You are on page 1of 25

SCALAR CURVATURE DEFORMATIONS WITH NON-COMPACT BOUNDARIES

HELGE FRERICHS

Abstract. We develop a general deformation principle for families of Riemannian metrics on smooth
manifolds with possibly non-compact boundary, preserving lower scalar curvature bounds. The principle
is used in order to strengthen boundary conditions, from mean convex to totally geodesic or doubling.
The deformation principle preserves further geometric properties such as completeness and a given
quasi-isometry type.
As an application, we prove non-existence results for Riemannian metrics with (uniformly) positive
scalar curvature and mean convex boundary.
arXiv:2403.03941v1 [math.DG] 6 Mar 2024

1. Introduction
By the h-principle for open Diff-invariant partial differential relations, each connected manifold of
dimension at least two and with non-empty boundary carries a metric of positive scalar curvature.
Furthermore, the space of all such metrics is contractible. If additional conditions are imposed on the
boundary, such as being mean convex or totally geodesic, the topology becomes non-trivial. This has
been systematically studied for manifolds with a compact boundary in work by Bär-Hanke [4].
The paper at hand presents a general deformation principle for Riemannian metrics with a lower
scalar curvature bound on manifolds with possibly non-compact boundary. This allows for comparison
of various boundary conditions. The method applies whenever one needs a deformation which is mean
curvature nonincreasing. A special case of our main result, Theorem 3.17, is the following:
Sample Theorem. Let n ≥ 2. Let M be an n-dimensional manifold with non-empty (possibly non-
compact) boundary and let g be a Riemannian metric on M with scalg > 0. Let k ∈ C ∞ (∂M ; T ∗ ∂M ⊗
T ∗ ∂M ) be a symmetric (0, 2)-tensor field along ∂M satisfying n−11
trg0 (k) ≤ Hg . Here g0 denotes the
metric induced on ∂M and Hg denotes the mean curvature of the boundary.
Then there exists a continuous path (fs )s∈[0,1] of Riemannian metrics on M such that
(a) scalfs > 0 for all s ∈ [0, 1];
(b) f0 = g;
(c) IIf1 = k where II denotes the second fundamental form of the boundary.
The space of Riemannian metrics on M is equipped with the weak C ∞ -topology.
We can add various features to this assertion:
▷ Instead of only considering positive scalar curvature (psc), an arbitrary continuous function σ : M →
R can be used as a lower scalar curvature bound. This means that if scalg > σ, then the metrics fs
can be constructed in such a way that scalfs > σ for all s ∈ [0, 1].
▷ The result also applies to families of metrics and in a relative version. The last aspect is a main
advantage of our deformation principle since it allows not only to answer existence questions for
metrics with boundary conditions, but also to treat spaces of such metrics and to study their
homotopy type.
▷ The deformations may be supported in an arbitrarily small neighbourhood of the boundary.
▷ Our method preserves completeness of metrics, i.e. if g(ξ) is complete for some ξ ∈ K then f (ξ, s)
is complete for all s ∈ [0, 1].
▷ Recall that two Riemannian metrics g and h on M are called quasi-isometric via the identity if
there exist constants A ≥ 1 and B ≥ 0 such that
1
dh (p, q) − B ≤ dg (p, q) ≤ A dh (p, q) + B for all p, q ∈ M .
A
2010 Mathematics Subject Classification. 53C21, 53C23; Secondary: 53C20, 57M25, 57R55.
Key words and phrases. Manifolds with non-compact boundary, lower scalar curvature bounds, lower mean curvature
bounds, deformations of Riemannian metrics, local flexibility, doubling metrics, Whitehead manifold.
Acknowledgment. This work is partly based on my master’s thesis under the guidance of Bernhard Hanke. It was
supported by the TopMath Program from the Elite Network of Bavaria. I am grateful to Shmuel Weinberger for advice
regarding the Whitehead manifold.
1
Here dg and dh denote the Riemannian distances associated to g and h. The above condition defines
an equivalence relation on the space of Riemannian metrics on M .
The deformation principle preserves a given quasi-isometry type, meaning that it is carried out
within a fixed equivalence class of metrics.
A boundary condition of high interest is what we call the doubling property for metrics on a manifold
with boundary. A doubling metric is defined as a metric whose twofold copy is a smooth Riemannian
metric on the double of the manifold. By working with doubling metrics, our deformation principle
provides a connection between scalar curvature geometry on manifolds with boundary and on manifolds
without boundary. The use of such a connection is a strategy that has already been employed in work
by Gromov-Lawson [13], Almeida [1], and Carlotto-Li [8, 9], for instance.
In a similar spirit, using work by Hanke-Kotschick-Roe-Schick [14], Cecchini [10] and Chang-
Weinberger-Yu [11], we prove some non-existence results for (uniform) psc-metrics with mean convex
boundary. The term ’mean convex’ indicates that the mean curvature is non-negative at each boundary
point. We state our applications from Corollary 4.1, 4.3-4.5.
Applications. (1) The right half-plane R2x1 ≥0 cannot carry any complete Riemannian metric of uni-
formly positive scalar curvature which has mean convex boundary.
(2) Let n ≥ 3. The right half-space Rnx1 ≥0 cannot carry any complete Riemannian metric of uniformly
positive scalar curvature which has mean convex boundary and which is quasi-isometric to the Eu-
clidean metric via the identity.
(3) Let M := Rnx1 ≥0 #T n be a connected sum of the right half-space with an n-torus attached in the
interior. Then M does not admit a complete Riemannian metric of positive scalar curvature which
has mean convex boundary.
(4) The Whitehead manifold has contractible submanifolds with non-compact boundary which do not ad-
mit complete Riemannian metrics of uniformly positive scalar curvature with mean convex boundary.
The paper is organized as follows: The first section discusses the geometry of doubling metrics. We
will spend some effort to prove the following seemingly tautological result: If g is a complete doubling
metric on a manifold with boundary, then the two-fold copy g ∪ g is a complete metric on the double of
the manifold. This result is of independent interest. The precise proof makes use of absolutely continuous
curves on manifolds with boundary, relying on work of Burtscher [7].
In the second section, we develop the main deformation principle for families of metrics with lower
scalar curvature bounds in a three-step process. The essential part (step two and three) is to generalize
the deformation scheme [4, Prop. 23 and 26] of Bär-Hanke to non-compact boundaries. Here we make
use of the local flexibility lemma in [3]. Our deformation principle combines the work of Bär-Hanke with
a closer study of the topology of the boundary.
In the last section we present our non-existence results that have been mentioned above.

2. Completeness of doubling metrics


Let M be a smooth manifold with non-empty boundary. Given a Riemannian metric g on M its
normal exponential map provides a collar neighbourhood
ϕg : V → U ⊂ M , ϕg (t, p) = expp (tNg (p))
where V ⊂ [0, ∞)×∂M is an open neighbourhood of {0}×∂M and Ng is the inward-pointing unit normal
vector field along ∂M . We call ϕg a geodesic collar neighbourhood. It induces a smooth structure on
the topological double DM = M ∪∂M M .
The resulting smooth manifold is sometimes referred to as DM g to emphasise the smooth structure.
Note that two distinct Riemannian metrics induce diffeomorphic smooth structures on DM since two
arbitrary collar neighbourhoods for ∂M are isotopic. However, the two smooth structures are not the
same in general. This will be important later.
Definition 2.1. We call g a doubling metric if the twofold copy g ∪ g is a smooth Riemannian metric
on DM g .
There is another characterization for the doubling property: According to the generalized Gauss
lemma, pulling back g along ϕg yields a generalized cylinder metric
g = dt2 + gt
2
where gt is given by g• : V → T ∗ ∂M ⊗ T ∗ ∂M , gt (p)(v, w) = (ϕ∗g g)(t, p)(v, w). The family g• gives rise to
(ℓ)
new families ġ• , g̈• , g• : V → T ∗ ∂M ⊗ T ∗ ∂M of (0, 2)-tensor fields on ∂M defined by
d
ġt (p)(v, w) := (gt (p)(v, w)),
dt
d2
g̈t (p)(v, w) := 2 (gt (p)(v, w)),
dt
(ℓ) d(ℓ)
gt (p)(v, w) := (ℓ) (gt (p)(v, w))
dt
(ℓ)
for all (t, p) ∈ V and v, w ∈ Tp ∂M . Then g is doubling iff g0 ≡ 0 for ℓ odd.
Apart from that, the second fundamental form of the boundary is given by IIg = − 12 ġ0 , see [2, Prop.
4.1]. This means that doubling metrics need to have a totally geodesic boundary. The aim of this section
is to prove the following theorem for a connected M .
Theorem 2.2. If g is a complete doubling metric on M then G := g ∪ g is a complete metric on DM .
By completeness we mean metric completeness, which is the convergence of all Cauchy sequences with
respect to the Riemannian distance. The next equation turns out to be crucial:
Proposition 2.3. It holds dg (p, q) = dG (p, q) for all p, q ∈ M .
Here dg (p, q) is the Riemannian distance between p and q in M with respect to g and dG (p, q) is
their distance in DM with respect to G. Note that we have to consider all admissible curves in DM in
order to determine dG (p, q) whereas for dg (p, q) we only look at curves in M . This is why the inequality
dg (p, q) ≥ dG (p, q) is obvious.
To address the other inequality, we use a reflection argument: Reflecting an admissible curve in DM
along the hypersurface ∂M results in a curve in M that retains its original length. Therefore, there are
no shorter curves between p and q in DM than in M .
Unfortunately, not every class of admissible curves is invariant under reflection. For instance, the
common class of piecewise smooth curves is not as can be seen from the right half-plane M = R2x1 ≥0 and
the Euclidean metric g = geucl . Its double DM g is diffeomorphic to R2 via the canonical map

2 [(x1 , x2 ), 1] , x1 ≥ 0
κ : R → DM , κ(x) = .
[(−x1 , x2 ), 2] , x1 ≤ 0
Then  2
(e−1/t sin( πt ), t), t ̸= 0
γ : [0, 1] → R2 , γ(t) =
(0, 0), t=0
is a piecewise smooth curve but its reflection to M is not. This is why we consider the larger class of
absolutely continuous curves. The following considerations are based on Burtscher [7].
Definition 2.4. Let I ⊂ R be an interval (open or closed) or an open subset. A function f : I → Rn is
called absolutely continuous if for every ε > 0, there exists some δ > 0 such P that for every m ∈ N and
m
every family {(ai , bi )}m
i=1 of disjoint intervals with [a ,
i ib ] ⊂ I and total length i=1 |bi − ai | < δ, it holds
m
X
∥f (bi ) − f (ai )∥ < ε.
i=1
Here ∥ · ∥ denotes the Euclidean norm in Rn .
A function f : I → Rn is said to be locally absolutely continuous if it is absolutely continuous on every
closed interval [a, b] ⊂ I.
A function f : I → Rnx1 ≥0 is called (locally) absolutely continuous if it is (locally) absolutely continuous
as a function I → Rn .
Now let M be a connected smooth manifold with boundary of dimension n ∈ N. The boundary may
or may not be empty.
Definition 2.5. Let I ⊂ R be a closed interval. A continuous curve γ : I → M is called absolutely
continuous if, for every chart (φ, U ) of M , the function
φ ◦ γ : γ −1 (U ) → φ(U ) ⊂ Rnx1 ≥0
is locally absolutely continuous. The set of all absolutely continuous curves γ : [0, 1] → M is denoted
by Aac (M ) respectively Aac . For any two points p, q ∈ M , we denote by Aac p,q
⊂ Aac the set of all
absolutely continuous curves γ : [0, 1] → M with γ(0) = p and γ(1) = q.
3
We show that absolute continuity is a well-behaved concept in differential topology.
Lemma 2.6. Let I ⊂ R be a closed interval. A continuous curve γ : I → M is absolutely continuous iff
for every p ∈ γ(I), there exists a local chart (φ, U ) of M around p such that φ ◦ γ : γ −1 (U ) → φ(U ) is
locally absolutely continuous.
Proof. Let (ψ, V ) be an arbitrary local chart of M with V ∩ γ(I) ̸= ∅. Let [a, b] ⊂ γ −1 (V ) be a closed
interval. We need to prove absolute continuity for ψ ◦ γ : [a, b] → ψ(V ).
By assumption there exists an atlas (φl , Ul )l∈I of M such that φl ◦ γ : γ −1 (Ul ) → φl (Ul ) is locally
absolutely continuous for every l ∈ I. Then we find a partition a = s0 < s1 < · · · < sN = b of [a, b] so
that γ|[sj−1 ,sj ] is entirely contained in one of the neighbourhoods Ul . To every 1 ≤ j ≤ N we assign such
a neighbourhood Uj .
Every transition map ψ ◦ φ−1 j : φj (V ∩ Uj ) → ψ(V ∩ Uj ) is smooth, hence Lipschitz continuous on the
compact subspace im(φj ◦ γ|[sj−1 ,sj ] ). Let (Lj )N j=1 be a family of associated positive Lipschitz constants.
Now we prove that ψ ◦ γ : [a, b] → ψ(V ) is absolutely continuous. To this end, let ε > 0 and let δ > 0
m
be small enough so that for every 1 ≤ j ≤ N , every Pm m ∈ N and every family {(ai , bi )}i=1 of disjoint
intervals with [ai , bi ] ⊂ [sj−1 , sj ] and total length i=1 |bi − ai | < δ, it holds
m
X ε
∥(φj ◦ γ)(bi ) − (φj ◦ γ)(ai )∥ < .
i=1
Lj N

Pm m ∈ N and let
Let {(ai , bi )}m
be a family of disjoint intervals with [ai , bi ] ⊂ [a, b] and total length
i=1
i=1 |bi −ai | < δ. After dividing the intervals (ai , bi ) if necessary, we can assume that for each 1 ≤ i ≤ m,
there is some 1 ≤ j ≤ N with [ai , bi ] ⊂ [sj−1 , sj ]. Then it holds
m
X
∥(ψ ◦ γ)(bi ) − (ψ ◦ γ)(ai )∥
i=1
N
X X
= ∥(ψ ◦ γ)(bi ) − (ψ ◦ γ)(ai )∥
j=1 [ai ,bi ]
⊂[sj−1 ,sj ]
N
X X
= ∥(ψ ◦ φ−1 −1
j )((φj ◦ γ)(bi )) − (ψ ◦ φj )((φj ◦ γ)(ai ))∥
j=1 [ai ,bi ]
⊂[sj−1 ,sj ]
N N
X X X ε
≤ Lj · ∥(φj ◦ γ)(bi ) − (φj ◦ γ)(ai )∥ < Lj · = ε. □
j=1 j=1
Lj N
[ai ,bi ]
⊂[sj−1 ,sj ]

Remark 2.7. Burtscher [7, Prop. 3.4] is a consequence of Lemma 2.6.


Proposition 2.8. Let I ⊂ R be a closed interval. Every absolutely continuous curve γ : I → M is
differentiable almost everywhere. If g is a Riemannian metric on M , then |γ̇|g : I → R is integrable.
Proof. See [7, Prop 3.7]. □
We fix a Riemannian metricR g on M . By Proposition 2.8 each absolutely continuous curve γ has a
well-defined length ℓg (γ) := I |γ̇(s)|g ds. Then (Aac , ℓg ) is a length structure on M and so induces a
distance function dac .
We want to show that dac coincides with the usual Riemannian distance dg which is determined by
piecewise smooth curves. Note that piecewise smooth curves are absolutely continuous since they are
absolutely continuous on a partition of their domain.
Definition 2.9. The variational metric on Aac is defined by
Z 1
Dac : Aac × Aac → R , Dac (γ1 , γ2 ) = sup dg (γ1 (s), γ2 (s)) + ||γ̇1 (s)|g − |γ̇2 (s)|g | ds.
s∈[0,1] 0

This is indeed a metric, i.e. (Aac , Dac ) is a metric space.


Remark 2.10. The length functional ℓg : Aac → R is 1-Lipschitz with respect to the variational metric
Dac .
4
Proposition 2.11. For every p, q ∈ M , the set A∞
p,q
of piecewise smooth curves γ : [0, 1] → M with
p,q
γ(0) = p and γ(1) = q is dense in (Aac , Dac ).
Proof. See [7, Thm. 3.11]. The proof also applies in the case where M has non-empty boundary. The
only thing to notice is the following: If γ : [0, 1] → Rnx1 ≥0 is a curve in the right half-plane, then so is the
convolution ρ ∗ γ with a Friedrichs mollifier ρ where the convolution is componentwise. □

R s 2.12. Given an absolutely continuous curve γ : [0, 1] → M and ε > 0 we find a δ > 0 such
Remark
that s12 |γ̇|g < ε for all s1 , s2 ∈ [0, 1] with |s1 − s2 | < δ. According to Burtscher, this is due to the
fundamental theorem of calculus for absolutely continuous functions. In fact, the statement holds by
integrability of |γ̇|g and the dominated convergence theorem.
However, the fundamental theorem of calculus can be used at another point: If f : R → R is an
absolutely continuous and bounded function, one can show that the convolution f ∗ ρ with a Friedrichs
mollifier ρ is smooth and that its derivative is given by (f ∗ ρ)′ = f ∗ ρ′ .
Then choose a, b ∈ R with supp ρ ⊂ [a, b] and let s ∈ R. By the fundamental theorem of calculus for
absolutely continuous functions, we find
(f ∗ρ′ )(s) − (f ′ ∗ ρ)(s)
Z b
= f (s − σ)ρ′ (σ) − f ′ (s − σ)ρ(σ) dσ
a
Z b
d
= (f (s − σ)ρ(σ)) dσ
a dσ
= f (s − b)ρ(b) − f (s − a)ρ(a) = 0
respectively (f ∗ ρ′ )(s) = (f ′ ∗ ρ)(s). Therefore (f ∗ ρ)′ = f ′ ∗ ρ. This fact is used in equation (3.5b) of
[7, Thm. 3.11].
By Remark 2.10 and Proposition 2.11, we obtain
Corollary 2.13. dac = dg .
The next step is to perform the reflection argument for absolutely continuous curves. From now on,
g is a doubling metric on a manifold M with non-empty boundary. Put G := g ∪ g as a metric on DM g .
Proposition 2.14. Let I ⊂ R be a closed interval and let γ : I → DM g be an absolutely continuous
curve.
(a) Then γ̃ : I → M, γ̃ := pr ◦ γ is an absolutely continuous curve where pr : DM → M denotes the
projection map onto a fixed copy of M .
˙
(b) It holds |γ̇(s)|G = |γ̃(s)| g for all s ∈ I where both γ and γ̃ are differentiable.

Proof. The two copies of M will be denoted by [M, 1] and [M, 2]. For the proof let pr : DM → [M, 1] be
the projection onto the first copy.
Regarding (a): Let p ∈ γ̃(I). If p ∈ / ∂M we find a local chart (φ, U ) of M around p with U ⊂ int(M ).
Let [a, b] ⊂ γ̃ −1 (U ). Since γ|[a,b] does not hit the boundary, it either holds im(γ|[a,b] ) ⊂ [U, 1] or
im γ|[a,b] ⊂ [U, 2] due to continuity. In any case, φ ◦ γ̃|[a,b] = φ ◦ γ|[a,b] is absolutely continuous.
In the following we consider p ∈ ∂M : The hypersurface ∂M ⊂ DM has a geodesic tubular neighbour-
hood
ϕG : V → U G , ϕG (t, p) = expG p (tNg (p))
where V ⊂ R × ∂M is an open neighbourhood of {0} × ∂M and Ng is the inward pointing unit normal
vector field along the boundary of [M, 1] with respect to g. Moreover, V is supposed to be symmetric
under reflection along {0} × ∂M .
We have ϕG (V ∩ {t ≥ 0}) ⊂ [M, 1] and ϕG (V ∩ {t ≤ 0}) ⊂ [M, 2], as well as
ϕG (t, p) = (pr ◦ ϕG )(−t, p) (1)
g G
for all (t, p) ∈ V with t ≥ 0 by symmetry of G. Put U := U ∩ [M, 1]. Then
g
ϕg : V ∩ {t ≥ 0} → U , ϕg (t, p) := ϕG (t, p)
is a geodesic collar neighbourhood of ∂M in [M, 1], see Figure 1. It holds γ̃ −1 (U g ) = γ −1 (U G ).
Within a tubular neighbourhood, γ is given by
ϕ−1
G ◦ γ: γ
−1
(U G ) → V ⊂ R × ∂M ,
(ϕ−1
G ◦ γ)(s) = (t(s), p(s))
5
V UG

∂M Ug
ϕG

ϕg

0 t

Figure 1. Tubular and collar neighbourhoods of ∂M

for some continuous functions t : γ −1 (U G ) → R and p : γ −1 (U G ) → ∂M . We find


ϕ−1
g ◦ γ̃ : γ
−1
(U G ) → V ∩ {t ≥ 0},
(ϕ−1
g ◦ γ̃(s) = (|t(s)|, p(s))

for γ̃ in the geodesic collar which is a consequence of (1).


Now we will provide the actual proof. To ensure clarity, we assume that ∂M is covered by a single
chart without significant loss of generality. Let [a, b] ⊂ γ −1 (U G ) and ε > 0. Choose δ > 0 small enough
so that the condition for absolute continuity of ϕ−1 G ◦ γ|[a,b] holds with ε and δ.
Let m ∈ N andPlet {(ai , bi )}m
i=1 be a family of disjoint intervals with [ai , bi ] ⊂ [a, b] for all 1 ≤ i ≤ m
m
and total length i=1 |bi − ai | < δ. Then we have
m
X
∥(ϕ−1 −1
g ◦ γ̃)(bi ) − (ϕg ◦ γ̃)(ai )∥
i=1
m
X
= ∥(|t(bi )|, p(bi )) − (|t(ai )|, p(ai ))∥
i=1
Xm
= ((|t(bi )| − |t(ai )|)2 + ∥p(bi ) − p(ai )∥2 )1/2
i=1
m
X
≤ (|t(bi ) − t(ai )|2 + ∥p(bi ) − p(ai )∥2 )1/2
i=1
m
X
= ∥(t(bi ), p(bi )) − (t(ai ), p(ai ))∥
i=1
Xm
= ∥(ϕ−1 −1
G ◦ γ)(bi ) − (ϕG ◦ γ)(ai )∥ < ε
i=1

meaning that ϕ−1


g ◦ γ̃|[a,b] is absolutely continuous. In total, γ̃ is absolutely continuous by Lemma 2.6.
Regarding (b): Let s0 ∈ I be a point where both γ and γ̃ are differentiable. According to proposi-
tion 2.8 and part (a), this is the case almost everywhere.
˙ 0 )|g is easy to prove if γ(s0 ) ∈
The claim |γ̇(s0 )|G = |γ̃(s / ∂M (cf. part (a)) so that we consider
γ(s0 ) ∈ ∂M . Then it holds t(s0 ) = 0. Since γ and γ̃ are differentiable in s0 , so are t, |t| and p. By the
generalized Gauss lemma, we have
ϕ∗G G = dt2 + Gt ,
ϕ∗g g = dt2 + gt
where G• : V → T ∗ ∂M ⊗ T ∗ ∂M and g• : V ∩ {t ≥ 0} → T ∗ ∂M ⊗ T ∗ ∂M coincide on V ∩ {t ≥ 0}.
It holds
d 2
|γ̇(s0 )|2G = t(s) + |ṗ(s0 )|2g0 ,
ds s=s0
d 2
˙ 0 )|2g =
|γ̃(s |t(s)| + |ṗ(s0 )|2g0 .
ds s=s0
6
We show that d/ds|s=s0 t(s) = d/ds|s=s0 |t(s)| = 0.
First case: There exists δ > 0 such that t(s) ≥ 0 for all s ∈ (s0 − δ, s0 + δ) or t(s) ≤ 0 for all
s ∈ (s0 − δ, s0 + δ). Assuming t(s) ≥ 0 for all such s, the left-sided derivative in s0 is ≤ 0 whereas the
right-sided derivative is ≥ 0. Hence d/ds|s=s0 t(s) = 0. As |t| coincides with t on (s0 − δ, s0 + δ), we also
have d/ds|s=s0 |t(s)| = 0.
Second case: There is no such δ as in the first case. We can pick a sequence (sk )k with sk → s0 so
that t(sk ) ≥ 0 for k even and t(sk ) ≤ 0 for k odd. For the even part, one has
|t(s2k )| t(s2k ) k→∞ d
= −−−−→ t(s)
s2k − s0 s2k − s0 ds s=s0

and for the odd part


|t(s2k+1 )| −t(s2k+1 ) k→∞ d
= −−−−→ − t(s).
s2k+1 − s0 s2k+1 − s0 ds s=s0
By differentiability of |t|, we conclude d/ds|s=s0 t(s) = −d/ds|s=s0 t(s) or d/ds|s=s0 t(s) = 0. Therefore,
d d d
|t(s)| = t(s) = − t(s) = 0.
ds s=s0 ds s=s0 ds s=s0

˙ 0 )|g .
In either case, it holds |γ̇(s0 )|G = |ṗ(s0 )|g0 = |γ̃(s □
Remark 2.15. The doubling property for g is actually not needed in part (a), since gluing together two
copies of an appropriate collar neighbourhood of ∂M ⊂ M yields a tubular neighbourhood for ∂M in
DM . This is by construction of the smooth structure on DM .
Now Proposition 2.3 follows immediately from Corollary 2.13 and Proposition 2.14.
Proof of Theorem 2.2. Let (pk )k be a Cauchy sequence in DM with respect to dG . Without loss of
generality, we can assume that all points pk , k ∈ N are contained in a single copy of M . If not, we select
an appropriate subsequence. Then (pk )k converges by Proposition 2.3 and the completeness of g. □

3. The deformation principle


Throughout this section we fix a smooth manifold M of dimension n ≥ 2 with non-empty (possibly
non-compact) boundary. For each Riemannian metric g on M we denote by
▷ scalg : M → R its scalar curvature;
▷ g0 ∈ C ∞ (∂M ; T ∗ ∂M ⊗ T ∗ ∂M ) the metric induced on ∂M ;
▷ Ng the inward pointing unit normal vector field along ∂M ;
▷ IIg the second fundamental form of ∂M ⊂ M with respect to Ng ;
1
▷ Hg = n−1 trg0 (IIg ) : ∂M → R the mean curvature of ∂M .
The space of Riemannian metrics on M is denoted by R(M ). It is endowed with the weak C ∞ -topology.
Let σ : M → R be a continuous function that also remains unchanged. We investigate metrics in the
subspace R>σ (M ) := {g ∈ R(M ) : scalg > σ} where the topology is inherited from R(M ).
The following refined versions of geodesic collar neighbourhoods are important for the deformation
principle: Given a metric g on M and a neighbourhood U ⊂ M of ∂M there exists a continuous positive
function η : ∂M → R such that U g = Uηg is the diffeomorphic image of
Vη := {(t, p) ∈ [0, ∞) × ∂M : t < η(p)}
under the normal exponential map, i.e.
ϕg : Vη → Uηg , ϕg (t, p) = expp (tNg (p))
is a collar neighbourhood. We can arrange Uηg ⊂ U by choosing η small enough. Working in families of
metrics is more complicated as the neighbourhoods Vη and Uηg may differ between members. However,
it is possible to fix either the domain or the target in the following sense:
Fact C1. Let K be a compact Hausdorff space and let g : K → R(M ) be a continuous family of Riemann-
ian metrics. For each neighbourhood U of ∂M there exists a continuous positive function η : ∂M → R
such that
g(ξ)
▷ ϕg(ξ) : Vη → Uη is a collar neighbourhood for each ξ ∈ K;

▷ ϕ : K → C (Vη ; M ), ξ 7→ ϕg(ξ) is continuous;
▷ Uηg ⊂ U for all ξ ∈ K.
7
Fact C2. Let K be a compact Hausdorff space and let g : K → R(M ) be a continuous family of Rie-
g(ξ)
mannian metrics. Let η : ∂M → R be a continuous positive function such that ϕg(ξ) : Vη → Uη is a
geodesic collar neighbourhood for all ξ ∈ K. Then there exists a neighbourhood U ⊂ M of ∂M with the
following properties:
g(ξ)
▷ U ⊂ Uη for all ξ ∈ K;
▷ ϕ : K → C ∞ (U ; Vη ), ξ 7→ ϕ−1
−1
g(ξ) is continuous.

As with R(M ), we equip all function spaces with the weak C ∞ -topology. Facts C1 and C2 are proven
in [12, Satz 2.14 and Prop. 2.15].
Let I ⊂ R be an interval with [0, 1] ⊂ I. A continuous family of time-dependent Riemannian metrics
is a continuous map g : K → C ∞ (M × I; T ∗ M ⊗ T ∗ M ) with g(ξ)(•, s) beeing a Riemannian metric for
each ξ ∈ K and s ∈ I. We require I to be open as well as g(ξ)(p, s) = g(ξ)(p, 0) for all s ≤ 0 and
g(ξ)(p, s) = g(ξ)(p, 1) for all s ≥ 1. This condition is not inalienable but for avoiding corners.
There is a need for adapted versions of C1 and C2 with families of time-dependent metrics. As an
g(ξ)(•,s)
abbreviation, we will denote ϕξ,s := ϕg(ξ)(•,s) and Uηξ,s := Uη .
Fact C1’. Let K be a compact Hausdorff space and let g : K → C ∞ (M × I; T ∗ M ⊗ T ∗ M ) be a con-
tinuous family of time-dependent Riemannian metrics. For each neighbourhood U of ∂M there exists a
continuous positive function η : ∂M → R such that
▷ ϕξ,s : Vη → Uηξ,s is a collar neighbourhood for each ξ ∈ K and s ∈ I;
▷ ϕ : K → C ∞ (Vη × I; M ), ξ 7→ [(t, p, s) 7→ ϕξ,s (t, p)] is well-defined and continuous;
▷ Uηξ,s ⊂ U for all ξ ∈ K and s ∈ I.
Fact C2’. Let K be a compact Hausdorff space and let g : K → C ∞ (M ×I; T ∗ M ⊗T ∗ M ) be a continuous
family of time-dependent Riemannian metrics. Let η : ∂M → R be a continuous positive function such
that ϕξ,s : Vη → Uηξ,s is a geodesic collar neighbourhood for all ξ ∈ K and s ∈ I. Then there exists a
neighbourhood U ⊂ M of ∂M with the following properties:
▷ U ⊂ U ξ,s for all ξ ∈ K and s ∈ I;
▷ ϕ−1 : K → C ∞ (U × I; Vη ), ξ 7→ [(p, s) 7→ ϕ−1
ξ,s (p)] is well-defined and continuous.

To say ϕ is well-defined means that the assignment (t, p, s) 7→ ϕξ,s (t, p) is jointly smooth with respect
to the point (t, p) ∈ Vη and the parameter s ∈ I (similar with ϕ−1 ). Facts C1’ and C2’ can be deduced
from [12, Satz 2.14 and Prop. 2.15] using autonomization techniques from ODE theory.
To begin the deformation process, it is useful to standardise the normal exponential maps of a given
continuous family g : K → R(M ). Uniformisation is even necessary for some applications, see Corollary
4.3. The process concludes with Corollary 3.5.
Lemma 3.1. Let r > 0. There exists a complete Riemannian metric m ∈ R(M ) and an open cover
(Ui )i∈I such that each Ui is relatively compact and that
Ii := {j ∈ I : B m (Ui , r) ∩ B m (Uj , r) ̸= ∅}
is finite for every i ∈ I. Here B m (Ui , r) = {p ∈ M : dm (p, Ui ) < r} denotes the closed r-neighbourhood
around Ui with respect to m.
Proof. We choose a proper smooth embedding Φ : M → RN of M into some sufficiently large Euclidean
space RN . The sets

Ui := Φ−1 (B(i, N )), i ∈ I := ZN
√ √
form an open cover of M where B(i, N ) denotes the open Euclidean ball of radius N around i ∈ I.
Each fixed Ui intersects only a√finite number of other Uj , j ∈ I. The sets are relatively compact since
each closure cl(U ) ⊂ Φ−1 (B(i, N )) is contained in a compact neighbourhood.
We set
m := Φ∗ geucl
where geucl denotes the Euclidean metric√ in RN . The metric m is complete because
√ Φ(M ) ⊂ RN is
closed. It holds Φ(B m (Ui , r)) ⊂ B(i, N + r) for every i ∈ I. Since the balls B(i, N + r) only meet in
finite families, the same is true for the neighbourhoods B m (Ui , r). □
We choose r = 1 and fix a background metric m ∈ R(M ) together with an open cover (Ui )i∈I of M
as in the lemma. We would like to mention that the balls B m (Ui , 1) are compact due to the Hopf-Rinow
8
theorem, see [6, Prop. 2.5.22 and Thm. 2.5.28]. This is applicable since Riemannian manifolds with
boundary are length spaces, see e.g. [5, Prop. 3.18].
Suppose there is a distinguished metric g(ξ0 ), ξ0 ∈ K in a given family g : K → R(M ). The starting
point for the uniformisation process is to join every metric g(ξ), ξ ∈ K with the metric g(ξ0 ) by a path
of metrics.
Let I ⊂ R be an open interval with [0, 1] ⊂ I and let ρ : R → [0, 1] be a smooth function with
▷ ρ(s) = 0 for all s ≤ 0;
▷ ρ(s) = 1 for all s ≥ 1;
▷ supp ρ′ ⊂ (0, 1).
Let U be a neighbourhood of ∂M . We consider the family of time-dependent Riemannian metrics given
by (1 − ρ(s))g(ξ0 ) + ρ(s)g(ξ) for ξ ∈ K and s ∈ I.
By Fact C1’, there exists an associated family of geodesic collar neighbourhoods ϕξ,s : V → U ξ,s that
satisfy the inclusion U ξ,s ⊂ U for all ξ ∈ K and s ∈ I. We drop the subscript η at this point.
Put
ω : K → C ∞ (U g(ξ0 ) × I; M ),
ω(ξ)(p, s) = (ϕξ,s ◦ ϕ−1
g(ξ0 ) )(p).

This map is continuous due to Fact C1’.


Our proofs will utilise time-dependent vector fields. Similar to time-dependent metrics, a time-
dependent vector field V ∈ C ∞ (M × I; T M ) is supposed to satisfy V (ξ)(p, s) = V (ξ)(p, 0) for all s ≤ 0
and V (ξ)(p, s) = V (ξ)(p, 1) for all s ≥ 1. We further require V (∂M × I) ⊂ T (∂M ).
The vector field is said to have bounded velocity if there exists a complete Riemannian metric g ∈ R(M )
and a constant L ≥ 0 such that |V (p, s)|g ≤ L for all p ∈ M and s ∈ I, cf. [15].

Lemma 3.2. Let (K, ξ0 ) be a compact pointed Hausdorff space and let g : K → R(M ) be a continuous
family of Riemannian metrics. Let δ : M → R be a continuous positive function.
For each neighbourhood U of ∂M , there exists a smaller neighbourhood ∂M ⊂ U ⊂ U and a contin-
uous family
V : K → C ∞ (M × I; T M )
of time-dependent vector fields such that the following holds for all ξ ∈ K and s ∈ I:
(a) V (ξ)(p, s) = 0 for all p ∈ ∂M ;
(b) V (ξ0 )(p, s) = 0 for all p ∈ M ;
(c) The curve ω(ξ)(p, •) : I → M is a solution to the IVP
γ̇(s) = V (ξ)(γ(s), s)
γ(0) = p
for all p ∈ U ;
(d) V (ξ)(p, s) = 0 for all p ∈ M \ cl(U );
(e) |V (ξ)(p, s)|m < δ(p) for all p ∈ M ;
(f ) |V (ξ)(p, s)|g(ξ) < δ(p) for all p ∈ M .

Proof. Let U be a neighbourhood of ∂M . We will start from the point before the lemma: There exist
neighbourhoods ∂M ⊂ U ′′ ⊂ U ′ ⊂ M with cl(U ′′ ) ⊂ U ′ ⊂ U ξ,s for all ξ ∈ K and s ∈ I as can be seen
from Fact C1’,C2’ and the later Lemma 3.15. Then
v : K → C ∞ (U ′ × I; T M ),
∂ω(ξ)
v(ξ)(p, s) = ((ϕg(ξ0 ) ◦ ϕ−1
ξ,s )(p), s)
∂s
is a local family of time-dependent vector fields. For each point p ∈ ∂M , it holds ω(ξ)(p, s) = p for all
ξ ∈ K and s ∈ I so that v ≡ 0 on ∂M . By continuity and compactness of K, we can assume that
|v(ξ)(p, s)|m < δ(p),
|v(ξ)(p, s)|g(ξ) < δ(p)

for all p ∈ U ′ , ξ ∈ K and s ∈ I. This is without loss of generality.


9
Next let V ′′ ⊂ [0, ∞) × ∂M be a neighbourhood of ∂M such that ϕξ,s (V ′′ ) ⊂ U ′′ for all ξ ∈ K and
s ∈ I. Put U := ϕg(ξ0 ) (V ′′ ). We then obtain the differential equation
∂ω(ξ)
(p, s) = v(ξ)(ω(ξ)(p, s), s)
∂s
for all p ∈ U, ξ ∈ K and s ∈ I. In order to globalize the construction let χ : M → [0, 1] be a smooth
bumping function with
▷ χ = 1 on U ′′ ;
▷ supp χ ⊂ U ′ .
Define a global time-dependent vector field by
V : K → C ∞ (M × I; T M ),
V (ξ)(p, s) = χ(p) · v(ξ)(p, s).
It can be verified that V satisfies properties (a) − (f ). We present (c):
It holds ω(ξ)(p, s) ∈ U ′′ for all p ∈ U, ξ ∈ K and s ∈ I by construction. We further know that ω
satisfies the IVP
∂ω(ξ)
(p, s) = v(ξ)(ω(ξ)(p, s), s),
∂s
ω(ξ)(p, 0) = p.
Since V and v coincide on U ′′ , the curves ω(ξ)(p, •) : I → M are solutions to (c). □
Proposition 3.3. Let (K, ξ0 ) be a compact pointed Hausdorff space and let g : K → R(M ) be a contin-
uous family of Riemannian metrics. Let δ : M → R be a continuous positive function.
For each neighbourhood U of ∂M there exists a smaller neighbourhood ∂M ⊂ U ⊂ U and a continuous
map
Ω : K × [0, 1] → Diff(M )
such that the following holds for all ξ ∈ K and s ∈ I:
(a) Ω(ξ, 0) = idM ;
(b) Ω(ξ0 , s) = idM ;
(c) Ω(ξ, 1) = ϕg(ξ) ◦ ϕ−1 g(ξ0 ) on U;
(d) Ω(ξ, s) = id and dΩ(ξ, s) = id on M \ U ;
(e) dm (Ω(ξ, s)(p), p) < δ(p) for all p ∈ M ;
(f ) dg(ξ) (Ω(ξ, s)(p), p) < δ(p) for all p ∈ M .
Proof. Let U be a neighbourhood of ∂M . We apply Lemma 3.2 for the function δ/2 to obtain a family
of vector fields V and a neighbourhood ∂M ⊂ UV ⊂ U with properties 3.2(a) − (f ). Now we use the
open cover (Ui )i∈I from Lemma 3.1:
Let i ∈ I. By compactness there exists some δ̃i > 0 such that
1
||V (ξ)(p, s)|m − |V (ξ)(q, s)|m | < min{δ(p′ ) : p′ ∈ B m (Ui , 1)} (2)
2
and
1
||V (ξ)(p, s)|g(ξ) − |V (ξ)(q, s)|g(ξ) | < min{δ(p′ ) : p′ ∈ B m (Ui , 1)} (3)
2
for all p, q ∈ B m (Ui , 1) with dm (p, q) < δ̃i and for all ξ ∈ K, s ∈ I. Let (ψi )i∈I be a partition of unity
subordinate to (Bm (Ui , 1))i∈I . Set
X
δ̃ : M → R , δ̃(p) = ψi (p) · min δ̃j .
j∈Ii
i∈I

This is a well-defined continuous function with δ̃(p) > 0 for all p ∈ M . We have δ(p) ≤ δ̃i for all i ∈ I
and p ∈ B m (Ui , 1).
Since V ≡ 0 on ∂M , there is a neighbourhood ∂M ⊂ U ′ ⊂ M such that
|V (ξ)(p, s)|m < min{δ̃(p), 1},
(4)
|V (ξ)(p, s)|g(ξ) < δ̃(p)
for all p ∈ U ′ , ξ ∈ K and s ∈ I. Choose a smooth bumping function χ : M → [0, 1] with
▷ χ = 1 on some neighbourhood U ′′ of ∂M ;
▷ supp χ ⊂ U ′ .
10
Then set
VΩ : K → C ∞ (M × I; T M ),
VΩ (ξ)(p, s) = χ(p) · V (ξ)(p, s).
This family satisfies the inequalities (4) globally and it satisfies |VΩ (ξ)| ≤ |V (ξ)| for every Riemannian
metric. In particular, VΩ (ξ) has bounded velocity for every ξ ∈ K.
According to [15, Thm. 8.1.1] and [12, Satz A.5], the solution curves to the IVP
γ̇(s) = VΩ (ξ)(γ(s), s)
γ(0) = p, p ∈ M
assemble in a well-defined and continuous solution map
Ω : K → C ∞ (M × I; M )
which consists of diffeomorphisms. From this we deduce a map
Ω : K × [0, 1] → Diff(M )
which is again denoted by Ω, and check properties (a) − (f ):
(a) is by definition of flows.
(b) follows from Lemma 3.2(b).
(c) As in the proof of Lemma 3.2, we find a neighbourhood U ⊂ UV ∩ U ′′ such that
∂ω(ξ)
(p, s) = VΩ (ω(ξ)(p, s), s),
∂s
ω(ξ)(p, 0) = p
for all p ∈ U, ξ ∈ K and s ∈ I. By uniqueness of integral curves, it holds Ω(ξ)(p, s) = ω(ξ)(p, s) for
p ∈ U . In particular, Ω(ξ, 1) = ϕg(ξ) ◦ ϕ−1
g(ξ0 ) on U for all ξ ∈ K.
(d) follows from the fact that VΩ ≡ 0 on M \ cl U .
(e) Let p ∈ M . Then it holds p ∈ Ui for some i ∈ I. For any ξ ∈ K and s ∈ [0, 1], we have
dm (Ω(ξ, s)(p), p) = dm (Ω(ξ, s)(p), Ω(ξ, 0)(p))
Z s
∂Ω(ξ)
≤ (p, t) dt
0 ∂t m
Z s
= |VΩ (ξ)(Ω(ξ, t)(p), t)|m dt
Z0 s
≤ 1 dt ≤ 1.
0

This means Ω(ξ, s)(p) ∈ B m (Ui , 1) for all ξ ∈ K and s ∈ [0, 1]. Therefore, it holds
Z s
dm (Ω(ξ, s)(p), p) ≤ |VΩ (ξ)(Ω(ξ, t)(p), t)|m dt
Z0 s
< δ̃(Ω(ξ, t)(p)) dt
0
Z s
≤ δ̃i dt ≤ δ̃i
0
for all ξ ∈ K and s ∈ [0, 1]. Finally, we obtain
Z s
dm (Ω(ξ, s)(p), p) ≤ |VΩ (ξ)(Ω(ξ, t)(p), t)|m dt
0
Z s
≤ |V (ξ)(Ω(ξ, t)(p), t)|m dt
Z0 s
≤ |V (ξ)(p, t)|m + ||V (ξ)(p, t)|m − |V (ξ)(Ω(ξ, t)(p), t)|m | dt
Z0 s
1 1
< δ(p) + min{δ(p′ ) : p′ ∈ B m (Ui , 1)} dt
0 2 2
≤ δ(p)
for all ξ ∈ K and s ∈ [0, 1] by inequality (2).
11
(f ) Let p ∈ M . Choose some i ∈ I with p ∈ Ui . We know that Ω(ξ, s)(p) ∈ B m (Ui , 1) and that
dm (Ω(ξ, s)(p), p) < δ̃i for all ξ ∈ K and s ∈ [0, 1]. The assertion follows from (3). □
Remark 3.4. Hirsch [15, Thm. 8.1.1] refers to time-dependent vector fields that are parametrized over
compact intervals. The result is also true for open intervals.
Corollary 3.5. Let K be a compact Hausdorff space and let g : K → R>σ (M ) be a continuous family
of Riemannian metrics of scalar curvature greater than σ. Let β ∈ R(M ) be another distinguished
Riemannian metric.
For each neighbourhood U of ∂M there exists a smaller neighbourhood ∂M ⊂ U ⊂ U and a continuous
map
f : K × [0, 1] → R>σ (M )
so that the following holds for all ξ ∈ K and s ∈ [0, 1]:
(a) f (ξ, 0) = g(ξ);
(b) Nf (ξ,1) = Nβ and ϕf (ξ,1) = ϕβ on ϕ−1 β (U );
(c) f (ξ, s)t = g(ξ)t near {0} × ∂M ;
(d) f (ξ, s) = g(ξ) on M \ U ;
(e) f (ξ, s) is quasi-isometric to g(ξ) via the identity;
(f ) if g(ξ) is complete then so is f (ξ, s).
Property (e) means that there exist constants A ≥ 1 and B ≥ 0 such that
1
dg(ξ) (p, q) − B ≤ df (ξ,s) (p, q) ≤ A dg(ξ) (p, q) + B
A
for all p, q ∈ M . In fact, we can arrange these inequalities for A = 1 and an arbitrary constant B > 0
independent from ξ and s. The proof below works with B = 1.
Proof. Let U be a neighbourhood of ∂M . Let ξ0 be an arbitrary object. Then K ∪ {ξ0 } with the
topology of disjoint union is again a compact Hausdorff space. We put g(ξ0 ) := β.
Each metric f (ξ, s) will be a pullback of g(ξ) along some diffeomorphism of M . For the construction
we consider the open cover (Ui )i∈I from Lemma 3.1. Another important quantity is the scalar curvature
surplus
D : K → C ∞ (M ; R) , D(ξ)(p) = scalg(ξ) (p) − σ(p).
Now let i ∈ I. By compactness, there exists some δi > 0 such that
|scalg(ξ) (p) − scalg(ξ) (q)| < min{D(ξ ′ )(p′ ) : p′ ∈ B m (Ui , 1), ξ ′ ∈ K}
for all p, q ∈ B m (Ui , 1) with dm (p, q) < δi and for all ξ ∈ K.
Let (ψi )i∈I be a partition of unity subordinate to (Ui )i∈I . Set
1X 
δ : M → R , δ(p) = ψi (p) · min min δj , 1 .
2 j∈Ii
i∈I

This is a well-defined continuous function with 0 < δ(p) ≤ 21 for all p ∈ M . Furthermore, it holds
δ(p) < δi for all i ∈ I and p ∈ Ui .
Proposition 3.3, applied for the compact Hausdorff space K ∪ {ξ0 } and the function δ, yields a
neighbourhood ∂M ⊂ U ⊂ U and a family Ω : (K ∪ {ξ0 }) × [0, 1] → Diff(M ) with its distinguished
properties 3.3(a) − (f ). We put
f : K × [0, 1] → R(M ) , f (ξ, s) := Ω(ξ, s)∗ g(ξ).
This operation preserves the lower scalar curvature bound σ which is shown as follows: Let p ∈ M, ξ ∈ K
and s ∈ [0, 1]. Then it holds scalf (ξ,s) (p) = scalg(ξ) (Ω(ξ, s)(p)).
Choose some i ∈ I with p ∈ Ui . By Proposition 3.3(e), we have dm (Ω(ξ, s)(p), p) < δ(p) < 1, i.e.
Ω(ξ, s)(p) ∈ B m (Ui , 1). The same inequality gives dm (Ω(ξ, s)(p), p) < δ(p) < δi so that
|scalg(ξ) (Ω(ξ, s)(p)) − scalg(ξ) (p)|
< min{D(ξ ′ )(p′ ) : p′ ∈ B m (Ui , 1), ξ ′ ∈ K} ≤ D(ξ)(p).
Hence, we obtain scalf (ξ,s) (p) > σ(p).
It remains to check properties (a) − (f ):
(a) follows from Proposition 3.3(a);
12
(b) Let ξ ∈ K and (t, p) ∈ ϕ−1
β (U ). The map Ω(ξ, 1) : (M, f (ξ, 1)) → (M, g(ξ)), as an isometry, maps
normal geodesics to normal geodesics. By Proposition 3.3(c), we have
Ω(ξ, 1)(ϕβ (t, p)) = ϕg(ξ) (t, p) = expg(ξ)
p (tNg(ξ) (p))
= Ω(ξ, 1)(exppf (ξ,1) (tNf (ξ,1) (p)))
= Ω(ξ, 1)(ϕf (ξ,1) (t, p))
since Ω preserves boundary points. Thus, ϕf (ξ,1) (t, p) = ϕβ (t, p). In particular,
∂ ∂
Nf (ξ,1) (p) = ϕf (ξ,1) (t, p) = ϕβ (t, p) = Nβ (p).
∂t t=0 ∂t t=0

(c) Let ξ ∈ K. We choose a neighbourhood V ⊂ [0, ∞) × ∂M of ∂M such that ϕf (ξ,s) : V → U f (ξ,s) is


a geodesic collar neighbourhood with U f (ξ,s) ⊂ U for all s ∈ [0, 1]. Consider the diagram
ϕg(ξ)
(V, dt2 + g(ξ)t ) (U g(ξ) , g(ξ))

id Ω(ξ,s)

ϕf (ξ,s)
(V, dt2 + f (ξ, s)t ) (U f (ξ,s) , f (ξ, s))

It commutes since Ω(ξ, s) maps normal geodesics to normal geodesics. As the horizontal maps and
the vertical map on the right are isometries, so is the identity on the left.
(d) follows from Proposition 3.3(d).
(e) Let p, q ∈ M, ξ ∈ K and s ∈ [0, 1]. According to Proposition 3.3(f ), it holds
df (ξ,s) (p, q) = dg(ξ) (Ω(ξ, s)(p), Ω(ξ, s)(q))
≤ dg(ξ) (p, q) + dg(ξ) (Ω(ξ, s)(p), p) + dg(ξ) (Ω(ξ, s)(q), q)
≤ dg(ξ) (p, q) + 1.
The other inequality is similar.
(f ) Let ξ ∈ K and s ∈ [0, 1]. Since Ω(ξ, s) : (M, f (ξ, s)) → (M, g(ξ)) is a Riemannian isometry, it is
also a metric isometry. Completeness is transferred from g(ξ) to f (ξ, s). □

In a second step, we establish a standard type of Riemannian metrics within the deformation principle.
These are called C-normal metrics. They provide better control on scalar curvature.
For the definition, we consider a Riemannian metric g on M with its representation g = dt2 + gt in
a geodesic collar neighbourhood ϕg : V → U g . Then it holds IIg = − 21 ġ0 as a (0, 2)-tensor field on ∂M .
The associated Weingarten map of ∂M is denoted by Wg . Its mean curvature is given by
1 1
Hg = n−1 tr(Wg ) = − 2(n−1) trg0 (ġ0 ).

More generally, for each (t, p) ∈ V we find a neighbourhood p ∈ Vp ⊂ ∂M such that {t} × Vp ⊂ V .
The second fundamental form for this slice is IIt = − 12 ġt (with respect to the normal ∂t

). They build
a map II• : V → T ∂M ⊗ T ∂M . So do the corresponding Weingarten maps W• : V → T ∗ ∂M ⊗ T ∂M ,
∗ ∗

uniquely defined by
1
⟨Wt (p)(v), w⟩gt = IIt (p)(v, w) = − ġt (p)(v, w).
2
The scalar curvature of (M, g) within U g can be computed by means of [2, Prop. 4.1]:
scalg = scalgt + 3 tr(Wt2 ) − tr(Wt )2 − trgt (g̈t ). (5)
Definition 3.6. Let C ∈ C ∞ (∂M ; R). A Riemannian metric g ∈ R(M ) is called C-normal if the
pullback metrics g• with respect to some small geodesic collar neighbourhood ϕg : V → U g are given by
gt (p) = g0 (p) + t · ġ0 (p) − C(p)t2 · g0 (p)
= g0 (p) − 2t · IIg (p) − C(p)t2 · g0 (p), (t, p) ∈ V.
Every Riemannian metric on M can be deformed into a C-normal metric while preserving lower scalar
curvature bounds, with the original 1-jet along the boundary:
13
Proposition 3.7. Let K be a compact Hausdorff space and let g : K → R>σ (M ) be a continuous family
of Riemannian metrics of scalar curvature greater than σ.
Then there exists a smooth function C0 ∈ C ∞ (∂M ; R) such that for each function C ∈ C ∞ (∂M ; R)
with C ≥ C0 and each neighbourhood U of ∂M there exists a continuous map
f : K × [0, 1] → R>σ (M )
such that the following holds for all ξ ∈ K and s ∈ [0, 1]:
(a) f (ξ, 0) = g(ξ);
(b) f (ξ, 1) is C-normal;
(c) if g(ξ) is C̃-normal then f (ξ, s) is ((1 − s)C̃ + sC)-normal;
(d) f (ξ, s)0 = g(ξ)0 and IIf (ξ,s) = IIg(ξ) ;
(e) f¨(ξ, s)0 = (1 − s)g̈(ξ)0 − 2sCg(ξ)0 ;
(ℓ) (ℓ)
(f ) f (ξ, s)0 = (1 − s)g(ξ)0 for all ℓ ≥ 3;
(g) f (ξ, s) = g(ξ) on M \ U ;
(h) f (ξ, s) is quasi-isometric to g(ξ) via the identity;
(i) if g(ξ) is complete then so is f (ξ, s).
Proof. Based on Theorem 3.5, it can be assumed that all metrics g(ξ), ξ ∈ K have the same normal
exponential map.
Let U be a neighbourhood of ∂M and let ϕ : V → U be a common geodesic collar neighbourhood
for all metrics in g which satisfies U ⊂ U . One can identify the metrics g(ξ), ξ ∈ K on U with their
associated generalized cylinder metrics on V .
Let (Ui )i∈I be a cover of ∂M consisting of open and relatively compact subsets, and let (ψi )i∈I be a
partition of unity subordinate to (Ui )i∈I . For each i ∈ I, we define
1
Ci := 2(n−1) max ∥trg(ξ)0 (g̈(ξ)0 )∥C 0 (Ui ) .
ξ∈K

It holds Ci < ∞ due to compactness of cl(Ui ). Then put


X
C0 : ∂M → R , C0 := ψi · Ci .
i∈I

This is a smooth function with


2(n − 1) · C0 (p) ≥ |trg(ξ)0 (p) (g̈(ξ)0 (p))|
for all ξ ∈ K and p ∈ ∂M .
Now let C : ∂M → R be a smooth function with C ≥ C0 . We consider the Taylor expansion of g•
around t = 0:
1
g(ξ)t = g(ξ)0 + ġ(ξ)0 · t + g̈(ξ)0 · t2 + R(ξ)t .
2
Here R(ξ)• is a smooth map R(ξ)• : V → T ∗ ∂M ⊗ T ∗ ∂M made out of symmetric (0, 2)-tensors on ∂M
that depends continuously on ξ. It holds
R(ξ)0 = Ṙ(ξ)0 = R̈(ξ)0 = 0
for all ξ ∈ K. Put
F (ξ, s) := g(ξ) − s 21 g̈(ξ)0 + C · g(ξ)0 · t2 + R(ξ)t
 

for ξ ∈ K and s ∈ [0, 1]. This defines a continuous family F : K → C ∞ (V ; T ∗ V ⊗ T ∗ V ) of symmetric


(0, 2)-tensor fields. It can also be seen as a family of (0, 2)-tensor fields on U .
Shrinking U if necessary, we can assume that F (ξ, s) is a Riemannian metric on U for each ξ ∈ K and
s ∈ [0, 1]. This is by continuity and compactness of K. Furthermore, (5) gives
scalF (ξ,s) |∂M = scalF (ξ,s)0 + 3 tr(WF2 (ξ,s) ) − tr(WF (ξ,s) )2 − trF (ξ,s)0 (F̈ (ξ, s)0 )
2
= scalg(ξ)0 + 3 tr(Wg(ξ) ) − tr(Wg(ξ) )2 − trg(ξ)0 (g̈(ξ)0 )
− trg(ξ)0 (−sg̈(ξ)0 − 2sC · g(ξ)0 )

= scalg(ξ) |∂M + s · trg(ξ)0 (g̈(ξ)0 ) + 2C · (n − 1) ≥ scalg(ξ) |∂M > σ|∂M .
Continuity allows to assume that F (ξ, s) ∈ R>σ (U ) for all ξ ∈ K and s ∈ [0, 1]. Along the boundary, we
find
▷ F (ξ, s)|∂M = g(ξ)|∂M ;
▷ Ḟ (ξ, s)0 = ġ(ξ)0 ;
14
▷ F̈ (ξ, s)0 = (1 − s)g̈(ξ)0 − 2sCg(ξ)0 ;
(ℓ) (ℓ)
▷ F (ξ, s)0 = (1 − s)g(ξ)0 for all ℓ ≥ 3.
Finally, since F (ξ, s)|∂M = g(ξ)|∂M , it can be arranged that
1
|v|g(ξ) ≤ |v|F (ξ,s) ≤ 2|v|g(ξ) (6)
2
for all v ∈ T U, ξ ∈ K and s ∈ [0, 1]. This time we use continuity for the maps
|v|F (ξ,s)(p)
K × [0, 1] × Ũ × S n−1 → R , (ξ, s, p, v) 7→
|v|g(ξ)(p)
where Ũ ⊂ U is an arbitraty coordinate neighbourhood. Observe that one can restrict to the Euclidean
sphere S n−1 which is compact and independent from ξ.
The condition scal > σ defines a second order open partial differential relation on the space of pointwise
metrics over M . The metrics in g solve it globally, the metrics in F solve it locally over U .
By the family version of the local flexibility lemma [3, Addendum 3.4] of Bär-Hanke, we obtain an
open neighbourhood ∂M ⊂ U0 ⊂ U and a family
f : K × [0, 1] → C ∞ (M ; T ∗ M ⊗ T ∗ M )
such that the following holds for all ξ ∈ K and s ∈ [0, 1]:
▷ f (ξ, s) ∈ R>σ (M );
▷ f (ξ, 0) = g(ξ);
▷ f (ξ, s)|U0 = F (ξ, s)|U0 ;
▷ f (ξ, s)|M \U = g(ξ)|M \U .
In particular, f (ξ, s)|M \U = g(ξ)|M \U . We conclude properties (a) − (g) for this f .
The remaining properties (h) − (i) can be deduced from the actual construction in the flexibility
lemma: Accordingly there exists a smooth function τ : M → [0, 1] such that

F (ξ, sτ (p))(p), p ∈ U,
f (ξ, s)(p) =
g(ξ)(p), else.
Together with (6), this yields
1
|v|g(ξ) ≤ |v|f (ξ,s) ≤ 2|v|g(ξ)
2
for all v ∈ T M, ξ ∈ K and s ∈ [0, 1]. Properties (h) − (i) follow immediately. □
Remark 3.8. The identification g = dt2 + gt refers to a geodesic collar neighbourhood ϕg : V → U g of
g. The deformations F (ξ, s) are defined in terms of the collars ϕg(ξ) . As soon as F (ξ, s) is a Riemannian
metric on U = U g(ξ) with its own geodesic collars, one may ask for the meaning of F (ξ, s)t . In fact,
there is no subtlety because normal geodesics of F (ξ, s) and g(ξ) coincide near the boundary.
Remark 3.9. Given L > 1 we can arrange the deformation so that
1
|v|g(ξ) ≤ |v|f (ξ,s) ≤ L|v|g(ξ) .
L
Remark 3.10. Strictly speaking, the local flexibility lemma from Bär-Hanke applies to smooth manifolds
without boundary. To achieve an admissible setting, we attach a small cylinder to ∂M and extend the
metrics g(ξ) to the new manifold M + in continuous dependence of ξ.
Specifically, there is a continuous map g + : K → R(M + ) such that the canonical inclusion (M, g(ξ)) →
(M + , g + (ξ)) is an isometric embedding for every ξ ∈ K. We can guarantee continuity for g + by means
of the Seeley extension theorem [17]. See [12, Prop. 2.8] for details.
The last deformation step is for adjusting the 1-jet along the boundary, again respecting lower scalar
curvature bounds. Several lemmas are required.
Lemma 3.11. Let V be a finite dimensional real vector space and let g1 and g0 be two Euclidean scalar
products on V such that ∥g1 − g0 ∥g0 ≤ 12 . Then
|trg1 (h) − trg0 (h)| ≤ 2 · ∥g1 − g0 ∥g0 · ∥h∥g0
holds for all symmetric bilinear forms h on V . Here ∥ · ∥g0 denotes the Frobenius norm on the space of
symmetric bilinear forms induced by g0 .
Proof. See [4, Lemma 24]. □
15
Lemma 3.12. There exists a constant c0 > 0 such that for each 0 < δ ≤ 21 there exists a smooth function
χδ : [0, ∞) → R with

▷ χδ (t) = t for t near 0, χδ (t) = 0 for t ≥ δ and 0 ≤ χδ (t) ≤ 2δ for all t,
▷ |χ̇δ (t)| ≤ c0 for all t, √
▷ − 2δ ≤ χ̈δ (t) ≤ 0 for all t ∈ [0, δ] and |χ̈δ (t)| ≤ c0 for all t ∈ [δ, δ].
Proof. See [4, Lemma 25]. □
Remark 3.13. Given 0 < δ ≤ 12 and a function χδ as in Lemma 3.12 it holds χδ (t) ≤ t for all t ∈ [0, ∞):
Since χ̈δ ≤ 0 on [0, δ] we find χ̇δ (t) ≤ χ̇δ (0) = 1 for all t ∈ [0, δ]. Thus, χδ (t) ≤ t by the mean value
theorem. For t ∈ (δ, ∞), we have χδ (t) ≤ 2δ ≤ t.
Lemma 3.14. There exist countable covers (Ui1 )i∈N ⊂ (Ui2 )i∈N ⊂ (Ui3 )i∈N of ∂M such that the following
holds for all i ∈ N:
▷ Each subset Ui1 , Ui2 , Ui3 ⊂ ∂M is open and relatively compact;
▷ Ii := {j ∈ N : Ui3 ∩ Uj3 ̸= ∅} is finite;
▷ cl(Ui1 ) ⊂ Ui2 and cl(Ui2 ) ⊂ Ui3 .
M → RN be a proper smooth embedding into
Proof. The proof is similar to that of Lemma 3.1. Let Φ : √
some Euclidean space. Then the sets Ui := Φ (B(i, k N )) for k = 1, 2, 3 and i ∈ ZN ∼
k −1
= N do the
job. □
We fix three covers (Ui1 )i∈N ⊂ (Ui2 )i∈N ⊂ (Ui3 )i∈N of ∂M as in Lemma 3.14 together with a partition
of unity ψ = (ψi )i∈N subordinate to (Ui1 )i∈N .
The next lemma concerns the construction of closed collar neighbourhoods in manifolds with boundary.
This is not a trivial matter, as demonstrated by the following example: Let g be a Riemannian metric
on M . Suppose there exists a geodesic collar ϕg : [0, ε) × ∂M → M of uniform width ε > 0.
If ∂M is compact, then ϕg ([0, 2ε ] × ∂M ) is compact, hence closed in M . However, if ∂M is non-
compact, then ϕg ([0, 2ε ] × ∂M ) is not closed in general, as can be seen with M = R2x1 ≥0 \ ({0} × R≤0 ) and
g = geucl . If the deformation was done within such a collar, the resulting metric would not be smooth
and not even continuous along the x1 -axis. Nevertheless, there does exist a closed collar in our example,
namely the set A = {(x, y) ∈ R2 : 0 < x ≤ y}. Another closed collar is given by
 [ h 1 i h 1 1 i h 1 i 
B= 0, × , ∪ 0, × [1, ∞) .
i i i−1 2
i≥2

We generalize the collar neighbourhood B to arbitrary manifolds.

∂M ∂M A
B

ε x1 x1

Figure 2. Collar neighbourhoods in M = R2x1 ≥0 \ ({0} × R≤0 ); A and B are closed


whereas the collar in the left picture is not

Lemma 3.15. Let g be a Riemannian metric on M and let ϕg : Vη → Uηg be a geodesic collar neigh-
1,g
bourhood. For each i ∈ N let 0 < εi < min{i−1 , inf p∈Ui3 η(p)} and U εi := ϕg ([0, εi ] × cl(Ui1 )). Then
1,g
∪i∈N U εi ⊂ M is a closed neighbourhoood of ∂M .
16
1,g 1,g
Proof. Let (qn )n∈N be a sequence in ∪i∈N U εi with qn → q ∈ M for n → ∞. We show q ∈ ∪i∈N U εi . To
this end, let (ϕ−1
g (qn ))n = (tn , pn ) denote the associated sequence in the preimage.
1,g
First case: There is N ∈ N such that pn ∈ ∪i≤N cl(Ui1 ) for all n ∈ N. Then it holds qn ∈ ∪j∈Ii ,i≤N U εj
1,g
for all n ∈ N. The last set is closed as it is a finite union of compact sets. Therefore, q ∈ ∪i∈N U εi .
Second case: There is no such N ∈ N. By the estimate εi < i−1 , we find a subsequence tφ(n) of tn
such that tφ(n) → 0 for n → ∞, that is dg (qφ(n) , ∂M ) → 0 for n → ∞. The triangle inequality gives

dg (q, ∂M ) ≤ dg (q, qφ(n) ) + dg (qφ(n) , ∂M ) for all n ∈ N


1,g
so that dg (q, ∂M ) = 0. Since ∂M ⊂ M is closed, this yields q ∈ ∂M ⊂ ∪i∈N U εi . □

Proposition 3.16. Let K be a compact Hausdorff space. Let g0 : K → C ∞ (∂M ; T ∗ ∂M ⊗ T ∗ ∂M ) be


a continuous family of Riemannian metrics on ∂M and let h, k : K → C ∞ (∂M ; T ∗ ∂M ⊗ T ∗ ∂M ) be
continuous families of symmetric (0, 2)-tensor fields satisfying trg0 (h) ≥ trg0 (k).
Then there exists a smooth function C0 = C0 (g, h, k) ∈ C ∞ (∂M ; R), C0 > 0 such that
▷ for every continuous family
g : K → R>σ (M )
of C-normal metrics of scalar curvature greater than σ with C ∈ C ∞ (∂M ; R), C ≥ C0 , g(ξ)0 = g0 (ξ)
and IIg(ξ) = h(ξ) for all ξ ∈ K and
▷ for each neighbourhood U of ∂M
there exists a continuous map
f : K × [0, 1] → R>σ (M )
so that the following holds for all ξ ∈ K and s ∈ [0, 1]:
(a) f (ξ, 0) = g(ξ);
(b) f (ξ, s) is C-normal;
(c) f (ξ, s)0 = g(ξ)0 ;
(d) IIf (ξ,s) = (1 − s)IIg(ξ) + sk(ξ);
(e) f (ξ, s) = g(ξ) on M \ U ;
(f ) f (ξ, s) is quasi-isometric to g(ξ) via the identity;
(g) if g(ξ) is complete then so is f (ξ, s).

Proof. Let U be a neighbourhood of ∂M . Let C ∈ C ∞ (∂M ; R) be a positive smooth function and let
g : K → R>σ (M ) be a continuous family of C-normal metrics of scalar curvature greater than σ with
g(ξ)0 = g0 (ξ) and IIg(ξ) = h(ξ) for all ξ ∈ K.
Given that this proposition follows Corollary 3.5 and Proposition 3.7 within the deformation principle,
we can assume that all metrics in g have the same normal exponential map and that they satisfy the
condition of C-normality on a common neighbourhood U0 of ∂M .
We choose a continuous positive function η : ∂M → R such that ϕ : Vη → Uη is a common geodesic
collar neighbourhood for g with Uη ⊂ U0 ∩ U . For each i ∈ N let 0 < ηi < inf p∈Ui3 η(p) be some fixed
distance from the boundary.
Later, we will consider the C k -norm ∥ · ∥C k for tensor fields on ∂M , which is supposed to be the
maximum of C k -norms induced by the metrics g0 (ξ), ξ ∈ K. Set

Ci := ∥C∥C 2 (Ui3 ) .

The deformation will be performed over each piece Ui2 ⊂ ∂M , gluing everything together by the partition
of unity ψ. To this end, let δ = (δi )i∈N be a family of real numbers satisfying
 
1 −2
0 < δi < min , i , min ηj2 , min Cj−2
2 j∈Ii j∈Ii

for all i ∈ N. For each i ∈ N, let χδi be a function as in Lemma 3.12. We define

f δi : K × [0, 1] → C ∞ (Uη2i ; T ∗ M ⊗ T ∗ M ),
f δi (ξ, s) = dt2 + (1 − Ct2 ) · g0 (ξ) − 2t · h(ξ) + 2sχδi (t) · (h(ξ) − k(ξ))
17
where Uη2i := ϕ([0, ηi ) × Ui2 ). Given (t, p) ∈ [0, ηi ) × Ui2 it holds
∥f δi (ξ, s)(t, p) − g(ξ)(t, p)∥g(ξ)(t,p) = 2sχδi (t) · ∥h(ξ)(p) − k(ξ)(p)∥g(ξ)t (p)
≤ δi · sup ∥h(ξ)(p) − k(ξ)(p)∥g(ξ)t (p) .
ξ∈K,t∈[0,ηi ]
p∈Ui2

Since positive definiteness is an open condition, f δi becomes a family of Riemannian metrics on Uη2i for
sufficiently small δi . Then put
f δ : K × [0, 1] → C ∞ (M ; T ∗ M ⊗ T ∗ M ),
 2
dt + (1 − Ct2 ) · g0 (ξ) − 2t · h(ξ) +
P
 ψi · 2sχδi (t) · (h(ξ) − k(ξ))
i∈N


 S 2
f δ (ξ, s) = on Uηi ,
 i∈N S 2
 g(ξ) on M − U ηi


i∈N

Uη2i , as
S
for δ = (δi )i∈N as small as above. This expression is positive definite at each point (t, p) ∈ i∈N
can be seen from the formula
X
f δ (ξ, s) = ψj · (dt2 + (1 − Ct2 ) · g0 (ξ) − 2t · h(ξ) + 2sχδj (t) · (h(ξ) − k(ξ))). (7)
j∈N

If p ∈ Uj2 for some j ∈ N and t < ηj , then the respective summand is positive definite by the above. If
t ≥ ηj , then χδj (t) = 0 and the other three summands equal g(ξ)(t, p). There are no more summands to
take into account because ψj (p) = 0 for all j ∈ N with p ∈/ Uj2 .
We also verify the smoothness of f (ξ, s) and the continuity of f δ with respect to (ξ, s). This is
δ

obvious once we know [ 1


f δ (ξ, s) = g(ξ) on M − U √δi ,
i∈N
1
the latter beeing open in M by Lemma 3.15. For justification, let (t, p) ∈ i∈N Uη2i − i∈N U √δi . That
S S

is (t, p) ∈ Uη2i for some i ∈ N. Every summand in (7) is either ψj (p) · g(ξ)(t, p) (if p ∈ Uj1 ) or zero (if
/ Uj1 ). Therefore, f δ (ξ, s)(t, p) = j∈N ψj (p) · g(ξ)(t, p) = g(ξ)(t, p).
P
p∈

U32 U12 U22

η1
η3 √ η2
δ1 √
√ δ2
δ3

U31 U11 U21

Figure 3. Construction for f δ by boxes

The main task is to arrange the deformation so that scalf δ (ξ,s) > σ for all ξ ∈ K and s ∈ [0, 1]. Let
i ∈ N. In the following we use the notation ≲ to mean that the left hand side is bounded by the right
hand side multiplied with a positive constant that only depends on g0 , h and k on Ui3 . In particular, it is
independent from δ, t, p, ξ, s and C. Furthermore, a statement is said to be satisfied for sufficiently small
(δj )j∈Ii if it is satisfied for all tuples (δj )j∈Ii whose maximum is smaller than a certain constant. This
constant only depends on g0 , h and k on Ui3 and on C on j∈Ii Uj3 .
S

On Ui3 we use the abbreviation


γ := f δ (ξ, s) and
X
γt := f (ξ, s)t = (1 − Ct2 ) · g0 (ξ) − 2t · h(ξ) +
δ
ψj · 2sχδj (t) · (h(ξ) − k(ξ))
j∈Ii

18
for t ∈ [0, ηi ). It holds
scalγ = scalγt + 3tr(Wt2 ) − tr(Wt )2 − trγt (γ̈t )
≥ scalγt − trγt (IIt )2 − trγt (γ̈t ) (8)
since Wt is a field of self-adjoint endomorphisms. Moreover,
1 X
IIt = − γ̇t = h(ξ) + Ct · g0 (ξ) − ψj · sχ̇δj (t) · (h(ξ) − k(ξ)),
2
j∈Ii
X
γ̈t = −2C · g0 (ξ) + ψj · 2sχ̈δj (t) · (h(ξ) − k(ξ)).
j∈Ii

Note that normal geodesics of f δ (ξ, s) and g(ξ) coincide (cf. Remark 3.8).
p We investigate each summand
in (8) on the set [0, ηi ) × Ui2 . One can even restrict to [0, maxj∈Ii δj ) × Ui2 since γ equals g(ξ) on
[maxj∈Ii δj , ηi ) × Ui2 .
p
p
Regarding scalγt : For every t ∈ [0, maxj∈Ii δj ), it holds
X
∥γt − γ0 ∥C 2 (Ui3 ) = −Ct2 · γ0 − 2t · h(ξ) + ψj · 2sχδj (t) · (h(ξ) − k(ξ)) 2 3
C (Ui )
j∈Ii
X
≤ t2 · ∥C · γ0 ∥C 2 (Ui3 ) + 2t · ∥h(ξ)∥C 2 (Ui3 ) + 2s χδj (t) · ∥ψj · (h(ξ) − k(ξ))∥C 2 (Ui3 )
j∈Ii
X
≤ 4t2 · Ci · ∥γ∥C 2 (Ui3 ) + 2t · ∥h(ξ)∥C 2 (Ui3 ) + 8s χδj (t) · ∥ψj ∥C 2 (Ui3 ) · ∥h(ξ) − k(ξ)∥C 2 (Ui3 ) .
j∈Ii

We have
t2 · Ci ≤ t · max
p
δj · Ci ≤ t.
j∈Ii

Together with the inequality 2sχδj (t) ≤ 2t from Remark 3.13, we find
X
∥γt − γ0 ∥C 2 (Ui3 ) ≤ 4t · ∥γ0 ∥C 2 (Ui3 ) + 2t · ∥h(ξ)∥C 2 (Ui3 ) + 8t ∥ψj ∥C 2 (Ui3 ) · ∥h(ξ) − k(ξ)∥C 2 (Ui3 ) .
j∈Ii
p
Hence ∥γt − γ0 ∥C 2 (Ui3 ) ≲ t + t + t = 3t ≤ 3 maxj∈Ii δj . This yields
scalγt ≳ −1 on Ui2 (9)
for sufficiently small (δj )j∈Ii .
1
p
Regarding trγt (IIt ): Let (δj )j∈Ii be sufficiently small so that ∥γt −γ0 ∥γ0 ≤ 2 for all (t, p) ∈ [0, maxj∈Ii δj )×
2
Ui . It holds
|trγt (IIt )| ≤ |trγ0 (IIt )| + |trγt (IIt ) − trγ0 (IIt )|
≲ |trγ0 (IIt )| + ∥γt − γ0 ∥γ0 · ∥IIt ∥γ0
≲ |trγ0 (IIt )| + ∥IIt ∥γ0 .
By the estimate |χ̇δj (t)| ≤ c0 (cf. Lemma 3.12), we obtain
X
∥IIt ∥γ0 = ∥h(ξ) + Ct · γ0 − ψj · sχ̇δj (t) · (h(ξ) − k(ξ))∥γ0
j∈Ii
√ X
≤ ∥h(ξ)∥γ0 + n − 1 · Ct + ψj · s|χ̇δj (t)| · ∥h(ξ) − k(ξ)∥γ0
j∈Ii
p X
≲ 1 + Ci · max δj + ψj
j∈Ii
j∈Ii
≤3

and |trγ0 (IIt )| ≤ n − 1 · ∥IIt ∥γ0 ≲ 1. Hence
|trγt (IIt )| ≲ |trγ0 (IIt )| + ∥IIt ∥γ0 ≲ 1. (10)
Regarding −trγt (γ̈t ): The trace is given by
X
−trγt (γ̈t ) = − ψj · 2sχ̈δj (t) · trγt (h(ξ) − k(ξ)) + 2C · trγt (γ0 ). (11)
j∈Ii
19
We consider the first sum: Let j0 ∈ Ii be fixed. Since trγ0 (h) ≥ trγ0 (k), χ̈δj (t) ≤ 0 and ∥γt − γ0 ∥γ0 ≲ t
for all (t, p) ∈ [0, δj0 ] × Ui2 , it holds
χ̈δj0 (t) · trγt (h(ξ) − k(ξ)) ≤ χ̈δj0 (t) · trγt (h(ξ) − k(ξ)) − χ̈δj0 (t) · trγ0 (h(ξ) − k(ξ))
≤ |χ̈δj0 (t)| · |trγt (h(ξ) − k(ξ)) − trγ0 (h(ξ) − k(ξ))|
≲ |χ̈δj0 (t)| · ∥γt − γ0 ∥γ0 · ∥h(ξ) − k(ξ)∥γ0
2
≲ ·t
δj0
≤2
for such t and p. This calculation is where the estimate χδj0 (t) ≤ t comes into play, contrary to
[4, Proposition p26].
For t ∈ [δj0 , δj0 ), we find
|χ̈δj0 (t) · trγt (h(ξ) − k(ξ))| ≲ |trγt (h(ξ) − k(ξ))|
≲ |trγ0 (h(ξ) − k(ξ))| + ∥γt − γ0 ∥γ0 · ∥h(ξ) − k(ξ)∥γ0
≲ 1 + 1 = 2.
p
Finally, χ̈δj0 (t) = 0 for t ≥ δj0 . In total, we have
χ̈δj0 (t) · trγt (h(ξ) − k(ξ)) ≲ 1
p
on [0, maxj∈Ii δj ) × Ui2 so that
X
ψj · 2sχ̈δj (t) · trγt (h(ξ) − k(ξ)) ≲ 1. (12)
j∈Ii

The second summand in (11) satisfies


|trγt (γ0 ) − n + 1| = |trγt (γ0 ) − trγ0 (γ0 )|
≲ ∥γt − γ0 ∥γ0 · ∥γ0 ∥γ0
p
≲ max δj .
j∈Ii

This means
1
|trγt (γ0 ) − n + 1| ≤ 2
for sufficiently small (δj )j∈Ii , respectively
2C · trγt (γ0 ) ≥ 2(n − 23 )C. (13)
Equations (11), (12) and (13) give
−trγt (γ̈t ) ≳ C (14)
if C is large enough on Ui2 . By (8), (9), (10) and (14), we obtain
scalγ ≳ C, (15)
provided that C is large enough on Ui2 .
Equations (14) and (15) make a condition on C on Ui2 . If we carry out the same calculation for every
set Uη2j , j ∈ N, this will add more conditions to C on Ui2 . In fact, only the finite number of sets Uη2j , j ∈ Ii
can make a contribution for Ui2 . This is why there exists a smooth function C00 ∈ C ∞ (∂M ) which is
sufficiently large on all pieces Uj2 , j ∈ N in the sense of the calculation above.
Given C ≥ C00 we impose finitely many conditions on (δj )j∈Ii , and in particular on δi , in order
to obtain equation (15). Repeating the procedure for all sets Uη2j , j ∈ N will add only finitely many
conditions on δi . This admits the following conclusion:
Let C ∈ C ∞ (∂M ; R), C ≥ C00 . There exists a sequence of positive constants (mi )i∈N =
mi (g0 |Ui3 , h|Ui3 , k|Ui3 ) such that for each collection δ = (δi )i∈N of sufficiently small numbers, it holds
scalγ ≥ mi · C on [0, max δj ) × Ui2 for all i ∈ N.
p
j∈Ii

Finally, X
2ψi · max m−1

C0 : ∂M → R , C0 = C00 + j · sup |σ|
j∈Ii [0,ηj ]×Uj2
i∈N
is a function as is required in the proposition.
20
It remains to arrange properties (f ) and (g): Given L > 1 we will be able to choose δ = (δi )i∈N small
enough so that
1
|v|g(ξ) ≤ |v|f (ξ,s) ≤ L|v|g(ξ)
L
for all v ∈ T M, ξ ∈ K and s ∈ [0, 1]. Let i ∈ N. Without great loss of generality, one can assume that
each closure cl(Uη2j ), j ∈ Ii is covered by a single chart. Then we add the following condition on δ:
0 < δi < min inf{|v|4g(ξ)(q) : ξ ∈ K, q ∈ Uη2j , v ∈ S n−1 }.
j∈Ii

For every q ∈ Uη2j , v ∈S n−1


, ξ ∈ K and s ∈ [0, 1], it holds
|v|2f (ξ,s) 1 X
=1+ ψj · 2sχδj (t) · (h(ξ) − k(ξ))(v T , v T )
|v|2g(ξ) |v|2g(ξ)
j∈Ii
T
where v is the tangential part of v with respect to dϕ. We estimate
|v|2f (ξ,s) 1 X
2 −1≥− 2 δj · |(h(ξ) − k(ξ))(v T , v T )|
|v|g(ξ) |v|g(ξ)
j∈Ii
Xp Xp
≥− δj · |(h(ξ) − k(ξ))(v T , v T )| ≳ − δj .
j∈Ii j∈Ii

This yields
|v|2f (ξ,s)  1 
−1≥− 1− 2
|v|2g(ξ) L
for sufficiently small (δj )j∈Ii . Hence |v|f (ξ,s) /|v|g(ξ) ≥ 1/L for such (δj )j∈Ii . Similarly,
|v|2f (ξ,s) 1 X
−1≤ δj · |(h(ξ) − k(ξ))(v T , v T )|
|v|2g(ξ) |v|2g(ξ)
j∈Ii
Xp
≲ δj .
j∈Ii

Choosing (δj )j∈Ii small enough, we obtain


|v|2f (ξ,s)
− 1 ≤ L2 − 1,
|v|2g(ξ)
respectively |v|f (ξ,s) /|v|g(ξ) ≤ L. □
The results from Proposition 3.7 and 3.16 are combined to the following theorem where we use two
auxiliary functions S1 , S2 : [0, 1] → [0, 1] defined by
0 ≤ t ≤ 21 1 − 2t, 0 ≤ t ≤ 21
 
1,
S1 (t) = 1 and S2 (t) = 1 .
2(1 − t), 2 ≤ t ≤ 1 0, 2 ≤t≤1

Theorem 3.17. Let K be a compact Hausdorff space and let


g : K → R>σ (M )
be continuous. Let k : K → C (∂M ; T ∂M ⊗ T ∗ ∂M ) be a continuous family of symmetric (0, 2)-tensor
∞ ∗
1
fields satisfying n−1 trg0 (k(ξ)) ≤ Hg(ξ) for all ξ ∈ K.
Then there exists a smooth positive function C0 ∈ C ∞ (∂M ) such that for each C ∈ C ∞ (∂M ) with
C ≥ C0 and for each neighbourhood U of ∂M there is a continuous map
f : K × [0, 1] → R>σ (M )
so that the following holds for all ξ ∈ K and s ∈ [0, 1]:
(a) f (ξ, 0) = g(ξ);
(b) f (ξ, 1) is C-normal;
(c) f (ξ, s)0 = g(ξ)0 ;
(d) IIf (ξ,s) = S1 (s)IIg(ξ) + (1 − S1 (s))k(ξ), in particular IIf (ξ,1) = k(ξ);
(e) if g(ξ) is C̃-normal then f (ξ, s) is Cs -normal for Cs = S2 (s)C̃ + (1 − S2 (s))C;
(f ) f¨(ξ, s)0 = S2 (s)g̈(ξ)0 − 2(1 − S2 (s))Cg(ξ)0 ;
(ℓ) (ℓ)
(g) for ℓ ≥ 3 we have f (ξ, s)0 = S2 (s) · g(ξ)0 ;
21
(h) f (ξ, s) = g(ξ) on M \ U ;
(i) f (ξ, s) is quasi-isometric to g(ξ) via the identity;
(j) if g(ξ) is complete then so is f (ξ, s).
Proof. The proof is the same as in [4, Theorem 27]. We apply Proposition 3.7 and Proposition 3.16 for
sufficiently large C, one after the other, and concatenate the arising homotopies. □

4. Applications
We present a couple of non-existence results for metrics with (uniformly) positive scalar curvature
and mean convex boundary.
Corollary 4.1. The right half-plane R2x1 ≥0 cannot carry any complete Riemannian metric of uniformly
positive scalar curvature which has mean convex boundary.
Proof. Set M := R2x1 ≥0 . Assume there is a Riemannian metric g as in the theorem. We apply Theo-
rem 3.17 to obtain a complete doubling metric f ∈ R(M ) of uniformly positive scalar curvature.
Set F := f ∪ f ∈ R(DM f ). This metric is complete by Theorem 2.2 and it has uniformly positive
scalar curvature. Since DM f is diffeomorphic to R2 , we obtain a complete metric of uniformly positive
scalar curvature on R2 . However, such a metric does not exist by the Bonnet-Myers theorem. □
In fact, there do exist metrics of positive scalar curvature and mean convex boundary on R2x1 ≥0 . An
example of this is the following Riemannian hypersurface with boundary, applied in the case n = 2:
 
1 2 2
Pn := xn+1 = , x + . . . + x < 1, x1 ≥ 0 ⊂ (Rn+1 , geucl ).
1 − x21 − . . . − x2n 1 n

Figure 4 shows P2 in R3 . For n ≥ 3, the level sets of Pn are half spheres of dimension two or greater,
indicating that Pn even has uniformly positive scalar curvature. This is why Corollary 4.1 cannot hold in
higher dimensions. However, if we fix a quasi-isometry type, we can find an analogous result for n ≥ 3.
It requires a preparatory lemma.

x3

P2

x2 x1
0 0,5
−0,5
−0,5

Figure 4. P2 in R3

Lemma 4.2. Let M be a smooth manifold with non-empty boundary. Let g1 and g2 be two complete
doubling metrics on M . Put G1 := g1 ∪ g1 and G2 := g2 ∪ g2 . If id : (M, dg1 ) → (M, dg2 ) is a quasi-
isometry, then so is id : (DM, dG1 ) → (DM, dG2 ).
Proof. We assume that M is connected and that id : (M, dg1 ) → (M, dg2 ) is a quasi-isometry. Then there
exist constants A ≥ 1 and B ≥ 0 such that
1
dg (p, q) − B ≤ dg2 (p, q) ≤ A dg1 (p, q) + B for all p, q ∈ M .
A 1
22
Let p, q ∈ DM . If p and q are contained in the same copy of M , then Proposition 2.3 yields the desired
quasi-isometry condition.
We focus on the case where p and q are contained in different copies. Since G2 is complete (see
Theorem 2.2), there exists a length-minimizing G2 -geodesic γ : [0, 1] → DM g2 with γ(0) = p and γ(1) = q.
By continuity there is s ∈ [0, 1] with γ(s) ∈ ∂M . The restricted curves γ|[0,s] and γ|[s,1] are again length-
minimizing G2 -geodesics. It holds
dG2 (p, q) = ℓG2 (γ) = ℓG2 (γ|[0,s] ) + ℓG2 (γ|[s,1] )
= dG2 (p, γ(s)) + dG2 (γ(s), q)
= dg2 (p, γ(s)) + dg2 (γ(s), q)
≥ ( A1 dg1 (p, γ(s)) − B) + ( A1 dg1 (γ(s), q) − B)
= ( A1 dG1 (p, γ(s)) − B) + ( A1 dG1 (γ(s), q) − B)
1
≥ A dG1 (p, q) − 2B.
The other inequality dG2 (p, q) ≤ A dG1 (p, q) + 2B can be shown in a similar manner. □

Corollary 4.3. Let n ≥ 3. The right half-space Rnx1 ≥0 cannot carry any complete Riemannian metric
of uniformly positive scalar curvature which has mean convex boundary and which is quasi-isometric to
the Euclidean metric via the identity.
Proof. Set M := Rnx1 ≥0 . Assume there is a Riemannian metric g as in the theorem. We apply Corol-
lary 3.5 and Theorem 3.17 for the metric g and the distinguished metric β := geucl . This yields a
Riemannian metric f ∈ R(M ) with the following properties:
▷ f has uniformly positive scalar curvature;
▷ f is doubling;
▷ ϕf = ϕβ on a neighbourhood of ∂M ⊂ [0, ∞) × ∂M ;
▷ f is quasi-isometric to geucl via the identity;
▷ f is complete.
From the third property, we deduce that f and β induce the same smooth structure on the double DM .
Both B = β ∪ β and F = f ∪ f are smooth metrics on DM . The metric F clearly has uniformly positive
scalar curvature. Furthermore, F is complete by Theorem 2.2 and the identity id : (DM, dB ) → (DM, dF )
is a quasi-isometry by Lemma 4.2.
Now we apply the canonical diffeomorphism

n [(x1 , x2 , . . . , xn ), 1] , x1 ≥ 0
κ : R → DM , κ(x) =
[(−x1 , x2 , . . . , xn ), 2] , x1 ≤ 0
to obtain the following commutative diagram of Riemannian manifolds and smooth maps:
id
(DM, B) (DM, F )

κ κ

id
(Rn , Geucl ) (Rn , κ∗ F )
This is where the unique smooth structure of DM comes into play because it guarantees smoothness for
the right vertical arrow. Both vertical arrows are Riemannian isometries.
The diagram above gives rise to a commutative diagram of metric spaces and continuous maps:
id
(DM, dB ) (DM, dF )

κ κ

id
(Rn , deucl ) (Rn , dκ∗ F )
The vertical arrows are metric isometries. As the upper horizontal arrow is a quasi-isometry, so is the
lower one. In total, κ∗ F violates the non-existence result in [14, Cor. 1.8]. □

The quasi-isometry condition can be dropped if we add some topology to the half-space.
23
Corollary 4.4. Let M := Rnx1 ≥0 #T n be a connected sum of the right half-space with an n-torus attached
in the interior. Then M does not admit a complete Riemannian metric of positive scalar curvature which
has mean convex boundary.

Proof. The strategy is as before. Since the double DM is diffeomorphic to a connected sum (T n #T n )\{pt}
of tori with a single point removed and since T n #T n is enlargeable, we can apply [10, Theorem C]. □

One can also drop the quasi-isometry condition by considering certain submanifolds with boundary of
the Whitehead manifold instead of half-spaces. The Whitehead manifold W is an open 3-manifold which
is contractible but not homeomorphic to R3 . Chang-Weinberger-Yu [11] have shown that W does not
admit any metric of uniformly positive scalar curvature. Subsequent work of Wang [18] even excludes
the existence of complete metrics of positive scalar curvature without uniform positive lower bound.
We look at two equivalent pictures for the Whitehead manifold: Consider the 3-sphere as a union
of two solid tori T and T0 . Then embed another solid torus T1 in T0 such that T1 forms a Whitehead
link with the meridian of T0 . Afterwards, we embed a solid torus T2 in T1 in the same way as T1
has been embedded in T0 and so on. The Whitehead manifold is defined as the open submanifold
W = S 3 − ∩∞ 3
k=1 Tk ⊂ S . It can also be seen as the union of T with infinitely many ”peeled” tori:
 
W = T ∪ T0 − T1 ∪ T1 − T2 ∪ . . .
This description motivates the second picture: Put A := T0 − int(T1 ). Then A is a manifold with two
boundary components where the outer component is denoted by ∂ 0 and the inner one is denoted by ∂ 1 .
Let φ : T0 → T1 be a homeomorphism twisting the torus T0 . We consider the sequence of spaces
W0 := A ∪∂ 1 T1 ,
W1 := A ∪∂ 1 φ(W0 ),
W2 := A ∪∂ 1 φ(W1 ),
..
.

together with the obvious inclusion maps. Furthermore, we construct a sequence of maps Wi → W as
follows: Let f : A → A be a self-homeomorphism with f |∂ 1 = φ−1 |∂T1 . Such a map exists since there
is an isotopy for the Whitehead link changing its components. For i = 0, we define f0 : W0 → W by
f0 |A = f and f0 |T1 = [T0 → T ] ◦ φ−1 where the second map swaps T0 and T . The next map f1 : W1 → W
is defined by f1 |A = φ ◦ f and f1 |φ(W0 ) = f0 ◦ φ−1 . In a nutshell, the maps fi send the innermost torus
of Wi to T and the peeled tori that are stacked to the outside of Wi to the peeled tori inside W .
As the collection {T, T0 − int(T1 ), T1 − int(T2 ), . . .} of closed subsets of W is locally finite, a map
W → Y to any topological space Y is continuous iff it is continuous on every subset in the collection.
Thus, W → Y is continuous iff Wi → W → Y is continuous for every i, showing that W is the
(topological) direct limit of the Wi . This is the second picture for the Whitehead manifold. Both
pictures give geometrical and topological insights: While one can immediately see the manifold structure
of W in the first picture, its contractability can be deduced from the second picture relying on the fact
that Wi → Wi+1 is nullhomotopic for every i and that the πn -functors commute with direct limits.
Now we remove the innermost torus of W , corresponding to the torus T in the first picture. As has
been shown by Newman-Whitehead [16, Theorem 6], the fundamental group π1 (W − T ) is not finitely
generated. This is sufficient to show that W is not homeomorphic to R3 : If there was a homeomorphism
Φ : W → R3 , then it would hold π1 (W − T ) ∼ = π1 (R3 − Φ(T )). However, π1 (R3 − Φ(T )) is finitely
generated (Wirtinger presentation).
Next, we pass to submanifolds with boundary of W . To this end, let R ≈ R3 be an open tube around
a properly embedded ray [0, ∞) in W − T . Then W − R is a contractible manifold with noncompact
boundary ∂(W − R) ≈ R2 . Indeed, we apply the Seifert-van Kampen theorem for the subsets W − R
and R of W to obtain
1 = π1 (W ) ∼
= π1 (W − R) ∗π1 (∂(W −R)) π1 (R) = π1 (W − R).
Thus, by the Whitehead and Hurewicz theorems, it suffices to show that W − R is homologically trivial.
The latter follows from a Mayer-Vietoris sequence for the same subsets as above.

Corollary 4.5. W − R does not admit any complete metric of uniformly positive scalar curvature which
has mean convex boundary.
24
Proof. Using the deformation principle and [11, Theorem 1], it suffices to show that the double D(W −R)
is contractible but not homeomorphic to R3 . The contractibility can be proven in a similar manner to
that of W − R. In the second part of the proof, we consider the manifold
X := D(W − R) − ι(T )
≈ (W − (T ∪ R)) ∪R (W − R)
where ι : W − R → D(W − R) is the canonical inclusion map for the first summand. According to
Seifert-van Kampen, the fundamental group of X is given by
π1 (X) ∼
= π1 (W − (T ∪ R)) ∗ π1 (W − R)
= π1 (W − (T ∪ R))

= π1 (W − T )
and hence, it is not finitely generated. If D(W − R) was homeomorphic to R3 , then π1 (X) would allow
a finite generating system, again using the Wirtinger presentation. □

References
[1] Almeida, S.: Minimal hypersurfaces of a positive scalar curvature manifold. Math. Z. 190 (1985), 73–82.
[2] Bär, C.; Gauduchon, P.; Moroianu, A.: Generalized cylinders in semi-Riemannian and Spin geometry. Math. Z. 249
(2005), 545–580.
[3] Bär, C.; Hanke, B.: Local flexibility for open partial differential relations. Comm. Pure Appl. Math. 75 (2022),
1377–1415.
[4] : Boundary conditions for scalar curvature. Perspectives in scalar curvature. Vol. 2., World Sci. Publ., Hacken-
sack, NJ, 325–377, 2023.
[5] Bridson, M. R.; Haefliger, A.: Metric Spaces of Non-Positive Curvature. Grundlehren der mathematischen Wis-
senschaften, vol. 319. Springer Berlin, Heidelberg, 1999.
[6] Burago, D.; Burago, Y.; Ivanov, S.: A course in metric geometry. Graduate Studies in Mathematics, vol. 33. American
Mathematical Society, Providence, RI, 2001.
[7] Burtscher, A. Y.: Length structures on manifolds with continuous Riemannian metrics. New York J. Math. 21 (2015),
273–296.
[8] Carlotto, A.; Li, Ch.: Constrained deformations of positive scalar curvature metrics (2021). Preprint available on
https://arxiv.org/abs/1903.11772, to appear in J. Diff. Geom.
[9] : Constrained deformations of positive scalar curvature metrics, II. Comm. Pure Appl. Math. 77 (2024), 795–
862.
[10] Cecchini, S.: A long neck principle for Riemannian spin manifolds with positive scalar curvature. Geom. Funct. Anal. 30
(2020), 1183–1223.
[11] Chang, S.; Weinberger, S.; Yu, G.: Taming 3-manifolds using scalar curvature. Geom. Dedicata 148 (2010), 3–14.
[12] Frerichs, H.: Skalarkrümmung auf Mannigfaltigkeiten mit nicht-kompaktem Rand (2022), available at https://
nbn-resolving.de/urn:nbn:de:bvb:384-opus4-1081965.
[13] Gromov, M.; Lawson, H. B.: Spin and scalar curvature in the presence of a fundamental group. I. Ann. of Math. (2)
111 (1980), 209–230.
[14] Hanke, B.; Kotschick, D.; Roe, J.; Schick, T.: Coarse topology, enlargeability, and essentialness. Ann. Sci. École
Norm. Sup. 41 (2008), 471–493.
[15] Hirsch, M. W.: Differential topology. Graduate Texts in Mathematics, vol. 33. Springer New York, NY, 1976.
[16] Newman, M. H. A.; Whitehead, J. H. C.: On the group of a certain linkage. Q. J. Math. 8 (1937), 14–21.
[17] Seeley, R. T.: Extension of C ∞ functions defined in a half space. Proc. Amer. Math. Soc. 15 (1964), 625–626.
[18] Wang, J.: Contractible 3-manifolds and Positive scalar curvature (I) (2023). Preprint available on https://arxiv.org/
abs/1901.04605, to appear in J. Diff. Geom.

Universität Augsburg, Institut für Mathematik, 86135 Augsburg, Germany


Email address: helge.frerichs@math.uni-augsburg.de

25

You might also like