Advances in Mathematics 281 (2015) 353–393

Contents lists available at ScienceDirect

Advances in Mathematics
www.elsevier.com/locate/aim

The Calderón–Zygmund inequality and Sobolev

spaces on noncompact Riemannian manifolds
Batu Güneysu a , Stefano Pigola b,∗
a
Institut für Mathematik, Humboldt-Universität zu Berlin, 12489 Berlin, Germany
b
Dipartimento di Scienza e Alta Tecnologia – Sezione di Matematica,
Università dell’Insubria, 22100 Como, Italy

a r t i c l e i n f o a b s t r a c t

Article history: We introduce the concept of Calderón–Zygmund inequalities

Received 21 July 2014 on Riemannian manifolds. For 1 < p < ∞, these are
Received in revised form 12 inequalities of the form
February 2015
Accepted 21 March 2015
Available online 28 May 2015 Hess(u)Lp ≤ C1 uLp + C2 ΔuLp ,
Communicated by Camillo De Lellis

MSC: valid a priori for all smooth functions u with compact

53C20 support, and constants C1 ≥ 0, C2 > 0. Such an inequality
46E35 can hold or fail, depending on the underlying Riemannian
58J50 geometry. After establishing some generally valid facts and
consequences of the Calderón–Zygmund inequality (like new
Keywords: denseness results for second order Lp -Sobolev spaces and
Global Calderón–Zygmund
gradient estimates), we establish sufficient geometric criteria
inequality
Sobolev spaces
for the validity of these inequalities on possibly noncompact
Noncompact Riemannian manifold Riemannian manifolds. These results in particular apply to
many noncompact hypersurfaces of constant mean curvature.
© 2015 Elsevier Inc. All rights reserved.

Contents

1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 354

* Corresponding author.
E-mail addresses: gueneysu@math.hu-berlin.de (B. Güneysu), stefano.pigola@uninsubria.it (S. Pigola).

http://dx.doi.org/10.1016/j.aim.2015.03.027
0001-8708/© 2015 Elsevier Inc. All rights reserved.
354 B. Güneysu, S. Pigola / Advances in Mathematics 281 (2015) 353–393

2. Setting and notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 358

3. Applications of the Calderón–Zygmund inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . . 360
3.1. Sobolev spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 360
3.2. New first and second order estimates from CZ(p) . . . . . . . . . . . . . . . . . . . . . . . . 363
4. Criteria for the validity of the Calderón–Zygmund inequality . . . . . . . . . . . . . . . . . . . . . 366
4.1. Local results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 366
4.2. Stability under compact perturbations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 370
4.3. Global results for CZ(2) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 371
4.4. Global results for CZ(p) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 380
4.5. Calderón–Zygmund inequalities on H-hypersurfaces . . . . . . . . . . . . . . . . . . . . . . 384
Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 386
Appendix A. Doubling vs nonnegative Ricci . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 387
Appendix B. Harmonic coordinates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 389
Appendix C. Gluing Riemannian manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 391
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 392

1. Introduction

Let M be a smooth possibly noncompact Riemannian m-manifold without boundary.

For an arbitrary p ∈ (1, ∞), let us consider the following canonically given problems for
second order Sobolev spaces on M , on the Lp -scale:

• Problem 1: Under which (geometric) assumptions on M does one have the denseness
H2,p
0 (M ) = H
2,p
(M )?
• Problem 2: Under which assumptions on M does one have the implication

f ∈ Lp (M ) ∩ C2 (M ), Δf ∈ Lp (M ) ⇒ f ∈ H2,p (M )

(that is, |Hess(f )| ∈ Lp (M ))?

• Problem 3: Under which assumptions on M does one have an inequality of the form

grad(f )Lp ≤ C(Δf Lp + f Lp ) for all f ∈ C∞

c (M )? (1)

Let us note here that Problem 1, the denseness of C∞ c (M ) in H

2,p
(M ), is a classical
problem which has been treated systematically in [18]. Here, we would like to stress
the fact that without a lower control on the injectivity results, nothing seems to be
known so far for the case p = 2. Furthermore, Problem 2 is obviously concerned with
Lp -estimates for solutions of the Poisson equation on M , and Problem 3 arises naturally
in Rellich–Kondrachov type compactness arguments: Here, typically one has given a
sequence of functions {fn } ⊂ C∞ c (M ) such that

sup max(fn  , Δfn ) < ∞,

n∈N

and one would like to know whether the sequence {fn } is bounded in H1,p (M ).
It turns that there is an inequality underlying all three problems simultaneously:
 B. Güneysu, S. Pigola / Advances in Mathematics 281 (2015) 353–393 355

Definition (Calderón–Zygmund inequality). Let 1 < p < ∞. We say that M ≡ (M, g)

satisfies the Lp -Calderón–Zygmund inequality, or in short CZ(p), if there are C1 ≥ 0,
C2 > 0, such that for all u ∈ C∞
c (M ) one has

Hess(u)Lp ≤ C1 uLp + C2 ΔuLp . (2)

Obviously, if the Calderón–Zygmund CZ(p) holds on M then, by restriction, the same

inequality holds on any open subset of M and with the same constants. Furthermore,
we have directly excluded the extremal cases p = 1 and p = ∞ in the definition since
the corresponding hypothetical elliptic estimates fail [25,11] for the Euclidean Laplace

operator i ∂i2 , thus (2) with p = 1 or p = ∞ cannot hold in general. Let us also remark
that in the Euclidean Rm , this inequality can be proved [15] by estimates on singular
integral operators which have been proved by Calderón and Zygmund [4]. Ultimately,
this was the motivation for us for calling1 CZ(p) the “Calderón–Zygmund inequality”.
Now one has the following implications which link CZ(p) to the above Problems 1–3:

(a) It has been observed in [17] that under CZ(p), if M is geodesically complete and
admits a sequence of Laplacian cut-off functions (this is the case e.g. if M has a
nonnegative Ricci curvature), then one has H2,p
0 (M ) = H
2,p
(M ).
(b) In Proposition 3.8 we show that under CZ(p), if M is geodesically complete and
admits a sequence of Hessian cut-off functions (this is the case e.g. if M has a
bounded curvature tensor), then any f ∈ H1,p (M ) ∩C2 (M ) with Δf ∈ Lp (M ) satisfies
f ∈ H2,p (M ).
(c) In Corollary 3.11 we prove that, on a geodesically complete manifold, CZ(p) always
implies (1).

Here, the results (a) and (b) follow from the existence of appropriate second order
cut-off functions (see also [17]), which we prove to exist under very weak assumptions
on the curvature and without positive injectivity radius. In this context, we will also
establish our first main result, Theorem 3.7 below, stating that for a geodesically complete
manifold M with a bounded curvature tensor one has H2,p 0 (M ) = H
2,p
(M ) for all 1 <
p < ∞. This result is entirely new for p = 2, for it does not require a positive injectivity
radius (cf. [18]).
We also mention that the statement (c) makes use of an appropriate Lp -interpolation
result, which should be of an independent interest (cf. Proposition 3.10).
In summary, these observations clearly motivate a systematic treatment of the follow-
ing problem:

Under which geometric assumptions on M does one have CZ(p), and how do the
CZ(p)-constants C1 , C2 depend on the underlying geometry?

1
The authors would like to thank Klaus Ecker in this context.
356 B. Güneysu, S. Pigola / Advances in Mathematics 281 (2015) 353–393

Let us start by taking a look at the local situation: We first prove in Theorem 4.1
that one always has CZ(p) on relatively compact domains Ω ⊂ M , where one can even
pick C1 = 0, if Ω has a smooth boundary. Whence, using a gluing procedure which again
relies on Lp -interpolation, we obtain that CZ(p) is stable under compact perturbations
(this is contained in Theorem 4.3). In particular, we have:

Theorem A. Let 1 < p < ∞. A complete Riemannian manifold supports CZ(p), if and
only if all its ends (with respect to any smooth, compact domain Ω) support CZ(p). In par-
ticular, CZ(p) holds on the Riemannian connected sum M = M1 #M2 of m-dimensional
complete Riemannian manifolds M1 and M2 if and only if CZ(p) holds on both M1
and M2 .

However, as one might expect, in the above cases the CZ(p)-constants depend rather
implicitly on the underlying geometry, which raises the question of more precise estimates
on geodesic balls: This problem is considered in Proposition 4.2, where we prove that
CZ(p) holds on sufficiently small geodesic balls, with constants only depending on the
radius, m = dim M , p, and a lower bound of an appropriate harmonic radius.
As for global results, we found the following rather complete answer for the p = 2 case:

Theorem B. a) If M has Ricci curvature bounded from below by a constant −D with

D > 0, then one has (a stronger infinitesimal variant) of CZ(2), with constants only
depending on D.
b) There exists a 2-dimensional geodesically complete Riemannian manifold N with
unbounded Gaussian curvature, such that CZ(2) fails on N .

Note here that Theorem B a) does not require any completeness and is in fact a simple
consequence of Bochner’s identity. The proof of part b) is rather involved, and is done
by explicitly constructing a strongly curved parabolic model surface.
For p = 2, things are more complicated. In this context, we prove the following two
results, which can also be considered as the main results of this paper:

Theorem C. Let 1 < p < ∞ and assume that M has a bounded Ricci curvature and a
positive injectivity radius.2 Then one has CZ(p), with constants depending only on m, p,
Ric∞ and the injectivity radius.

Theorem D. Let 1 < p ≤ 2. Assume that M is geodesically complete with a first order
bounded geometry, and that there are constants D ≥ 1, 0 ≤ δ < 2 with

δ
+r δ
vol(Btr (x)) ≤ DtD et vol(Br (x)) for all x ∈ M, r > 0, t ≥ 1. (3)

2
Thus M is automatically geodesically complete.
 B. Güneysu, S. Pigola / Advances in Mathematics 281 (2015) 353–393 357

Then one has CZ(p), with constants depending only on m, p, R∞ , ∇R∞ , D, δ, with
R the curvature tensor.

The proof of Theorem C uses appropriate elliptic estimates, in combination with

harmonic-radius bounds on the Riemannian structure, a method which requires bounds
on the injectivity radius, but has the advantage of working for arbitrary p. The latter
result in particular applies to noncompact hypersurfaces of constant mean curvature, as
we will prove:

Theorem E. Let f : M → Mm+1 (c) be a complete, noncompact, oriented, H-hypersurface

with finite total curvature, embedded into the complete, simply connected space-form
Mm+1 (c) of constant sectional curvature c ≤ 0. Then M satisfies the assumptions of
Theorem C, in particular, for every 1 < p < ∞, the Calderón–Zygmund inequality CZ(p)
holds on M .

The proof of Theorem D is very different from that of Theorem C: It uses deep
boundedness-results on covariant Riesz-transforms (cf. Proposition 4.18) which are es-
sentially due to Thalmaier and Wang [29], and which ultimately follow from covariant
probabilistic heat-semigroup derivative formulas. This technique makes it possible to
avoid assumptions on the injectivity radius. We stress the fact our proof uses bound-
edness results on the more complicated Riesz-transforms of the type ∇(Δ1 + a)−1/2 on
1-forms, and not simply d(Δ1 + a)−1/2 , the latter of which being rather simple to obtain
as d and Δ1 commute. Let us also note that under geodesic completeness, the generalized
doubling assumption (3) is implied by Ric ≥ 0 (though Ric ≥ 0 is not necessary at all
for (3); cf. Appendix A).
The paper is organized as follows:
In Section 2 we establish some Riemann geometric notation.
Section 3 is devoted to the applications of Calderón–Zygmund inequality. Here, in
Section 3.1 we establish some new facts concerning the second order Lp -Sobolev spaces,
as well as their connections to the Calderón–Zygmund inequality. In particular, the initial
Problem 1, respectively, its solution (a) is considered here. Section 3.2 is concerned with
producing gradient estimates, as well as global estimates for Poisson’s equation by using
the Calderón–Zygmund inequality (in particular, it is dealt with the above pair given by
(Problem 2, (b)) and the pair given by (Problem 3, (c)). It is also here where we prove
the above mentioned interpolation result (Proposition 3.10).
In Section 4 we prove several geometric criteria for the validity of CZ(p): Here, we
start in Section 4.1 with generally valid (that is, without assuming curvature bounds)
local, as well as stability results. In particular, the reader may find the proof of Theo-
rem A there. The rest of this section is devoted to curvature criteria that imply global
Calderón–Zygmund inequalities. In particular, in Section 4.3 the p = 2 case is consid-
ered, where we prove the above Theorem B, and in Section 4.4 the general p situation
is considered, with the above Theorem C and Theorem D. Section 4.5 is devoted to
358 B. Güneysu, S. Pigola / Advances in Mathematics 281 (2015) 353–393

H-hypersurfaces with the above Theorem E, as well as an abstraction of the latter re-
sult.
We have also included an appendix, where an example of a Riemannian manifold
with the volume assumption (3) but without Ric ≥ 0 is given, and where some facts on
harmonic-radius bounds as well as on abstract Riemannian gluings have been collected,
for the convenience of the reader.

2. Setting and notation

We fix an arbitrary smooth Riemannian m-manifold3 M ≡ (M, g) with ∇ the Levi-

Civita connection on M . We will denote the corresponding distance function with d(•, •),
the open balls with Ba (x), x ∈ M , a > 0, and the volume measure with μ(dx) := vol(dx),

where whenever there is no danger of confusion we shall simply write f dμ instead of

M
f dμ. The symbol rinj (x) ∈ (0, ∞] will stand for the injectivity radius at x, with

rinj (M ) := inf rinj (x) ∈ [0, ∞]

x∈M

the global injectivity radius.

Let us develop some further geometric notation which will be used in the sequel:
If E → M is a smooth Euclidean vector bundle, then whenever there is no danger of
 we will denote the underlying Euclidean structure simply with (•, •)x , x ∈ M ,
confusion
|•|x := (•, •)x will stand for the corresponding norm on Ex . Using μ, for any 1 ≤ p ≤ ∞
we get the corresponding Lp -spaces of equivalence classes of Borel sections ΓLp (M, E),
with their norms
    1/p
f (x)p μ(dx) , if p < ∞
f p := M x
inf{C | C ≥ 0, |f | ≤ C μ-a.e.}, if p = ∞.

The symbol •, • will stand for the canonical inner product on the Hilbert space
ΓL2 (M, E), and ‘†’ will denote the formal adjoint with respect to •, • of a smooth
linear partial differential operator that acts on some E → M as above.
We equip T∗ M with its canonical Euclidean structure

(α1 , α2 ) := g(α1 , α2 ), α1 , α2 ∈ ΓC∞ (M, T∗ M ),

where αj stands for the vector field which is defined by α in terms of g. This produces
canonical Euclidean metrics on all bundles of k-times contravariant and l-times covariant
tensors Tk,l M → M . Next, ∇ induces a Euclidean covariant derivative on T∗ M = T0,1 →
M through

3
In the sequel a manifold will always be understood to be without boundary, unless otherwise stated.
 B. Güneysu, S. Pigola / Advances in Mathematics 281 (2015) 353–393 359

∇X1 α(X2 ) := X1 (α(X2 )) − α(∇X1 X2 ),

for any smooth 1-form α and any smooth vector fields X1 , X2 on M , which of course
means nothing but (∇X1 α) = ∇X1 α , and these data are tensored to give a Euclidean
covariant derivative on Tk,l M → M which, by the above abuse of notation, is always
denoted with ∇. Thus, for any u ∈ C∞ (M ) we have

Hess(u) := ∇du ∈ ΓC∞ (M, T0,2 ).

The gradient grad(u) ∈ ΓC∞ (M, TM ), is defined by

(grad(u), X) := du(X) for any X ∈ ΓC∞ (M, TM ),

with

d : ΓC∞ (M, ∧• T∗ M ) −→ ΓC∞ (M, ∧•+1 T∗ M )

the exterior differential, and where as usual 0-forms are identified with functions. Then
the divergence div(X) ∈ C∞ (M ) of a smooth vector field X on M is given by div(X) =
d† X  , where X  stands for the 1-form which is defined by X in terms of g = (•, •). Let
us denote with

Δ• := d† d + dd† : ΓC∞ (M, ∧• T∗ M ) −→ ΓC∞ (M, ∧• T∗ M )

the Laplace–Beltrami operator on differential forms. Note that our sign convention is
such that Δ• is nonnegative, and the Friedrichs realization of Δ• in ΓL2 (M, ∧• T∗ M )
will be denoted with the same symbol. In the sequel, we will freely use the formulas

Hess(u1 u2 ) = u2 Hess(u1 ) + du1 ⊗ du2 + du2 ⊗ du1 + u1 Hess(u2 ),

Δu1 = −tr(Hess(u1 )) = −div(grad(u1 )),
Δ(u1 u2 ) = Δ(u1 )u2 − 2(grad(u1 ), grad(u2 )) + Δ(u2 )u1 ,
div(u1 X) = (grad(u1 ), X) + u1 div(X),

valid for all smooth functions u1 , u2 and smooth vector fields X on M .

We close this section with some conventions and notation which concerns curvature
data: The curvature tensor R is read as a section

R ∈ ΓC∞ (M, T1,3 M ),

given for smooth vector fields X, Y , Z by

R(X, Y, Z) := ∇X ∇Y Z − ∇Y ∇X Z − ∇[X,Y ] Z ∈ ΓC∞ (M, TM ).

360 B. Güneysu, S. Pigola / Advances in Mathematics 281 (2015) 353–393

Then the Ricci curvature

Ric ∈ ΓC∞ (M, T0,2 M )

is the section given by the fiberwise trace

Ric(Z, Y ) = tr(X → R(X, Y, Z)) ∈ C∞ (M ).

If m ≥ 2, then for any x ∈ M , the sectional curvature of a 2-dimensional subspace

A = span(X, Y ) of Tx M is well-defined by

(R(X, Y, Y ), X)
Sec(A) := ∈ R.
|X ∧ Y |2

Finally, we mention that whenever we write C = C(a1 , . . . , al ) for a constant, this means
that C only depends on the parameters a1 , . . . , al , and nothing else.

3. Applications of the Calderón–Zygmund inequality

3.1. Sobolev spaces

Let us start with some remarks concerning the connection between CZ(p) and second
order Lp -Sobolev spaces. In fact, precisely this context was the original motivation for
our study of Calderón–Zygmund inequalities. To this end, we first list some conventions
and notation on Sobolev spaces: For any 1 ≤ p < ∞, the Banach space H1,p (M ) is
defined by

H1,p (M ) := {u | u ∈ Lp (M ), |grad(u)| ∈ Lp (M ) as distr.} ,

with its natural norm

u1,p := up + grad(u)p .

Likewise, one has the Banach space

H2,p (M ) := {u | u ∈ Lp (M ), |grad(u)|, |Hess(u)| ∈ Lp (M ) as distr.} ,

with its natural norm u2,p . By a generalized Meyers–Serrin type theorem [16], one has
that the linear space

C∞ (M ) ∩ Hk,p (M ) is dense in Hk,p (M )

(a fact which is actually true for all k ∈ N with the natural definition of higher order
Sobolev spaces). Finally, we define Hk,p0 (M ) ⊂ H
k,p
(M ) as usual to be the closure of
∞ k,p
Cc (M ) in H (M ).
 B. Güneysu, S. Pigola / Advances in Mathematics 281 (2015) 353–393 361

Remark 3.1. Let 1 ≤ p < ∞.

1. Every u ∈ H2,p (M ) satisfies Δu ∈ Lp (M ) in the sense of distributions. Indeed,

integrating by parts and using

Hess† ◦ tr† = (tr ◦ Hess)† ,

where we consider tr(•) as a smooth zeroth order linear differential operator, one
gets that the distribution Δu is in fact a Borel function which coincides with

Δu(x) = −trx (Hess(u)|x ),

so that
√
|Δu| ≤ m |Hess(u)| μ-a.e. in M.

∞
2. If u ∈ H2,p
0 (M ), and if {uk } ⊂ Cc (M ) is a sequence such that u − uk 2,p → 0, then
obviously {uk } is Cauchy in Lp (M ), {Hess(uk )} is Cauchy in ΓLp (M, T0,2 M ), and
using
√
|Δψ| ≤ m |Hess(ψ)|, for all ψ ∈ C∞ (M ), (4)

it also follows that {Δuk } is Cauchy in Lp (M ). In particular, one necessarily has

u − uk p → 0, Hess(u) − Hess(uk )p → 0, Δu − Δuk p → 0.

Remark 3.1.2 immediately implies that Calderón–Zygmund inequalities always extend

to H2,p
0 (M ) in the following sense:

Corollary 3.2. Let 1 < p < ∞. If one has (2), then this inequality extends from C∞
c (M )
to H2,p
0 (M ) with the same constants.

The following definition will be convenient (cf. [17]):

Definition 3.3. a) M is said to admit a sequence (χn ) ⊂ C∞

c (M ) of Laplacian cut-off
functions, if (χn ) has the following properties:

(C1) 0 ≤ χn (x) ≤ 1 for all n ∈ N, x ∈ M ,

(C2) for all compact K ⊂ M , there is an n0 (K) ∈ N such that for all n ≥ n0 (K) one
has χn |K = 1,
(C3) supx∈M |dχn (x)|x → 0 as n → ∞,
(C4) supx∈M |Δχn (x)| → 0 as n → ∞.
362 B. Güneysu, S. Pigola / Advances in Mathematics 281 (2015) 353–393

b) M is said to admit a sequence (χn ) ⊂ C∞c (M ) of Hessian cut-off functions, if (χn )

has the above properties (C1), (C2), (C3), and in addition

(C4 ) supx∈M |Hess(χn )(x)|x → 0 as n → ∞.

Remark 3.4. By (4), any sequence of Hessian cut-off functions is automatically a sequence
of Laplacian cut-off functions.

One has:

Proposition 3.5. a) If M is geodesically complete with Ric ≥ 0, then M admits a sequence

of Laplacian cut-off functions.
b) If M is geodesically complete with R∞ < ∞, then M admits a sequence of
Hessian cut-off functions.

Proof. a) This result is included in [17]. It relies on a rigidity result by Cheeger and
Colding [6] (see also [30]).
b) By a result of L.-F. Tam,4 see Proposition 26.49 in [9], one has that there is a
constant C = C(R∞ , m) > 0, such that for any x0 ∈ M there is a smooth function
d˜ = d˜x0 : M → [0, ∞) satisfying

d(•, x0 ) + 1 ≤ d˜ ≤ d(•, x0 ) + C, |grad(d)|

˜ ≤ C, |Hess(d)|
˜ ≤ C. (5)

Pick now a smooth function t : R → [0, 1] which is compactly supported, equal to 1 in

˜
[0, C + 12 ], and zero on [C + 1, ∞). Then χn (x) := t(d(x)/n) has the required proper-
ties. 2

Proposition 3.6. a) Assume that (2) holds for some 1 < p < ∞ and that M admits a
sequence of Laplacian cut-off functions. Then one has H2,p
0 (M ) = H
2,p
(M ), in particular
2,p
(by Corollary 3.2), (2) extends to H (M ) with the same constants.
b) If M admits a sequence of Hessian cut-off functions, then one has H2,p 0 (M ) =
H2,p (M ) for all 1 < p < ∞.

Proof. Part a) has been observed in [17]. Part b) follows from the same argument: Given a
smooth f ∈ H2,p (M ), pick a sequence (χn ) of Hessian cut-off functions. With fn := χn f ,
using

Hess(fn ) = f Hess(χn ) + dχn ⊗ df + df ⊗ dχn + χn Hess(f ),

one easily gets that fn converges to f in H2,p (M ). 2

4
A completely different construction of an exhaustion function with bounded gradient and Hessian is also
contained in [7].
 B. Güneysu, S. Pigola / Advances in Mathematics 281 (2015) 353–393 363

As a consequence, we point out the following density result whose relevance relies on
the fact that no injectivity assumption is required on underlying complete manifold.

Theorem 3.7. If M is geodesically complete with R∞ < ∞, then one has H2,p
0 (M ) =
H2,p (M ) for all 1 < p < ∞.

3.2. New first and second order estimates from CZ(p)

Let us continue with a connection between Calderón–Zygmund inequalities and global

control of solutions to the Poisson equation: Classically, one uses local Lp -Calderón–
Zygmund type inequalities in order to get higher local regularity of solutions of the
1,p
Poisson equation Δu = f . Indeed, if u is in Wloc and f is in Lploc then a local Calderón–
Zygmund type inequality shows ∂i ∂j u is in Lploc , proving that u is in Wloc2,p
. In fact, one
can use this way of concluding to derive global estimates to solutions of the Poisson
equation:

Proposition 3.8. Let 1 < p < ∞, and assume that M satisfies CZ(p) and admits a se-
quence of Hessian cut-off functions. Let u ∈ C2 (M ) be a solution of the Poisson equation

Δu = f.

If u, |grad(u)|, f ∈ Lp (M ) then |Hess(u)| ∈ Lp (M ), and one has

Hess(u)p ≤ Cf p ,

where C > 0 only depends on the CZ(p) constants.

Remark 3.9. It is not completely clear to what extent the Lp -assumption on the gradient
is technical and related to the method of proof. In any case, it would be interesting to
find situations where it is automatically satisfied. Compare also with Corollary 3.11.

Proof of Proposition 3.8. Pick a sequence of Hessian cut-off functions (ϕk ) and define
the corresponding sequence of compactly supported C2 -functions uk := uϕk . Then using

Hess(uk ) = uHess(ϕk ) + dϕk ⊗ du + du ⊗ dϕk + ϕk Hess(u),

Δuk = uΔϕk − 2(grad(u), grad(ϕk )) + ϕk Δu,

and applying to Hess(uk ) the Calderón–Zygmund inequality, we obtain (using Corol-

lary 3.2)
 
ϕk Hess(u)p ≤ C uHess (ϕk )p + |grad(u)| |grad(ϕk )|p + uΔϕk p + ϕk Δup .

Whence, taking the limit as k → ∞, the claim follows from Fatou’s Lemma. 2
364 B. Güneysu, S. Pigola / Advances in Mathematics 281 (2015) 353–393

In order to prove our next application of CZ(p), a gradient estimate, we will need the
following Lp -interpolation result, which should be of an independent interest, and which
will also be used later on to prove local CZ(p) inequalities:

Proposition 3.10. a) For any 2 ≤ p < ∞ there is a constant C = C(p) > 0 such that for
any ε > 0, u ∈ C∞
c (M ) one has

C
grad(u)p ≤ up + Cε Hess(u)p . (6)
ε

b) Assume that either M is geodesically complete or that M is a relatively compact

open subset of an arbitrary smooth Riemannian manifold. Then for any 1 < p ≤ 2 there
is a constant C = C(p) > 0 such that for any ε > 0, u ∈ C∞
c (M ) one has

C
grad(u)p ≤ up + CεΔup . (7)
ε

Proof. a) Let u ∈ C∞c (M ) and, having fixed α > 0, consider the smooth, compactly
supported vector field

  p−2
2
2
X := u · |grad(u)| + α grad(u).

Using the divergence theorem and elaborating, we obtain

  p−2
2
2 2
|grad(u)| + α |grad(u)| dμ
  p−4
2
2 2
≤ |p − 2| |u| |grad(u)| + α |grad(u)| |Hess(u)|dμ
  p−2
2
2
+ |u||Δu| |grad(u)| + α dμ.

Letting α → 0 and applying the monotone and dominated convergence theorems we get5

p
|grad(u)| dμ

p−2 p−2
≤ |p − 2| |u||Hess(u)| |grad(u)| dμ + |u||Δu| |grad(u)| dμ. (8)

Now, in both the integrands appearing in the right-hand side of (8), we use the Young
inequality

5
Obviously, by monotone convergence, the same integral inequality holds if p < 2. However, in this case,
the right-hand side could be infinite. For instance, in R2 , this happens if p = 1 as one can see by taking
u(x, y) = (x2 + 1)ϕ(x, y), where 0 ≤ ϕ ≤ 1 is a cut-off function satisfying ϕ = 1 on [0, 1] × [0, 1].
 B. Güneysu, S. Pigola / Advances in Mathematics 281 (2015) 353–393 365


1 p εq q
ab ≤ a +  b , a, b ≥ 0
εp p q

with

p = p/2, q  = p/(p − 2).

We obtain that the right-hand side of (8) is


1 p p 1 p p εq p
≤ |u| 2 |Hess(u)| 2 dμ + |u| 2 |Δu| 2 dμ + 2 |grad(u)| . (9)
ε  p
p ε  p
p q

Choose 0 < ε = ε(p) < 1 so small that


εq
Ap := 1 − 2(p − 2)  > 0,
q

and let Bp := 1/(εp p ). Then, inserting (7) into (8) we deduce

p p p p p
Ap |grad(u)| dμ ≤ (p − 2)Bp |u| 2 |Hess(u)| 2 dμ + Bp |u| 2 |Δu| 2 dμ.

Whence, using twice the Young inequality

1 2 ε2 2
ab ≤ a + b (10)
2ε2 2
with any arbitrary ε > 0 we conclude

|grad(u)| dμ ≤ ε−2 Cp
p
|u|p dμ + ε2 Dp |Hess(u)|p dμ + |Δu|p dμ

where we have set

(p − 1)Bp max(p − 2, 1)Bp

Cp := , Dp := .
2Ap 2Ap

b) Assume first that M is geodesically complete. Then by Theorem 4.1 of [10] we have
the multiplicative inequality
1 1
grad(u)p ≤ C(p)up2 Δup2 , (11)

which completes the proof in this case, using once more inequality (10).
Assume now that M is a relatively compact open subset of an arbitrary smooth and
geodesically incomplete Riemannian manifold (M , g). Then as above it is sufficient to
prove that (11) remains valid on M , which can be seen, for instance, from the usual
366 B. Güneysu, S. Pigola / Advances in Mathematics 281 (2015) 353–393

construction of complete metrics in a given conformal class: Assume that M is an in-

complete open manifold, otherwise the conclusion is trivial by restriction. We consider
a smooth, relatively compact exhaustion Mk ⊂⊂ Mk+1  M such that M ⊂⊂ M1 and
we pick any smooth function λ : M → R≥0 with the following properties: (i) λ(x) = 0
on M1 ; (ii) for every k ≥ 1, λ(x) = ck on M2k+1 \ M2k , where a sequence {ck } ⊂ (0, ∞)
with ck  ∞ will be specified later. Next, we define a new metric on M by gλ := e2λ g.
By construction, gλ = g on M . Let rk = distg (∂M2k+1 , ∂M2k ) and choose {ck } in such a

way that rk · eck = +∞. Then, gλ is geodesically complete. Indeed, if γ : [0, ∞) → M
is a divergent path, γ is forced to pass across every annulus M2k+1 \ M2k . Therefore, its

length satisfies (γ) ≥ rk · eck = ∞ and this characterizes the geodesic completeness.
Since (11) holds on (M , gλ ), by restriction it holds on its open subset set M , as claimed.
Another way to prove the validity of (11) on relatively compact open subset of pos-
sibly incomplete smooth Riemannian manifold is to use the (less elementary) double
construction as explained in the proof of Theorem 4.1 a). 2

We immediately get the following corollary, a gradient estimate which is a variant of

(7) for arbitrary p, and which is useful in establishing compactness results for solutions
of the Dirichlet problem for the Poisson equation (cf. the proof of Theorem 4.1 a) below):

Corollary 3.11. In the situation of Proposition 3.10, assume that CZ(p) holds on M for
some 1 < p < ∞. Then, there exists a constant C > 0, which only depends on the
CZ(p)-constants and p, such that the following inequality

grad(u)p ≤ C(Δup + up )

holds for every u ∈ C∞

c (M ).

4. Criteria for the validity of the Calderón–Zygmund inequality

4.1. Local results

This section is devoted to general (that is, without any particular curvature as-
sumptions) local results concerning the validity of CZ(p) inequalities. We start with
the following result, where it is claimed that CZ(p) with 1 < p < ∞ always holds on
relatively compact open subsets (where of course the corresponding constants cannot be
controlled explicitly):

Theorem 4.1. Let 1 < p < ∞. Then, CZ(p) holds on any relatively compact open subset
Ω ⊂ M . Moreover, if Ω ⊂ M is a relatively compact domain with smooth boundary ∂Ω,
then CZ(p) holds in the stronger form

Hess(u)p ≤ CΔup ,
 B. Güneysu, S. Pigola / Advances in Mathematics 281 (2015) 353–393 367

for every u ∈ C∞c (Ω) and for a constant C > 0. In both cases, the constants depend quite
implicitly on the geometry of Ω (cf. the proof).

The proof relies on the validity of the corresponding CZ(p) on a closed Riemannian
manifold. This reduction procedure is obtained by using the Riemannian double of a
manifold with boundary; see Appendix C. See also Remark 4.16 for a different and
somewhat more direct argument.

Proof. Let N be a relatively compact domain of M such that ∂N is a smooth hyper-

surface and Ω ⊂ N . “The” Riemannian double D(N ) of N is a compact Riemannian
manifold without boundary. Moreover, by its very construction, it is always possible to
assume that D(N ) contains an isometric copy ΩN of the original domain Ω; see Ap-
pendix C. It is clear that every closed manifold supports CZ(p) (this follows, e.g., from
Theorem 4.11 below). In particular, this applies to D(N ), namely, there exist suitable
constants C1 > 0 and C2 ≥ 0, depending on the geometry of D(N ), such that

Hess(u)p ≤ C1 Δup + C2 up , (12)

for every u ∈ C∞ (D(N )). In particular, the same inequality holds for every u ∈ C∞
c (ΩN ).
Since, up to Riemannian isometries, ΩN is just the original domain Ω, we conclude that
(12) (with the same constants) holds on Ω, as required.
We now assume that Ω as above has smooth boundary and is connected. Then, in
spirit of the proof of Lemma 9.17 in [15] we obtain that there exists a constant C3 > 0
such that

up ≤ C3 Δup , (13)

for every u ∈ C∞
c (Ω). Inserting this latter into (12) concludes the proof of part a). For
the sake of completeness, let us provide a self-contained proof of (13). By contradiction,
suppose that there exists a sequence {uk } ⊂ C∞ c (Ω) satisfying

(a) uk p = 1, (b) Δuk p → 0. (14)

Note that, by Corollary 3.11, {uk } is bounded in H1,p

0 (Ω). Therefore, the Rellich–
Kondrachov compactness theorem yields the existence of a subsequence {uk } that
converges strongly in Lp to a function u ∈ Lp (Ω). In fact, u ∈ C0 (Ω) if p > m. It
follows from (14)(a) that

up = 1. (15)

Now, by (12) and Corollary 3.11, {uk } is bounded in the reflexive Banach space H2,p
0 (Ω).
Therefore, a subsequence {uk } converges weakly in H2,p 0 (Ω) and the weak limit is
368 B. Güneysu, S. Pigola / Advances in Mathematics 281 (2015) 353–393

u ∈ H2,p
0 (Ω). In particular, by Remark 3.1.1, we have that the distributional Laplacian
of u is a Borel function Δu ∈ Lp (Ω) and, furthermore, for every ϕ ∈ C∞
c (Ω),

ϕΔuk dμ = − tr ◦ Hess(uk ) · ϕ dμ
Ω Ω

= −Hess(uk ), tr† (ϕ) dμ

Ω

→ −Hess(u), tr† (ϕ) dμ

Ω

=− tr ◦ Hess(u) · ϕ dμ
Ω

= ϕΔu dμ.
Ω

On the other hand, by (14)(b),

ϕΔuk dμ → 0,
Ω

thus proving that u ∈ H2,p

0 (Ω) is a strong solution of the Laplace equation:

Δu = 0 a.e. in Ω.

By elliptic regularity, Theorem 9.19 in [15], since the Laplace–Beltrami operator is uni-
formly elliptic with smooth coefficients in Ω, we deduce that u ∈ C∞ (Ω) and u = 0 on
∂Ω. The usual maximum principle then implies that u = 0. Obviously, this contradicts
(15). 2

We have seen that CZ(p) always holds on relatively compact domains, however, in
general one may have a rough control on the constants. We close this section with the
following Proposition 4.2 where we prove a much more precise CZ(p) on sufficiently small
geodesic balls. To this end recall the definition of rQ,k,α (x), the Ck,α -harmonic radius
with accuracy Q at x (cf. Appendix B).

Proposition 4.2. Fix an arbitrary x ∈ M . Then for all 1 < p < ∞, all real numbers r
with

0 < 2r < r∗ (x) := min(r2,1,1/2 (x), 1)/2 ∈ (0, ∞), (16)

and all real numbers D with

 B. Güneysu, S. Pigola / Advances in Mathematics 281 (2015) 353–393 369

r2,1,1/2 Br∗ (x) (x) = inf r2,1,1/2 (z) ≥ D > 0,

z∈Br∗ (x) (x)

there is a constant C = C(r, p, m, D) > 0, such that for all u ∈ C∞

c (M ) one has

        
 
1Br/2 (x) Hess(u) ≤ C 1B2r (x) up + 1B2r (x) Δup + 1B2r (x) grad(u)p . (17)
p

In particular, with some C̃ = C̃(r, p, m, D) > 0, for all u ∈ C∞

c (Br/2 (x)) one has

Hess(u)p ≤ C(up + Δup ). (18)

Proof. Let u ∈ C∞
c (M ), and pick a C
1,1/2
-harmonic coordinate system

φ = (y 1 , . . . , y m ) : Br∗ (x) (x) −→ Rm

with accuracy Q = 2. Then by the properties (B.1) and (B.2) of φ and by Remark B.3
we have the following inequalities on Br∗ (x) (x),

2−1 (δij ) ≤ (gij ) ≤ 2(δij ) as bilinear-forms, (19)

 
max max(gij ∞ , g ij ∞ ) ≤ C1 (m), (20)
i,j∈{1,...,m}
 
max max(∂l gij ∞ , ∂l g ij ∞ ) ≤ C2 (D, m). (21)
i,j,l∈{1,...,m}

Since by (19) and (16) we have

φ(Br/2 (x)) ⊂ Beucl

√ (0) ⊂ Beucl
r/ 2
√ (0) ⊂ φ(B2r (x)) ⊂ φ(B ∗
2r r (x) (x)),

applying Theorem 9.11 from [15] with L = Δ and with the Euclidean balls Ω := Beucl
√ (0),
2r
 eucl
Ω := Br/√2 (0) implies the existence of a C3 = C3 (p, r, m, D) > 0 such that


|∂i ∂j u(y)|p dy ≤ C3 |Δu(y)|p dy + C3 |u(y)|p dy. (22)
i,j B2r(x)
Br/2 (x) B2r (x)

One can deduce from (19)–(21) the following pointwise estimate in Br∗ (x) (x),

|Hess(u)|p ≤ C8 |∂i ∂j u|p + C5 |grad(u)|p , (23)
i,j


with some C8 = C8 (r, m, D, p) > 0. Indeed, let |A|2HS = ij A2ij denote the Hilbert–
Schmidt norm of a real-valued matrix A = (Aij ). Then with H := (Hess(u)ij ),

h := (∂i ∂j u), Γ := (− l Γlij ∂l u), G := (gij ) one has H = h + Γ and
370 B. Güneysu, S. Pigola / Advances in Mathematics 281 (2015) 353–393

   
|Hess(u)| = G−1 H HS = G−1 h + G−1 ΓHS
 
≤ G−1 HS (|h|HS + |Γ|HS ) ≤ C4 (|h|HS + |Γ|HS ) in Br∗ (x) (x)

and
  2  2 2
|Γ|2HS ≤ Γlij · (∂l u) ≤ C5 |grad(u)| in Br∗ (x) (x),
ij l l

for some C4 , C5 > 0 depending only on D and m. Whence, we get the estimate on
Br∗ (x) (x)

|Hess(u)| ≤ C6 |∂i ∂j u|2 + C6 |grad(u)|,
i,j

|Hess(u)|p ≤ C7 |∂i ∂j u|p + C7 |grad(u)|p
i,j

for some C6 , C7 > 0 depending only on D, m and p. This proves the validity of (23).
Using this latter in combination with (19) and (22) gives us a C9 = C9 (r, m, D, p) > 0
such that
        
 
1Br/2 (x) Hess(u) ≤ C9 1B2r (x) up + 1B2r (x) Δup + 1B2r (x) grad(u)p .
p

Finally, (18) follows by interpolation using Proposition 3.10 a). 2

Let us remark here that an essential point of the estimate from Proposition 4.2 is
that C depends on x only through a lower bound D on the local harmonic radius, a fact
which makes it possible to use this result in order to derive CZ(p) on a large class of
noncompact Riemannian manifolds (cf. the proof of Theorem 4.11).

4.2. Stability under compact perturbations

In this section we prove that the validity of a Calderón–Zygmund inequality depends

(with possibly different constants) only on the geometry at infinity of the underlying
manifold. Said in equivalent terms, the Riemannian structure of a manifold supporting
CZ(p) can be perturbed on a compact domain without violating CZ(p).

Theorem 4.3. Assume that either p ≥ 2 or that 1 < p < 2 and M is geodesically complete.
If there is a relatively compact open subset Ω ⊂ M such that CZ(p) holds on M \ Ω, then
CZ(p) also holds on M . Here, a possible choice of Calderón–Zygmund constants on M
depends on those of M \ Ω, those of a relatively compact open neighborhood of Ω, and
the choice of a gluing function (cf. the proof for the details).
 B. Güneysu, S. Pigola / Advances in Mathematics 281 (2015) 353–393 371

As we will see momentarily, the proof relies on the Lp -interpolation inequality from
Proposition 3.10 a), which makes gluing methods accessible to Calderón–Zygmund in-
equalities at all.

Proof of Theorem 4.3. Let u ∈ C∞ c (M ). Take an open subset Ω1 ⊂ M such that Ω ⊂⊂

Ω1 ⊂⊂ M and a function ξ ∈ C∞ c (Ω1 ) with 0 ≤ ξ ≤ 1, and ξ = 1 on Ω, which we are
going to use in order to glue two CZ(p)’s together. To this end, set φ := (1 − ξ). By
the validity of CZ(p) on Ω1 (by part a)) and on M \ Ω, and writing u = ξu + φu, with
ξu ∈ C∞ ∞
c (Ω1 ), φu ∈ Cc (M \ Ω), we get A, B > 0, independent of u, such that

Hess(u)p ≤ Hess(ξu)p + Hess(φu)p

   
≤ A Δ(ξu)p + ξup + B Δ(φu)p + φup
 
≤ A uΔξp + ξΔup + |grad(ξ)| · |grad(u)|p + ξup
 
+ B uΔφp + φΔup + |grad(φ)| · |grad(u)|p + φup
 
≤ C Δup + grad(u)p + up ,

with

0 < C := A max Δξ∞ , ξ∞ , grad(ξ)∞

+ B max Δφ∞ , φ∞ , grad(φ)∞ .

Interpolating with Proposition 3.10 a), the latter inequality completes the proof. 2

In the next corollary we essentially rephrase Theorem 4.3 in more geometric terms.
This formulation involves two geometric objects: (1) the connected sum of Riemannian
manifolds, whose construction will be recalled in Appendix C, and (2) the notion of an
end E of a complete Riemannian manifold M with respect to a compact domain Ω: E is
any of the unbounded connected components of M \ Ω.

Corollary 4.4 (= Theorem A). Given 1 < p < ∞, a complete Riemannian mani-
fold supports CZ(p), if and only if each of its ends E1 , . . . , Ek , with respect to any
smooth, compact domain Ω, supports the same Calderón–Zygmund inequality. In par-
ticular, CZ(p) holds on the Riemannian connected sum M = M1 #M2 of m-dimensional
complete Riemannian manifolds M1 and M2 if and only if the same inequality holds on
both M1 and M2 .

4.3. Global results for CZ(2)

This section is devoted to the relation between global curvature assumptions and the
validity of global CZ(2) inequalities. Here, in view of Bochner’s equality, it is easy to give
372 B. Güneysu, S. Pigola / Advances in Mathematics 281 (2015) 353–393

a rather complete answer: CZ(2) always holds globally in a strong, “infinitesimial” way,
under a global lower bound on the Ricci curvature, and furthermore this result does not
even require geodesic completeness:

Proposition 4.5 (= Theorem B a)). Assume that Ric ≥ −C 2 for some constant C ∈ R,
meaning as usual that

Ric(X, X) ≥ −C 2 |X|2 for all vector fields X ∈ ΓC∞ (M, TM ).

Then CZ(2) holds in the following “infinitesimal” way: For every ε > 0 and every u ∈
C∞
c (M ) one has

2 Cε2 C2
Hess(u)2 ≤ u22 + 1 + 2 Δu22 .
2 2ε

Proof. By Bochner’s equality we have

1
|Hess(u)|2 = − Δ|grad(u)|2 + (grad(u), grad(Δu)) − Ric(grad(u), grad(u))
2
1
= − Δ|grad(u)|2 + (du, dΔu) − Ric(grad(u), grad(u)).
2
Now the claim follows easily from integrating this identity, using integration by parts,
Δu = d† du and the inequality

a2 ε2 b2
ab ≤ 2
+ ,
2ε 2
valid for a, b ≥ 0. 2

On the other hand, it is necessary for CZ(2) to have some control on the curvature,
as can be seen from:

Theorem 4.6 (= Theorem B b)). There exists a 2-dimensional, geodesically complete

Riemannian manifold N with unbounded Gaussian curvature and such that CZ(2) fails
on N .

The rest of this section is devoted to the proof of Theorem 4.6, that is, we are going
to construct an explicit example of a complete Riemannian surface M with unbounded
curvature and that does not support the global L2 -Calderón–Zygmund inequality

CZ(2) Hess(u)2 ≤ C(Δu2 + u2 ), u ∈ C∞

c (M ) .

Roughly speaking, in order to violate CZ(2), the idea is to minimize the contribution
of Δu with respect to Hess(u). Clearly, the best way to do this would be to choose u
 B. Güneysu, S. Pigola / Advances in Mathematics 281 (2015) 353–393 373

harmonic (and not affine) but this is impossible because u has compact support. To
overcome the problem, we can take u as the composition of a proper harmonic function
with a singularity at the origin and a cut-off function of R, compactly supported in (0, ∞).
Using this composition we get rid of the singularity and produce a smooth, compactly
supported function whose L2 -norm of the Laplacian can be small when compared with
that of the Hessian. We shall implement this construction on a model manifold where,
for rotationally symmetric functions, the expressions of the L2 -norms involved in CZ(2)
are very explicit and directly related to the geometry of the underlying space.
By an m-dimensional model manifold Rm σ we mean the Euclidean space R
m
endowed
with the smooth, complete Riemannian metric that, in polar coordinates, writes as

g = dr ⊗ dr + σ 2 (r)gSm−1 ,

where gSm−1 is the standard metric of Sm−1 and σ : [0, ∞) → [0, ∞) is a smooth function
satisfying the following structural conditions:

(a) σ (2k) (0) = 0, ∀k = 0, 1, . . .

(b) σ  (0) = 1
(c) σ(t) > 0, ∀t > 0.

We can always identify σ with its smooth, odd extension σ : R → R such that σ(t) =
−σ(−t) for every t ≤ 0. Recall that the sectional curvatures of Rm
σ are given by

σ 
Sec(X ∧ ∇r) = −
σ
1 − (σ  )2
Sec(X ∧ Y ) = ,
σ2

for every g-orthonormal vectors X, Y ∈ ∇r⊥ , where ∇r represents the radial direction.
Moreover, observe that the Riemannian measure of Rmσ is given by

dμ = σ m−1 (r) · dr · dμSm−1 ,

where dμSm−1 denotes the canonical Riemannian measure on Sm−1 .

Let us assume that

∞
dt
= ∞.
σ m−1 (t)

From the potential theoretic viewpoint, this means that Rmσ is parabolic, namely, the
minimal positive Green kernel of the Laplace–Beltrami operator of Rm
σ is identically ∞.
374 B. Güneysu, S. Pigola / Advances in Mathematics 281 (2015) 353–393

Then,

r
dt
G(r) =
σ m−1 (t)
1

is a smooth, positive, strictly-increasing function on (0, ∞) satisfying

⎧
⎪ =∞ if r=∞
⎪
⎪
⎪
⎨ 0
> if r>1
G(r) = 0 if r=1
⎪
⎪
⎪
⎪
⎩< 0 if 0<r<1
= −∞ if r = 0+ .

Moreover, G(r) gives rise to a smooth, rotationally symmetric harmonic function G(x)
on Rm
σ \{0}. In particular:

σ 
ΔG = G + (m − 1) G = 0, on Rm
σ \{0}.
σ

We need the following computational lemma.

Lemma 4.7. Let Rm

σ be a complete, parabolic, model manifold so that σ
1−m
∈
/ L1 (+∞).
Let, as above,

r
1
G(r) = dt,
σ m−1 (t)
1

and let {αk }, {βk } ⊂ (0, ∞) be two sequences such that 1 < αk < βk . Assume further
that for any k one has given a function φk ∈ C∞ c (0, ∞) which satisfies supp(φk ) ⊂
[αk , βk ], and define uk ∈ C∞ (Rmσ ) by setting u k (x) = φk (G(x)). Then, each uk is in
fact compactly supported in

{αk ≤ G ≤ βk } ⊂ Rm
σ ,

and one has

βk
σ
2
(φk (s)) G−1 (s)
2 2
Hess (uk )2 ≥ ωm ds,
σ
αk

βk
(φk (s))
2
2
Δuk 2 = ωm 2(m−1)
ds,
(σ (G−1 (s)))
αk
 B. Güneysu, S. Pigola / Advances in Mathematics 281 (2015) 353–393 375

βk
2(m−1)
σ G−1 (s)
2 2
uk 2 = ωm (φk (s)) ds,
αk

where ωm > 0 is a dimensional constant.

Proof. Recall that

σ 
Hess (uk ) = uk · dr ⊗ dr + u · σ 2 gSm−1 .
σ k

Therefore, we have

σ
2
|Hess (uk )| = (uk ) + (m − 1) (uk )
2 2 2
σ
σ
2
≥ (uk )
2
.
σ

Since

uk (r) = φk (G)G

and

1
G =
σ m−1
we get

σ
2
|Hess (uk )| ≥ (uk )
2 2
σ
σ
2
= (φk (G)) (G )
2 2
σ
σ G
2
= (φk (G))
2
.
σ σ m−1

In particular, letting ωm be the volume of the standard (m − 1)-sphere,

∞
2 2
Hess (uk )2 = ωm |Hess (uk )| σ m−1 dt
0
∞
σ
2
(φk (G)) G dt
2
≥ ωm
σ
0
376 B. Güneysu, S. Pigola / Advances in Mathematics 281 (2015) 353–393

G−1 (βk )
σ
2
(φk (G)) G dt
2
= ωm
σ
G−1 (αk )

βk
σ
2
(φk (s)) G−1
2
= ωm ds
σ
αk

where, in the last equality, we have used the change of variable G(t) = s. Similarly, on
noting that

uk (r) = φk (G) (G ) + φk (G)G

2

1
= φk (G) 2(m−1) + φk (G)G ,
σ
using also the harmonicity of G, we compute

σ 
Δuk = uk + (m − 1) u
σ k
1 σ 
= φk (G) + φk (G) G + (m − 1) G
σ 2(m−1) σ
1
= φk (G) .
σ 2(m−1)
It follows that
∞
1
(φk (G))
2 2
Δuk 2 = ωm dt
σ 3(m−1)
0

G−1 (βk )
1
(φk (G)) G dt
2
= ωm
σ 2(m−1)
G−1 (αk )

βk
1
(φk (s))
2
= ωm 2(m−1)
ds.
(σ (G−1 ))
αk

Finally, we compute

∞
2 2
uk 2 = ωm (φk (G)) σ m−1 dt
0

G−1 (βk )

(φk (G)) σ 2(m−1) G dt

2
= ωm
G−1 (αk )
 B. Güneysu, S. Pigola / Advances in Mathematics 281 (2015) 353–393 377

βk
2(m−1)
σ G−1
2
= ωm (φk (s)) ds.
αk

This completes the proof. 2

Now we proceed with the choice of the warping function σ and of the cut-off func-
tions φk in such a way that CZ(2) is violated along the corresponding sequence of
test-functions uk . To this end, we begin by taking

m = 2, αk = k, βk = k + 1.

Next, we choose σ(t) in such a way that

t ≤ σ(t) ≤ t + 1, t > 1.

Remark 4.8. We explicitly note that, by definition of G,

t+1
log ≤ G(t) ≤ log(t), t > 1.
2

It follows that

es ≤ G−1 (s) ≤ 2es − 1, s > 0.

In particular,

ek ≤ G−1 (s) ≤ 2ek+1 − 1, on [k, k + 1].

Moreover, fix 0 < δ < 1/2 sufficiently close to 0 in such a way that

e1−2δ − 2 ≥ 0.

Since, for k  1,

G−1 (k + 1 − δ) − G−1 (k + δ) ≥ ek+1−δ − 2ek+δ + 1 = ek+δ (e1−2δ − 2) + 1 > 2

we also deduce that, for each k  1, there exists some integer h = h(k) > k such that

[h, h + 1] ⊆ [G−1 (k + δ), G−1 (k + 1 − δ)].

Finally,

es ≤ σ G−1 (s) ≤ 2es .

These estimates will be used repeatedly in the sequel.

378 B. Güneysu, S. Pigola / Advances in Mathematics 281 (2015) 353–393

We also require that σ(t) oscillates in each interval [k, k +1] with a slope that increases
with k. More precisely, let N ⊇ (nk )  +∞ and, for each k, divide [k, k + 1] into
nk subintervals of the same length 2εk where, obviously, εk = 1/2nk . Next, in each
subinterval [t̄, t̄ + 2εk ], with t̄ = k, k + 2εk , k + 4εk , · · · , k + 1 − 2εk , consider the piecewise
rectilinear function

εk + 1
t → t̄ + (t − t̄) on [t̄, t̄ + εk ],
εk

and

1 − εk
t → t̄ + εk + 1 + (t̄ + εk − t) on [t̄ + εk , t̄ + 2εk ].
εk

The upper and lower angles are smoothened out in regions as close to the vertices as we
desire via concave (resp. convex) functions; see [14]. For instance, we can assume that σ
is rectilinear with slope 1/εk in subintervals of length εk − 2ε3k .

Remark 4.9. By construction, in each rectilinear portion, (σ  )2 grows like 1/ε2k on a

subinterval of length εk − 2ε3k ∼ εk , as k → +∞.

Remark 4.10. The smoothing is obtained via functions of increasingly high second deriva-
tive. It follows that the Gaussian curvature of R2σ explodes to ∞ as the oscillatory part
becomes closer and closer to vertical segments. We also point out that, in dimensions
m ≥ 3 this construction gives rise to a model manifold whose sectional curvatures, both
radial and tangential, explode to ∞. Finally, observe that, due to the profile of σ, the
manifold should have vanishing injectivity radius (although, at the pole 0 ∈ R2σ , it holds
that rinj (0) = ∞).

To conclude, we choose φk (t) = φ(t − k) where φ(t) is compactly supported in [0, 1]

and satisfies φ(t) = t on [δ, 1 − δ]. In this way, φk ≡ 1 on the interval of fixed length
[k + δ, k + 1 − δ] so to capture some of the oscillations of σ(G−1 ). Note also that φk ∞
and φk ∞ are uniformly bounded.
Now, according to Lemma 4.7, we have the following estimates:

k+1
2
σ G−1 (s)
2 2
uk 2 = ω2 (φk (s)) ds ≤ Ce2k ,
k

and

k+1
(φk (s))
2
2
Δuk 2 = ω2 2 ds
(σ (G−1 (s)))
k
 B. Güneysu, S. Pigola / Advances in Mathematics 281 (2015) 353–393 379

k+1
ds C
≤C 2s
= 2k
e e
k

and, finally,

k+1
σ
2
(φk (s)) G−1 (s)
2 2
Hess (uk )2 ≥ ω2 ds
σ
k

k+1
4ω2 2
(φk (s)) σ  G−1 (s)
2
≥ ds
e2(k+1)
k

k+1−δ
4ω2
≥ (σ  (G−1 (s))2 ds
e2(k+1)
k+δ

k+1−δ
4ω2 (G−1 (s))
= (σ  (G−1 (s))2 ds
e2(k+1) (G−1 (s))
k+δ

k+1−δ
4ω2
= (σ  (G−1 (s))2 σ(G−1 (s))(G−1 (s)) ds
e2(k+1)
k+δ

k+1−δ
4ω2 ek+δ
≥ 2(k+1) (σ  (G−1 (s))2 (G−1 (s)) ds
e
k+δ

G−1 (k+1−δ)
4ω2 eδ−1
= (σ  (t))2 dt
ek
G−1 (k+δ)

h+1
4ω2 eδ−1
≥ (σ  (t))2 dt, h = h(k) > k,
ek
h

δ−1
4ω2 e 1
≥ · (εh − 2ε3h )
ek ε2h
C 1 C 1
≥ · ≥ k · .
e k εh e εk

Whence, we deduce that we can choose εk  0 in such a way that CZ(2) is violated.
This completes the proof of Theorem 4.6.
380 B. Güneysu, S. Pigola / Advances in Mathematics 281 (2015) 353–393

4.4. Global results for CZ(p)

This section is devoted to Riemann geometric criteria for the validity of global CZ(p)
inequalities, for arbitrary values of p. As one might guess, this situation is much more
complicated than the previously considered p = 2 case. We found the following two
criteria, which can also be considered as the main result of this paper.
The first result covers the whole Lp -scale in a great generality:

Theorem 4.11 (= Theorem C). Let 1 < p < ∞ and assume Ric∞ < ∞, rinj (M ) > 0.
Then there is a

C = C(m, p, Ric∞ , rinj (M )) > 0,

such that for all u ∈ C∞

c (M ) one has

Hess(u)p ≤ C(up + Δup ). (24)

Remark 4.12. Note that, using the usual definition of geodesic completeness in terms
of the exponential function, it is elementary to see that a positive injectivity radius
automatically implies geodesically completeness.

The second result is concerned with the 1 < p ≤ 2 case in a slightly different setting:
the geometry of the manifold is bounded up to order one but the injectivity radius
condition is replaced with a kind of generalized volume doubling assumption. The proof
of this result is of independent interest because it points out a deep relation between
Calderón–Zygmund inequalities and covariant Riesz transforms:

Theorem 4.13 (= Theorem D). Let 1 < p ≤ 2. Assume that M is geodesically complete
with R∞ < ∞, ∇R∞ < ∞, and that there are constants D ≥ 1, 0 ≤ δ < 2 with
δ
+r δ
μ(Btr (x)) ≤ DtD et μ(Br (x)) for all x ∈ M, r > 0, t ≥ 1. (25)

Then there is a

C = C(m, p, R∞ , ∇R∞ , D, δ) > 0,

such that for all u ∈ C∞

c (M ) one has

Hess(u)p ≤ C(up + Δup ). (26)

Remark 4.14. If M is geodesically complete with Ric ≥ 0, then one has the doubling
condition

μ(B2r (x)) ≤ 2m μ(Br (x)) for all r > 0, x ∈ M,

 B. Güneysu, S. Pigola / Advances in Mathematics 281 (2015) 353–393 381

which easily implies

μ(Btr (x)) ≤ 22m tm μ(Br (x)) for all r > 0, t ≥ 1, x ∈ M,

so that (25) is satisfied in this situation (with constants that only depend on m). On
the other hand, nonnegative Ricci curvature is not necessary for (25). An example is
provided in Appendix A.

The rest of this section is devoted to the proof of Theorem 4.11 and of Theorem 4.13,
respectively.
We will need the following auxiliary result (see for example Lemma 1.6 in [18] and its
proof) for the former:

Lemma 4.15. Assume that M is geodesically complete with Ric ≥ −C for some C > 0.
Then for any r > 0 there exists a sequence of points {xi } ⊂ M and a natural number
N = N (m, r, C) < ∞, such that

• Br/4 (xi ) ∩ Br/4 (xj ) = ∅ for all i, j ∈ N with i = j,


• i∈N Br/2 (xi ) = M ,
• the intersection multiplicity of the system {B2r (xi ) | i ∈ N} is ≤ N .

Now we can give the

Proof of Theorem 4.11. By Theorem B.4 there is a

D = D(m, rinj (M ), Ric∞ ) > 0

such that r2,1,1/2 (M ) ≥ D. Let 0 < r := min(D, 1)/2. We take a covering ∪i∈N Br/2 (xi ) =
M as in Lemma 4.15. By Proposition 4.2 we have a c = c(r, p, m, D) > 0 such that, for
all i ∈ N, all u ∈ C∞c (M ),

p p
|Hess(u)| dμ ≤ c |Δu|p dμ + c |grad(u)| dμ + c |u|p dμ,
Br/2 (xi ) B2r (xi ) B2r (xi ) B2r (xi )

so summing over i and using monotone convergence we get

p
 p
|Hess(u)| dμ ≤ |Hess(u)| dμ
M i
Br/2 (xi )
  p
≤c 1B2r (xi ) |Δu|p dμ + c 1B2r (xi ) |grad(u)| dμ
M i M i


+c 1B2r (xi ) |u|p dμ,
M i
382 B. Güneysu, S. Pigola / Advances in Mathematics 281 (2015) 353–393

which by Lemma 4.15 gives

 
Hess(u)p ≤ (cN )1/p Δup + grad(u)p + up .

A use of Proposition 3.10 a) completes the proof. 2

Remark 4.16. Obviously, a similar argument can be used to prove Theorem 4.1. Sim-
ply cover the compact domain Ω with a finite number of balls Br/2 with 0 < 2r <
min(1, r2,1,1/2 (Ω))/2.

Finally, we give the proof of Theorem 4.13, which as we have already remarked in
the introduction, uses the machinery of covariant Riesz-transforms. We will need the
following auxiliary Hilbert space lemma:

Lemma 4.17. Let S be a densely defined closed linear operator from a Hilbert space H1
to a Hilbert space H2 , and let T be a bounded self-adjoint operator in H1 . Then for any
λ > 0 with T ≥ −λ one has
 
 
S(S ∗ S + T + λ + 1)−1/2  ≤ 1.

Proof. Firstly, the polar decomposition (cf. Theorem VIII.32 in [28]) of S reads S =
U (S ∗ S)1/2 , with a partial isometry U from H1 to H2 whose domain of isometry contains
the range of (S ∗ S)1/2 . Secondly, we have

S ∗ S + T + λ + 1 ≥ S ∗ S + 1,

which in this case means nothing but

   
 ∗   
(S S + T + λ + 1)1/2 f  ≥ (S ∗ S + 1)1/2 f 

for all f in the domain of definition of (S ∗ S)1/2 , in particular,

 
 ∗ 
(S S + 1)1/2 (S ∗ S + T + λ + 1)−1/2 h ≤ h for all h ∈ H2

so
 
 ∗ 
(S S + 1)1/2 (S ∗ S + T + λ + 1)−1/2  ≤ 1.

Now we can estimate as follows

 
 1
S(S ∗ S + T + λ + 1)− 2 
 
 1
= S(S ∗ S + 1)− 2 (S ∗ S + 1) 2 (S ∗ S + T + λ + 1)− 2 
1 1
 B. Güneysu, S. Pigola / Advances in Mathematics 281 (2015) 353–393 383

 
 1
≤ (S ∗ S) 2 (S ∗ S + 1)− 2 (S ∗ S + 1) 2 (S ∗ S + T + λ + 1)− 2 
1 1 1

  √ √
 1
≤ (S ∗ S) 2 (S ∗ S + 1)− 2  ≤ sup t/ t + 1 ≤ 1,
1

t≥0

where we have used the spectral calculus (cf. Theorem VIII.5 in [28]) for the last norm
bound. 2

Using Lemma 4.17, we can now prove the boundedness of covariant Riesz transforms:

Proposition 4.18. Under the assumptions of Theorem 4.13, there exists a constant

C2 = C2 (m, p, R∞ , ∇R∞ , D, δ) > 0,

such that for any C3 = C3 (R∞ , m) > 0 with Ric ≥ −C3 , the covariant Riesz-transform
∇(Δ1 + C3 + 1)−1/2 is bounded in the sense of
 
 
∇(Δ1 + C3 + 1)−1/2  ≤ C2 .
p,p

Proof. We are going to use Theorem 4.1 in [29] (an abstract result concerning Lp -bounds
for P1 (P2 + a)−1/2 with appropriate differential operators Pj that are compatible in a
certain sense) in combination with Example 2.6 therein (which states that the required
compatibility assumptions indeed are satisfied for P1 = ∇, P2 = Δ1 ). To this end, we
pick a constant C4 = C4 (R∞ , m) > 0 with

−C3 ≤ Ric ≤ C4 , (27)

and we remark that

Δ1 = ∇† ∇ + Ric(, ).

Thus, applying Lemma 4.17 (where we omit obvious essential self-adjointness arguments)
with S = ∇ (on 1-forms), and T = Ric(, ), which is read as a self-adjoint multiplication
operator, bounded by assumption (27), we get that the operator

Tσ := ∇(Δ1 + C3 + 1)−1/2

from Theorem 4.1 in [29] is bounded in the L2 -sense, with operator norm ≤ 1. It re-
mains to check the corresponding assumptions A and B from [29]: Here, the validity of
assumption A follows immediately from our global C1 -curvature assumptions and their
consequence (27), cf. Example 2.6 from [29]. Assumption B1 follows from the Lapla-
cian comparison theorem and (27), and assumption B3 is implied by the usual Li–Yau
heat kernel estimates, using again (27). Finally, B2 is precisely our volume assumption
(25). 2
384 B. Güneysu, S. Pigola / Advances in Mathematics 281 (2015) 353–393

Proof of Theorem 4.13. Pick C3 as in Proposition 4.18. We are going to prove the
existence of a

C = C(m, p, R∞ , ∇R∞ , D, δ) > 0

such that for all u ∈ C∞

c (M ) one has

 
∇d(Δ0 + C3 + 1)−1 u ≤ Cup . (28)
p

To this end, let us first note that under geodesic completeness (28) is equivalent to
 
 
∇(Δ1 + C3 + 1)−1/2 d(Δ0 + C3 + 1)−1/2 u ≤ Cup . (29)
p

Indeed, it is well-known that (e.g. Remark B.9 in [12])

de−tΔ0 = e−tΔ1 d for all t > 0

in the sense of closed operators, so that taking Laplace transforms we also have

d(Δ0 + C3 + 1)−1/2 = (Δ1 + C3 + 1)−1/2 d,

and we end up with

d(Δ0 + C3 + 1)−1 u = d(Δ0 + C3 + 1)−1/2 (Δ0 + C3 + 1)−1/2 u

= (Δ1 + C3 + 1)−1/2 d(Δ0 + C3 + 1)−1/2 u.

Now we use that by a classical result on Riesz-transforms of functions by Bakry [2] (see
also [22,23] for the weighted case) we can pick a constant C1 = C1 (p) > 0 with
 
 
d(Δ0 + C3 + 1)−1/2  ≤ C1 , (30)
p,p

so that by picking a constant C2 > 0 as in Proposition 4.18, the proof of (29) is completed
by setting C := C1 C2 . 2

4.5. Calderón–Zygmund inequalities on H-hypersurfaces

Let Mm+1 (c) denote the complete, simply connected space-form of constant sectional
curvature c ≤ 0. In this section we explore the validity of the Lp -Calderón–Zygmund
inequalities on a largely investigated class of submanifolds of Mm+1 (c): the hypersurfaces
of constant mean curvature H ∈ R (H-hypersurfaces for short) with finite total scalar
curvature.
 B. Güneysu, S. Pigola / Advances in Mathematics 281 (2015) 353–393 385

Let f : M → Mm+1 (c) be a complete, connected, oriented, isometrically immersed

submanifold of dimension dim M = m ≥ 3. Its second fundamental tensor, with respect
to a chosen Gauss map ν, is denoted by II. The corresponding mean curvature vector
field is H = trace(II)/m. We write H = Hν, where the smooth function H is the mean
curvature function of the hypersurface, and we assume that H is constant. The total
curvature of the constant mean curvature hypersurface M is the Lm -norm of its traceless
second fundamental tensor Φ = II − Hg. We say that M has finite total curvature
if Φm < ∞. In case H = 0 the hypersurface is called minimal and the finite total
curvature condition reduces to IIm < ∞.
A complete, oriented H-hypersurface M in Mm+1 (c) of finite total curvature must be
necessarily closed provided H 2 + c > 0. Indeed, according to [1,3], the traceless tensor
Φ satisfies the decay condition

sup |Φ| → 0 as R → ∞. (31)

M \BM
R

See also [27]. Therefore, by Gauss equations, the Ricci curvature of M is positively
pinched outside a compact set, [21], and the compactness conclusion follows from the
Bonnet–Myers type theorems in [13].
Since, on the one hand, the condition H 2 + c > 0 implies obvious non-existence results
and, on the other hand, we are mainly interested in noncompact situations, from now
on we assume that

H 2 + c ≤ 0. (32)

In particular, if c = 0, then M is minimal.

Under the compatibility condition (32), it is a well-known consequence of the curvature
estimate (31) that the H-hypersurface is properly immersed and it has a finite number
of ends, each of which is diffeomorphic to a cylinder over some compact hypersurface;
[1,5]. Moreover, up to imposing a more stringent pinching on H when c < 0, the volume
of each end is subjected to a certain growth; [1,27]. Actually, it is known that any
complete Riemannian manifold isometrically immersed with bounded mean curvature
into a Cartan–Hadamard manifold satisfies the non-collapsing condition at infinity

inf μ(B1 (x)) = v > 0. (33)

x∈M

Indeed, according to [20], such a submanifold enjoys the L1 -Sobolev inequality

u m−1
m ≤ C(grad(u)1 + u1 ),

for every u ∈ C∞
c (M ) and for some constant C > 0 depending on m and H∞ < +∞.
Whence, it is standard to deduce the validity of (33) by integrating the differential
inequality
386 B. Güneysu, S. Pigola / Advances in Mathematics 281 (2015) 353–393

m−1 d
μ(Br (x)) m ≤C μ(Br (x)) + μ(Br (x))
dr

that arises from a suitable choice of the (radial) cut-off functions u and a standard ap-
plication of the co-area formula. This is part of the classical Federer–Fleming argument.
Note that, using a rescaling procedure, the unit ball in (33) can be replaced by any ball
of fixed radius r > 0. Obviously, in this case, the constant v will depend on r.
We are now ready to prove the main result of the section.

Theorem 4.19. Let f : M → Mm+1 (c) be a complete, noncompact, oriented, H-hyper-

surface with finite total curvature, embedded into the complete, simply connected space-
form Mm+1 (c) of constant sectional curvature c ≤ 0. Then M satisfies the assumptions
of Theorem 4.11, in particular, for every 1 < p < ∞, the Calderón–Zygmund inequality
CZ(p) holds on M .

Proof. Combining the Gauss equations with estimate (31) on the traceless second funda-
mental tensor, we deduce that M has bounded sectional curvature. On the other hand,
M satisfies the non-collapsing condition (33). It follows from Theorem 4.7 in [8] that
rinj (M ) > 0. Therefore, we can apply Theorem 4.11 above and conclude the validity of
CZ(p). 2

Remark 4.20. The decay of the traceless second fundamental tensor of the H-hypersurface
M of Mm+1 (c) holds provided M has finite Lp -total curvature Φp < ∞ for some
m ≤ p < ∞, [27]. The conclusion of Theorem 4.19 can be extended accordingly.

Remark 4.21. As a matter of fact, inspection of the proof of Theorem 4.19 shows that it
relies on two facts: (a) by Gauss equations, the sectional curvature of a manifold M is
bounded if M is isometrically immersed, with bounded second fundamental form, into an
ambient manifold of bounded curvature; (b) the injectivity radius of M is bounded from
below by a positive constant provided M is isometrically immersed, with bounded mean
curvature, into a Cartan-Hadamard manifold. Therefore, Theorem 4.19 can be extended
in the following more abstract form:

Theorem 4.22. Let f : M → N be a complete Riemannian manifold isometrically

immersed into a complete, simply connected manifold N with sectional curvatures sat-
isfying −A2 ≤ SecN ≤ 0. If the second fundamental tensor of the immersion satisfies
II∞ < ∞, then M satisfies the assumptions of Theorem 4.11, in particular CZ(p)
holds on M , for every 1 < p < ∞.

Acknowledgments

The authors are indebted to Baptiste Devyver for pointing out the multiplicative
inequality contained in Theorem 4.1 of [10]. Furthermore, the authors would like to thank
 B. Güneysu, S. Pigola / Advances in Mathematics 281 (2015) 353–393 387

Gilles Carron, Xiang-Dong Li, and Feng-Yu Wang for very helpful correspondences.
Finally, the authors would like to thank the anonymous referee for helpful comments.

Appendix A. Doubling vs nonnegative Ricci

In this section we construct an example of a complete manifold satisfying the assump-

tions of Theorem 4.13 (actually, also those of Theorem 4.11) but whose Ricci curvature
can be negative somewhere (although it remains bounded from below).
Let (N, h) be any compact Riemannian manifold of dimension m − 1 and let (M, g) =
(N × R, h + dt ⊗ dt). As M is co-compact, it has bounded geometry up to order ∞. On
the other hand, there exists a constant C > 0 depending on the geometry of N such
that, for every (p0 , t0 ) ∈ M , R > 0 and t ≥ 1

μM BM
tR ((p0 , t0 ))
≤ Ctm . (A.1)
μM BM
R ((p 0 , t 0 ))

Indeed, since
 √
max(dN , dR ) ≤ dM = d2N + d2R ≤ 2 max(dN , dR )

we have

BN √ (p0 ) × BR √ (t0 ) ⊆ BM ((p0 , t0 )) ⊆ BN (p0 ) × BR (t0 )

R/ 2 R/ 2 R R R

proving that

μM BM
tR ((p0 , t0 ))
√ μN BN (p0 )
≤ 2t ·  tR .
μM BM
R ((p ,
0 0t )) N
μN BR/√2 (p0 )

Therefore, we are reduced to show that

μN BN (p0 )
 tR  ≤ Ctm−1 .
N √
μN BR/ 2 (p0 )

To this end, note that, if

rinj (N )
tR ≤ ,
2

then, the desired inequality follows from volume comparison. Indeed, let

−K 2 ≤ SecN ≤ K 2 ,
388 B. Güneysu, S. Pigola / Advances in Mathematics 281 (2015) 353–393

then the continuous functions α1 , α2 : [0, rinj (N )/2] → (0, ∞) defined by (P. Petersen
notation, [26])
r
0
snm−2
K 2 (s)ds
α1 (r) =
rm−1
r
0
snm−2
−K 2 (s)ds
α2 (r) =
rm−1

satisfy

Ai ≤ αi (r) ≤ Bi

where Ai , Bi > 0 are constants depending only on K, m and rinj (N ). It follows that

m−1
μN BN (p0 ) B (tR)
 tR ≤ 2 √ m−1 = Ct
m−1
,
N
μN BR/√2 (p0 ) A 1 R/ 2

as claimed. On the other hand, if

rinj (N )
tR > ,
2
since

μN BN (p0 ) μN (N )
 tR ≤  
N
μN BR/√2 (p0 ) N
μN Br (N )/(t2√2 ) (p0 )
inj

and

rinj (N ) r (N )
√ ≤ inj
t2 2 2

using again volume comparison we get

  rinj (N )
m−1
μN BN √ (p0 ) ≥ A1
rinj (N )/t2 2
√
t2 2

and, hence,
√ m−1
μN BN (p0 ) μ (N ) 2 2
 tR ≤ N · tm−1 .
N
μN BR/√2 (p0 ) A1 rinj (N )

This completes the proof of (A.1).

 B. Güneysu, S. Pigola / Advances in Mathematics 281 (2015) 353–393 389

Appendix B. Harmonic coordinates

In this section, we collect some facts concerning harmonic coordinates. Let again
M ≡ (M, g) be an arbitrary smooth Riemannian m-manifold without boundary, let ∇
be the Levi-Civita connection and Δ the Laplace–Beltrami operator.

Definition B.1. Let x ∈ M , Q ∈ (1, ∞), k ∈ N≥0 , α ∈ (0, 1). The Ck,α -harmonic radius
of M with accuracy Q at x is defined to be the largest real number rQ,k,α (x) with
the following property: The ball BrQ,k,α (x) (x) admits a centered harmonic coordinate
system

φ : BrQ,k,α (x) (x) −→ Rm

(that is, φ(x) = 0 and Δφj = 0 on BrQ,k,α (x) (x) for each j), such that

Q−1 (δij ) ≤ (gij ) ≤ Q(δij ) in BrQ,k,α (x) (x) as symmetric bilinear forms, (B.1)

and for all i, j ∈ {1, . . . , m},


rQ,k,α (x)|β| sup |∂β gij (x )|
x ∈B rQ,k,α (x) (x)
β∈Nm , 1≤|β|≤k

 |∂β gij (x ) − ∂β gij (x )|

+ rQ,k,α (x)k+α sup
x ,x ∈BrQ,k,α (x) (x), x =x d(x , x )α
β∈Nm , |β|=k

≤ Q − 1. (B.2)

We shall refer to a coordinate system as above as a Ck,α -harmonic coordinate system

with accuracy Q on BrQ,k,α (x) (x).

Remark B.2. 1. Note that, when compared with the corresponding definition from [18],
we additionally require φ(x) = 0 here.
2. It is easily checked that the function x → min(rQ,k,α (x), δ) is globally 1-Lipschitz,
for any fixed δ > 0.
3. By polarization, the inequality (B.1) implies that for some C = C(m, Q) > 0 it
holds that

max sup |gij (x )| ≤ C,

i,j∈{1,...,m} x ∈Br (x)
Q,k,α (x)

in particular, putting everything together, there is a continuous decreasing function

F = FQ,k,α,m : (0, ∞) −→ (0, ∞),

390 B. Güneysu, S. Pigola / Advances in Mathematics 281 (2015) 353–393

such that the Euclidean Ck,α -norm of the metric in this coordinates satisfies

max gij Ck,α ≤ F (rQ,k,α (x)). (B.3)

i,j∈{1,...,m}

This justifies the name Ck,α -(harmonic) coordinate system.

Remark B.3. The natural differential operators of M (such as the gradient, the Laplacian
and the Hessian of a given function) are defined in terms of the inverse metric coeffi-
cients g ij . It is easy to see that, within the coordinate ball BrQ,k,α (x) (x), they inherit
the Ck,α -type control, in terms of Q, m, α, k, of the metric coefficients gij . Indeed, the
Cramer formula states that
det Gji
g ij = (−1)i+j , (B.4)
det (gij )

where Gij denotes the (m −1) ×(m −1) matrix obtained from (gij ) by deleting the ith-row
and the jth-column. Both the numerator and the denominator of (B.4) are obtained as
the sum of products of C 0,α -controlled functions and, by (B.1), Q−m ≤ det(gij ) ≤ Qm .
Therefore, we can obtain a C0,α control of the functions g ij by using the C0,α estimates
of gij in combination with the following elementary fact:

Assume that f, h : U ⊆ Rm → R satisfy C −1 ≤ h ≤ C and |f | ≤ D, for some

constants C, D > 0. Then:

|Δx,y (f /h)| ≤ C 3 |Δx,y (f )| + C 2 D |Δx,y (h)|

|Δx,y (f h)| ≤ C |Δx,y (f )| + D |Δx,y (h)|
|Δx,y (f + h)| ≤ |Δx,y (f )| + |Δx,y (h)| ,

where, to simplify the writings, we have set Δx,y (•) = •(y) − •(x).

Now, differentiating the identity g ik · gkj = δij we get

∂t g ij = −g ih · ∂t ghj · g jk . (B.5)

It follows that a C0,α control of ∂g ij is obtained from those of g ij and ∂gij . Proceeding
inductively on the derivatives of (B.5) we finally deduce the claimed Ck,α estimate of g ij .

The main result in this context states that control on the Ricci curvature up to order
k together with control on the injectivity radius imply control on rQ,k+1,α (x). To this
end, for any Ω ⊂ M open and any ε > 0 let

Ωε := {x | x ∈ M, d(x, Ω) < ε} ⊂ M

be the ε-neighborhood of Ω.
 B. Güneysu, S. Pigola / Advances in Mathematics 281 (2015) 353–393 391

Theorem B.4. Let Q ∈ (1, ∞), α ∈ (0, 1). Assume that there is an open subset Ω ⊂ M ,
and numbers k ∈ N≥0 , ε > 0, r > 0, c0 , . . . , ck > 0 with
 j 
∇ Ric(x) ≤ cj , rinj (x) ≥ r for all x ∈ Ωε , j ∈ {0, . . . , k}.
x

Then there is a constant

C = C(m, Q, k, α, ε, r, c1 , . . . , ck ) > 0,

such that for all x ∈ Ω one has rQ,k+1,α (x) ≥ C.

Proof. Except the additional assumption φ(x) = 0 that we have made for harmonic
coordinates, this result can be found in [18] and the references therein (cf. Theorem 1.3
therein). However, since translations do not effect the required estimates, this is not a
restriction. 2

In particular, the latter result implies rQ,j,α (x) > 0 for all x ∈ M , a fact which a priori
is not clear at all. One calls the number

rQ,j,α (M ) := inf rQ,j,α (x)

x∈M

the Ck,α -harmonic radius for the accuracy Q.

Appendix C. Gluing Riemannian manifolds

Suppose we are given two (smooth) Riemannian manifolds (M1 , g1 ) and (M2 , g2 ) with
compact diffeomorphic boundaries and let f : ∂M1 → ∂M2 be a fixed diffeomorphism.6
The Riemannian gluing M = M1 ∪f M2 of M1 and M2 along f is the Riemannian
manifold (M, g) defined as follows.
As a topological manifold, M is the quotient space obtained from the disjoint union
M1  M2 under the identification x ∼ f (x), for every x ∈ ∂M1 . It turns out that the
natural inclusions ij : Mj → M, j = 1, 2, are continuous embeddings.
Next, having fixed arbitrarily small collar neighborhoods αj : ∂Mj × [0, 2) → Mj
of ∂Mj , j = 1, 2, we consider the homeomorphism α : ∂M1 × (−2, 2) → M onto a
neighborhood V of i1 (∂M1 ) = i2 (∂M2 ) defined as follows:

i1 ◦ α1 (x, −t) t≤0
α(x, t) =
i2 ◦ α2 (f (x), t) t ≥ 0.

The original differentiable structures on M1 and M2 are then extended to a (unique up to

diffeomorphisms) differentiable structure on M by requiring that the natural inclusions

6
Different choices of f could produce non-diffeomorphic gluings as the exotic twisted spheres show.
392 B. Güneysu, S. Pigola / Advances in Mathematics 281 (2015) 353–393

ij , j = 1, 2, are smooth embeddings and by pretending that α is a smooth diffeomorphism.

See e.g. Chapter 8 of [19] and Chapter 5 of [24].
Finally, let W = α(∂M1 × (−1, 1)), and fix any Riemannian metric g3 on W. For
instance, we can pull-back on W via α−1 a product metric h+dt⊗dt on ∂M1 ×(−1, 1). Let
ξ1 , ξ2 and ξ denote a partition of unity subordinated to the open covering i1(M1 \∂M1 ),
i2 (M2 \∂M2 ), and W of M. A Riemannian metric on M is defined by setting

g = ξ1 · (i−1 ∗ −1 ∗
1 ) g1 + ξ2 · (i2 ) g2 + ξ · g3 .

Note that, outside the compact neighborhood W of i1 (∂M1 ) = i2 (∂M2 ), M is isometric

to the original open manifolds Mj \αj (∂Mj × [0, 1]). In particular, different choices of
g3 will leave the corresponding Riemannian structure of M in the same bi-Lipschitz
class. Moreover, if Ω is a domain compactly contained, e.g., in M1 \ ∂M1 , then the collar
neighborhood α1 (∂M1 × [0, 2)) of ∂M1 can be chosen so to have empty intersection
with Ω. Therefore, the neighborhood V ⊃ W of i1 (M1 ) = i2 (M2 ) does not intersect
i1 (Ω). Whence, it follows that Ω can be identified with its isometric copy i1 (Ω) into the
glued space.
Now, if we specialize this construction to the case where M1 = M2 and h = id we
obtain “the” Riemannian double M = D(M1 ) of M1 . On the other hand, if M1 and M2
are obtained by deleting a disk from the manifolds without boundaries M1 and M2 then
we get the (rough and un-oriented) Riemannian connected sum M1 #M2 .

References

[1] M. Anderson, The compactification of a minimal submanifold, IHES/Caltech Preprint, 1984–86,

unpublished manuscript, http://www.math.sunysb.edu/~anderson/compactif.pdf.
[2] D. Bakry, Etude des transformations de Riesz dans les varietes riemanniennes a courbure de Ricci
minoree, in: Seminaire de Probabilites, XXI, in: Lecture Notes in Math., vol. 1247, Springer, Berlin,
1987, pp. 137–172.
[3] P. Bérard, M.P. do Carmo, W. Santos, Complete hypersurfaces with constant mean curvature and
finite total curvature, Ann. Global Anal. Geom. 16 (1998) 273–290.
[4] A.P. Calderon, A. Zygmund, On the existence of certain singular integrals, Acta Math. 88 (1) (1952)
85–139.
[5] P. Castillon, Spectral properties of constant mean curvature submanifolds in hyperbolic space, Ann.
Global Anal. Geom. 17 (1999) 563–580.
[6] J. Cheeger, T.H. Colding, Lower bounds on Ricci curvature and the almost rigidity of warped
products, Ann. of Math. 144 (1996) 189–237.
[7] J. Cheeger, M. Gromov, Chopping Riemannian manifolds, in: Differential Geometry, in: Pitman
Monogr. Surv. Pure Appl. Math., vol. 52, Longman Sci. Tech., Harlow, 1991, pp. 85–94.
[8] J. Cheeger, M. Gromov, M. Taylor, Finite propagation speed, kernel estimates for functions of the
Laplace operator, and the geometry of complete Riemannian manifolds, J. Differential Geom. 17
(1982) 15–53.
[9] B. Chow, S.-C. Chu, D. Glickenstein, C. Guenther, J. Isenberg, T. Ivey, D. Knopf, P. Lu, F. Luo,
L. Ni, The Ricci Flow: Techniques and Applications. Part III. Geometric–Analytic Aspects, Math.
Surveys Monogr., vol. 163, American Mathematical Society, Providence, RI, 2010.
[10] T. Coulhon, X.T. Duong, Riesz transform and related inequalities on noncompact Riemannian
manifolds, Comm. Pure Appl. Math. 56 (2003) 1728–1751.
[11] K. de Leeuw, H. Mirkil, Majorations dans L∞ des operateurs differentiels a coefficients constants,
C. R. Acad. Sci. Paris 254 (1962) 2286–2288.
 B. Güneysu, S. Pigola / Advances in Mathematics 281 (2015) 353–393 393

[12] B.K. Driver, A. Thalmaier, Heat equation derivative formulas for vector bundles, J. Funct. Anal.
183 (1) (2001) 42–108.
[13] G.J. Galloway, Compactness criteria for Riemannian manifolds, Proc. Amer. Math. Soc. 84 (1982)
106–110.
[14] M. Ghomi, The problem of optimal smoothing for convex functions, Proc. Amer. Math. Soc. 130
(2002) 2255–2259.
[15] D. Gilbarg, N.S. Trudinger, Elliptic Partial Differential Equations of Second Order, Springer-Verlag,
2001.
[16] D. Guidetti, B. Güneysu, D. Pallara, H = W on the whole Lp -scale on arbitrary manifolds, preprint,
2014.
[17] B. Güneysu, Sequences of Laplacian cut-off functions, J. Geom. Anal. (2014), http://dx.doi.org/
10.1007/s12220-014-9543-9.
[18] E. Hebey, Sobolev Spaces on Riemannian Manifolds, Lecture Notes in Math., vol. 1635, Springer-
Verlag, Berlin, 1996.
[19] M.W. Hirsch, Differential Topology, Grad. Texts in Math., vol. 33, Springer-Verlag, New York, 1994.
[20] D. Hoffman, J. Spruck, Sobolev and isoperimetric inequalities for Riemannian submanifolds, Comm.
Pure Appl. Math. 27 (1974) 715–727.
[21] P.-F. Leung, An estimate on the Ricci curvature of a submanifold and some applications, Proc.
Amer. Math. Soc. 114 (1992) 1051–1061.
[22] X.-D. Li, Riesz transforms on forms and Lp-Hodge decomposition on complete Riemannian mani-
folds, Rev. Mat. Iberoam. 26 (2010) 481–528.
[23] X.-D. Li, On the Lp -estimates of Riesz transforms on forms over complete Riemanian manifolds,
arXiv:1304.3150.
[24] J.R. Munkres, Elementary Differential Topology, Ann. of Math. Stud., vol. 54, Princeton University
Press, Princeton, NJ, 1966.
[25] D. Ornstein, A non-equality for differential operators in the L1 norm, Arch. Ration. Mech. Anal.
11 (1962) 40–49.
[26] P. Petersen, Riemannian Geometry, Grad. Texts in Math., vol. 171, Springer, 2006.
[27] S. Pigola, G. Veronelli, Uniform decay estimates for finite-energy solutions of semi-linear elliptic
inequalities and geometric applications, Differential Geom. Appl. 29 (2011) 35–54.
[28] M. Reed, B. Simon, Methods of Modern Mathematical Physics. I. Functional Analysis, second ed.,
Academic Press, Inc., 1980.
[29] A. Thalmaier, F.-Y. Wang, Derivative estimates of semigroups and Riesz transforms on vector
bundles, Potential Anal. 20 (2004) 105–123.
[30] F. Wang, X. Zhu, On the structure of spaces with Bakry–Emery Ricci curvature bounded below,
arXiv:1304.4490, 2013.