A “MAXIMUM PRINCIPLE FOR SEMICONTINUOUS FUNCTIONS”

APPLICABLE TO INTEGRO-PARTIAL DIFFERENTIAL EQUATIONS

ESPEN R. JAKOBSEN AND KENNETH H. KARLSEN

Abstract. We formulate and prove a non-local “maximum principle for semicontinuous

functions” in the setting of fully nonlinear degenerate elliptic integro-partial differential equa-
tions with integro operators of second order. Similar results have been used implicitly by
several researchers to obtain comparison/uniqueness results for integro-partial differential
equations, but proofs have so far been lacking.

1. Introduction
The theory of viscosity solutions (existence, uniqueness, stability, regularity etc.) for fully
nonlinear degenerate second order partial differential equations is now highly developed [4,
5, 13, 15]. In recent years there has been an interest in extending viscosity solution theory
to integro-partial differential equations (integro-PDEs henceforth) [1, 2, 3, 6, 8, 9, 10, 29,
30, 31, 32, 33, 34]. Such non-local equations occur in the theory of optimal control of Lévy
(jump-diffusion) processes and find many applications in mathematical finance, see, e.g.,
[1, 2, 3, 9, 8, 10, 17] and the references cited therein. We refer to [18, 19] for a deep investigation
of integro-PDEs in the framework of Green functions and regular solutions, see also [20].
In this paper we are interested in comparison/uniqueness results for viscosity solutions of
fully nonlinear degenerate elliptic integro-PDEs on a possibly unbounded domain Ω ⊂ R N .
To be as general as possible, we write these equations in the form
(1.1) F (x, u(x), Du(x), D 2 u(x), u(·)) = 0 in Ω,
where F : Ω × R × RN × SN × Cp2 (Ω) → R is a given functional. Here SN denotes the space
of symmetric N × N real valued matrices, and C p2 (Ω) denotes the space of C 2 (Ω) functions
with polynomial growth of order p ≥ 0 at infinity. At this stage we simply assume that the
non-local part of F is well defined on Ω.
These equations are non-local as is indicated by the u(·)-term in (1.1). A simple example
of such an equation is
Z
 
(1.2) −ε∆u + λu − u(x + z) − u(x) − zDu(x) 1|z|<1 m(dz) = f (x) in RN ,
RN \{0}

Date: February 19, 2004.

Key words and phrases. nonlinear degenerate elliptic integro-partial differential equation, Bellman equation,
Isaacs equation, viscosity solution, comparison principle, uniqueness.
Espen R. Jakobsen is supported by the Research Council of Norway under grant 151608/432. Kenneth H.
Karlsen is supported in part by the BeMatA program of the Research Council of Norway and the European
network HYKE, funded by the EC as contract HPRN-CT-2002-00282. This work was done while Espen R.
Jakobsen visited the Center of Mathematics for Applications (CMA) and the Department of Mathematics at
the University of Oslo, Norway, and he is grateful for the kind hospitality. We are grateful to Guy Barles for
his critical comments and to Harald Hanche Olsen for pointing out some mathematical errors.
1
2 JAKOBSEN AND KARLSEN

where λ > 0, ε ≥ 0, f : RN → R is uniformly continuous, and m(dz) is a non-negative Radon

measure on RN \ {0} (the so-called Lévy measure) with a singularity at the origin satisfying
Z


(1.3) z 2 1|z|<1 + |z|p 1|z|≥1 m(dz) < ∞.
RN \{0}

Note that the Lévy measure integrates functions with p-th order polynomial growth at infinity.
In view of (1.3), a simple Taylor expansion of the integrand shows that u has to belong to
Cp2 (RN ) for the integro operator in (1.2) to be well defined. From this we also see that the
integro operator in (1.2) acts as a non-local second order term, and for that reason the “order”
of the integro operator is said to be two. If |z| 2 in (1.3) is replaced by |z|, this changes the
order of the integro operator from two to one, since then it acts just like a non-local first
order term. Finally, if |z|2 in (1.3) is replaced by 1 (i.e., m(dz) is a bounded measure), then
the integro operator in (1.2) is said to be bounded or of order zero. In the bounded case, the
integro operator acts just like a non-local zero order term.
A significant example of a non-local equation of the form (1.1) is the non-convex Isaacs
equations associated with zero-sum, two-player stochastic differential games (see, e.g., [16] for
the case without jumps)
n o
(1.4) inf sup −Lα,β u(x) − I α,β u(x) + f α,β (x) = 0 in RN ,
α∈A β∈B

where A and B are compact metric spaces and for any sufficiently regular φ,
 h i

 Lα,β φ(x) = tr aα,β (x)D 2 φ + bα,β (x)Dφ + cα,β (x)φ,



 α,β 1 T
(1.5) a (x) = σ α,β (x)σ α,β (x) ≥ 0,


 Z2 h i

 I α,β φ(x) =
 φ(x + η α,β (x, z)) − φ(x) − η α,β (x, z)Dφ(x) 1|z|<1 m(dz).
RM \{0}

Here tr and T denote the trace and transpose of matrices. The Lévy measure m(dz) is a
nonnegative Radon measure on RM \ {0}, 1 ≤ M ≤ N , satisfying a condition similar to (1.3),
see (A0) and (A4) in Section 3. Also see Section 3 for the (standard) regularity assumptions
on the coefficients, σ α,β (x), bα,β (x), cα,β (x), and η α,β (x, z). We remark that if A is a singleton,
then equation (1.4) becomes the convex Bellman equation associated with optimal control of
Lévy (jump-diffusion) processes over an infinite horizon (see, e.g., [29, 30] and the references
therein). Henceforth we call equation (1.4) for the Bellman/Isaacs equation.
Rather general existence and comparison/uniqueness results for viscosity solutions of first
order integro-PDEs (no local second order term) can be found in [33, 34, 31, 32], see also [9]
for the Bellman equation associated with a singular control problem arising in finance.
Depending on the order of the integro operator (i.e., the assumptions on the singularity of
the Lévy measure m(dz) at the origin), the case of degenerate elliptic (or parabolic) integro-
PDEs is more complicated. When the integro operator is of order zero (bounded), general
existence and comparison/uniqueness results for (semicontinuous and unbounded) viscosity
solutions are given in [1, 2, 3]. When the integro operator is of second order (i.e., the Lévy
measure m(dz) is unbounded near the origin as in (1.3)), the existence and uniqueness of
unbounded viscosity solutions of systems of semilinear degenerate parabolic integro-PDEs
is proved in Barles, Buckdahn, and Pardoux [6]. Pham [30] proved an existence result and
a comparison principle among uniformly continuous (and hence at most linearly growing)
 INTEGRO-PARTIAL DIFFERENTIAL EQUATIONS 3

viscosity sub- and supersolutions of parabolic integro-PDEs of the Bellman type (i.e. (1.4)
with singleton A). Motivated by singular stochastic control applications in finance, the papers
[8, 10] provide existence and comparison results for non-local degenerate elliptic free boundary
problems with state-constraint boundary conditions.
The main contribution of the present paper is to provide “non-local” versions of Propo-
sition 5.1 in Ishii [22] (see also Proposition II.3 in Ishii and Lions [23]) and Theorem 1 in
Crandall [11], which are properly adapted to integro-PDEs of the form (1.1). This “non-local
maximum principle” is used to obtain comparison principles for semicontinuous viscosity sub-
and supersolutions of (1.1). Although there exist already comparison results for some integro-
PDEs with a second order integro operator, see [6, 30, 8, 10], they are all based on the by now
standard approach that uses the maximum principle for semicontinuous functions [12, 13].
As we argue for in Section 2, it is in general not clear how to implement this approach for
non-local equations. After all the maximum principle for semicontinuous functions [12, 13] is
a local result! This was one of our motivations for writing this paper, which in contrast to
[6, 30, 8, 10] advocates the use of original approach due to Jensen, Ishii, Lions [27, 22, 23] for
proving comparison results for non-local equations. Although our main result (see Theorem
4.8) is not surprising, and the tools used in the proof are nowadays standard in the viscosity
solution theory, it has not appeared in the literature before and in our opinion it seem to
provide the “natural” framework for deriving general comparison results for fully nonlinear
degenerate second order integro-PDEs. Moreover, we stress that our Theorem 4.8 cannot be
derived directly from the maximum principle for semicontinuous functions [12, 13] (although
it is well known that this can be done in the pure PDE case).
In addition to the main result mentioned above, our paper complements the existing liter-
ature [6, 30] (see also [8, 10]) on second order PDEs with integro operators of second order
in the following ways: (i) Our formulation is abstract and more general, (ii) we consider only
semicontinuous sub- and supersolutions, and (iii) we consider (slightly) more general integro
operators (see Remark 6.1).
The remaining part of this paper is organized as follows: In Section 2 we discuss our
main result (Theorem 4.8) in the simplest possible context of (1.2) and relate it to some of
the existing literature on integro-PDEs. In Section 3 we first list our assumptions on the
coefficients in the Bellman/Isaacs equation (1.4). Then we state and discuss a comparison
theorem for these equations. In Section 4 we first give two equivalent definitions of a viscosity
solution for (1.1) and illustrate them on the Bellman/Isaacs equation. Then we state our main
result (Theorem 4.8). In Section 5 we list structure conditions on (1.1) implying, via Theorem
4.8, comparison/uniqueness results on unbounded domains. The structure conditions are
illustrated on the Bellman/Isaacs equation. Section 6 contains the proof of the comparison
theorem for the Bellman/Isaacs equation, and the proof of Theorem 4.8 is given in Section 7.
Although we only discuss elliptic integro-PDEs, it is not hard to formulate “parabolic”
versions of our main results (for example Theorem 4.8), see [26].
We end this introduction by collecting some notations that will be used throughout this
paper. If x belong to U ⊂ Rn and r > 0, then B(x, r) = {x ∈ U : |x| < r}. We use the
notation 1U for the function that is 1 in U and 0 outside. By a modulus ω, we mean a
positive, nondecreasing, continuous, sub-additive function which is zero at the origin. Let
C n (Ω) n = 0, 1, 2 denote the spaces of n times continuously differentiable functions on Ω.
We let U SC(Ω) and LSC(Ω) denote the spaces of upper and lower semicontinuous functions
on Ω, and SC(Ω) = U SC(Ω) ∪ LSC(Ω). A lower index p denotes the polynomial growth at
4 JAKOBSEN AND KARLSEN

infinity, so Cpn (Ω), U SCp (Ω), LSCp (Ω), SCp (Ω) consist of functions f from C n (Ω), U SC(Ω),
LSC(Ω), SC(Ω) satisfying the growth condition
|f (x)| ≤ C(1 + |x|p ) for all x ∈ Ω.
Finally, in the space of symmetric matrices S N we denote by ≤ the usual ordering (i.e.
X ∈ SN , 0 ≤ X means that X positive semidefinite) and by | · | the spectral radius norm (i.e.
the maximum of the absolute values of the eigenvalues).

2. Discussion of main result

To explain the contribution of the present paper and put it in a proper perspective with
regards some of the existing literature [6, 30, 8, 10], let us elaborate on a difficulty related to
proving comparison/uniqueness results arising from the very notion of a viscosity solution.
For illustrative purposes, we focus on the simple equation (1.2). The general case (1.1) will
be treated in the sections that follow.
First of all, since the equation is non-local it is necessary to use a global formulation of
viscosity solutions: A function u ∈ U SC(R N ) is a viscosity subsolution of (1.2) if
Z
 
(2.1) −ε∆φ(x)+λu(x)− φ(x + z) − φ(x) − zDφ(x) 1|z|<1 m(dz) ≤ f (x) in RN ,
RM \{0}

for any x ∈ RN and φ ∈ Cp2 (RN ) such that x is a global maximum point for u − φ. Note
that (2.1) makes sense in view of (1.3) and the C p2 regularity of φ. A viscosity supersolution
u ∈ LSC(RN ) is defined similarly.
One can dispense with the growth restrictions on the test functions by replacing the defini-
tion of a viscosity subsolution (2.1) by the following equivalent one: A function u ∈ U SC p (RN )
is a viscosity subsolution of (1.2) if for any κ > 0
Z
 
−ε∆φ(x) + λu(x) − φ(x + z) − φ(x) − zDφ(x) 1|z|<1 m(dz)
0<|z|<κ
(2.2) Z
 
− u(x + z) − u(x) − zDφ(x) 1|z|<1 m(dz) ≤ f (x),
|z|≥κ

for any x ∈ RN and φ ∈ C 2 (RN ) such that x is a global maximum point for u − φ. We have
a similar equivalent formulation of a viscosity supersolution u ∈ LSC p (RN ). It is this second
formulation that is used to prove comparison/uniqueness results for (1.2).
In the pure PDE setting (m(dz) ≡ 0), nowadays comparison principles are most effectively
proved using the so-called maximum principle for semicontinuous functions [12, 13]. However,
this result is not formulated in terms of test functions, but rather in terms of the second order
2,+ 2,−
semijets J 2,+ , J 2,− , or more precisely their closures J , J (see [13] for definitions of the
2,+
semijets). Let u be a viscosity subsolution of (1.2). If (q, X) ∈ J u(x), then by definition
there exists a sequence of triples (x k , qk , Xk ) such that (qk , Xk ) ∈ J 2,+ u(xk ) for each k and
(2.3) (xk , u(xk )) → (x, u(x)), qk → q, Xk → X, as k → ∞,
and u − φ has a global maximum at x = xk for each k. According to a construction by Evans
(see, e.g., [15, Proposition V.4.1]), for each k there is a C 2 function φk : RN → R such that
φk (xk ) = u(xk ), Dφk (xk ) = qk , D 2 φk (xk ) = Xk ,
 INTEGRO-PARTIAL DIFFERENTIAL EQUATIONS 5

and u − φ has a global maximum at x = xk . Applying (2.2) we thus get

Z
 
−εXk + λu(xk ) − φk (xk + z) − φk (xk ) − zqk 1|z|<1 m(dz)
0<|z|<κ
(2.4) Z
 
− u(xk + z) − u(xk ) − zqk 1|z|<1 m(dz) ≤ f (xk ),
|z|≥κ

for each k. In view of (2.3), in the pure PDE setting (m(dz) ≡ 0) one can send k → ∞
in (2.4), the result being a formulation of the subsolution inequality (2.2) in terms of the
2,+ 2,−
elements (q, X) in J u(x). A similar formulation (in terms of the elements in J u(x))
can be given for a supersolution u. Consequently, a comparison principle in the pure PDE
case can then be deduced using the maximum principle for semicontinuous functions [12, 13].
The situation is less clear in the non-local case. When m(dz) 6= 0, we can easily send
k → ∞ in the second integral term in (2.4) thanks to u ∈ U SC p (RN ). To handle the first
integral term, suppose for the moment that the sequence {φ k }∞ 2 N
k=1 ⊂ C (R ) converges (say,
N 2 N
uniformly on compact subsets of R ) to a limit φ that belongs to C (R ) and Dφ(x) = q.
It is then clear that
Z
 
−εX + λu(x) − φ(x + z) − φ(x) − zDφ(x) 1|z|<1 m(dz)
0<|z|<κ
(2.5) Z
 
− u(x + z) − u(x) − zq 1|z|<1 m(dz) ≤ f (x),
|z|≥κ

2,+
where (q, X) ∈ J u(x) and with a similar inequality for supersolutions. Now we could
again prove comparison/uniqueness results using the maximum principle for semicontinuous
functions [12, 13]. This approach to proving a comparison principle for viscosity solutions of
integro-PDEs was first suggested by Pham [30], and later used in [9, 10]. Indeed, converted to
our setting, Lemma 2.2 in [30] states that one can find a C 2 function φ such that (2.5) holds
and another C 2 function such that the corresponding inequality for a viscosity supersolution
holds. However, there is no proof of this lemma in [30], and neither is it clear to us how to
prove (2.5) in general. To be more precise, we do not know how to prove that the sequence
{φk }∞ 2 N 2 N 2
k=1 ⊂ C (R ) has a limit point φ that in general belongs to C (R ). The C requirement
of such a limit is necessary if we want to make sense to (2.5) when the integro operator is of
second order.
We will take a different approach to proving comparison/uniqueness results for integro-
PDEs. Namely, following the original of Jensen [27], Ishii [22], Ishii and Lions [23], and
Crandall [11], we establish some sort of “non-local” maximum principle for semicontinuous
viscosity sub- and supersolutions of (1.1) (see Theorem 4.8), and then various comparison
principles can be derived from this result. Let us illustrate our approach on (1.2).
Let u and v be respectively viscosity sub- and supersolutions of (1.2). A standard trick in
viscosity solution theory for dealing with the low regularity of the solutions is the doubling
of variables device. Instead of studying directly a global maximum point of u(x) − v(x), we
consider a global maximum point (x̄, ȳ) of

u(x) − v(y) − φ(x, y),

6 JAKOBSEN AND KARLSEN

where φ is a suitable C 2 penalization term. Our main result (Theorem 4.8) applied to (1.2)
yields the following: For any γ ∈ (0, 21 ) there exists matrices X, Y ∈ SN satisfying
 
X 0 1
≤ D 2 φ(x̄, ȳ),
0 −Y 1 − 2γ
such that the following two inequalities hold:
Z
 
−εX + λu(x̄) − φ(x̄ + z, ȳ) − φ(x̄, ȳ) − zDx φ(x̄, ȳ) 1|z|<1 m(dz)
0<|z|<κ
Z
 
− u(x̄ + z) − u(x̄) − zDx φ(x̄, ȳ) 1|z|<1 m(dz) ≤ f (x̄),
|z|≥κ
Z
 
−εY + λv(ȳ) − φ(x̄, ȳ + z) − φ(x̄, ȳ) + zDy φ(x̄, ȳ) 1|z|<1 m(dz)
0<|z|<κ
Z
 
− v(ȳ + z) − v(ȳ) + zDy φ(x̄, ȳ) 1|z|<1 m(dz) ≥ f (ȳ).
|z|≥κ

The key point is that the C 2 penalization function φ(x, y) used in the doubling of variables
device occupies the slots in integro operator near the origin. Indeed, equipped with the above
result, it is possible to derive comparison/uniqueness results for (1.2) as in, e.g., Pham [30].
In [6, Proof of Theorem 3.5], Barles, Buckdahn, and Pardoux used the maximum principle
for semicontinuous functions [12, 13] and a result very much in the spirit of the above result
(or Theorem 4.8) for proving uniqueness of viscosity solutions for parabolic integro-PDEs.
However, the authors give no proof of such a result. We also stress that for the first order
version of (1.2) (ε = 0), the above two inequalities come for free from the nature of the point
(x̄, ȳ) and the definition of a viscosity solution, see, e.g., [34, 9]. However, in the second order
case (ε > 0), the proof of this result, or more generally Theorem 4.8, is more involved in
the sense that it consists of adapting the chain of arguments developed by Jensen [27], Ishii
[22], Ishii and Lions [23], and Crandall [11] to our non-local situation. The above result (or
more generally Theorem 4.8) can be viewed as some sort of “non-local”maximum principle
for semicontinuous viscosity sub- and supersolutions. It should be compared with the “local”
maximum principle for semicontinuous functions [11, 12, 13].

3. The Bellman/Isaacs equation

In this section we will give natural assumptions on the coefficients in the Bellman/Isaacs
equation (1.4) that leads to comparison results for bounded semicontinuous viscosity sub- and
supersolutions on RN . We state the comparison results, but postpone the proof to Section 6.
We remark here that Pham [30] presents a comparison principle for uniformly continuous sub-
and supersolutions of the parabolic Bellman equation. The results in this section can be seen
as slight extensions of his result (to more general non-linearities, semicontinuous sub/super
solutions and slightly more general integro operators), but the techniques are essentially the
same. The purpose of this section is simply to provide an example where we may use our “non-
local maximum principle” to obtain comparison results. Furthermore, the Bellman/Isaacs
equation, under the assumptions stated below, will serve as examples in the abstract and
more general theory developed in the sections that follow.
The following conditions are natural and standard for (1.4) in view of the connections to
the theory of stochastic control and differential games (see, e.g., [15, 16, 28, 30]):
 INTEGRO-PARTIAL DIFFERENTIAL EQUATIONS 7

There are constants K1 , Kx ≥ 0, a function ρ ≥ 0, and a modulus of continuity ω, such

that the following statements hold for every x, y ∈ R N , α ∈ A, β ∈ B and z ∈ RM \ {0}:
(A0) σ, b, c, f, η are continuous w.r.t. x, α, β and Borel measurable w.r.t. z; A, B are
compact metric spaces; and m(dz) is a positive Radon measure on R M \ {0}
satisfying
Z


ρ(z)2 1|z|<1 + 1|z|≥0 m(dz) < ∞.
RM \{0}

(A1) |σ α,β (x) − σ α,β (y)| + |b

α,β
(x) − bα,β (y)| ≤ K1 |x − y|.
(A2) |cα,β (x) − cα,β (y)| + |f α,β (x) − f α,β (y)| ≤ ω(|x − y|).
(A3) |η α,β (x, z) − η α,β (y, z)| ≤ ρ(z)|x − y|.
(A4) |η α,β (x, z)| ≤ ρ(z)(1 + |x|) and |η α,β (x, z)| 1|z|<1 ≤ Kx .
(A5) − cα,β ≥ λ > 0.
The Lévy measure m(dz) may have a singularity at z = 0. As an example in R 1 , take ρ(z) = |z|
and m(dz) = z −2−δ 1|z|<1 where δ ∈ (0, 1). According to our definition, this integro operator
has order 2. Furthermore, according to (A0), the Lévy measure m(dz) integrates bounded
functions away from the origin. Compared to Section 1 this means that p = 0 (see also (C1)
in Section 4). Because of this, (A0), (A4), and a Taylor expansion of the integrand shows that
the integro part of the Bellman/Isaacs equation (1.4) is well defined for C 02 (RN ) functions,
see [18, 19, 30].
We also remark that (A1) and (A2) imply
|σ α,β (x)| + |bα,β (x)| + |cα,β (x)| + |f α,β (x)| ≤ C(1 + |x|),
for some constant C > 0. It is the growth of f at infinity that determines the growth of the
solutions at infinity, so if f is bounded so are the solutions.
Theorem 3.1. Assume (A0) – (A5) hold, |f α,β | is bounded, and u, −v ∈ U SC0 (RN ). If u is
a viscosity subsolution and v a viscosity supersolution of (1.4), then u ≤ v in R N .
As an immediate consequence we have uniqueness of bounded viscosity solutions of (1.4).
The notion of viscosity solutions will be defined in Section 4, and Theorem 3.1 will be proved
in Section 6 using the abstract comparison result Theorem 5.2.
Remark 3.2 (Growth at infinity). If the integrability condition in (A0) is replaced by
Z


ρ(z)2 1|z|<1 + (1 + ρ(z)) 1|z|≥1 m(dz) < ∞,
RM \{0}

and we drop the assumption that f is bounded, then the above assumptions leads to problems
where the solutions may have linear growth at infinity. This case seems to be more difficult,
and we do not know if (the modified) assumptions (A0) – (A5) are sufficient to have a
comparison result. However, there are two special cases where we may have comparison
results:
• If σ, b, and η are bounded, and c is constant, then a comparison result can be obtained
by adapting the techniques of Ishii in [22] (Theorem 7.1).
8 JAKOBSEN AND KARLSEN

• If λ is sufficiently large compared to |Dσ|, |Dη|, and |Db|, then we have a comparison
result because of cancellation effects in the proof, cf. Pham [30] where this technique
is used in the parabolic case.
If the above assumption are modified appropriately and λ is big enough, then we can have
comparison results for solutions with arbitrary polynomial growth. For parabolic problems
(see, e.g., [30]) we are always in this situation since we can have λ arbitrary large after an
exponential-in-time scaling of the solution.
Remark 3.3. It is possible to consider Radon measures m depending on x, α, β under assump-
tions similar to those used in Soner [34] for first order integro-PDEs.
Remark 3.4. It is also possible to consider bounded domains Ω, but then you need a condition
on the jumps so that the jump-process does not leave Ω.

4. The main result

In this section we state two equivalent definitions of a viscosity solution and our main result
(Theorem 4.8). As we go along, we use the Bellman/Isaacs equation (1.4) for illustrative
purposes.
For every x, y ∈ Ω, r, s ∈ R, X, Y ∈ SN , and φ, φk , ψ ∈ Cp2 (Ω) we will use the following
assumptions on (1.1):
(C1) The function (x, r, q, X) 7→ F (x, r, q, X, φ(·)) is continuous, and
if xk → x, D n φk → D n φ locally uniformly in Ω for n = 0, 1, 2, and
|φk (x)| ≤ C(1 + |x|p ) (C independent of k and x), then
F (xk , r, q, X, φk (·)) → F (x, r, q, X, φ(·)).
(C2) If X ≤ Y and φ − ψ has a global maximum at x, then
F (x, r, q, X, φ(·)) ≥ F (x, r, q, Y, ψ(·)).
(C3) If r ≤ s, then F (x, r, q, X, φ(·)) ≤ F (x, s, q, X, φ(·)).
(C4) For every constant C ∈ R, F (x, r, q, X, φ(·) + C) = F (x, r, q, X, φ(·)).
Example 4.1. The Bellman/Isaacs equation (1.4) satisfies conditions (C1) – (C4) with Ω = R N
and p = 0 when assumptions (A0) and (A4) hold. For this equation
(
h i
F (x, r, q, X, φ(·)) = inf sup − tr aα,β (x) X + bα,β (x) q + cα,β (x) r + f α,β (x)
α∈A β∈B
Z )
h i
α,β α,β
− φ(x + η (x, z)) − φ(x) − η (x, z) q 1|z|<1 m(dz) .
RM \{0}

Definition 4.2. A locally bounded function u ∈ U SC(Ω) (u ∈ LSC(Ω)) is a viscosity

subsolution (viscosity supersolution) of (1.1) if for every x ∈ Ω and φ ∈ C p2 (Ω) such that x is
a global maximizer (global minimizer) for u − φ,
F (x, u(x), Dφ(x), D 2 φ(x), φ(·)) ≤ 0 (≥ 0).
We say that u is a viscosity solution of (1.1) if u is both a sub- and supersolution of (1.1).
Note that viscosity solutions according to this definition are continuous. Without changing
the solutions, we may change this definition in the following two standard ways:
 INTEGRO-PARTIAL DIFFERENTIAL EQUATIONS 9

Lemma 4.3. (i) If (C4) holds, we may assume that φ(x) = u(x) in Definition 4.2.
(ii) If (C2) holds, we may replace global extremum by global strict extremum in Definition
4.2.
Proof. We only prove (ii) and here we only consider maxima. Assume φ ∈ C p2 (Ω) is such that
u − φ has a global maximum at x ∈ Ω. Pick a non-negative ψ ∈ C 2 (Ω) with compact support
such that ψ|B(x,δ) (y) = |x − y|4 for some 0 < δ < dist(x, ∂Ω). Now u − (φ + ψ) has a global
strict maximum at x, and D(φ+ψ) = Dφ and D 2 (φ+ψ) = D 2 φ at x. Since φ−(φ+ψ) = −ψ
also has a global maximum at x, by (C2) and the above considerations we have
F (x, u(x), D(φ + ψ)(x), D 2 (φ + ψ)(x), (φ + ψ)(·))
≤ F (x, u(x), Dφ(x), D 2 φ(x), φ(·)) (≤ 0),
and the proof is complete. 
The concept of a solution in Definition 4.2 is an extension of the classical solution concept.
Lemma 4.4. (i) If (C2) holds, then a classical subsolution u of (1.1) belonging to C p2 (Ω) is
a viscosity subsolution of (1.1).
(ii) A viscosity subsolution u of (1.1) belonging to C p2 (Ω) is a classical subsolution of (1.1).
Next we introduce an alternative definition of viscosity solutions that is needed for proving
comparison and uniqueness results. For every κ ∈ (0, 1), assume that we have a function
Fκ : Ω × R × RN × SN × SCp (Ω) × C 2 (Ω) → R
satisfying the following list of assumptions for every κ ∈ (0, 1), x, y ∈ Ω, r, s ∈ R, q ∈
RN , X, Y ∈ SN , u, −v ∈ U SCp (Ω), w ∈ SCp (Ω), and φ, φk , ψ, ψk ∈ Cp2 (Ω).
(F0) Fκ (x, r, q, X, φ(·), φ(·)) = F (x, r, q, X, φ(·)).
(F1) The function F in (F0) satisfies (C1).
(F2) If X ≤ Y and both u − v and φ − ψ have global maxima at x, then
Fκ (x, r, q, X, u(·), φ(·)) ≥ Fκ (x, r, q, Y, v(·), ψ(·)).
(F3) The function F in (F0) satisfies (C3).
(F4) For all constants C1 , C2 ∈ R,
Fκ (x, r, q, X, w(·) + C1 , φ(·) + C2 ) = Fκ (x, r, q, X, w(·), φ(·)).
(F5) If ψk → w a.e. in Ω and |ψk (x)| ≤ C(1 + |x|p ) (C independent of k and x), then
Fκ (x, r, q, X, ψk (·), φ(·)) → Fκ (x, r, q, X, u(·), φ(·)).
Remark 4.5. If (F0) – (F4) hold, then (C1) – (C4) hold.
Example 4.6. For the Bellman/Isaacs equation (1.4),
n h i
Fκ (x, r, q, X, u(·), φ(·)) = inf sup −tr aα,β (x) X + bα,β (x) q + cα,β (x) r
α∈A β∈B
o
+ f α,β (x) − Bκα,β (x, q, φ(·)) − B α,β,κ (x, q, u(·)) ,
where
Z h i
Bκα,β (x, q, φ(·)) = φ(x + η α,β (x, z)) − φ(x) − η α,β (x, z) q m(dz),
B(0,κ)\{0}
10 JAKOBSEN AND KARLSEN
Z h i
B α,β,κ (x, q, u(·)) = u(x + η α,β (x, z)) − u(x) − η α,β (x, z) q 1|z|<1 m(dz).
RM \B(0,κ)

If κ < 1 and conditions (A0) and (A4) hold, then (F0) – (F5) hold for (1.4) with Ω = R N
and p = 0.
Lemma 4.7 (Alternative definition). Assume (F0), (F2), (F4), and (F5) hold. u ∈ U SC p (Ω)
(u ∈ LSCp (Ω)) is a viscosity subsolution (viscosity supersolution) of (1.1) if and only if for
every x ∈ Ω and φ ∈ C 2 (Ω) such that x is a global maximizer (global minimizer) for u − φ,
Fκ (x, u(x), Dφ(x), D 2 φ(x), u(·), φ(·)) ≤ 0 (≥ 0) for every κ ∈ (0, 1).

Proof. The proof follows [31], see also [6].

If. Let u − φ have a global maximum at x for some φ ∈ C p2 (Ω). Using (F0), (F2) and the
assumptions of the lemma we have
F (x, u(x), Dφ(x), D 2 φ(x), φ(·)) = Fκ (x, u(x), Dφ(x), D 2 φ(x), φ(·), φ(·))
≤ Fκ (x, u(x), Dφ(x), D 2 φ(x), u(·), φ(·)) ≤ 0.

Only if. Let φ ∈ C 2 (Ω) be such that u − φ has a global maximum at x. By an argument
similar to the one in the proof of Lemma 4.3 with (F4) replacing (C4), we can assume that
(u − φ)(x) = 0. Pick a sequence of Cp2 (Ω) functions {φε }ε such that u ≤ φε ≤ φ and φε → u
a.e. as ε → 0. It follows that u − φε and φε − φ also have global maxima at x. The last
maximum implies that at x, D(φε −φ) = 0 and D 2 (φε −φ) ≤ 0. By (F2), (F0), and Definition
4.2 we have
Fκ (x, u(x), Dφ(x), D 2 φ(x), φε (·), φ(·))
≤ Fκ (x, u(x), Dφε (x), D 2 φε (x), φε (·), φε (·))
= F (x, u(x), Dφε (x), D 2 φε (x), φε (·)) ≤ 0.
Since φε → u a.e., sending ε → 0 in the above inequality and using (F5) yields the “≤”
inequality in the lemma, and the only if part is proved. 

We have now come to our main theorem. It is this theorem that should replace the
maximum principle for semicontinuous functions [12, 13] when proving comparison results for
integro-PDEs.
Theorem 4.8. Let u, −v ∈ U SCp (Ω) satisfy u(x), −v(x) ≤ C(1 + |x|2 ) and solve in the
viscosity solution sense
F (x, u, Du, D 2 u, u(·)) ≤ 0 and G(x, v, Dv, D 2 v, v(·)) ≥ 0,
where F and G satisfy (C1) – (C4). Let φ ∈ C 2 (Ω × Ω) and (x̄, ȳ) ∈ Ω × Ω be such that
(x, y) 7→ u(x) − v(y) − φ(x, y)
has a global maximum at (x̄, ȳ). Furthermore assume that in a neighborhood of (x̄, ȳ), there
are continuous functions g0 : R2N → R, g1 , g2 : RN → SN with g0 (x̄, ȳ) > 0 satisfying
   
2 I −I g1 (x) 0
D φ ≤ g0 (x, y) + .
−I I 0 g2 (y)
 INTEGRO-PARTIAL DIFFERENTIAL EQUATIONS 11

If, in addition, for each κ ∈ (0, 1) there exist F κ and Gκ satisfying (F0) – (F5), then for
any γ ∈ (0, 21 ) there are two matrices X, Y ∈ SN satisfying
       
g0 (x̄, ȳ) I 0 X 0 g1 (x̄) 0 g0 (x̄, ȳ) I −I
(4.1) − ≤ − ≤ .
γ 0 I 0 −Y 0 g2 (ȳ) 1 − 2γ −I I
such that
(4.2) Fκ (x̄, u(x̄), Dx φ(x̄, ȳ), X, u(·), φ(·, ȳ)) ≤ 0,
(4.3) Gκ (x̄, v(ȳ), −Dy φ(x̄, ȳ), Y, v(·), −φ(x̄, ·)) ≥ 0.
The proof of Theorem 4.8 is given in Section 7. We underline that the key point in Theorem
4.8 is the validity of the inequalities (4.2) and (4.3). The proof of Theorem 4.8 shows that
2,+ 2,+
(Dx φ(x̄, ȳ), X) ∈ J u(x̄) and (−Dy φ(x̄, ȳ), Y ) ∈ J v(ȳ). This information alone would in
the pure PDE case, under certain (semi)continuity assumptions on the equations, imply that
the viscosity solution inequalities hold. In the non-local case, the situation is more delicate
and we refer to Section 2 for a discussion of this point.
The technical assumption u(x), −v(x) ≤ C(1 + |x| 2 ) is an artifact of the method of proof,
and it does not seem so easy to remove. However, in applications this condition does not
create any difficulties. The assumptions on the test-function is satisfied in most practical
cases. For test-functions like
1
φ(x, y) = |x − y|q , δ > 0, q > 0,
δ
the assumptions hold for all (x, y) ∈ R 2N when q ≥ 2 and for all (x, y) ∈ {x 6= y} otherwise.
Finally, we remark that it is possible to have a result without restrictions on D 2 φ. Such
a result (Lemma 7.4) is actually used to prove Theorem 4.8. But this result is indirect in
the sense that it is not the function (x, y) 7→ u(x) − v(y) − φ(x, y) that is considered directly
but rather its “sup-convoluted” version (x, y) 7→ u ε (x) − vε (y) − φ(x, y). In fact, this was the
original approach to uniqueness of viscosity solutions for second order PDEs, cf. Jensen, Ishii,
Lions [27, 22, 23].

5. Comparison for unbounded Ω

In this section we use Theorem 4.8 to prove a general comparison result for non-local
equations of the form (1.1). We need two additional assumptions on the equation, and we
state them for the Fκ function.
For every κ ∈ (0, 1), x, y ∈ Ω, r ∈ R, p ∈ RN , X, Y ∈ SN , u, −v ∈ U SCp (Ω), and φ ∈ C 2 (Ω)
the following statements hold:
(F6) There is a λ > 0 such that if s ≤ r, then
Fκ (x, r, p, X, u(·), φ(·)) − Fκ (x, s, p, X, u(·), φ(·)) ≥ λ(r − s).
(F7) For any δ, ε > 0, define
1
φ(x, y) = |x − y|2 − ε(|x|2 + |y|2 ).
δ
If u(x), −v(x) ≤ C(1 + |x|2 ) in Ω, and (x̄, ȳ) ∈ Ω × Ω is such that
(x, y) 7→ u(x) − v(y) − φ(x, y)
has a global maximum at (x̄, ȳ), then for any κ > 0 there are numbers m κ,δ,ε
12 JAKOBSEN AND KARLSEN

satisfying lim lim lim mκ,δ,ε = 0 and a modulus ω such that

ε→0δ→0κ→0
   
1 1
Fκ ȳ, r, (x̄ − ȳ) − εȳ, Y, v(·), −φ(x̄, ·) − Fκ x̄, r, (x̄ − ȳ) + εx̄, X, u(·), φ(·, ȳ)
δ δ
 
1
≤ ω |x̄ − ȳ| + |x̄ − ȳ|2 + ε(1 + |x̄|2 + |ȳ|2 ) + mκ,δ,ε,
δ
for every X, Y satisfying
     
X 0 4 I −I I 0
(5.1) ≤ + 2ε .
0 −Y δ −I I 0 I
Condition (F7) is a version for non-local equations (in an unbounded domain) of the stan-
dard condition (3.14) in [13]. The inequality (5.1) corresponds to the second inequality in
(4.1) with γ = 1/4.
Example 5.1. If (A0) – (A5) are satisfied, then (F6) and (F7) are satisfied for the Bell-
man/Isaacs equation (1.4) when Fκ is defined as in Example 4.6. We will show this in the
next section.
Theorem 5.2. Assume that for every κ ∈ (0, 1) there exists F κ satisfying (F0) – (F7), that
u, −v ∈ U SCp (Ω) are bounded from above, and that for every z ∈ ∂Ω, x ∈ Ω,
(5.2) u(z) ≥ u(x) − ω0 (|x − z|) and v(z) ≤ v(x) + ω0 (|x − z|),
where ω0 is a modulus.
If u and v are respectively sub- and supersolutions of (1.1), and u ≤ v on ∂Ω, then u ≤ v
in Ω.
Remark 5.3. The above result implies uniqueness of bounded viscosity solutions on a possibly
unbounded domain Ω. The “ω0 condition” means that the semicontinuous viscosity sub- and
supersolutions u and v are uniformly semicontinuous up to the boundary. Any viscosity
solution satisfying this condition attains its boundary values uniformly continuously.
Proof. Define Ψ(x, y) = u(x) − v(y) − φ(x, y), where φ is defined in (F7). By standard
arguments there is a point (x̄, ȳ) ∈ Ω × Ω (depending on δ and ε) such that Ψ attains its
supremum over Ω × Ω here. Define σ := supΩ×Ω Ψ = Ψ(x̄, ȳ). Note that for any x ∈ Ω, (u −
v)(x)−2ε|x|2 ≤ σ. So, obviously we are done if we can prove that lim ε→0 limδ→0 limκ→0 σ ≤ 0.
To prove this, we will first derive a positive upper bound for σ. We may assume σ > 0
since otherwise any positive upper bound trivially holds. Since u and −v are bounded from
above, we have the following bounds
1
(5.3) ε(|x̄|2 + |ȳ|2 ) ≤ ω1 (ε) and |x̄ − ȳ|2 ≤ ω2 (δ),
δ
where ωi , i = 1, 2, are moduli not depending on any of the parameters κ, ε, δ. These are
standard results, see, e.g., [13, Lemma 3.1] for the proofs. Now, either (i) (x̄, ȳ) ∈ ∂(Ω × Ω),
or (ii) (x̄, ȳ) ∈ Ω × Ω. In case (i), (5.2) and u ≤ v on ∂Ω implies that u(x̄) − v(ȳ) ≤ ω 0 (|x̄ − ȳ|),
and hence
1
σ ≤ ω0 (|x̄ − ȳ|) − |x̄ − ȳ|2 − ε(|x̄|2 + |ȳ|2 ) ≤ ω0 (|x̄ − ȳ|) =: I.
δ
In case (ii), we apply Theorem 4.8 to find matrices X, Y ∈ S N , satisfying (4.1), such that
(4.2) and (4.3) hold. Since σ > 0 implies that u(x̄) ≤ v(ȳ), subtracting the above inequalities
 INTEGRO-PARTIAL DIFFERENTIAL EQUATIONS 13

and using (F6) and (F7) yield

λ(u(x̄) − v(ȳ)) ≤ Fκ (ȳ, v(ȳ), −Dy φ(x̄, ȳ), Y, vε (·), −φ(x̄, ·))
− Fκ (x̄, v(ȳ), Dx φ(x̄, ȳ), X, uε (·), φ(·, ȳ))
 
1 2 2 2
≤ ω (|x̄ − ȳ| + |x̄ − ȳ| ) + ε(1 + |x̄| + |ȳ| ) + mκ,δ,ε =: II.
δ
So we have σ ≤ u(x̄) − v(ȳ) ≤ II/λ. To complete the proof, we combine cases (i) and (ii) to
obtain the following upper bound on σ,
σ ≤ max(I, II/λ),
where by (5.3) I and II only depends on κ, ε, δ, and lim ε→0 limδ→0 limκ→0 σ ≤ 0. 
Remark 5.4. The case when the solutions have linear growth at infinity seems to be more
difficult, and we do not know the optimal conditions for having a comparison result in this
case. However, there are two special cases where we may have a comparison result:
• If Fκ , F are uniformly continuous in all variables, then a comparison result can be
obtained by adapting the techniques of Ishii in [22, Theorem 7.1].
• If λ is sufficiently large, then we have a comparison result due to “cancellation effects”
in the proof.
If λ is big enough, then we can handle arbitrary polynomial growth in the solutions by
slightly changing assumption (F7). For parabolic problems we are always in this situation
since we can have λ arbitrary large after an exponential in time scaling of the solution.

6. Proof of comparison for the Bellman/Isaacs equation, Theorem 3.1

As an application of the general results presented in the previous sections, we prove in
this section a comparison result for semicontinuous sub- and supersolutions of the elliptic
Bellman/Isaacs equation (Theorem 3.1).
In view of Examples 4.1 and 4.6 and the abstract comparison result in the previous section
(Theorem 5.2), Theorem 3.1 follows if we can verify that (F0) – (F7) hold for the functions
Fκ defined in Example 4.6. The only difficult part is to show that (F7) holds, so we restrict
our discussion to this condition. In the pure PDE case this is proved by Ishii [22]. Although
not stated as such, in the integro-PDE case this is essentially proved in [30] for uniformly
continuous u, v (see also [34, 8, 10]). To give the reader some ideas how this is done (for
semicontinuous u, v), we consider briefly the integro operator of the Bellman/Isaacs equation
(1.4). According to Example 4.6, it can be decomposed into B κα,β and B α,β,κ , and thanks to
(A0) and (A4), the Bκα,β term goes to zero as κ → 0. Let us now consider the other term.
For (F7) to be satisfied, it is necessary that
   
α,β,κ 1 α,β,κ 1
−B ȳ, (x̄ − ȳ) − εȳ, v(·) + B x̄, (x̄ − ȳ) + εx̄, u(·)
δ δ
(6.1)  
1 2 2 2
≤ ω |x̄ − ȳ| + |x̄ − ȳ| + ε(1 + |x̄| + |ȳ| ) + mκ,δ,ε .
δ

Let us write B α,β,κ = B1α,β,κ +B2α,β , where B1α,β,κ is the part where z is integrated over the set
κ ≤ |z| ≤ 1 and B2α,β is the part where z is integrated over the set |z| ≥ 1. The part of (6.1)
14 JAKOBSEN AND KARLSEN

corresponding to B1α,β,κ can be handled as follows. If we let ψ(x, y) = u(x) − v(y) − φ(x, y),
then a simple calculation shows that the integrand of this part equals

ψ(x̄ + η α,β (x̄, z), ȳ + η α,β (ȳ, z)) − ψ(x̄, ȳ)

1
+ |η α,β (x̄, z) − η α,β (ȳ, z)|2 + ε(|η α,β (x̄, z)|2 + |η α,β (ȳ, z)|2 ).
δ
Since ψ has a global maximum at (x̄, ȳ) the two first terms are non-positive, so by (A0), (A3),
and (A4), we get
   
α,β,κ 1 α,β,κ 1
−B1 ȳ, (x̄ − ȳ) − εȳ, v(·) + B1 x̄, (x̄ − ȳ) + εx̄, u(·)
δ δ
 
1 2 2 2
≤C |x̄ − ȳ| + ε(|x̄| + |ȳ| ) ,
δ

for some constant C. To handle the part of (6.1) corresponding to B 2α,β,κ , we introduce

Mδ,ε := sup ψ(x, y) and M := sup(u − v),

x,y∈R x∈R

and remark that lim lim Mδ,ε = M . Then it follows that

ε→0δ→0
   
1 1
− B2α,β,κ ȳ, (x̄ − ȳ) − εȳ, v(·) + B2α,β,κ x̄, (x̄ − ȳ) + εx̄, u(·)
δ δ
Z h   i
= u(x̄ + η α,β (x̄, z)) − u(x̄) − v(ȳ + η α,β (ȳ, z)) − v(ȳ) m(dz)
|z|≥1
Z  
α,β α,β
= u(x̄ + η (x̄, z)) − v(ȳ + η (ȳ, z)) −Mδ,ε + φ(x̄, ȳ) m(dz).
|z|≥1 | {z }
g(x̄, ȳ, z)

The last equality follows from the definition of M δ,ε since (x̄, ȳ) is a maximum point of ψ by
assumption. As we have seen before (see (5.3)), |x̄ − ȳ| → 0 as δ → 0 and |x̄|, |ȳ| are bounded
as long as ε > 0 is kept fixed, and since g(x̄, ȳ, z) is upper semicontinuous in x and y this
leads to
lim sup g(x̄, ȳ, z) ≤ M.
δ→0

An other application of (5.3) shows that lim lim φ(x̄, ȳ) = 0. So sending first δ → 0 (taking
ε→0δ→0
limit superior) and then ε → 0, we see that the above integrand and hence also the integral
(by Lebesgue dominated convergence theorem), is upper bounded by 0. We can conclude that
there is an upper bound mδ,ε of the difference in the B2α,β,κ terms such that lim lim mδ,ε = 0,
ε→0δ→0
and the proof of (6.1) is complete.

Remark 6.1. The trick of dividing B α,β,κ into two terms [10] Rallows one to consider more
general Lévy measures than in [30]. In [30] it is required that RM \{0} ρ(z)2 m(dz) is finite,
while we assume the weaker condition (A0).
 INTEGRO-PARTIAL DIFFERENTIAL EQUATIONS 15

7. Proof of the main result, Theorem 4.8

The outline of the proof of Theorem 4.8 is as follows. First we regularize the sub- and
supersolutions using the ε-sup and ε-inf convolutions, thereby yielding approximate sub- and
supersolutions of the original equations that are twice differentiable a.e. Using the classical
maximum principle, we derive for these approximate sub- and supersolutions an analogous
result to Theorem 4.8. In this result (Lemma 7.4) the lower bounds in the matrix inequality
corresponding to (4.1) depends on the regularization parameter ε. A transformation of these
matrices (Lemma 7.7) leads to new matrices satisfying (4.1) independently of ε. Furthermore,
the viscosity solution inequalities for the approximate sub- and supersolutions are still satisfied
with these new matrices. We can then go to the limit along a subsequence of ε → 0 and obtain
Theorem 4.8.
The first part of this approach corresponds to the original approach of Jensen [27], giving
the first uniqueness results for viscosity solutions of second order PDEs. Actually, we follow
the more refined approach of Ishii [22] and Ishii and Lions [23]. In the second part, the key
ingredient is a matrix lemma of Crandall [11]. We remark that our approach deviates from
the by now standard approach based on the maximum principle for semicontinuous functions
[12, 13]. As explained in Section 2, it appears that a “local” approach based on the maximum
principle for semicontinuous functions is not straightforward to implement for the non-local
equation (1.1).
We start by defining the sup and inf convolutions and stating some of their properties.
Definition 7.1. Let f ∈ U SC(Ω) satisfy f (x) ≤ C(1 + |x| 2 ) in Ω and 0 < ε < (2C)−1 . The
sup convolution f ε is defined as
 
|x − y|2
f ε (x) = sup u(y) − .
y∈Ω ε
Let f ∈ LSC(Ω) satisfy f (x) ≥ −C(1 + |x| 2 ) in Ω and 0 < ε < (2C)−1 . The inf convolution
fε is defined as  
|x − y|2
fε (x) = inf u(y) + .
y∈Ω ε
Lemma 7.2. Let f ∈ U SC(Ω) satisfy f (x) ≤ C(1 + |x| 2 ) in Ω and 0 < ε < (2C)−1 .
(i) f ε (x) ≤ 2C(1 + |x|2 ) and f ε (x) + 1ε |x|2 is convex and locally Lipschitz in Ω.
(ii) f ≤ f ε ≤ f ε̄ for 0 < ε ≤ ε̄ < (2C)−1 and f ε → f pointwise as ε → 0.

1/2
(iii) Let ε < (4C)−1 and define C(x) := 4C(1 + |x|2 ) − 2u(x) . If x ∈ Ω is such that
1/2 1/2
dist(x, ∂Ω) > ε C(x), then there exists x̄ ∈ Ω such that |x − x̄| ≤ ε C(x) and
1
uε (x) = u(x̄) − |x − x̄|2 .
ε
ε
Since fε = −(−f ) , we immediately get the corresponding properties for the inf-convolution.
We refer to [4, 15] for proofs of results like those in Lemma 7.2. Now if f is a function satisfying
the assumptions of Lemma 7.2, we define
n o
f 1/2
Ωε = x ∈ Ω : dist(x, ∂Ω) > ε C(x) ,
where C(x) is defined in Lemma 7.2. Moreover, let τ h denote the shift operator defined by
(τh φ)(x) = φ(x + h)
for any function φ and x, x + h in the domain of definition of φ.
16 JAKOBSEN AND KARLSEN

Lemma 7.3. Assume (C3) holds, u, −v ∈ U SC p (Ω) satisfy u(x), −v(x) ≤ C(1 + |x|2 ), and
0 < ε < (4C)−1 .
(a) If u is a viscosity subsolution of (1.1), then u ε solves
Fε (x, uε (x), Duε (x), D 2 uε (x), uε (·)) ≤ 0 in Ωuε ,
in the viscosity solution sense, where
Fε (x, r, p, X, φ(·)) = inf F (y, r, p, X, τx−y φ(·)).
|x−y|≤C(x)ε1/2

(b) If v is a viscosity supersolution of (1.1), then v ε solves

F ε (x, vε (x), Dvε (x), D 2 vε (x), vε (·)) ≥ 0 in Ω−v
ε ,
in the viscosity solution sense, where
F ε (x, r, p, X, φ(·)) = sup F (y, r, p, X, τx−y φ(·)).
|x−y|≤C(x)ε1/2

Proof. We only prove (a), the proof of (b) is similar. Let φ ∈ C p2 (Ω) and x̄ ∈ Ωuε be such
that uε − φ has a global maximum at x̄. According to Lemma 7.2 (iii) there is a ȳ ∈ Ω such
that |x̄ − ȳ| ≤ C(x̄)ε1/2 and uε (x̄) = u(ȳ) − 1ε |x̄ − ȳ|2 . Now it is not so difficult to see that
y 7→ (u − τx̄−ȳ φ) (y) has a global maximum at ȳ (cf. [22, Proof of Proposition 4.2]). Since u
is a viscosity subsolution of (1.1),
F (ȳ, u(ȳ), D(τx̄−ȳ φ)(ȳ), D 2 (τx̄−ȳ φ)(ȳ), (τx̄−ȳ φ)(·)) ≤ 0.
Since |x̄ − ȳ| ≤ C(x)ε1/2 , D n (τx̄−ȳ φ)(ȳ) = D n φ(x̄) for n = 1, 2, and uε (x̄) ≤ u(ȳ), it follows
using (C3) and the above inequality that
Fε (x̄, uε (x̄), Dφ(x̄), D 2 φ(x̄), φ(·)) ≤ 0,
and the proof is complete 
Now we have come to one of the main technical results in this paper. It is a version for
integro-PDEs of Proposition 5.1 in Ishii [22] (see also Proposition II.3 in Ishii and Lions [23]).
Lemma 7.4. Let u, −v ∈ U SCp (Ω) satisfy u(x), −v(x) ≤ C(1 + |x|2 ), and solve in the
viscosity solution sense
F (x, u(x), Du(x), D 2 u(x), u(·)) ≤ 0 and G(x, v(x), Dv(x), D 2 v(x), v(·)) ≥ 0,
where F, G satisfy (C1) – (C4). For 0 < ε < (4C) −1 , let φ ∈ Cp2 (Ω×Ω) and (x̄, ȳ) ∈ Ωuε ×Ω−v
ε
be such that
(x, y) 7→ uε (x) − vε (y) − φ(x, y)
has a global maximum over Ωuε × Ω−v ε at (x̄, ȳ). Then there exist two matrices X, Y ∈ S N
satisfying
   
2 I 0 X 0
(7.1) − ≤ ≤ D 2 φ(x̄, ȳ),
ε 0 I 0 −Y
such that
(7.2) Fε (x̄, uε (x̄), Dx φ(x̄, ȳ), X, φ(·, ȳ)) ≤ 0,
(7.3) Gε (ȳ, vε (ȳ), −Dy φ(x̄, ȳ), Y, −φ(x̄, ·)) ≥ 0.
Remark 7.5. Compared with Ishii [22, Proposition 5.1], the main feature of Lemma 7.4 is the
inclusion of the penalization function φ(·, ·) in the non-local slots in (7.2) and (7.3).
 INTEGRO-PARTIAL DIFFERENTIAL EQUATIONS 17

Remark 7.6. The condition u(x), −v(x) ≤ C(1 + |x| 2 ) in Lemma 7.4 is necessary for uε and
vε to be well-defined according to Definition 7.1.

Proof. 1. Let w(x, y) = uε (x) − vε (y). By Lemma 7.2 w is locally Lipschitz continuous and
semi-convex in Ω × Ω. By Alexandroff’s theorem, w is twice differentiable a.e. in Ω × Ω
(cf. [13, 15]).
2. By the assumptions, w − φ has a global maximum over Ω uε × Ω−v ε at (x̄, ȳ). By (C4),
we may assume that w = φ at (x̄, ȳ) by adding a constant to φ if necessary. Furthermore,
(x, y) 7→ w(x, y) − φ(x, y) − δ|(x, y) − (x̄, ȳ)| 4 has a strict maximum over Ωuε × Ω−v ε at (x̄, ȳ)
for every δ > 0, and this maximum takes the value 0.
3. The crucial step in this proof is the application of Jensen’s lemma, see Lemmas 3.10 and
3.15 in Jensen [27] or Lemma 5.3 in Ishii [22]. Pick a r > 0 such that B((x̄, ȳ), r) ⊂ Ω uε × Ω−v ε ,
then by 1 and 2 we may apply Jensen’s lemma to w − φ − δ|(x, y) − (x̄, ȳ)| 4 on B((x̄, ȳ), r).
By this Lemma there are sequences {(x k , yk )}k ⊂ B((x̄, ȳ), r) and {(pk , qk )}k ⊂ RN × RN
such that (i) (xk , yk ) → (x̄, ȳ) and (pk , qk ) → (0, 0) as k → ∞, (ii) w is twice differentiable at
(xk , yk ), and (iii) the function

(x, y) 7→ w(x, y) − φ(x, y) − δ|(x, y) − (x̄, ȳ)| 4 − (pk , qk ) · (x, y)

attains its maximum over B((x̄, ȳ), r) at (x k , yk ). Note that for notational reasons, we suppress
the dependence on δ in xk , yk , pk , qk .
4. Now, let φ̄k,δ (x, y) = φ(x, y) + δ|(x, y) − (x̄, ȳ)|4 + (pk , qk ) · (x, y) + Ck,δ for some constant
Ck,δ , and note that by 3, w − φ̄k,δ attains its maximum over B((x̄, ȳ), r) at (x k , yk ). Hence
the differentiability of w implies that at (x k , yk ), Dw = D φ̄k,δ and D 2 w ≤ D 2 φ̄k,δ . Finally,
choose Ck,δ such that (w − φ̄k,δ )(xk , yk ) = 0.
5. Pick a non-negative function θ ∈ C ∞ (Ωuε × Ω−v ε ) which is 1 in B((x̄, ȳ), r/2) and 0
outside of B((x̄, ȳ), r). Now define

φk,δ = θ φ̄k,δ + (1 − θ) φ.

Obviously φk,δ ∈ Cp2 (Ωuε × Ω−v ε ), and we claim that w − φk,δ has a global maximum at (xk , yk ).
This follows since by 2, w ≤ φ in Ωuε × Ω−v ε , and by 4, w ≤ φ̄k,δ in B((x̄, ȳ), r) and w = φ̄k,δ
at (xk , yk ).
6. There exists a function ψk,δ ∈ Cp2 (Ωuε × Ω−v n n
ε ) such that D ψk,δ (xk , yk ) = D w(xk , yk )
for n = 0, 1, 2 and w ≤ ψk,δ ≤ φk,δ in Ωuε × Ω−v ε . In particular, ψk,δ − φk,δ (also) attains its
maximum over Ωuε × Ω−v ε at (x ,
k ky ).
To prove the above claim we consider separately the following cases: (i) D 2 w = D 2 φk,δ
at (xk , yk ) and (ii) D 2 w < D 2 φk,δ at (xk , yk ) (note that trivially D 2 w ≤ D 2 φk,δ at (xk , yk )).
In case (i) we simply set ψk,δ = φk,δ . In case (ii) we pick a φ̄ ∈ C 2 (Ωuε × Ω−v ε ) such that
D n φ̄ = D n w at (xk , yk ) for n = 0, 1, 2, and w − φ̄ ≤ 0 in Ωuε × Ω−v ε . This can be done by a
construction of Evans, see e.g. [15, Proposition V.4.1]. It follows that at (x k , yk ), φ̄ = φk,δ ,
D φ̄ = Dφk,δ , and D 2 φ̄ < D 2 φk,δ . This means that we can find a δ̄ > 0 such that φ̄ < φk,δ in
¯ Now we define ψk,δ in the following way:
the ball B((xk , yk ), δ).

ψk,δ = θ φ̄ + (1 − θ) φk,δ ,

where θ ∈ C ∞ (Ωuε ×Ω−v

ε ) is non-negative, 1 in B((x k , yk ), δ̄/2), and 0 outside of B((x k , yk ), δ̄).
This function θ is not to be confused with the θ in 5.
18 JAKOBSEN AND KARLSEN

7. By 6, (w − ψk,δ )(x, yk ) has a maximum over Ωuε at xk , and (w − ψk,δ )(xk , y) has a
minimum over Ω−v
ε at yk . Hence Lemma 7.3 yields

Fε (xk , uε (xk ), Dx ψk,δ (xk , yk ), Dx2 ψk,δ (xk , yk ), ψk,δ (·, yk )) ≤ 0,

Gε (yk , vε (yk ), −Dy ψk,δ (xk , yk ), −Dy2 ψk,δ (xk , yk ), −ψk,δ (xk , ·)) ≥ 0.
By the properties of ψk,δ , (C2), and since ψk,δ − φk,δ has its global maximum at (xk , yk ), we
get
Fε (xk , uε (xk ), Duε (xk ), Xk,δ , φk,δ (·, yk )) ≤ 0,
Gε (yk , vε (yk ), Dvε (yk ), Yk,δ , −φk,δ (xk , ·)) ≥ 0,
where Xk,δ = D 2 uε (xk ) and Yk,δ = D 2 vε (yk ).
8. Since w − φk,δ has a maximum at (xk , yk ) and by the semi-convexity of w, see Lemma
7.2 (i), the following inequality holds
   
2 I 0 Xk,δ 0
− ≤ ≤ D 2 φk,δ (xk , yk ).
ε 0 I 0 −Yk,δ
Furthermore, if 0 < δ < 1 and ε fixed then this inequality implies that −CI ≤ X k,δ , Yk,δ ≤ CI
for some constant C > 0. The set of such matrixes is compact by Lemma 5.3 in Ishii [22],
so we may pick converging subsequences, also denoted by {X k,δ , Yk,δ }k , converging to some
Xδ , Yδ ∈ SN . By the above inequality and since D 2 φk,δ (xk , yk ) → D 2 φ(x̄, ȳ) as k → ∞, we
see that the limits Xδ , Yδ satisfy (7.1).
9. The next step of the proof is to send k → ∞ (along the subsequence in 8) in the
inequalities at the end of 7, and conclude by continuity of all arguments and (C1) that
Fε (x̄, uε (x̄), Dx φ(x̄, ȳ), Xδ , φ(·, ȳ) + δθ(·, ȳ)|(·, ȳ) − (x̄, ȳ)| 4 ) ≤ 0,
Gε (ȳ, vε (ȳ), −Dy φ(x̄, ȳ), Yδ , −φ(x̄, ·) − δθ(x̄, ·)|(x̄, ·) − (x̄, ȳ)| 4 ) ≥ 0.
Let us verify the assumptions of (C1). First note that |φ k,δ (x, y)| ≤ C(1 + |x|p + |y|p ) with
C independent of δ, k. This bound follows from the definition of φ k,δ (see 4 and 5) since
φ ∈ Cp2 . Then we claim that D n φk,δ (·, ȳ) → D n (φ(·, ȳ) + δ|(·, ȳ) − (x̄, ȳ)|4 ) locally uniformly
for n = 0, 1, 2 as k → ∞ (and similarly for φ k,δ (x̄, ·)). By its definition (see 5) it is enough to
check this for φ̄k,δ in B((x̄, ȳ), r). But by the definition of φ̄k,δ (see 4) this easily follows since
by 6, pk , qk → 0, and by 2, Ck,δ = sup(w − φ̄k,δ ) → sup(w − φ − δ|(x, y) − (x̄, ȳ)|4 ) = 0.
10. The finial step is to send δ → 0. Because X δ , Yδ satisfy (7.1), we have compactness as
in 8, so we pick a subsequence δ → 0 such that the matrices converge to some X, Y ∈ S N .
Of course X, Y still satisfy (7.1). Furthermore, by continuity of all arguments and (C1) we
conclude that the inequalities in 9 become (7.2) and (7.3) as δ → 0 along this subsequence. 

The next result is a matrix lemma due to Crandall [11].

Lemma 7.7. Let X, Y ∈ SN satisfy
   
X 0 I −I
(7.4) ≤ .
0 −Y −I I
Then for γ ∈ (0, 12 ), (I − γX) and (I + γY ) are invertible, and if
X γ = X(I − γX)−1 and Yγ = Y (I + γY )−1
 INTEGRO-PARTIAL DIFFERENTIAL EQUATIONS 19

then
(7.5) X ≤ X γ ≤ Yγ ≤ Y
and
     
1 I 0 Xγ 0 1 I −I
(7.6) − ≤ ≤ .
γ 0 I 0 −Yγ 1 − 2γ −I I
Using Lemmas 7.4 and 7.7, we now prove a version of Theorem 4.8 where the F/G-
formulation is used instead of the F κ /Gκ -formulation. Theorem 4.8 is an easy consequence
of this result (see below).
Lemma 7.8. Let u, −v ∈ U SCp (Ω) satisfy u(x), −v(x) ≤ C(1+|x|2 ) and solve in the viscosity
solution sense
F (x, u, Du, D 2 u, u(·)) ≤ 0 and G(x, v, Dv, D 2 v, v(·)) ≥ 0,
where F, G satisfies (C1) – (C4). Let φ ∈ C p2 (Ω × Ω) and (x̄, ȳ) ∈ Ω × Ω be such that
(x, y) 7→ u(x) − v(y) − φ(x, y)
has a global strict maximum at (x̄, ȳ). Furthermore assume that in a neighborhood of (x̄, ȳ),
there are continuous functions g0 : R2N → R, g1 , g2 : RN → SN with g0 (x̄, ȳ) > 0 satisfying
   
2 I −I g1 (x) 0
D φ ≤ g0 (x, y) + .
−I I 0 g2 (y)
Then for each γ ∈ (0, 12 ) there exist matrices X, Y ∈ SN satisfying
       
g0 (x̄, ȳ) I 0 X 0 g1 (x̄) 0 g0 (x̄, ȳ) I −I
(7.7) − ≤ − ≤ .
γ 0 I 0 −Y 0 g2 (ȳ) 1 − 2γ −I I
such that
(7.8) F (x̄, u(x̄), Dx φ(x̄, ȳ), X, φ(·, ȳ)) ≤ 0,
(7.9) G(x̄, v(ȳ), −Dy φ(x̄, ȳ), Y, −φ(x̄, ·)) ≥ 0.
Remark 7.9. Compared with Crandall [11, Theorem 1], the main feature in Lemma 7.8
is the inclusion of the inequalities (7.8) and (7.9). In the pure PDE case, under certain
(semi)continuity assumptions on the equation (1.1), the corresponding inequalities come for
free. We refer to Section 2 for a discussion of this point.
Proof. For all sufficiently small ε > 0, (x, y) 7→ u ε (x) − vε (y) − φ(x, y) has a global maximum
at some point (xε , yε ) ∈ Ω × Ω, and as ε → 0, (xε , yε ) → (x̄, ȳ), uε (xε ) → u(x̄), vε (yε ) → v(ȳ).
Moreover, we may find a ε0 > 0 and a r > 0 such that for all ε < ε0 , (i) (xε , yε ) ∈ B((x̄, ȳ), r),
(ii) B((x̄, ȳ), r) ⊂ Ωuε × Ω−v
ε , and (iii) g0 > 0 in B((x̄, ȳ), r).
By Lemma 7.4 there exist two matrices X, Y ∈ S N satisfying
 
2 X 0
− I≤ ≤ D 2 φ(xε , yε ),
ε 0 −Y
and furthermore
Fε (xε , uε (xε ), Dx φ(xε , yε ), X, φ(·, yε )) ≤ 0,
Gε (yε , vε (yε ), −Dy φ(xε , yε ), Y, −φ(xε , ·)) ≥ 0.
20 JAKOBSEN AND KARLSEN

By the assumptions, we may rewrite the left hand side of the above matrix inequality as
follows,
   
X̃ 0 I −I
≤ ,
0 −Ỹ −I I
where
1 1
X̃ = (X − g1 (xε )) and Ỹ = (Y + g2 (yε )) .
g0 (xε , yε ) g0 (xε , yε )
These two matrices satisfies the assumptions of Lemma 7.7, so we can conclude that inequal-
ities corresponding to (7.5) and (7.6) hold. Now define
X̄ = g0 (xε , yε )X̃ γ + g1 (xε ) and Ȳ = g0 (xε , yε )Ỹγ − g2 (yε ).
The conclusions of Lemma 7.7 can then be written as follows,
X ≤ X̄, Ȳ ≤ Y,
and
       
g0 (xε , yε ) I 0 X̄ 0 g (x ) 0 g0 (xε , yε ) I −I
(7.10) − ≤ − 1 ε ≤ .
γ 0 I 0 −Ȳ 0 g2 (yε ) 1 − 2γ −I I
By degenerate ellipticity (C2),
(7.11) Fε (xε , uε (xε ), Dx φ(xε , yε ), X̄, φ(·, yε )) ≤ 0,
(7.12) Gε (yε , vε (yε ), −Dy φ(xε , yε ), Ȳ , −φ(xε , ·)) ≥ 0.
By continuity of g0 , g1 , g2 on B((x̄, ȳ), r), we see from (7.10) that { X̄}ε>0 and {Ȳ }ε>0 are
compact in S(RN ). Hence we may pick subsequences converging as ε → 0 to limit matrices
(still) called X̄ and Ȳ . Moreover, sending ε → 0 in (7.10) along such a subsequence gives the
matrix inequality (7.7). Passing to the limit ε → 0 along the same subsequence in (7.11) and
(7.12) we obtain (7.8) and (7.9) using (C1) and continuity. The proof is complete. 
We are will now prove Theorem 4.8.
Proof of Theorem 4.8. After an application of Lemma 4.8, this proof is similar to the proof of
Lemma 4.7. First note that we may assume that the maximum is strict and that the maximal
value is 0. Then pick a sequence of Cp2 (Ω × Ω) functions {φε }ε>0 such that u(x) − v(y) ≤
φε (x, y) ≤ φ(x, y) and φε (x, y) → u(x) − v(y) (pointwise) in Ω × Ω. Note that φ ε − φ has
a global maximum at (x̄, ȳ), and hence D(φ ε − φ)(x̄, ȳ) = 0 and D 2 (φε − φ)(x̄, ȳ) ≤ 0. In
particular, it follows that (x, y) 7→ u(x) − v(y) − φ ε (x, y) satisfies the assumptions in Lemma
7.8, so we have matrices X, Y ∈ SN satisfying (7.7) (which equals (4.1)) and
F (x̄, u(x̄), Dx φε (x̄, ȳ), X, φε (·, ȳ)) ≤ 0,
G(ȳ, v(ȳ), −Dy φε (x̄, ȳ), Y, −φε (x̄, ·)) ≥ 0.
Applying (F0) and then (F2) to the above inequalities for F and G yield
Fκ (x̄, u(x̄), Dx φ(x̄, ȳ), X, φε (·, ȳ), φ(·, ȳ)) ≤ 0,
Gκ (ȳ, v(ȳ), −Dy φ(x̄, ȳ), Y, −φε (x̄, ·), −φ(x̄, ·)) ≥ 0.
Using (F5), (F4), and sending ε → 0, yield (4.2) and (4.3). The proof is complete. 
 INTEGRO-PARTIAL DIFFERENTIAL EQUATIONS 21

References
[1] O. Alvarez and A. Tourin, Viscosity solutions of nonlinear integro-differential equations, Ann. Inst. H.
Poincaré Anal. Non Linéaire, 13(3):293–317, 1996.
[2] A. L. Amadori. The obstacle problem for nonlinear integro-differential operators arising in option pricing.
Quaderno IAC Q21-000, 2000.
[3] A. L. Amadori. Nonlinear integro-differential evolution problems arising in option pricing: a viscosity
solutions approach. Journal of Differential and Integral Equations. To appear.
[4] M. Bardi and I. Capuzzo-Dolcetta. Optimal Control and Viscosity Solutions of Hamilton-Jacobi-Bellman
Equations. Birkhäuser, Boston 1997.
[5] G. Barles, Solutions de viscosité des équations de Hamilton-Jacobi, Springer, Paris, 1994.
[6] G. Barles, R. Buckdahn, and E. Pardoux. Backward stochastic differential equations and integral-partial
differential equations. Stochastics Stochastics Rep., 60(1-2):57–83, 1997.
[7] G. Barles and E. R. Jakobsen. On the convergence rate of approximation schemes for Hamilton-Jacobi-
Bellman equations. M2AN Math. Model. Numer. Anal., 36(1):33–54, 2002.
[8] F. E. Benth, K. H. Karlsen, and K. Reikvam. Optimal portfolio management rules in a non-Gaussian
market with durability and intertemporal substitution. Finance Stoch., 5(4):447–467, 2001.
[9] F. E. Benth, K. H. Karlsen, and K. Reikvam. Optimal portfolio selection with consumption and nonlinear
integro-differential equations with gradient constraint: a viscosity solution approach. Finance Stoch.,
5(3):275–303, 2001.
[10] F. E. Benth, K. H. Karlsen, and K. Reikvam. Portfolio optimization in a Lévy market with intertemporal
substitution and transaction costs. Stoch. Stoch. Rep., 74(3-4):517–569, 2002.
[11] M. G. Crandall. Semidifferentials, quadratic forms and fully nonlinear elliptic equations of second order.
Ann. Inst. H. Poincar Anal. Non Linaire, 6(6):419–435, 1989.
[12] M. G. Crandall and H. Ishii. The maximum principle for semicontinuous functions. Differential Integral
Equations, 3(6):1001–1014, 1990.
[13] M. G. Crandall, H. Ishii, and P.-L. Lions. User’s guide to viscosity solutions of second order partial
differential equations. Bull. Amer. Math. Soc. (N.S.), 27(1):1–67, 1992.
[14] M. G. Crandall and P.-L. Lions. Viscosity solutions of Hamilton-Jacobi equations. Trans. Amer. Math.
Soc., 277(1):1–42, 1983.
[15] W. H. Fleming and H. M. Soner. Controlled Markov processes and viscosity solutions. Springer-Verlag,
New York, 1993.
[16] W. H. Fleming and P. E. Souganidis. On the existence of value functions of two-player, zero-sum stochastic
differential games. Indiana Univ. Math. J., 38(2):293–314, 1989.
[17] N. C. Framstad, B. Øksendal, and A. Sulem. Optimal consumption and portfolio in a jump diffusion
market with proportional transaction costs. J. Math. Econom., 35(2):233–257, 2001.
[18] M. G. Garroni and J.-L. Menaldi. Green functions for second order parabolic integro-differential problems,
volume 275 of Pitman Research Notes in Mathematics Series. Longman Scientific & Technical, Harlow,
1992.
[19] M. G. Garroni and J. L. Menaldi. Second order elliptic integro-differential problems, volume 430 of Chap-
man & Hall/CRC Research Notes in Mathematics. Chapman & Hall/CRC, Boca Raton, FL, 2002.
[20] F. Gimbert and P.-L. Lions. Existence and regularity results for solutions of second-order, elliptic integro-
differential operators. Ricerche Mat., 33(2):315–358, 1984.
[21] H. Ishii. A boundary value problem of the Dirichlet type for Hamilton-Jacobi equations. Ann. Scuola
Norm. Sup. Pisa Cl. Sci. (4), 16(1):105–135, 1989.
[22] H. Ishii. On uniqueness and existence of viscosity solutions of fully nonlinear second-order elliptic PDEs.
Comm. Pure Appl. Math., 42(1):15–45, 1989.
[23] H. Ishii and P.-L. Lions. Viscosity solutions of fully non-linear second-order partial differential equations.
J. Differential Equations, 83:26-78, 1990.
[24] E. R. Jakobsen and K. H. Karlsen. Continuous dependence estimates for viscosity solutions of fully
nonlinear degenerate parabolic equations. J. Differential Equations, 183:497-525, 2002.
[25] E. R. Jakobsen and K. H. Karlsen. Continuous dependence estimates for viscosity solutions of fully
nonlinear degenerate elliptic equations. Electron. J. Diff. Eqns., 2002(39):1–10, 2002.
[26] E. R. Jakobsen and K. H. Karlsen. Continuous dependence estimates for viscosity solutions of fully
nonlinear degenerate parabolic integro-PDEs. In preparation.
22 JAKOBSEN AND KARLSEN

[27] R. Jensen. The maximum principle for viscosity solutions of fully nonlinear second order partial differential
equations. Arch. Rational Mech. Anal., 101(1):1-27, 1988.
[28] N. V. Krylov. Controlled diffusion processes. Springer-Verlag, New York, 1980.
[29] R. Mikulyavichyus and G. Pragarauskas. Nonlinear potentials of the Cauchy-Dirichlet problem for the
Bellman integro-differential equation. Liet. Mat. Rink., 36(2):178–218, 1996.
[30] H. Pham. Optimal stopping of controlled jump diffusion processes: A viscosity solution approach. J. Math.
Systems Estim. Control, 8(1), 1998.
[31] A. Sayah. Équations d’Hamilton-Jacobi du premier ordre avec termes intégro-différentiels. I. Unicité des
solutions de viscosité. Comm. Partial Differential Equations, 16(6-7):1057–1074, 1991.
[32] A. Sayah. Équations d’Hamilton-Jacobi du premier ordre avec termes intégro-différentiels. II. Existence
de solutions de viscosité. Comm. Partial Differential Equations, 16(6-7):1075–1093, 1991.
[33] H. M. Soner. Optimal control with state-space constraint. II. SIAM J. Control Optim., 24(6):1110–1122,
1986.
[34] H. M. Soner. Optimal control of jump-Markov processes and viscosity solutions. Stochastic differential
systems, stochastic control theory and applications (Minneapolis, Minn., 1986), IMA Vol. Math. Appl.,
10, 501–511, Springer, New York, 1988.

(Espen R. Jakobsen)
Department of Mathematical Sciences
Norwegian University of Science and Technology
N–7491 Trondheim, Norway
E-mail address: [email protected]
URL: http://www.math.ntnu.no/~erj/

(Kenneth Hvistendahl Karlsen)

Department of Mathematics
University of Bergen
Johs. Brunsgt. 12
N–5008 Bergen, Norway
and
Center of Mathematics for Applications
Department of Mathematics
University of Oslo
P.O. Box 1053, Blindern
N–0316 Oslo, Norway
E-mail address: [email protected]
URL: http://www.mi.uib.no/~kennethk/