_t \circ P_{\Delta x} \left( (u_0, \mu _0, \nu _0 )\right) \) for \(t \in [0, \infty )\). Then
Proof
Following the same line of reasoning as in the proof of (4.11b), yields (4.51).
Before proving (4.52), recall that by Definition 2.4, we have
and likewise for \(G_{\Delta x}(t,x)\). Thus, the change of variables \(x=y(t,\xi )\) yields
Following the steps in the proof of [13, (2.15)], we can establish
for almost every \(\xi \in \mathbb {R}\), which implies that
where we used (4.40a) and (4.12).
For (4.53a), note that, by construction, \(y_{\Delta x}(t,\xi )\) and \(H_{\Delta x}(t,\xi )\) are piecewise linear functions with nodes at the points \({\hat{\xi }}_j\) with \(j\in {\mathbb {Z}}\). Furthermore, \(H_{\Delta x}(t, \xi ) - G_{\Delta x}(t, y_{\Delta x}(t, \xi ))>0\) on an interval I if and only if \(y_{\Delta x}(t,\xi ) \) is constant on this interval. Assuming that such an interval I exists, an upper bound on its length can be established as follows. Observe that
and hence, due to (4.54),
Thus, it holds that
To estimate (4.53b), we exploit that \(G_{\Delta x}(t)\in \text {BV}(\mathbb {R})\), since it is monotonically increasing and bounded. Introducing \(\Theta = \Vert y(t) - y_{\Delta x}(t) \Vert _{\infty }\), then yields
where we applied Tonelli’s theorem to exchange the order of the sum and integral.
Combining (4.53), (4.55), (4.56), and (4.57), we end up with (4.52). \(\square \)
Theorem 4.12
Given \((u_0, \mu _0, \nu _0) \in \mathcal {D}_0\), let \((u, \mu , \nu )(t)= T_t \left( (u_0, \mu _0, \nu _0)\right) \) and \((u_{\Delta x}, \mu _{\Delta x}, \nu _{\Delta x})(t) = T_t \circ P_{\Delta x} \left( (u_0, \mu _0, \nu _0)\right) \) for \(t \in [0, \infty )\). Then, as \(\Delta x\rightarrow 0\)Footnote 1,
Furthermore, there exists a subsequence \(\{\Delta x_m \}_{m \in {\mathbb {N}}}\) with \(\Delta x_m \rightarrow 0\) as \(m \rightarrow \infty \), and a null set \(N \subset [0, \infty )\) such that for all \(t \in N^c\) we have as \(\Delta x_m\rightarrow 0\)
Note that in the case of conservative solutions, \(F_{\Delta x} = G_{\Delta x}\) and hence (4.52) implies that we have \(L^1(\mathbb {R})\)-convergence of \(F_{\Delta x}(t)\) for all \(t \ge 0\).
Proof
Note that \(U_{\Delta x}(t) \rightarrow U(t)\) and \(y_{\Delta x}(t) \rightarrow y(t)\) in \(L^{\infty }(\mathbb {R})\) as \(\Delta x\rightarrow 0\) by Theorem 4.9, which combined with (4.51) yields (4.58).
To prove (4.59), observe that for any \(\psi \in C_c^{\infty }(\mathbb {R})\),
and hence, using (4.58), the right hand side tends to zero as \(\Delta x\rightarrow 0\). Since \(C_c^\infty (\mathbb {R})\) is a dense subset of \(L^2(\mathbb {R})\), we end up with \(u_{\Delta x, x}(t) \rightharpoonup u_x(t)\) in \(L^2(\mathbb {R})\) as \(\Delta x \rightarrow 0\).
Next, note that \(y_{\Delta x}(t) \rightarrow y(t)\) in \(L^{\infty }(\mathbb {R})\) and \(y_{\Delta x,\xi }(t) \rightarrow y_\xi (t)\) in \(L^{2}(\mathbb {R})\) as \(\Delta x\rightarrow 0\) by Theorem 4.9, and hence (4.60) follows from (4.52)
To prove (4.61), let \(\phi \in C_b(\mathbb {R})\). Then, by Definition 2.4, we have
Since \(y_{\Delta x}(t) \rightarrow y(t)\) in \(L^{\infty }(\mathbb {R})\) by (4.39a) and \(\phi \in C_b(\mathbb {R})\), it follows that \(\phi (y_{\Delta x}(t)) \rightarrow \phi (y(t))\) pointwise a.e.. Furthermore, \(|\phi (y(t)) - \phi (y_{\Delta x}(t))|H_{\xi }(t) \le 2\Vert \phi \Vert _{\infty }H_{\xi }(t)\) and hence, by the dominated convergence theorem, (4.65a) vanishes as \(\Delta x \rightarrow 0\).
For (4.65b) on the other hand, we observe that \(\phi (y_{\Delta x}(t))\in L^{\infty }(\mathbb {R})\) and \(H_{\Delta x, \xi }(t)\rightarrow H_{\xi }(t)\) in \(L^1(\mathbb {R})\), see (4.39e), imply
as \(\Delta x\rightarrow 0\). Thus the left hand side of (4.65) tends to zero as \(\Delta x \rightarrow 0\) for any \(\phi \in C_b(\mathbb {R})\) and we have shown (4.61).
As \(g(X_{\Delta x}(t)) \rightarrow g(X(t))\) in \(L^1(\mathbb {R})\) by (4.39f), the proof of (4.62) is completely analogous to the one of (4.61).
To show the remaining part of the theorem, recall Corollary 4.10, which ensures the existence of a subsequence \(\{\Delta x_m \}_{m \in {\mathbb {N}}}\) with \(\Delta x_m \rightarrow 0\) as \(m \rightarrow \infty \), such that \(V_{\Delta x_m, \xi } (t) \rightarrow V_{\xi }(t)\) in \(L^1(\mathbb {R})\) for all \(t \in N^c\), where \(N\subset [0, T]\) is a null set. Therefore, applying a similar argument to the one used for proving (4.61), we can establish that \(\mu _{\Delta x_m}(t) \Longrightarrow \mu (t)\) for all \(t\in N^c\). Choosing \(\phi = 1 \in C_b(\mathbb {R})\), we obtain (4.64). Furthermore, as \(C_0(\mathbb {R}) \subset C_b(\mathbb {R})\), we have \(\mu _{\Delta x_m}(t) \rightarrow \mu (t)\) vaguely for all \(t \in N^c\) and (4.63) holds by [7, Prop. 7.19]. \(\square \)
Remark 4.13
For each \(t\in [0,T]\) and \(\Delta x>0\), \(F_{\Delta x}(t)\in \text {BV}(\mathbb {R})\). Thus, by Helly’s selection principle, see [2, App. II.], there exists for every \(t \in N\) a subsequence \(\{F_{\Delta x_k}(t) \}_{k \in {\mathbb {N}}}\) such that \(F_{\Delta x_k}(t)\) converges pointwise almost everywhere to a function of bounded variation.
Remark 4.14
The projection operator \(P_{\Delta x}\) is constructed with focus on accurately approximating the wave profile u, preserving the total energy and ensuring that
Therefore, the control over the size of the sets \(\{x: u_x(x) \le - \frac{2}{t}< u_{\Delta x, x}(x) < 0\}\) and \(\{x: u_{\Delta x, x}(x) \le - \frac{2}{t}< u_{x}(x) < 0\}\) is limited and as a consequence it prevents the convergence of the energy density \(\mu _{\Delta x}\) for every fixed time.
5 Numerical Experiments
In this section we present three examples which highlight different challenges for the numerical algorithm. The first two examples are also considered in [5, 14] and display solutions with multipeakon and cusp initial data, respectively. In the first example, wave breaking happens twice; initially, and also at a later time, and at each occurrence a finite amount of energy concentrates on a set of Lebesgue measure zero. For the second example on the other hand, the wave breaking times accumulate in the interval [0, 3], but at every time only an infinitesimal amount of energy concentrates. The behavior of this latter example is thoroughly studied in [5, Sec. 5.3]. In both cases, expressions for the exact solutions in Eulerian coordinates can be found.
In the third example, we consider a cosine wave profile as initial data. Here, an expression for the Lagrangian solution is known, but the exact expression for u in Eulerian coordinates is not available for comparison. The exact solution possesses accumulating wave breaking times, and, similarly to the cusp initial data, experiences wave breaking continuously in time, but now over the unbounded interval \([\frac{2}{\pi }, \infty )\). Furthermore, in contrast to the cusp example, the rate at which the total energy dissipates is nonlinear, making this an interesting challenge for the algorithm.
We present plots of both the time evolution and of the errors \(\sup _{t \in [0,T]}\Vert u(t) - u_{\Delta x}(t) \Vert _{\infty }\) and \(|F_{\infty }(T) - F_{\Delta x, \infty }(T)|\) for a chosen time T. In the first two examples, the errors are computed by comparing the exact and numerical solutions at the gridpoints of a uniform mesh which is a refinement of the finest mesh we use for the numerical approximations. The value of \(u_{\Delta x}\) at the gridpoints is found by using (3.15). As we do not have a closed form expression for the exact solution in Eulerian coordinates in the third example, Example 5.3, we instead use
to measure the error introduced by the projection operator \(P_{{\Delta x}}\). Note that due to (4.51), this is an upper bound for the \(L^\infty \)-error of \(u_{\Delta x}\) in Eulerian coordinates.
Example 5.1
(Multipeakon initial data) Consider the Cauchy problem from Example 3.4 with \(\alpha =\frac{1}{2}\), but now set \(\nu _0 = \mu _0\), i.e.,
The exact solution experiences wave breaking at \(t=2\).
In Fig. 4 the numerical solutions \((u_{\Delta x}, F_{\Delta x})\) with \(\Delta x = 2.35 \cdot 10^{-1}\) and \(\Delta x = 8.02 \cdot 10^{-3}\) are compared to the exact solution (u, F) at \(t=0\), at the wave breaking time \(t=2\), and at \(t=4\).
Figure 5 displays the numerically computed errors. In this case the numerically computed convergence order, which is \(\ge 0.99\) for both u and F, is greater than the general order of the initial approximation error, see Proposition 4.1. We also computed the errors using (5.1), which led to an order of 0.56. Thus, the estimate in (4.51) is not optimal for this example.
Example 5.2
(Cusp initial data) Let \(\alpha =\frac{2}{5}\) and consider the initial data
The wave breaking times accumulate. For each \(t \in [0, 3]\) an infinitesimal amount of energy concentrates. Furthermore, \(u_{0,x}\) is not in \(L^{\infty }(\mathbb {R})\).
In Fig. 6\((u_{\Delta x}, F_{\Delta x})\) with \(\Delta x = 1.76 \cdot 10^{-1}\) and \(\Delta x = 6.01 \cdot 10^{-3}\), and the exact solution, are displayed at \(t=0\), \(t=\frac{3}{2}\), and \(t=3\). Figure 7 shows the errors as we refine the mesh. We also computed the errors using (5.1), which led to an order of 0.24. Thus, in this case the estimate in (4.51) is optimal.
As can be observed in Fig. 7 of Example 5.2, the convergence order is lower than in Example 5.1, suggesting that accumulating wave breaking times deteriorates the convergence rate of the numerical method. However, the upcoming example indicates that accumulation is of less significance than regularity of the initial data.
Example 5.3
(Cosine wave initial data) Let \(\alpha =\frac{3}{5}\) and consider the initial data
Note that in contrast to the previous example, the derivative \(u_{0, x}\) is Lipschitz continuous. Moreover, for each \(t \in \left( \tfrac{2}{\pi }, \infty \right) \), wave breaking occurs at four distinct isolated points both in Eulerian and Lagrangian coordinates, and happens continuously in the time interval \([\frac{2}{\pi }, \infty )\). One can compute the solution in Lagrangian coordinates exactly, as well as the total energy \(F_{\infty }(t)\) for all \(t \ge 0\). The Lagrangian solution is then numerically mapped into Eulerian coordinates and compared with two different numerical approximations at the times \(t=0\), \(\frac{2}{\pi }\), and \(\frac{4}{\pi }\) in Fig. 8.
Figure 9 shows the approximation errors. We observe that the convergence order is higher than in Example 5.2, although the wave breaking times still accumulate.
Data Availability
The datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request.
Notes
We say that \(\eta _{\Delta x} \Longrightarrow \eta \) if \(\int _{\mathbb {R}}\phi d\eta _{\Delta x} \rightarrow \int _{\mathbb {R}}\phi d \eta \) for all \(\phi \in C_b(\mathbb {R}):= C(\mathbb {R}) \cap L^{\infty }(\mathbb {R})\).
References
Beals, R., Sattinger, D.H., Szmigielski, J.: Inverse scattering solutions of the Hunter–Saxton equation. Appl. Anal. 78(3–4), 255–269 (2001)
Billingsley, P.: Convergence of Probability Measures. Wiley, New York-London-Sydney (1968)
Bressan, A., Constantin, A.: Global solutions of the Hunter–Saxton equation. SIAM J. Math. Anal. 37(3), 996–1026 (2005)
Bressan, A., Holden, H., Raynaud, X.: Lipschitz metric for the Hunter–Saxton equation. J. Math. Pures Appl. 94(1), 68–92 (2010)
Christiansen, T.: Traveling wave solutions for the Hunter–Saxton equation. Master’s thesis, Norwegian University of Science and Technology (NTNU), (2021)
Dafermos, C.M.: Generalized characteristics and the Hunter–Saxton equation. J. Hyperb. Differ. Equ. 8(1), 159–168 (2011)
Folland, G.B: Real analysis. Pure and Applied Mathematics (New York). John Wiley & Sons, Inc., New York, second edition edition. Modern techniques and their applications, A Wiley-Interscience Publication (1999)
Grasmair, M., Grunert, K., Holden, H.: On the equivalence of Eulerian and Lagrangian variables for the two-component Camassa-Holm system. In Current research in nonlinear analysis, volume 135 of Springer Optim. Appl., pp. 157–201. Springer, Cham (2018)
Grunert, K., Holden, H.: Uniqueness of conservative solutions for the Hunter–Saxton equation. Res. Math. Sci. 9(2), 19 (2022)
Grunert, K., Holden, H., Raynaud, X.: Global conservative solutions to the Camassa–Holm equation for initial data with nonvanishing asymptotics. Discret. Contin. Dyn. Syst. 32(12), 4209–4227 (2012)
Grunert, K., Holden, H., Raynaud, X.: Global dissipative solutions of the two-component Camassa–Holm system for initial data with nonvanishing asymptotics. Nonlinear Anal. Real World Appl. 17, 203–244 (2014)
Grunert, K., Holden, H., Raynaud, X.: A continuous interpolation between conservative and dissipative solutions for the two-component Camassa–Holm system. Forum Math. Sigma, vol. 3, Paper No. e1, 73, (2015)
Grunert, K., Nordli, A.: Existence and Lipschitz stability for \(\alpha \)-dissipative solutions of the two-component Hunter-Saxton system. J. Hyperb. Differ. Equ. 15(3), 559–597 (2018)
Grunert, K., Nordli, A., Solem, S.: Numerical conservative solutions of the Hunter–Saxton equation. BIT 61(2), 441–471 (2021)
Grunert, K., Tandy, M.: Lipschitz stability for the Hunter–Saxton equation. J. Hyperb. Differ. Equ. 19(2), 275–310 (2022)
Grunert, K., Tandy, M.: A Lipschitz metric for \(\alpha \)-dissipative solutions to the Hunter–Saxton equation, arXiv:2302.07150
Holden, H., Karlsen, K.H., Risebro, N.H.: Convergent difference schemes for the Hunter–Saxton equation. Math. Comp. 76(258), 699–744 (2007)
Holden, H., Raynaud, X.: Global conservative solutions of the Camassa–Holm equation: a Lagrangian point of view. Comm. Part. Differ. Equ. 32(10–12), 1511–1549 (2007)
Hunter, J.K., Saxton, R.: Dynamics of director fields. SIAM J. Appl. Math. 51(6), 1498–1521 (1991)
Hunter, J.K., Zheng, Y.: On a completely integrable nonlinear hyperbolic variational equation. Phys. D 79(2–4), 361–386 (1994)
Hunter, J.K., Zheng, Y.: On a nonlinear hyperbolic variational equation. I. Global existence of weak solutions. Arch. Ration. Mech. Anal 129(4), 305–353 (1995)
Khesin, B., Lenells, J., Misiołek, G.: Generalized Hunter–Saxton equation and the geometry of the group of circle diffeomorphisms. Math. Ann. 342(3), 617–656 (2008)
Lenells, J.: The Hunter–Saxton equation describes the geodesic flow on a sphere. J. Geom. Phys. 57(10), 2049–2064 (2007)
Lenells, J.: The Hunter–Saxton equation: a geometric approach. SIAM J. Math. Anal. 40(1), 266–277 (2008)
LeVeque, R.J.: Finite Volume Methods for Hyperbolic Problems. Cambridge Texts in Applied Mathematics. Cambridge University Press, Cambridge (2002)
Liu, J., Yin, Z.: Global weak solutions for a periodic two-component \(\mu \)-Hunter–Saxton system. Monatsh. Math. 168(3–4), 503–521 (2012)
McDonald, J.N., Weiss, N.A.: A course in real analysis. Academic Press Inc. San Diego CA, second edition edition (2013)
Nordli, A.: A Lipschitz metric for conservative solutions of the two-component Hunter–Saxton system. Methods Appl. Anal. 23(3), 215–232 (2016)
Nordli, A.: On the two-component Hunter–Saxton system. PhD thesis, Norwegian University of Science and Technology (NTNU) (2017)
Sato, S.: Stability and convergence of a conservative finite difference scheme for the modified Hunter–Saxton equation. BIT 59(1), 213–241 (2019)
Wunsch, M.: The generalized Hunter–Saxton system. SIAM J. Math. Anal. 42(3), 1286–1304 (2010)
Yan, X., Shu, C.-W.: Dissipative numerical methods for the Hunter–Saxton equation. J. Comput. Math. 28(5), 606–620 (2010)
Funding
Open access funding provided by NTNU Norwegian University of Science and Technology (incl St. Olavs Hospital - Trondheim University Hospital). The funding was provided by Norges Forskningsråd (Grant numbers: 286822, 250070).
Author information
Authors and Affiliations
Corresponding author
Ethics declarations
Conflict of interest
The authors declare that they have no conflict of interest.
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Research supported by the grants Waves and Nonlinear Phenomena (WaNP) and Wave Phenomena and Stability — a Shocking Combination (WaPheS) from the Research Council of Norway.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Christiansen, T., Grunert, K., Nordli, A. et al. A Convergent Numerical Algorithm for \(\alpha \)-Dissipative Solutions of the Hunter–Saxton Equation. J Sci Comput 99, 14 (2024). https://doi.org/10.1007/s10915-024-02479-4
Received:
Revised:
Accepted:
Published:
DOI: https://doi.org/10.1007/s10915-024-02479-4
Keywords
- Hunter–Saxton equation
- Projection operator
- Conservative solutions
- Numerical method
- Convergence
- \(\alpha \)-Dissipative solutions