Discontinuous eikonal equations
in metric measure spaces

In this paper we study the eikonal equation in a metric measure space $({\mathbf{X}},d,\mu)$ with a possibly discontinuous inhomogenous term, where $\mu$ is assumed to be a nonnegative, locally finite Borel measure. We assume that ${\mathbf{X}}$ is a complete length space and $\Omega\subsetneq{\mathbf{X}}$ is a bounded domain. We consider

with the Dirichlet boundary condition

where $f$ is a given positive Borel measurable function in $\overline{\Omega}$ and $g$ is a given bounded function on $\partial\Omega$. Further assumptions on $f$ and $g$ will be introduced later.

Well-posedness of the eikonal equation and more general Hamilton-Jacobi equations is a classical topic. The theory of viscosity solutions provides a successful tool to solve such fully nonlinear PDEs in the Euclidean space. It is well known that when $\Omega\subset{\mathbb{R}}^{n}$ and $f\in C(\overline{\Omega})$ satisfies the lower bound condition

there exists a unique viscosity solution $u\in C(\overline{\Omega})$ of (1.1) with Dirichlet data (1.2) under appropriate regularity condition on $g$. The uniqueness result is obtained by establishing a comparison principle while the existence of solutions can be proved via several different approaches including Perron’s method and a control-based formula.

Let us give more details on the connection with the optimal control theory, as it largely motivates our exploration and plays a central role in this work. It is widely known [40, 6] that, under appropriate conditions on $f$ and $g$, the unique viscosity solution $u$ of (1.1) with the Dirichlet boundary condition (1.2) in the Euclidean space can be expressed by

This formula represents the so-called value function of an optimal control problem, where one tries to move a point from $x\in\overline{\Omega}$ toward the boundary and minimize the total payoff comprising a running cost $f$ integrated along the trajectory and a terminal cost $g$ at the first hit on the boundary. See [38, 31] and references therein for more recent developments on representation formulas for general Hamilton-Jacobi equations.

The representation formula not only provides a clear understanding of the solution and promotes various control-related practical applications, but also suggests possible generalization of the theory for a possibly discontinuous function $f$. In fact, much progress has been made in this direction in the Euclidean space. The study on discontinuous Hamilton-Jacobi equations in the Euclidean space was initiated in the work [36] and later found applications in geometric optics with different layers, image processing, shape from shading, etc. Further development in different contexts to treat discontinuous Hamiltonians including the time-dependent case can be found in [55, 47, 48, 53, 12, 11, 19, 8, 13, 54, 17, 27, 28]. The aforementioned papers in the Euclidean case, reduced to our eikonal problem, seem to need at least local essential boundedness of $f$ for their well-posedness results except for the papers [55, 53, 54]. In these three papers, the results regarding uniqueness and existence results rely on regularity conditions on the discontinuity set of $f$ or certain controllability assumptions for the optimal control formulation. Such conditions can hardly be extended further to our relaxed geometric setting due to the lack of smooth structure in general metric spaces.

In the present paper we aim to study in a direct manner the eikonal equation with potentially wilder discontinuity than those considered in [55, 53, 54]. We drop the local boundedness of the inhomogeneous data $f$ and allow it to be as discontinuous as a function in $L^{p}(\Omega)$. It is worth adding that an important tool called $L^{p}$ viscosity solution theory has been developed for second order uniformly elliptic equations with measurable data in [10, 18], but it does not seem to apply to the first order case. For the Dirichlet problem (1.1)(1.2), our primary observation is that the formula (1.4) actually requires nothing more than the path integrability of $f$ along a curve connecting $x$ to the boundary, which is a much weaker assumption than the continuity or even boundedness of $f$. Let us present a simple one-dimensional example to clarify this point.

Let $\Omega=(-1,1)\subset{\mathbb{R}}$ and set $f:\Omega\to{\mathbb{R}}$ such that

The value $f(0)$ can be chosen to be any positive real number in order to meet the requirement (1.3). Such type of singular eikonal equations finds applications in lens design for wireless communication systems [16]. It is clear that $f\in L^{p}(\Omega)$ with $p\in[1,2)$. Therefore the formula in (1.4) still makes sense for given boundary data $g$ at $\pm 1$. For instance, setting $g(\pm 1)=0$, we see that the formula immediately yields

It seems to be the correct solution of (1.1)(1.2), for it satisfies the standard definition of viscosity solutions if we consider $f(0)=\infty$. Also, one can verify the validity of approximation via truncation of the running cost. Indeed, taking $f_{M}=\min\{f,M\}$ for $M>0$ large, we can obtain a unique viscosity solution for the truncated problem with the bounded function $f_{M}$. By direct calculations, we easily see that the approximate solution converges uniformly to $u$ in $[-1,1]$. Moreover, although in general we cannot expect Lipschitz regularity of solutions due to the loss of boundness of $f$, the example shows that the solution enjoys Hölder regularity that depends on the integrability of $f$. This can be viewed as a natural consequence corresponding to the Morrey-Sobolev inequality, as the eikonal equation suggests $u\in W^{1,p}(\Omega)$ at least formally.

To the best of our knowledge, the well-posedness and regularity issue for this type of discontinuous eikonal equations have not been carefully examined even in the Euclidean case. We are therefore motivated to tackle these problems in a general way that permits applications to a broad class of metric measure spaces with length structure. We develop the notion of Monge solutions used in [47, 8, 43] and show uniqueness and existence of solutions in this general setting. We also establish Hölder regularity of the solution under certain regularity assumptions on $f$ and ${\mathbf{X}}$. More details about our results will be given momentarily.

First-order Hamilton-Jacobi equations in metric spaces have recently attracted great attention for applications in optimal transport [5, 57], mean field games [14], topological networks [49, 35, 1, 33, 34] etc. Concerning the equations that depend on the gradient in terms of its norm, we refer to [4, 29, 26, 45] for several different notions of viscosity solutions as well as the associated uniqueness and existence results. In our previous paper [43], we proved the equivalence of these solutions and introduce another alternative notion, extending the approach of Monge solutions proposed in [47] to the eikonal equation in complete length spaces; see also an application of this Monge-type notion to the study of the first eigenvalue problem for the infinity Laplacian in metric spaces [41]. Our current work further develops the notion of Monge solutions so that viscosity solution theory will be available to handle a class of nonlinear equations in general metric measure spaces with data having a wider range of discontinuities.

Our work is new even from the point of view of pure analysis. There has been great progress on various aspects of first order analysis on metric measure spaces, including first order Sobolev spaces and their relation to variational problems and partial differential equations. One central notion that generalizes the norm of the gradient of a Sobolev function in Euclidean spaces is the so-called upper gradient. We refer the reader to [30] for a survey of the Sobolev spaces theory built upon this notion. From this point of view, the Dirichlet problem (1.1) with boundary data (1.2) can be interpreted as an extension problem for a function $u$ to $\overline{\Omega}$ with boundary value $g$ assigned on $\partial\Omega$ and the upper gradient $f$ prescribed in $\Omega$. In general, many possible choices can be expected, among which the viscosity solution constructed by the formula (1.4) is usually considered to be the most “physical”. This perspective can also be found in the work [15]. There is certainly no need to restrict the prescribed gradient to a continuous or bounded function class. We are thus inspired to address the extension problem for $f\in L^{p}(\Omega)$ and investigate properties of the extended function $u$.

Our presentation can be divided into two parts. The first part, consisting of Sections 2–4, is a general PDE theory with only the length structure and curve-wise integrals of $f$ involved. In our main well-posedness results, we impose certain key conditions on the closure of the domain $\overline{\Omega}$ and the inhomogeneous term $f$. In the second part, adding a measure structure to the space, we provide more specific sufficient conditions for those assumptions on $\overline{\Omega}$ and $f$.

One of the crucial steps in the first part is to find an appropriate notion of viscosity solutions in this setting that agrees with the representation formula (1.4) and justifies a comparison principle for us to prove the uniqueness. As mentioned above, we adopt the notion of Monge solutions, which in the case of a continuous $f$, requires

where, for a locally Lipschitz function $u$, the subslope $|\nabla^{-}u|$ is defined by

In the Euclidean space, this definition is known to be equivalent to the usual viscosity solutions [47]. One of its advantages is that it avoids using smooth test functions, whose lack of availability is indeed a key difficulty in the study of Hamilton-Jacobi equations in general metric spaces.

When $f$ is of class $L^{p}(\Omega)$, the definition in (1.7) may not apply directly. For example, if one changes the value of $f(0)$ to be finite in the aforementioned example, then the expected solution (1.6) does not satisfy definition (1.7) at $0$. We thus need to somehow exploit collective information of $f$ rather than its pointwise value. Also, we always assume that the lower bound condition (1.3) holds for $f$ in order to obtain uniqueness. Heuristically speaking, since $f$ is uniformly positive in $\Omega$, we can rewrite the equation (1.1) as

and adopt a new pointwise slope by incorporating the value of $f(x)$ into the metric. This approach enables us to study the eikonal equation as if the right hand side is a constant. Such connection can be rigorously realized by introducing the following new metric, which is also called optical length function in [47, 8] etc.,

where $\Gamma_{f}(x,y)$ denotes the collection of all curves $\gamma$ connecting $x$ and $y$ in $\overline{\Omega}$ on which the path integral is finite, namely,

We say that the curves $\gamma$ in $\Gamma_{f}(x,y)$ are admissible with respect to $f$. We shall assume throughout Sections 2–4 that

so that the formula in (1.4) is well defined for any $x\in\overline{\Omega}$. In Section 5, we discuss some sufficient assumptions on $\overline{\Omega}$ implying (1.9).

Adopting the subslope with respect to this newly defined metric $L_{f}$, defined by

one can simply define the Monge solution of (1.1) in Definition 3.1 by asking that $u$ satisfies

The same idea of changing metrics was also outlined in [9] in connection with metric geometry and in [56] for applications in homogenization.

This new definition of Monge solutions is consistent with the well-studied case where $f$ is continuous. In fact, if $f$ is continuous at $x_{0}\in\Omega$, then

It is then easily seen that $|\nabla_{f}^{-}u|(x_{0})=1$ and $|\nabla^{-}u|(x_{0})=f(x_{0})$ are equivalent.

Switching the metric from $d$ to $L_{f}$ enables us to reduce the possibly discontinuous inhomogeneous data in (1.7) to the continuous standard form (1.11). As a consequence, most of our arguments in [43] can be applied to establish the comparison principle and prove existence of solutions of (1.1), (1.2) under the compatibility condition

However, a major challenge concerning the topological change of the space appears in the proof of comparison principle. Note that the metrics $L_{f}$ and $d$ are not topologically equivalent in general, especially when $f$ is unbounded. One can find examples where $d(x,y)\to 0$ fails to imply $L_{f}(x,y)\to 0$; see Example 2.4. For most of our analysis, we impose the following key compatibility condition

which ensures that the topology induced by $L_{f}$ is in agreement with the metric topology inherited from $({\mathbf{X}},d)$, and enables us to obtain not only the uniqueness but also the uniform continuity of Monge solutions.

In general, if (1.14) fails to hold, then, because of the change in the metric, the domain $\Omega$ that is bounded in $d$ may become unbounded in $L_{f}$. The notions of interior and boundary points of $\Omega$ may also change accordingly. However, it is still possible to show uniqueness and existence of Monge solutions if we have boundedness of $\overline{\Omega}$ with respect to $L_{f}$ as well as a weaker version of (1.14) as below:

where $\Omega_{r}=\{x\in\Omega:d(x,\partial\Omega)>r\}$. Here, we denote $d(x,\partial\Omega)=\inf_{y\in\partial\Omega}d(x,y)$. In this case, we can get a unique Monge solution that is Lipschitz continuous in $L_{f}$ but possibly discontinuous in $d$; see Example 4.4 for a concrete example. Note that the completeness of $\overline{\Omega}$ is preserved under our metric change (Lemma 2.5), which enables us to establish a comparison principle based on Ekeland’s variational principle.

Our main result can be summarized as follows. Its proof is given in Section 4.

The interpretation (1.17) is automatically guaranteed by (1.2) if $(\overline{\Omega},L_{f})$ is assumed to be compact. Our comparison results with and without compactness of $(\overline{\Omega},L_{f})$ are presented in Section 3.2. We also include a discussion in Section 4.2 about the case when the compatibility condition (1.13) on $g$ fails to hold. In the Euclidean space, this corresponds to loss of Dirichlet boundary data, and one needs to relax the boundary condition in the viscosity sense; see for example [51, 52, 37] for the well-posedness results under certain regularity conditions on $\partial\Omega$. Thanks to the general setting considered in this present paper, we treat this problem in a more direct way. If the boundary data $g$ on a part of $\partial\Omega$ is lost, we take the remaining piece, denoted by $\Sigma_{g}$, as the real boundary and regard $\overline{\Omega}\setminus\Sigma_{g}$ as the interior of the domain. In fact, the formula (1.4) does yield the Monge solution property of $u$ in $\overline{\Omega}\setminus\Sigma_{g}$. Such flexibility in changing the notion of boundary and interior points leads us to uniqueness and existence of solutions to the reduced Dirichlet problem.

The second half of our presentation is devoted to understanding more deeply the key assumptions (1.9) and (1.14) in Theorem 1.1. We study this regularity problem in the setting of a general metric measure space $({\mathbf{X}},d,\mu)$ with the measure $\mu$ assumed to be doubling. We recall that a locally finite Borel measure $\mu$ is doubling if there exists $C_{d}\geq 1$ such that $\mu(B_{2r}(x))\leq C_{d}\mu(B_{r}(x))$ for all $r>0$, where $B_{r}(x)$ denotes the open metric ball centered at $x\in{\mathbf{X}}$ with radius $r>0$, i.e.,

We also call $Q=\log_{2}C_{d}$ the homogeneous dimension of the doubling metric measure space. It is not difficult to see that by the doubling property of $\mu$, we have, for any metric balls $B_{r}(x)\subset B_{R}(x)\subset{\mathbf{X}}$,

We prove the property (1.14) and Hölder regularity with respect to the metric $d$ of the solution $u$ defined by (1.16) under either of the following two assumptions.

• (A1)

  $\overline{\Omega}$ satisfies a $p$-Poincaré inequality for some finite $p>\max\{1,Q\}$ and $f\in L^{p}(\overline{\Omega})$, where $Q>0$ is the homogeneous dimension from (1.18) above.

• (A2)

  $\overline{\Omega}$ satisfies an $\infty$-Poincaré inequality and $f\in L^{\infty}(\overline{\Omega})$.

Here, the $p$-Poincaré inequality is a condition that indicates the richness of curves of a space; see Section 5 for a more precise description and [30] for more detailed introduction. In our assumptions, a balance is taken between the regularity of the space $\overline{\Omega}$ and the function $f$. In order to consider merely $p$-integrable $f$ with a finite $p$, we assume $p$-Poincaré inequality in (A1), which is stronger than the $\infty$-Poincaré inequality in (A2),

Our results in Theorem 1.2 resemble the Morrey-Sobolev embedding theorem. In the proof, we apply a geometric characterization of the Poincaré inequality provided by [21] for both cases of (A1) and (A2) with adaptations for our eikonal equation. We can also show that $u$ is in the Sobolev class $N^{1,p}(\overline{\Omega})$ under the conditions of Theorem 1.2. While (A1) and (A2) seem close to optimal for us to obtain (1.14), both of them are actually too strong to directly show (1.9), which is far weaker than (1.14) and simply requires the existence of one curve on which $f$ is integrable. We construct an example in ${\mathbb{R}}^{2}$, showing that (1.9) can be obtained even for some $f\notin L^{2}(\Omega)$; see Example 5.4. It would be interesting to find more general sufficient conditions for (1.9).

It is worth remarking that the $L^{p}$ class in our study is only used to describe the extent of discontinuity of $f$ and cannot be understood as the usual $p$-integrable function space. Since the Monge solution obtained by (1.4) substantially depends on path integrals of $f$, changing, especially decreasing, the value of $f$ even on a null set will drastically affect the solution. In other words, our Monge solution strongly depends on the choice of representatives of the equivalence class for $f\in L^{p}(\overline{\Omega})$.

Section 6 is devoted to discussions on how to alleviate the instability of solutions with respect to $f$ increasing on null sets. Our method generalizes the approach in [12, 8]. Instead of taking the solution for $L^{p}$ function $f$, we generate a class of solutions for different representatives from its equivalence class under the assumption (A1) or (A2). More precisely, for any null set $N\subset\overline{\Omega}$, we can find a Monge solution of (1.1)(1.2) with $f$ replaced by $f_{N}=f+\infty\chi_{N}$, where $\chi_{N}$ denotes the characteristic function of $N$. This change in $f$ amounts to restricting $f$ to be integrable only on curves transversal to $N$. Here, a curve is said to be transversal to $N$ if $\mathcal{H}^{1}(\gamma^{-1}(N))=0$, where $\mathcal{H}^{1}$ stands for the one-dimensional Hausdorff measure. By substituting $\Gamma_{f}(x,y)$ for each pair of points $x,y\in\overline{\Omega}$ with the class of transversal curves

and taking the associated optical length function

we can obtain the same existence and uniqueness results for any given null set $N$.

We are particularly interested in the maximal solution over all null sets $N$, as it corresponds to the choice of the largest function in the equivalence class of $f$ and represents the strongest possible instability. It can be defined by

where $\tilde{L}_{f}$ is the maximal optical length function over all $N$, that is,

We proved in Proposition 6.9 and Remark 6.11 that $\tilde{u}$ is the maximal weak solution of (1.1), (1.2) under (A2) together with appropriate assumptions on $\overline{\Omega}$ and boundary data $g$ as well as almost everywhere continuity of $f$. Here, the notion of weak solutions is a generalization of the Euclidean counterpart, requiring $\tilde{u}$ to satisfy $|\nabla\tilde{u}|=f$ almost everywhere in $\Omega$, where $|\nabla\tilde{u}|$ denotes the slope of $u$, i.e., for $x\in\Omega$,

This result extends the discussion about the well-known maximal weak solution characterization of viscosity solutions in the Euclidean space [47, 8] to the setting of metric measure spaces and possibly unbounded discontinuous inhomogeneous term $f$. It is however not clear to us how to handle the case when $f$ is not continuous almost everywhere. Another important open question is under what assumptions can we can eliminate the instability, i.e., $u=\tilde{u}$ in $\overline{\Omega}$. Some discussions regarding this issue can be found in [53] for the Euclidean space but the case for general metric spaces seems more difficult.

Let us conclude the introduction by mentioning that in this work we choose not to pursue more general Hamilton-Jacobi equations. One can easily generalize our method if a control-based formula is available and a new metric incorporating the discontinuity of the Hamiltonian can be found. Even for the eikonal equation itself, our general setup actually implicitly allows more general dependence of the Hamiltonian on the space variable. Typical examples of ${\mathbf{X}}$ include the sub-Riemannian manifolds. For instance, when $({\mathbf{X}},d)$ is taken to be the first Heisenberg group with the Carnot-Carathéodory metric, the eikonal equation (1.1) can be written in the Euclidean coordinates as

We refer to [7, 2, 3] etc. for discussions on the eikonal equation with $f\equiv 1$ in the sub-Riemannian setting.

The authors thank Professors Yoshikazu Giga, Nao Hamamuki and Atsushi Nakayasu for valuable comments on the first draft of the paper. The work of the first author was supported by JSPS Grant-in-Aid for Scientific Research (No. 19K03574, No. 22K03396). The work of the second author is partially supported by the grant DMS-#2054960 of the National Science Foundation (U.S.A.). The work of the third author was supported by JSPS Grant-in-Aid for Research Activity Start-up (No. 20K22315) and JSPS Grant-in-Aid for Early-Career Scientists (No. 22K13947).

Let us now study the Dirichlet problem (1.1), (1.2). We begin with the definition of Monge solutions to (1.1).

We stress that the topology in the local boundedness of $u$ and limsup in the definition above is taken with respect to the metric $L_{f}$. Since $L_{f}(x,x_{0})\to 0$ implies $d(x,x_{0})\to 0$ by (2.1), we see that

implies

In general, we cannot expect that the reverse inequality of (3.3) holds without further assumptions like (1.14).

We also emphasize that our definition does not require any continuity or semicontinuity of $u$. This relaxation constitutes a difference from the standard definition of discontinuous viscosity solutions. Recall that when defining a (possibly discontinuous) subsolution $u$, we usually assume $u\in USC(\Omega)$ (and $u\in LSC(\Omega)$ for the symmetric supersolution definition). Here, $USC(\Omega)$ and $LSC(\Omega)$ denote the classes of upper and lower semicontinuous functions in $\Omega$ respectively with respect to $d$. We do not adopt such restrictions here.

Using the subslope defined in (1.10), we can see in a straightforward manner that our Monge solutions (resp., subsolutions, supersolutions) of (1.1) can simply be understood as Monge solutions (resp., subsolutions, supersolutions) of

On the other hand, Definition 3.1, together with (1.14), does not guarantee the upper semicontinuity of a Monge solution. However, for Monge solutions that arise from the associated Dirichlet problem, we will later use its optimal control interpretation to show the continuity of solutions.

Thanks to the metric change, one can essentially apply the results in [43] to (3.4) in the complete metric space $(\overline{\Omega},L_{f})$ to show both uniqueness and existence of Monge solutions of (1.1)(1.2) under appropriate assumptions on the boundary data $g$. The compatibility condition (1.13) we will impose on $g$, combined with (1.14), basically ensures that the topology stays equivalent when converting the metric from $d$ to $L_{f}$.

Let us prove a comparison principle, where we assume the semicontinuity of Monge sub- and supersolutions with respect to the metric $L_{f}$. We say that $u:\overline{\Omega}\to{\mathbb{R}}$ is upper (resp., lower) semicontinuous with respect to $L_{f}$, denoted by $u\in USC_{L}(\overline{\Omega})$ (resp., $u\in LSC_{L}(\overline{\Omega})$), if for every fixed $x\in\overline{\Omega}$, we have

It is not difficult to see that $u\in USC_{L}(\overline{\Omega})$ (resp., $u\in LSC_{L}(\overline{\Omega})$) if $u\in USC(\overline{\Omega})$ (resp., $u\in LSC(\overline{\Omega})$) with respect to $d$, since $L_{f}(x,y)\to 0$ implies that $d(x,y)\to 0$.