2020 Mathematics Subject Classification. 53C12, 53C42.
The first named author was partially supported by CNPq-Brazil grants 303375/2025-6, 404611/2025-7 and 405618/2023-9. The second named author was partially supported by CNPq-Brazil grants 400043/2025-4 and 405618/2023-9. This work was supported by the Serrapilheira Institute (grant number Serra – R-2110-39576).
Introduction
Let M n + 1 M^{n+1} M n + 1 , n ≥ 1 n \ge1 n ≥ 1 , be a complete connected Riemannian manifold, and let F \mathcal{F} F be a smooth codimension one foliation of M M M by connected hypersurfaces. We call F \mathcal{F} F a CMC foliation if every leaf has constant mean curvature, the constant possibly depending on the leaf. Every codimension one foliation becomes transversely orientable after passing to a covering of M M M of degree at most two; the lifted foliation is again a CMC foliation whose leaves cover those of F \mathcal{F} F and have the same mean curvature up to sign, and completeness and lower sectional curvature bounds are inherited by Riemannian coverings. Since only ∣ H ∣ |H| ∣ H ∣ will be estimated, we assume throughout that F \mathcal{F} F is transversely oriented. Let N N N be a smooth unit vector field normal to the leaves and define H ∈ C ∞ ( M ) H \in C^\infty(M) H ∈ C ∞ ( M ) by
div N = − n H . (1) \operatorname{div} N=-nH. \tag*{(1)} div N = − n H . (1)
With the shape operator A : = − ∇ N A:=-\nabla N A := − ∇ N , the restriction of A A A to the tangent space of a leaf has trace n H nH n H (see § 2 ), so H H H is the mean curvature of the leaves with respect to N N N .
In 2008, Meeks III, Pérez, and Ros [12 Remark 5.21] proposed the following conjecture.
Conjecture 1.1. For each integer n ≥ 1 n \ge1 n ≥ 1 there is A = A ( n + 1 ) > 0 A=A(n+1)>0 A = A ( n + 1 ) > 0 with the following property: for every complete Riemannian manifold M n + 1 M^{n+1} M n + 1 with absolute sectional curvature bounded from above by 1 and every codimension one CMC foliation F \mathcal{F} F of M M M , the mean curvature H L H_L H L of each leaf L L L of F \mathcal{F} F satisfies ∣ H L ∣ ≤ A |H_L| \le A ∣ H L ∣ ≤ A .
They also conjectured that one can take A ( n + 1 ) = 1 A(n+1)=1 A ( n + 1 ) = 1 [12 Conjecture 5.1(2)] :
Conjecture 1.2. Let F \mathcal{F} F be a codimension one CMC foliation of a complete Riemannian manifold M n + 1 M^{n+1} M n + 1 with absolute sectional curvature bounded from above by 1. Then ∣ H L ∣ ≤ 1 |H_L| \le1 ∣ H L ∣ ≤ 1 for every leaf L L L of F \mathcal{F} F .
Meeks III, Pérez, and Ros proved Conjecture 1.2 when M M M is the hyperbolic space H 3 \mathbb{H}^3 H 3 [12 Corollary 5.10] , and they provided in [12 Theorem 5.23] partial positive answers to Conjecture 1.1, they proved that when M M M is a homogeneously regular manifold with absolute sectional curvature at most 1 and n = 3 n=3 n = 3 or 4, the mean curvature of the leaves of a CMC foliation of M M M is bounded by a constant depending only on n n n . When dim M = 3 \dim M=3 dim M = 3 , a bound for ∣ H ∣ |H| ∣ H ∣ depending only on an upper bound for ∣ sec M ∣ |\sec_M| ∣ sec M ∣ can also be derived from the universal curvature estimate for CMC foliations of compact Riemannian 3-manifolds with boundary [13 Theorem 1.3] : applied to a compact domain of M M M containing a ball of radius π \pi π about a point p p p , it bounds the norm of the second fundamental form of the leaves at p p p , hence their mean curvature, by a universal constant. When M M M is compact, the sharp bound ∣ H ∣ ≤ K 0 |H| \le\sqrt{K_0} ∣ H ∣ ≤ K 0 is known even under the weaker assumption Ric M ≥ − n K 0 \operatorname{Ric}_{M} \ge-nK_{0} Ric M ≥ − n K 0 ; see [1 Theorem 3.1] for K 0 = 0 K_{0}=0 K 0 = 0 , [12 Theorem 5.8.C] , [11 Remark 1.4(i)] , and [17 Theorem 4.4] . Recently, the authors proved in [18 ] that if sec M ≡ 0 \operatorname{sec}_{M} \equiv0 sec M ≡ 0 , then H ≡ 0 H \equiv0 H ≡ 0 .
The aim of this article is to provide a complete positive answer to Conjecture 1.1. Indeed, we obtain the following more general result.
Theorem A. Let M n + 1 M^{n+1} M n + 1 be a complete Riemannian manifold with sec M ≥ − K 0 \operatorname{sec}_{M} \ge-K_{0} sec M ≥ − K 0 for some K 0 ≥ 0 K_{0} \ge0 K 0 ≥ 0 , and let F \mathcal{F} F be a codimension one CMC foliation of M M M . Then
∣ H ∣ ≤ ( n + 1 ) K 0 on M . \lvert H\rvert\le(n+1)\sqrt{K_{0}} \qquad\text{on } M. ∣ H ∣ ≤ ( n + 1 ) K 0 on M . In particular, if sec M ≥ 0 \operatorname{sec}_{M} \ge0 sec M ≥ 0 , every leaf of F \mathcal{F} F is minimal.
Corollary 1.3. Conjecture 1.1 holds with A ( n + 1 ) = n + 1 A(n+1)=n+1 A ( n + 1 ) = n + 1 . More generally, if ∣ sec M ∣ ≤ K \lvert\operatorname{sec}_{M}\rvert\le K ∣ sec M ∣ ≤ K , then ∣ H ∣ ≤ ( n + 1 ) K \lvert H\rvert\le(n+1)\sqrt{K} ∣ H ∣ ≤ ( n + 1 ) K .
Proof. ∣ sec M ∣ ≤ K \lvert\operatorname{sec}_{M}\rvert\le K ∣ sec M ∣ ≤ K implies sec M ≥ − K \operatorname{sec}_{M} \ge-K sec M ≥ − K ; apply Theorem A with K 0 = K K_{0}=K K 0 = K . □ \square □
Strategy and organization
Fix x 0 ∈ M x_{0} \in M x 0 ∈ M and let r : = dist M ( x 0 , ⋅ ) r := \operatorname{dist}_{M}(x_{0},\mathord\cdot) r := dist M ( x 0 , ⋅ ) . For c 0 > K 0 c_{0} > \sqrt{K_{0}} c 0 > K 0 , we consider the vector field V = G s N + 2 s n H G s − 1 ∇ N N V = G^{s}N + \frac{2s}{n}HG^{s-1}\nabla_{N}N V = G s N + n 2 s H G s − 1 ∇ N N , G = ( H + 2 − c 0 2 ) + G=(H_{+}^{2}-c_{0}^{2})_{+} G = ( H + 2 − c 0 2 ) + , whose divergence controls a positive quantity Y Y Y supported in { H > c 0 } \{H>c_{0}\} { H > c 0 } (Lemma 2.14 ). Testing div V ≥ Y \operatorname{div}V \ge Y div V ≥ Y against powers of cutoff functions of r r r and iterating the resulting flow inequality shows that if { H > c 0 } ≠ ∅ \{H>c_{0}\}\ne\varnothing { H > c 0 } = ∅ , the integral of Y Y Y over { r ≤ R } \{r\le R\} { r ≤ R } must grow exponentially in R R R at a rate that, for c 0 > c ∗ ( n , K 0 ; C 1 ) c_{0}>c_{*}(n,K_{0};C_{1}) c 0 > c ∗ ( n , K 0 ; C 1 ) , exceeds the rate n K 0 n\sqrt{K_{0}} n K 0 allowed by the Bishop–Gromov inequality. Here C 1 C_{1} C 1 is an upper bound for Hess r \operatorname{Hess}r Hess r far from x 0 x_{0} x 0 : the uncontrolled field ∇ N N \nabla_{N}N ∇ N N is traded for second derivatives of the cutoff functions (Lemma 2.4), and only an upper bound for Hess r \operatorname{Hess}r Hess r is needed. Since r r r is not smooth, this bound is used in the support sense of Calabi; Proposition 2.12 shows that it suffices for the integration by parts. By the Hessian comparison theorem, one can take C 1 = K 0 coth ( K 0 R ∗ ) C_{1}=\sqrt{K_{0}}\coth(\sqrt{K_{0}}R_{*}) C 1 = K 0 coth ( K 0 R ∗ ) outside B R ∗ ( x 0 ) B_{R_{*}}(x_{0}) B R ∗ ( x 0 ) , and letting R ∗ → ∞ R_{*}\to\infty R ∗ → ∞ gives the constant c ∗ ( n , K 0 ; K 0 ) = ( n + 1 ) K 0 c_{*}(n,K_{0};\sqrt{K_{0}})=(n+1)\sqrt{K_{0}} c ∗ ( n , K 0 ; K 0 ) = ( n + 1 ) K 0 .
Section 2 collects the identities relating the foliation to the ambient divergence, the Hessian comparison for the distance function in the support sense, the regularization argument, and the flow field. Section 3 proves Theorem A.
Preliminaries
Notation and identities
We write ∇ \nabla ∇ , g ( ⋅ , ⋅ ) : = ⟨ ⋅ , ⋅ ⟩ g(\mathord\cdot,\mathord\cdot):=\langle\mathord\cdot,\mathord\cdot\rangle g ( ⋅ , ⋅ ) := ⟨ ⋅ , ⋅ ⟩ , div \operatorname{div} div , R R R for the Levi-Civita connection, metric, divergence, and curvature tensor of M M M , with R ( X , Y ) Z = ∇ X ∇ Y Z − ∇ Y ∇ X Z − ∇ [ X , Y ] Z R(X,Y)Z=\nabla_{X}\nabla_{Y}Z-\nabla_{Y}\nabla_{X}Z-\nabla_{[X,Y]}Z R ( X , Y ) Z = ∇ X ∇ Y Z − ∇ Y ∇ X Z − ∇ [ X , Y ] Z , so that sec ( X , Y ) = ⟨ R ( X , Y ) Y , X ⟩ \operatorname{sec}(X,Y)=\langle R(X,Y)Y,X\rangle sec ( X , Y ) = ⟨ R ( X , Y ) Y , X ⟩ for orthonormal X , Y X,Y X , Y . Extend the shape operator to T M TM T M by A : = − ∇ N A:=-\nabla N A := − ∇ N . Since ⟨ ∇ X N , N ⟩ = 1 2 X ⟨ N , N ⟩ = 0 \langle\nabla_{X}N,N\rangle=\frac{1}{2}X\langle N,N\rangle=0 ⟨ ∇ X N , N ⟩ = 2 1 X ⟨ N , N ⟩ = 0 , A A A takes values in N ⊥ N^{\perp} N ⊥ . For X , Y X,Y X , Y tangent to a leaf L L L , ⟨ A X , Y ⟩ = ⟨ N , ∇ X Y ⟩ \langle AX,Y\rangle=\langle N,\nabla_{X}Y\rangle ⟨ A X , Y ⟩ = ⟨ N , ∇ X Y ⟩ , which is symmetric in X , Y X,Y X , Y because [ X , Y ] [X,Y] [ X , Y ] is tangent to L L L ; hence A ∣ T L A|_{TL} A ∣ T L is a symmetric endomorphism of T L TL T L and, computing div N \operatorname{div} N div N in an orthonormal frame { e 1 , … , e n , N } \{e_1,\ldots,e_n,N\} { e 1 , … , e n , N } ,
tr A ∣ T L = − ∑ i ⟨ ∇ e i N , e i ⟩ = − div N + ⟨ ∇ N N , N ⟩ = n H . \operatorname{tr} A|_{TL}=-\sum_i\langle\nabla_{e_i}N,e_i\rangle=-\operatorname{div}N+\langle\nabla_NN,N\rangle=nH. tr A ∣ T L = − i ∑ ⟨ ∇ e i N , e i ⟩ = − div N + ⟨ ∇ N N , N ⟩ = n H .
Put
g : = N ( H ) , T : = ∇ N N , ∣ A ∣ 2 : = tr ( ( A ∣ T L ) 2 ) , \mathbf{g}:=N(H),\qquad T:=\nabla_NN,\qquad|A|^2:=\operatorname{tr}\left((A|_{TL})^2\right), g := N ( H ) , T := ∇ N N , ∣ A ∣ 2 := tr ( ( A ∣ T L ) 2 ) ,
and let ∇ L \nabla_L ∇ L , ∇ L \nabla^L ∇ L , div L \operatorname{div}_L div L , and Δ L \Delta_L Δ L denote the gradient, covariant derivative, divergence, and Laplacian on a leaf, with respect to its induced metric, acting on the restrictions of the corresponding objects to the leaves. Then T T T is tangent to the leaves, A N = − T AN=-T A N = − T , and if k 1 , … , k n k_1,\ldots,k_n k 1 , … , k n are the eigenvalues of A ∣ T L A|_{TL} A ∣ T L ,
∣ A ∣ 2 = ∑ i k i 2 ≥ 1 n ( ∑ i k i ) 2 = n H 2 . (2) |A|^2=\sum_i k_i^2\geq\frac{1}{n}\left(\sum_i k_i\right)^2=nH^2. \tag*{(2)} ∣ A ∣ 2 = i ∑ k i 2 ≥ n 1 ( i ∑ k i ) 2 = n H 2 . (2)
Lemma 2.1. If κ ∈ C 1 ( M ) \kappa\in C^1(M) κ ∈ C 1 ( M ) is constant along the leaves, then ∇ κ = N ( κ ) N \nabla\kappa=N(\kappa)N ∇ κ = N ( κ ) N . In particular ∇ H = g N \nabla H=gN ∇ H = g N .
Proof. X ( κ ) = 0 X(\kappa)=0 X ( κ ) = 0 for every X X X tangent to the leaves, so ∇ κ \nabla\kappa ∇ κ is parallel to N N N . □
The following lemma was proved by Barbosa, Kenmotsu, and Oshikiri in [1 ]
Lemma 2.2 (Proposition 2.14 in [1 ] ). For a C 1 C^1 C 1 vector field X X X tangent to the leaves, div X = div L X − ⟨ X , T ⟩ \operatorname{div}X=\operatorname{div}_L X-\langle X,T\rangle div X = div L X − ⟨ X , T ⟩ and div T = ∣ A ∣ 2 + Ric ( N , N ) − n g \operatorname{div}T=|A|^2+\operatorname{Ric}(N,N)-ng div T = ∣ A ∣ 2 + Ric ( N , N ) − n g .
The following lemma was proved by Sampaio and Silva (see [18 ] ) using similar ideas. For the sake of completeness, we provide a proof.
Lemma 2.3 (Lemma 2.5 in [18 ] ). For f ∈ C 2 ( M ) f\in C^2(M) f ∈ C 2 ( M ) ,
Δ L f = tr T L Hess f + n H N ( f ) , tr T L Hess f : = ∑ i = 1 n Hess f ( e i , e i ) , \Delta_Lf=\operatorname{tr}_{TL}\operatorname{Hess}f+nHN(f),\qquad\operatorname{tr}_{TL}\operatorname{Hess}f:=\sum_{i=1}^{n}\operatorname{Hess}f(e_i,e_i), Δ L f = tr T L Hess f + n H N ( f ) , tr T L Hess f := i = 1 ∑ n Hess f ( e i , e i ) , where { e i } \{e_i\} { e i } is an orthonormal basis of the tangent space of the leaf.
Proof. In a frame { e 1 , … , e n , N } \{e_1,\ldots,e_n,N\} { e 1 , … , e n , N } as above, the e i e_i e i pertain to local vector fields tangent to the leaves. Then Δ L f = ∑ i ( e i e i f − ( ∇ e i L e i ) f ) \Delta_Lf=\sum_i(e_ie_if-(\nabla^L_{e_i}e_i)f) Δ L f = ∑ i ( e i e i f − ( ∇ e i L e i ) f ) , Hess f ( e i , e i ) = e i e i f − ( ∇ e i e i ) f \operatorname{Hess}f(e_i,e_i)=e_ie_if-(\nabla_{e_i}e_i)f Hess f ( e i , e i ) = e i e i f − ( ∇ e i e i ) f , and ∇ e i e i = ∇ e i L e i + ⟨ ∇ e i e i , N ⟩ N = ∇ e i L e i + ⟨ A e i , e i ⟩ N \nabla_{e_i}e_i=\nabla^L_{e_i}e_i+\langle\nabla_{e_i}e_i,N\rangle N=\nabla^L_{e_i}e_i+\langle Ae_i,e_i\rangle N ∇ e i e i = ∇ e i L e i + ⟨ ∇ e i e i , N ⟩ N = ∇ e i L e i + ⟨ A e i , e i ⟩ N . Summing over i i i gives tr T L Hess f = Δ L f − n H N ( f ) \operatorname{tr}_{TL}\operatorname{Hess}f=\Delta_Lf-nHN(f) tr T L Hess f = Δ L f − n H N ( f ) . □
The following lemma was proved in [18 Lemma 3.2] when M = R n + 1 M=\mathbb{R}^{n+1} M = R n + 1 . The proof works along the same lines, but since it is short, for completeness, we present it here.
Lemma 2.4 (trading T T T for Hessians). For χ ∈ C c 2 ( M ) \chi\in C_c^2(M) χ ∈ C c 2 ( M ) and κ ∈ C 1 ( M ) \kappa\in C^1(M) κ ∈ C 1 ( M ) constant along the leaves,
∫ M κ ⟨ ∇ χ , T ⟩ = ∫ M κ tr T L Hess χ + n ∫ M H κ N ( χ ) . (3) \int_M\kappa\langle\nabla\chi,T\rangle=\int_M\kappa\operatorname{tr}_{TL}\operatorname{Hess}\chi+n\int_MH\kappa N(\chi). \tag*{(3)} ∫ M κ ⟨ ∇ χ , T ⟩ = ∫ M κ tr T L Hess χ + n ∫ M H κ N ( χ ) . (3)
Proof. The field ∇ L χ : = ∇ χ − N ( χ ) N \nabla_L\chi:= \nabla\chi-N(\chi)N ∇ L χ := ∇ χ − N ( χ ) N is C 1 C^1 C 1 , compactly supported and tangent to the leaves, and its restriction to each leaf is the gradient of χ ∣ L \chi|_L χ ∣ L , so div L ( ∇ L χ ) = Δ L χ \operatorname{div}_L(\nabla_L\chi)=\Delta_L\chi div L ( ∇ L χ ) = Δ L χ . Since T T T is tangent to the leaves, ⟨ ∇ L χ , T ⟩ = ⟨ ∇ χ , T ⟩ \langle\nabla_L\chi,T\rangle=\langle\nabla\chi,T\rangle ⟨ ∇ L χ , T ⟩ = ⟨ ∇ χ , T ⟩ . By Lemmas 2.2 and 2.3,
⟨ ∇ χ , T ⟩ = Δ L χ − div ( ∇ L χ ) = tr T Hess χ + n H N ( χ ) − div ( ∇ L χ ) . \langle\nabla\chi,T\rangle=\Delta_L\chi-\operatorname{div}(\nabla_L\chi)=\operatorname{tr}_T\operatorname{Hess}\chi+nH\,N(\chi)-\operatorname{div}(\nabla_L\chi). ⟨ ∇ χ , T ⟩ = Δ L χ − div ( ∇ L χ ) = tr T Hess χ + n H N ( χ ) − div ( ∇ L χ ) .
Multiply by κ \kappa κ . By Lemma 2.1, ⟨ ∇ κ , ∇ L χ ⟩ = 0 \langle\nabla\kappa,\nabla_L\chi\rangle=0 ⟨ ∇ κ , ∇ L χ ⟩ = 0 , so κ div ( ∇ L χ ) = div ( κ ∇ L χ ) \kappa\operatorname{div}(\nabla_L\chi)=\operatorname{div}(\kappa\nabla_L\chi) κ div ( ∇ L χ ) = div ( κ ∇ L χ ) , and the divergence of the compactly supported C 1 C^1 C 1 field κ ∇ L χ \kappa\nabla_L\chi κ ∇ L χ has integral zero. □
The distance function and the Hessian comparison in the support sense
The following notion goes back to Calabi [2 ] ; see also [19 ] .
Definition 2.5. Let W ⊂ M W\subset M W ⊂ M be open, f ∈ C 0 ( W ) f\in C^0(W) f ∈ C 0 ( W ) , x ∈ W x\in W x ∈ W and a ∈ R a\in\mathbb{R} a ∈ R . We say that Hess f ( x ) ≤ a ⟨ ⋅ , ⋅ ⟩ \operatorname{Hess}f(x)\le a\langle\cdot,\cdot\rangle Hess f ( x ) ≤ a ⟨ ⋅ , ⋅ ⟩ in the support sense if for every ε > 0 \varepsilon>0 ε > 0 there are a neighbourhood U ⊂ W U\subset W U ⊂ W of x x x and φ ∈ C 2 ( U ) \varphi\in C^2(U) φ ∈ C 2 ( U ) (an upper support function of f f f at x x x ) such that φ ≥ f \varphi\ge f φ ≥ f on U U U , φ ( x ) = f ( x ) \varphi(x)=f(x) φ ( x ) = f ( x ) ,
Hess φ ( x ) ≤ ( a + ε ) ⟨ ⋅ , ⋅ ⟩ . \operatorname{Hess}\varphi(x)\le(a+\varepsilon)\langle\cdot,\cdot\rangle. Hess φ ( x ) ≤ ( a + ε ) ⟨ ⋅ , ⋅ ⟩ . If f f f is C 2 C^2 C 2 near x x x , this is equivalent to Hess f ( x ) ≤ a ⟨ ⋅ , ⋅ ⟩ \operatorname{Hess}f(x)\le a\langle\cdot,\cdot\rangle Hess f ( x ) ≤ a ⟨ ⋅ , ⋅ ⟩ : f f f itself is an upper support function, and conversely φ − f \varphi-f φ − f has a local minimum at x x x , so Hess f ( x ) ≤ Hess φ ( x ) \operatorname{Hess}f(x)\le\operatorname{Hess}\varphi(x) Hess f ( x ) ≤ Hess φ ( x ) .
For K 0 ≥ 0 K_0\ge0 K 0 ≥ 0 and t > 0 t>0 t > 0 , let ct K 0 ( t ) \operatorname{ct}_{K_0}(t) ct K 0 ( t ) be the function defined by:
ct K 0 ( t ) = { K 0 coth ( K 0 t ) , K 0 > 0 , 1 t , K 0 = 0. \operatorname{ct}_{K_0}(t)=
\begin{cases}
\sqrt{K_0}\coth(\sqrt{K_0}t), & K_0>0,\\
\frac{1}{t}, & K_0=0.
\end{cases} ct K 0 ( t ) = { K 0 coth ( K 0 t ) , t 1 , K 0 > 0 , K 0 = 0.
The function ct K 0 \operatorname{ct}_{K_0} ct K 0 is continuous and decreasing on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) . Moreover,
lim t → ∞ ct K 0 ( t ) = { K 0 , if K 0 > 0 , 0 , if K 0 = 0. \lim_{t\to\infty}\operatorname{ct}_{K_0}(t)=
\begin{cases}
\sqrt{K_0}, & \text{if }K_0>0,\\
0, & \text{if }K_0=0.
\end{cases} t → ∞ lim ct K 0 ( t ) = { K 0 , 0 , if K 0 > 0 , if K 0 = 0.
We denote by Cut ( p ) \operatorname{Cut}(p) Cut ( p ) the cut locus of p ∈ M p\in M p ∈ M and define the distance function r p : M → [ 0 , ∞ ) r_p:M\to[0,\infty) r p : M → [ 0 , ∞ ) by
r p ( x ) : = dist M ( p , x ) . r_p(x):=\operatorname{dist}_M(p,x). r p ( x ) := dist M ( p , x ) .
Theorem 2.6 (Hessian comparison; see [14 ] , [10 ] , [5 ] ). Let M M M be complete with sec M ≥ − K 0 \operatorname{sec}_M\ge-K_0 sec M ≥ − K 0 , p ∈ M p\in M p ∈ M and y ∈ M ∖ ( Cut ( p ) ∪ { p } ) y\in M\setminus(\operatorname{Cut}(p)\cup\{p\}) y ∈ M ∖ ( Cut ( p ) ∪ { p }) . Then r p r_p r p is smooth near y y y and
Hess r p ( y ) ≤ ct K 0 ( r p ( y ) ) ( ⟨ ⋅ , ⋅ ⟩ − d r p ⊗ d r p ) ≤ ct K 0 ( r p ( y ) ) ⟨ ⋅ , ⋅ ⟩ . \operatorname{Hess}r_p(y)\le\operatorname{ct}_{K_0}(r_p(y))\left(\langle\cdot,\cdot\rangle-dr_p\otimes dr_p\right)\le\operatorname{ct}_{K_0}(r_p(y))\langle\cdot,\cdot\rangle. Hess r p ( y ) ≤ ct K 0 ( r p ( y )) ( ⟨ ⋅ , ⋅ ⟩ − d r p ⊗ d r p ) ≤ ct K 0 ( r p ( y )) ⟨ ⋅ , ⋅ ⟩ .
We also use two standard facts about cut loci (see [4 Ch. XIII, §2] , [3 Ch. 5] ):
(C1) if γ : [ 0 , ℓ ] → M \gamma:[0,\ell]\to M γ : [ 0 , ℓ ] → M is a minimizing geodesic from p = γ ( 0 ) p=\gamma(0) p = γ ( 0 ) , then γ ( t ) ∉ Cut ( p ) \gamma(t)\notin\operatorname{Cut}(p) γ ( t ) ∈ / Cut ( p ) for 0 < t < ℓ 0<t<\ell 0 < t < ℓ ;
(C2) q ∈ Cut ( p ) q\in\operatorname{Cut}(p) q ∈ Cut ( p ) if and only if p ∈ Cut ( q ) p\in\operatorname{Cut}(q) p ∈ Cut ( q ) .
The following lemma follows from Calabi’s trick in [2 ] , see also [19 ] and [5 ] . For the sake of completeness, we provide a proof.
Lemma 2.7. Let M M M be complete with sec M ≥ − K 0 \operatorname{sec}_{M} \ge-K_{0} sec M ≥ − K 0 , x 0 ∈ M x_{0} \in M x 0 ∈ M , and r : = dist M ( x 0 , ⋅ ) r := \operatorname{dist}_{M}(x_{0},\cdot) r := dist M ( x 0 , ⋅ ) . Then for every x ∈ M ∖ { x 0 } x \in M \setminus\{x_{0}\} x ∈ M ∖ { x 0 } ,
Hess r ( x ) ≤ ct K 0 ( r ( x ) ) ⟨ ⋅ , ⋅ ⟩ \operatorname{Hess} r(x) \le\operatorname{ct}_{K_{0}}(r(x))\langle\cdot,\cdot\rangle Hess r ( x ) ≤ ct K 0 ( r ( x )) ⟨ ⋅ , ⋅ ⟩ in the support sense. In particular, for every R ∗ > 0 R_{*} > 0 R ∗ > 0 , Hess r ≤ ct K 0 ( R ∗ ) ⟨ ⋅ , ⋅ ⟩ \operatorname{Hess} r \le\operatorname{ct}_{K_{0}}(R_{*})\langle\cdot,\cdot\rangle Hess r ≤ ct K 0 ( R ∗ ) ⟨ ⋅ , ⋅ ⟩ in the support sense at every point of { r > R ∗ } \{r > R_{*}\} { r > R ∗ } .
Proof. Fix x ∈ M ∖ { x 0 } x \in M \setminus\{x_{0}\} x ∈ M ∖ { x 0 } and let r ( x ) = dist M ( x 0 , x ) > 0 r(x) = \operatorname{dist}_{M}(x_{0},x) > 0 r ( x ) = dist M ( x 0 , x ) > 0 . By completeness and the Hopf-Rinow theorem, there exists a unit-speed minimizing geodesic γ : [ 0 , r ( x ) ] → M \gamma: [0,r(x)] \to M γ : [ 0 , r ( x )] → M such that γ ( 0 ) = x 0 \gamma(0) = x_{0} γ ( 0 ) = x 0 and γ ( r ( x ) ) = x \gamma(r(x)) = x γ ( r ( x )) = x . Fix ε ∈ ( 0 , r ( x ) ) \varepsilon\in(0,r(x)) ε ∈ ( 0 , r ( x )) and set
p ε : = γ ( ε ) , f ε : = ε + r p ε , p_{\varepsilon} := \gamma(\varepsilon), \qquad f_{\varepsilon} := \varepsilon+ r_{p_{\varepsilon}}, p ε := γ ( ε ) , f ε := ε + r p ε ,
where r p ε ( y ) : = dist M ( p ε , y ) r_{p_{\varepsilon}}(y) := \operatorname{dist}_{M}(p_{\varepsilon},y) r p ε ( y ) := dist M ( p ε , y ) . By the triangle inequality, for every y ∈ M y \in M y ∈ M ,
r ( y ) = dist M ( x 0 , y ) ≤ dist M ( x 0 , p ε ) + dist M ( p ε , y ) = f ε ( y ) , r(y) = \operatorname{dist}_{M}(x_{0},y) \le\operatorname{dist}_{M}(x_{0},p_{\varepsilon}) + \operatorname{dist}_{M}(p_{\varepsilon},y) = f_{\varepsilon}(y), r ( y ) = dist M ( x 0 , y ) ≤ dist M ( x 0 , p ε ) + dist M ( p ε , y ) = f ε ( y ) ,
since dist M ( x 0 , p ε ) = ε \operatorname{dist}_{M}(x_{0},p_{\varepsilon}) = \varepsilon dist M ( x 0 , p ε ) = ε . Moreover, as γ \gamma γ is minimizing, its restriction to [ ε , r ( x ) ] [\varepsilon,r(x)] [ ε , r ( x )] is also minimizing. Hence
r p ε ( x ) = r ( x ) − ε , r_{p_{\varepsilon}}(x) = r(x) - \varepsilon, r p ε ( x ) = r ( x ) − ε ,
and consequently
f ε ( x ) = ε + r p ε ( x ) = r ( x ) . f_{\varepsilon}(x) = \varepsilon+ r_{p_{\varepsilon}}(x) = r(x). f ε ( x ) = ε + r p ε ( x ) = r ( x ) .
Thus f ε ≥ r f_{\varepsilon} \ge r f ε ≥ r on M M M and f ε ( x ) = r ( x ) f_{\varepsilon}(x) = r(x) f ε ( x ) = r ( x ) . We next show that x ∉ Cut ( p ε ) x \notin\operatorname{Cut}(p_{\varepsilon}) x ∈ / Cut ( p ε ) . Consider the reversed geodesic
γ ~ ( t ) : = γ ( r ( x ) − t ) , t ∈ [ 0 , r ( x ) ] . \widetilde{\gamma}(t) := \gamma(r(x)-t), \qquad t \in[0,r(x)]. γ ( t ) := γ ( r ( x ) − t ) , t ∈ [ 0 , r ( x )] .
It is a minimizing geodesic from x x x to x 0 x_{0} x 0 , and p ε = γ ~ ( r ( x ) − ε ) p_{\varepsilon} = \widetilde{\gamma}(r(x)-\varepsilon) p ε = γ ( r ( x ) − ε ) is an interior point of this geodesic. By property (C1), p ε ∉ Cut ( x ) p_{\varepsilon} \notin\operatorname{Cut}(x) p ε ∈ / Cut ( x ) . The symmetry of the cut locus, property (C2), then yields x ∉ Cut ( p ε ) x \notin\operatorname{Cut}(p_{\varepsilon}) x ∈ / Cut ( p ε ) . Since x ≠ p ε x \ne p_{\varepsilon} x = p ε , the distance function r p ε r_{p_{\varepsilon}} r p ε is smooth in a neighborhood of x x x . Therefore, f ε f_{\varepsilon} f ε is smooth near x x x , and
Hess f ε ( x ) = Hess r p ε ( x ) . \operatorname{Hess} f_{\varepsilon}(x) = \operatorname{Hess} r_{p_{\varepsilon}}(x). Hess f ε ( x ) = Hess r p ε ( x ) .
Applying the Hessian comparison theorem at x x x , with base point p ε p_{\varepsilon} p ε , gives
Hess f ε ( x ) ≤ ct K 0 ( r p ε ( x ) ) ⟨ ⋅ , ⋅ ⟩ x = ct K 0 ( r ( x ) − ε ) ⟨ ⋅ , ⋅ ⟩ x . \operatorname{Hess} f_{\varepsilon}(x) \le\operatorname{ct}_{K_{0}}(r_{p_{\varepsilon}}(x))\langle\cdot,\cdot\rangle_{x} = \operatorname{ct}_{K_{0}}(r(x)-\varepsilon)\langle\cdot,\cdot\rangle_{x}. Hess f ε ( x ) ≤ ct K 0 ( r p ε ( x )) ⟨ ⋅ , ⋅ ⟩ x = ct K 0 ( r ( x ) − ε ) ⟨ ⋅ , ⋅ ⟩ x .
Now let δ > 0 \delta> 0 δ > 0 be arbitrary. Since r ( x ) > 0 r(x) > 0 r ( x ) > 0 and ct K 0 \operatorname{ct}_{K_{0}} ct K 0 is continuous on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) , we may choose ε ∈ ( 0 , r ( x ) ) \varepsilon\in(0,r(x)) ε ∈ ( 0 , r ( x )) sufficiently small so that
ct K 0 ( r ( x ) − ε ) ≤ ct K 0 ( r ( x ) ) + δ . \operatorname{ct}_{K_{0}}(r(x)-\varepsilon) \le\operatorname{ct}_{K_{0}}(r(x))+\delta. ct K 0 ( r ( x ) − ε ) ≤ ct K 0 ( r ( x )) + δ .
For this choice of ε \varepsilon ε , the smooth function f ε f_{\varepsilon} f ε satisfies
f ε ≥ r , f ε ( x ) = r ( x ) , f_{\varepsilon} \ge r, \qquad f_{\varepsilon}(x) = r(x), f ε ≥ r , f ε ( x ) = r ( x ) ,
and
Hess f ε ( x ) ≤ ( ct K 0 ( r ( x ) ) + δ ) ⟨ ⋅ , ⋅ ⟩ x . \operatorname{Hess} f_{\varepsilon}(x) \le\left(\operatorname{ct}_{K_{0}}(r(x))+\delta\right)\langle\cdot,\cdot\rangle_{x}. Hess f ε ( x ) ≤ ( ct K 0 ( r ( x )) + δ ) ⟨ ⋅ , ⋅ ⟩ x .
Since δ > 0 \delta> 0 δ > 0 is arbitrary, the definition of the support sense implies that
Hess r ( x ) ≤ ct K 0 ( r ( x ) ) ⟨ ⋅ , ⋅ ⟩ x \operatorname{Hess} r(x) \leq\operatorname{ct}_{K_0}(r(x))\langle\cdot,\cdot\rangle_x Hess r ( x ) ≤ ct K 0 ( r ( x )) ⟨ ⋅ , ⋅ ⟩ x
in the support sense. As x ∈ M ∖ { x 0 } x \in M \setminus\{x_0\} x ∈ M ∖ { x 0 } was arbitrary, the first assertion follows.
Finally, let R ∗ > 0 R_* > 0 R ∗ > 0 and suppose that r ( x ) > R ∗ r(x) > R_* r ( x ) > R ∗ . Since ct K 0 \operatorname{ct}_{K_0} ct K 0 is decreasing,
ct K 0 ( r ( x ) ) ≤ ct K 0 ( R ∗ ) . \operatorname{ct}_{K_0}(r(x)) \leq\operatorname{ct}_{K_0}(R_*). ct K 0 ( r ( x )) ≤ ct K 0 ( R ∗ ) .
The first assertion therefore implies
Hess r ( x ) ≤ ct K 0 ( R ∗ ) ⟨ ⋅ , ⋅ ⟩ x \operatorname{Hess} r(x) \leq\operatorname{ct}_{K_0}(R_*)\langle\cdot,\cdot\rangle_x Hess r ( x ) ≤ ct K 0 ( R ∗ ) ⟨ ⋅ , ⋅ ⟩ x
in the support sense at every x ∈ { r > R ∗ } x \in\{r > R_*\} x ∈ { r > R ∗ } . This proves the second assertion. □
Regularization and the cutoff inequality
The following result follows from the divergence theorem and its extension to vector fields with locally Lipschitz components; see [15 , 6 , 7 ] .
Lemma 2.8. Let W W W be a compactly supported vector field on M M M whose components in local coordinates are locally Lipschitz. Then div W ∈ L ∞ ( M ) \operatorname{div} W \in L^\infty(M) div W ∈ L ∞ ( M ) is defined almost everywhere and ∫ M div W = 0 \int_M \operatorname{div} W = 0 ∫ M div W = 0 .
Remark 2.9. Lemmas 2.10 and 2.11 below are essentially known. For β ≡ 0 \beta\equiv0 β ≡ 0 , Lemma 2.10 is the classical fact that a continuous function of one variable admitting upper support functions with arbitrarily small second derivative is concave; see [19 Section 1, Lemmas 2 and 4] , where the proofs are omitted as standard. The general case reduces to this one by subtracting a C 2 C^2 C 2 function B B B with B ′ ′ = β B'' = \beta B ′′ = β .
Lemma 2.11 is a variant, with Euclidean convolution in normal coordinates in place of Riemannian convolution, of the smoothing results of Greene and Wu [9 §2] . Indeed, if f f f is locally L L L -Lipschitz and Hess f ≤ a ⟨ ⋅ , ⋅ ⟩ \operatorname{Hess} f \leq a\langle\cdot,\cdot\rangle Hess f ≤ a ⟨ ⋅ , ⋅ ⟩ in the support sense on W W W , then for every η > 0 \eta> 0 η > 0 the function − f -f − f is ( − a − η ) (-a-\eta) ( − a − η ) -convex on W W W in the sense of [9 p. 60] . To see this, apply [19 Lemmas 2 and 4] near each p ∈ W p \in W p ∈ W to − f + 1 2 ( a + η / 2 ) dist M ( p , ⋅ ) 2 -f + \frac{1}{2}(a+\eta/2)\operatorname{dist}_M(p,\cdot)^2 − f + 2 1 ( a + η /2 ) dist M ( p , ⋅ ) 2 . Hence, by [9 pp. 57, 59, 60–61] , the Riemannian convolutions f ε f_\varepsilon f ε of f f f are smooth near any compact set K ⊂ W K \subset W K ⊂ W , converge to f f f uniformly on K K K , and, for all sufficiently small ε \varepsilon ε , satisfy ∣ ∇ f ε ∣ < L + η |\nabla f_\varepsilon| < L+\eta ∣∇ f ε ∣ < L + η and Hess f ε ≤ ( a + η ) ⟨ ⋅ , ⋅ ⟩ \operatorname{Hess} f_\varepsilon\leq(a+\eta)\langle\cdot,\cdot\rangle Hess f ε ≤ ( a + η ) ⟨ ⋅ , ⋅ ⟩ near K K K (see also [8 ] ).
We also use the almost everywhere convergence ∇ f ε → ∇ f \nabla f_\varepsilon\to\nabla f ∇ f ε → ∇ f , which is not stated in [9 ] . It follows from the uniform semiconcavity of the approximations in local charts and [16 Theorem 24.5] . Since we need these facts in the precise form stated above, we give self-contained proofs of Lemmas 2.10 and 2.11 in Appendix A.
Lemma 2.10. Let I ⊂ R I \subset\mathbb{R} I ⊂ R be an open interval, and let v , β ∈ C 0 ( I ) v,\beta\in C^0(I) v , β ∈ C 0 ( I ) . Assume that for every t 0 ∈ I t_0 \in I t 0 ∈ I and every ε > 0 \varepsilon> 0 ε > 0 , there exists a C 2 C^2 C 2 function ϕ \phi ϕ defined on a neighborhood of t 0 t_0 t 0 such that
ϕ ≥ v , ϕ ( t 0 ) = v ( t 0 ) , ϕ ′ ′ ( t 0 ) ≤ β ( t 0 ) + ε . \phi\geq v,\qquad\phi(t_0)=v(t_0),\qquad\phi''(t_0)\leq\beta(t_0)+\varepsilon. ϕ ≥ v , ϕ ( t 0 ) = v ( t 0 ) , ϕ ′′ ( t 0 ) ≤ β ( t 0 ) + ε .
Then v ′ ′ ≤ β v^{\prime\prime}\leq\beta v ′′ ≤ β in the sense of distributions on I I I ; that is,
∫ I v ( t ) ζ ′ ′ ( t ) d t ≤ ∫ I β ( t ) ζ ( t ) d t \int_I v(t)\zeta^{\prime\prime}(t)\,\mathrm{d}t\leq\int_I\beta(t)\zeta(t)\,\mathrm{d}t ∫ I v ( t ) ζ ′′ ( t ) d t ≤ ∫ I β ( t ) ζ ( t ) d t
for every nonnegative ζ ∈ C c ∞ ( I ) \zeta\in C^\infty_c(I) ζ ∈ C c ∞ ( I ) .
Fix a mollifier μ ∈ C c ∞ ( R n + 1 ) \mu\in C^\infty_c(\mathbb{R}^{n+1}) μ ∈ C c ∞ ( R n + 1 ) , μ ≥ 0 \mu\geq0 μ ≥ 0 , supp μ ⊂ B ( 0 , 1 ) \operatorname{supp}\mu\subset B(0,1) supp μ ⊂ B ( 0 , 1 ) , ∫ μ = 1 \int\mu=1 ∫ μ = 1 , and put μ δ ( z ) : = δ − n − 1 μ ( z / δ ) \mu_\delta(z):=\delta^{-n-1}\mu(z/\delta) μ δ ( z ) := δ − n − 1 μ ( z / δ ) . For a symmetric matrix S S S we write ∥ S ∥ \lVert S\rVert ∥ S ∥ for its operator norm, and for the Christoffel symbols in a chart we put ∥ Γ ( y ) ∥ : = sup { ∣ Γ i j k ( y ) ξ i ξ j ζ k ∣ : ∣ ξ ∣ = ∣ ζ ∣ = 1 } \lVert\Gamma(y)\rVert:=\sup\{|\Gamma^k_{ij}(y)\xi^i\xi^j\zeta_k|:|\xi|=|\zeta|=1\} ∥ Γ ( y )∥ := sup { ∣ Γ ij k ( y ) ξ i ξ j ζ k ∣ : ∣ ξ ∣ = ∣ ζ ∣ = 1 } (Euclidean norms).
Lemma 2.11 (regularization in normal coordinates). Let W ⊂ M W\subset M W ⊂ M be open, a ∈ R a\in\mathbb{R} a ∈ R , L > 0 L>0 L > 0 , and let f ∈ C 0 ( W ) f\in C^0(W) f ∈ C 0 ( W ) be such that every point of W W W has a neighbourhood on which f f f is L L L -Lipschitz, and Hess f ≤ a ⟨ ⋅ , ⋅ ⟩ \operatorname{Hess}f\leq a\langle\cdot,\cdot\rangle Hess f ≤ a ⟨ ⋅ , ⋅ ⟩ in the support sense at every point of W W W . Let K ⊂ W K\subset W K ⊂ W be compact and η > 0 \eta>0 η > 0 . Then there are σ > 0 \sigma>0 σ > 0 and points c 1 , … , c m ∈ K c_1,\ldots,c_m\in K c 1 , … , c m ∈ K such that, with U i : = B 2 σ ( c i ) U_i:=B_{2\sigma}(c_i) U i := B 2 σ ( c i ) :
(i) K ⊂ U 1 ∪ ⋯ ∪ U m K\subset U_1\cup\cdots\cup U_m K ⊂ U 1 ∪ ⋯ ∪ U m , and exp c i \exp_{c_i} exp c i maps the Euclidean ball B ( 0 , 4 σ ) ⊂ T c i M ≅ R n + 1 B(0,4\sigma)\subset T_{c_i}M\cong\mathbb{R}^{n+1} B ( 0 , 4 σ ) ⊂ T c i M ≅ R n + 1 (identified by an orthonormal basis) diffeomorphically onto the geodesic ball B 4 σ ( c i ) ⊂ W B_{4\sigma}(c_i)\subset W B 4 σ ( c i ) ⊂ W ;
(ii) for 0 < δ < σ 0<\delta<\sigma 0 < δ < σ the functions f i , δ : = ( u i ∗ μ δ ) ∘ exp c i − 1 f_{i,\delta}:=(u_i*\mu_\delta)\circ\exp_{c_i}^{-1} f i , δ := ( u i ∗ μ δ ) ∘ exp c i − 1 , u i : = f ∘ exp c i u_i:=f\circ\exp_{c_i} u i := f ∘ exp c i , are smooth on B 3 σ ( c i ) B_{3\sigma}(c_i) B 3 σ ( c i ) , ∣ ∇ f i , δ ∣ ≤ 2 L |\nabla f_{i,\delta}|\leq2L ∣∇ f i , δ ∣ ≤ 2 L there, f i , δ → f f_{i,\delta}\to f f i , δ → f uniformly on U i ‾ \overline{U_i} U i and ∇ f i , δ → ∇ f \nabla f_{i,\delta}\to\nabla f ∇ f i , δ → ∇ f a.e. on U i U_i U i as δ → 0 \delta\to0 δ → 0 ;
(iii) there are ω i ( δ ) → 0 \omega_i(\delta)\to0 ω i ( δ ) → 0 such that Hess f i , δ ≤ ( a + η + ω i ( δ ) ) ⟨ ⋅ , ⋅ ⟩ \operatorname{Hess}f_{i,\delta}\leq(a+\eta+\omega_i(\delta))\langle\cdot,\cdot\rangle Hess f i , δ ≤ ( a + η + ω i ( δ )) ⟨ ⋅ , ⋅ ⟩ on U i ‾ \overline{U_i} U i .
The next proposition is the form of Lemma 2.4 that will be used with the cutoff functions of Section 3.
Proposition 2.12 (cutoff inequality). Let M M M be complete with sec M ≥ − K 0 \operatorname{sec}_M\geq-K_0 sec M ≥ − K 0 , x 0 ∈ M x_0\in M x 0 ∈ M , r : = dist M ( x 0 , ⋅ ) r:=\operatorname{dist}_M(x_0,\cdot) r := dist M ( x 0 , ⋅ ) , 0 < R ∗ < R 1 < R 2 0<R_*<R_1<R_2 0 < R ∗ < R 1 < R 2 and C 1 : = ct K 0 ( R ∗ ) C_1:=\operatorname{ct}_{K_0}(R_*) C 1 := ct K 0 ( R ∗ ) . Let Φ ∈ C ∞ ( R ) \Phi\in C^\infty(\mathbb{R}) Φ ∈ C ∞ ( R ) satisfy Φ ′ ≤ 0 \Phi'\leq0 Φ ′ ≤ 0 , Φ ≡ Φ − \Phi\equiv\Phi_- Φ ≡ Φ − on ( − ∞ , R 1 ] (-\infty,R_1] ( − ∞ , R 1 ] for a constant Φ − \Phi_- Φ − , and Φ ≡ 0 \Phi\equiv0 Φ ≡ 0 on [ R 2 , ∞ ) [R_2,\infty) [ R 2 , ∞ ) , and let λ ∈ C 1 ( M ) \lambda\in C^1(M) λ ∈ C 1 ( M ) , λ ≥ 0 \lambda\geq0 λ ≥ 0 , be constant along the leaves. Then φ : = Φ ∘ r \varphi:=\Phi\circ r φ := Φ ∘ r is Lipschitz with compact support, ∇ φ = Φ ′ ( r ) ∇ r \nabla\varphi=\Phi'(r)\nabla r ∇ φ = Φ ′ ( r ) ∇ r a.e., and
∫ M λ ⟨ ∇ φ , T ⟩ − n ∫ M H λ N ( φ ) ≥ ∫ M λ [ Φ ′ ′ ( r ) ∣ ∇ L r ∣ 2 + n C 1 Φ ′ ( r ) ] , (4) \int_M\lambda\langle\nabla\varphi,T\rangle-n\int_MH\lambda N(\varphi)\geq\int_M\lambda\left[\Phi^{\prime\prime}(r)|\nabla_Lr|^2+nC_1\Phi'(r)\right], \tag*{(4)} ∫ M λ ⟨ ∇ φ , T ⟩ − n ∫ M H λ N ( φ ) ≥ ∫ M λ [ Φ ′′ ( r ) ∣ ∇ L r ∣ 2 + n C 1 Φ ′ ( r ) ] , (4) where ∇ L r : = ∇ r − N ( r ) N \nabla_Lr:=\nabla r-N(r)N ∇ L r := ∇ r − N ( r ) N is defined a.e., and the integrand on the right vanishes outside S : = { R 1 ≤ r ≤ R 2 } S:=\{R_1\leq r\leq R_2\} S := { R 1 ≤ r ≤ R 2 } .
Proof. By the Hopf-Rinow theorem, the closed metric balls of M M M are compact. Since the distance function r = dist M ( x 0 , ⋅ ) r=\operatorname{dist}_M(x_0,\cdot) r = dist M ( x 0 , ⋅ ) is 1 1 1 -Lipschitz, the function φ = Φ ∘ r \varphi=\Phi\circ r φ = Φ ∘ r is Lipschitz. Moreover, φ = 0 \varphi=0 φ = 0 on { r ≥ R 2 } \{r\geq R_2\} { r ≥ R 2 } , and hence φ \varphi φ has compact support. The chain rule for Lipschitz functions gives
∇ φ = Φ ′ ( r ) ∇ r almost everywhere on M . \nabla\varphi=\Phi'(r)\nabla r\quad\text{almost everywhere on }M. ∇ φ = Φ ′ ( r ) ∇ r almost everywhere on M .
In particular, ∇ φ = 0 \nabla\varphi=0 ∇ φ = 0 almost everywhere outside S : = { x ∈ M : R 1 ≤ r ( x ) ≤ R 2 } S:=\{x\in M:R_1\le r(x)\le R_2\} S := { x ∈ M : R 1 ≤ r ( x ) ≤ R 2 } . The set S S S is compact and is contained in the open set W : = { r > R ∗ } W:=\{r>R_*\} W := { r > R ∗ } . For every compactly supported Lipschitz function ψ \psi ψ , set
Q ( ψ ) : = ∫ M λ ⟨ ∇ ψ , T ⟩ − n ∫ M H λ N ( ψ ) . Q(\psi):=\int_M\lambda\langle\nabla\psi,T\rangle-n\int_M H\lambda N(\psi). Q ( ψ ) := ∫ M λ ⟨ ∇ ψ , T ⟩ − n ∫ M H λ N ( ψ ) .
We first establish the desired inequality by approximation.
By Lemma 2.7 , the distance function satisfies
Hess r ≤ C 1 g \operatorname{Hess}r\le C_1g Hess r ≤ C 1 g
in the support sense on W W W , where C 1 = c t K 0 ( R ∗ ) C_1=c t K_0(R_*) C 1 = c t K 0 ( R ∗ ) . Fix η > 0 \eta>0 η > 0 . Applying Lemma 2.11 to f = r f=r f = r , L = 1 L=1 L = 1 , a = C 1 a=C_1 a = C 1 , and the compact set S S S , we obtain finitely many open sets U i = B 2 σ ( c i ) U_i=B_{2\sigma}(c_i) U i = B 2 σ ( c i ) covering S S S , with U i ‾ ⊂ W \overline{U_i}\subset W U i ⊂ W , and smooth functions r i , δ r_{i,\delta} r i , δ on B 3 σ ( c i ) B_{3\sigma}(c_i) B 3 σ ( c i ) , for all sufficiently small δ > 0 \delta>0 δ > 0 , such that
r i , δ ⟶ r uniformly on U i ‾ , r_{i,\delta}\longrightarrow r\quad\text{uniformly on }\overline{U_i}, r i , δ ⟶ r uniformly on U i ,
∇ r i , δ ⟶ ∇ r almost everywhere on U i , \nabla r_{i,\delta}\longrightarrow\nabla r\quad\text{almost everywhere on }U_i, ∇ r i , δ ⟶ ∇ r almost everywhere on U i ,
and
Hess r i , δ ≤ ( C 1 + η + ω i ( δ ) ) g on U i ‾ , ω i ( δ ) ⟶ 0. \operatorname{Hess}r_{i,\delta}\le(C_1+\eta+\omega_i(\delta))g\quad\text{on }\overline{U_i},\qquad\omega_i(\delta)\longrightarrow0. Hess r i , δ ≤ ( C 1 + η + ω i ( δ )) g on U i , ω i ( δ ) ⟶ 0.
Here and below, δ \delta δ is taken sufficiently small that all the regularizations are defined on the indicated sets.
Let O − : = { r < R 1 } O_-:=\{r<R_1\} O − := { r < R 1 } , O + : = { r > R 2 } O_+:=\{r>R_2\} O + := { r > R 2 } . Choose a smooth partition of unity ζ − + ζ + + ∑ i = 1 m ζ i = 1 \zeta_-+\zeta_++\sum_{i=1}^{m}\zeta_i=1 ζ − + ζ + + ∑ i = 1 m ζ i = 1 subordinate to the cover { O − , O + , U 1 , … , U m } \{O_-,O_+,U_1,\ldots,U_m\} { O − , O + , U 1 , … , U m } , with supp ζ − \operatorname{supp}\zeta_- supp ζ − and supp ζ i \operatorname{supp}\zeta_i supp ζ i compact. Define
φ δ : = ζ − Φ − + ∑ i = 1 m ζ i Φ ( r i , δ ) . \varphi_\delta:=\zeta_-\Phi_-+\sum_{i=1}^{m}\zeta_i\Phi(r_{i,\delta}). φ δ := ζ − Φ − + i = 1 ∑ m ζ i Φ ( r i , δ ) .
Each product ζ i Φ ( r i , δ ) \zeta_i\Phi(r_{i,\delta}) ζ i Φ ( r i , δ ) is understood on U i U_i U i and extended by zero outside U i U_i U i ; this extension is smooth because supp ζ i ⋐ U i \operatorname{supp}\zeta_i\Subset U_i supp ζ i ⋐ U i . Thus φ δ ∈ C ∞ ( M ) \varphi_\delta\in C^\infty(M) φ δ ∈ C ∞ ( M ) , and all the functions φ δ \varphi_\delta φ δ have support in a fixed compact subset of M M M .
Since Φ ( r ) = Φ − \Phi(r)=\Phi_- Φ ( r ) = Φ − on O − O_- O − and Φ ( r ) = 0 \Phi(r)=0 Φ ( r ) = 0 on O + O_+ O + , the partition of unity gives
φ = ζ − Φ − + ∑ i = 1 m ζ i Φ ( r ) . \varphi=\zeta_-\Phi_-+\sum_{i=1}^{m}\zeta_i\Phi(r). φ = ζ − Φ − + i = 1 ∑ m ζ i Φ ( r ) .
The uniform convergence r i , δ → r r_{i,\delta}\to r r i , δ → r , together with the boundedness of Φ ′ \Phi' Φ ′ , implies φ δ → φ \varphi_\delta\to\varphi φ δ → φ uniformly. Moreover, the chain rule gives
∇ ( Φ ( r i , δ ) ) = Φ ′ ( r i , δ ) ∇ r i , δ . \nabla\bigl(\Phi(r_{i,\delta})\bigr)=\Phi'(r_{i,\delta})\nabla r_{i,\delta}. ∇ ( Φ ( r i , δ ) ) = Φ ′ ( r i , δ ) ∇ r i , δ .
Consequently, the almost-everywhere convergence of ∇ r i , δ \nabla r_{i,\delta} ∇ r i , δ and the uniform convergence of r i , δ r_{i,\delta} r i , δ imply
∇ φ δ ⟶ ∇ φ almost everywhere on M . \nabla\varphi_{\delta} \longrightarrow\nabla\varphi\quad\text{almost everywhere on } M. ∇ φ δ ⟶ ∇ φ almost everywhere on M .
The gradients ∇ φ δ \nabla\varphi_{\delta} ∇ φ δ are uniformly bounded, and their supports lie in a fixed compact set. Since λ \lambda λ , T T T , and H H H are bounded on that compact set, the dominated convergence theorem yields
Q ( φ δ ) ⟶ Q ( φ ) . (5) Q(\varphi_{\delta}) \longrightarrow Q(\varphi). \tag*{(5)} Q ( φ δ ) ⟶ Q ( φ ) . (5)
By Lemma 2.4, applied to the smooth compactly supported function φ δ \varphi_{\delta} φ δ , we have
Q ( φ δ ) = ∫ M λ tr T L Hess φ δ . Q(\varphi_{\delta})=\int_M \lambda\,\operatorname{tr}_{TL}\operatorname{Hess}\varphi_{\delta}. Q ( φ δ ) = ∫ M λ tr T L Hess φ δ .
The product rule for the Hessian gives
Hess φ δ = ∑ i = 1 m ζ i Hess ( Φ ( r i , δ ) ) + E δ . \operatorname{Hess}\varphi_{\delta}=\sum_{i=1}^{m}\zeta_i\operatorname{Hess}\bigl(\Phi(r_{i,\delta})\bigr)+E_{\delta}. Hess φ δ = i = 1 ∑ m ζ i Hess ( Φ ( r i , δ ) ) + E δ .
where
E δ = Φ − Hess ζ − + ∑ i = 1 m [ d ζ i ⊗ d ( Φ ( r i , δ ) ) + d ( Φ ( r i , δ ) ) ⊗ d ζ i + Φ ( r i , δ ) Hess ζ i ] . E_{\delta}=\Phi_{-}\operatorname{Hess}\zeta_{-}+\sum_{i=1}^{m}\left[d\zeta_i\otimes d\bigl(\Phi(r_{i,\delta})\bigr)+d\bigl(\Phi(r_{i,\delta})\bigr)\otimes d\zeta_i+\Phi(r_{i,\delta})\operatorname{Hess}\zeta_i\right]. E δ = Φ − Hess ζ − + i = 1 ∑ m [ d ζ i ⊗ d ( Φ ( r i , δ ) ) + d ( Φ ( r i , δ ) ) ⊗ d ζ i + Φ ( r i , δ ) Hess ζ i ] .
By the convergence already established, E δ E_{\delta} E δ converges almost everywhere to
E 0 = Φ − Hess ζ − + ∑ i = 1 m [ d ζ i ⊗ d φ + d φ ⊗ d ζ i + φ Hess ζ i ] . E_{0}=\Phi_{-}\operatorname{Hess}\zeta_{-}+\sum_{i=1}^{m}\left[d\zeta_i\otimes d\varphi+d\varphi\otimes d\zeta_i+\varphi\operatorname{Hess}\zeta_i\right]. E 0 = Φ − Hess ζ − + i = 1 ∑ m [ d ζ i ⊗ d φ + d φ ⊗ d ζ i + φ Hess ζ i ] .
Furthermore, E δ E_{\delta} E δ is uniformly bounded and supported in a fixed compact set. We claim that E 0 = 0 E_{0}=0 E 0 = 0 almost everywhere. Indeed,
∑ i = 1 m ζ i = 1 − ζ − − ζ + , \sum_{i=1}^{m}\zeta_i=1-\zeta_{-}-\zeta_{+}, i = 1 ∑ m ζ i = 1 − ζ − − ζ + ,
so
∑ i = 1 m d ζ i = − d ζ − − d ζ + , ∑ i = 1 m Hess ζ i = − Hess ζ − − Hess ζ + . \sum_{i=1}^{m}d\zeta_i=-d\zeta_{-}-d\zeta_{+},\qquad\sum_{i=1}^{m}\operatorname{Hess}\zeta_i=-\operatorname{Hess}\zeta_{-}-\operatorname{Hess}\zeta_{+}. i = 1 ∑ m d ζ i = − d ζ − − d ζ + , i = 1 ∑ m Hess ζ i = − Hess ζ − − Hess ζ + .
Since supp ζ − ⊂ O − \operatorname{supp}\zeta_{-}\subset O_{-} supp ζ − ⊂ O − and supp ζ + ⊂ O + \operatorname{supp}\zeta_{+}\subset O_{+} supp ζ + ⊂ O + , we have
d φ = 0 almost everywhere on supp d ζ − ∪ supp d ζ + , d\varphi=0\quad\text{almost everywhere on }\operatorname{supp}d\zeta_{-}\cup\operatorname{supp}d\zeta_{+}, d φ = 0 almost everywhere on supp d ζ − ∪ supp d ζ + ,
and φ = Φ − \varphi=\Phi_{-} φ = Φ − on supp ζ − \operatorname{supp}\zeta_{-} supp ζ − , φ = 0 \varphi=0 φ = 0 on supp ζ + \operatorname{supp}\zeta_{+} supp ζ + . It follows that the terms involving d φ d\varphi d φ vanish almost everywhere, while
E 0 = Φ − Hess ζ − − φ Hess ζ − − φ Hess ζ + = 0 almost everywhere . E_{0}=\Phi_{-}\operatorname{Hess}\zeta_{-}-\varphi\operatorname{Hess}\zeta_{-}-\varphi\operatorname{Hess}\zeta_{+}=0\quad\text{almost everywhere}. E 0 = Φ − Hess ζ − − φ Hess ζ − − φ Hess ζ + = 0 almost everywhere .
The dominated convergence theorem therefore gives
∫ M λ tr T L E δ ⟶ 0. (6) \int_M \lambda\,\operatorname{tr}_{TL}E_{\delta}\longrightarrow0. \tag*{(6)} ∫ M λ tr T L E δ ⟶ 0. (6)
It remains to estimate the main Hessian terms. For each i i i , the chain rule for the Hessian yields
Hess ( Φ ( r i , δ ) ) = Φ ′ ′ ( r i , δ ) d r i , δ ⊗ d r i , δ + Φ ′ ( r i , δ ) Hess r i , δ . \operatorname{Hess}(\Phi(r_{i,\delta}))=\Phi^{\prime\prime}(r_{i,\delta})\,dr_{i,\delta}\otimes dr_{i,\delta}+\Phi^\prime(r_{i,\delta})\operatorname{Hess}r_{i,\delta}. Hess ( Φ ( r i , δ )) = Φ ′′ ( r i , δ ) d r i , δ ⊗ d r i , δ + Φ ′ ( r i , δ ) Hess r i , δ .
Taking the trace over the tangent space T L TL T L of a leaf, whose dimension is n n n , gives
tr T L Hess ( Φ ( r i , δ ) ) = Φ ′ ′ ( r i , δ ) ∣ ∇ L r i , δ ∣ 2 + Φ ′ ( r i , δ ) tr T L Hess r i , δ , \operatorname{tr}_{TL}\operatorname{Hess}(\Phi(r_{i,\delta}))=\Phi^{\prime\prime}(r_{i,\delta})\lvert\nabla_Lr_{i,\delta}\rvert^2+\Phi^\prime(r_{i,\delta})\operatorname{tr}_{TL}\operatorname{Hess}r_{i,\delta}, tr T L Hess ( Φ ( r i , δ )) = Φ ′′ ( r i , δ ) ∣ ∇ L r i , δ ∣ 2 + Φ ′ ( r i , δ ) tr T L Hess r i , δ ,
where ∇ L r i , δ : = ∇ r i , δ − N ( r i , δ ) N \nabla_Lr_{i,\delta}:=\nabla r_{i,\delta}-N(r_{i,\delta})N ∇ L r i , δ := ∇ r i , δ − N ( r i , δ ) N . By Lemma 2.11(iii),
tr T L Hess r i , δ ≤ n ( C 1 + η + ω i ( δ ) ) . \operatorname{tr}_{TL}\operatorname{Hess}r_{i,\delta}\le n(C_1+\eta+\omega_i(\delta)). tr T L Hess r i , δ ≤ n ( C 1 + η + ω i ( δ )) .
Because Φ ′ ≤ 0 \Phi^\prime\le0 Φ ′ ≤ 0 , multiplying this inequality by Φ ′ ( r i , δ ) \Phi^\prime(r_{i,\delta}) Φ ′ ( r i , δ ) reverses its direction. Hence
tr T L Hess ( Φ ( r i , δ ) ) ≥ Φ ′ ′ ( r i , δ ) ∣ ∇ L r i , δ ∣ 2 + n Φ ′ ( r i , δ ) ( C 1 + η + ω i ( δ ) ) . \operatorname{tr}_{TL}\operatorname{Hess}(\Phi(r_{i,\delta}))\ge\Phi^{\prime\prime}(r_{i,\delta})\lvert\nabla_Lr_{i,\delta}\rvert^2+n\Phi^\prime(r_{i,\delta})(C_1+\eta+\omega_i(\delta)). tr T L Hess ( Φ ( r i , δ )) ≥ Φ ′′ ( r i , δ ) ∣ ∇ L r i , δ ∣ 2 + n Φ ′ ( r i , δ ) ( C 1 + η + ω i ( δ )) .
Multiplying by λ ζ i ≥ 0 \lambda\zeta_i\ge0 λ ζ i ≥ 0 , integrating, and summing over i i i , we obtain
Q ( φ δ ) ≥ ∑ i = 1 m ∫ M λ ζ i Φ ′ ′ ( r i , δ ) ∣ ∇ L r i , δ ∣ 2 + n ∑ i = 1 m ∫ M λ ζ i Φ ′ ( r i , δ ) ( C 1 + η + ω i ( δ ) ) + ∫ M λ tr T L E δ . \begin{aligned}
Q(\varphi_\delta)&\ge\sum_{i=1}^{m}\int_M\lambda\zeta_i\Phi^{\prime\prime}(r_{i,\delta})\lvert\nabla_Lr_{i,\delta}\rvert^2\\
&\quad+n\sum_{i=1}^{m}\int_M\lambda\zeta_i\Phi^\prime(r_{i,\delta})(C_1+\eta+\omega_i(\delta))\\
&\quad+\int_M\lambda\operatorname{tr}_{TL}E_\delta.
\end{aligned} Q ( φ δ ) ≥ i = 1 ∑ m ∫ M λ ζ i Φ ′′ ( r i , δ ) ∣ ∇ L r i , δ ∣ 2 + n i = 1 ∑ m ∫ M λ ζ i Φ ′ ( r i , δ ) ( C 1 + η + ω i ( δ )) + ∫ M λ tr T L E δ .
For almost every point of U i U_i U i , the convergence of the gradients implies
∇ L r i , δ ⟶ ∇ L r : = ∇ r − N ( r ) N . \nabla_Lr_{i,\delta}\longrightarrow\nabla_Lr:=\nabla r-N(r)N. ∇ L r i , δ ⟶ ∇ L r := ∇ r − N ( r ) N .
All terms in the preceding integrals are dominated by integrable functions independent of sufficiently small δ \delta δ : this follows from the uniform gradient bounds in Lemma 2.11(ii), the boundedness of Φ ′ \Phi^\prime Φ ′ and Φ ′ ′ \Phi^{\prime\prime} Φ ′′ , the convergence ω i ( δ ) → 0 \omega_i(\delta)\to0 ω i ( δ ) → 0 , and the compactness of the supports. Therefore, by dominated convergence, (5) , and (6) , we obtain
Q ( φ ) ≥ ∑ i = 1 m ∫ M λ ζ i [ Φ ′ ′ ( r ) ∣ ∇ L r ∣ 2 + n Φ ′ ( r ) ( C 1 + η ) ] . Q(\varphi)\ge\sum_{i=1}^{m}\int_M\lambda\zeta_i\left[\Phi^{\prime\prime}(r)\lvert\nabla_Lr\rvert^2+n\Phi^\prime(r)(C_1+\eta)\right]. Q ( φ ) ≥ i = 1 ∑ m ∫ M λ ζ i [ Φ ′′ ( r ) ∣ ∇ L r ∣ 2 + n Φ ′ ( r ) ( C 1 + η ) ] .
Since Φ ′ \Phi^\prime Φ ′ and Φ ′ ′ \Phi^{\prime\prime} Φ ′′ vanish outside S S S , and ζ − = ζ + = 0 \zeta_-=\zeta_+=0 ζ − = ζ + = 0 on S S S , we have ∑ i = 1 m ζ i = 1 \sum_{i=1}^{m}\zeta_i=1 ∑ i = 1 m ζ i = 1 on S S S . Recalling that ∇ L r = ∇ r − N ( r ) N = ∇ L r \nabla_Lr=\nabla r-N(r)N=\nabla_Lr ∇ L r = ∇ r − N ( r ) N = ∇ L r almost everywhere, it follows that
Q ( φ ) ≥ ∫ M λ [ Φ ′ ′ ( r ) ∣ ∇ L r ∣ 2 + n Φ ′ ( r ) ( C 1 + η ) ] . Q(\varphi)\ge\int_M\lambda\left[\Phi^{\prime\prime}(r)\lvert\nabla_Lr\rvert^2+n\Phi^\prime(r)(C_1+\eta)\right]. Q ( φ ) ≥ ∫ M λ [ Φ ′′ ( r ) ∣ ∇ L r ∣ 2 + n Φ ′ ( r ) ( C 1 + η ) ] .
Finally, letting η → 0 + \eta\to0^+ η → 0 + and using dominated convergence on the compact set S S S , we conclude that
Q ( φ ) ≥ ∫ M λ [ Φ ′ ′ ( r ) ∣ ∇ L r ∣ 2 + n C 1 Φ ′ ( r ) ] , Q(\varphi)\ge\int_M\lambda\left[\Phi^{\prime\prime}(r)\lvert\nabla_Lr\rvert^2+nC_1\Phi^\prime(r)\right], Q ( φ ) ≥ ∫ M λ [ Φ ′′ ( r ) ∣ ∇ L r ∣ 2 + n C 1 Φ ′ ( r ) ] ,
which is precisely (4) . □
Remark 2.13. If r r r were C 2 C^2 C 2 on a neighbourhood of S S S , Proposition 2.12 would follow directly from Lemma 2.4, since then φ ∈ C c 2 ( M ) \varphi\in C_c^2(M) φ ∈ C c 2 ( M ) and tr T L Hess φ = Φ ′ ′ ( r ) ∣ ∇ L r ∣ 2 + Φ ′ ( r ) tr T L Hess r ≥ Φ ′ ′ ( r ) ∣ ∇ L r ∣ 2 + n C 1 Φ ′ ( r ) \operatorname{tr}_{T_L}\operatorname{Hess}\varphi= \Phi''(r)|\nabla_Lr|^2 + \Phi'(r)\operatorname{tr}_{T_L}\operatorname{Hess}r \ge\Phi''(r)|\nabla_Lr|^2 + nC_1\Phi'(r) tr T L Hess φ = Φ ′′ ( r ) ∣ ∇ L r ∣ 2 + Φ ′ ( r ) tr T L Hess r ≥ Φ ′′ ( r ) ∣ ∇ L r ∣ 2 + n C 1 Φ ′ ( r ) . The approximation argument is needed because r r r is not differentiable on the cut locus of x 0 x_0 x 0 . Alternatively one could smooth r r r by the Riemannian convolution of Greene and Wu [9 ] .
The flow field
Fix a constant c 0 > K 0 c_0 > \sqrt{K_0} c 0 > K 0 and an integer s ≥ 3 s \ge3 s ≥ 3 such that 2 s > n 2s > n 2 s > n . Set δ 0 2 : = c 0 2 − K 0 > 0 \delta_0^2 := c_0^2 - K_0 > 0 δ 0 2 := c 0 2 − K 0 > 0 . For h ∈ R h \in\mathbb{R} h ∈ R , define
ϕ ( h ) : = { 0 , h ≤ c 0 , h 2 − c 0 2 , h > c 0 . \phi(h) :=
\begin{cases}
0, & h \le c_0,\\
h^2-c_0^2, & h > c_0.
\end{cases} ϕ ( h ) := { 0 , h 2 − c 0 2 , h ≤ c 0 , h > c 0 .
Since c 0 > 0 c_0 > 0 c 0 > 0 , this is equivalent to ϕ ( h ) = ( ( h + ) 2 − c 0 2 ) + \phi(h) = ((h_+)^2-c_0^2)_+ ϕ ( h ) = (( h + ) 2 − c 0 2 ) + , where h + : = max { h , 0 } h_+ := \max\{h,0\} h + := max { h , 0 } . In particular, ϕ \phi ϕ is nonnegative and locally Lipschitz, and it is smooth on R ∖ { c 0 } \mathbb{R} \setminus\{c_0\} R ∖ { c 0 } .
Define
G : = ϕ ( H ) = ( ( H + ) 2 − c 0 2 ) + , F : = H G s , Y : = ( 2 s − n ) H G s + 2 s δ 0 2 H G s − 1 . G := \phi(H) = ((H_+)^2-c_0^2)_+, \qquad F := HG^s, \qquad Y := (2s-n)HG^s + 2s\delta_0^2HG^{s-1}. G := ϕ ( H ) = (( H + ) 2 − c 0 2 ) + , F := H G s , Y := ( 2 s − n ) H G s + 2 s δ 0 2 H G s − 1 .
By the definition of ϕ \phi ϕ , we have
G ( x ) > 0 ⟺ H ( x ) > c 0 . G(x)>0 \quad\Longleftrightarrow\quad H(x)>c_0. G ( x ) > 0 ⟺ H ( x ) > c 0 .
Thus G > 0 G>0 G > 0 precisely on the open set U : = { x ∈ M : H ( x ) > c 0 } U := \{x \in M : H(x)>c_0\} U := { x ∈ M : H ( x ) > c 0 } , and G = 0 G=0 G = 0 on M ∖ U M \setminus U M ∖ U . We next record the regularity of these functions. Since
ϕ ( h ) s = ( h − c 0 ) + s ( h + c 0 ) s , \phi(h)^s = (h-c_0)_+^s(h+c_0)^s, ϕ ( h ) s = ( h − c 0 ) + s ( h + c 0 ) s ,
Leibniz’s rule shows that, for every k = 0 , … , s − 1 k=0,\ldots,s-1 k = 0 , … , s − 1 , the k k k -th derivative on ( c 0 , ∞ ) (c_0,\infty) ( c 0 , ∞ ) tends to zero as h → c 0 + h \to c_0^+ h → c 0 + , because each term contains a factor ( h − c 0 ) s − j (h-c_0)^{s-j} ( h − c 0 ) s − j with j ≤ k < s j \le k < s j ≤ k < s . Since ϕ s ≡ 0 \phi^s \equiv0 ϕ s ≡ 0 on ( − ∞ , c 0 ] (-\infty,c_0] ( − ∞ , c 0 ] , these derivatives match continuously at h = c 0 h=c_0 h = c 0 . Hence ϕ s ∈ C s − 1 ( R ) \phi^s \in C^{s-1}(\mathbb{R}) ϕ s ∈ C s − 1 ( R ) . Similarly, ϕ s − 1 \phi^{s-1} ϕ s − 1 vanishes to order s − 1 s-1 s − 1 at c 0 c_0 c 0 , and hence ϕ s − 1 ∈ C s − 2 ( R ) \phi^{s-1} \in C^{s-2}(\mathbb{R}) ϕ s − 1 ∈ C s − 2 ( R ) . Since s ≥ 3 s \ge3 s ≥ 3 , both ϕ s \phi^s ϕ s and ϕ s − 1 \phi^{s-1} ϕ s − 1 are of class C 1 C^1 C 1 . Assuming, as in the present setting, that H ∈ C 1 ( M ) H \in C^1(M) H ∈ C 1 ( M ) , the chain rule therefore implies that
G s = ϕ ( H ) s ∈ C 1 ( M ) , G s − 1 = ϕ ( H ) s − 1 ∈ C 1 ( M ) . G^s = \phi(H)^s \in C^1(M), \qquad G^{s-1} = \phi(H)^{s-1} \in C^1(M). G s = ϕ ( H ) s ∈ C 1 ( M ) , G s − 1 = ϕ ( H ) s − 1 ∈ C 1 ( M ) .
It follows that F , Y ∈ C 1 ( M ) F,Y \in C^1(M) F , Y ∈ C 1 ( M ) .
Furthermore, F F F and Y Y Y vanish identically on M ∖ U M \setminus U M ∖ U . On U U U , we have H > c 0 > 0 H>c_0>0 H > c 0 > 0 , G > 0 G>0 G > 0 , and δ 0 2 > 0 \delta_0^2>0 δ 0 2 > 0 . Since 2 s > n 2s>n 2 s > n , both coefficients 2 s − n 2s-n 2 s − n and 2 s 2s 2 s are positive. Consequently,
F ≥ 0 , Y ≥ 0 on M . F \ge0, \qquad Y \ge0 \quad\text{on } M. F ≥ 0 , Y ≥ 0 on M .
Finally, define
V : = κ N + λ T , κ : = G s , λ : = 2 s n H G s − 1 . (7) V := \kappa N + \lambda T, \qquad\kappa:= G^s, \qquad\lambda:= \frac{2s}{n}HG^{s-1}. \tag*{(7)} V := κ N + λ T , κ := G s , λ := n 2 s H G s − 1 . (7)
The regularity established above implies that κ , λ ∈ C 1 ( M ) \kappa,\lambda\in C^{1}(M) κ , λ ∈ C 1 ( M ) . Since H H H is constant along each leaf, and κ \kappa κ and λ \lambda λ are functions of H H H , both κ \kappa κ and λ \lambda λ are constant along the leaves. Moreover, λ = 0 \lambda=0 λ = 0 on M ∖ U M\setminus U M ∖ U , whereas on U U U we have H > c 0 > 0 H>c_{0}>0 H > c 0 > 0 and G > 0 G>0 G > 0 . Therefore, λ ≥ 0 \lambda\geq0 λ ≥ 0 on M M M . In particular, provided N N N and T T T are of class C 1 C^{1} C 1 , the vector field V V V is of class C 1 C^{1} C 1 .
Lemma 2.14. If Ric ( N , N ) ≥ − n K 0 \operatorname{Ric}(N,N)\geq-nK_{0} Ric ( N , N ) ≥ − n K 0 , then div V ≥ Y ≥ 0 \operatorname{div}V\geq Y\geq0 div V ≥ Y ≥ 0 on M M M .
Proof. Recall that
κ = G s , λ = 2 s n H G s − 1 , G = ( ( H + ) 2 − c 0 2 ) + , \kappa=G^{s},\qquad\lambda=\frac{2s}{n}HG^{s-1},\qquad G=\left((H_{+})^{2}-c_{0}^{2}\right)_{+}, κ = G s , λ = n 2 s H G s − 1 , G = ( ( H + ) 2 − c 0 2 ) + , where c 0 > 0 c_{0}>0 c 0 > 0 , s ≥ 3 s\geq3 s ≥ 3 , and 2 s > n 2s>n 2 s > n . By the regularity established above, κ , λ ∈ C 1 ( M ) \kappa,\lambda\in C^{1}(M) κ , λ ∈ C 1 ( M ) . Moreover, differentiating κ \kappa κ with respect to H H H gives
κ H : = d κ d H = s G s − 1 d G d H = 2 s H G s − 1 = n λ \kappa_{H}:=\frac{d\kappa}{dH}=sG^{s-1}\frac{dG}{dH}=2sHG^{s-1}=n\lambda κ H := d H d κ = s G s − 1 d H d G = 2 sH G s − 1 = nλ on { H > c 0 } \{H>c_{0}\} { H > c 0 } . On the complement of this set, both sides vanish. By continuity, we therefore have κ H = n λ \kappa_{H}=n\lambda κ H = nλ on M M M .
By Lemma 2.1, the gradients of κ \kappa κ and λ \lambda λ satisfy
∇ κ = κ H g N , ∇ λ = λ H g N , \nabla\kappa=\kappa_{H}\mathfrak{g}N,\qquad\nabla\lambda=\lambda_{H}\mathfrak{g}N, ∇ κ = κ H g N , ∇ λ = λ H g N , where g = N ( H ) \mathfrak{g}=N(H) g = N ( H ) denotes the scalar appearing in Lemma 2.1. In particular, ⟨ ∇ λ , T ⟩ = 0 \langle\nabla\lambda,T\rangle=0 ⟨ ∇ λ , T ⟩ = 0 , because T T T is tangent to the leaves and hence orthogonal to N N N . Using the product rule for divergence, we obtain
div V = div ( κ N ) + div ( λ T ) = ⟨ ∇ κ , N ⟩ + κ div N + ⟨ ∇ λ , T ⟩ + λ div T . \operatorname{div}V=\operatorname{div}(\kappa N)+\operatorname{div}(\lambda T)
=\langle\nabla\kappa,N\rangle+\kappa\operatorname{div}N+\langle\nabla\lambda,T\rangle+\lambda\operatorname{div}T. div V = div ( κ N ) + div ( λ T ) = ⟨ ∇ κ , N ⟩ + κ div N + ⟨ ∇ λ , T ⟩ + λ div T . By the identities in (1.1) and Lemma 2.2,
div N = − n H , div T = ∣ A ∣ 2 + Ric ( N , N ) − n g . \operatorname{div}N=-nH,\qquad\operatorname{div}T=\lvert A\rvert^{2}+\operatorname{Ric}(N,N)-n\mathfrak{g}. div N = − n H , div T = ∣ A ∣ 2 + Ric ( N , N ) − n g . Consequently,
div V = κ H g − n H κ + λ ( ∣ A ∣ 2 + Ric ( N , N ) − n g ) = ( κ H − n λ ) g + λ ( ∣ A ∣ 2 + Ric ( N , N ) ) − n H κ . \operatorname{div}V=\kappa_{H}\mathfrak{g}-nH\kappa+\lambda\left(\lvert A\rvert^{2}+\operatorname{Ric}(N,N)-n\mathfrak{g}\right)
=(\kappa_{H}-n\lambda)\mathfrak{g}+\lambda\left(\lvert A\rvert^{2}+\operatorname{Ric}(N,N)\right)-nH\kappa. div V = κ H g − n H κ + λ ( ∣ A ∣ 2 + Ric ( N , N ) − n g ) = ( κ H − nλ ) g + λ ( ∣ A ∣ 2 + Ric ( N , N ) ) − n H κ . Since κ H = n λ \kappa_{H}=n\lambda κ H = nλ on M M M , the terms involving g \mathfrak{g} g cancel, and we conclude that
div V = λ ( ∣ A ∣ 2 + Ric ( N , N ) ) − n H κ . (8) \operatorname{div}V=\lambda\left(\lvert A\rvert^{2}+\operatorname{Ric}(N,N)\right)-nH\kappa. \tag*{(8)} div V = λ ( ∣ A ∣ 2 + Ric ( N , N ) ) − n H κ . (8) We now prove the desired inequality separately on M ∖ U M\setminus U M ∖ U and on U U U , where U = { H > c 0 } U=\{H>c_{0}\} U = { H > c 0 } .
On M ∖ U M\setminus U M ∖ U , we have G = 0 G=0 G = 0 , and therefore κ = λ = 0 \kappa=\lambda=0 κ = λ = 0 . Since κ \kappa κ is C 1 C^{1} C 1 and vanishes on this set, its derivative also vanishes at every interior point of M ∖ U M\setminus U M ∖ U . At points of ∂ U \partial U ∂ U , the identity κ H = n λ \kappa_{H}=n\lambda κ H = nλ and continuity give the same conclusion. It follows from (2.7) that div V = 0 \operatorname{div}V=0 div V = 0 on M ∖ U M\setminus U M ∖ U . By the definition of Y Y Y , we likewise have Y = 0 Y=0 Y = 0 on M ∖ U M\setminus U M ∖ U .
Hence div V = Y = 0 \operatorname{div} V = Y = 0 div V = Y = 0 there. On U U U , we have H > c 0 > 0 H > c_0 > 0 H > c 0 > 0 , G > 0 G > 0 G > 0 , and λ = 2 s n H G s − 1 > 0 \lambda= \frac{2s}{n}HG^{s-1} > 0 λ = n 2 s H G s − 1 > 0 . The Cauchy–Schwarz inequality for the principal curvatures implies ∣ A ∣ 2 ≥ n H 2 |A|^2 \ge nH^2 ∣ A ∣ 2 ≥ n H 2 . Combining this estimate with the assumed lower bound Ric ( N , N ) ≥ − n K 0 \operatorname{Ric}(N,N) \ge-nK_0 Ric ( N , N ) ≥ − n K 0 , we obtain
∣ A ∣ 2 + Ric ( N , N ) ≥ n ( H 2 − K 0 ) . |A|^2+\operatorname{Ric}(N,N)\ge n(H^2-K_0). ∣ A ∣ 2 + Ric ( N , N ) ≥ n ( H 2 − K 0 ) .
Since λ ≥ 0 \lambda\ge0 λ ≥ 0 , identity (8) yields
div V ≥ n λ ( H 2 − K 0 ) − n H κ = 2 s H G s − 1 ( H 2 − K 0 ) − n H G s . \begin{aligned}
\operatorname{div} V &\ge n\lambda(H^2-K_0)-nH\kappa\\
&=2sHG^{s-1}(H^2-K_0)-nHG^s.
\end{aligned} div V ≥ nλ ( H 2 − K 0 ) − n H κ = 2 sH G s − 1 ( H 2 − K 0 ) − n H G s .
On U U U , the definition of G G G gives G = H 2 − c 0 2 G=H^2-c_0^2 G = H 2 − c 0 2 . Recalling that δ 0 2 = c 0 2 − K 0 \delta_0^2=c_0^2-K_0 δ 0 2 = c 0 2 − K 0 , we have
H 2 − K 0 = G + δ 0 2 . H^2-K_0=G+\delta_0^2. H 2 − K 0 = G + δ 0 2 .
Substituting this identity into the preceding estimate, we find
div V ≥ 2 s H G s − 1 ( G + δ 0 2 ) − n H G s = ( 2 s − n ) H G s + 2 s δ 0 2 H G s − 1 = Y . \begin{aligned}
\operatorname{div} V &\ge2sHG^{s-1}(G+\delta_0^2)-nHG^s \\
&=(2s-n)HG^s+2s\delta_0^2HG^{s-1} \\
&=Y.
\end{aligned} div V ≥ 2 sH G s − 1 ( G + δ 0 2 ) − n H G s = ( 2 s − n ) H G s + 2 s δ 0 2 H G s − 1 = Y .
Finally, since 2 s > n 2s>n 2 s > n , δ 0 2 > 0 \delta_0^2>0 δ 0 2 > 0 , and H , G > 0 H,G>0 H , G > 0 on U U U , the definition of Y Y Y implies Y ≥ 0 Y\ge0 Y ≥ 0 on U U U . Together with Y = 0 Y=0 Y = 0 on M ∖ U M\setminus U M ∖ U , this proves that div V ≥ Y ≥ 0 \operatorname{div} V\ge Y\ge0 div V ≥ Y ≥ 0 on M M M . □
Proof of Theorem A
Throughout this section, M n + 1 M^{n+1} M n + 1 is complete with sec M ≥ − K 0 \sec_M\ge-K_0 sec M ≥ − K 0 , F \mathcal{F} F is a codimension one CMC foliation of M M M , x 0 ∈ M x_0\in M x 0 ∈ M is fixed and r : = dist M ( x 0 , ⋅ ) r:=\operatorname{dist}_M(x_0,\cdot) r := dist M ( x 0 , ⋅ ) . If v ∈ T x M v\in T_xM v ∈ T x M is a unit vector and { v , e 1 , … , e n } \{v,e_1,\ldots,e_n\} { v , e 1 , … , e n } is an orthonormal basis, then Ric ( v , v ) = ∑ i sec ( v , e i ) ≥ − n K 0 \operatorname{Ric}(v,v)=\sum_i\sec(v,e_i)\ge-nK_0 Ric ( v , v ) = ∑ i sec ( v , e i ) ≥ − n K 0 ; thus
Ric M ≥ − n K 0 (9) \operatorname{Ric}_M\ge-nK_0 \tag*{(9)} Ric M ≥ − n K 0 (9)
which is what Lemma 2.14 and the volume comparison below require. We fix R ∗ > 0 R_* > 0 R ∗ > 0 , put C 1 : = ct K 0 ( R ∗ ) C_1:=\operatorname{ct}_{K_0}(R_*) C 1 := ct K 0 ( R ∗ ) and R ♭ : = R ∗ + 1 R_\flat:=R_*+1 R ♭ := R ∗ + 1 , and let c 0 , s , G , δ 0 , F , Y , κ , λ , V c_0,s,G,\delta_0,F,Y,\kappa,\lambda,V c 0 , s , G , δ 0 , F , Y , κ , λ , V be as in Subsection 2.4.
Cutoff functions and volume
Let ψ ∈ C ∞ ( R ) \psi\in C^\infty(\mathbb{R}) ψ ∈ C ∞ ( R ) be a nonincreasing function such that ψ ( t ) = 1 \psi(t)=1 ψ ( t ) = 1 for t ≤ 0 t\le0 t ≤ 0 , and ψ ( t ) = 0 \psi(t)=0 ψ ( t ) = 0 for t ≥ 1 t\ge1 t ≥ 1 . Set
C ψ : = sup t ∈ R ∣ ψ ′ ( t ) ∣ , C ψ ′ : = sup t ∈ R ∣ ψ ′ ′ ( t ) ∣ . C_\psi:=\sup_{t\in\mathbb{R}}|\psi'(t)|,\qquad C'_\psi:=\sup_{t\in\mathbb{R}}|\psi''(t)|. C ψ := t ∈ R sup ∣ ψ ′ ( t ) ∣ , C ψ ′ := t ∈ R sup ∣ ψ ′′ ( t ) ∣.
Since ψ ( 0 ) = 1 \psi(0)=1 ψ ( 0 ) = 1 and ψ ( 1 ) = 0 \psi(1)=0 ψ ( 1 ) = 0 , the mean value theorem implies that there exists ξ ∈ ( 0 , 1 ) \xi\in(0,1) ξ ∈ ( 0 , 1 ) such that ∣ ψ ′ ( ξ ) ∣ = 1 |\psi'(\xi)|=1 ∣ ψ ′ ( ξ ) ∣ = 1 . Consequently, C ψ ≥ 1 C_\psi\ge1 C ψ ≥ 1 . Moreover, for every τ > 0 \tau>0 τ > 0 , one can choose such a smooth nonincreasing cutoff function ψ \psi ψ with C ψ ≤ 1 + τ C_\psi\le1+\tau C ψ ≤ 1 + τ .
For ε ∈ ( 0 , 1 ] \varepsilon\in(0,1] ε ∈ ( 0 , 1 ] and R ≥ R b R \ge R_b R ≥ R b , define
χ ( x ) : = ψ ( r ( x ) − R ε R ) , A R : = { x ∈ M : R ≤ r ( x ) ≤ ( 1 + ε ) R } . \chi(x):=\psi\left(\frac{r(x)-R}{\varepsilon R}\right), \qquad A_R:=\{x\in M:R\le r(x)\le(1+\varepsilon)R\}. χ ( x ) := ψ ( εR r ( x ) − R ) , A R := { x ∈ M : R ≤ r ( x ) ≤ ( 1 + ε ) R } .
Since r r r is 1 1 1 -Lipschitz and ψ \psi ψ is smooth with bounded derivative, χ \chi χ is Lipschitz. The monotonicity and prescribed values of ψ \psi ψ imply
0 ≤ χ ≤ 1 , χ = 1 on { r ≤ R } , χ = 0 on { r ≥ ( 1 + ε ) R } . 0\le\chi\le1,\qquad\chi=1\text{ on }\{r\le R\},\qquad\chi=0\text{ on }\{r\ge(1+\varepsilon)R\}. 0 ≤ χ ≤ 1 , χ = 1 on { r ≤ R } , χ = 0 on { r ≥ ( 1 + ε ) R } .
In particular, supp χ ⊂ { r ≤ ( 1 + ε ) R } \operatorname{supp}\chi\subset\{r\le(1+\varepsilon)R\} supp χ ⊂ { r ≤ ( 1 + ε ) R } . The chain rule for Lipschitz functions yields, almost everywhere,
∇ χ = 1 ε R ψ ′ ( r − R ε R ) ∇ r . \nabla\chi=\frac{1}{\varepsilon R}\psi'\left(\frac{r-R}{\varepsilon R}\right)\nabla r. ∇ χ = εR 1 ψ ′ ( εR r − R ) ∇ r .
Since ∣ ∇ r ∣ ≤ 1 |\nabla r|\le1 ∣∇ r ∣ ≤ 1 almost everywhere, we obtain
∣ ∇ χ ∣ ≤ C ψ ε R almost everywhere on M . |\nabla\chi|\le\frac{C_\psi}{\varepsilon R}\quad\text{almost everywhere on }M. ∣∇ χ ∣ ≤ εR C ψ almost everywhere on M .
Furthermore, ψ ′ = 0 \psi'=0 ψ ′ = 0 on ( − ∞ , 0 ) ∪ ( 1 , ∞ ) (-\infty,0)\cup(1,\infty) ( − ∞ , 0 ) ∪ ( 1 , ∞ ) , so ∇ χ = 0 \nabla\chi=0 ∇ χ = 0 almost everywhere outside A R A_R A R . Define
C R ′ : = n C 1 + C ψ ′ C ψ ε R . (10) C'_R:=nC_1+\frac{C'_\psi}{C_\psi\varepsilon R}. \tag*{(10)} C R ′ := n C 1 + C ψ εR C ψ ′ . (10)
Since C ψ ≥ 1 C_\psi\ge1 C ψ ≥ 1 and C ψ ′ ≥ 0 C'_\psi\ge0 C ψ ′ ≥ 0 are independent of R R R , the function R ↦ C R ′ R\mapsto C'_R R ↦ C R ′ is nonincreasing and lim R → ∞ C R ′ = n C 1 \lim_{R\to\infty}C'_R=nC_1 lim R → ∞ C R ′ = n C 1 .
We next establish a volume bound. By the Ricci curvature lower bound in (9) , the Bishop-Gromov volume comparison theorem (see [14 ] ) gives
vol B t ( x 0 ) ≤ ∣ S n ∣ ∫ 0 t ( sinh ( K 0 s ) K 0 ) n d s . \operatorname{vol} B_t(x_0)\le|S^n|\int_0^t\left(\frac{\sinh(\sqrt{K_0}s)}{\sqrt{K_0}}\right)^n\,ds. vol B t ( x 0 ) ≤ ∣ S n ∣ ∫ 0 t ( K 0 sinh ( K 0 s ) ) n d s .
where the integrand is interpreted as s n s^n s n when K 0 = 0 K_0=0 K 0 = 0 . Using sinh x ≤ x e x \sinh x\le xe^x sinh x ≤ x e x for x ≥ 0 x\ge0 x ≥ 0 , we find
( sinh ( K 0 s ) K 0 ) n ≤ s n e n K 0 s . \left(\frac{\sinh(\sqrt{K_0}s)}{\sqrt{K_0}}\right)^n\le s^n e^{n\sqrt{K_0}s}. ( K 0 sinh ( K 0 s ) ) n ≤ s n e n K 0 s .
Hence
vol B t ( x 0 ) ≤ ∣ S n ∣ ∫ 0 t s n e n K 0 s d s ≤ ∣ S n ∣ n + 1 t n + 1 e n K 0 t . \operatorname{vol} B_t(x_0)\le|S^n|\int_0^t s^n e^{n\sqrt{K_0}s}\,ds
\le\frac{|S^n|}{n+1}t^{n+1}e^{n\sqrt{K_0}t}. vol B t ( x 0 ) ≤ ∣ S n ∣ ∫ 0 t s n e n K 0 s d s ≤ n + 1 ∣ S n ∣ t n + 1 e n K 0 t .
Writing ω n : = ∣ S n ∣ n + 1 \omega_n:=\frac{|S^n|}{n+1} ω n := n + 1 ∣ S n ∣ , we conclude that
vol B t ( x 0 ) ≤ ω n t n + 1 e n K 0 t . \operatorname{vol} B_t(x_0)\le\omega_n t^{n+1}e^{n\sqrt{K_0}t}. vol B t ( x 0 ) ≤ ω n t n + 1 e n K 0 t .
Since { r ≤ t } ⊂ B t ′ ( x 0 ) \{r\le t\}\subset B_{t'}(x_0) { r ≤ t } ⊂ B t ′ ( x 0 ) for every t ′ > t t'>t t ′ > t , letting t ′ ↓ t t'\downarrow t t ′ ↓ t in the preceding estimate gives
vol { r ≤ ( 1 + ε ) R } ≤ V ( R ) , (11) \operatorname{vol}\{r\le(1+\varepsilon)R\}\le\mathcal{V}(R), \tag*{(11)} vol { r ≤ ( 1 + ε ) R } ≤ V ( R ) , (11)
where V ( R ) : = ω n ( ( 1 + ε ) R ) n + 1 e n K 0 ( 1 + ε ) R \mathcal{V}(R):=\omega_n\bigl((1+\varepsilon)R\bigr)^{n+1}e^{n\sqrt{K_0}(1+\varepsilon)R} V ( R ) := ω n ( ( 1 + ε ) R ) n + 1 e n K 0 ( 1 + ε ) R .
Finally, define
I j : = ∫ M χ j Y ( j ≥ 1 ) , I 0 : = ∫ { r ≤ ( 1 + ε ) R } Y , J ( t ) : = ∫ { r ≤ t } Y . I_j:=\int_M \chi^jY\quad(j\ge1),\qquad I_0:=\int_{\{r\le(1+\varepsilon)R\}}Y,\qquad J(t):=\int_{\{r\le t\}}Y. I j := ∫ M χ j Y ( j ≥ 1 ) , I 0 := ∫ { r ≤ ( 1 + ε ) R } Y , J ( t ) := ∫ { r ≤ t } Y .
Because Y Y Y is continuous and the closed metric balls are compact by the Hopf-Rinow theorem, these integrals are finite. Moreover, Y ≥ 0 Y\ge0 Y ≥ 0 by Lemma 2.14, and 0 ≤ χ ≤ 1 0\le\chi\le1 0 ≤ χ ≤ 1 . Thus, for every integer j ≥ 1 j\ge1 j ≥ 1 ,
0 ≤ χ j ≤ χ j − 1 ≤ 1 , 0\le\chi^j\le\chi^{j-1}\le1, 0 ≤ χ j ≤ χ j − 1 ≤ 1 ,
where for j = 1 j=1 j = 1 we interpret χ 0 = 1 \chi^0=1 χ 0 = 1 on { r ≤ ( 1 + ε ) R } \{r\le(1+\varepsilon)R\} { r ≤ ( 1 + ε ) R } in the integral defining I 0 I_0 I 0 . Since χ = 1 \chi=1 χ = 1 on { r ≤ R } \{r\le R\} { r ≤ R } and supp χ ⊂ { r ≤ ( 1 + ε ) R } \operatorname{supp}\chi\subset\{r\le(1+\varepsilon)R\} supp χ ⊂ { r ≤ ( 1 + ε ) R } , we obtain
J ( R ) = ∫ { r ≤ R } Y ≤ ∫ M χ j Y = I j ≤ I j − 1 ≤ ∫ { r ≤ ( 1 + ε ) R } Y = I 0 = J ( ( 1 + ε ) R ) . \begin{aligned}
J(R)&=\int_{\{r\le R\}}Y\\
&\le\int_M\chi^jY\\
&=I_j\\
&\le I_{j-1}\\
&\le\int_{\{r\le(1+\varepsilon)R\}}Y\\
&=I_0\\
&=J((1+\varepsilon)R).
\end{aligned} J ( R ) = ∫ { r ≤ R } Y ≤ ∫ M χ j Y = I j ≤ I j − 1 ≤ ∫ { r ≤ ( 1 + ε ) R } Y = I 0 = J (( 1 + ε ) R ) .
Therefore,
J ( R ) ≤ I j ≤ I j − 1 ≤ I 0 = J ( ( 1 + ε ) R ) , j ≥ 1. (12) J(R)\le I_j\le I_{j-1}\le I_0=J((1+\varepsilon)R),\qquad j\ge1. \tag*{(12)} J ( R ) ≤ I j ≤ I j − 1 ≤ I 0 = J (( 1 + ε ) R ) , j ≥ 1. (12)
The flow inequality and its iteration
Proposition 3.1 (flow inequality). For R ≥ R ♭ R\ge R_{\flat} R ≥ R ♭ and every integer j ≥ 1 j\ge1 j ≥ 1 ,
I j ≤ j C ψ ε R ∫ A R χ j − 1 B R , B R : = G s + 2 s H 2 G s − 1 + 2 s n C R ′ H G s − 1 . (13) I_j\le\frac{jC_{\psi}}{\varepsilon R}\int_{A_R}\chi^{j-1}\mathcal{B}_R,\qquad\mathcal{B}_R:=G^s+2sH^2G^{s-1}+\frac{2s}{n}C'_RH G^{s-1}. \tag*{(13)} I j ≤ εR j C ψ ∫ A R χ j − 1 B R , B R := G s + 2 s H 2 G s − 1 + n 2 s C R ′ H G s − 1 . (13)
Proof. Fix R ≥ R ♭ R\ge R_{\flat} R ≥ R ♭ and an integer j ≥ 1 j\ge1 j ≥ 1 . Recall that
V = κ N + λ T , κ = G s , λ = 2 s n H G s − 1 , V=\kappa N+\lambda T,\qquad\kappa=G^s,\qquad\lambda=\frac{2s}{n}HG^{s-1}, V = κ N + λ T , κ = G s , λ = n 2 s H G s − 1 ,
and that div V ≥ Y ≥ 0 \operatorname{div}V\ge Y\ge0 div V ≥ Y ≥ 0 by Lemma 2.14. Since χ \chi χ is Lipschitz and compactly supported, the vector field χ j V \chi^jV χ j V is also Lipschitz and compactly supported. Lemma 2.8 therefore yields ∫ M div ( χ j V ) = 0 \int_M\operatorname{div}(\chi^jV)=0 ∫ M div ( χ j V ) = 0 . Using the product rule for the divergence, we obtain
∫ M χ j div V = − ∫ M ⟨ ∇ ( χ j ) , V ⟩ = − ∫ M κ N ( χ j ) − ∫ M λ ⟨ ∇ ( χ j ) , T ⟩ . \begin{aligned}
\int_M\chi^j\operatorname{div}V=-\int_M\langle\nabla(\chi^j),V\rangle\\
&=-\int_M\kappa N(\chi^j)-\int_M\lambda\langle\nabla(\chi^j),T\rangle.
\end{aligned} ∫ M χ j div V = − ∫ M ⟨ ∇ ( χ j ) , V ⟩ = − ∫ M κ N ( χ j ) − ∫ M λ ⟨ ∇ ( χ j ) , T ⟩ .
We now apply Proposition 2.12 to the function Φ ( t ) : = ψ ( t − R ε R ) j \Phi(t):=\psi\left(\frac{t-R}{\varepsilon R}\right)^j Φ ( t ) := ψ ( εR t − R ) j . Since ψ \psi ψ is non-increasing, equals 1 on ( − ∞ , 0 ] (-\infty,0] ( − ∞ , 0 ] , and vanishes on [ 1 , ∞ ) [1,\infty) [ 1 , ∞ ) , the function Φ \Phi Φ is smooth and nonincreasing, with
Φ ≡ 1 on ( − ∞ , R ] , Φ ≡ 0 on [ ( 1 + ε ) R , ∞ ) . \Phi\equiv1\quad\text{on }(-\infty,R],\qquad\Phi\equiv0\quad\text{on }[(1+\varepsilon)R,\infty). Φ ≡ 1 on ( − ∞ , R ] , Φ ≡ 0 on [( 1 + ε ) R , ∞ ) .
Moreover, R > R ∗ R>R_\ast R > R ∗ by the choice of R b R_b R b . Thus all the hypotheses of Proposition 2.12 are satisfied, with R 1 = R R_1=R R 1 = R , R 2 = ( 1 + ε ) R R_2=(1+\varepsilon)R R 2 = ( 1 + ε ) R , and Φ − = 1 \Phi_-=1 Φ − = 1 . Since χ j = Φ ∘ r \chi^j=\Phi\circ r χ j = Φ ∘ r , that proposition gives
− ∫ M λ ⟨ ∇ ( χ j ) , T ⟩ ≤ − n ∫ M H λ N ( χ j ) − ∫ M λ [ Φ ′ ′ ( r ) ∣ ∇ L r ∣ 2 + n C 1 Φ ′ ( r ) ] . -\int_M \lambda\langle\nabla(\chi^j),T\rangle\leq-n\int_M H\lambda N(\chi^j)-\int_M\lambda\left[\Phi''(r)|\nabla_Lr|^2+nC_1\Phi'(r)\right]. − ∫ M λ ⟨ ∇ ( χ j ) , T ⟩ ≤ − n ∫ M H λ N ( χ j ) − ∫ M λ [ Φ ′′ ( r ) ∣ ∇ L r ∣ 2 + n C 1 Φ ′ ( r ) ] .
Consequently,
∫ M χ j div V ≤ − ∫ M ( κ + n H λ ) N ( χ j ) − ∫ M λ Φ ′ ′ ( r ) ∣ ∇ L r ∣ 2 − n C 1 ∫ M λ Φ ′ ( r ) . (14) \int_M\chi^j\operatorname{div}V\leq-\int_M(\kappa+nH\lambda)N(\chi^j)-\int_M\lambda\Phi''(r)|\nabla_Lr|^2-nC_1\int_M\lambda\Phi'(r). \tag*{(14)} ∫ M χ j div V ≤ − ∫ M ( κ + n H λ ) N ( χ j ) − ∫ M λ Φ ′′ ( r ) ∣ ∇ L r ∣ 2 − n C 1 ∫ M λ Φ ′ ( r ) . (14)
Moreover, differentiating Φ \Phi Φ gives
Φ ′ ( t ) = j ε R ψ ( t − R ε R ) j − 1 ψ ′ ( t − R ε R ) \Phi'(t)=\frac{j}{\varepsilon R}\psi\left(\frac{t-R}{\varepsilon R}\right)^{j-1}\psi'\left(\frac{t-R}{\varepsilon R}\right) Φ ′ ( t ) = εR j ψ ( εR t − R ) j − 1 ψ ′ ( εR t − R )
and
Φ ′ ′ ( t ) = j ( j − 1 ) ( ε R ) 2 ψ ( t − R ε R ) j − 2 ψ ′ ( t − R ε R ) 2 + j ( ε R ) 2 ψ ( t − R ε R ) j − 1 ψ ′ ′ ( t − R ε R ) , \Phi''(t)=\frac{j(j-1)}{(\varepsilon R)^2}\psi\left(\frac{t-R}{\varepsilon R}\right)^{j-2}\psi'\left(\frac{t-R}{\varepsilon R}\right)^2+\frac{j}{(\varepsilon R)^2}\psi\left(\frac{t-R}{\varepsilon R}\right)^{j-1}\psi''\left(\frac{t-R}{\varepsilon R}\right), Φ ′′ ( t ) = ( εR ) 2 j ( j − 1 ) ψ ( εR t − R ) j − 2 ψ ′ ( εR t − R ) 2 + ( εR ) 2 j ψ ( εR t − R ) j − 1 ψ ′′ ( εR t − R ) ,
where the first term is omitted when j = 1 j=1 j = 1 .
Because ψ ′ ≤ 0 \psi'\leq0 ψ ′ ≤ 0 and ∣ ψ ′ ∣ ≤ C ψ |\psi'|\leq C_\psi ∣ ψ ′ ∣ ≤ C ψ , we have
∣ Φ ′ ( r ) ∣ ≤ j C ψ ε R χ j − 1 . |\Phi'(r)|\leq\frac{jC_\psi}{\varepsilon R}\chi^{j-1}. ∣ Φ ′ ( r ) ∣ ≤ εR j C ψ χ j − 1 .
The same formula shows that Φ ′ ( r ) = 0 \Phi'(r)=0 Φ ′ ( r ) = 0 almost everywhere outside A R A_R A R . Since ∣ ∇ r ∣ ≤ 1 |\nabla r|\leq1 ∣∇ r ∣ ≤ 1 almost everywhere,
∣ N ( χ j ) ∣ = j χ j − 1 ∣ N ( χ ) ∣ ≤ j C ψ ε R χ j − 1 almost everywhere on A R , |N(\chi^j)|=j\chi^{j-1}|N(\chi)|\leq\frac{jC_\psi}{\varepsilon R}\chi^{j-1}\quad\text{almost everywhere on }A_R, ∣ N ( χ j ) ∣ = j χ j − 1 ∣ N ( χ ) ∣ ≤ εR j C ψ χ j − 1 almost everywhere on A R ,
and N ( χ j ) = 0 N(\chi^j)=0 N ( χ j ) = 0 almost everywhere outside A R A_R A R .
Next, the first term in the formula for Φ ′ ′ \Phi'' Φ ′′ is nonnegative, because j ≥ 2 j\geq2 j ≥ 2 whenever this term is present. It follows that
− Φ ′ ′ ( r ) ≤ j C ψ ′ ( ε R ) 2 χ j − 1 . -\Phi''(r)\leq\frac{jC'_\psi}{(\varepsilon R)^2}\chi^{j-1}. − Φ ′′ ( r ) ≤ ( εR ) 2 j C ψ ′ χ j − 1 .
Since ∣ ∇ L r ∣ ≤ ∣ ∇ r ∣ ≤ 1 |\nabla_Lr|\leq|\nabla r|\leq1 ∣ ∇ L r ∣ ≤ ∣∇ r ∣ ≤ 1 almost everywhere, we obtain
− Φ ′ ′ ( r ) ∣ ∇ L r ∣ 2 ≤ j C ψ ′ ( ε R ) 2 χ j − 1 almost everywhere on A R . -\Phi''(r)|\nabla_Lr|^2\leq\frac{jC'_\psi}{(\varepsilon R)^2}\chi^{j-1}\quad\text{almost everywhere on }A_R. − Φ ′′ ( r ) ∣ ∇ L r ∣ 2 ≤ ( εR ) 2 j C ψ ′ χ j − 1 almost everywhere on A R .
Furthermore, Φ ′ ′ ( r ) = 0 \Phi''(r)=0 Φ ′′ ( r ) = 0 almost everywhere outside A R A_R A R . Finally, since C 1 ≥ 0 C_1\geq0 C 1 ≥ 0 and Φ ′ ≤ 0 \Phi'\leq0 Φ ′ ≤ 0 ,
− n C 1 Φ ′ ( r ) ≤ j n C 1 C ψ ε R χ j − 1 , -nC_1\Phi'(r)\leq\frac{jnC_1C_\psi}{\varepsilon R}\chi^{j-1}, − n C 1 Φ ′ ( r ) ≤ εR j n C 1 C ψ χ j − 1 ,
and this term also vanishes almost everywhere outside A R A_R A R .
By the definitions of κ \kappa κ and λ \lambda λ ,
κ + n H λ = G s + 2 s H 2 G s − 1 ≥ 0. \kappa+nH\lambda=G^s+2sH^2G^{s-1}\ge0. κ + n H λ = G s + 2 s H 2 G s − 1 ≥ 0.
Using the preceding estimates in (14) , together with λ ≥ 0 \lambda\ge0 λ ≥ 0 , we conclude that
∫ M χ j div V ≤ j C ψ ε R ∫ A R χ j − 1 ( κ + n H λ ) + ∫ A R λ χ j − 1 ( j C ψ ′ ( ε R ) 2 + j n C 1 C ψ ε R ) = j C ψ ε R ∫ A R χ j − 1 [ G s + 2 s H 2 G s − 1 + λ ( n C 1 + C ψ ′ C ψ ε R ) ] . \begin{aligned}
\int_M\chi^j\operatorname{div}V\le\frac{jC_\psi}{\varepsilon R}\int_{A_R}\chi^{j-1}(\kappa+nH\lambda)+\int_{A_R}\lambda\chi^{j-1}\left(\frac{jC'_\psi}{(\varepsilon R)^2}+\frac{jnC_1C_\psi}{\varepsilon R}\right) \\
&=\frac{jC_\psi}{\varepsilon R}\int_{A_R}\chi^{j-1}\left[G^s+2sH^2G^{s-1}+\lambda\left(nC_1+\frac{C'_\psi}{C_\psi\varepsilon R}\right)\right].
\end{aligned} ∫ M χ j div V ≤ εR j C ψ ∫ A R χ j − 1 ( κ + n H λ ) + ∫ A R λ χ j − 1 ( ( εR ) 2 j C ψ ′ + εR j n C 1 C ψ ) = εR j C ψ ∫ A R χ j − 1 [ G s + 2 s H 2 G s − 1 + λ ( n C 1 + C ψ εR C ψ ′ ) ] .
Recalling the definition C R ′ = n C 1 + C ψ ′ C ψ ε R C'_R=nC_1+\frac{C'_\psi}{C_\psi\varepsilon R} C R ′ = n C 1 + C ψ εR C ψ ′ and substituting λ = 2 s n H G s − 1 \lambda=\frac{2s}{n}HG^{s-1} λ = n 2 s H G s − 1 , we obtain
∫ M χ j div V ≤ j C ψ ε R ∫ A R χ j − 1 [ G s + 2 s H 2 G s − 1 + 2 s n C R ′ H G s − 1 ] . \int_M\chi^j\operatorname{div}V\le\frac{jC_\psi}{\varepsilon R}\int_{A_R}\chi^{j-1}\left[G^s+2sH^2G^{s-1}+\frac{2s}{n}C'_RHG^{s-1}\right]. ∫ M χ j div V ≤ εR j C ψ ∫ A R χ j − 1 [ G s + 2 s H 2 G s − 1 + n 2 s C R ′ H G s − 1 ] .
The expression in brackets is precisely B R B_R B R . Finally, since div V ≥ Y \operatorname{div}V\ge Y div V ≥ Y and χ j ≥ 0 \chi^j\ge0 χ j ≥ 0 ,
I j = ∫ M χ j Y ≤ ∫ M χ j div V . I_j=\int_M\chi^jY\le\int_M\chi^j\operatorname{div}V. I j = ∫ M χ j Y ≤ ∫ M χ j div V .
Combining the last two inequalities proves
I j ≤ j C ψ ε R ∫ A R χ j − 1 B R , I_j\le\frac{jC_\psi}{\varepsilon R}\int_{A_R}\chi^{j-1}B_R, I j ≤ εR j C ψ ∫ A R χ j − 1 B R ,
as claimed. □
Lemma 3.2 (top-order bound). On M M M , B R ≤ θ R ′ Y B_R\le\theta'_RY B R ≤ θ R ′ Y , where
θ R ′ = θ ′ ( s , c 0 , C R ′ ) : = max { 2 s + 1 ( 2 s − n ) c 0 , ( 2 s + 1 ) c 0 + 2 s n C R ′ 2 s δ 0 2 } . (15) \theta'_R=\theta'(s,c_0,C'_R):=\max\left\{\frac{2s+1}{(2s-n)c_0},\frac{(2s+1)c_0+\frac{2s}{n}C'_R}{2s\delta_0^2}\right\}. \tag*{(15)} θ R ′ = θ ′ ( s , c 0 , C R ′ ) := max { ( 2 s − n ) c 0 2 s + 1 , 2 s δ 0 2 ( 2 s + 1 ) c 0 + n 2 s C R ′ } . (15) Proof. Recall that B R = G s + 2 s H 2 G s − 1 + 2 s n C R ′ H G s − 1 B_R=G^s+2sH^2G^{s-1}+\frac{2s}{n}C'_RHG^{s-1} B R = G s + 2 s H 2 G s − 1 + n 2 s C R ′ H G s − 1 , whereas Y = ( 2 s − n ) H G s + 2 s δ 0 2 H G s − 1 Y=(2s-n)HG^s+2s\delta_0^2HG^{s-1} Y = ( 2 s − n ) H G s + 2 s δ 0 2 H G s − 1 . Both B R B_R B R and Y Y Y vanish on M ∖ U M\setminus U M ∖ U , where U = { H > c 0 } U=\{H>c_0\} U = { H > c 0 } . Thus, it suffices to prove the desired inequality on U U U .
Fix x ∈ U x\in U x ∈ U . Since H ( x ) > c 0 > 0 H(x)>c_0>0 H ( x ) > c 0 > 0 , we have H G s − 1 > 0 HG^{s-1}>0 H G s − 1 > 0 . Dividing the inequality B R ≤ θ R ′ Y B_R\le\theta'_RY B R ≤ θ R ′ Y by this positive quantity, we see that it is equivalent to
G H + 2 s H + 2 s n C R ′ ≤ θ R ′ ( ( 2 s − n ) G + 2 s δ 0 2 ) . \frac{G}{H}+2sH+\frac{2s}{n}C'_R\le\theta'_R\left((2s-n)G+2s\delta_0^2\right). H G + 2 sH + n 2 s C R ′ ≤ θ R ′ ( ( 2 s − n ) G + 2 s δ 0 2 ) . On U U U , the identity G = H 2 − c 0 2 G=H^2-c_0^2 G = H 2 − c 0 2 gives 0 < G < H 2 0<G<H^2 0 < G < H 2 . Consequently,
G H ≤ H . \frac{G}{H}\le H. H G ≤ H . Moreover, since H > 0 H>0 H > 0 and G = H 2 − c 0 2 G=H^2-c_0^2 G = H 2 − c 0 2 , we have
H = G + c 0 2 H = G H + c 0 2 H ≤ G c 0 + c 0 , H=\frac{G+c_0^2}{H}=\frac{G}{H}+\frac{c_0^2}{H}\le\frac{G}{c_0}+c_0, H = H G + c 0 2 = H G + H c 0 2 ≤ c 0 G + c 0 ,
where we used H ≥ c 0 H \ge c_0 H ≥ c 0 in the last inequality. It follows that
G H + 2 s H + 2 s n C R ′ ≤ ( 2 s + 1 ) H + 2 s n C R ′ ≤ 2 s + 1 c 0 G + ( 2 s + 1 ) c 0 + 2 s n C R ′ . \frac{G}{H}+2sH+\frac{2s}{n}C'_R \le(2s+1)H+\frac{2s}{n}C'_R
\le\frac{2s+1}{c_0}G+(2s+1)c_0+\frac{2s}{n}C'_R. H G + 2 sH + n 2 s C R ′ ≤ ( 2 s + 1 ) H + n 2 s C R ′ ≤ c 0 2 s + 1 G + ( 2 s + 1 ) c 0 + n 2 s C R ′ .
By the definition of θ R ′ \theta'_R θ R ′ ,
θ R ′ ≥ 2 s + 1 ( 2 s − n ) c 0 and θ R ′ ≥ ( 2 s + 1 ) c 0 + 2 s n C R ′ 2 s δ 0 2 . \theta'_R \ge\frac{2s+1}{(2s-n)c_0}
\qquad\text{and}\qquad
\theta'_R \ge\frac{(2s+1)c_0+\frac{2s}{n}C'_R}{2s\delta_0^2}. θ R ′ ≥ ( 2 s − n ) c 0 2 s + 1 and θ R ′ ≥ 2 s δ 0 2 ( 2 s + 1 ) c 0 + n 2 s C R ′ .
Since 2 s > n 2s>n 2 s > n , both ( 2 s − n ) G (2s-n)G ( 2 s − n ) G and 2 s δ 0 2 2s\delta_0^2 2 s δ 0 2 are nonnegative. Hence
2 s + 1 c 0 G ≤ θ R ′ ( 2 s − n ) G \frac{2s+1}{c_0}G \le\theta'_R(2s-n)G c 0 2 s + 1 G ≤ θ R ′ ( 2 s − n ) G
and
( 2 s + 1 ) c 0 + 2 s n C R ′ ≤ θ R ′ 2 s δ 0 2 . (2s+1)c_0+\frac{2s}{n}C'_R \le\theta'_R 2s\delta_0^2. ( 2 s + 1 ) c 0 + n 2 s C R ′ ≤ θ R ′ 2 s δ 0 2 .
Adding these inequalities yields
G H + 2 s H + 2 s n C R ′ ≤ θ R ′ ( ( 2 s − n ) G + 2 s δ 0 2 ) . \frac{G}{H}+2sH+\frac{2s}{n}C'_R \le\theta'_R\bigl((2s-n)G+2s\delta_0^2\bigr). H G + 2 sH + n 2 s C R ′ ≤ θ R ′ ( ( 2 s − n ) G + 2 s δ 0 2 ) .
Multiplying by H G s − 1 > 0 HG^{s-1}>0 H G s − 1 > 0 , we conclude that B R ≤ θ R ′ Y \mathcal{B}_R \le\theta'_R Y B R ≤ θ R ′ Y on U U U . Together with the vanishing of both sides on M ∖ U M\setminus U M ∖ U , this proves the lemma. □
Corollary 3.3. For R ≥ R 0 ≥ R ♭ R\ge R_0\ge R_\flat R ≥ R 0 ≥ R ♭ and j ≥ 1 j\ge1 j ≥ 1 , I j ≤ j C ψ θ R 0 ′ ε R I j − 1 I_j\le\frac{jC\psi\theta'_{R_0}}{\varepsilon R}I_{j-1} I j ≤ εR j C ψ θ R 0 ′ I j − 1 .
Proof. By Proposition 3.1 and Lemma 3.2, we obtain
I j ≤ j C ψ θ R ′ ε R ∫ A R χ j − 1 Y . I_j\le\frac{jC\psi\theta'_R}{\varepsilon R}\int_{A_R}\chi^{j-1}Y. I j ≤ εR j C ψ θ R ′ ∫ A R χ j − 1 Y .
Since C R ′ C'_R C R ′ is nonincreasing in R R R and θ R ′ \theta'_R θ R ′ is nondecreasing as a function of C R ′ C'_R C R ′ , the condition R ≥ R 0 R\ge R_0 R ≥ R 0 implies θ R ′ ≤ θ R 0 ′ \theta'_R\le\theta'_{R_0} θ R ′ ≤ θ R 0 ′ . Therefore,
I j ≤ j C ψ θ R 0 ′ ε R ∫ A R χ j − 1 Y . I_j\le\frac{jC\psi\theta'_{R_0}}{\varepsilon R}\int_{A_R}\chi^{j-1}Y. I j ≤ εR j C ψ θ R 0 ′ ∫ A R χ j − 1 Y .
If j ≥ 2 j\ge2 j ≥ 2 , then 0 ≤ χ ≤ 1 0\le\chi\le1 0 ≤ χ ≤ 1 and Y ≥ 0 Y\ge0 Y ≥ 0 , so
∫ A R χ j − 1 Y ≤ ∫ M χ j − 1 Y = I j − 1 . \int_{A_R}\chi^{j-1}Y\le\int_M\chi^{j-1}Y=I_{j-1}. ∫ A R χ j − 1 Y ≤ ∫ M χ j − 1 Y = I j − 1 .
If j = 1 j=1 j = 1 , the definition of I 0 I_0 I 0 and the inclusion A R ⊂ { r ≤ ( 1 + ε ) R } A_R\subset\{r\le(1+\varepsilon)R\} A R ⊂ { r ≤ ( 1 + ε ) R } give
∫ A R Y ≤ ∫ { r ≤ ( 1 + ε ) R } Y = I 0 . \int_{A_R}Y\le\int_{\{r\le(1+\varepsilon)R\}}Y=I_0. ∫ A R Y ≤ ∫ { r ≤ ( 1 + ε ) R } Y = I 0 .
Combining these estimates yields
I j ≤ j C ψ θ R 0 ′ ε R I j − 1 I_j\le\frac{jC\psi\theta'_{R_0}}{\varepsilon R}I_{j-1} I j ≤ εR j C ψ θ R 0 ′ I j − 1
for every j ≥ 1 j\ge1 j ≥ 1 , as claimed. □
Lemma 3.4 (a priori bound). There is a constant C 2 C_2 C 2 , depending only on n n n , s s s , c 0 c_0 c 0 , K 0 K_0 K 0 , ψ \psi ψ , ε \varepsilon ε and R ∗ R_\ast R ∗ , such that J ( R ) ≤ C 2 V ( R ) J(R) \le C_2 V(R) J ( R ) ≤ C 2 V ( R ) for all R ≥ R b R \ge R_b R ≥ R b .
Proof. Set q : = 2 s + 1 q := 2s+1 q := 2 s + 1 and fix R ≥ R b R \ge R_b R ≥ R b . By Proposition 3.1, applied with j = q j=q j = q , we have
I q ≤ q C ψ ε R ∫ A R χ q − 1 B R . I_q \le\frac{qC_\psi}{\varepsilon R}\int_{A_R}\chi^{q-1}B_R. I q ≤ εR q C ψ ∫ A R χ q − 1 B R . Recall that B R = 0 B_R=0 B R = 0 on M ∖ U M\setminus U M ∖ U , where U = { H > c 0 } U=\{H>c_0\} U = { H > c 0 } . We split U = U 1 ∪ U 2 U=U_1\cup U_2 U = U 1 ∪ U 2 , where
U 1 : = { c 0 < H < 2 c 0 } , U 2 : = { H ≥ 2 c 0 } . U_1:=\{c_0<H<2c_0\},\qquad U_2:=\{H\ge2c_0\}. U 1 := { c 0 < H < 2 c 0 } , U 2 := { H ≥ 2 c 0 } . On U 1 U_1 U 1 , we have G < 3 c 0 2 G<3c_0^2 G < 3 c 0 2 and H < 2 c 0 H<2c_0 H < 2 c 0 . Moreover, C R ′ ≤ C R b ′ C'_R\le C'_{R_b} C R ′ ≤ C R b ′ . Consequently, there exists a constant B 0 > 0 B_0>0 B 0 > 0 , depending only on n n n , s s s , c 0 c_0 c 0 , K 0 K_0 K 0 , ψ \psi ψ , ε \varepsilon ε and R ∗ R_\ast R ∗ , such that
B R ≤ B 0 on U 1 . B_R\le B_0\qquad\text{on }U_1. B R ≤ B 0 on U 1 . On U 2 U_2 U 2 , the inequalities
G = H 2 − c 0 2 ≥ 3 4 H 2 , F = H G s ≤ H 2 s + 1 = H q G=H^2-c_0^2\ge\frac{3}{4}H^2,\qquad F=HG^s\le H^{2s+1}=H^q G = H 2 − c 0 2 ≥ 4 3 H 2 , F = H G s ≤ H 2 s + 1 = H q imply H ≥ F 1 / q H\ge F^{1/q} H ≥ F 1/ q . Hence
G s = F H ≤ F q − 1 q , G^s=\frac{F}{H}\le F^{\frac{q-1}{q}}, G s = H F ≤ F q q − 1 , and
H 2 G s − 1 = F H G ≤ 4 F 3 H ≤ 4 3 F q − 1 q . H^2G^{s-1}=F\frac{H}{G}\le\frac{4F}{3H}\le\frac{4}{3}F^{\frac{q-1}{q}}. H 2 G s − 1 = F G H ≤ 3 H 4 F ≤ 3 4 F q q − 1 . Similarly,
H G s − 1 = F G ≤ 4 F 3 H 2 ≤ 4 3 F q − 2 q . HG^{s-1}=\frac{F}{G}\le\frac{4F}{3H^2}\le\frac{4}{3}F^{\frac{q-2}{q}}. H G s − 1 = G F ≤ 3 H 2 4 F ≤ 3 4 F q q − 2 . It follows that
B R ≤ B 1 ( F q − 1 q + F q − 2 q ) on U 2 , B_R\le B_1\left(F^{\frac{q-1}{q}}+F^{\frac{q-2}{q}}\right)\qquad\text{on }U_2, B R ≤ B 1 ( F q q − 1 + F q q − 2 ) on U 2 , where B 1 : = 1 + 8 s 3 + 8 s 3 C R b ′ B_1:=1+\frac{8s}{3}+\frac{8s}{3}C'_{R_b} B 1 := 1 + 3 8 s + 3 8 s C R b ′ .
Define X : = χ q F X:=\chi^qF X := χ q F . Since 0 ≤ χ ≤ 1 0\le\chi\le1 0 ≤ χ ≤ 1 , we obtain
χ q − 1 F q − 1 q = X q − 1 q , χ q − 1 F q − 2 q ≤ X q − 2 q . \chi^{q-1}F^{\frac{q-1}{q}}=X^{\frac{q-1}{q}},\qquad\chi^{q-1}F^{\frac{q-2}{q}}\le X^{\frac{q-2}{q}}. χ q − 1 F q q − 1 = X q q − 1 , χ q − 1 F q q − 2 ≤ X q q − 2 . Combining the estimates on U 1 U_1 U 1 and U 2 U_2 U 2 , we conclude that
I q ≤ q C ψ B 0 ε R vol ( A R ) + m ∫ A R ( X q − 1 q + X q − 2 q ) , I_q\le\frac{qC_\psi B_0}{\varepsilon R}\operatorname{vol}(A_R)+m\int_{A_R}\left(X^{\frac{q-1}{q}}+X^{\frac{q-2}{q}}\right), I q ≤ εR q C ψ B 0 vol ( A R ) + m ∫ A R ( X q q − 1 + X q q − 2 ) , where m : = q C ψ B 1 ε R ≤ m 1 : = q C ψ B 1 ε R b m:=\frac{qC_\psi B_1}{\varepsilon R}\le m_1:=\frac{qC_\psi B_1}{\varepsilon R_b} m := εR q C ψ B 1 ≤ m 1 := ε R b q C ψ B 1 . For every 0 < ϑ < 1 0<\vartheta<1 0 < ϑ < 1 , Young’s inequality gives
m X ϑ ≤ η X + C ( η , ϑ ) m 1 1 − ϑ , mX^\vartheta\le\eta X+C(\eta,\vartheta)m^{\frac{1}{1-\vartheta}}, m X ϑ ≤ η X + C ( η , ϑ ) m 1 − ϑ 1 , for any η > 0 \eta>0 η > 0 , where C ( η , ϑ ) C(\eta,\vartheta) C ( η , ϑ ) depends only on η \eta η and ϑ \vartheta ϑ . Applying this inequality with ϑ = ( q − 1 ) / q \vartheta=(q-1)/q ϑ = ( q − 1 ) / q and ϑ = ( q − 2 ) / q \vartheta=(q-2)/q ϑ = ( q − 2 ) / q , we find
I q ≤ [ q C ψ B 0 ε R b + C η ( m 1 q + m 1 q / 2 ) ] vol ( A R ) + 2 η ∫ M X , I_q\le\left[\frac{qC_\psi B_0}{\varepsilon R_b}+C_\eta\left(m_1^q+m_1^{q/2}\right)\right]\operatorname{vol}(A_R)+2\eta\int_M X, I q ≤ [ ε R b q C ψ B 0 + C η ( m 1 q + m 1 q /2 ) ] vol ( A R ) + 2 η ∫ M X ,
where C η C_{\eta} C η depends only on η \eta η and q q q . Since Y = ( 2 s − n ) F + 2 s δ 0 2 H G s − 1 ≥ ( 2 s − n ) F Y=(2s-n)F+2s\delta_0^2HG^{s-1}\ge(2s-n)F Y = ( 2 s − n ) F + 2 s δ 0 2 H G s − 1 ≥ ( 2 s − n ) F and X = χ q F X=\chi^qF X = χ q F , we have
∫ M X ≤ I q 2 s − n . \int_M X\le\frac{I_q}{2s-n}. ∫ M X ≤ 2 s − n I q .
Choose η = ( 2 s − n ) / 4 > 0 \eta=(2s-n)/4>0 η = ( 2 s − n ) /4 > 0 . Then
I q ≤ C vol ( A R ) + 1 2 I q . I_q\le C\operatorname{vol}(A_R)+\frac{1}{2}I_q. I q ≤ C vol ( A R ) + 2 1 I q .
where C C C depends only on n n n , s s s , c 0 c_0 c 0 , K 0 K_0 K 0 , ψ \psi ψ , ε \varepsilon ε and R ∗ R_* R ∗ . Absorbing the last term into the left-hand side yields
I q ≤ 2 C vol ( A R ) ≤ 2 C vol { r ≤ ( 1 + ε ) R } . I_q\le2C\operatorname{vol}(A_R)\le2C\operatorname{vol}\{r\le(1+\varepsilon)R\}. I q ≤ 2 C vol ( A R ) ≤ 2 C vol { r ≤ ( 1 + ε ) R } .
Finally, since χ = 1 \chi=1 χ = 1 on { r ≤ R } \{r\le R\} { r ≤ R } and Y ≥ 0 Y\ge0 Y ≥ 0 ,
J ( R ) = ∫ { r ≤ R } Y ≤ ∫ M χ q Y = I q . J(R)=\int_{\{r\le R\}}Y\le\int_M\chi^qY=I_q. J ( R ) = ∫ { r ≤ R } Y ≤ ∫ M χ q Y = I q .
Using the volume estimate (3.3), we conclude that
J ( R ) ≤ C 2 V ( R ) , J(R)\le C_2V(R), J ( R ) ≤ C 2 V ( R ) ,
after enlarging the constant C 2 C_2 C 2 if necessary. This proves the lemma. □
Conclusion of the proof
For n ≥ 1 n\ge1 n ≥ 1 and K 0 , C 1 ≥ 0 K_0,C_1\ge0 K 0 , C 1 ≥ 0 , define
c ∗ ( n , K 0 ; C 1 ) : = 1 2 [ n K 0 + ( n 2 + 4 ) K 0 + 4 n K 0 C 1 ] (16) c_*(n,K_0;C_1):=\frac{1}{2}\left[n\sqrt{K_0}+\sqrt{(n^2+4)K_0+4n\sqrt{K_0}C_1}\right] \tag*{(16)} c ∗ ( n , K 0 ; C 1 ) := 2 1 [ n K 0 + ( n 2 + 4 ) K 0 + 4 n K 0 C 1 ] (16)
This is the nonnegative root of
c 2 − n K 0 c − K 0 − n K 0 C 1 = 0. c^2-n\sqrt{K_0}c-K_0-n\sqrt{K_0}C_1=0. c 2 − n K 0 c − K 0 − n K 0 C 1 = 0.
The function c ∗ c_* c ∗ is continuous and nondecreasing in C 1 C_1 C 1 . Moreover, since n ≥ 1 n\ge1 n ≥ 1 and C 1 ≥ 0 C_1\ge0 C 1 ≥ 0 , we have
c ∗ ( n , K 0 ; C 1 ) ≥ 1 2 ( n + n 2 + 4 ) K 0 ≥ K 0 . c_*(n,K_0;C_1)\ge\frac{1}{2}\left(n+\sqrt{n^2+4}\right)\sqrt{K_0}\ge\sqrt{K_0}. c ∗ ( n , K 0 ; C 1 ) ≥ 2 1 ( n + n 2 + 4 ) K 0 ≥ K 0 .
When K 0 = 0 K_0=0 K 0 = 0 , the defining formula gives
c ∗ ( n , 0 ; C 1 ) = 0. c_*(n,0;C_1)=0. c ∗ ( n , 0 ; C 1 ) = 0.
Finally, taking C 1 = K 0 C_1=\sqrt{K_0} C 1 = K 0 in (3.8), we obtain
c ∗ ( n , K 0 ; K 0 ) = 1 2 [ n K 0 + ( n 2 + 4 + 4 n ) K 0 ] = ( n + 1 ) K 0 . (17) \begin{aligned}
c_*(n,K_0;\sqrt{K_0})=\frac{1}{2}\left[n\sqrt{K_0}+\sqrt{(n^2+4+4n)K_0}\right] \\
&=(n+1)\sqrt{K_0}. \tag*{(17)}
\end{aligned} c ∗ ( n , K 0 ; K 0 ) = 2 1 [ n K 0 + ( n 2 + 4 + 4 n ) K 0 ] = ( n + 1 ) K 0 . (17)
Proposition 3.5. For every R ∗ > 0 R_*>0 R ∗ > 0 , ∣ H ∣ ≤ c ∗ ( n , K 0 ; ct K 0 ( R ∗ ) ) |H|\le c_*(n,K_0;\operatorname{ct}_{K_0}(R_*)) ∣ H ∣ ≤ c ∗ ( n , K 0 ; ct K 0 ( R ∗ )) on M M M .
Proof. We divide the proof into four steps. Fix R ∗ > 0 R_* > 0 R ∗ > 0 and set C 1 : = ct K 0 ( R ∗ ) C_1 := \operatorname{ct}_{K_0}(R_*) C 1 := ct K 0 ( R ∗ ) , c ∗ : = c ∗ ( n , K 0 ; C 1 ) c_* := c_*(n,K_0;C_1) c ∗ := c ∗ ( n , K 0 ; C 1 ) . We prove first that H ≤ c ∗ H \le c_* H ≤ c ∗ on M M M . The opposite inequality will follow by reversing the orientation of the foliation.
Step 1: Forced exponential growth. Fix R 0 ≥ R b R_0 \ge R_b R 0 ≥ R b and set θ : = C ψ θ R 0 ′ > 0 \theta:= C_{\psi}\theta'_{R_0} > 0 θ := C ψ θ R 0 ′ > 0 . For every R ≥ max { R 0 , θ ε } R \ge\max\left\{R_0,\frac{\theta}{\varepsilon}\right\} R ≥ max { R 0 , ε θ } , let q : = ⌈ ε R θ ⌉ ≥ 1 q := \left\lceil\frac{\varepsilon R}{\theta}\right\rceil\ge1 q := ⌈ θ εR ⌉ ≥ 1 . By iterating Corollary 3.3, we obtain
I q ≤ q ! ( C ψ θ R 0 ′ ε R ) q I 0 = q ! ( θ ε R ) q I 0 . I_q \le q!\left(\frac{C_{\psi}\theta'_{R_0}}{\varepsilon R}\right)^q I_0 = q!\left(\frac{\theta}{\varepsilon R}\right)^q I_0. I q ≤ q ! ( εR C ψ θ R 0 ′ ) q I 0 = q ! ( εR θ ) q I 0 .
Since q θ ≤ ε R q\theta\le\varepsilon R q θ ≤ εR , it follows that
I q ≤ q ! q q I 0 . I_q \le\frac{q!}{q^q}I_0. I q ≤ q q q ! I 0 .
By (3.4) and the definition of I 0 I_0 I 0 ,
J ( R ) ≤ I q , I 0 = J ( ( 1 + ε ) R ) . J(R) \le I_q,\qquad I_0=J((1+\varepsilon)R). J ( R ) ≤ I q , I 0 = J (( 1 + ε ) R ) .
Consequently,
J ( R ) ≤ q ! q q J ( ( 1 + ε ) R ) . J(R) \le\frac{q!}{q^q}J((1+\varepsilon)R). J ( R ) ≤ q q q ! J (( 1 + ε ) R ) .
Using Stirling’s estimate
q ! ≤ e q ( q e ) q , q! \le e\sqrt{q}\left(\frac{q}{e}\right)^q, q ! ≤ e q ( e q ) q ,
we obtain
J ( R ) ≤ e q e − q J ( ( 1 + ε ) R ) . J(R) \le e\sqrt{q}e^{-q}J((1+\varepsilon)R). J ( R ) ≤ e q e − q J (( 1 + ε ) R ) .
Moreover, the definition of q q q gives
ε R θ − 1 < q ≤ ε R θ . \frac{\varepsilon R}{\theta}-1 < q \le\frac{\varepsilon R}{\theta}. θ εR − 1 < q ≤ θ εR .
In particular,
q ≤ ε R θ and e − q < e exp ( − ε R θ ) . \sqrt{q}\le\sqrt{\frac{\varepsilon R}{\theta}}
\qquad\text{and}\qquad
e^{-q}<e\exp\left(-\frac{\varepsilon R}{\theta}\right). q ≤ θ εR and e − q < e exp ( − θ εR ) .
Thus
J ( R ) ≤ e 2 ε R θ exp ( − ε R θ ) J ( ( 1 + ε ) R ) (18) J(R) \le e^2\sqrt{\frac{\varepsilon R}{\theta}}\exp\left(-\frac{\varepsilon R}{\theta}\right)J((1+\varepsilon)R) \tag*{(18)} J ( R ) ≤ e 2 θ εR exp ( − θ εR ) J (( 1 + ε ) R ) (18)
for every R ≥ max { R 0 , θ / ε } R \ge\max\{R_0,\theta/\varepsilon\} R ≥ max { R 0 , θ / ε } .
Step 2: Comparison with the volume growth. Suppose that
n K 0 ( 1 + ε ) θ < 1. (19) n\sqrt{K_0}(1+\varepsilon)\theta< 1. \tag*{(19)} n K 0 ( 1 + ε ) θ < 1. (19)
We show that this condition forces J ≡ 0 J\equiv0 J ≡ 0 .
Fix R ˉ ≥ max { R 0 , θ ε } \bar{R}\ge\max\left\{R_0,\frac{\theta}{\varepsilon}\right\} R ˉ ≥ max { R 0 , ε θ } and define
ϱ i : = ( 1 + ε ) i R ˉ , i = 0 , 1 , 2 , … . \varrho_i := (1+\varepsilon)^i\bar{R},\qquad i=0,1,2,\ldots. ϱ i := ( 1 + ε ) i R ˉ , i = 0 , 1 , 2 , … .
Applying (3.10) successively at R = ϱ 0 , … , ϱ k − 1 R=\varrho_0,\ldots,\varrho_{k-1} R = ϱ 0 , … , ϱ k − 1 , we find
J ( R ˉ ) ≤ [ ∏ i = 0 k − 1 e 2 ε ϱ i θ ] exp ( − ε θ ∑ i = 0 k − 1 ϱ i ) J ( ϱ k ) . J(\bar{R}) \le\left[\prod_{i=0}^{k-1}e^2\sqrt{\frac{\varepsilon\varrho_i}{\theta}}\right]\exp\left(-\frac{\varepsilon}{\theta}\sum_{i=0}^{k-1}\varrho_i\right)J(\varrho_k). J ( R ˉ ) ≤ [ i = 0 ∏ k − 1 e 2 θ ε ϱ i ] exp ( − θ ε i = 0 ∑ k − 1 ϱ i ) J ( ϱ k ) .
The geometric-series identity gives
∑ i = 0 k − 1 ϱ i = ϱ k − R ˉ ε . \sum_{i=0}^{k-1} \varrho_i = \frac{\varrho_k-\bar{R}}{\varepsilon}. i = 0 ∑ k − 1 ϱ i = ε ϱ k − R ˉ .
Furthermore, ϱ i ≤ ϱ k \varrho_i \leq\varrho_k ϱ i ≤ ϱ k for i < k i<k i < k , and hence
∏ i = 0 k − 1 e 2 ε ϱ i θ ≤ ( e 2 ε ϱ k θ ) k . \prod_{i=0}^{k-1} e^2\sqrt{\frac{\varepsilon\varrho_i}{\theta}} \leq\left(e^2\sqrt{\frac{\varepsilon\varrho_k}{\theta}}\right)^k. i = 0 ∏ k − 1 e 2 θ ε ϱ i ≤ ( e 2 θ ε ϱ k ) k .
It follows that
J ( R ˉ ) ≤ ( e 2 ε ϱ k θ ) k exp ( − ϱ k − R ˉ θ ) J ( ϱ k ) . J(\bar{R}) \leq\left(e^2\sqrt{\frac{\varepsilon\varrho_k}{\theta}}\right)^k \exp\left(-\frac{\varrho_k-\bar{R}}{\theta}\right)J(\varrho_k). J ( R ˉ ) ≤ ( e 2 θ ε ϱ k ) k exp ( − θ ϱ k − R ˉ ) J ( ϱ k ) .
By Lemma 3.4 and the volume estimate (3.3),
J ( ϱ k ) ≤ C 2 ω n ( ( 1 + ε ) ϱ k ) n + 1 exp ( n K 0 ( 1 + ε ) ϱ k ) . J(\varrho_k) \leq C_2\omega_n\left((1+\varepsilon)\varrho_k\right)^{n+1}\exp\left(n\sqrt{K_0}(1+\varepsilon)\varrho_k\right). J ( ϱ k ) ≤ C 2 ω n ( ( 1 + ε ) ϱ k ) n + 1 exp ( n K 0 ( 1 + ε ) ϱ k ) .
Therefore,
J ( R ˉ ) ≤ ( e 2 ε ϱ k θ ) k exp ( − ϱ k − R ˉ θ ) C 2 ω n ( ( 1 + ε ) ϱ k ) n + 1 exp ( n K 0 ( 1 + ε ) ϱ k ) . J(\bar{R}) \leq\left(e^2\sqrt{\frac{\varepsilon\varrho_k}{\theta}}\right)^k \exp\left(-\frac{\varrho_k-\bar{R}}{\theta}\right)C_2\omega_n\left((1+\varepsilon)\varrho_k\right)^{n+1}\exp\left(n\sqrt{K_0}(1+\varepsilon)\varrho_k\right). J ( R ˉ ) ≤ ( e 2 θ ε ϱ k ) k exp ( − θ ϱ k − R ˉ ) C 2 ω n ( ( 1 + ε ) ϱ k ) n + 1 exp ( n K 0 ( 1 + ε ) ϱ k ) .
Since k = log ( ϱ k / R ˉ ) log ( 1 + ε ) k=\frac{\log(\varrho_k/\bar{R})}{\log(1+\varepsilon)} k = l o g ( 1 + ε ) l o g ( ϱ k / R ˉ ) , we have k = O ( log ϱ k ) k=O(\log\varrho_k) k = O ( log ϱ k ) as k → ∞ k\to\infty k → ∞ . Taking logarithms of the positive expression on the right-hand side above, its logarithm is
− [ 1 θ − n K 0 ( 1 + ε ) ] ϱ k + O ( ( log ϱ k ) 2 ) . -\left[\frac{1}{\theta}-n\sqrt{K_0}(1+\varepsilon)\right]\varrho_k+O\left((\log\varrho_k)^2\right). − [ θ 1 − n K 0 ( 1 + ε ) ] ϱ k + O ( ( log ϱ k ) 2 ) .
Indeed, the product of the k k k factors contributes O ( k log ϱ k ) = O ( ( log ϱ k ) 2 ) O(k\log\varrho_k)=O((\log\varrho_k)^2) O ( k log ϱ k ) = O (( log ϱ k ) 2 ) , while the remaining polynomial factors contribute only O ( log ϱ k ) O(\log\varrho_k) O ( log ϱ k ) .
By (3.11), the coefficient
1 θ − n K 0 ( 1 + ε ) \frac{1}{\theta}-n\sqrt{K_0}(1+\varepsilon) θ 1 − n K 0 ( 1 + ε )
is strictly positive. Since
( log ϱ k ) 2 ϱ k ⟶ 0 , \frac{(\log\varrho_k)^2}{\varrho_k}\longrightarrow0, ϱ k ( log ϱ k ) 2 ⟶ 0 ,
the right-hand side of the estimate for J ( R ˉ ) J(\bar{R}) J ( R ˉ ) tends to zero as k → ∞ k\to\infty k → ∞ . Hence J ( R ˉ ) = 0 J(\bar{R})=0 J ( R ˉ ) = 0 . As R ˉ \bar{R} R ˉ was arbitrary above the fixed threshold, J ( R ) = 0 J(R)=0 J ( R ) = 0 for all sufficiently large R R R . Because J J J is nonnegative and nondecreasing, this implies
J ( R ) = 0 for every R > 0. J(R)=0 \qquad\text{for every } R>0. J ( R ) = 0 for every R > 0.
Thus J ≡ 0 J\equiv0 J ≡ 0 . Since Y ≥ 0 Y\geq0 Y ≥ 0 is continuous and
J ( R ) = ∫ { r ≤ R } Y , J(R)=\int_{\{r\leq R\}}Y, J ( R ) = ∫ { r ≤ R } Y ,
we conclude that Y ≡ 0 Y\equiv0 Y ≡ 0 on M M M .
Recall that
Y = ( 2 s − n ) H G s + 2 s δ 0 2 H G s − 1 , δ 0 2 = c 0 2 − K 0 > 0. Y=(2s-n)HG^s+2s\delta_0^2HG^{s-1}, \qquad\delta_0^2=c_0^2-K_0>0. Y = ( 2 s − n ) H G s + 2 s δ 0 2 H G s − 1 , δ 0 2 = c 0 2 − K 0 > 0.
On U = { H > c 0 } U = \{H > c_0\} U = { H > c 0 } , we have H > 0 H > 0 H > 0 and G > 0 G > 0 G > 0 . Consequently,
Y ≥ 2 s δ 0 2 H G s − 1 > 0 on U , Y \ge2s\delta_0^2HG^{s-1} > 0 \qquad\text{on } U, Y ≥ 2 s δ 0 2 H G s − 1 > 0 on U ,
contradicting Y ≡ 0 Y \equiv0 Y ≡ 0 unless U = ∅ U = \varnothing U = ∅ . Therefore,
H ≤ c 0 on M , H \le c_0 \qquad\text{on } M, H ≤ c 0 on M ,
provided the parameters satisfy (19) .
Step 3: Choice of the parameters. Let c 0 > c ∗ c_0 > c_* c 0 > c ∗ . By the properties of c ∗ c_* c ∗ ,
c 0 > K 0 , δ 0 2 : = c 0 2 − K 0 > 0. c_0 > \sqrt{K_0}, \qquad\delta_0^2 := c_0^2 - K_0 > 0. c 0 > K 0 , δ 0 2 := c 0 2 − K 0 > 0.
We show that the parameters in the preceding steps can be chosen so that (19) holds.
If K 0 = 0 K_0 = 0 K 0 = 0 , then (19) holds automatically, since its left-hand side is zero. We may therefore assume that K 0 > 0 K_0 > 0 K 0 > 0 .
By the definition of c ∗ c_* c ∗ , the inequality c 0 > c ∗ c_0 > c_* c 0 > c ∗ is equivalent to
c 0 2 − n K 0 c 0 − K 0 − n K 0 C 1 > 0. c_0^2 - n\sqrt{K_0}c_0 - K_0 - n\sqrt{K_0}C_1 > 0. c 0 2 − n K 0 c 0 − K 0 − n K 0 C 1 > 0.
Equivalently,
(20)
n K 0 c 0 + C 1 c 0 2 − K 0 < 1. (20) n\sqrt{K_0}\frac{c_0 + C_1}{c_0^2 - K_0} < 1. \tag*{(20)} n K 0 c 0 2 − K 0 c 0 + C 1 < 1. (20)
Moreover,
1 c 0 ≤ c 0 + C 1 c 0 2 − K 0 , \frac{1}{c_0} \le\frac{c_0 + C_1}{c_0^2 - K_0}, c 0 1 ≤ c 0 2 − K 0 c 0 + C 1 ,
because this inequality is equivalent to
c 0 ( c 0 + C 1 ) − ( c 0 2 − K 0 ) = c 0 C 1 + K 0 ≥ 0. c_0(c_0 + C_1) - (c_0^2 - K_0) = c_0C_1 + K_0 \ge0. c 0 ( c 0 + C 1 ) − ( c 0 2 − K 0 ) = c 0 C 1 + K 0 ≥ 0.
Recall that
θ ′ ( s , c 0 , C R ′ ) = max { 2 s + 1 ( 2 s − n ) c 0 , ( 2 s + 1 ) c 0 + 2 s n C R ′ 2 s ( c 0 2 − K 0 ) } . \theta'(s,c_0,C'_R) = \max\left\{\frac{2s+1}{(2s-n)c_0},\frac{(2s+1)c_0+\frac{2s}{n}C'_R}{2s(c_0^2-K_0)}\right\}. θ ′ ( s , c 0 , C R ′ ) = max { ( 2 s − n ) c 0 2 s + 1 , 2 s ( c 0 2 − K 0 ) ( 2 s + 1 ) c 0 + n 2 s C R ′ } .
For fixed σ > 0 \sigma> 0 σ > 0 , as s → ∞ s \to\infty s → ∞ , the two entries in this maximum, with C R ′ = n C 1 + σ C'_R = nC_1 + \sigma C R ′ = n C 1 + σ , converge respectively to
1 c 0 and c 0 + C 1 + σ / n c 0 2 − K 0 . \frac{1}{c_0} \qquad\text{and} \qquad\frac{c_0+C_1+\sigma/n}{c_0^2-K_0}. c 0 1 and c 0 2 − K 0 c 0 + C 1 + σ / n .
By (20) , we can first choose σ > 0 \sigma> 0 σ > 0 sufficiently small so that
n K 0 c 0 + C 1 + σ / n c 0 2 − K 0 < 1. n\sqrt{K_0}\frac{c_0+C_1+\sigma/n}{c_0^2-K_0} < 1. n K 0 c 0 2 − K 0 c 0 + C 1 + σ / n < 1.
Together with
1 c 0 ≤ c 0 + C 1 c 0 2 − K 0 , \frac{1}{c_0} \le\frac{c_0+C_1}{c_0^2-K_0}, c 0 1 ≤ c 0 2 − K 0 c 0 + C 1 ,
this ensures that both limiting entries satisfy the strict inequality required by the rate condition.
It follows that we can choose τ > 0 \tau> 0 τ > 0 sufficiently small, ε ∈ ( 0 , 1 ] \varepsilon\in(0,1] ε ∈ ( 0 , 1 ] sufficiently small, and an integer s s s sufficiently large with s ≥ 3 s \ge3 s ≥ 3 and 2 s > n 2s > n 2 s > n , such that
n K 0 ( 1 + ε ) ( 1 + τ ) θ ′ ( s , c 0 , n C 1 + σ ) < 1. (21) n\sqrt{K_0}(1+\varepsilon)(1+\tau)\theta'(s,c_0,nC_1+\sigma)<1. \tag*{(21)} n K 0 ( 1 + ε ) ( 1 + τ ) θ ′ ( s , c 0 , n C 1 + σ ) < 1. (21)
Choose a cutoff function ψ \psi ψ with C ψ ≤ 1 + τ C_\psi\le1+\tau C ψ ≤ 1 + τ . Finally, by the definition
C R ′ = n C 1 + C ψ ′ C ψ ε R , C'_R=nC_1+\frac{C'_\psi}{C_\psi\varepsilon R}, C R ′ = n C 1 + C ψ εR C ψ ′ ,
we can choose R 0 ≥ R ♭ R_0 \ge R_\flat R 0 ≥ R ♭ sufficiently large that
C R 0 ′ ≤ n C 1 + σ . C'_{R_0}\le nC_1+\sigma. C R 0 ′ ≤ n C 1 + σ .
Since θ ′ \theta' θ ′ is nondecreasing in its last argument,
θ R 0 ′ = θ ′ ( s , c 0 , C R 0 ′ ) ≤ θ ′ ( s , c 0 , n C 1 + σ ) . \theta'_{R_0}=\theta'(s,c_0,C'_{R_0})\le\theta'(s,c_0,nC_1+\sigma). θ R 0 ′ = θ ′ ( s , c 0 , C R 0 ′ ) ≤ θ ′ ( s , c 0 , n C 1 + σ ) .
Thus, with θ = C ψ θ R 0 ′ \theta=C_\psi\theta'_{R_0} θ = C ψ θ R 0 ′ , we obtain
n K 0 ( 1 + ε ) θ ≤ n K 0 ( 1 + ε ) ( 1 + τ ) θ ′ ( s , c 0 , n C 1 + σ ) < 1. n\sqrt{K_0}(1+\varepsilon)\theta\le n\sqrt{K_0}(1+\varepsilon)(1+\tau)\theta'(s,c_0,nC_1+\sigma)<1. n K 0 ( 1 + ε ) θ ≤ n K 0 ( 1 + ε ) ( 1 + τ ) θ ′ ( s , c 0 , n C 1 + σ ) < 1.
Hence (19) holds. Step 2 now yields
H ≤ c 0 on M . H\le c_0 \qquad\text{on }M. H ≤ c 0 on M .
Since this conclusion holds for every c 0 > c ∗ c_0>c_* c 0 > c ∗ , letting c 0 ↓ c ∗ c_0\downarrow c_* c 0 ↓ c ∗ gives
H ≤ c ∗ on M . H\le c_* \qquad\text{on }M. H ≤ c ∗ on M .
Step 4: The opposite inequality. Reverse the orientation of the foliation by replacing N N N with − N -N − N . The corresponding mean curvature becomes − H -H − H , while
∇ − N ( − N ) = ∇ N N = T . \nabla_{-N}(-N)=\nabla_NN=T. ∇ − N ( − N ) = ∇ N N = T .
Thus the vector field T T T , the underlying foliation, and the sectional-curvature assumption remain unchanged. Applying the upper bound just proved to the reversed orientation gives
− H ≤ c ∗ on M . -H\le c_* \qquad\text{on }M. − H ≤ c ∗ on M .
Combining this with H ≤ c ∗ H\le c_* H ≤ c ∗ , we conclude that
∣ H ∣ ≤ c ∗ = c ∗ ( n , K 0 ; ct K 0 ( R ∗ ) ) on M . |H|\le c_*=c_*\left(n,K_0;\operatorname{ct}_{K_0}(R_*)\right) \qquad\text{on }M. ∣ H ∣ ≤ c ∗ = c ∗ ( n , K 0 ; ct K 0 ( R ∗ ) ) on M .
This completes the proof. □
Proof of Theorem A. By Proposition 3.5 , for every R ∗ > 0 R_*>0 R ∗ > 0 ,
∣ H ∣ ≤ c ∗ ( n , K 0 ; ct K 0 ( R ∗ ) ) on M . |H|\le c_*\left(n,K_0;\operatorname{ct}_{K_0}(R_*)\right) \qquad\text{on }M. ∣ H ∣ ≤ c ∗ ( n , K 0 ; ct K 0 ( R ∗ ) ) on M .
Since
lim R ∗ → ∞ ct K 0 ( R ∗ ) = K 0 , \lim_{R_*\to\infty}\operatorname{ct}_{K_0}(R_*)=\sqrt{K_0}, R ∗ → ∞ lim ct K 0 ( R ∗ ) = K 0 ,
the continuity of c ∗ ( n , K 0 ; ⋅ ) c_\ast(n,K_0;\cdot) c ∗ ( n , K 0 ; ⋅ ) implies that
lim R ∗ → ∞ c ∗ ( n , K 0 ; c t K 0 ( R ∗ ) ) = c ∗ ( n , K 0 ; K 0 ) . \lim_{R^\ast\to\infty}c_\ast\left(n,K_0;c t_{K_0}(R^\ast)\right)=c_\ast\left(n,K_0;\sqrt{K_0}\right). R ∗ → ∞ lim c ∗ ( n , K 0 ; c t K 0 ( R ∗ ) ) = c ∗ ( n , K 0 ; K 0 ) .
Passing to the limit in the estimate above, we obtain
∣ H ∣ ≤ c ∗ ( n , K 0 ; K 0 ) . |H|\le c_\ast\left(n,K_0;\sqrt{K_0}\right). ∣ H ∣ ≤ c ∗ ( n , K 0 ; K 0 ) .
By (17) , c ∗ ( n , K 0 ; K 0 ) = ( n + 1 ) K 0 c_\ast\left(n,K_0;\sqrt{K_0}\right)=(n+1)\sqrt{K_0} c ∗ ( n , K 0 ; K 0 ) = ( n + 1 ) K 0 . Therefore,
∣ H ∣ ≤ ( n + 1 ) K 0 on M . |H|\le(n+1)\sqrt{K_0}\quad\text{on }M. ∣ H ∣ ≤ ( n + 1 ) K 0 on M .
In particular, if K 0 = 0 K_0=0 K 0 = 0 , then ∣ H ∣ ≤ 0 |H|\le0 ∣ H ∣ ≤ 0 , and hence H ≡ 0 H\equiv0 H ≡ 0 on M M M . Thus every leaf of the foliation is minimal. □
Appendix A. Proofs of Lemmas 2.10 and 2.11
Proof of Lemma 2.10
Lemma 2.10. Let I ⊂ R I\subset\mathbb{R} I ⊂ R be an open interval, and let v , β ∈ C 0 ( I ) v,\beta\in C^0(I) v , β ∈ C 0 ( I ) . Assume that for every t 0 ∈ I t_0\in I t 0 ∈ I and every ε > 0 \varepsilon>0 ε > 0 , there exists a C 2 C^2 C 2 function ϕ \phi ϕ defined on a neighborhood of t 0 t_0 t 0 such that
ϕ ≥ v , ϕ ( t 0 ) = v ( t 0 ) , ϕ ′ ′ ( t 0 ) ≤ β ( t 0 ) + ε . \phi\ge v,\qquad\phi(t_0)=v(t_0),\qquad\phi''(t_0)\le\beta(t_0)+\varepsilon. ϕ ≥ v , ϕ ( t 0 ) = v ( t 0 ) , ϕ ′′ ( t 0 ) ≤ β ( t 0 ) + ε . Then v ′ ′ ≤ β v''\le\beta v ′′ ≤ β in the sense of distributions on I I I , that is,
∫ I v ( t ) ζ ′ ′ ( t ) d t ≤ ∫ I β ( t ) ζ ( t ) d t \int_I v(t)\zeta''(t)\,\mathrm{d}t\le\int_I\beta(t)\zeta(t)\,\mathrm{d}t ∫ I v ( t ) ζ ′′ ( t ) d t ≤ ∫ I β ( t ) ζ ( t ) d t for every nonnegative ζ ∈ C c ∞ ( I ) \zeta\in C^\infty_c(I) ζ ∈ C c ∞ ( I ) .
Proof. Since β \beta β is continuous on I I I , there exists a function B ∈ C 2 ( I ) B\in C^2(I) B ∈ C 2 ( I ) such that B ′ ′ = β B''=\beta B ′′ = β . Set w : = v − B w:=v-B w := v − B . We first observe that w w w satisfies the following upper support property: for every t 0 ∈ I t_0\in I t 0 ∈ I and every ε > 0 \varepsilon>0 ε > 0 , there exists a C 2 C^2 C 2 function φ \varphi φ defined on a neighborhood of t 0 t_0 t 0 such that
φ ≥ w , φ ( t 0 ) = w ( t 0 ) , φ ′ ′ ( t 0 ) ≤ ε . \varphi\ge w,\qquad\varphi(t_0)=w(t_0),\qquad\varphi''(t_0)\le\varepsilon. φ ≥ w , φ ( t 0 ) = w ( t 0 ) , φ ′′ ( t 0 ) ≤ ε .
Indeed, given t 0 ∈ I t_0\in I t 0 ∈ I and ε > 0 \varepsilon>0 ε > 0 , apply the hypothesis to v v v with the same ε \varepsilon ε . Let φ 0 \varphi_0 φ 0 be the resulting upper support function, so that
φ 0 ≥ v , φ 0 ( t 0 ) = v ( t 0 ) , φ 0 ′ ′ ( t 0 ) ≤ β ( t 0 ) + ε . \varphi_0\ge v,\qquad\varphi_0(t_0)=v(t_0),\qquad\varphi_0''(t_0)\le\beta(t_0)+\varepsilon. φ 0 ≥ v , φ 0 ( t 0 ) = v ( t 0 ) , φ 0 ′′ ( t 0 ) ≤ β ( t 0 ) + ε .
Define φ : = φ 0 − B \varphi:=\varphi_0-B φ := φ 0 − B . Then φ ≥ v − B = w \varphi\ge v-B=w φ ≥ v − B = w , φ ( t 0 ) = w ( t 0 ) \varphi(t_0)=w(t_0) φ ( t 0 ) = w ( t 0 ) , and
φ ′ ′ ( t 0 ) = φ 0 ′ ′ ( t 0 ) − B ′ ′ ( t 0 ) ≤ β ( t 0 ) + ε − β ( t 0 ) = ε . \begin{aligned}
\varphi''(t_0)=\varphi_0''(t_0)-B''(t_0) \\
&\le\beta(t_0)+\varepsilon-\beta(t_0)=\varepsilon.
\end{aligned} φ ′′ ( t 0 ) = φ 0 ′′ ( t 0 ) − B ′′ ( t 0 ) ≤ β ( t 0 ) + ε − β ( t 0 ) = ε .
This establishes the asserted upper support property of w w w .
We claim that w w w is concave on I I I . Suppose, to the contrary, that w w w is not concave. Then there exist a , b ∈ I a,b\in I a , b ∈ I , with a < b a<b a < b , such that w w w lies strictly below the chord joining ( a , w ( a ) ) (a,w(a)) ( a , w ( a )) and ( b , w ( b ) ) (b,w(b)) ( b , w ( b )) at some point of ( a , b ) (a,b) ( a , b ) . Let ℓ \ell ℓ be the affine function satisfying ℓ ( a ) = w ( a ) \ell(a)=w(a) ℓ ( a ) = w ( a ) , ℓ ( b ) = w ( b ) \ell(b)=w(b) ℓ ( b ) = w ( b ) . Since w w w is not concave, we have
m : = max t ∈ [ a , b ] ( ℓ ( t ) − w ( t ) ) > 0. m:=\max_{t\in[a,b]}(\ell(t)-w(t))>0. m := t ∈ [ a , b ] max ( ℓ ( t ) − w ( t )) > 0.
Choose δ > 0 \delta>0 δ > 0 sufficiently small that δ ( b − a ) 2 4 < m \frac{\delta(b-a)^2}{4}<m 4 δ ( b − a ) 2 < m . Define
h ~ ( t ) : = ℓ ( t ) − δ ( t − a ) ( b − t ) − w ( t ) , \widetilde{h}(t):=\ell(t)-\delta(t-a)(b-t)-w(t), h ( t ) := ℓ ( t ) − δ ( t − a ) ( b − t ) − w ( t ) ,
t ∈ [ a , b ] t\in[a,b] t ∈ [ a , b ] . Since w w w and ℓ \ell ℓ are continuous, h ~ \widetilde{h} h is continuous on [ a , b ] [a,b] [ a , b ] . Moreover, h ~ ( a ) = h ~ ( b ) = 0 \widetilde{h}(a)=\widetilde{h}(b)=0 h ( a ) = h ( b ) = 0 . Since 0 ≤ ( t − a ) ( b − t ) ≤ ( b − a ) 2 4 0\leq(t-a)(b-t)\leq\frac{(b-a)^2}{4} 0 ≤ ( t − a ) ( b − t ) ≤ 4 ( b − a ) 2 for all t ∈ [ a , b ] t\in[a,b] t ∈ [ a , b ] , we obtain
max t ∈ [ a , b ] h ~ ( t ) ≥ m − δ ( b − a ) 2 4 > 0. \max_{t\in[a,b]}\widetilde{h}(t)\geq m-\frac{\delta(b-a)^2}{4}>0. t ∈ [ a , b ] max h ( t ) ≥ m − 4 δ ( b − a ) 2 > 0.
Consequently, h h h attains its positive maximum at some t ∗ ∈ ( a , b ) t^*\in(a,b) t ∗ ∈ ( a , b ) .
Now define the C 2 C^2 C 2 function
ψ ( t ) : = ℓ ( t ) − δ ( t − a ) ( b − t ) − h ~ ( t ∗ ) . \psi(t):=\ell(t)-\delta(t-a)(b-t)-\widetilde{h}(t^*). ψ ( t ) := ℓ ( t ) − δ ( t − a ) ( b − t ) − h ( t ∗ ) .
By the definition of h h h and the maximality of h ~ ( t ∗ ) \widetilde{h}(t^*) h ( t ∗ ) , we have
ψ ( t ) − w ( t ) = ℓ ( t ) − δ ( t − a ) ( b − t ) − h ~ ( t ∗ ) − w ( t ) = h ~ ( t ) − h ~ ( t ∗ ) ≤ 0 \begin{aligned}
\psi(t)-w(t)&=\ell(t)-\delta(t-a)(b-t)-\widetilde{h}(t^*)-w(t)\\
&=\widetilde{h}(t)-\widetilde{h}(t^*)\leq0
\end{aligned} ψ ( t ) − w ( t ) = ℓ ( t ) − δ ( t − a ) ( b − t ) − h ( t ∗ ) − w ( t ) = h ( t ) − h ( t ∗ ) ≤ 0
for every t ∈ [ a , b ] t\in[a,b] t ∈ [ a , b ] . Hence ψ ≤ w \psi\leq w ψ ≤ w on [ a , b ] [a,b] [ a , b ] , ψ ( t ∗ ) = w ( t ∗ ) \psi(t^*)=w(t^*) ψ ( t ∗ ) = w ( t ∗ ) . Furthermore, since ℓ \ell ℓ is affine, ψ ′ ′ ( t ) = 2 δ \psi''(t)=2\delta ψ ′′ ( t ) = 2 δ for every t ∈ [ a , b ] t\in[a,b] t ∈ [ a , b ] .
We now apply the upper support property of w w w at t ∗ t^* t ∗ with ε = δ \varepsilon=\delta ε = δ . There exists a C 2 C^2 C 2 function φ 1 \varphi_1 φ 1 defined on a neighborhood of t ∗ t^* t ∗ such that
φ 1 ≥ w , φ 1 ( t ∗ ) = w ( t ∗ ) , φ 1 ′ ′ ( t ∗ ) ≤ δ . \varphi_1\geq w,\qquad\varphi_1(t^*)=w(t^*),\qquad\varphi_1''(t^*)\leq\delta. φ 1 ≥ w , φ 1 ( t ∗ ) = w ( t ∗ ) , φ 1 ′′ ( t ∗ ) ≤ δ .
Since ψ ≤ w \psi\leq w ψ ≤ w on [ a , b ] [a,b] [ a , b ] and t ∗ ∈ ( a , b ) t^*\in(a,b) t ∗ ∈ ( a , b ) , we have φ 1 ≥ ψ \varphi_1\geq\psi φ 1 ≥ ψ on a neighborhood of t ∗ t^* t ∗ , with equality at t ∗ t^* t ∗ . Thus φ 1 − ψ \varphi_1-\psi φ 1 − ψ has a local minimum at t ∗ t^* t ∗ , and therefore
( φ 1 − ψ ) ′ ′ ( t ∗ ) ≥ 0. (\varphi_1-\psi)''(t^*)\geq0. ( φ 1 − ψ ) ′′ ( t ∗ ) ≥ 0.
It follows that
φ 1 ′ ′ ( t ∗ ) ≥ ψ ′ ′ ( t ∗ ) = 2 δ , \varphi_1''(t^*)\geq\psi''(t^*)=2\delta, φ 1 ′′ ( t ∗ ) ≥ ψ ′′ ( t ∗ ) = 2 δ ,
contradicting φ 1 ′ ′ ( t ∗ ) ≤ δ \varphi_1''(t^*)\leq\delta φ 1 ′′ ( t ∗ ) ≤ δ , since δ > 0 \delta>0 δ > 0 . Therefore, w w w is concave on I I I .
It remains to show that the concavity of w w w implies
∫ I w ( t ) ζ ′ ′ ( t ) d t ≤ 0 \int_I w(t)\zeta''(t)\,\mathrm{d}t\leq0 ∫ I w ( t ) ζ ′′ ( t ) d t ≤ 0
for every nonnegative ζ ∈ C c ∞ ( I ) \zeta\in C_c^\infty(I) ζ ∈ C c ∞ ( I ) . Fix such a function ζ \zeta ζ . For every sufficiently small h > 0 h>0 h > 0 , the concavity of w w w gives
w ( t + h ) − 2 w ( t ) + w ( t − h ) ≤ 0 w(t+h)-2w(t)+w(t-h)\leq0 w ( t + h ) − 2 w ( t ) + w ( t − h ) ≤ 0
whenever t − h , t , t + h ∈ I t-h,t,t+h\in I t − h , t , t + h ∈ I . Since ζ \zeta ζ has compact support, we may choose h > 0 h>0 h > 0 sufficiently small so that all relevant points lie in I I I . Multiplying by ζ ( t ) ≥ 0 \zeta(t)\geq0 ζ ( t ) ≥ 0 and integrating, we obtain
∫ I ζ ( t ) w ( t + h ) − 2 w ( t ) + w ( t − h ) h 2 d t ≤ 0. \int_I \zeta(t)\frac{w(t+h)-2w(t)+w(t-h)}{h^2}\,\mathrm{d}t\leq0. ∫ I ζ ( t ) h 2 w ( t + h ) − 2 w ( t ) + w ( t − h ) d t ≤ 0.
A change of variables in the first and third terms, together with the compact support of ζ \zeta ζ , yields
∫ I w ( t ) ζ ( t + h ) − 2 ζ ( t ) + ζ ( t − h ) h 2 d t ≤ 0. \int_I w(t)\frac{\zeta(t+h)-2\zeta(t)+\zeta(t-h)}{h^2}\,\mathrm{d}t\leq0. ∫ I w ( t ) h 2 ζ ( t + h ) − 2 ζ ( t ) + ζ ( t − h ) d t ≤ 0.
As h → 0 h\to0 h → 0 , the central second difference of ζ \zeta ζ converges uniformly to ζ ′ ′ \zeta'' ζ ′′ on a compact subset of I I I containing the supports of these functions. Since w w w is continuous, we may pass to the limit and conclude that
∫ I w ( t ) ζ ′ ′ ( t ) d t ≤ 0. \int_I w(t)\zeta''(t)\,\mathrm{d}t\leq0. ∫ I w ( t ) ζ ′′ ( t ) d t ≤ 0.
Finally, recalling that v = w + B v=w+B v = w + B , we obtain
∫ I v ( t ) ζ ′ ′ ( t ) d t = ∫ I w ( t ) ζ ′ ′ ( t ) d t + ∫ I B ( t ) ζ ′ ′ ( t ) d t ≤ ∫ I B ( t ) ζ ′ ′ ( t ) d t . \begin{aligned}
\int_I v(t)\zeta''(t)\,\mathrm{d}t
&=\int_I w(t)\zeta''(t)\,\mathrm{d}t+\int_I B(t)\zeta''(t)\,\mathrm{d}t\\
&\leq\int_I B(t)\zeta''(t)\,\mathrm{d}t.
\end{aligned} ∫ I v ( t ) ζ ′′ ( t ) d t = ∫ I w ( t ) ζ ′′ ( t ) d t + ∫ I B ( t ) ζ ′′ ( t ) d t ≤ ∫ I B ( t ) ζ ′′ ( t ) d t .
Since ζ \zeta ζ has compact support, integration by parts twice gives
∫ I B ( t ) ζ ′ ′ ( t ) d t = ∫ I B ′ ′ ( t ) ζ ( t ) d t = ∫ I β ( t ) ζ ( t ) d t . \int_I B(t)\zeta''(t)\,\mathrm{d}t=\int_I B''(t)\zeta(t)\,\mathrm{d}t=\int_I \beta(t)\zeta(t)\,\mathrm{d}t. ∫ I B ( t ) ζ ′′ ( t ) d t = ∫ I B ′′ ( t ) ζ ( t ) d t = ∫ I β ( t ) ζ ( t ) d t .
Therefore,
∫ I v ( t ) ζ ′ ′ ( t ) d t ≤ ∫ I β ( t ) ζ ( t ) d t \int_I v(t)\zeta''(t)\,\mathrm{d}t\leq\int_I \beta(t)\zeta(t)\,\mathrm{d}t ∫ I v ( t ) ζ ′′ ( t ) d t ≤ ∫ I β ( t ) ζ ( t ) d t
for every nonnegative ζ ∈ C c ∞ ( I ) \zeta\in C^\infty_c(I) ζ ∈ C c ∞ ( I ) , as desired. □ \square □
Proof of Lemma 2.11
Lemma A.1. Let f f f be L L L -Lipschitz on a neighbourhood of x x x and let φ \varphi φ be C 1 C^1 C 1 near x x x with φ ≥ f \varphi\geq f φ ≥ f near x x x and φ ( x ) = f ( x ) \varphi(x)=f(x) φ ( x ) = f ( x ) . Then ∣ ∇ φ ( x ) ∣ ≤ L \lvert\nabla\varphi(x)\rvert\leq L ∣ ∇ φ ( x )∣ ≤ L .
Proof. Let v ∈ T x M v\in T_xM v ∈ T x M be a unit vector. For all sufficiently small t > 0 t>0 t > 0 , the minimizing property of radial geodesics gives dist M ( x , exp x ( t v ) ) = t \operatorname{dist}_M(x,\exp_x(tv))=t dist M ( x , exp x ( t v )) = t . Since f f f is L L L -Lipschitz, we have ∣ f ( y ) − f ( x ) ∣ ≤ L dist M ( x , y ) \lvert f(y)-f(x)\rvert\leq L\operatorname{dist}_M(x,y) ∣ f ( y ) − f ( x )∣ ≤ L dist M ( x , y ) for all y y y sufficiently close to x x x . In particular, f ( y ) ≥ f ( x ) − L dist M ( x , y ) f(y)\geq f(x)-L\operatorname{dist}_M(x,y) f ( y ) ≥ f ( x ) − L dist M ( x , y ) . Taking y = exp x ( t v ) y=\exp_x(tv) y = exp x ( t v ) and using φ ≥ f \varphi\geq f φ ≥ f and φ ( x ) = f ( x ) \varphi(x)=f(x) φ ( x ) = f ( x ) , we obtain
φ ( exp x ( t v ) ) ≥ f ( exp x ( t v ) ) ≥ f ( x ) − L dist M ( x , exp x ( t v ) ) = f ( x ) − L t = φ ( x ) − L t . \begin{aligned}
\varphi(\exp_x(tv))&\geq f(\exp_x(tv))\\
&\geq f(x)-L\operatorname{dist}_M(x,\exp_x(tv))\\
&=f(x)-Lt\\
&=\varphi(x)-Lt.
\end{aligned} φ ( exp x ( t v )) ≥ f ( exp x ( t v )) ≥ f ( x ) − L dist M ( x , exp x ( t v )) = f ( x ) − L t = φ ( x ) − L t .
Thus,
φ ( exp x ( t v ) ) − φ ( x ) t ≥ − L . \frac{\varphi(\exp_x(tv))-\varphi(x)}{t}\ge-L. t φ ( exp x ( t v )) − φ ( x ) ≥ − L .
Letting t → 0 + t\to0^{+} t → 0 + and using the differentiability of φ \varphi φ , we conclude that
⟨ ∇ φ ( x ) , v ⟩ ≥ − L . \langle\nabla\varphi(x),v\rangle\ge-L. ⟨ ∇ φ ( x ) , v ⟩ ≥ − L .
Applying the same argument to − v -v − v yields ⟨ ∇ φ ( x ) , v ⟩ ≤ L \langle\nabla\varphi(x),v\rangle\le L ⟨ ∇ φ ( x ) , v ⟩ ≤ L . Therefore, ∣ ⟨ ∇ φ ( x ) , v ⟩ ∣ ≤ L |\langle\nabla\varphi(x),v\rangle|\le L ∣ ⟨ ∇ φ ( x ) , v ⟩ ∣ ≤ L for every unit vector v ∈ T x M v\in T_xM v ∈ T x M . Taking the supremum over all unit vectors, we obtain ∣ ∇ φ ( x ) ∣ ≤ L |\nabla\varphi(x)|\le L ∣∇ φ ( x ) ∣ ≤ L . This completes the proof. □ \square □
Lemma 2.11. Let W ⊂ M W\subset M W ⊂ M be open, a ∈ R a\in\mathbb{R} a ∈ R , L > 0 L>0 L > 0 , and let f ∈ C 0 ( W ) f\in C^{0}(W) f ∈ C 0 ( W ) be such that every point of W W W has a neighbourhood on which f f f is L L L -Lipschitz, and Hess f ≤ a ⟨ ⋅ , ⋅ ⟩ \operatorname{Hess}f\le a\langle\cdot,\cdot\rangle Hess f ≤ a ⟨ ⋅ , ⋅ ⟩ in the support sense at every point of W W W . Let K ⊂ W K\subset W K ⊂ W be compact and η > 0 \eta>0 η > 0 . Then there are σ > 0 \sigma>0 σ > 0 and points c 1 , … , c m ∈ K c_1,\ldots,c_m\in K c 1 , … , c m ∈ K such that, with U i : = B 2 σ ( c i ) U_i:=B_{2\sigma}(c_i) U i := B 2 σ ( c i ) :
(i) K ⊂ U 1 ∪ ⋯ ∪ U m K\subset U_1\cup\cdots\cup U_m K ⊂ U 1 ∪ ⋯ ∪ U m , and exp c i \exp_{c_i} exp c i maps the Euclidean ball B ( 0 , 4 σ ) ⊂ T c i M ≅ R n + 1 B(0,4\sigma)\subset T_{c_i}M\cong\mathbb{R}^{n+1} B ( 0 , 4 σ ) ⊂ T c i M ≅ R n + 1 (identified by an orthonormal basis) diffeomorphically onto the geodesic ball B 4 σ ( c i ) ⊂ W B_{4\sigma}(c_i)\subset W B 4 σ ( c i ) ⊂ W ;
(ii) for 0 < δ < σ 0<\delta<\sigma 0 < δ < σ the functions f i , δ : = ( u i ∗ μ δ ) ∘ exp c i − 1 f_{i,\delta}:=(u_i*\mu_\delta)\circ\exp_{c_i}^{-1} f i , δ := ( u i ∗ μ δ ) ∘ exp c i − 1 , u i : = f ∘ exp c i u_i:=f\circ\exp_{c_i} u i := f ∘ exp c i , are smooth on B 3 σ ( c i ) B_{3\sigma}(c_i) B 3 σ ( c i ) , ∣ ∇ f i , δ ∣ ≤ 2 L |\nabla f_{i,\delta}|\le2L ∣∇ f i , δ ∣ ≤ 2 L there, f i , δ → f f_{i,\delta}\to f f i , δ → f uniformly on U i ‾ \overline{U_i} U i and ∇ f i , δ → ∇ f \nabla f_{i,\delta}\to\nabla f ∇ f i , δ → ∇ f a.e. on U i U_i U i as δ → 0 \delta\to0 δ → 0 ;
(iii) there are ω i ( δ ) → 0 \omega_i(\delta)\to0 ω i ( δ ) → 0 such that Hess f i , δ ≤ ( a + η + ω i ( δ ) ) ⟨ ⋅ , ⋅ ⟩ \operatorname{Hess}f_{i,\delta}\le(a+\eta+\omega_i(\delta))\langle\cdot,\cdot\rangle Hess f i , δ ≤ ( a + η + ω i ( δ )) ⟨ ⋅ , ⋅ ⟩ on U i ‾ \overline{U_i} U i .
Proof. We divide the proof into three steps, corresponding to assertions (i), (ii), and (iii).
Proof of item (i).
Since K ⊂ W K\subset W K ⊂ W is compact and W W W is open, there exists r W > 0 r_W>0 r W > 0 such that B r W ( c ) ⊂ W B_{r_W}(c)\subset W B r W ( c ) ⊂ W for every c ∈ K c\in K c ∈ K . Moreover, the injectivity radius function c ↦ inj ( c ) c\mapsto\operatorname{inj}(c) c ↦ inj ( c ) is positive and continuous. Consequently, r K : = min c ∈ K inj ( c ) > 0 r_K:=\min_{c\in K}\operatorname{inj}(c)>0 r K := min c ∈ K inj ( c ) > 0 . Thus, for every c ∈ K c\in K c ∈ K , the exponential map is a diffeomorphism from B ( 0 , r ) B(0,r) B ( 0 , r ) onto B r ( c ) B_r(c) B r ( c ) whenever 0 < r < min { r W , r K } 0<r<\min\{r_W,r_K\} 0 < r < min { r W , r K } .
We next establish uniform estimates in normal coordinates. The associated normal coordinate maps depend smoothly on the base point and the coordinate variables. Since the metric coefficients satisfy g i j ( 0 ) = δ i j g_{ij}(0)=\delta_{ij} g ij ( 0 ) = δ ij at the origin of each normal chart, compactness and smooth dependence imply that, after decreasing r > 0 r>0 r > 0 if necessary,
1 2 I ≤ ( g i j ( y ) ) ≤ 2 I \frac{1}{2}I\le(g_{ij}(y))\le2I 2 1 I ≤ ( g ij ( y )) ≤ 2 I
uniformly for all base points c ∈ K c\in K c ∈ K and all y ∈ B ( 0 , r ) y\in B(0,r) y ∈ B ( 0 , r ) .
Furthermore, the Christoffel symbols vanish at the origin of normal coordinates and their first derivatives with respect to the coordinate variables are uniformly bounded on these coordinate balls. Indeed, cover K K K by finitely many coordinate
::: neighborhoods V 1 , … , V N V_1,\ldots,V_N V 1 , … , V N , each equipped with a smooth orthonormal frame. For each α ∈ { 1 , … , N } \alpha\in\{1,\ldots,N\} α ∈ { 1 , … , N } , choose a compact set K α ⊂ V α K_\alpha\subset V_\alpha K α ⊂ V α such that
K ⊂ ⋃ α = 1 N K α . K \subset\bigcup_{\alpha=1}^{N} K_\alpha. K ⊂ α = 1 ⋃ N K α .
Let F α ( p , y ) : = exp p ( E α ( p ) y ) F_\alpha(p,y) := \exp_p(E_\alpha(p)y) F α ( p , y ) := exp p ( E α ( p ) y ) , where E α ( p ) : R n + 1 → T p M E_\alpha(p): \mathbb{R}^{n+1} \to T_pM E α ( p ) : R n + 1 → T p M is the isometry determined by the chosen orthonormal frame. Since the Riemannian metric is smooth, the maps F α F_\alpha F α and the corresponding metric coefficients g i j ( α ) ( p , y ) g_{ij}^{(\alpha)}(p,y) g ij ( α ) ( p , y ) depend smoothly on ( p , y ) (p,y) ( p , y ) wherever they are defined. Moreover, the matrices g ( α ) ( p , y ) g^{(\alpha)}(p,y) g ( α ) ( p , y ) are positive definite, so their inverses also depend smoothly on ( p , y ) (p,y) ( p , y ) . It follows from the formula
Γ i j ( α ) , k ( p , y ) = 1 2 ∑ ℓ = 1 n + 1 g ( α ) , k ℓ ( p , y ) ( ∂ i g j ℓ ( α ) ( p , y ) + ∂ j g i ℓ ( α ) ( p , y ) − ∂ ℓ g i j ( α ) ( p , y ) ) . \Gamma_{ij}^{(\alpha),k}(p,y) = \frac{1}{2}\sum_{\ell=1}^{n+1} g^{(\alpha),k\ell}(p,y)\left(\partial_i g_{j\ell}^{(\alpha)}(p,y) + \partial_j g_{i\ell}^{(\alpha)}(p,y) - \partial_\ell g_{ij}^{(\alpha)}(p,y)\right). Γ ij ( α ) , k ( p , y ) = 2 1 ℓ = 1 ∑ n + 1 g ( α ) , k ℓ ( p , y ) ( ∂ i g j ℓ ( α ) ( p , y ) + ∂ j g i ℓ ( α ) ( p , y ) − ∂ ℓ g ij ( α ) ( p , y ) ) .
that the Christoffel symbols depend smoothly on ( p , y ) (p,y) ( p , y ) . In particular, the map
( p , y ) ⟼ D y Γ ( α ) ( p , y ) (p,y) \longmapsto D_y\Gamma^{(\alpha)}(p,y) ( p , y ) ⟼ D y Γ ( α ) ( p , y )
is continuous.
After decreasing the radius r > 0 r>0 r > 0 , if necessary, the coordinate maps are defined on a neighborhood of K α × B ( 0 , r ) ‾ K_\alpha\times\overline{B(0,r)} K α × B ( 0 , r ) for every α \alpha α . Since each set K α × B ( 0 , r ) ‾ K_\alpha\times\overline{B(0,r)} K α × B ( 0 , r ) is compact, the continuity of D y Γ ( α ) D_y\Gamma^{(\alpha)} D y Γ ( α ) implies that there exists a constant C α ≥ 0 C_\alpha\ge0 C α ≥ 0 such that
∥ D y Γ ( α ) ( p , y ) ∥ ≤ C α \left\|D_y\Gamma^{(\alpha)}(p,y)\right\| \le C_\alpha D y Γ ( α ) ( p , y ) ≤ C α
for every p ∈ K α p \in K_\alpha p ∈ K α and y ∈ B ( 0 , r ) ‾ y \in\overline{B(0,r)} y ∈ B ( 0 , r ) . Since the cover is finite, setting C : = max 1 ≤ α ≤ N C α C := \max_{1\le\alpha\le N} C_\alpha C := max 1 ≤ α ≤ N C α gives a uniform bound independent of the base point and the chosen coordinate neighborhood. Finally, since the Christoffel symbols vanish at the origin of normal coordinates, we have Γ ( α ) ( p , 0 ) = 0 \Gamma^{(\alpha)}(p,0)=0 Γ ( α ) ( p , 0 ) = 0 . Thus, by the fundamental theorem of calculus, for every p ∈ K α p \in K_\alpha p ∈ K α and y ∈ B ( 0 , r ) y \in B(0,r) y ∈ B ( 0 , r ) ,
Γ ( α ) ( p , y ) = Γ ( α ) ( p , y ) − Γ ( α ) ( p , 0 ) = ∫ 0 1 D y Γ ( α ) ( p , t y ) [ y ] d t . \begin{aligned}
\Gamma^{(\alpha)}(p,y) = \Gamma^{(\alpha)}(p,y) - \Gamma^{(\alpha)}(p,0) \\
&= \int_0^1 D_y\Gamma^{(\alpha)}(p,ty)[y]\,dt.
\end{aligned} Γ ( α ) ( p , y ) = Γ ( α ) ( p , y ) − Γ ( α ) ( p , 0 ) = ∫ 0 1 D y Γ ( α ) ( p , t y ) [ y ] d t .
Consequently,
∥ Γ ( α ) ( p , y ) ∥ ≤ ∫ 0 1 ∥ D y Γ ( α ) ( p , t y ) ∥ ∣ y ∣ d t ≤ C ∣ y ∣ . \begin{aligned}
\left\|\Gamma^{(\alpha)}(p,y)\right\| \le\int_0^1 \left\|D_y\Gamma^{(\alpha)}(p,ty)\right\|\,\lvert y\rvert\,dt \\
&\le C\lvert y\rvert.
\end{aligned} Γ ( α ) ( p , y ) ≤ ∫ 0 1 D y Γ ( α ) ( p , t y ) ∣ y ∣ d t ≤ C ∣ y ∣ .
Taking Λ : = C \Lambda:= C Λ := C , we conclude that
∥ Γ ( α ) ( p , y ) ∥ ≤ Λ ∣ y ∣ \left\|\Gamma^{(\alpha)}(p,y)\right\| \le\Lambda\lvert y\rvert Γ ( α ) ( p , y ) ≤ Λ ∣ y ∣
uniformly for all base points p ∈ K p \in K p ∈ K and all y ∈ B ( 0 , r ) y \in B(0,r) y ∈ B ( 0 , r ) in the corresponding normal coordinates. Hence there exists Λ ≥ 0 \Lambda\ge0 Λ ≥ 0 , independent of c ∈ K c \in K c ∈ K , such that
∥ Γ ( y ) ∥ ≤ Λ ∣ y ∣ for all y ∈ B ( 0 , r ) . \|\Gamma(y)\| \le\Lambda\lvert y\rvert\qquad\text{for all } y \in B(0,r). ∥Γ ( y ) ∥ ≤ Λ ∣ y ∣ for all y ∈ B ( 0 , r ) .
Choose σ 0 > 0 \sigma_0 > 0 σ 0 > 0 sufficiently small that 4 σ 0 < r 4\sigma_0 < r 4 σ 0 < r . Then, for every c ∈ K c \in K c ∈ K ,
exp c : B ( 0 , 4 σ 0 ) ⟶ B 4 σ 0 ( c ) \exp_c:B(0,4\sigma_0)\longrightarrow B_{4\sigma_0}(c) exp c : B ( 0 , 4 σ 0 ) ⟶ B 4 σ 0 ( c )
is a diffeomorphism onto a ball contained in W W W , and the estimates
1 2 I ≤ ( g i j ( y ) ) ≤ 2 I , ∥ Γ ( y ) ∥ ≤ Λ ∣ y ∣ (22) \frac{1}{2}I \leq(g_{ij}(y)) \leq2I,\qquad\|\Gamma(y)\| \leq\Lambda|y| \tag*{(22)} 2 1 I ≤ ( g ij ( y )) ≤ 2 I , ∥Γ ( y ) ∥ ≤ Λ∣ y ∣ (22)
hold uniformly.
Now choose 0 < σ ≤ σ 0 0 < \sigma\leq\sigma_0 0 < σ ≤ σ 0 so small that 14 2 Λ L σ ≤ η 14\sqrt{2}\Lambda L\sigma\leq\eta 14 2 Λ L σ ≤ η . If Λ = 0 \Lambda= 0 Λ = 0 , this restriction is automatic. By compactness of K K K , there exist finitely many points c 1 , … , c m ∈ K c_1,\ldots,c_m \in K c 1 , … , c m ∈ K such that
K ⊂ ⋃ i = 1 m B σ ( c i ) . K \subset\bigcup_{i=1}^{m} B_\sigma(c_i). K ⊂ i = 1 ⋃ m B σ ( c i ) .
Set U i : = B 2 σ ( c i ) U_i := B_{2\sigma}(c_i) U i := B 2 σ ( c i ) , for i = 1 , … , m i=1,\ldots,m i = 1 , … , m . Then
K ⊂ ⋃ i = 1 m U i , K \subset\bigcup_{i=1}^{m} U_i, K ⊂ i = 1 ⋃ m U i ,
and each exponential map exp c i : B ( 0 , 4 σ ) ⟶ B 4 σ ( c i ) \exp_{c_i}:B(0,4\sigma)\longrightarrow B_{4\sigma}(c_i) exp c i : B ( 0 , 4 σ ) ⟶ B 4 σ ( c i ) is a diffeomorphism onto a ball contained in W W W . This proves (i).
Proof of item (ii).
Fix i ∈ { 1 , … , m } i \in\{1,\ldots,m\} i ∈ { 1 , … , m } . For simplicity, write
exp : = exp c i , u : = f ∘ exp , D : = B ( 0 , 4 σ ) . \exp:= \exp_{c_i},\qquad u := f \circ\exp,\qquad D := B(0,4\sigma). exp := exp c i , u := f ∘ exp , D := B ( 0 , 4 σ ) .
Let g j k g_{jk} g j k and Γ j l k \Gamma^k_{jl} Γ j l k denote the metric coefficients and Christoffel symbols in these normal coordinates. The function u = f ∘ exp c i u=f\circ\exp_{c_i} u = f ∘ exp c i is locally Lipschitz on D D D . By Rademacher’s theorem, u u u is differentiable almost everywhere, and its weak Euclidean gradient D u Du D u agrees almost everywhere with its classical gradient. Since f f f is locally L L L -Lipschitz with respect to the Riemannian distance, its weak Riemannian gradient satisfies
∣ ∇ f ∣ g ≤ L |\nabla f|_g \leq L ∣∇ f ∣ g ≤ L
almost everywhere.
At almost every point y ∈ D y \in D y ∈ D , the chain rule gives
D u ( y ) = d ( f ∘ exp c i ) y . Du(y)=d(f\circ\exp_{c_i})_y. D u ( y ) = d ( f ∘ exp c i ) y .
Writing g ( y ) g(y) g ( y ) for the matrix of the metric in normal coordinates and identifying T c i M T_{c_i}M T c i M with R n + 1 \mathbb{R}^{n+1} R n + 1 via an orthonormal basis, we have
∣ ∇ f ∣ g 2 = D u ( y ) T g ( y ) − 1 D u ( y ) . |\nabla f|_g^2=Du(y)^Tg(y)^{-1}Du(y). ∣∇ f ∣ g 2 = D u ( y ) T g ( y ) − 1 D u ( y ) .
The metric bounds in (A.1), imply g ( y ) − 1 ≥ 1 2 I g(y)^{-1}\geq\frac{1}{2}I g ( y ) − 1 ≥ 2 1 I . Consequently,
1 2 ∣ D u ( y ) ∣ 2 ≤ D u ( y ) T g ( y ) − 1 D u ( y ) = ∣ ∇ f ∣ g 2 ≤ L 2 \frac{1}{2}|Du(y)|^2\leq Du(y)^Tg(y)^{-1}Du(y)=|\nabla f|_g^2\leq L^2 2 1 ∣ D u ( y ) ∣ 2 ≤ D u ( y ) T g ( y ) − 1 D u ( y ) = ∣∇ f ∣ g 2 ≤ L 2
for almost every y ∈ D y \in\mathrm{D} y ∈ D . Therefore,
∣ D u ∣ ≤ 2 L almost everywhere in D . (23) |Du| \le\sqrt{2}L \quad\text{almost everywhere in D}. \tag*{(23)} ∣ D u ∣ ≤ 2 L almost everywhere in D . (23)
We first prove a distributional upper bound for the Euclidean Hessian of u u u . Fix y 0 ∈ D y_0 \in\mathrm{D} y 0 ∈ D and put x 0 = exp ( y 0 ) x_0 = \exp(y_0) x 0 = exp ( y 0 ) . Given ε > 0 \varepsilon> 0 ε > 0 , the support-sense assumption on Hess f \operatorname{Hess} f Hess f provides a C 2 C^2 C 2 upper support function ϕ \phi ϕ for f f f at x 0 x_0 x 0 such that
ϕ ≥ f , ϕ ( x 0 ) = f ( x 0 ) , Hess ϕ ( x 0 ) ≤ ( a + ε ) g x 0 . \phi\ge f,\qquad\phi(x_0) = f(x_0),\qquad\operatorname{Hess}\phi(x_0) \le(a+\varepsilon)g_{x_0}. ϕ ≥ f , ϕ ( x 0 ) = f ( x 0 ) , Hess ϕ ( x 0 ) ≤ ( a + ε ) g x 0 .
Set v : = ϕ ∘ exp v := \phi\circ\exp v := ϕ ∘ exp . Then v v v is an upper support function for u u u at y 0 y_0 y 0 . By Lemma A.1,
∣ ∇ ϕ ( x 0 ) ∣ ≤ L , |\nabla\phi(x_0)| \le L, ∣∇ ϕ ( x 0 ) ∣ ≤ L ,
and therefore
∣ D v ( y 0 ) ∣ ≤ 2 L . |Dv(y_0)| \le\sqrt{2}L. ∣ D v ( y 0 ) ∣ ≤ 2 L .
The coordinate expression for the Riemannian Hessian is
Hess ϕ ( ∂ j , ∂ l ) = ∂ j l v − Γ j l k ∂ k v . \operatorname{Hess}\phi(\partial_j,\partial_l) = \partial_{jl}v-\Gamma^k_{jl}\partial_kv. Hess ϕ ( ∂ j , ∂ l ) = ∂ j l v − Γ j l k ∂ k v .
Consequently, for every ξ ∈ R n + 1 \xi\in\mathbb{R}^{n+1} ξ ∈ R n + 1 ,
D 2 v ( y 0 ) ( ξ , ξ ) = Hess ϕ ( x 0 ) ( ξ , ξ ) + Γ j l k ( y 0 ) ξ j ξ l ∂ k v ( y 0 ) ≤ ( a + ε ) g y 0 ( ξ , ξ ) + ∥ Γ ( y 0 ) ∥ ∣ ξ ∣ 2 ∣ D v ( y 0 ) ∣ ≤ a g y 0 ( ξ , ξ ) + 2 ε ∣ ξ ∣ 2 + 4 2 Λ L σ ∣ ξ ∣ 2 . \begin{aligned}
D^2v(y_0)(\xi,\xi)=\operatorname{Hess}\phi(x_0)(\xi,\xi)+\Gamma^k_{jl}(y_0)\xi^j\xi^l\partial_kv(y_0) \\
&\le(a+\varepsilon)g_{y_0}(\xi,\xi)+\|\Gamma(y_0)\||\xi|^2|Dv(y_0)| \\
&\le ag_{y_0}(\xi,\xi)+2\varepsilon|\xi|^2+4\sqrt{2}\Lambda L\sigma|\xi|^2.
\end{aligned} D 2 v ( y 0 ) ( ξ , ξ ) = Hess ϕ ( x 0 ) ( ξ , ξ ) + Γ j l k ( y 0 ) ξ j ξ l ∂ k v ( y 0 ) ≤ ( a + ε ) g y 0 ( ξ , ξ ) + ∥Γ ( y 0 ) ∥∣ ξ ∣ 2 ∣ D v ( y 0 ) ∣ ≤ a g y 0 ( ξ , ξ ) + 2 ε ∣ ξ ∣ 2 + 4 2 Λ L σ ∣ ξ ∣ 2 .
Here we used g ≤ 2 I g \le2I g ≤ 2 I , ∣ y 0 ∣ < 4 σ |y_0|<4\sigma ∣ y 0 ∣ < 4 σ , and ∣ D v ( y 0 ) ∣ ≤ 2 L |Dv(y_0)|\le\sqrt{2}L ∣ D v ( y 0 ) ∣ ≤ 2 L .
Define the continuous symmetric matrix-valued function
B ( y ) : = a g ( y ) + 4 2 Λ L σ I . B(y):=ag(y)+4\sqrt{2}\Lambda L\sigma I. B ( y ) := a g ( y ) + 4 2 Λ L σ I .
The preceding estimate implies that, for every ξ ∈ R n + 1 \xi\in\mathbb{R}^{n+1} ξ ∈ R n + 1 and every ε > 0 \varepsilon>0 ε > 0 ,
D 2 v ( y 0 ) ( ξ , ξ ) ≤ ξ T B ( y 0 ) ξ + 2 ε ∣ ξ ∣ 2 . D^2v(y_0)(\xi,\xi)\le\xi^{\mathsf{T}}B(y_0)\xi+2\varepsilon|\xi|^2. D 2 v ( y 0 ) ( ξ , ξ ) ≤ ξ T B ( y 0 ) ξ + 2 ε ∣ ξ ∣ 2 .
We now apply Lemma 2.10 to the restrictions of u u u to straight lines. Fix a unit vector ξ \xi ξ and write
q ( t ) : = u ( z + t ξ ) , z ∈ ξ ⊥ . q(t):=u(z+t\xi),\qquad z\in\xi^\perp. q ( t ) := u ( z + t ξ ) , z ∈ ξ ⊥ .
For each point t 0 t_0 t 0 in the interval where the line lies in D \mathrm{D} D , the preceding support inequality shows that q q q admits C 2 C^2 C 2 upper support functions whose second derivatives at t 0 t_0 t 0 are bounded above by
β ξ ( t 0 ) + ρ , β ξ ( t ) : = ξ T B ( z + t ξ ) ξ , \beta_\xi(t_0)+\rho,\qquad\beta_\xi(t):=\xi^{\mathsf{T}}B(z+t\xi)\xi, β ξ ( t 0 ) + ρ , β ξ ( t ) := ξ T B ( z + t ξ ) ξ ,
for every ρ > 0 \rho>0 ρ > 0 . Lemma 2.10 therefore yields
∫ J q ( t ) θ ′ ′ ( t ) d t ≤ ∫ J β ξ ( t ) θ ( t ) d t \int_J q(t)\theta''(t)\,\mathrm{d}t\le\int_J\beta_\xi(t)\theta(t)\,\mathrm{d}t ∫ J q ( t ) θ ′′ ( t ) d t ≤ ∫ J β ξ ( t ) θ ( t ) d t
for every nonnegative $\theta\in C^\infty_c(J).
Integrating over the orthogonal complement ξ ⊥ \xi^\perp ξ ⊥ and applying Fubini’s theorem, we conclude that
∫ D u ∂ ξ 2 ζ d y ≤ ∫ D ( ξ T B ( y ) ξ ) ζ ( y ) d y (24) \int_D u\,\partial_{\xi}^{2}\zeta\,dy \leq\int_D \left(\xi^{\mathsf{T}}\mathbf{B}(y)\xi\right)\zeta(y)\,dy \tag*{(24)} ∫ D u ∂ ξ 2 ζ d y ≤ ∫ D ( ξ T B ( y ) ξ ) ζ ( y ) d y (24)
for every nonnegative ζ ∈ C c ∞ ( D ) \zeta\in C^\infty_c(\mathrm{D}) ζ ∈ C c ∞ ( D ) . By homogeneity, this inequality holds for every ξ ∈ R n + 1 \xi\in\mathbb{R}^{n+1} ξ ∈ R n + 1 .
Let μ ∈ C c ∞ ( R n + 1 ) \mu\in C^\infty_c(\mathbb{R}^{n+1}) μ ∈ C c ∞ ( R n + 1 ) be a nonnegative mollifier supported in the unit ball, with ∫ μ = 1 \int\mu=1 ∫ μ = 1 , and set
μ δ ( y ) : = δ − ( n + 1 ) μ ( y / δ ) , u δ : = u ∗ μ δ . \mu_\delta(y):=\delta^{-(n+1)}\mu(y/\delta), \qquad u_\delta:=u\ast\mu_\delta. μ δ ( y ) := δ − ( n + 1 ) μ ( y / δ ) , u δ := u ∗ μ δ .
For 0 < δ < σ 0<\delta<\sigma 0 < δ < σ , the function u δ u_\delta u δ is smooth on B ( 0 , 3 σ ) B(0,3\sigma) B ( 0 , 3 σ ) . Fix y ∈ B ( 0 , 3 σ ) y\in B(0,3\sigma) y ∈ B ( 0 , 3 σ ) . Applying (A.3) to the test function z ↦ μ δ ( y − z ) z\mapsto\mu_\delta(y-z) z ↦ μ δ ( y − z ) gives
D 2 u δ ( y ) ( ξ , ξ ) = ∫ D u ( z ) ∂ ξ 2 [ μ δ ( y − z ) ] d z ≤ ∫ D ( ξ T B ( z ) ξ ) μ δ ( y − z ) d z = ξ T ( B ∗ μ δ ) ( y ) ξ . \begin{aligned}
D^2u_\delta(y)(\xi,\xi)=\int_D u(z)\,\partial_\xi^2[\mu_\delta(y-z)]\,dz \\
&\leq\int_D\left(\xi^{\mathsf{T}}\mathbf{B}(z)\xi\right)\mu_\delta(y-z)\,dz \\
&=\xi^{\mathsf{T}}(\mathbf{B}\ast\mu_\delta)(y)\xi.
\end{aligned} D 2 u δ ( y ) ( ξ , ξ ) = ∫ D u ( z ) ∂ ξ 2 [ μ δ ( y − z )] d z ≤ ∫ D ( ξ T B ( z ) ξ ) μ δ ( y − z ) d z = ξ T ( B ∗ μ δ ) ( y ) ξ .
Since u u u is locally Lipschitz, it is Lipschitz on B ( 0 , 3 σ ) ‾ \overline{B(0,3\sigma)} B ( 0 , 3 σ ) , with some Lipschitz constant C > 0 C>0 C > 0 . For y ∈ B ( 0 , 2 σ ) y\in B(0,2\sigma) y ∈ B ( 0 , 2 σ ) and 0 < δ < σ 0<\delta<\sigma 0 < δ < σ , the support condition supp ( μ δ ) ⊂ B ( 0 , δ ) \operatorname{supp}(\mu_\delta)\subset B(0,\delta) supp ( μ δ ) ⊂ B ( 0 , δ ) ensures that y − z ∈ B ( 0 , 3 σ ) y-z\in B(0,3\sigma) y − z ∈ B ( 0 , 3 σ ) whenever z ∈ supp ( μ δ ) z\in\operatorname{supp}(\mu_\delta) z ∈ supp ( μ δ ) . Since ∫ R n + 1 μ δ ( z ) d z = 1 \int_{\mathbb{R}^{n+1}}\mu_\delta(z)\,dz=1 ∫ R n + 1 μ δ ( z ) d z = 1 , we obtain
∣ u δ ( y ) − u ( y ) ∣ = ∣ ∫ R n + 1 ( u ( y − z ) − u ( y ) ) μ δ ( z ) d z ∣ ≤ C ∫ R n + 1 ∣ z ∣ μ δ ( z ) d z = C δ ∫ R n + 1 ∣ w ∣ μ ( w ) d w . \begin{aligned}
\lvert u_\delta(y)-u(y)\rvert=\left\lvert\int_{\mathbb{R}^{n+1}}\left(u(y-z)-u(y)\right)\mu_\delta(z)\,dz\right\rvert\\
&\leq C\int_{\mathbb{R}^{n+1}}\lvert z\rvert\mu_\delta(z)\,dz \\
&=C\delta\int_{\mathbb{R}^{n+1}}\lvert w\rvert\mu(w)\,dw.
\end{aligned} ∣ u δ ( y ) − u ( y )∣ = ∫ R n + 1 ( u ( y − z ) − u ( y ) ) μ δ ( z ) d z ≤ C ∫ R n + 1 ∣ z ∣ μ δ ( z ) d z = C δ ∫ R n + 1 ∣ w ∣ μ ( w ) d w .
The right-hand side is independent of y y y and tends to zero as δ → 0 \delta\to0 δ → 0 . Hence
u δ ⟶ u uniformly on B ( 0 , 2 σ ) ‾ . u_\delta\longrightarrow u\quad\text{uniformly on }\overline{B(0,2\sigma)}. u δ ⟶ u uniformly on B ( 0 , 2 σ ) .
Since u ∈ W l o c 1 , ∞ ( D ) u\in W^{1,\infty}_{\mathrm{loc}}(\mathrm{D}) u ∈ W loc 1 , ∞ ( D ) , its weak gradient D u Du D u belongs to L l o c 1 ( D ; R n + 1 ) L^1_{\mathrm{loc}}(\mathrm{D};\mathbb{R}^{n+1}) L loc 1 ( D ; R n + 1 ) . For every compact subset of D \mathrm{D} D and all sufficiently small δ > 0 \delta>0 δ > 0 , differentiation of convolutions gives
D u δ = ( D u ) ∗ μ δ . Du_\delta=(Du)\ast\mu_\delta. D u δ = ( D u ) ∗ μ δ .
Consequently,
D u δ ( y ) = ∫ R n + 1 D u ( y − δ z ) μ ( z ) d z . Du_\delta(y)=\int_{\mathbb{R}^{n+1}}Du(y-\delta z)\mu(z)\,dz. D u δ ( y ) = ∫ R n + 1 D u ( y − δ z ) μ ( z ) d z .
In particular, by (A.2) and the nonnegativity of μ δ \mu_\delta μ δ , ∣ D u δ ( y ) ∣ ≤ 2 L \lvert Du_\delta(y)\rvert\leq\sqrt{2}L ∣ D u δ ( y )∣ ≤ 2 L .
By the Lebesgue differentiation theorem, almost every y ∈ B ( 0 , 2 σ ) y \in B(0,2\sigma) y ∈ B ( 0 , 2 σ ) is a Lebesgue point of D u Du D u . Fix such a point y y y . Since μ ≥ 0 \mu\ge0 μ ≥ 0 , supp μ ⊂ B ( 0 , 1 ) \operatorname{supp}\mu\subset B(0,1) supp μ ⊂ B ( 0 , 1 ) , and μ ∈ L ∞ ( R n + 1 ) \mu\in L^\infty(\mathbb{R}^{n+1}) μ ∈ L ∞ ( R n + 1 ) , we obtain
∣ D u δ ( y ) − D u ( y ) ∣ ≤ ∥ μ ∥ L ∞ ∫ B ( 0 , 1 ) ∣ D u ( y − δ z ) − D u ( y ) ∣ d z = ∥ μ ∥ L ∞ δ − ( n + 1 ) ∫ B ( y , δ ) ∣ D u ( x ) − D u ( y ) ∣ d x = ∥ μ ∥ L ∞ ∣ B ( 0 , 1 ) ∣ 1 ∣ B ( y , δ ) ∣ ∫ B ( y , δ ) ∣ D u ( x ) − D u ( y ) ∣ d x . \begin{aligned}
|Du_\delta(y)-Du(y)| &\le\|\mu\|_{L^\infty}\int_{B(0,1)}|Du(y-\delta z)-Du(y)|\,dz\\
&=\|\mu\|_{L^\infty}\delta^{-(n+1)}\int_{B(y,\delta)}|Du(x)-Du(y)|\,dx\\
&=\|\mu\|_{L^\infty}|B(0,1)|\frac{1}{|B(y,\delta)|}\int_{B(y,\delta)}|Du(x)-Du(y)|\,dx.
\end{aligned} ∣ D u δ ( y ) − D u ( y ) ∣ ≤ ∥ μ ∥ L ∞ ∫ B ( 0 , 1 ) ∣ D u ( y − δ z ) − D u ( y ) ∣ d z = ∥ μ ∥ L ∞ δ − ( n + 1 ) ∫ B ( y , δ ) ∣ D u ( x ) − D u ( y ) ∣ d x = ∥ μ ∥ L ∞ ∣ B ( 0 , 1 ) ∣ ∣ B ( y , δ ) ∣ 1 ∫ B ( y , δ ) ∣ D u ( x ) − D u ( y ) ∣ d x .
The last expression tends to zero as δ → 0 + \delta\to0^+ δ → 0 + , by the defining property of a Lebesgue point. Hence
D u δ ( y ) ⟶ D u ( y ) for almost every y ∈ B ( 0 , 2 σ ) . Du_\delta(y)\longrightarrow Du(y)\quad\text{for almost every }y\in B(0,2\sigma). D u δ ( y ) ⟶ D u ( y ) for almost every y ∈ B ( 0 , 2 σ ) .
Define f i , δ : = u δ ∘ exp c i − 1 f_{i,\delta}:=u_\delta\circ\exp_{c_i}^{-1} f i , δ := u δ ∘ exp c i − 1 on B 3 σ ( c i ) B_{3\sigma}(c_i) B 3 σ ( c i ) . Then f i , δ f_{i,\delta} f i , δ is smooth. Its Riemannian gradient satisfies
∣ ∇ f i , δ ∣ 2 = ( D u δ ) T g − 1 D u δ ≤ 2 ∣ D u δ ∣ 2 ≤ 4 L 2 . |\nabla f_{i,\delta}|^2=(Du_\delta)^{\mathsf T}g^{-1}Du_\delta\le2|Du_\delta|^2\le4L^2. ∣∇ f i , δ ∣ 2 = ( D u δ ) T g − 1 D u δ ≤ 2∣ D u δ ∣ 2 ≤ 4 L 2 .
Thus
∣ ∇ f i , δ ∣ ≤ 2 L on B 3 σ ( c i ) . |\nabla f_{i,\delta}|\le2L\quad\text{on }B_{3\sigma}(c_i). ∣∇ f i , δ ∣ ≤ 2 L on B 3 σ ( c i ) .
The smooth dependence of the metric coefficients and the almost-everywhere convergence of D u δ Du_\delta D u δ imply
∇ f i , δ ⟶ ∇ f almost everywhere on U i . \nabla f_{i,\delta}\longrightarrow\nabla f\quad\text{almost everywhere on }U_i. ∇ f i , δ ⟶ ∇ f almost everywhere on U i .
Finally, the uniform convergence of u δ u_\delta u δ gives
f i , δ ⟶ f uniformly on U i ‾ . f_{i,\delta}\longrightarrow f\quad\text{uniformly on }\overline{U_i}. f i , δ ⟶ f uniformly on U i .
This proves (ii).
Proof of item (iii). We continue to use the normal coordinates centered at c i c_i c i . From the distributional Hessian estimate established above,
D 2 u δ ( y ) ( ξ , ξ ) ≤ a ( g ∗ μ δ ) y ( ξ , ξ ) + 4 2 Λ L σ ∣ ξ ∣ 2 . D^2u_\delta(y)(\xi,\xi)\le a\,(g*\mu_\delta)_y(\xi,\xi)+4\sqrt{2}\,\Lambda L\sigma|\xi|^2. D 2 u δ ( y ) ( ξ , ξ ) ≤ a ( g ∗ μ δ ) y ( ξ , ξ ) + 4 2 Λ L σ ∣ ξ ∣ 2 .
On the other hand, the coordinate formula for the Riemannian Hessian gives
Hess f i , δ ( ξ , ξ ) = D 2 u δ ( y ) ( ξ , ξ ) − Γ j l k ( y ) ξ j ξ l ∂ k u δ ( y ) ≤ a ( g ∗ μ δ ) y ( ξ , ξ ) + 4 2 Λ L σ ∣ ξ ∣ 2 + ∥ Γ ( y ) ∥ ∣ ξ ∣ 2 ∣ D u δ ( y ) ∣ . \begin{aligned}
\operatorname{Hess} f_{i,\delta}(\xi,\xi)
&=D^2u_\delta(y)(\xi,\xi)-\Gamma^k_{jl}(y)\xi^j\xi^l\partial_k u_\delta(y)\\
&\le a\,(g*\mu_\delta)_y(\xi,\xi)+4\sqrt{2}\,\Lambda L\sigma|\xi|^2+\|\Gamma(y)\||\xi|^2|Du_\delta(y)|.
\end{aligned} Hess f i , δ ( ξ , ξ ) = D 2 u δ ( y ) ( ξ , ξ ) − Γ j l k ( y ) ξ j ξ l ∂ k u δ ( y ) ≤ a ( g ∗ μ δ ) y ( ξ , ξ ) + 4 2 Λ L σ ∣ ξ ∣ 2 + ∥Γ ( y ) ∥∣ ξ ∣ 2 ∣ D u δ ( y ) ∣.
For y ∈ B ( 0 , 2 σ ) y\in B(0,2\sigma) y ∈ B ( 0 , 2 σ ) , we have
∥ Γ ( y ) ∥ ≤ 2 Λ σ . \|\Gamma(y)\|\le2\Lambda\sigma. ∥Γ ( y ) ∥ ≤ 2Λ σ .
Using ∣ D u δ ∣ ≤ 2 L |Du_\delta|\le\sqrt{2}L ∣ D u δ ∣ ≤ 2 L , we obtain
Hess f i , δ ( ξ , ξ ) ≤ a ( g ∗ μ δ ) y ( ξ , ξ ) + 6 2 Λ L σ ∣ ξ ∣ 2 . (25) \operatorname{Hess} f_{i,\delta}(\xi,\xi)\le a\,(g*\mu_\delta)_y(\xi,\xi)+6\sqrt{2}\,\Lambda L\sigma|\xi|^2. \tag*{(25)} Hess f i , δ ( ξ , ξ ) ≤ a ( g ∗ μ δ ) y ( ξ , ξ ) + 6 2 Λ L σ ∣ ξ ∣ 2 . (25)
Suppose that g y ( ξ , ξ ) = 1 g_y(\xi,\xi)=1 g y ( ξ , ξ ) = 1 . Since g y ≥ 1 2 I g_y \ge\frac{1}{2}I g y ≥ 2 1 I , we have ∣ ξ ∣ 2 ≤ 2 |\xi|^2 \le2 ∣ ξ ∣ 2 ≤ 2 . Consequently,
Hess f i , δ ( ξ , ξ ) ≤ a + ∣ a ∣ ∣ ( ( g ∗ μ δ ) y − g y ) ( ξ , ξ ) ∣ + 12 2 Λ L σ ≤ a + 2 ∣ a ∣ sup B ( 0 , 2 σ ) ‾ ∥ g ∗ μ δ − g ∥ + 12 2 Λ L σ . \begin{aligned}
\operatorname{Hess} f_{i,\delta}(\xi,\xi) &\le a+|a|\left|((g*\mu_\delta)_y-g_y)(\xi,\xi)\right|+12\sqrt{2}\Lambda L\sigma\\
&\le a+2|a|\sup_{\overline{B(0,2\sigma)}}\|g*\mu_\delta-g\|+12\sqrt{2}\Lambda L\sigma.
\end{aligned} Hess f i , δ ( ξ , ξ ) ≤ a + ∣ a ∣ ∣ (( g ∗ μ δ ) y − g y ) ( ξ , ξ ) ∣ + 12 2 Λ L σ ≤ a + 2∣ a ∣ B ( 0 , 2 σ ) sup ∥ g ∗ μ δ − g ∥ + 12 2 Λ L σ .
Here the first inequality follows by writing
a ( g ∗ μ δ ) y ( ξ , ξ ) = a g y ( ξ , ξ ) + a ( ( g ∗ μ δ ) y − g y ) ( ξ , ξ ) a(g*\mu_\delta)_y(\xi,\xi)=a g_y(\xi,\xi)+a((g*\mu_\delta)_y-g_y)(\xi,\xi) a ( g ∗ μ δ ) y ( ξ , ξ ) = a g y ( ξ , ξ ) + a (( g ∗ μ δ ) y − g y ) ( ξ , ξ )
and using g y ( ξ , ξ ) = 1 g_y(\xi,\xi)=1 g y ( ξ , ξ ) = 1 . Since the metric coefficients are uniformly continuous on a compact neighborhood of B ( 0 , 2 σ ) ‾ \overline{B(0,2\sigma)} B ( 0 , 2 σ ) , the approximate identity property implies that
sup B ( 0 , 2 σ ) ‾ ∥ g ∗ μ δ − g ∥ ⟶ 0 as δ → 0. \sup_{\overline{B(0,2\sigma)}}\|g*\mu_\delta-g\|\longrightarrow0 \qquad\text{as }\delta\to0. B ( 0 , 2 σ ) sup ∥ g ∗ μ δ − g ∥ ⟶ 0 as δ → 0.
Define ω i ( δ ) : = 2 ∣ a ∣ sup B ( 0 , 2 σ ) ‾ ∥ g ∗ μ δ − g ∥ \omega_i(\delta):=2|a|\sup_{\overline{B(0,2\sigma)}}\|g*\mu_\delta-g\| ω i ( δ ) := 2∣ a ∣ sup B ( 0 , 2 σ ) ∥ g ∗ μ δ − g ∥ . Then ω i ( δ ) → 0 \omega_i(\delta)\to0 ω i ( δ ) → 0 as δ → 0 \delta\to0 δ → 0 . Moreover, our choice of σ \sigma σ ensures that 12 2 Λ L σ ≤ 14 2 Λ L σ ≤ η 12\sqrt{2}\Lambda L\sigma\le14\sqrt{2}\Lambda L\sigma\le\eta 12 2 Λ L σ ≤ 14 2 Λ L σ ≤ η . It follows that
Hess f i , δ ( ξ , ξ ) ≤ a + η + ω i ( δ ) \operatorname{Hess} f_{i,\delta}(\xi,\xi)\le a+\eta+\omega_i(\delta) Hess f i , δ ( ξ , ξ ) ≤ a + η + ω i ( δ )
whenever g y ( ξ , ξ ) = 1 g_y(\xi,\xi)=1 g y ( ξ , ξ ) = 1 . By homogeneity, we conclude that
Hess f i , δ ≤ ( a + η + ω i ( δ ) ) g \operatorname{Hess} f_{i,\delta}\le(a+\eta+\omega_i(\delta))g Hess f i , δ ≤ ( a + η + ω i ( δ )) g
on U i U_i U i . Since f i , δ f_{i,\delta} f i , δ and the metric are smooth on a neighborhood of U i ‾ \overline{U_i} U i , the inequality extends to U i ‾ \overline{U_i} U i by continuity. This proves (iii) and completes the proof. □
Acknowledgments: During the preparation of this work, the authors used Claude, Gemini, and ChatGPT for refining arguments, locating references, and drafting and revising the manuscript. All AI-generated suggestions were independently verified by the authors, who take full responsibility for the content and accuracy of the final manuscript.
References
José Edson Sampaio: Departamento de Matemática, Universidade Federal do Ceará, Rua Campus do Pici, s/n, Bloco 914, Pici, 60440-900, Fortaleza-CE, Brazil. E-mail: edsonsampaio@mat.ufc.br
Euripedes Carvalho da Silva: Departamento de Matemática, Instituto Federal de Educação, Ciência e Tecnologia do Ceará, Av. Parque Central, 1315, Distrito Industrial I, 61939-140, Maracanaú-CE, Brazil. and Departamento de Matemática, Universidade Federal do Ceará, Rua Campus do Pici, s/n, Bloco 914, Pici, 60440-900, Fortaleza-CE, Brazil. E-mail: euripedes.carvalho@ifce.edu.br