Abstract
We prove the existence of a unique viscosity solution to certain systems of fully nonlinear parabolic partial differential equations with interconnected obstacles in the setting of Neumann boundary conditions. The method of proof builds on the classical viscosity solution technique adapted to the setting of interconnected obstacles and construction of explicit viscosity sub- and supersolutions as bounds for Perron’s method. Our motivation stems from so called optimal switching problems on bounded domains.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
This note deals with existence and uniqueness of viscosity solutions of systems of non-linear partial differential equations (PDE) with interconnected obstacles and Neumann boundary conditions. Let \(\Omega \subset {\mathbb {R}}^{n}\) be a domain, i.e. an open connected set, and let \(i\in \{1,2,\ldots ,m\}\) for some positive integer m. The problem considered can in general be stated as
The solution is a vector-valued function \(u:=(u_1,u_2,\ldots ,u_m)\) where the components are interconnected through the obstacle \({{\mathcal {M}}}_i\) in the sense that component i lies above the obstacle
which itself depends on all other components. Here and in the following, \(Du_i \in {{\mathbb {R}}}^n\) and \(D^2u_i \in \mathcal {S}^n\) (\(\mathcal {S}^n\) being the set of symmetric \(n \times n\) matrices) denote the gradient and Hessian matrix of \(u_i\) w.r.t. x.
In the standard setting of a single PDE without obstacle, the well-renowned paper Crandall–Ishii–Lions [7] proves existence and uniqueness of viscosity solutions of fully nonlinear degenerate elliptic PDEs with Neumann boundary conditions in smooth domains satisfying an exterior ball condition. Dupuis–Ishii [9, 10] consider more general domains such as non-smooth boundaries and domains with corners, and Barles [3] proved a comparison principle and existence of unique solutions to degenerate elliptic and parabolic boundary value problems with non-linear Neumann type boundary conditions in bounded domains with \(W^{3,\infty }\)-boundary. Ishii and Sato [16] proved similar theorems for boundary value problems for some singular degenerate parabolic partial differential equations with non-linear oblique derivative boundary conditions in bounded \(C^1\)-domains. Further, in bounded domains with \(W^{3,\infty }\)-boundary, Bourgoing [4] considered singular degenerate parabolic equations and equations having \(L^1\)-dependence in time. Lundström–Önskog [23] recently extended parts of [9] by establishing existence and uniqueness of viscosity solutions to a general parabolic PDE in non-smooth, time-dependent domains.
For systems of variational inequalities with interconnected obstacles as in (BVP), the literature on bounded spatial domains \(\Omega \) with boundary conditions is, to the authors’ knowledge, scarce. Instead much focus has been directed to unbounded domains (essentially \(\Omega \equiv {\mathbb {R}}^n\)) and often linear PDEs. In this setting the literature is, on the contrary, rather rich, see, e.g., El-Asri–Hamadene [11], Djehiche–Hamadene–Popier [8], Hu–Tang [14], Biswas–Jakobsen–Karlsen [4], Hamadene-Morlais [13], Lundström–Nyström–Olofsson [19, 20], Lundström–Olofsson–Önskog [21], Fuhrman-Morlais [12], Reisinger–Zhang [24] and the many references listed therein. Extensions to more general operators have been studied in Klimsiak [18]. Recently, Boufoussi-Hamadene-Jakani [6] considered a similar setting to ours and proved existence and uniqueness of a viscosity solution to a system of PDEs with interconnected obstacles and Neumann boundary conditions. In contrast to the present paper, their method is based on a probabilistic formulation in terms of reflected backward stochastic differential equations.
A reason for this large amount of interest is the close connection to stochastic optimization and so called “optimal switching” problems. Under assumptions, the solution to (BVP) then represents the value function of an optimal switching problem in which component i of the underlying stochastic process has the second order operator \(F_i\) as infinitesimal generator. With this interpretation, the domain \(\Omega \) represents the constraints put on the underlying stochastic process and, in particular, \(\Omega \equiv {\mathbb {R}}^n\) means no constraints.
However, when studying applications of optimal switching theory, one can easily think of examples where constraints are not only natural, but also necessary for the model to be meaningful. For example, consider a hydropower plant with a dam in which the underlying (stochastic) process \(X_t\) represents the amount of water currently available in the dam. Then, naturally, the process can only take values between 0 and some \(X_{max}\), the latter being the capacity of the dam. (In case the power plant is of run-of-river type, i.e., has no dam, the problem can be reduced to \(\Omega ={\mathbb {R}}^n\) and an application of the theory in this setting was recently studied in Lundström–Olofsson–Önskog [22].) When the dam is full, additional increase of water must be spilled, keeping the amount of water in the dam constant. Such situation may be modelled through a reflection of the underlying stochastic process \(X_t\). If this reflection is assumed in the normal direction, then one obtains a Neumann boundary condition, i.e.
where n(x) is the normal of \(\partial \Omega \) at x. A more general reflection model puts us in the oblique derivative problem in which n(x) in (BC) should be replaced by \(\nu (x)\), a vector field satisfying \(\langle \nu (x), n(x)\rangle > 0\) on \(\partial \Omega \). The Dirichlet setting, which is studied in Barkhudaryan–Gomes–Shahgholian–Salehi [2], can be given a similar interpretation as above but then in the sense that the game ends when the process hits the boundary \(\partial \Omega \).
Optimal switching problems with reflection is not as well-studied as the non-reflected counterpart. However, given the close connection to real-world applications and the growing interest of optimal switching problems under constraints, see, e.g., Kharroubi [17], we expect this to change and a thorough analysis of the related PDE-theory on bounded domains is motivated. Moreover, although the study is motivated by applications, the results are of independent mathematical interest, enlarging the class of problem for which the viscosity theory is fully investigated.
2 Mathematical Problem Formulation, Assumptions and Main Results
2.1 Notation and Assumptions
We let \(\Omega \subset {{\mathbb {R}}}^n\) be an open connected bounded set with closure \({\bar{\Omega }}\) and boundary \(\partial \Omega := {\bar{\Omega }} \setminus \Omega \) and consider \(t \in [0,T]\). We denote by \(\mathcal C^{\alpha , \beta }\) the class of functions which are \(\alpha \) and \(\beta \) times continuously differentiable on \([0,T] \times {\bar{\Omega }}\) w.r.t. the first (time t) and second (space x) variable, respectively. If \(f(\cdot )\) only depends on the spatial variable we simply write \(f\in {\mathcal {C}}^\beta \). We assume that
i.e., that \(\partial \Omega \) is once continuously differentiable and that
where \(B(x,r):= \{y: |y-x|<r\}\). (A simple example of such a domain is \(\partial \Omega \in {\mathcal {C}}^{1}\) with Lipschitz continuous derivative, as such domains satisfy both the exterior and interior ball condition, see Aikawa–Kilpeläinen–Shanmugalingam–Zhong [1, Lemma 2.2].)
We assume that \(F_i(t,x,r,p,X) : [0,T] \times {\bar{\Omega }} \times {{\mathbb {R}}}\times {{\mathbb {R}}}^n \times \mathcal {S}^n \rightarrow {{\mathbb {R}}}\) is a continuous function satisfying
and that
for every \(t \in [0,T)\) fixed and whenever
where \(I\equiv I^n\) denotes the \(n \times n\) identity matrix. Note that these assumptions on \(F_i\) imply degenerate ellipticity.
Regarding the boundary condition \(B_i\), we assume that
Last, we assume that the obstacle functions \(c_{ij}(t,x)\) are continuous on \([0,T]\times {\bar{\Omega }}\) and satisfy
and that the so called “no-loop”-condition holds;
We need to strengthen the assumption on the obstacle slightly to prove existence. In particular, we then also demand that
which is an additional structural condition compared to (O2).
Last, we require that the initial data is continuous and compatible with the obstacle functions, in particular that
Remark 2.1
In the parabolic setting studied here, it is a standard task to show that one can assume \(\lambda >0\) in (F1) w.l.o.g. Indeed, if \(\lambda \le 0\), then for \({\bar{\lambda }} < \lambda \), u(t, x) is a solution to (BVP) if and only if \(e^{{\bar{\lambda }} t}u(t,x)\) is a solution to (BVP) with \(F_i\) and \(f_i\) replaced by
and with \(c_{ij}\) similarly scaled. The function to the left in the above display is strictly increasing and hence satisfies (F1) with \(\lambda \) positive. A consequence of this is that we may as well assume that \(F_i\) is strictly increasing in r and that there exists \(\gamma >0\) s.t. for \(r>s\)
We are looking for solutions in the viscosity sense and will use the classical definitions of (parabolic) sub- and superjets, provided here for convenience.
Definition 2.2
(Definition (8.1) of [7]) For \(({\hat{t}}, {\hat{x}}) \in (0,T) \times {\mathcal {O}} \), the triplet \((a,p,X) \in {{\mathbb {R}}}\times {{\mathbb {R}}}^n \times \mathcal {S}^n\) lies in the parabolic superjet of u at \(({\hat{t}}, {\hat{x}})\), written \((a,p,X) \in {\mathcal {P}}^{2,+}_{{\mathcal {O}}} u({\hat{t}}, {\hat{x}})\), if
as \((0,T) \times {\mathcal {O}} \ni (t,x) \rightarrow ({\hat{t}}, {\hat{x}})\). Analogously, the parabolic subjet is defined by \(\mathcal P^{2,-}_{{\mathcal {O}}} u({\hat{t}}, {\hat{x}}) := -\mathcal P^{2,+}_{{\mathcal {O}}} -u({\hat{t}}, {\hat{x}})\).
The closure of \({\mathcal {P}}^{2,\cdot }_{{\mathcal {O}}} u ({\hat{t}}, {\hat{x}})\), denoted \(\bar{{\mathcal {P}}}^{2,\cdot }_{{\mathcal {O}}} u({\hat{t}}, {\hat{x}})\), is defined as
Definition 2.3
A function \(u=(u_1, \dots , u_m) \in USC([0,T]\times {\bar{\Omega }})\) is a viscosity subsolution to (BVP) if, for all \(i \in \{1,\dots ,m\}\),
Viscosity supersolutions are defined analogously with USC replaced by LSC, \({{\bar{\mathcal {P}}}^{2,+}_{\cdot }}\) replaced by \({{\bar{\mathcal {P}}}^{2,-}_{\cdot }}\), \(\wedge \) replaced by \(\vee \), and \(\le \) replaced by \(\ge \). A function \(u \in C([0,T]\times {\bar{\Omega }})\) is a viscosity solution to (BVP) if it is both a sub- and a supersolution.
2.2 Main Results
With the above preliminaries set, we are ready to state our main results, which concern comparison of sub- and supersolutions and existence and uniqueness of viscosity solutions to (BVP).
Theorem 2.4
Assume that (D1), (F1)–(F3), (BC1), and (O1)–(O2) hold. If u and v are viscosity sub- and supersolutions to (BVP), respectively, then \(u(t,x) \le v(t,x)\) for \((t,x) \in [0,T) \times {\bar{\Omega }}\).
Theorem 2.5
Assume that (D1), (F1)–(F3), (BC1), (O1)–(O3), and (I1) hold. Then there exists a unique viscosity solution to (BVP).
Our proofs rely on by now classical techniques in the theory of viscosity solutions, using doubling of variables and the maximum principle for semi-continuous functions for Theorem 2.4 and Perron’s method together with a construction of certain sub- and supersolutions for Theorem 2.5. From the proof of Theorem 2.4 we also get the following corollary, useful in its own merit.
Corollary 2.6
Let u and v be viscosity sub- and supersolutions, respectively, to (BVP). If u and v satisfy Definition 2.3 (ii) on some open region \(G \subset (0,T) \times \partial \Omega \) and \(u \le v\) on \(\left( (0,T) \times \partial \Omega \right) \setminus G\), then \(u(t,x) \le v(t,x)\) for all \((t,x) \in [0,T) \times {\bar{\Omega }}\).
Remark 2.7
We have chosen to present our results and proofs in a rather simplistic setting, with directional derivative in the normal direction and a smooth, stationary domain \(\Omega \). However, our proofs reveal that the additional arguments needed to treat systems of PDEs with interconnected obstacles (rather than a single PDE) are more or less decoupled from the standard arguments of Crandall–Ishii–Lions [7]. Hence, by combining these additional arguments with existence and uniqueness proofs for more general PDEs and domains, one should be able to prove existence and uniqueness for systems of PDEs with interconnected obstacles in similar generality. For tractability, we refrain from such generalizations here.
3 Proof of Theorem 2.4: The Comparison Principle
Throughout this section, we assume that the assumptions of Theorem 2.4 hold, i.e., that (D1), (F1)–(F3), (BC1), and (O1)–(O2) hold. Let \(u=(u_1, u_2, \dots , u_m)\) and \(v=(v_1, v_2, \dots , v_m)\) be sub- and supersolutions to (BVP), respectively. Our proof follows the classical outline of Crandall–Ishii–Lions [7] and consists of four major steps.
3.1 Step 1. An \(\epsilon \)-room on the Boundary
Lemma 3.1
(Lemma 7.6 of [7]) For any continuous \(\nu (x): \partial \Omega \rightarrow {{\mathbb {R}}}^n\) satisfying \(\langle n(x), \nu (x) \rangle > 0\) there exists a \(C^2({\bar{\Omega }})\) function \(\varphi (x)\) s.t.
Let
where \(\varphi (x)\) is as given by the lemma above (with \(\nu (x) = n(x)\)) and \(C_\eta >0\) is a constant to be specified later. Note that \(u_i^\eta < u_i\) so that for any \((a,p,X) \in {\bar{\mathcal {P}}}^{2,+}_{\Omega }u_i^\eta (t,x)\) we have from (F1) and Remark 2.1 (i.e., from (F1*)) that
for some \(\gamma >0\) and thus
where \(M= \sup _{{\bar{\Omega }}}\{D\varphi (x) +||D^2\varphi (x)||\}\) and where the last inequality comes from (F2). Since \((a,p,X) \in {\bar{\mathcal {P}}}^{2,+}_{\Omega }u^\eta (t,x)\) and \(u^\eta = u - \eta \varphi (x)- {{\mathcal {C}}}_\eta \) we have
and therefore choosing \(C_\eta = \frac{\omega (\eta M)}{\gamma }\) we get that
whenever \((t,x) \in (0,T) \times \Omega \) and \( u_i(t,x) > {{\mathcal {M}}}_i u(t,x) \) (since u is a subsolution). However, \(- \eta \varphi (x) - C_\eta \) is independent of i and thus
and we can conclude that
whenever \((a,p,X)\in {\bar{\mathcal {P}}}^{2,+}_{\Omega }u^\eta _i(t,x)\), \((t,x) \in (0,T) \times \Omega \). On the other hand, if \(x \in \partial \Omega \), we have by (BC1)
and thus \(u^\eta \) is a subsolution to (BVP) with the boundary condition
A similar calculation shows that \(v^\eta \) is a supersolution to (BVP) with boundary condition \({\hat{B}}_i(t,x,r,u): = B_i(t,x,r,p) - \eta \ge 0\). Consequently, it suffices to prove the comparison \(u^\eta \le v^\eta \) when \(u^\eta \) and \(v^\eta \) are sub- and supersolutions to (BVP) with boundary conditions \({\check{B}}_i(t,x,r,p)\) and \({\hat{B}}_i(t,x,r,p)\), respectively, since we retrieve our result in the limit as \(\eta \rightarrow 0\).
3.2 Step 2. Avoiding the Neumann Boundary Condition
We now construct a test function which allows us to discard the Neumann boundary condition in (BVP). We will need the following maximum principle whose proof, which is similar to the current, is postponed to the end of the section.
Proposition 3.2
Let u and v be viscosity sub- and supersolutions, respectively, to
on \([0,T) \times {\bar{\Omega }}\) in the sense of Definition 2.3 (i) and (iii). Then,
Let us now assume the opposite of what we seek to prove, i.e., that there exists a non-empty set \({{\mathcal {I}}}\subset \{1,\dots , m\}\) and \(({\hat{t}}, {\hat{x}}) \in [0,T] \times {\bar{\Omega }}\) such that
for any \(i\in {{\mathcal {I}}}\). Since \(u_i \le g_i \le v_i\) on \(\{0\} \times {\bar{\Omega }}\) by definition, we have \({\hat{t}} >0\) and by Proposition 3.2 we can then also assume \({\hat{x}} \in \partial \Omega \). Moreover, if we set
for \(\theta >0\) arbitrary, we have \(u^\theta < u\),
and
It follows immediately that \(u^\theta \) is a subsolution to (BVP) with \(F_i\) replaced by
and that \(u^\theta \rightarrow -\infty \) as \(t \rightarrow T\). We may therefore also assume that \({\hat{t}} < T\), since if not, we can prove \(u^\theta \le v\) as follows and then retrieve our result in the limit as \(\theta \rightarrow 0\).
For \(\epsilon >0\) (tending to 0 in the final argument) and \(i \in {\mathcal {I}}\) arbitrary but fixed, let
and consider the function
which is USC by construction. Let \((t_\epsilon , x_\epsilon , y_\epsilon )\) be the maximum point of \(\Phi _\epsilon \) on \([0,T) \times {\bar{\Omega }} \times {\bar{\Omega }}\). (This maximum point exists for \(\epsilon \) small as \(\Phi _\epsilon \) is USC, \({\hat{t}} <T\) and \({\bar{\Omega }}\) is compact.) Clearly, \(|x_\epsilon - y_\epsilon | \rightarrow 0\) as \(\epsilon \rightarrow 0\) as \((t_\epsilon , x_\epsilon , y_\epsilon )\) is a maximum point. Since
we have
where the last inequality holds since \(u_i-v_i\) is USC. This gives that also \(\frac{1}{\epsilon }|x_\epsilon -y_\epsilon |^2 \rightarrow 0\) as \(\epsilon \rightarrow 0\) and \((t_\epsilon ,x_\epsilon ) \rightarrow ({\hat{t}}, {\hat{x}})\) since \(u_i(t,x) - v_i(t,x) - \varphi _\epsilon (t,x,x)\) has a maximum at \(({\hat{t}}, {\hat{x}})\). Moreover, from the upper- and lower semi-continuity of u and v we get
In particular, from the definition of \(\varphi _\epsilon \) and \((t_\epsilon , x_\epsilon , y_\epsilon )\) we get
Note that \(u_i \in USC ([0,T]\times {\bar{\Omega }})\) so we have \(\limsup _{\epsilon \rightarrow 0} u_i(t_\epsilon , x_\epsilon ) \le u_i({\hat{t}}, {\hat{x}})\). Since \((t_\epsilon , x_\epsilon , y_\epsilon ) \rightarrow ({\hat{t}}, {\hat{x}}, {\hat{x}})\) as \(\epsilon \rightarrow 0\), this inequality cannot be strict; if it were, (3.2) shows that we must necessarily have \(\liminf _{\epsilon \rightarrow 0} v_i(t_\epsilon ,x_\epsilon )< v({\hat{t}}, {\hat{x}})\) as well, but this contradicts \(v_i \in LSC([0,T]\times {\bar{\Omega }})\). An analogous argument shows \(v_i(t_\epsilon , x_\epsilon ) \rightarrow v_i({\hat{t}}, {\hat{x}})\).
We now invoke the exterior sphere condition which implies the existence of \(r>0\) s.t.
Differentiating gives
and thus, if \(x_\epsilon \in \partial \Omega \) we have
Hence, as \(\epsilon \rightarrow 0\) we have
for some \(D\ge 0\). Similarly, for the supersolution v we get for \(y_\epsilon \in \partial \Omega \)
and thus that, as \(\epsilon \rightarrow 0\),
for some \({\tilde{D}} \le 0\) (since \(u_i > v_i\) at \(({\hat{t}}, {\hat{x}})\) and \(f_i\) is non-decreasing by assumption (BC1)). Hence, if
is a positive maximum of \(u_i-v_i\) over \((0,T) \times {\bar{\Omega }}\) we must have
provided \(\epsilon \) is small enough. Indeed, this holds by continuity of \(F_i\) and since, by Step 1, we can consider \(u_i\) and \(v_i\) to be sub- and supersolutions to (BVP) with boundary conditions \({\check{B}}_i : =B_i + \eta \le 0\) and \({\hat{B}}_i : =B_i -\eta \ge 0 \), respectively. Thus, since we have (3.3) and (3.4), the boundary conditions cannot be satisfied and therefore the equation must hold on the boundary (in the sub- /supersolution sense).
3.3 Step 3. Avoiding the Obstacle
We will now argue as in Ishii–Koike [15] to ensure that the subsolution (more precisely, at least one component of it) lies strictly above its obstacle at \(({\hat{t}}, {\hat{x}})\). To do this, recall the set \({{\mathcal {I}}}\) and assume that \(u_i({\hat{t}}, {\hat{x}}) \le {\mathcal {M}}_i u({\hat{t}}, {\hat{x}})\), i.e.,
whenever \(i \in {{\mathcal {I}}}\). This implies the existence of \(k \in \{1, \dots , i-1, i+1,\dots , m\}\) s.t.
Moreover, since v is a supersolution we have
Combining the above two inequalities yield
but since \(({\hat{t}}, {\hat{x}})\) is a maximum of \(u_i-v_i\) and \(i \in {{\mathcal {I}}}\) this must in fact be an equality and \(k \in I\) as well. Repeating this as many times as necessary, we find the existence of a sequence of indices \(\{i_1, i_2, \dots , i_p, i_1\}\), \(i_p \ne i_{p+1}\) such that
which implies
a contradiction to (O1). Thus, there exists at least one index \(i \in {{\mathcal {I}}}\) s.t.
Since \(u_i(t_\epsilon , x_\epsilon ) \rightarrow u_i(\hat{t}, \hat{x})\) and \(v_i(t_\epsilon , x_\epsilon )\rightarrow v_i(\hat{t}, \hat{x})\) for all \(i \in \{1, \dots , m\}\), we can then conclude from this and Step 2 that
for at least one \(i \in {{\mathcal {I}}}\) and \(\epsilon \) small enough.
3.4 Step 4. Reaching the Contradiction
We are now ready to reach our final contradiction. To do this, we will use the following lemma, the so called maximum principle for semi-continuous functions. Lemma 3.3 corresponds to Theorem 8.3 of [7] in a less general but for our purposes sufficient form.
Lemma 3.3
Suppose that \((t_\epsilon , x_\epsilon , y_\epsilon )\) is a maximum point of
over \((0,T) \times {\bar{\Omega }}\). Then, for each \(\theta >0\) there are \(X, Y \in \mathcal {S}^n\) such that
where \(A :=D_x^2 \varphi _\epsilon (t_\epsilon , x_\epsilon , y_\epsilon )\) is the Hessian matrix of \(\varphi (t,x,y)\) (w.r.t. x and y).
We first note that
which gives
Choosing \(\theta = \epsilon \) in Lemma 3.3 (ii) gives the existence of
with
Note in particular that this implies that for any \(\xi >0\) fixed there exists \(\epsilon >0\) such that the above holds with
Moreover, we have from (3.5) that
which implies
by Lemma 3.3 (iii).
What remains is to show that (3.7) is inconsistent with (3.1), (3.6) and the assumptions (F1) - (F3). To see this, note that
where we have used, in turn, (F2),
and (F1) (more precisely, (F1*)). Since (3.6) holds we get from (F3) that
and from (F2) that
and by combining these displays we have
Putting (3.7), (3.8), and (3.9) together gives
Taking the limit as \(\epsilon \rightarrow 0\) and recalling \(\frac{1}{\epsilon }|x_\epsilon -y_\epsilon |^2 \rightarrow 0\) we arrive at
which is a contradiction since \(\xi >0\) was arbitrary and \(\gamma \delta > 0\). \(\square \)
Proof
(Proof of Proposition 3.2) Assume first that \(u_i \le v_i\) on \((0,T) \times \partial \Omega \) for all \(i \in \{1,\dots , m\}\). As above, assume the existence of \({{\mathcal {I}}}\) and \(({\hat{t}}, {\hat{x}})\) such that
for all \(i \in {{\mathcal {I}}}\). Now, let
and construct \(\Phi _\epsilon (t,x,y) := u_i(t,x) -v_i (t,y) - \varphi _\epsilon (t,x,y)\). Calculations analogous to those in the first half of Step 2 above shows
and we can thus simply repeat Steps 3-4 above to get the desired contradiction.
For the general case, let
and note that \({\tilde{u}}:=u-K \le u\) is a subsolution to (BVP) in the sense of Definition 2.3 (i) and (iii) since \({\tilde{u}}_i(0,\cdot ) \le u_i(0,\cdot ) \le g_i\), K is independent of i and
by (F1). Moreover, \({\tilde{u}} \le v\) on \((0,T) \times \partial \Omega \) so we can apply the above result to conclude that
and we are done. \(\square \)
Proof
(Proof of Corollary 2.6) If u is a viscosity subsolution, then so is \(u-K\) for all \(K>0\). It thus suffices to prove that if \(u\le v\) on \((0,T) \times \partial \Omega )\setminus G\), then \(u\le v\) in \([0,T) \times {\bar{\Omega }}\). If \(G=(0,T) \times \partial \Omega \), this implication and its proof is identical to Theorem 2.4 and if \(G= \emptyset \) it is identical to Proposition 3.2. If \(G \subset (0,T) \times \partial \Omega \) is a non-empty proper subset, we know by assumption that \(u\le v\) on \(((0,T) \times \partial \Omega ) \setminus G\) and so the maximum point \(({\hat{t}}, {\hat{x}})\) defined in (3.1) must belong to the set G where the boundary condition is satisfied. Hence, we can follow the proof of Theorem 2.4 to conclude that \(u\le v\) in \([0,T) \times {\bar{\Omega }}\). \(\square \)
4 Proof of Theorem 2.5: Perron’s Method and Barrier Construction
Throughout this section, we assume that the assumptions of Theorem 2.5 hold, i.e., that (D1), (F1)–(F3), (BC1), (O1)–(O3) and (I1) hold. Our proof of existence follows the machinery of Perron’s method. In particular, we have the following result.
Proposition 4.1
Assume that, for each \(i \in \{1, \dots , m\}\) and \({\hat{x}} \in {\bar{\Omega }}\) there exist families of continuous viscosity sub- and supersolutions, \(\{u^{i,{\hat{x}},\epsilon }\}_{\epsilon >0}\) and \(\{v^{i,{\hat{x}},\epsilon }\}_{\epsilon >0}\), to (BVP) such that
Then,
is a viscosity solution to (BVP).
With this result given, what remains is to construct appropriate barriers, i.e., families of viscosity sub- and supersolutions taking on the correct initial data. More specifically, we prove the following.
Proposition 4.2
For any \({\hat{x}} \in {\bar{\Omega }}\) and \(\epsilon > 0\), there exist non-negative constants A, B, C and \(\kappa \) such that \(U^{{\hat{x}}, \epsilon } :=(U_1^{{\hat{x}}, \epsilon }, \ldots , U^{{\hat{x}}, \epsilon }_m)\) and \(V^{i, {\hat{x}}, \epsilon }:=(V_1^{i, {\hat{x}}, \epsilon }, \dots , V^{i, {\hat{x}}, \epsilon }_m)\),
where \(\varphi (x)\) is given in Lemma 3.1 and \(i\in \{1,\dots , m\}\), are viscosity sub- and supersolutions to (BVP), respectively. Moreover,
Combining Propositions 4.1 and 4.2 above proves the existence part of Theorem 2.5. Uniqueness follows immediately from Theorem 2.4 and the definition of sub- and supersolutions. What needs to be done is thus to prove the above propositions. Being the non-standard one, we start with Proposition 4.2.
Proof
(Proof of Proposition 4.2) We prove only the supersolution property (the subsolution property is proven analogously but without the need to deal with the obstacle as in (4.2) below). We need to show that \((V_1^{i, {\hat{x}}, \epsilon }, \cdots , V^{i, {\hat{x}}, \epsilon }_m)\) satisfies condition \((i)-(iii)\) of Definition 2.3. To ease notation, we suppress the superindices \(i, {\hat{x}}, \epsilon \) and write \(V_j\) in place of \(V^{i,{\hat{x}},\epsilon }_j\).
We begin with condition (iii). For any \(\epsilon >0\) and A given, we can ensure \(V_j(0,x) \ge g_j(x), \forall x \in {\bar{\Omega }}\) and all \(j \in \{1,\dots ,m\}\) by choosing B sufficiently large. Indeed, this is possible since \(\varphi , g\), and \(c_{ij}\) are continuous and
where the last inequality is due to Assumption (I1). Moreover, we can let \(\epsilon \rightarrow 0\) (by letting \(B \rightarrow \infty \) if necessary) and since \(c_{ii}(t,x)\equiv 0\) we thus have that
Note that the choice of \(B=B(\epsilon ,A)\) can be made with \(C= \kappa = 0\); later increasing \(\kappa \) and/or C will only increase V further while keeping (4.1).
We next turn to condition (i) of Definition 2.3. Starting with the obstacle, we have
where the last inequality is by assumption (O3). Concerning the second part we observe that, for some C large enough, it holds that
for \((a,p,X) \in \mathcal {P}^{2,-}_{ \Omega }V_i(t,x)\), for all \((t,x) \in (0,T)\times \Omega \) and all \(i \in \{1,\dots ,m\}\). Indeed, this follows after noticing that \(V_i\) smooth and \(F_i\) continuous give a lower bound for \(F_i(t,x,V_i(t,x), D V_i(t,x), D^2 V_i(t,x))\) on the compact region \([0,T]\times {\bar{\Omega }}\). This lower bound can be made independent from C (and \(\epsilon \)) since \(F_i\) is assumed non-decreasing with \(V_i\) and \(V_i\) is non-decreasing in C (and \(\epsilon \)); it may however depend on A, B and the parameter \(\kappa \) to be chosen later. Since \(\partial _t V_j(t,x) = C + \partial _t c_{ij}(t,x)\) and the latter term is bounded in \([0,T]\times {\bar{\Omega }}\), we conclude that inequality (4.3) holds for large enough \(C=C(A,B,\kappa )\).
What remains to verify is condition (ii) of Definition 2.3. Note that, for \(x \in \partial \Omega \), satisfying (4.3) in the classical sense does not ensure ditto in the viscosity sense. We therefore instead focus on the Neumann condition and intend to prove that
for \((t,x) \in (0,T)\times \partial \Omega \) and for all \(j \in \{1,\dots ,m\}\). Recall that we have chosen B s.t. \(V_j(0,x) \ge g_j(x)\) in \({\bar{\Omega }}\). Since \(\partial _t c_{ij}(t,x)\) is bounded from below, we can also ensure that \( V_j(t,x) \ge g_j(x) \) holds for all \((t,x) \in [0,T]\times {\bar{\Omega }}\) by increasing C if necessary. By the monotonicity property of f (non-decreasing in r) we then have
Let \({\tilde{A}}\) be such that
The Neumann boundary condition thus follows if we can show that
for all \((t,x)\in (0,T)\times \partial \Omega \). Differentiating \(V_j(t,x)\) gives
and thus
where \(|\langle n(x), D c_{ij}\rangle |\) is bounded since \(c_{ij}\) is smooth. The exterior ball condition implies that
where r depends on \(\Omega \). From this we see that, for \(\kappa \) large enough and depending only on \(\Omega \), the last term in (4.5) is non-negative since \(\langle n(x), D\varphi (x) \rangle \ge 1\) by Lemma 3.1.
Using \(\langle n(x), D\varphi (x) \rangle \ge 1\) once more it now only remains to pick A such that \(A - \max _{i,j \in \{1, \dots , m\}}|D c_{ij}| \ge \tilde{A}\) in order to fulfill (4.4). Observe that this choice of A depends only on the data of the problem. With \(\kappa \) and A now fixed we conclude that for any \(\epsilon > 0\) we can choose \(B = B(\epsilon , A)\) and then \(C = C(A,B,\kappa )\) such that the above calculations hold.
Last, we note that the constructed barrier now satisfies the boundary condition in the classical sense. This suffices as Proposition 7.2 of [7] and the specific form of the function \(B_i\) (cf. (7.4) of [7]) then gives that the boundary condition also holds in the viscosity sense. \(\square \)
Proof (Proof of Proposition 4.1)
[Proof of Proposition 4.1] Note first that w is well defined and bounded by the assumption of existence of sub- and supersolutions. Let \(w_*:=(w_{*, 1}, \ldots , w_{*,m})\) and \(w^*:=(w^*_1, \ldots , w^*_m)\) denote the lower- and upper semi-continuous envelopes of \(w=(w_1, \dots , w_m)\), respectively, i.e., the largest LSC function that is dominated by w and the smallest USC function that dominates w, respectively. By definition and Theorem 2.4 we have, for any \(i, {\hat{x}}\) and \(\epsilon >0\) fixed,
The essence of Perron’s method is to now prove that \(w_*\) is a supersolution and \(w^*\) a subsolution to (BVP), implying that also \(w^*\le w_*\) by Theorem 2.4 and thus that \(w = w_*=w^*\) is a solution to (BVP).
To establish the subsolution property, we need to show
Statement (iii) follows immediately from the assumptions and Theorem 2.4. Indeed, \(w_i \le v^{i,{\hat{x}}, \epsilon }_i\) for any \(i, {\hat{x}},\) and \(\epsilon \) and thus \(w^*_i(0,{\hat{x}}) \le \inf _{\epsilon } (v_i^{i,{\hat{x}}, \epsilon })^*(0,{\hat{x}}) =g_i({\hat{x}})\) (where the last equality is by assumption).
Concerning statement (i), we first note that by the definition of \(w^*_{i}\) there exists a sequence
with each \(u^n\) being a subsolution of (BVP). Assume now that \((a,p,X) \in {\bar{\mathcal {P}}}^{2,+}_{\Omega }w^*_i({\hat{t}},{\hat{x}})\) for \(({\hat{t}},{\hat{x}}) \in (0,T)\times \Omega \). Then, by the existence of the sequence (4.6) and the fact that \(w_i^*\) is USC we get from Proposition 4.3 of [7] that there exists a sequence
Since \(({\hat{t}},{\hat{x}}) \in (0,T)\times \Omega \) we will have \((t_n, x_n) \in (0,T)\times \Omega \) as well for n large enough and by the subsolution property of \(u^n\) and the fact that F is continuous we get
and thus \(w^*\) satisfies (i) above. If \((a,p,X) \in {\bar{\mathcal {P}}}^{2,+}_{\Omega }w^*_i({\hat{t}},{\hat{x}})\) for \(({\hat{t}}, {\hat{x}}) \in (0,T)\times \partial \Omega \), analogous reasoning gives the existence of a sequence such that
for all n. Taking the limit as \(n \rightarrow \infty \), now using also that \(f_i\) is continuous, we conclude that also (ii) holds. In particular, we then have \(w^*=w\).
We now prove that \(w_*\) is a supersolution following a classical argument by contradiction. More precisely, we show that if \(w_*\) is not a supersolution, then there exists a subsolution strictly greater than \(w ^*\), contradicting the very definition of w.
Starting with the initial condition (iii) we have \(w_i (0,{\hat{x}}) \ge u_i^{i,{\hat{x}}, \epsilon } (0,{\hat{x}})\) for all \(\epsilon >0\) and thus \(w_{*,i}(0,{\hat{x}}) \ge \sup _\epsilon (u_i^{i,{\hat{x}}, \epsilon })_*(0,{\hat{x}})= \sup _\epsilon u_i^{i,{\hat{x}}, \epsilon }(0,{\hat{x}}) =g_i({\hat{x}})\). Assume now that \(w_*\) is not a supersolution by violating Definition 2.3 (i), i.e., that for some \(({\hat{t}},{\hat{x}}) \in (0,T)\times \Omega \) and \(i \in \{1,\dots , m\}\) we have
for some \((a,p,X) \in {\bar{\mathcal {P}}}^{2,-}_{\Omega }w_{*,i}({\hat{t}}, {\hat{x}})\). For \(\delta >0\) and such (a, p, X) fixed, construct the function
Since \({{\mathcal {M}}}_i w_{*} \le {{\mathcal {M}}}_i w^*\) it follows from continuity that the function \({\tilde{w}}(t,x)\) is a viscosity subsolution to
for \((t,x) \in Q_R := \{(t,x) : |t-{\hat{t}}| + |x-{\hat{x}}|^2 <R\}\) and \(\delta \), \(\beta \) and R sufficiently small. (If \(t \ne {\hat{t}}\), \({\tilde{w}}\) satisfies (4.7) in the classical sense. If \(t={\hat{t}}\), \({\bar{\mathcal {P}}}^{2,+}_{\Omega }{\tilde{w}}(t,x) = \{(a+\beta \eta , p, X) : \eta \in [-1,1]\}\) and the contribution from \(\beta \eta \) is harmless if \(\beta \) is small enough.) By definition of \({\bar{\mathcal {P}}}^{2,-}_{\Omega }w_{*,i}({\hat{t}}, {\hat{x}})\) we have
and thus, if we let \(\delta = \frac{\beta R}{4}\) and consider \((t,x) \in Q_R \setminus Q_{\frac{R}{2}}\), we get
Now let \({\check{u}} = \{{\check{u}}_1, \dots , {\check{u}}_m\}\) where
Note that by (4.8), there is no jump in \({\check{u}}_i\) at \(\partial Q_R\) if R is small enough. Since \({\check{u}}_i \ge w^*_i\) we have
for any \(j \ne i\). Recalling that \(w^*_i\) is a subsolution, the above shows that \({\check{u}}\) is a subsolution to (BVP) as well (after decreasing R even further if necessary to ensure \(Q_R \in (0,T)\times \Omega \)). We now note that, since \((w^*_i)_*= w_{*,i}\), there exists by definition a sequence \((t_n, x_n, w^*_i(t_n, x_n)) \rightarrow ({\hat{t}}, {\hat{x}}, w_{*,i}({\hat{t}}, {\hat{x}}))\). If we follow this sequence we thus find that \({\check{u}}_i (t_n, x_n) = {\tilde{w}}_i (t_n, x_n) > w^*_i\) for some point \((t_n,x_n)\) sufficiently close to \(({\hat{t}}, {\hat{x}})\) (since \({\tilde{w}}_i \rightarrow w_{*,i} + \delta > w_{*,i}\)). Thus, we have constructed a subsolution which is strictly greater than w, a contradiction.
What remains is to consider if \(w_*\) fails to be a supersolution by violating condition (ii), i.e., if there exists \(({\hat{t}}, {\hat{x}}) \in (0,T)\times \partial \Omega \) and \((a,p,X) \in {\bar{\mathcal {P}}}^{2,-}_{\Omega }w_*({\hat{t}}, {\hat{x}})\) s.t.
However, if (4.9) holds, continuity of \(f_i\) and the smoothness of \(\partial \Omega \) gives that \(B_i(t, x, {\tilde{w}}(t,x), p) \le 0\) (in the classical sense and thus in the viscosity sense by Proposition 7.2 of [7]) for \((t,x) \in \partial \Omega \) and sufficiently close to \(({\hat{t}}, {\hat{x}})\). We then conclude as above that there exists \(\delta , \beta \), and R s.t. \({\tilde{w}}\) satisfies the subsolution property (4.7) for \((t,x) \in (0,T)\times \Omega \) sufficiently close to \(({\hat{t}}, {\hat{x}})\) and thus that \({\check{u}}\) is a subsolution. Again, we have constructed a subsolution which dominates \(w^*\), contradicting the very definition of w. The proof is complete. \(\square \)
References
Aikawa, H., Kilpeläinen, T., Shanmugalingam, N., Zhong, X.: Boundary Harnack principle for pharmonic functions in smooth Euclidean domains. Potential Anal. 26, 281–301 (2007)
Barkhudaryan, R., Gomes, D.A., Shahgholian, H., Salehi, M.: System of variational inequalities with interconnected obstacles. Appl. Anal. (2020). https://doi.org/10.1080/00036811.2020.1757076
Barles, G.: Fully non-linear Neumann type boundary conditions for second-order elliptic and parabolic equations. J. Differ. Equ. 106(1), 90–106 (1993)
Biswas, I.H., Jakobsen, E.R., Karlsen, K.H.: Viscosity solutions for a system of integro-PDEs and connections to optimal switching and control of jump-diffusion processes. Appl. Math. Optim. 62, 47–80 (2010)
Bourgoing, M.: Viscosity solutions of fully non-linear second order parabolic equations with \(L^1\) dependence in time and Neumann boundary conditions. Discret. Contin. Dyn. Syst. 21(3), 763–800 (2008)
Boufoussi B., Hamadene. S., Jakani, M.: Viscosity Solutions of Systems of PDEs with Interconnected Obstacles and Nonlinear Neumann Boundary Conditions, hal-03120824v2 (2022)
Crandall, M.G., Ishii, H., Lions, P.-L.: User’s guide to viscosity solutions of second order partial differential equations. Bull. Am. Math. Soc. 27, 1–67 (1992)
Djehiche, B., Hamadene, S., Popier, A.: A finite horizon optimal multiple switching problem. SIAM J. Control Optim. 48, 2751–2770 (2010)
Dupuis, P., Ishii, H.: On oblique derivative problems for fully non-linear second-order elliptic equations on non-smooth domains. Nonlinear Anal. 15, 1123–1138 (1990)
Dupuis, P., Ishii, H.: On oblique derivative problems for fully non-linear second-order elliptic PDEs on domains with corners. Hokkaido Math. J. 20, 135–164 (1991)
El-Asri, B., Hamadene, S.: The finite horizon optimal multi-modes switching problem: the viscosity solution approach. Appl. Math. Optim. 60, 213–235 (2009)
Fuhrman, M., Morlais, M.-A.: Optimal switching problems with an infinite set of modes: an approach by randomization and constrained backward SDEs. Stoch. Process. Appl. 130(5), 3120–3153 (2020)
Hamadene, S., Morlais, M.A.: Viscosity solutions of systems of PDEs with interconnected obstacles and switching problem. Appl. Math. Optim. 67, 163–196 (2013)
Hu, Y., Tang, S.: Multi-dimensional BSDE with oblique reflection and optimal switching. Probab. Theory Relat. Fields 147, 89–121 (2010)
Ishii, H., Koike, S.: Viscosity solutions of a system on non-linear second-order elliptic PDEs arising in switching games. Funkcialaj Ekvacioj 34, 143–155 (1991)
Ishii, H., Sato, M.-H.: Nonlinear oblique derivative problems for singular degenerate parabolic equations on a general domain. Nonlinear Anal. 57, 1077–1098 (2004)
Kharroubi, I.: Optimal switching in finite horizon under state constraints. SIAM J. Control Optim. 54, 2202–2233 (2016)
Klimsiak, T.: Systems of quasi-variational inequalities related to the switching problem. Stoch. Process. Appl. 129, 1259–1286 (2019)
Lundström, N.L.P., Nyström, K., Olofsson, M.: Systems of variational inequlities for non-local operators related to optimal switching problems: existence and uniqueness. Manuscr. Math. 145, 407–432 (2014)
Lundström, N.L.P., Nyström, K., Olofsson, M.: Systems of variational inequalities in the context of optimal switching problems and operators of Kolmogorov type. Ann. Math. Pura Appl. 193, 1213–1247 (2014)
Lundström, N.L.P., Olofsson, M., Önskog, T.: Existence, uniqueness and regularity of solutions to systems of nonlocal obstacle problems related to optimal switching. J. Math. Anal. Appl. 475(1), 13–31 (2019)
Lundström N.L.P., Olofsson M., Önskog T.: Management strategies for run-of-river hydropower plants: an optimal switching approach. Optim. Eng. 1–25 (2021)
Lundström, N.L.P., Önskog, T.: Stochastic and partial differential equations on non-smooth time-dependent domains. Stoch. Process. Appl. 129(4), 1097–1131 (2019)
Reisinger, C., Zhang, Y.: Error estimates of penalty schemes for quasi-variational inequalities arising from impulse control problems. SIAM J. Control Optim. 58(1), 243–276 (2020)
Acknowledgements
We thank two anonymous referees for careful reading of the manuscript and for comments which have significantly improved the presentation. Marcus Olofsson was financed in full and Niklas Lundström in part by the Swedish research council (grant no. 2018-03743). We gratefully acknowledge this support.
Funding
Open access funding provided by Umea University.
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
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
Lundström, N.L.P., Olofsson, M. Systems of Fully Nonlinear Parabolic Obstacle Problems with Neumann Boundary Conditions. Appl Math Optim 86, 27 (2022). https://doi.org/10.1007/s00245-022-09890-z
Accepted:
Published:
DOI: https://doi.org/10.1007/s00245-022-09890-z
Keywords
- Optimal switching
- Multi modes switching
- Constraints
- Fully nonlinear PDE
- Neumann boundary condition
- Viscosity solution
- Barrier