J.

Differential Equations 245 (2008) 307–336

www.elsevier.com/locate/jde

Viscosity solutions to second order partial differential

equations on Riemannian manifolds ✩
Daniel Azagra ∗ , Juan Ferrera, Beatriz Sanz
Departamento de Análisis Matemático, Facultad de Matemáticas, Universidad Complutense, 28040 Madrid, Spain
Received 19 February 2007; revised 28 March 2008
Available online 7 May 2008

Abstract
We prove comparison, uniqueness and existence results for viscosity solutions to a wide class of fully
nonlinear second order partial differential equations F (x, u, du, d 2 u) = 0 defined on a finite-dimensional
Riemannian manifold M. Finest results (with hypothesis that require the function F to be degenerate ellip-
tic, that is nonincreasing in the second order derivative variable, and uniformly continuous with respect to
the variable x) are obtained under the assumption that M has nonnegative sectional curvature, while, if one
additionally requires F to depend on d 2 u in a uniformly continuous manner, then comparison results are
established with no restrictive assumptions on curvature.
 2008 Elsevier Inc. All rights reserved.

MSC: 58J32; 49J52; 49L25; 35D05; 35J70

Keywords: Degenerate elliptic second order PDEs; Hamilton–Jacobi equations; Viscosity solution; Riemannian
manifold

1. Introduction

The theory of viscosity solutions to nonlinear PDEs on Rn (and on infinite-dimensional Ba-

nach spaces) was introduced by M.G. Crandall and P.L. Lions in the 1980s. This theory quickly
gained popularity and was enriched and expanded with numerous and important contributions
from many mathematicians. We cannot mention all of the significant papers in the vast literature

✩
The authors were supported by grants MTM-2006-03531 and UCM-CAM-910626.
* Corresponding author.
E-mail addresses: [email protected] (D. Azagra), [email protected] (J. Ferrera), [email protected]
(B. Sanz).

0022-0396/$ – see front matter  2008 Elsevier Inc. All rights reserved.
doi:10.1016/j.jde.2008.03.030
308 D. Azagra et al. / J. Differential Equations 245 (2008) 307–336

concerning viscosity solutions and Hamilton–Jacobi equations, so we will content ourselves with
referring the reader to [4] and the references given therein.
More recently there have been various approaches to extend the theory of viscosity solutions
of first order Hamilton–Jacobi equations, and the corresponding nonsmooth calculus, to the set-
ting of Riemannian manifolds. This is a natural thing to do, because many functions arising from
geometrical problems, such as the distance function to a given set of a Riemannian manifold, are
not differentiable. Also, many important nonlinear equations full of geometrical meaning, such
as the eikonal equations, have no classical solutions, and their natural solutions, which in this
case we think are the viscosity solutions, are not differentiable (if some readers disagree with our
saying that viscosity solutions are the natural notion of solution for eikonal equations, they might
change their mind if they have a look at the recent paper [5], where the authors construct a 1-
Lipschitz function u defined on the closed unit ball B of Rn , n  2, which is differentiable on the
open ball B, and such that ∇u(x) = 1 almost everywhere, but ∇u(0) = 0; that is, the eikonal
equation ∇u(x) = 1 in B, u = 0 on ∂B, admits some exotic almost everywhere solutions
which are everywhere differentiable and are very different from its unique viscosity solution,
namely the distance function to the boundary ∂B, which is not everywhere differentiable but is
much more natural from a geometric point of view).
Mantegazza and Mennucci [9] studied viscosity solutions to eikonal equations on Riemannian
manifolds, in connection with regularity properties of the distance function to a compact subset
of the manifold. In [2] a theory of (first order) nonsmooth calculus for Riemannian manifolds
(possibly of infinite dimension) was introduced and applied to show existence and uniqueness of
viscosity solutions to Hamilton–Jacobi equations on such manifolds. Simultaneously, Ledyaev
and Zhu [8] developed a (first order) nonsmooth calculus on finite-dimensional Riemannian
manifolds and applied it to the study of Hamilton–Jacobi equations from a somewhat different
approach, related to control theory and differential inclusions.
The usefulness of nonsmooth analysis on Riemannian manifolds has been shown in [6], where
viscosity solutions are employed as a technical tool to prove important results in conformal ge-
ometry.
However, to the best of our knowledge, no one has yet carried out a systematic study of
second order viscosity subdifferentials and viscosity solutions to second order partial differential
equations on Riemannian manifolds.
In this paper we will initiate such a study by establishing comparison, uniqueness and exis-
tence of viscosity solutions to second order PDEs of the form
 
F x, u, du, d 2 u = 0,

where u : M → R and M is a finite-dimensional complete Riemannian manifold. We will study

the Dirichlet problem with a simple boundary condition of the type u = f on ∂Ω, where Ω is
an open subset of M; and also the same equation, with no boundary conditions, on all of M.
Let us briefly describe the results of this paper. We begin with the natural definition of second
order subjet of a function u : M → R, that is J 2,− u(x) = {(dϕ(x), d 2 ϕ(x)): ϕ ∈ C 2 (M, R),
f − ϕ attains a local minimum at x}. This is a nice definition from a geometric point of view,
but it would be complicated and uneconomic to develop a nonsmooth calculus exclusively based
on this definition. It is more profitable to try to localize the definition through charts and then
use the second order nonsmooth calculus on Rn to establish the corresponding results on M.
However, second derivatives of composite functions are complicated, so not every chart serves
 D. Azagra et al. / J. Differential Equations 245 (2008) 307–336 309

this purpose, and we have to work only with the exponential chart. It is not difficult to see that
(ζ, A) ∈ J 2,− u(x) if and only if (ζ, A) ∈ J 2,− (u ◦ expx )(0).
When one turns to the limiting subjet J 2,− u(x) (defined as the set of limits of sequences
(ζn , An ), where (ζn , An ) ∈ J 2,− u(xn ) and xn converges to x), things become less obvious but,
with the help of a lemma which relates the second derivatives of a function ϕ : M → R to those
of the function ψ = ϕ ◦ expx (at points near the origin in T Mx ), one can still show that (ζ, A) ∈
J 2,− u(x) if and only if (ζ, A) ∈ J 2,− (u ◦ expx )(0).
By using this characterization we can extend Theorem 3.2 of [4] to the Riemannian set-
ting. This kind of result can be regarded as a sophisticated nonsmooth fuzzy rule for the
superdifferential of the sum of two functions, and is the key to the proof of all the compar-
ison results in [4] and in this paper. The result essentially says that if u1 , u2 are two upper
semicontinuous functions on M, ϕ is a C 2 smooth function on M × M, and we assume that
ω(x1 , x2 ) = u1 (x1 ) + u2 (x2 ) − ϕ(x1 , x2 ) attains a local maximum at (x̂1 , x̂2 ), then, for each
ε > 0 there exist bilinear forms Bi ∈ L2s ((T M)x̂i , R), i = 1, 2, such that
 
∂
ϕ(x̂1 , x̂2 ), Bi ∈ J 2,+ ui (x̂i )
∂xi

for i = 1, . . . , k, and the block diagonal matrix with entries Bi satisfies

   
1 B1 0
− + A I   A + εA2 , (∗)
ε 0 B2

where A = d 2 ϕ(x̂1 , x̂2 ) ∈ L2s (T Mx̂1 × T Mx̂2 , R). This is all done in Section 2 of the paper.
In the case M = Rn this result is usually applied with ϕ(x, y) = α2 x − y2 , whose second
order derivative is given by the matrix
 
I −I
α .
−I I

When applied to vectors of the form (v, v) in Rn × Rn this derivative vanishes, which allows one
to derive from (∗) that B1  B2 (as quadratic forms). This in turn provides a very general form
of comparison result for viscosity solutions of the equation F (x, u, du, d 2 u) = 0 in which the
continuous function F is assumed to be degenerate elliptic (that is nonincreasing in the variable
d 2 u), strongly increasing in the variable u, and uniformly continuous with respect to x.
The natural approach in the Riemannian setting is then to consider ϕ(x, y) = α2 d(x, y)2 ,
where d is the Riemannian distance in M. Two problems immediately arise. First, the func-
tion ϕ is not differentiable in general if the points x, y are not suitably close to each other. This
is unimportant because, in the proof of the main comparison result, we only need ϕ to be C 2
smooth on a ball of small radius around a point x0 which is the limit of two different sequences
xα and yα , and we have to evaluate d 2 ϕ at the points (xα , yα ).
The second problem, however, is substantial. The second derivative of the function ϕ is a
quadratic form defined on T Mx × T My , and what we would like is that, when applied to a
vector of the form (v, Lxy v), where Lxy is the parallel transport from T Mx to T My along the
unique minimizing geodesic connecting x to y, this derivative is less than or equal to zero. This
way condition (∗) would imply that Lx̂2 x̂1 (B2 )  B1 , where Lx̂2 x̂1 (B2 ) is the parallel transport
310 D. Azagra et al. / J. Differential Equations 245 (2008) 307–336

of the quadratic form B2 from T Mx̂2 to T Mx̂1 along the unique minimizing geodesic connecting
x̂2 to x̂1 , defined by
   
Lx̂2 x̂1 (B2 )v, v := B2 (Lx̂1 x̂2 v), Lx̂1 x̂2 v .

And therefore we should be able to conclude that, if F is continuous, strongly increasing in the
variable u, and degenerate elliptic (that is F (x, r, ζ, B)  F (x, r, ζ, A) whenever A  B), then a
natural extension to X := {(x, r, ζ, A): x ∈ M, r ∈ R, ζ ∈ T Mx , A ∈ L2s (T Mx )} of the notion
of uniform continuity of F (x, r, ζ, A) with respect to the variable x (namely, that
 
F (y, r, Lxy ζ, Lxy P ) − F (x, r, ζ, P ) → 0 uniformly as y → x,

which we abbreviate by saying that F is intrinsically uniformly continuous with respect to x)

would be enough to show that comparison holds.
However, as we will show in Section 3, one has that

d 2 ϕ(x, y)(v, Lxy v)2  0

for all v ∈ T Mx if and only if M has nonnegative sectional curvature. Therefore, with this choice
of ϕ, one can get results as sharp as those in Rn only when one deals with manifolds of nonnega-
tive curvature. Nevertheless, if the sectional curvature K of M is bounded below, say K  −K0 ,
then one can show that

d 2 ϕ(x, y)(v, Lxy v)2  2K0 d(x, y)2 v2

for all v ∈ T Mx , and by using this estimation it is possible to deduce that, if one additionally
assumes that F satisfies a certain uniform continuity assumption with respect to the variables x
and D 2 u of the kind “for every ε > 0 there exists δ > 0 such that d(x, y)  δ and P −Lyx Q  δI
imply F (y, r, Lxy ζ, Q) − F (x, r, ζ, P )  ε for all ζ ∈ T Mx∗ , P ∈ L2s (T Mx ), Q ∈ L2s (T My ),
r ∈ R,” then the comparison principle holds for the equation F = 0 (either with the boundary
condition u = 0 on ∂Ω, or with the assumption that M has no boundary and the functions u, v
for which one seeks comparison are bounded). This is all shown in Sections 4 and 5.
In Section 6 we see that Perron’s method works perfectly well in the Riemannian setting. For
instance one can show existence of viscosity solutions to the equation u + G(x, du, d 2 u) = 0
on compact manifolds under the same continuity assumptions on G as those that we require for
comparison.
In particular, we get the following: if M is a compact manifold and G is degenerate elliptic
and uniformly continuous in the above sense, then there exists a unique viscosity solution of
u + G(x, du, d 2 u) = 0 on M. If one additionally assumes that M has nonnegative sectional
curvature then the above uniform continuity assumption can be relaxed: it is enough to require
that G is intrinsically uniformly continuous with respect to x, meaning that “for every ε > 0 there
exists δ > 0 such that d(x, y)  δ implies G(y, Lxy ζ, Lxy P ) − G(x, ζ, P )  ε for all ζ ∈ T Mx∗ ,
P ∈ L2s (T Mx ).”
We end the paper by discussing the applicability of the above theory to some particular exam-
ples of equations.
The notation we use is standard. M = (M, g) will always be a finite-dimensional Riemannian
manifold. The letters X, Y, Z, V , W will stand for smooth vector fields on the Riemannian man-
ifold M, and ∇Y X will always denote the covariant derivative of X along Y . The Riemannian
 D. Azagra et al. / J. Differential Equations 245 (2008) 307–336 311

curvature of M will be denoted by R. Geodesics in M will be denoted by γ , σ , and their velocity

D
fields by γ ′ , σ ′ . If X is a vector field along γ we will often denote X ′ (t) = dt X(t) = ∇γ ′ (t) X(t).
Recall that X is said to be parallel along γ if X ′ (t) = 0 for all t. The Riemannian distance in M
will always be denoted by d(x, y) (defined as the infimum of the lengths of all curves joining x
to y in M).
We will often identify (via the Riemannian metric) the tangent space of M at a point x,
denoted by T Mx , with the cotangent space at x, denoted by T Mx∗ . The space of bilinear forms
on T Mx (respectively symmetric bilinear forms) will be denoted by L2 (T Mx ) or L2 (T Mx , R)
(respectively L2s (T Mx ) or L2s (T Mx , R)). Elements of L2 (T Mx ) will be denoted by the letters
A, B, P , Q, and those of T Mx∗ by ζ , η, etc. Also, we will denote by T2,s (M) the tensor bundle
of symmetric bilinear forms, that is

T2,s (M) = L2s (T Mx , R)

x∈M

and T2,s (M)x = L2s (T Mx , R).

We will make extensive use of the exponential mapping expx and the parallel translation along
a geodesic γ throughout the paper, and of Jacobi fields along γ only in Section 3. Recall that for
every x ∈ M there exists a mapping expx , defined on a neighborhood of 0 in the tangent space
T Mx , and taking values in M, which is a local diffeomorphism and maps straight line segments
passing through 0 onto geodesic segments in M passing through x. The exponential mapping
also induces a local diffeomorphism on the cotangent space T Mx∗ , via the identification given by
the metric, that will be denoted by expx as well.
On the other hand, for a minimizing geodesic γ : [0, ℓ] → M connecting x to y in M, and for a
vector v ∈ T Mx there is a unique parallel vector field P along γ such that P (0) = v, this is called
the parallel translation of v along γ . The mapping T Mx ∋ v → P (ℓ) ∈ T My is a linear isometry
from T Mx onto T My which we will denote by Lxy . Its inverse is of course Lyx . This isometry
naturally induces an isometry (which we will still denote by Lxy ), T Mx∗ ∋ ζ → Lxy ζ ∈ T My∗ ,
defined by

Lxy ζ, v y := ζ, Lyx v x .

Similarly, Lxy induces an isometry L2 (T Mx , R) ∋ A → Lxy (A) ∈ L2 (T My , R) defined by

   
Lxy (A)v, v y := A(Lyx v), Lyx v x .

By iM (x) we will denote the injectivity radius of M at x, that is the supremum of the radius r
of all balls B(0x , r) in T Mx for which expx is a diffeomorphism from B(0x , r) onto B(x, r).
Similarly, i(M) will denote the global injectivity radius of M, that is i(M) = inf{iM (x): x ∈ M}.
Recall that the function x → iM (x) is continuous. In particular, if M is compact, we always have
i(M) > 0.
For Jacobi fields and any other unexplained terms of Riemannian geometry used in Section 3,
we refer the reader to [3,10].
312 D. Azagra et al. / J. Differential Equations 245 (2008) 307–336

2. Second order viscosity subdifferentials on Riemannian manifolds

Recall that the Hessian D 2 ϕ of a C 2 smooth function ϕ on a Riemannian manifold M is

defined by

D 2 ϕ(X, Y ) = ∇X ∇ϕ, Y ,

where ∇ϕ is the gradient of ϕ and X, Y are vector fields on M (see [10, p. 31]). The Hessian
is a symmetric tensor field of type (0, 2) and, for a point p ∈ M, the value D 2 ϕ(X, Y )(p) only
depends of f and the vectors X(p), Y (p) ∈ T Mp . So we can define the second derivative of ϕ
at p as the symmetric bilinear form d 2 ϕ(p) : T Mp × T Mp → R

(v, w) → d 2 ϕ(p)(v, w) := D 2 ϕ(X, Y )(p),

where X, Y are any vector fields such that X(p) = v, Y (p) = w. A useful way to compute
d 2 ϕ(p)(v, v) is to take a geodesic γ with γ ′ (0) = v and calculate

d2  
 ,
ϕ γ (t) 
dt 2 t=0

which equals d 2 ϕ(p)(v, v). We will often write d 2 ϕ(p)(v)2 instead of d 2 ϕ(p)(v, v).

Definition 2.1. Let M be a finite-dimensional Riemannian manifold, and f : M → (−∞, +∞]

a lower semicontinuous function. We define the second order subjet of f at a point x ∈ M by
 
J 2,− f (x) = dϕ(x), d 2 ϕ(x) : ϕ ∈ C 2 (M, R), f − ϕ attains a local minimum at x .

If (ζ, A) ∈ J 2,− f (x), we will say that ζ is a first order subdifferential of f , and A is a second
order subdifferential of f at x.
Similarly, for an upper semicontinuous function g : M → [−∞, +∞), we define the second
order superjet of f at x by
 
J 2,+ f (x) = dϕ(x), d 2 ϕ(x) : ϕ ∈ C 2 (M, R), f − ϕ attains a local maximum at x .

Observe that J 2,− f (x) and J 2,+ f (x) are subsets of T Mx∗ × L2s (T Mx , R), where L2s (T Mx , R) ≡
L2s (T Mx ) stands for the symmetric bilinear forms on T Mx . It is also clear that J 2,− f (x) =
−J 2,+ (−f )(x), and that we obtain the same definitions if we replace the condition
“ϕ ∈ C 2 (M, R)” with “ϕ is C 2 smooth on a neighborhood of x.”

By using the fact that a lower semicontinuous function f is bounded below on a neighbor-
hood B of any point x with f (x) < ∞, one can easily find a function ϕ ∈ C 2 (M, R) such that
infy∈∂B (f − ϕ)(y) > f (x), hence f − ϕ attains a local minimum at some point z ∈ B, and
(dϕ(z), d 2 ϕ(z)) ∈ J 2,− f (z). This shows that the set

z ∈ M: J 2,− f (z) = ∅

is dense in the set {x ∈ M: f (x) < ∞}. A similar statement is true of upper semicontinuous
functions. Therefore, when dealing with semicontinuous functions, one has lots of points where
 D. Azagra et al. / J. Differential Equations 245 (2008) 307–336 313

these subjets or superjets are nonempty, that is lots of points of second order sub- or super-
differentiability.
In the sequel M will always denote an n-dimensional Riemannian manifold. We next state
and prove several results for subjets which also hold, with obvious modifications, for superjets.

Proposition 2.2. Let f : M → (−∞, +∞] be a lower semicontinuous function. Let ζ ∈ T Mx∗ ,
A ∈ L2s (T Mx , R), x ∈ M. The following statements are equivalent:

(1) (ζ, A) ∈ J 2,− f (x).

1
(2) f (expx (v))  f (x) + ζ, v x + 2 Av, v x + o(v2 ).

Proof. (1) ⇒ (2) If (ζ, A) ∈ J 2,− f (x), by definition there exists ϕ ∈ C 2 (M, R) such that f − ϕ
attains a local minimum at x and ζ = dϕ(x), A = d 2 ϕ(x). We may obviously assume that ϕ(x) =
f (x), so we have

f (y) − ϕ(y)  0

on a neighborhood of x. Let us consider the function h(v) = ϕ(expx (v)) defined on a neighbor-
hood of 0x in T Mx . We have that
  1   
h(v) = h(0) + dh(0), v x + d 2 h(0)v, v x + o v2 .
2
By taking y = expx (v) and combining this with the above inequality we get

    1   
f expx (v)  f (x) + dh(0), v x + d 2 h(0)v, v x + o v2 ,
2

so we only need to show that ζ = dh(0) and A = d 2 h(0). To see this, let us fix v ∈ T Mx and
consider the geodesic γ (t) = expx (tv) and the function t → ϕ(γ (t)) = h(tv). We have that

d    
h(tv) = dϕ γ (t) , γ ′ (t)
dt
and

d2  2   ′ ′

h(tv) = d ϕ γ (t) γ (t), γ (t) .
dt 2
In particular, for t = 0, we get

d   
dh(0)(v) = h(tv) = dϕ(x), v = ζ, v ,
dt t=0

that is dh(0) = ζ ; and also


 2  d2   
d h(0)v, v = 2 h(tv) = d 2 ϕ(x)v, v = Av, v ,
dt t=0

that is A = d 2 h(0).
314 D. Azagra et al. / J. Differential Equations 245 (2008) 307–336

(2) ⇒ (1) Define F (v) = f (expx (v)) for v in a neighborhood of 0x ∈ T Mx . We have that

1  
F (v)  F (0) + ζ, v x + Av, v x + o v2 .
2

The result we want to prove is known to be true in the case when M = Rn , so there exists
ψ : T Mx → R such that F − ψ attains a minimum at 0 and dψ(0) = ζ , d 2 ψ(0) = A. Since
minima are preserved by composition with diffemorphisms, the function ϕ := ψ ◦ exp−1
x , defined
on an open neighborhood of x ∈ M, has the property that f − ϕ = (F − ψ) ◦ exp−1 x attains a
local minimum at x = exp−1
x (0). Moreover, according to (1) ⇒ (2) above, we have that

dϕ(x) = dψ(0) and d 2 ϕ(x) = d 2 ψ(0),

so we get dϕ(x) = ζ and d 2 ϕ(x) = A. Finally, by using smooth partitions of unity we can extend
ϕ from an open neighborhood of x to all of M. ✷

Corollary 2.3. Let f : M → (−∞, +∞] be a lower semicontinuous function, and consider
ζ ∈ T Mx∗ , A ∈ L2s (T Mx , R), x ∈ M. Then

(ζ, A) ∈ J 2,− f (x) ⇐⇒ (ζ, A) ∈ J 2,− (f ◦ expx )(0x ).

Making use of the above characterization, one can easily extend many known properties of
the sets J 2,− f (x) and J 2,+ f (x) from the Euclidean to the Riemannian setting. For instance,
one can immediately see that J 2,− f (x) and J 2,+ f (x) are convex subsets of T Mx∗ × L2s (T Mx ).
They are not necessarily closed, but if one fixes a ζ ∈ T Mx∗ then the set {A: (ζ, A) ∈ J 2,− f (x)}
is closed. A useful property that also extends from Euclidean to Riemannian is the following: if
ψ is C 2 smooth on a neighborhood of x then
 
J 2,− (f − ψ)(x) = ζ − dψ(x), A − d 2 ψ(x) : (ζ, A) ∈ J 2,− f (x) .

One can also see that f is twice differentiable at a point x ∈ M (in the sense that for some unique
ζ ∈ T Mx∗ , A ∈ L2s (T Mx , R) we have that f (expx (v)) = f (x) + ζ, v + 21 Av, v + o(v2 ) as
v → 0) if and only if J 2,− f (x) ∩ J 2,+ f (x) is nonempty (in which case J 2,− f (x) ∩ J 2,+ f (x) =
{(ζ, A)}).
Next we have to define the closures of these set-valued mappings. Let us first recall that a
sequence (An ) with An ∈ L2s (T Mxn ) is said to converge to A ∈ L2s (T Mx ) provided xn converges
to x in M and for every vector field V defined on an open neighborhood of x we have that
An V (xn ), V (xn ) converges to AV (x), V (x) . Since we have AV , W = 21 ( A(V + W ), V +
W − AV , V − AW, W ), it is clear that this is equivalent to saying that An V (xn ), W (xn )
converges to AV (x), W (x) for all vector fields V , W on a neighborhood of x in M.
Similarly, a sequence (ζn ) with ζn ∈ T Mx∗n converges to ζ provided that xn → x and
ζn , V (xn ) → ζ, V (x) for every vector field V defined on an open neighborhood of x.

Remark 2.4. It is not difficult to see that, if M = Rn , then An (respectively ζn ) converges to A

(respectively ζ ) in the above sense if and only if An − A → 0 (respectively ζn − ζ  → 0) in
L2s (Rn , R) (respectively in Rn ).
 D. Azagra et al. / J. Differential Equations 245 (2008) 307–336 315

It is also worth noting that An − A → 0 (respectively ζn − ζ  → 0) in L2s (Rn , R) (respec-
tively in Rn ) if and only if An v, v → Av, v (respectively ζn , v → ζ, v ) for every v ∈ Rn ,
that is pointwise convergence is equivalent to uniform convergence on bounded sets, as far as
linear or bilinear maps on Rn are concerned.

Definition 2.5. Let f be a lower semicontinuous function defined on a Riemannian manifold M,

and x ∈ M. We define

J 2,− f (x) = (ζ, A) ∈ T Mx∗ × Ls (T Mx ): ∃xn ∈ M, ∃(ζn , An ) ∈ J 2,− f (xn )

   
s.t. xn , f (xn ), ζn , An → x, f (x), ζ, A ,

and for an upper semicontinuous function g on M we define J 2,+ g(x) in an obvious similar
way.

Remark 2.6. According to Remark 2.4, we have that, in the case M = Rn , the sets J 2,− g(x) and
J 2,+ g(x) coincide with the subjets and superjets defined in [4].

In order to establish the analogue of Corollary 2.3 for the closure J 2,− g(x), we will use the
following fact.

Lemma 2.7. Let ϕ : M → R be a C 2 smooth function, and define ψ = ϕ ◦ expx on a neighbor-

hood of 0 ∈ T Mx . Let V be a vector field defined on a neighborhood of 0 in T Mx , and consider
the vector field defined by V (y) = d expx (wy )(V (wy )) on a neighborhood of x in M, where
wy := expx−1 (y), and let
 
σy (t) = expx wy + t V (wy ) .

Then we have that

 
D 2 ψ(V , V )(wy ) = D 2 ϕ(V , V )(y) + ∇ϕ(y), σy′′ (0) .

Observe that σx′′ (0) = 0 so, when y = x, we obtain

d 2 ψ(0)(v, v) = d 2 ϕ(x)(v, v)

for every v ∈ T Mx .

Proof of Lemma 2.7. Fix y near x. We have that

d   d      
ψ wy + t V (wy ) = ϕ σy (t) = ∇ϕ σy (t) , σy′ (t)
dt dt

and

d2   d2          
ψ w y + t V (w y ) = ϕ σy (t) = ∇σy′ (t) ∇ϕ σy (t) , σy′ (t) + ∇ϕ σy (t) , σy′′ (t) .
dt 2 dt 2
316 D. Azagra et al. / J. Differential Equations 245 (2008) 307–336

Note that σy′ (0) = V (y), hence by taking t = 0 we get the equality in the statement. Observe that
when y = x the curve σx is a geodesic, so σx′′ (0) = 0. ✷

Proposition 2.8. Let f : M → (−∞, +∞] be a lower semicontinuous function, and consider
ζ ∈ T Mx∗ , A ∈ L2s (T Mx , R), x ∈ M. Then

(ζ, A) ∈ J 2,− f (x) ⇐⇒ (ζ, A) ∈ J 2,− (f ◦ expx )(0x ).

Proof. (⇒) If (ζ, A) ∈ J 2,− f (x) there exist xn → x and (ζn , An ) ∈ J 2,− f (xn ) so that ζn → ζ ,
An → A, f (xn ) → f (x). Take ϕn ∈ C 2 (M) such that f − ϕn attains a minimum at xn and
ζn = dϕn (xn ), An = d 2 ϕn (xn ). Define ψn = ϕn ◦ expx on a neighborhood of 0 in T Mx , and
vn = exp−1x (xn ). It is clear that f ◦ expx −ψn attains a minimum at vn . We then have that
(dψn (vn ), d 2 ψn (vn )) ∈ J 2,− (f ◦ expx )(vn ), and since vn → 0 and f ◦ expx (vn ) → f (x), we
only have to show that dψn (vn ) → ζ and d 2 ψn (vn ) → A.
Take a vector field V on T Mx , and define a corresponding vector field V on a neighborhood
of x in M by
 
V (y) = d expx (wy ) V (wy ) ,

where wy = exp−1
x (y). We have that
     
dψn (vn ), V (vn ) = dϕn (xn ) ◦ d expx (vn ), V (vn ) = dϕn (xn ), V (xn ) ,

so we get
       
dψn (vn ), V (vn ) = ζn , V (xn ) → ζ, V (x) = ζ, V (0) ,

which shows dψn (vn ) → ζ . On the other hand, according to the preceding lemma, we also have
that
     
d 2 ψn (vn ) V (vn ), V (vn ) = An V (xn ), V (xn ) + ζn , σx′′n (0) ,

where σy (t) = expx (wy + t V (wy )).

Notice that the mapping y → σy′′ (0) defines a smooth vector field on a neighborhood of x
in M (and in particular σx′′n (0) → σx′′ (0) = 0 as n → ∞). Since An → A, ζn → ζ , we get, by
taking limits as n → ∞ in the above equality, that
     
d 2 ψn (vn ) V (vn ), V (vn ) → A V (x), V (x) + 0 = A V (0), V (0) ,

which proves that d 2 ψn (vn ) → A.

(⇐) If (ζ, A) ∈ J 2,− (f ◦ expx )(0) there exist vn → 0 and (ζ̃n , An ) ∈ J 2,− (f ◦ expx )(vn ) so
that ζ̃n → ζ , An → A, f (xn ) → f (x), where xn = expx (vn ). Take ψn ∈ C 2 (T Mx ) such that f ◦
expx −ψn attains a minimum at vn and ζ̃n = dψn (vn ), An = d 2 ψ(vn ). Define ϕn = ψn ◦ exp−1 x
on a neighborhood of x in M. Then f − ϕn attains a minimum at xn , so (dϕn (xn ), d 2 ϕn (xn )) ∈
J 2,− f (xn ), and we only have to show that dϕn (xn ) → ζ and d 2 ϕn (xn ) → A. Take a vector field
 D. Azagra et al. / J. Differential Equations 245 (2008) 307–336 317

V on a neighborhood of x in M, and define a corresponding vector field V on a neighborhood

of 0 in T Mx by
 
V (wy ) = d exp−1
x (y) V (y) ,

where wy = exp−1
x (y). Now we have that
   
dψn (vn ), V (vn ) = dϕn (xn ), V (xn ) ,

from which we deduce that dϕn (xn ) → ζ ; and also, by using this fact and the preceding lemma,

     
d 2 ϕn V (xn ), V (xn ) = An V (vn ), V (vn ) − dϕn (xn ), σx′′n (0)
     
→ AV (x), V (x) − ζ, σx′′ (0) = AV (x), V (x) − 0,

concluding the proof. ✷

Remark 2.9. One can see, as in the case of J 2,− f (x), that if ψ is C 2 smooth on a neighborhood
of x then
 
J 2,− (f − ψ)(x) = ζ − dψ(x), A − d 2 ψ(x) : (ζ, A) ∈ J 2,− f (x) .

The following result is the Riemannian version of Theorem 3.2 in [4] and, as in that paper, will
be the key to the proofs of comparison and uniqueness results for viscosity solutions of second
order PDEs on Riemannian manifolds.

Theorem 2.10. Let M1 , . . . , Mk be Riemannian manifolds, and Ωi ⊂ Mi open subsets. Define

Ω = Ω1 × · · · × Ωk ⊂ M1 × · · · × Mk = M. Let ui be upper semicontinuous functions on Ωi ,
i = 1, . . . , k; let ϕ be a C 2 smooth function on Ω and set

ω(x) = u1 (x1 ) + · · · + uk (xk )

for x = (x1 , . . . , xk ) ∈ Ω. Assume that x̂ = (x̂1 , . . . , x̂k ) is a local maximum of ω − ϕ. Then, for
each ε > 0 there exist bilinear forms Bi ∈ L2s ((T Mi )x̂i , R), i = 1, . . . , k, such that
 
∂
ϕ(x̂), Bi ∈ J 2,+ ui (x̂i )
∂xi

for i = 1, . . . , k, and the block diagonal matrix with entries Bi satisfies

⎛ ⎞
  B1 ... 0
1 . .. .. ⎠
− + A I  ⎝ .. . .  A + εA2 ,
ε
0 ... Bk

where A = d 2 ϕ(x̂) ∈ L2s (T Mx̂ , R).

318 D. Azagra et al. / J. Differential Equations 245 (2008) 307–336

Recall that, for ζ ∈ T M ∗ , A ∈ L(T Mx × T Mx , R), the norms ζ x and Ax are defined by

ζ x = sup ζ, v x : v ∈ T Mx , vx  1

and
 
Ax = sup  Av, v x : v ∈ T Mx , vx  1 = sup |λ|: λ is an eigenvalue of A .

Proof of Theorem 2.10. The result is proved in [4] in the case when all the manifolds Mi are
Euclidean spaces, and we are going to reduce the problem to this situation. By taking smaller
neighborhoods of the xi if necessary, we can assume that the Ωi are diffeomorphic images of
balls by the exponential mappings expx̂i : B(0, ri ) → Ωi = B(x̂i , ri ), and that expx̂ maps diffeo-
morphically a ball in T Mx̂ onto a ball containing Ω. The exponential map expx̂ from this ball in
T Mx̂ = (T M1 )x̂1 × · · · × (T Mk )x̂k into M is given by
 
expx̂ (v1 , . . . , vk ) = expx̂1 (v1 ), . . . , expx̂k (vk ) .

Now define functions on open subsets of Euclidean spaces by ω̃(v) = ω(expx̂ (v)) and ũi (vi ) =
ui (expx̂i (vi )). We have that ω̃(v1 , . . . , vk ) = ũ1 (v1 ) + · · · + ũk (vk ), and 0x̂ = (0x̂1 , . . . , 0x̂k ) is a
local maximum of ω̃ − ψ , where ψ = ϕ ◦ expx̂ .
Then, by the known result for Euclidean spaces, for each ε > 0 there exist bilinear forms
Bi ∈ L2s ((T Mi )x̂i , R), i = 1, . . . , k, such that
 
∂
ψ(0x̂ ), Bi ∈ J 2,+ ũi (0x̂i )
∂vi

for i = 1, . . . , k, and the block diagonal matrix with entries Bi satisfies

⎛ ⎞
  B1 ... 0
1 . .. .. ⎠
− + A I  ⎝ .. . .  A + εA2 ,
ε
0 ... Bk

where A = d 2 ψ(0x̂ ) ∈ L2s (T Mx̂ , R). According to Proposition 2.8 we have that
   
∂ ∂
ψ(0x̂ ), Bi ∈ J 2,+ ũi (0x̂i ) ⇐⇒ ψ(0x̂ ), Bi ∈ J 2,+ ui (x̂i ),
∂vi ∂vi

so we are done if we only see that

∂ ∂
ψ(0x̂ ) = ϕ(x̂) and d 2 ψ(0x̂ ) = d 2 ϕ(x̂).
∂vi ∂xi

But this is a consequence of Lemma 2.7. ✷

Now we extend the notion of viscosity solution to a Hamilton–Jacobi equation on a Rieman-

nian manifold. In the sequel we will denote

X := (x, r, ζ, A): x ∈ M, r ∈ R, ζ ∈ T Mx , A ∈ L2s (T Mx ) .

 D. Azagra et al. / J. Differential Equations 245 (2008) 307–336 319

Definition 2.11 (Viscosity solution). Let M be a Riemannian manifold, and F : X → R. We

say that an upper semicontinuous function u : M → R is a viscosity subsolution of the equation
F = 0 provided that
 
F x, u(x), ζ, A  0

for all x ∈ M and (ζ, A) ∈ J 2,+ u(x). Similarly, a viscosity supersolution of F = 0 on M is a

lower semicontinuous function u : M → R such that
 
F x, u(x), ζ, A  0

for every x ∈ M and (ζ, A) ∈ J 2,− u(x). If u is both a viscosity subsolution and a viscosity
supersolution of F = 0, we say that u is a viscosity solution of F = 0 on M.

Remark 2.12. If u is a solution of F  0 and F is continuous on X then F (x, u(x), ζ, A)  0

for every (ζ, A) ∈ J 2,+ u(x). A similar observation applies to solutions of F  0 and solutions
of F = 0.

Definition 2.13 (Degenerate ellipticity). We will say that a function F : X → R is degenerate

elliptic provided that

AB ⇒ F (x, r, ζ, B)  F (x, r, ζ, A)

for all x ∈ M, r ∈ R, ζ ∈ T Mx , A, B ∈ L2s (T Mx ).

Example 2.14. If we canonically identify the space of symmetric bilinear forms on T Mx with
the space of self-adjoint linear mappings from T Mx into T Mx , we have that

Lyx (Q) = L−1

xy QLxy .

Hence
 
trace(Lyx Q) = trace(Q) and det+ Lyx (Q) = det+ (Q)

(where det+ A is defined as the product of the nonnegative eigenvalues of A), and it is immedi-
ately seen that the functions G(r, ζ, A) = − det+ (A) and H (r, ζ, A) = − trace(A) are degenerate
elliptic and, moreover, are invariant by parallel translation, in the sense that

G(r, ζ, A) = G(r, Lxy ζ, Lxy A)

for all r ∈ R, ζ ∈ T Mx , A ∈ L2s (T Mx ). The same can be said of all nonincreasing, symmetric
functions of the eigenvalues of A. Thus one may combine such functions to construct many
interesting examples of equations to which our results apply, as we will see later on.

Remark 2.15. If the function F is degenerate elliptic, then every classical solution of F = 0 is
a viscosity solution of F = 0, as is immediately seen. However this may be not true if F is not
degenerate elliptic; for instance when M = R the function u(x) = x 2 − 2 is a classical solution
of u′′ (x) + u(x) − x 2 = 0 but is not a viscosity solution.
320 D. Azagra et al. / J. Differential Equations 245 (2008) 307–336

In order that the theory of viscosity solutions applies to an equation F = 0, the following
condition is usually required.

Definition 2.16 (Properness). We will say that a function F : X → R, (x, r, ζ, A) →

F (x, r, ζ, A), is proper provided

(i) F is degenerate elliptic, and

(ii) F is nondecreasing in the variable r.

3. A key property of the Hessian of the function (x, y) → d(x, y)2

When trying to establish comparison results for viscosity solutions of second order PDEs on a
Riemannian manifold M we will need to know how the Hessian of the function ϕ : M × M → R,

ϕ(x, y) = d(x, y)2

behaves. More precisely we will need to know on which manifolds M one has that

d 2 ϕ(x, y)(v, Lxy v)2  0 (♯)

for all v ∈ T Mx , with x, y ∈ M close enough to each other so that d(x, y) < min{iM (x), iM (y)}.
Let us calculate this derivative. We have that
∂ϕ ∂d
(x, y) = 2d(x, y) (x, y) = −2 exp−1
x (y). (1)
∂x ∂x
The second equality can be checked, for instance, by using the first variation formula of the
arc-length (see [10, p. 90]). Indeed, if α(t, s) is a variation through geodesics of a minimizing
geodesic γ (t) with y = γ (0) and x = γ (ℓ), where ℓ = d(x, y), and if L(s) denotes the length of
the geodesic t → α(t, s), then

  ℓ 
d     
L(s) = V , T |0 −
ℓ
V , ∇T T dt = V (ℓ), T (ℓ) − V (0), T (0) ,
ds s=0
0

where T = ∂α/∂t (so ∇T T = 0) and V = ∂α/∂s. Taking an α such that V is the Jacobi field
along γ satisfying V (0) = 0, V (ℓ) = v, we get

∂d d  1 
(x, y)(v) = L(s) = v, − exp−1
x (y) .
∂x ds s=0 ℓ

Similarly, we have

∂ϕ ∂d
(x, y) = 2d(x, y) (x, y) = −2 exp−1
y (x). (2)
∂y ∂y

Observe that
   
∂ϕ ∂ϕ ∂d ∂d
(x, y) + Lxy (x, y) = 0 = (x, y) + Lxy (x, y) . (3)
∂y ∂x ∂y ∂x
 D. Azagra et al. / J. Differential Equations 245 (2008) 307–336 321

By differentiating again in (1) and (2), we get

 2
∂ 2ϕ 2 ∂d ∂ 2d
(x, y)(v) = 2 (x, y)(v) + 2d(x, y) (x, y)(v)2 ,
∂x 2 ∂x ∂x 2
∂ 2ϕ ∂d ∂d ∂ 2d
(x, y)(v, w) = 2 (x, y)(v) (x, y)(w) + 2d(x, y) (x, y)(v, w),
∂x∂y ∂x ∂y ∂x∂y
 2
∂ 2ϕ 2 ∂d ∂ 2d
(x, y)(w) = 2 (x, y)(w) + 2d(x, y) (x, y)(w)2 ,
∂y 2 ∂y ∂y 2

so, if we take w = Lxy v and we sum the two first equations, and then we use (3), we get that
 2 
∂ 2ϕ 2 ∂ 2ϕ ∂ d 2 ∂ 2d
(x, y)(v) + (x, y)(v, L xy v) = 2d(x, y) (x, y)(v) + (x, y)(v, L xy v)
∂x 2 ∂x∂y ∂x 2 ∂x∂y

and we get a similar equation by changing x for y. By summing these two equations we get

d 2 ϕ(x, y)(v, Lxy v)2 = 2d(x, y)d 2 (d)(x, y)(v, Lxy v)2 ,

so it is clear that condition (♯) holds if and only if

d 2 (d)(x, y)(v, Lxy v)2  0 (♭)

for all v ∈ T Mx .
Another way to write conditions (♯) or (♭) is

d2   

d σ x (t), σ y (t)   0, (♮)
dt 2 t=0

where σx and σy are geodesics with σx (0) = x, σy (0) = y, σx′ (0) = v and σy′ (0) = Lxy v. The
function t → h(t) := d(σx (t), σy (t)) measures the distance between the geodesics σx and σy
(which have the same velocity and are parallel at t = 0) evaluated at a point moving along any of
these geodesics.
We are going to show that the second derivative h′′ (0) is negative (that is, condition (♮) holds)
if and only if M has positive sectional curvature.
In particular, by combining this fact with Eq. (3) (which tells us that h′ (0) = 0), we see that the
function h(t) attains a local maximum at t = 0 if and only if M has positive sectional curvature.
This corresponds to the intuitive notion that two geodesics that are parallel at their starting points
will get closer if the sectional curvature is positive, while they will spread apart if the sectional
curvature is negative.

Proposition 3.1. Condition (♯) (equivalently (♭), or (♮)) holds for a Riemannian manifold M if
and only if M has nonnegative sectional curvature. In fact one has, for the function ϕ(x, y) =
d(x, y)2 on M × M, that:
322 D. Azagra et al. / J. Differential Equations 245 (2008) 307–336

(1) If M has nonnegative sectional curvature then

d 2 ϕ(x, y)(v, Lxy v)2  0

for all v ∈ T Mx , with x, y ∈ M close enough to each other so that d(x, y) < min{iM (x),
iM (y)}.
(2) If M has nonpositive sectional curvature then

d 2 ϕ(x, y)(v, Lxy v)2  0

for all v ∈ T Mx , x, y ∈ M such that d(x, y) < min{iM (x), iM (y)}.

This fact must be known to the specialists in Riemannian geometry, but we have been unable
to find a reference for part (1), so we provide a proof. Let us begin by reviewing some standard
facts about the second variation of the arc-length and the energy functionals.
Take two points x0 , y0 ∈ M with d(x0 , y0 ) < min{iM (x0 ), iM (y0 )}, and let γ be the unique
minimizing geodesic, parameterized by arc-length, connecting x0 to y0 . Denote ℓ = d(x0 , y0 ),
the length of γ . Consider α(t, s), a smooth variation of γ , that is a smooth mapping
α : [0, ℓ] × [−ε, ε] → M such that α(t, 0) = γ (t) for all t ∈ [0, ℓ]. Consider the length and
the energy functionals, defined by

ℓ
 ′ 
L(s) = L(αs ) = α (t) dt
s
0

and

ℓ
 ′ 2
E(s) = E(αs ) = α (t) dt,
s
0

where αs is the variation curve defined by αs (t) = α(t, s) for every t ∈ [0, ℓ]. According to the
Cauchy–Schwarz inequality (applied to the functions f ≡ 1 and g(t) = αs′ (t) on the interval
[0, ℓ]) we have that

L(s)2  ℓE(s),

with equality if and only if αs′ (t) is constant. Therefore, in the case when αs is a geodesic for
each s (that is α is a variation of γ through geodesics) we have that

L(s)2 = ℓE(s)

for every s ∈ [−ε, ε].

Now take a vector v ∈ T Mx0 , set w = Lx0 y0 v, and consider the geodesics σx0 , σy0 defined by

σx0 (s) = expx0 (sv), σy0 (s) = expy0 (sw).

 D. Azagra et al. / J. Differential Equations 245 (2008) 307–336 323

We want to calculate

d2  
 ,
d 2 ϕ(x0 , y0 )(v, w)2 = ϕ σ x0 (s), σ y0 (s) 
ds 2 s=0

where ϕ(x, y) = d(x, y)2 . To this end let us denote by αs : [0, ℓ] → M the unique minimizing
geodesic joining the point σx0 (s) to the point σy0 (s) (now, for s = 0, αs is not necessarily param-
eterized by arc-length), and let us define α : [0, ℓ] × [−ε, ε] → M by α(t, s) = αs (t). Then α is
a smooth variation through geodesics of γ (t) = α(t, 0) and, according to the above discussion,
we have
 
ϕ σx0 (s), σy0 (s) = L(s)2 = ℓE(s),

and therefore

d 2 ϕ(x0 , y0 )(v, w)2 = ℓE ′′ (0). (4)

If we denote X(t) = ∂α(t, 0)/∂s, the variational field of α, then the formula for the second
variation of energy (see [3, p. 197]) tells us that

ℓ  t=ℓ
1 ′′    t=ℓ D ∂α 
E (0) = − X, X ′′ + R(γ ′ , X)γ ′ dt + X(t), X ′ (t) t=0 + (t, 0), γ ′ (t)  , (5)
2 ds ∂s t=0
0

or equivalently

ℓ  t=ℓ
1 ′′   D ∂α  
E (0) = ′ ′ ′
X , X − R(γ , X)γ , X dt + ′
(t, 0), γ (t)  ,
′
(6)
2 ds ∂s t=0
0

where we denote X ′ = ∇γ ′ (t) X, and X ′′ = ∇γ ′ (t) X ′ .

Note that, since the variation field of a variation through geodesics is always a Jacobi field, and
since the points x0 and y0 are not conjugate, the field X is in fact the unique Jacobi field along γ
satisfying that X(0) = v, X(ℓ) = w, that is X is the unique vector field along γ satisfying
 
X ′′ (t) + R γ ′ (t), X(t) γ ′ (t) = 0 and X(0) = v, X(ℓ) = w,

where R is the curvature of M. On the other hand, since the curves s → α(0, s) = σx0 (s) and
s → α(ℓ, s) = σy0 (s) are geodesics, we have that
 t=ℓ
D ∂α 
(t, 0), γ ′ (t)  = 0.
ds ∂s t=0

These observations allow us to simplify (5) and (6) by dropping the terms that vanish, thus
obtaining that

1 ′′    
E (0) = X(ℓ), X ′ (ℓ) − X(0), X ′ (0) (7)
2
324 D. Azagra et al. / J. Differential Equations 245 (2008) 307–336

and also

ℓ
1 ′′   
E (0) = X ′ , X ′ − R(γ ′ , X)γ ′ , X dt. (8)
2
0

Recall that the right-hand side of (8) is called the index form and is denoted by I (X, X).
By combining (4), (7) and (8) we get
   
d 2 ϕ(x0 , y0 )(v, w)2 = 2ℓ X(ℓ), X ′ (ℓ) − X(0), X ′ (0)
ℓ
  
= 2ℓ X ′ , X ′ − R(γ ′ , X)γ ′ , X dt.
0

Therefore condition (♯) holds if and only if, for every Jacobi field X along γ with X(0) = v,
X(ℓ) = w = Lx0 y0 v, one has that
   
X(ℓ), X ′ (ℓ) − X(0), X ′ (0)  0 (♦)

or, equivalently,

ℓ
  
X ′ , X ′ − R(γ ′ , X)γ ′ , X dt  0
0

for the same Jacobi fields.

Proof of Proposition 3.1. The proof of (2) is immediate and is well referenced (see for instance
[7, Theorem IX.4.3]): if M has nonpositive sectional curvature then we have R(γ ′ , X)γ ′ , X  0,
hence, according to the above formulas,

ℓ ℓ
  
2 2
d ϕ(x0 , y0 )(v, w) = 2ℓ X , X − R(γ ′ , X)γ ′ , X dt  2ℓ
′ ′
X ′ , X ′ dt  0,
0 0

which proves (2). Note that in this case we do not use that w = Lx0 y0 v, so this holds for all v, w.
Our proof of (1) uses the following lemma, which is a restatement of Corollary 10 in Chapter 8
of [11].

Lemma 3.2. Let γ : [0, ℓ] → M be a geodesic without conjugate points, X a Jacobi field along γ ,
and Z a piecewise smooth vector field along γ such that X(0) = Z(0) and X(ℓ) = Z(ℓ). Then

I (X, X)  I (Z, Z),

and equality holds only when Z = X.

 D. Azagra et al. / J. Differential Equations 245 (2008) 307–336 325

That is, among all vector fields along γ with the same boundary conditions, the unique Jacobi
field along γ determined by those conditions minimizes the index form. Recall that

ℓ
  
I (Z, Z) = Z ′ , Z ′ − R(γ ′ , Z)γ ′ , Z dt,
0

but this number is not equal to Z(ℓ), Z ′ (ℓ) − Z(0), Z ′ (0) unless Z is a Jacobi field.
Let X be the unique Jacobi field with X(0) = v, X(ℓ) = w = Lx0 y0 (v). Define Z = P (t),
where P (t) is the parallel translation along γ with P (0) = v (hence P (ℓ) = w). The field Z is
not necessarily a Jacobi field, but it has the considerable advantage that Z ′ (t) = 0 for all t, so we
have that

ℓ ℓ
    
I (Z, Z) = Z , Z − R(γ ′ , Z)γ ′ , Z dt = −
′ ′
R(γ ′ , Z)γ ′ , Z dt  0
0 0

because M has nonnegative sectional curvature. We then deduce from the above lemma that
   
X(ℓ), X ′ (ℓ) − X(0), X ′ (0) = I (X, X)  I (Z, Z)  0,

which, according to the above remarks (see (♦)), concludes the proof. ✷

Even though we will not have d 2 ϕ(x, y)(v, Lxy v)2  0 when M has negative curvature, we
can estimate this quantity and show that it is bounded by a term of the order of d(x, y)2 , pro-
vided that the curvature is bounded below. This will also be used in the next section to deduce
a comparison result which holds for all Riemannian manifolds (assuming that F is uniformly
continuous).

Proposition 3.3. Let M be a Riemannian manifold. Consider the function ϕ(x, y) = d(x, y)2 ,
defined on M × M. Assume that the sectional curvature K of M is bounded below, say K  −K0 .
Then

d 2 ϕ(x, y)(v, Lxy v)2  2K0 d(x, y)2 v2

for all v ∈ T Mx and x, y ∈ M with d(x, y) < min{iM (x), iM (y)}.

Note that for K0 = 0 we recover part (1) of Proposition 3.1.

Proof. Let X, Z be as in the proof of (1) of the preceding proposition. With the same notations,
we have that

d 2 ϕ(x, y)(v, Lxy v)2 = 2ℓI (X, X)  2ℓI (Z, Z)

ℓ ℓ
   2
= −2ℓ R(γ ′ , Z)γ ′ , Z dt  2ℓ K0 γ ′ (t) ∧ Z(t) dt
0 0
326 D. Azagra et al. / J. Differential Equations 245 (2008) 307–336

ℓ ℓ
 2  2
 2ℓ K0 γ ′ (t) Z(t) = 2ℓ K0 v2 dt = 2ℓ2 K0 v2 ,
0 0

which proves the result. ✷

4. Comparison results for the Dirichlet problem

In this section and throughout the rest of the paper we will often abbreviate saying that u is
an upper semicontinuous function on a set Ω by writing u ∈ USC(Ω). Similarly, LSC(Ω) will
stand for the set of lower semicontinuous functions on Ω.
The following lemma will be used in the proof of the main comparison result for the Dirichlet
problem
 
F x, u(x), du(x), d 2 u(x) = 0 on Ω, u=f on ∂Ω. (DP)

Lemma 4.1. Let Ω be a subset of a Riemannian manifold M, u ∈ USC(Ω), v ∈ LSC(Ω) and

 
α
mα := sup u(x) − v(y) − d(x, y)2
Ω×Ω 2

for α > 0. Let mα < ∞ for large α and (xα , yα ) be such that
  
α
lim mα − u(xα ) − v(yα ) − d(xα , yα )2 = 0.
α→∞ 2

Then we have:

(1) limα→∞ αd(xα , yα )2 = 0, and

(2) limα→∞ mα = u(x̂) − v(x̂) = supx∈Ω (u(x) − v(x)) whenever x̂ ∈ Ω is a limit point of xα
as α → ∞.

Proof. The result is proved in [4, Lemma 3.1] in the case M = Rn , and the same proof clearly
works in the generality of the statement (in fact this holds in any metric space). ✷

Now we can prove the main comparison result for the Dirichlet problem.

Theorem 4.2. Let Ω be a bounded open subset of a complete finite-dimensional Riemannian

manifold M, and F : X → R be proper, continuous, and satisfy:

(1) there exists γ > 0 such that

γ (r − s)  F (x, r, ζ, Q) − F (x, s, ζ, Q)

for r  s; and
 D. Azagra et al. / J. Differential Equations 245 (2008) 307–336 327

(2) there exists a function ω : [0, ∞] → [0, ∞] with limt→0+ ω(t) = 0 and such that
     
F y, r, α exp−1 −1 2
y (x), Q − F x, r, −α expx (y), P  ω αd(x, y) + d(x, y)

for all x, y ∈ Ω, r ∈ R, P ∈ T2,s (M)x , Q ∈ T2,s (M)y with

    
1 I 0 P 0
− + Aα    Aα + εα A2α , (∗)
εα 0 I 0 −Q

where Aα is the second derivative of the function ϕα (x, y) = α2 d(x, y)2 at the point (x, y) ∈
M × M,

1
εα = ,
2(1 + Aα )

and the points x, y are assumed to be close enough to each other so that d(x, y) <
min{iM (x), iM (y)}.

Let u ∈ USC(Ω) be a subsolution and v ∈ LSC(Ω) a supersolution of F = 0 on Ω, and u  v

on ∂Ω.
Then u  v holds on all of Ω.
In particular the Dirichlet problem (DP) has at most one viscosity solution.

Proof. Assume to the contrary that there exists z ∈ Ω with u(z) > v(z). By compactness of Ω
and upper semicontinuity of u − v, and according to Lemma 4.1, there exist xα , yα so that, with
the notation of Lemma 4.1,
α
u(xα ) − v(yα ) − d(xα , yα )2 = mα  δ := u(z) − v(z) > 0 (3)
2
and

αd(xα , yα )2 → 0 as α → ∞. (4)

Again by compactness of Ω we can assume that a subsequence of (xα , yα ), which we will still
denote (xα , yα ) (and suppose α ∈ N), converges to a point (x0 , y0 ) ∈ Ω × Ω. By Lemma 4.1 we
have that x0 = y0 and
 
δ  lim mα = u(x0 ) − v(x0 ) = sup u(x) − v(x) ,
α→∞
Ω

and in view of the condition u  v on ∂Ω we have that x0 ∈ Ω, and xα , yα ∈ Ω for large α.

Fix r0 > 0 and R0 > 0 such that, for every x ∈ B(x0 , r0 ), expx is a diffeomorphism from
B(0, R0 ) ⊂ T Mx onto B(x, R0 ) ⊃ B(x0 , r0 ) (see [3, Theorem 3.7 of Chapter 3]). Then, for ev-
ery x, y ∈ B(x0 , r0 ) we have that d(x, y) < min{iM (x), iM (y)}, the vectors exp−1 x (y) ∈ T Mx ≡
T Mx∗ and exp−1y (x) ∈ T M y ≡ T M ∗ are well defined, and the function ϕ(x, y) = d(x, y)2 is C 2
y
smooth on B(x0 , r0 ) × B(x0 , r0 ) ∈ M × M. Taking a subsequence if necessary, we can assume
that xα , yα ∈ B(x0 , r0 ) for all α.
328 D. Azagra et al. / J. Differential Equations 245 (2008) 307–336

Now, for each α, we can apply Theorem 2.10 with Ω1 = Ω2 = B(x0 , r0 ), u1 = u, u2 = −v,
ϕ(x, y) = ϕα (x, y) := α2 d(x, y)2 , and for

1
ε = εα := .
2(1 + d 2 ϕα (xα , yα ))

Since (xα , yα ) is a local maximum of the function (x, y) → u(x) − v(y) − ϕ(x, y), we obtain
bilinear forms P ∈ L2s (T Mxα , R), and Q ∈ L2s (T Myα , R) such that
 
∂
ϕ(xα , yα ), P ∈ J 2,+ u(xα ),
∂x
 
∂
− ϕ(xα , yα ), Q ∈ J 2,− v(yα )
∂y

(recall that J 2,− v(yα ) = −J 2,+ (−v)(yα )), and

   
1 P 0
− + Aα  I   Aα + εα A2α ,
εα 0 −Q

where Aα = d 2 ϕ(xα , yα ) ∈ L2s (T M(x,y) , R), so we get that condition (∗) holds for x = xα ,
y = yα . Therefore, according to condition (2), we have that
     
F yα , r, α exp−1 −1 2
yα (xα ), Q − F xα , r, −α expxα (yα ), P  ω αd(xα , yα ) + d(xα , yα ) . (5)

On the other hand, from Eq. (1) in the preceding section we have that

∂ ∂
ϕ(xα , yα ) = −α exp−1
xα (yα ) and − ϕ(xα , yα ) = α exp−1
yα (xα ),
∂x ∂y

hence (−α expx−1α

(yα ), P ) ∈ J 2,+ u(xα ), (α exp−1
yα (xα ), Q) ∈ J
2,− v(y ). Since u is subsolution
α
and v is supersolution, and F is continuous we then have, according to Remark 2.12, that
   
F xα , u(xα ), −α exp−1 −1
xα (yα ), P  0  F yα , v(yα ), α expyα (xα ), Q . (6)

By combining Eqs. (3)–(6) above, and using condition (1) too, we finally get
 
0 < γ δ  γ u(xα ) − v(yα )
   
 F xα , u(xα ), −α exp−1 −1
xα (yα ), P − F xα , v(yα ), −α expxα (yα ), P
   
 F xα , u(xα ), −α exp−1 −1
xα (yα ), P − F yα , v(yα ), α expyα (xα ), Q
   
+ F yα , v(yα ), α exp−1 −1
yα (xα ), Q − F xα , v(yα ), −α expxα (yα ), P
 
 ω αd(xα , yα )2 + d(xα , yα ) ,

and the contradiction follows by letting α → ∞. ✷

 D. Azagra et al. / J. Differential Equations 245 (2008) 307–336 329

Remark 4.3. Observe that, since α exp−1 −1

y (x) = Lxy (−α expx (y)), condition (2) of Theorem 4.2
can be replaced with a stronger but simpler assumption, namely that
 
F (y, r, Lxy ζ, Q) − F (x, r, ζ, P )  ω αd(x, y)2 + d(x, y)

for all x, y ∈ Ω, r ∈ R, P ∈ T2,s (M)x , Q ∈ T2,s (M)y , ζ ∈ T Mx∗ satisfying (∗).

Remark 4.4. If we want to compare two solutions u and v of F = 0 and we know that these
functions are bounded by some R > 0 (e.g. when M is compact) then it is obvious from the
above proof that it suffices to require that conditions (1) and (2) of Theorem 4.2 be satisfied for
all r, s in the interval [−R, R].

Proposition 4.5. If M has nonnegative sectional curvature, then condition (∗) implies that
P  Lyx (Q).

Proof. Let λ1 , . . . , λn be the eigenvalues of the restriction of Aα to the subspace D =

{(v, Lxy v): v ∈ T Mx } of T Mx × T My . By Proposition 3.1 we have that Aα (v, Lxy v)2  0
for all v ∈ T Mx , that is (Aα )|D  0, or equivalently λi  0 for i = 1, . . . , n. With our choice
of εα , this implies that

1 |λi | λi
λi + εα λ2i  λi + λ2i  λi + = 0
2(1 + sup1j n |λj |) 2 2

and since λi + εα λ2i , i = 1, . . . , n, are the eigenvalues of (Aα + εα A2α )|D , this means that
 
Aα + εα A2α (v, Lxy v)2  0.

Then condition (∗) implies that

 
P (v)2 − Q(Lxy v)2  Aα + εα A2α (v, Lxy v)2  0

for all v ∈ T Mx , which means that P  Lyx (Q). ✷

Therefore, if M has nonnegative curvature and F is degenerate elliptic then (∗) automatically
implies that

F (x, r, ζ, Lyx Q) − F (x, r, ζ, P )  0,

hence

F (y, r, Lxy ζ, Q) − F (x, r, ζ, P )

= F (y, r, Lxy ζ, Q) − F (x, r, ζ, Lyx Q) + F (x, r, ζ, Lyx Q) − F (x, r, ζ, P )
 F (y, r, Lxy ζ, Q) − F (x, r, ζ, Lyx Q),
330 D. Azagra et al. / J. Differential Equations 245 (2008) 307–336

and we see that condition (2) of the theorem is satisfied if we additionally require, for instance,
that
 
F (y, r, η, Q) − F (x, r, Lyx η, Lyx Q)  ω d(x, y) . (2♯)

Note that, in the case M = Rn we have Lyx η ≡ η and Lyx Q ≡ Q, and condition (2♯) simply
means that F (x, u, du, d 2 u) is uniformly continuous with respect to the variable x. Therefore
we can regard condition (2♯) as the natural extension to the Riemannian setting of the Euclidean
notion of uniform continuity of F with respect to x. This justifies the following

Definition 4.6. We will say that F : X → R is intrinsically uniformly continuous with respect to
the variable x if condition (2♯) above is satisfied.

Remark 4.7. As we saw in Example 2.14 above, many interesting examples of equations in-
volving nonincreasing symmetric functions of the eigenvalues of d 2 u (such as the trace and the
positive determinant det+ ) automatically satisfy condition (2♯) as long as they do not depend
on x. In fact, since the eigenvalues of A are the same as those of Lyx ALxy , any function of the
form
 
F (x, r, ζ, A) = G r, ζ x , eigenvalues of A

is intrinsically uniformly continuous with respect to x.

Therefore, for manifolds of nonnegative curvature, we do not need to impose that F depends
on d 2 u(x) in a uniformly continuous manner: the assumptions that F is degenerate elliptic and
intrinsically uniformly continuous with respect to x are sufficient. Let us sum up what we have
just shown.

Corollary 4.8. Let Ω be a bounded open subset of a complete finite-dimensional Riemannian

manifold M with nonnegative sectional curvature, and F : X → R be continuous, degenerate
elliptic, and satisfy:

(1) F is strongly increasing, that is there exists γ > 0 such that, if r  s then

γ (r − s)  F (x, r, ζ, Q) − F (x, s, ζ, Q);

(2) F is intrinsically uniformly continuous with respect to x (that is there exists a function
ω : [0, ∞] → [0, ∞] with limt→0+ ω(t) = 0 and such that
 
F (y, r, η, Q) − F (x, r, Lyx ζ, Lyx Q)  ω d(x, y) (2♯)

for all x, y, r, ζ, Q with d(x, y) < min{iM (x), iM (y)}).

Let u ∈ USC(Ω) be a subsolution and v ∈ LSC(Ω) a supersolution of F = 0 on Ω, and u  v

on ∂Ω.
Then u  v on all of Ω.
In particular, the Dirichlet problem (DP) has at most one viscosity solution.
 D. Azagra et al. / J. Differential Equations 245 (2008) 307–336 331

When M has negative curvature, condition (∗) does not imply P  Lyx Q, and degenerate
ellipticity together with fulfillment of (2♯) is not enough to ensure that condition (2) of The-
orem 4.2 is satisfied. In this case condition (2) of Theorem 4.2 involves kind of a uniform
continuity assumption on the dependence of F with respect to d 2 u(x). Let us be more explicit.

Proposition 4.9. Assume that M has sectional curvature bounded below by some constant
−K0  0. Then condition (∗) in Theorem 4.2 implies that

3
P − Lyx (Q)  K0 αd(x, y)2 I,
2

where I (v)2 = v, v = v2 .

Proof. We have that Aα = (α/2)d 2 ϕ(x, y), where ϕ(x, y) = d(x, y)2 . According to Proposi-
tion 3.3 we have

d 2 ϕ(x, y)(v, Lxy v)2  2K0 d(x, y)2 v2

for all v ∈ T Mx and x, y ∈ M with d(x, y) < min{iM (x), iM (y)}. Therefore

Aα (v, Lxy v)2  αK0 d(x, y)2 v2 .

This means that the maximum eigenvalue of the restriction of Aα to D := {(v, Lxy v): v ∈ T Mx },
which we denote λn , satisfies

λn  αK0 d(x, y)2 .

If λ1 , . . . , λn are the eigenvalues of (Aα )|D then λi + εα λ2i , i = 1, . . . , n, are those of (Aα +
εα A2α )|D . For a given i = 1, . . . , n, if λi  0 then λi + εα λ2i  0 as in the proof of Proposition 4.5.
In particular, if λn  0 then λj  0 for all j = 1, . . . , n, so we get that λj + εα λ2j  0 for all
j = 1, . . . , n, which means that
 
Aα + εα A2α (v, Lxy v)2  0.

On the other hand, if λn  0 then λn + εα λ2n  0, and because the function [0, +∞) ∋ s → s +
εα s 2 ∈ [0, +∞) is increasing, the maximum eigenvalue of (Aα +εα A2α )|D is precisely λn +εα λ2n .
This means that
 
Aα + εα A2α (v, Lxy v)2  λn + εα λ2n

for all v ∈ T Mx . Besides we have, by the choice of εα , that

1 λn 3
λn + εα λ2n  λn + λ2n  λn + = λn ,
2(1 + sup1j n |λn |) 2 2

hence
  3 3
Aα + εα A2α (v, Lxy v)2  λn  αK0 d(x, y)2 v2 .
2 2
332 D. Azagra et al. / J. Differential Equations 245 (2008) 307–336

In any case (no matter what the sign of λn is) we get that the above inequality holds. Therefore
condition (∗) implies
 2   3
P (v)2 − Q Lxy (v)  Aα + εα A2α (v, Lxy v)2  K0 αd(x, y)2 v2 . ✷
2
Corollary 4.10. Let M be a complete Riemannian manifold (no assumption on curvature), and
Ω be a bounded open subset of M. Suppose that F : X → R is proper, continuous, and satisfies
the following uniform continuity assumption: for every ε > 0 there exists δ > 0 such that

d(x, y)  δ, P − Lyx (Q)  δI ⇒ F (y, r, Lxy ζ, Q) − F (x, r, ζ, P )  ε (2♭)

for all x, y ∈ M with d(x, y) < iΩ , r ∈ R, ζ ∈ T Mx∗ , P ∈ L2s (T Mx , R), and Q ∈ L2s (T My , R).
Assume also that there is γ > 0 such that

γ (r − s)  F (x, r, ζ, Q) − F (x, s, ζ, Q) for all r  s.

Then there is at most one viscosity solution of the Dirichlet problem (DP).

Proof. Since M is complete, we know from the Hopf–Rinow Theorem that Ω is compact,
hence we have that iΩ = infx∈Ω iM (x) > 0. Take a number r with 0 < 2r < infx∈Ω iM (x).
Also by compactness of Ω, there exists K0 > 0 such that the sectional curvature is bounded
below by −K0 on Ω. Therefore we have that ϕ(x, y) = d(x, y)2 is C ∞ smooth on the set
{(x, y) ∈ Ω × Ω: d(x, y) < r} and, according to the preceding remark, if P , Q satisfy con-
dition (∗) of Theorem 4.2 we get P − Lyx (Q)  23 K0 α d(x, y)2 I whenever d(x, y) < r. Then
the uniform continuity assumption on F yields the existence of a function ω : [0, ∞] → [0, ∞]
with limt→0+ ω(t) = 0 and such that
     
F y, r, α exp−1 −1 2
y (x), Q − F x, r, −α expx (y), P  ω αd(x, y) + d(x, y) ,

hence the result follows from Theorem 4.2. ✷

Remark 4.11. As is usual with comparison principles, the proof of Theorem 4.2 can easily be
adapted to show that the viscosity solutions u of the equations F = 0 depend continuously on F .
That is, if u is solution of F = 0 and v is solution of G = 0, then
   
supu(x) − v(x)  sup F (x, r, ζ, A) − G(x, r, ζ, A) + sup |u − v|.
Ω (x,r,ζ,A)∈X ∂Ω

5. Comparison results without boundary conditions

The same argument as in the proof of Theorem 4.2, with some small changes, yields the
following.

Theorem 5.1. Let M be a connected, complete Riemannian manifold (without boundary) such
that i(M) > 0, and F : X → R be proper, continuous, and satisfy assumptions (1) and (2) of
Theorem 4.2. Let u be a subsolution, and v a supersolution, of F = 0. Assume that u and v are
uniformly continuous and limx→∞ u(x) − v(x)  0. Then u  v on M. In particular, if M is
compact, there is at most one viscosity solution of F = 0 on M.
 D. Azagra et al. / J. Differential Equations 245 (2008) 307–336 333

Uniform continuity and the inequality at infinity guarantee that the mα are attained, so the only
difference with the proof of Theorem 4.2 is that now we cannot assume that xα and yα converge
to some point x0 , but we do have that d(xα , yα ) < i(M) for large α, hence all the computations
and estimations in the proof of Theorem 4.2 are still valid.

Corollary 5.2. Let M be a compact Riemannian manifold of nonnegative sectional curvature,

and let F : X → R be continuous, degenerate elliptic, strongly increasing in u, and intrinsically
uniformly continuous with respect to x (that is, F satisfies conditions (1 − 2) of Corollary 4.8).
Let u be a subsolution, and v a supersolution of F = 0. Then u  v on M.

Proof. The same considerations as in Proposition 4.5 apply. ✷

Example 5.3. As we remarked above, condition (2) of Corollary 4.8 is easily satisfied when
F (x, r, ζ, A) does not depend on ζ and A themselves, but on ζ  and the eigenvalues of A. For
instance, the function
 3  5
F (x, r, ζ, A) = r − det+ (A) ζ 2 − f (x) trace(A)

satisfies (1) and (2) of the above corollary provided that f  0 and f is uniformly continuous.
Therefore the equation
  3
u − det+ D 2 u ∇u2 − (u)5 f = 0

has at most one viscosity solution on any compact manifold of positive curvature if we only
require that f is continuous and nonnegative.

Corollary 5.4. Let M be a compact Riemannian manifold (no assumption on curvature). Sup-
pose that F : X → R satisfies the uniform continuity assumption, and the growth assumption, of
Corollary 4.10. Let u be a subsolution, and v a supersolution of F = 0. Then u  v on M. In
particular there is at most one viscosity solution of F = 0.

Proof. Since M is compact the sectional curvature of M is bounded on all M, say K  −K0 .
Take a number r with 0 < 2r < i(M). The function ϕ(x, y) = d(x, y)2 is C ∞ on the set
{(x, y) ∈ M × M: d(x, y)  2r}. Suppose that P and Q satisfy (∗) of Theorem 4.2. Then,
from Proposition 4.9, we get that P − Lyx Q  32 K0 α d(x, y)2 I provided that d(x, y) < r.
Therefore the uniform continuity property of F gives us a function ω : [0, ∞] → [0, ∞] with
limt→0+ ω(t) = 0 and such that
     
F y, r, α exp−1 −1 2
y (x), Q − F x, r, −α expx (y), P  ω αd(x, y) + d(x, y) .

Hence we can apply Theorem 5.1 and conclude the result. ✷

6. Existence results

Perron’s method can easily be adapted to the Riemannian setting to establish existence of
viscosity solutions to the Dirichlet problem. The proof goes exactly as in [4] with appropriate
changes. The only step which is not completely obvious is the proof of the following
334 D. Azagra et al. / J. Differential Equations 245 (2008) 307–336

Proposition 6.1. Let (ζ, A) ∈ J 2,+ f (z) Suppose that fn is a sequence of upper semicontinuous
functions such that

(i) there exists xn such that (xn , fn (xn )) → (x, f (x)), and
(ii) if yn → y, then lim supn→∞ fn (yn )  f (y).

Then there exist x̂n and (ζn , An ) ∈ J 2,+ fn (x̂n ) such that (x̂n fn (x̂n ), ζn , An ) → (x, f (x), ζ, A).

Proof. Consider the functions f ◦ expx and fn ◦ expx defined on a neighborhood of 0 in T Mx .

These functions satisfy properties (i) and (ii) of the statement (when they take the roles of f and
fn and M is replaced with T Mx ). By Corollary 2.3 we have that (ζ, A) ∈ J 2,+ (f ◦ expx )(0).
And of course the result is known in the case when M = Rn , so we get a sequence v̂n and
(ζ̃n , An ) ∈ J 2,+ (fn ◦ expx )(v̂n ) such that
   
v̂n , fn ◦ expx (v̂n ), ζ̃n , An → 0, f ◦ expx (0), ζ, A .

Set x̂n = expx (v̂n ). We have that x̂n → x and fn (x̂n ) → f (x). Since (ζ̃n , An ) ∈ J 2,+ (fn ◦
expx )(v̂n ) there exist functions ψn such that fn ◦ expx −ψn attains a maximum at v̂n , ζ̃n =
dψn (v̂n ) and An = d 2 ψn (v̂n ). Let us define ϕ = ψn ◦ exp−1 x on a neighborhood of x. Then
fn − ϕn attains a maximum at x̂n so, if we set ζn = dϕ(x̂n ), An = d 2 ϕ(x̂n ), we have that
(ζn , An ) ∈ J 2,+ fn (x̂n ). It only remains to show that ζn → ζ and An → A. But this is exactly
what was shown in (⇐) of the proof of Proposition 2.8. ✷

By using this proposition one can prove, as in [4], existence of viscosity solutions to the
Dirichlet problem
 
F x, u, du, d 2 u = 0 in Ω, u = f on ∂Ω, (DP)

where Ω is an open bounded subset of a complete Riemannian manifold M.

Theorem 6.2. Let comparison hold for (DP), i.e., if w is a subsolution of (DP) and v is a super-
solution of (DP), then w  v. Suppose also that there exists a subsolution u and a supersolution
u of (DP) that satisfy the boundary condition u∗ (x) = u∗ (x) = f (x) for x ∈ ∂Ω. Then

W (x) = sup w(x): u  w  u and w is a subsolution of (DP)

is a solution of (DP).

Here we used the following notation:

u∗ (x) = lim sup u(y): y ∈ Ω and d(y, x)  r ,

r↓0

u∗ (x) = lim inf u(y): y ∈ Ω and d(y, x)  r ,

r↓0

that is u∗ denotes the upper semicontinuous envelope of u (the smallest upper semicontinuous
function, with values in [−∞, ∞], satisfying u  u∗ ), and similarly u∗ stands for the lower
semicontinuous envelope of u.
 D. Azagra et al. / J. Differential Equations 245 (2008) 307–336 335

One can also easily adapt the proof of [2, Theorem 6.17] to the second order situation, obtain-
ing the following.

Corollary 6.3. Let M be a compact Riemannian manifold, and G(x, du, d 2 u) be degenerate
elliptic and uniformly continuous in the sense of Corollary 4.10. Then there exists a unique
viscosity solution of the equation u + G(x, du, d 2 u) = 0 on M.

Again, if M has nonnegative curvature, the assumptions that F is elliptic and intrinsically
uniformly continuous with respect to x are sufficient in order to get an analogous result.

7. Examples

Most of the examples of proper F ’s given in [4] remain valid in the Riemannian setting. In
particular, as we have already seen, the functions (x, r, ζ, A) → − det+ (A) and (x, r, ζ, A) →
− trace(A) are degenerate elliptic and intrinsically uniformly continuous with respect to x. The
same is true of all many symmetric functions of the eigenvalues of A, such as minus the minimum
(or the maximum) eigenvalue, and of course nondecreasing combinations and sums of these are
degenerate elliptic too. One can find lots of examples of nonlinear equations for which the results
of this paper yield existence and uniqueness of viscosity solutions. For instance, one can easily
show that, for every compact manifold of positive curvature, the equation
    2k+1 2
max u − λ1 D 2 u ∇up − (u)2q+1 ∇ur − det+ D 2 u f , u−g =0

(where λ1 denotes the minimum eigenvalue function and p, q, r, k ∈ N) has a unique viscosity
solution if we only require that f and g are continuous. This gives an idea of the generality of
the above results.
Of course this example is rather unnatural. Let us finish this paper by examining what our
results yield in the case of a classic equation, that of Yamabe’s, which has been extensively
studied and completely solved by using variational methods. We do not claim that the following
discussion gives any new insight into Yamabe’s problem, we only want to study, from the point
of view of the viscosity solutions theory, a well-known example of a nonlinear equation arising
from an important geometrical problem.

Example 7.1 (The Yamabe equation). A fundamental problem in conformal geometry is to know
whether or not there exists a conformal metric g ′ with constant scalar curvature S ′ on a given
compact n-dimensional Riemannian manifold (M, g), with n  3, see [1,12]. This is equivalent
to solving the equation

n−1 n+2
−4 u + S(x)u = S ′ u n−2 , (Y )
n−2
where S is the scalar curvature of g. One can write this equation in the form F = 0, where

n+2 n−1
F (x, r, ζ, A) = S(x)r − S ′ r n−2 − 4 trace(A) = 0.
n−2

It is clear that F is degenerate elliptic. Assume that S is everywhere positive and that S ′  0.
Then, by compactness, there exists γ > 0 such that S(x)  γ for all x ∈ M. According to Re-
336 D. Azagra et al. / J. Differential Equations 245 (2008) 307–336

mark 4.4, in order to check conditions (1) and (2) of Theorem 5.1 we may assume that r, s lie on
a bounded interval. We have that
 
F (y, r, η, Q) − F (x, r, Lyx η, Lyx Q)  r S(y) − S(x),

hence, because S is uniformly continuous on M and r is bounded, we deduce that F satisfies (2)
of Corollary 5.2. On the other hand, if r  s then
 n+2 n+2 
F (x, r, ζ, A) − F (x, s, ζ, A) = S(x)(r − s) − S ′ r n−2 − s n−2  γ (r − s),

so condition (1) is also satisfied. It follows that there is at most one viscosity solution of F = 0.
Existence can be shown by using Perron’s method. In all, we see that if S is everywhere positive
and S ′  0 then there exists a unique viscosity solution u of (Y ).

Acknowledgment

We thank the referee for his useful suggestions which helped us improve the exposition.

References

[1] T. Aubin, Some Nonlinear Problems in Riemannian Geometry, Springer Monogr. Math., Springer-Verlag, Berlin,
1998.
[2] D. Azagra, J. Ferrera, F. López-Mesas, Nonsmooth analysis and Hamilton–Jacobi equations on Riemannian mani-
folds, J. Funct. Anal. 220 (2) (2005) 304–361.
[3] M.P. do Carmo, Riemannian Geometry, Math. Theory Appl., Birkhäuser, Boston, 1992.
[4] M.G. Crandall, H. Ishii, P.-L. Lions, User’s guide to viscosity solutions of second order partial differential equations,
Bull. Amer. Math. Soc. 27 (1992) 1–67.
[5] R. Deville, E. Matheron, Infinite games, Banach space geometry and the eikonal equation, Proc. London Math. Soc.
(3) 95 (1) (2007) 49–68.
[6] M.J. Gursky, J.A. Viaclovsky, Prescribing symmetric functions of the eigenvalues of the Ricci tensor, Ann. of Math.
(2) 166 (2) (2007) 475–531.
[7] S. Lang, Fundamentals of Differential Geometry, Grad. Texts in Math., vol. 191, Springer-Verlag, New York, 1999.
[8] Yu.S. Ledyaev, Q.J. Zhu, Nonsmooth analysis on smooth manifolds, Trans. Amer. Math. Soc. 359 (8) (2007) 3687–
3732.
[9] C. Mantegazza, A.C. Mennucci, Hamilton–Jacobi equations and distance functions on Riemannian manifolds, Appl.
Math. Optim. 47 (1) (2003) 1–25.
[10] T. Sakai, Riemannian Geometry, Transl. Math. Monogr., vol. 149, Amer. Math. Soc., 1992.
[11] M. Spivak, A Comprehensive Introduction to Differential Geometry, vol. IV, second ed., Publish or Perish, Wilm-
ington, DE, 1979.
[12] J. Viaclovsky, Conformal geometry and fully nonlinear equations, World Sci. Memorial Volume for S.S. Chern,
2006.