Introduction
Rigidity phenomena for compact Riemannian manifolds with boundary have been studied extensively from both geometric and analytic points of view. A fundamental result in this direction is a theorem by Hang and Wang, [1 ] . In one of its standard forms, their theorems asserts that if ( M n , g ) (M^n,g) ( M n , g ) is a compact Riemannian manifold with smooth boundary such that
Ric g ≥ ( n − 1 ) g , ( ∂ M , g ∣ T ∂ M ) ≅ S n − 1 , h ≥ 0 , \operatorname{Ric}_g \ge(n-1)g,\qquad(\partial M,g|_{T\partial M}) \cong S^{n-1},\qquad h \ge0, Ric g ≥ ( n − 1 ) g , ( ∂ M , g ∣ T ∂ M ) ≅ S n − 1 , h ≥ 0 ,
where h h h denotes the second fundamental form of ∂ M \partial M ∂ M with respect to the outward unit normal and the convention h ( X , Y ) = g ( ∇ X M ν , Y ) h(X,Y)=g(\nabla^{M}_{X}\nu,Y) h ( X , Y ) = g ( ∇ X M ν , Y ) , then ( M , g ) (M,g) ( M , g ) is isometric to the standard hemisphere S + n \mathbb{S}^{n}_{+} S + n . Generalizations to spherical domains bounded by Clifford tori appear in [3 ] .
The prescribed boundary metric and second fundamental form can also be viewed as geometric analogues of Dirichlet and Neumann data. This viewpoint connects hemisphere rigidity with overdetermined boundary problems, such as Serrin’s symmetry theorem [Ser71], and with rigidity results for the Obata equation ∇ 2 f + f g = 0 \nabla^{2}f+fg=0 ∇ 2 f + f g = 0 ; see [5 Esc90, Xia91, CLW21] . We ask whether the stronger interior condition sec g ≥ 1 \sec_{g}\geq1 sec g ≥ 1 allows one to replace boundary convexity by a condition that does not prescribe the signs of the principal curvatures.
Grove and Petersen [GP22] studied the related problem for Alexandrov spaces of curvature at least 1. They prove that maximal boundary volume, vol ( ∂ M ) = vol ( S n − 1 ) \operatorname{vol}(\partial M)=\operatorname{vol}(\mathbb{S}^{n-1}) vol ( ∂ M ) = vol ( S n − 1 ) , forces the space to be a spherical lens L α n L_{\alpha}^{n} L α n , bounded by two hemispheres meeting at an angle α ∈ ( 0 , π ] \alpha\in(0,\pi] α ∈ ( 0 , π ] . This does not directly resolve the smooth Riemannian problem: without local convexity of the boundary, a lower sectional curvature bound need not make the intrinsic distance space an Alexandrov space with the same lower curvature bound.
Our additional boundary assumption is the Codazzi equation. After identifying the boundary with the unit sphere, it reads
( ∇ X h ) ( Y , Z ) = ( ∇ Y h ) ( X , Z ) ( X , Y , Z ∈ T S n − 1 ) . (\nabla_{X}h)(Y,Z)=(\nabla_{Y}h)(X,Z)\qquad(X,Y,Z\in T\mathbb{S}^{n-1}). ( ∇ X h ) ( Y , Z ) = ( ∇ Y h ) ( X , Z ) ( X , Y , Z ∈ T S n − 1 ) .
Together with the Gauss equation, this condition gives a boundary problem on the round sphere that can be studied independently of the interior metric. The resulting rigidity holds in dimensions at least four.
Theorem A. Let ( M n , g ) (M^{n},g) ( M n , g ) , n ≥ 4 n\geq4 n ≥ 4 , be a compact connected Riemannian manifold with smooth boundary. Assume that
sec g ≥ 1 \sec_{g}\geq1 sec g ≥ 1 ;
( ∂ M , g ∣ T ∂ M ) (\partial M,g|_{T\partial M}) ( ∂ M , g ∣ T ∂ M ) is isometric to the standard unit sphere S n − 1 \mathbb{S}^{n-1} S n − 1 ;
the second fundamental form h h h of ∂ M \partial M ∂ M is a Codazzi tensor.
Then ( M , g ) (M,g) ( M , g ) is isometric to the standard hemisphere S + n \mathbb{S}^{n}_{+} S + n .
Dimension three admits a different behavior. Panina’s construction [Pan05] provides nonzero Codazzi tensors on S 2 S^2 S 2 satisfying the sign condition required by the Gauss equation. We show that these boundary data can be realized by metrics whose sectional curvatures remain at least 1 throughout the interior.
Theorem B. There exists a compact Riemannian manifold ( M 3 , g ) (M^3,g) ( M 3 , g ) , diffeomorphic to the closed three-dimensional ball, such that
sec g ≥ 1 , ( ∂ M , g ∣ T ∂ M ) ≅ S 2 , \operatorname{sec}_g \ge1,\qquad(\partial M,g|_{T\partial M}) \cong S^2, sec g ≥ 1 , ( ∂ M , g ∣ T ∂ M ) ≅ S 2 , and the second fundamental form of ∂ M \partial M ∂ M is a nonzero Codazzi tensor. In particular, ( M , g ) (M,g) ( M , g ) is not isometric to the standard hemisphere.
Theorem B is related to the study of saddle spheres in the standard S 3 S^3 S 3 . Gálvez, Mira, and Tassi [GMT24] prove that every real analytic immersed saddle two-sphere in S 3 S^3 S 3 is equatorial, and show that this rigidity fails in the smooth category. In their setting, the ambient metric is round, while the induced metric on the sphere is allowed to vary. Our construction fixes the induced boundary metric to be the standard metric on S 2 S^2 S 2 and realizes nontrivial saddle Codazzi tensors, after scaling, as its second fundamental form by deforming the ambient metric while maintaining sec g ≥ 1 \operatorname{sec}_g \ge1 sec g ≥ 1 .
Although these boundary conditions do not imply rigidity in dimension three, they give a sharp bound for the distance from an interior point to the boundary. Equality restores rigidity.
Theorem C. Let ( M 3 , g ) (M^3,g) ( M 3 , g ) be a compact connected Riemannian manifold with smooth boundary. Assume that
sec g ≥ 1 \operatorname{sec}_g \ge1 sec g ≥ 1 ;
( ∂ M , g ∣ T ∂ M ) (\partial M,g|_{T\partial M}) ( ∂ M , g ∣ T ∂ M ) is isometric to the standard unit sphere S 2 S^2 S 2 ;
the second fundamental form h h h of ∂ M \partial M ∂ M is a Codazzi tensor.
Then
r i n ( M ) : = max p ∈ M d g ( p , ∂ M ) ≤ π 2 , r_{\mathrm{in}}(M) := \max_{p\in M} d_g(p,\partial M) \le\frac{\pi}{2}, r in ( M ) := p ∈ M max d g ( p , ∂ M ) ≤ 2 π , with equality if and only if ( M , g ) (M,g) ( M , g ) is isometric to the standard hemisphere S + 3 S^3_+ S + 3 .
The common geometric mechanism behind our proofs is the study of saddle Codazzi tensors on round spheres. These are symmetric Codazzi tensors whose restrictions to all tangent two-planes have nonpositive determinant. Put m = n − 1 m=n-1 m = n − 1 and identify a symmetric boundary tensor with its self-adjoint endomorphism using the round metric. Every Codazzi tensor on S m S^m S m , m ≥ 2 m\ge2 m ≥ 2 , is of the form A f = ∇ 2 f + f g S m A_f=\nabla^2f+fg_{S^m} A f = ∇ 2 f + f g S m . When the boundary is round and sec g ≥ 1 \operatorname{sec}_g\ge1 sec g ≥ 1 , the Gauss equation shows that its second fundamental form satisfies precisely the saddle condition. The boundary geometry thus leads naturally to the question of whether nontrivial saddle Codazzi tensors can exist on S m S^m S m . Pointwise, the saddle condition allows at most one positive and one negative eigenvalue, so rank A f ≤ 2 A_f \le2 A f ≤ 2 . This rank restriction is the starting point of our analysis of the associated support map.
The potential f f f determines a map into Euclidean space,
F f : S m ⟶ R m + 1 , F f ( x ) = ∇ f ( x ) + f ( x ) x , d F f = A f . F_f:S^m\longrightarrow\mathbb{R}^{m+1},\qquad F_f(x)=\nabla f(x)+f(x)x,\qquad dF_f=A_f. F f : S m ⟶ R m + 1 , F f ( x ) = ∇ f ( x ) + f ( x ) x , d F f = A f .
It parametrizes the envelope of the planes ⟨ y , x ⟩ = f ( x ) \langle y,x\rangle=f(x) ⟨ y , x ⟩ = f ( x ) , with singularities allowed. On a sufficiently small constant-rank neighborhood its image is a submanifold, and x x x is a unit normal to that image. If r = max rank A f > 0 r=\max\operatorname{rank} A_f>0 r = max rank A f > 0 and m ≥ 3 m\ge3 m ≥ 3 , then r ≤ 2 < m r\le2<m r ≤ 2 < m . A suitable linear projection of F f F_f F f has a compact regular fiber component that is a full unit normal sphere of a local image. This sphere has dimension m − r ≥ 1 m-r\ge1 m − r ≥ 1 , so it contains a complete great circle on which F f F_f F f is constant.
Such a circle gives a global obstruction. Restrict f f f to any great 2-sphere containing the circle and subtract the linear function determined by the constant value of F f F_f F f . The resulting function u u u and its first derivatives vanish along the circle. Vanishing of u u u alone makes the orthogonal projection of F u F_u F u onto the plane of the circle zero there. Stokes’ theorem gives zero projected signed area on each hemisphere, so the saddle condition forces det A u = 0 \det A_u=0 det A u = 0 . An energy identity then realizes equality in the first Dirichlet eigenvalue inequality on each hemisphere and shows that u u u is linear. Its vanishing first derivatives along the circle now give u = 0 u=0 u = 0 . Varying the great 2-sphere gives A f = 0 A_f=0 A f = 0 on S m S^m S m . Thus h = 0 h=0 h = 0 in every dimension n ≥ 4 n\ge4 n ≥ 4 , and Hang–Wang rigidity applies.
The same map also explains the boundary data in dimension three. Its signed area vector gives the balance identity needed to construct the counterexamples. We first realize the data on an exact boundary collar. A correction supported in the interior then gives the global curvature bound while preserving the collar.
For the inradius equality case, we separate distance comparison from a rigidity statement for contact sets on S 2 S^2 S 2 (Proposition 4.5). Distance comparison produces a nonnegative boundary function u u u satisfying ∇ 2 u + b h ≤ 0 \nabla^2u+bh\le0 ∇ 2 u + bh ≤ 0 , with b > 0 b>0 b > 0 , near its compact zero set C C C . At contact points, h ≤ 0 h\le0 h ≤ 0 and rank h ≤ 1 h\le1 h ≤ 1 . The two distinct support-map values supplied by Gálvez–Mira [GM26], together with a convex barrier argument, force a scalar projection of F f F_f F f to have an interval of values on C C C . A regular level curve meeting C C C stays in C C C and becomes a complete great circle on which F f F_f F f is constant. Lemma 2.4 then gives h = 0 h=0 h = 0 . Thus the two rigidity arguments end with the same circle obstruction; in dimension three, the contact inequality and the Gálvez–Mira theorem supply the circle.
§ 2 develops support maps and the geometry of their normal fibers, which yields non-existence of nontrivial saddle Codazzi tensors on S m S^m S m for m ≥ 3 m\ge3 m ≥ 3 . Using this, we complete the proof of Theorem A. § 3 realizes the nonzero saddle data on S 2 S^2 S 2 by metrics with sec ≥ 1 \operatorname{sec} \ge1 sec ≥ 1 . § 5 proves the inradius estimate and its equality case.
Classification of Saddle Codazzi tensors on S m S^m S m (m ≥ 3 m \ge3 m ≥ 3 )
On the unit round sphere ( S m , g S m ) (S^m,g_{S^m}) ( S m , g S m ) , m ≥ 2 m \ge2 m ≥ 2 , we identify a symmetric bilinear form h h h with its self-adjoint endomorphism A A A by h ( X , Y ) = ⟨ A X , Y ⟩ h(X,Y)=\langle AX,Y\rangle h ( X , Y ) = ⟨ A X , Y ⟩ . We say that A A A is saddle if
det ( h ∣ Π ) ≤ 0 for every two-plane Π ⊂ T x S m and every x ∈ S m , \det(h|_{\Pi}) \le0 \qquad\text{for every two-plane } \Pi\subset T_xS^m \text{ and every } x \in S^m, det ( h ∣ Π ) ≤ 0 for every two-plane Π ⊂ T x S m and every x ∈ S m ,
where the determinant is computed in an orthonormal basis of Π \Pi Π . For a function f f f , we write A f A_f A f for the endomorphism determined by
⟨ A f X , Y ⟩ = ∇ 2 f ( X , Y ) + f ⟨ X , Y ⟩ . \langle A_fX,Y\rangle=\nabla^2f(X,Y)+f\langle X,Y\rangle. ⟨ A f X , Y ⟩ = ∇ 2 f ( X , Y ) + f ⟨ X , Y ⟩ .
Thus A f A_f A f represents the spherical Hessian of f f f under this identification. Pointwise evaluation is denoted by A ( x ) A(x) A ( x ) or A f ( x ) A_f(x) A f ( x ) ; the subscript in A f A_f A f records the potential function. For the boundary in Theorem A, the Gauss equation gives
det ( h ∣ Π ) = 1 − sec M ( Π ) ≤ 0. (1) \det(h|_{\Pi})=1-\operatorname{sec}_M(\Pi)\le0. \tag*{(1)} det ( h ∣ Π ) = 1 − sec M ( Π ) ≤ 0. (1)
The essential rigidity statement is therefore intrinsic to the sphere. We will prove that this boundary tensor vanishes in every dimension m ≥ 3 m \ge3 m ≥ 3 . We first describe the support map and the geometry of its compact fibers.
Lemma 2.1. A self-adjoint endomorphism is saddle if and only if it has at most one positive and at most one negative eigenvalue. In particular, its rank is at most two.
Proof. Applying the saddle condition to the plane spanned by any two orthonormal eigenvectors gives λ i λ j ≤ 0 \lambda_i\lambda_j \le0 λ i λ j ≤ 0 for i ≠ j i \ne j i = j . Conversely, if there is at most one eigenvalue of each sign, the restriction to any two-plane is neither positive nor negative definite, and hence has nonpositive determinant. □ \square □
Lemma 2.2. Let A A A be a smooth self-adjoint Codazzi tensor on S m S^m S m , m ≥ 2 m \ge2 m ≥ 2 . Then A = A f A=A_f A = A f for a smooth function f f f , and the map
F : S m ⟶ R m + 1 , F ( x ) = ∇ f ( x ) + f ( x ) x , F:S^m\longrightarrow\mathbb{R}^{m+1},\qquad F(x)=\nabla f(x)+f(x)x, F : S m ⟶ R m + 1 , F ( x ) = ∇ f ( x ) + f ( x ) x , satisfies d F x ( X ) = A ( x ) X dF_x(X)=A(x)X d F x ( X ) = A ( x ) X , viewed as an identity in R m + 1 \mathbb{R}^{m+1} R m + 1 . Conversely, A f A_f A f is Codazzi for every smooth f f f . The potential f f f is unique modulo functions x ↦ ⟨ a , x ⟩ x\mapsto\langle a,x\rangle x ↦ ⟨ a , x ⟩ .
Proof. Regard ω ( X ) = A X \omega(X)=AX ω ( X ) = A X as an R m + 1 \mathbb{R}^{m+1} R m + 1 -valued one-form. The Euclidean Gauss formula gives
d ω ( X , Y ) = ( ∇ X A ) Y − ( ∇ Y A ) X − ( ⟨ X , A Y ⟩ − ⟨ Y , A X ⟩ ) x = 0. d\omega(X,Y)=(\nabla_XA)Y-(\nabla_YA)X-\bigl(\langle X,AY\rangle-\langle Y,AX\rangle\bigr)x=0. d ω ( X , Y ) = ( ∇ X A ) Y − ( ∇ Y A ) X − ( ⟨ X , A Y ⟩ − ⟨ Y , A X ⟩ ) x = 0.
Codazzi cancels the tangential part, and self-adjointness cancels the normal part. Since S m S^m S m is simply connected, ω = d F \omega=dF ω = d F for a smooth map F : S m → R m + 1 F:S^m\to\mathbb{R}^{m+1} F : S m → R m + 1 . Put f ( x ) = ⟨ F ( x ) , x ⟩ f(x)=\langle F(x),x\rangle f ( x ) = ⟨ F ( x ) , x ⟩ . As d F ( X ) = A X dF(X)=AX d F ( X ) = A X is tangent to the sphere, d f ( X ) = ⟨ F , X ⟩ df(X)=\langle F,X\rangle df ( X ) = ⟨ F , X ⟩ , so F = ∇ f + f x F=\nabla f+fx F = ∇ f + f x . Differentiating gives
d F ( X ) = ∇ X ∇ f + f X = A X . (2) dF(X)=\nabla_X\nabla f+fX=AX. \tag*{(2)} d F ( X ) = ∇ X ∇ f + f X = A X . (2)
Conversely, for any smooth f f f , the map F = ∇ f + f x F=\nabla f+fx F = ∇ f + f x has differential A f A_f A f . Its self-adjointness and d ( d F ) = 0 d(dF)=0 d ( d F ) = 0 imply Codazzi by the same calculation. Finally, F F F is unique up to a constant vector, giving the stated freedom in f f f . □
We call F = ∇ f + f x F=\nabla f+fx F = ∇ f + f x the support map . The affine hyperplane
{ y ∈ R m + 1 : ⟨ y , x ⟩ = f ( x ) } \{y\in\mathbb{R}^{m+1}:\langle y,x\rangle=f(x)\} { y ∈ R m + 1 : ⟨ y , x ⟩ = f ( x )}
passes through F ( x ) F(x) F ( x ) and has normal x x x , which is orthogonal to im d F x \operatorname{im}dF_x im d F x . This is the support-function description of a hedgehog, without any convexity assumption; see [Pan05, Section 2] for this viewpoint in R 3 \mathbb{R}^3 R 3 . Although F ( S m ) F(S^m) F ( S m ) is compact, it may have rank-drop singularities. We will use only its smooth local images near the compact fiber components constructed below. Also, restriction to a totally geodesic subsphere commutes with the spherical Hessian, so every great subsphere inherits the saddle condition.
We identify Euclidean tangent spaces with R m + 1 \mathbb{R}^{m+1} R m + 1 by translation. Thus the normal vectors to a local image of F F F and the points of the source sphere are regarded as vectors in the same space.
Lemma 2.3. Let A A A be a smooth self-adjoint Codazzi tensor on S m S^m S m , and let F F F be its support map. Suppose r = max S m rank A r=\max_{S^m}\operatorname{rank} A r = max S m rank A satisfies 0 < r < m 0<r<m 0 < r < m . There is a great ( m − r ) (m-r) ( m − r ) -sphere K K K in the rank-r r r set on which F F F is constant. Along K K K , the map F F F has a common smooth local image S S S of dimension r r r . Write F ∣ K ≡ z F|_K\equiv z F ∣ K ≡ z , and put P = T z S P=T_zS P = T z S and N = P ⊥ N=P^\perp N = P ⊥ . Then K = S m ∩ N K=S^m\cap N K = S m ∩ N : the fiber is the entire unit normal sphere of S S S at z z z .
Proof. Choose p ∈ S m p\in S^m p ∈ S m with rank A ( p ) = r \operatorname{rank}A(p)=r rank A ( p ) = r and a linear map L : R m + 1 → R r L:\mathbb{R}^{m+1}\to\mathbb{R}^r L : R m + 1 → R r such that L ∘ A ( p ) L\circ A(p) L ∘ A ( p ) is surjective. Then G = L ∘ F G=L\circ F G = L ∘ F is a submersion near p p p , so its image contains a nonempty open subset of R r \mathbb{R}^r R r . By Sard’s theorem, choose in this subset a regular value y y y of the global map G : S m → R r G:S^m\to\mathbb{R}^r G : S m → R r , and let K K K be a connected component of G − 1 ( y ) G^{-1}(y) G − 1 ( y ) . It is a compact submanifold without boundary, and
r = rank d G x ≤ rank A ( x ) ≤ r , T x K = ker d G x = ker A ( x ) ( x ∈ K ) . r=\operatorname{rank}dG_x\leq\operatorname{rank}A(x)\leq r,\qquad T_xK=\ker dG_x=\ker A(x)\quad(x\in K). r = rank d G x ≤ rank A ( x ) ≤ r , T x K = ker d G x = ker A ( x ) ( x ∈ K ) .
Thus F ∣ K F|_K F ∣ K is constant; denote its value by z z z .
The rank of F F F is r r r near K K K . By the constant-rank theorem, F F F has an embedded local image of dimension r r r near each point of K K K . If two such neighborhoods meet on K K K , their image germs at z z z agree: the restriction of F F F to their intersection still has rank r r r , and its image contains a neighborhood of z z z in each local image. Connectedness of K K K makes these image germs agree along K K K . A finite cover of the compact set K K K then gives a common local image S S S near z z z , after shrinking the neighborhoods. Consequently P = T z S = im A ( x ) P=T_zS=\operatorname{im} A(x) P = T z S = im A ( x ) is independent of x ∈ K x\in K x ∈ K . Self-adjointness gives the ambient normal space
N = R x ⊕ ker A ( x ) , x ∈ K . N=\mathbb{R}x\oplus\ker A(x),\qquad x\in K. N = R x ⊕ ker A ( x ) , x ∈ K .
Since P ⊂ x ⊥ P\subset x^\perp P ⊂ x ⊥ , we have K ⊂ S m ∩ P ⊥ K\subset\mathbb{S}^m\cap P^\perp K ⊂ S m ∩ P ⊥ . Both sides have dimension m − r m-r m − r ; hence K K K is open in this great sphere, and compactness makes it closed. Since m − r ≥ 1 m-r\ge1 m − r ≥ 1 , the great sphere is connected, proving K = S m ∩ N K=\mathbb{S}^m\cap N K = S m ∩ N .
The normal-fiber description supplies a complete great circle on which the support map is constant. To use this global information, we restrict to a great two-sphere containing the circle and translate the constant support value to the origin. The support function then vanishes on the circle. This condition alone forces its restriction to be a first spherical harmonic: Stokes’ theorem makes the projected signed area vanish on each hemisphere, and the saddle condition turns this into equality in the first Dirichlet eigenvalue inequality.
Lemma 2.4. Let u ∈ C 2 ( S 2 ) u\in C^2(\mathbb{S}^2) u ∈ C 2 ( S 2 ) satisfy det A u ≤ 0 \det A_u\le0 det A u ≤ 0 . Suppose u = 0 u=0 u = 0 on a complete great circle C = S 2 ∩ e ⊥ C=\mathbb{S}^2\cap e^\perp C = S 2 ∩ e ⊥ , where e ∈ R 3 e\in\mathbb{R}^3 e ∈ R 3 is a unit vector. Then, for some a ∈ R a\in\mathbb{R} a ∈ R ,
u ( x ) = a ⟨ e , x ⟩ on S 2 . u(x)=a\langle e,x\rangle\qquad\text{on }\mathbb{S}^2. u ( x ) = a ⟨ e , x ⟩ on S 2 . In particular, A u = 0 A_u=0 A u = 0 .
Proof. Rotate the sphere so that e = e 3 e=e_3 e = e 3 . Write A = A u A=A_u A = A u and F = ∇ u + u x F=\nabla u+ux F = ∇ u + ux , and let H ± = { ± x 3 ≥ 0 } H_\pm=\{\pm x_3\ge0\} H ± = { ± x 3 ≥ 0 } be the two closed hemispheres. Since u = 0 u=0 u = 0 on the equator C C C , its gradient there is parallel to e 3 e_3 e 3 . Hence F 1 = F 2 = 0 F_1=F_2=0 F 1 = F 2 = 0 on C C C . The identity d F = A dF=A d F = A gives
d F 1 ∧ d F 2 = x 3 det A d μ , dF_1\wedge dF_2=x_3\det A\,d\mu, d F 1 ∧ d F 2 = x 3 det A d μ ,
where d μ d\mu d μ is the oriented round area form. This identity is on the source sphere and does not require F F F to be an immersion. Stokes’ theorem therefore yields
∫ H ± x 3 det A d μ = ∫ ∂ H ± F 1 d F 2 = 0. \int_{H_\pm}x_3\det A\,d\mu=\int_{\partial H_\pm}F_1\,dF_2=0. ∫ H ± x 3 det A d μ = ∫ ∂ H ± F 1 d F 2 = 0.
Since x 3 x_3 x 3 has a fixed nonzero sign in the interior of each hemisphere and det A ≤ 0 \det A\le0 det A ≤ 0 , we obtain det A = 0 \det A=0 det A = 0 everywhere.
Let H H H be either hemisphere and ν \nu ν its outward unit conormal. The equator is totally geodesic and u ∣ C = 0 u|_C=0 u ∣ C = 0 , so Δ u − ∇ 2 u ( ν , ν ) = 0 \Delta u-\nabla^2u(\nu,\nu)=0 Δ u − ∇ 2 u ( ν , ν ) = 0 on C C C . The boundary term in the integrated Bochner formula is therefore u ν ( Δ u − ∇ 2 u ( ν , ν ) ) = 0 u_\nu(\Delta u-\nabla^2u(\nu,\nu))=0 u ν ( Δ u − ∇ 2 u ( ν , ν )) = 0 , giving
∫ H ( ( Δ u ) 2 − ∣ ∇ 2 u ∣ 2 ) d μ = ∫ H ∣ ∇ u ∣ 2 d μ . \int_H\left((\Delta u)^2-\lvert\nabla^2u\rvert^2\right)d\mu=\int_H\lvert\nabla u\rvert^2d\mu. ∫ H ( ( Δ u ) 2 − ∣ ∇ 2 u ∣ 2 ) d μ = ∫ H ∣ ∇ u ∣ 2 d μ .
Using the two-dimensional identity
2 det A = ( Δ u ) 2 − ∣ ∇ 2 u ∣ 2 + 2 u Δ u + 2 u 2 2\det A = (\Delta u)^2 - \lvert\nabla^2 u\rvert^2 + 2u\Delta u + 2u^2 2 det A = ( Δ u ) 2 − ∣ ∇ 2 u ∣ 2 + 2 u Δ u + 2 u 2
and integrating u Δ u u\Delta u u Δ u by parts, we conclude that
0 = 2 ∫ H det A d μ = 2 ∫ H u 2 d μ − ∫ H ∣ ∇ u ∣ 2 d μ . 0 = 2\int_H \det A\,d\mu= 2\int_H u^2\,d\mu- \int_H \lvert\nabla u\rvert^2\,d\mu. 0 = 2 ∫ H det A d μ = 2 ∫ H u 2 d μ − ∫ H ∣ ∇ u ∣ 2 d μ .
The integral identities above are valid for C 2 C^2 C 2 functions by smooth approximation; their interior and boundary terms involve derivatives of order at most two.
The first Dirichlet eigenvalue of a unit hemisphere is 2 2 2 , and its eigenspace is spanned by x 3 x_3 x 3 . The last identity is precisely the equality case in the Rayleigh inequality. Thus u = a + x 3 u=a_+x_3 u = a + x 3 on H + H_+ H + and u = a − x 3 u=a_-x_3 u = a − x 3 on H − H_- H − . Continuity of ∇ u \nabla u ∇ u across C C C , where ∇ x 3 = e 3 \nabla x_3=e_3 ∇ x 3 = e 3 , gives a + = a − a_+=a_- a + = a − . This proves the claim. □
Theorem 2.5. Let m ≥ 3 m\ge3 m ≥ 3 . Every smooth self-adjoint saddle Codazzi tensor on S m \mathbb{S}^m S m vanishes identically. Equivalently, if f ∈ C ∞ ( S m ) f\in C^\infty(\mathbb{S}^m) f ∈ C ∞ ( S m ) and A f A_f A f is saddle, then f ( x ) = ⟨ z , x ⟩ f(x)=\langle z,x\rangle f ( x ) = ⟨ z , x ⟩ for some z ∈ R m + 1 z\in\mathbb{R}^{m+1} z ∈ R m + 1 .
Proof. Write A = A f A=A_f A = A f and F = ∇ f + f x F=\nabla f+fx F = ∇ f + f x . Suppose A ≠ 0 A\ne0 A = 0 . By Lemma 2.1 , its maximal rank satisfies 0 < r ≤ 2 < m 0<r\le2<m 0 < r ≤ 2 < m . Lemma 2.3 therefore gives a great ( m − r ) (m-r) ( m − r ) -sphere K K K on which F F F is constant, say F = z F=z F = z . Since m − r ≥ 1 m-r\ge1 m − r ≥ 1 , choose a complete great circle C ⊂ K C\subset K C ⊂ K . For f ^ ( x ) = f ( x ) − ⟨ z , x ⟩ \hat f(x)=f(x)-\langle z,x\rangle f ^ ( x ) = f ( x ) − ⟨ z , x ⟩ , we have
f ^ = 0 , d f ^ = 0 on C . \hat f=0,\qquad d\hat f=0 \quad\text{on } C. f ^ = 0 , d f ^ = 0 on C .
Let Σ ≃ S 2 \Sigma\simeq\mathbb{S}^2 Σ ≃ S 2 be any great two-sphere containing C C C , and set u = f ^ ∣ Σ u=\hat f|_\Sigma u = f ^ ∣ Σ . Since Σ \Sigma Σ is totally geodesic, A u A_u A u represents the restriction of the bilinear form associated with A A A to T Σ T\Sigma T Σ . Hence det A u ≤ 0 \det A_u\le0 det A u ≤ 0 , and u = 0 u=0 u = 0 on C C C . Lemma 2.4 makes u u u a linear function on Σ \Sigma Σ . Since also d u = 0 du=0 d u = 0 on C C C , this linear function is zero. Thus u = 0 u=0 u = 0 on Σ \Sigma Σ . Every point of S m \mathbb{S}^m S m lies on a great two-sphere containing C C C : adjoin its position vector to the two-plane spanning C C C , and extend to a three-plane if necessary. Thus f ^ = 0 \hat f=0 f ^ = 0 on S m \mathbb{S}^m S m , so A = 0 A=0 A = 0 , a contradiction. □
The dimension threshold is now explicit. For m ≥ 3 m\ge3 m ≥ 3 , the bound r ≤ 2 < m r\le2<m r ≤ 2 < m supplies a positive-dimensional normal fiber and hence a complete circle. On S 2 \mathbb{S}^2 S 2 , maximal rank two supplies no such circle; this is the case in which the nontrivial saddle data used below occur.
Proof of Theorem A. Identify ∂ M \partial M ∂ M with S n − 1 \mathbb{S}^{n-1} S n − 1 . Its second fundamental form is Codazzi by assumption, and (1) gives the saddle condition for the associated self-adjoint endomorphism. Theorem 2.5 therefore gives h = 0 h=0 h = 0 . Since sec g ≥ 1 \operatorname{sec}_g\ge1 sec g ≥ 1 implies Ric g ≥ ( n − 1 ) g \operatorname{Ric}_g\ge(n-1)g Ric g ≥ ( n − 1 ) g , the hemisphere rigidity theorem of Hang–Wang [1 ] gives ( M n , g ) ≅ S + n (M^n,g)\cong\mathbb{S}^n_+ ( M n , g ) ≅ S + n . □
Construction of counterexamples in dimension three
Unlike the case when m ≥ 3 m \ge3 m ≥ 3 , there do exist nontrivial saddle Codazzi tensors on S 2 \mathbb{S}^{2} S 2 , see e.g., [Pan05, Theorem 6.1]. We shall realize the nontrivial saddle data available on S 2 \mathbb{S}^{2} S 2 , proving Theorem B.
Saddle data and the metric realization
Let H = S + 3 H = \mathbb{S}^{3}_{+} H = S + 3 be the closed northern hemisphere, equipped with the standard round metric g 0 g_{0} g 0 , and let σ \sigma σ be the induced unit round metric on ∂ H = S 2 \partial H = \mathbb{S}^{2} ∂ H = S 2 . Panina’s construction provides a function f ∈ C ∞ ( S 2 ) f \in C^{\infty}(\mathbb{S}^{2}) f ∈ C ∞ ( S 2 ) such that
A = ∇ σ 2 f + f σ , ρ = − det σ A ≥ 0 , ρ ≠ 0. (3) A = \nabla_{\sigma}^{2} f + f\sigma,\qquad\rho= -\det_{\sigma} A \ge0,\qquad\rho\ne0. \tag*{(3)} A = ∇ σ 2 f + f σ , ρ = − σ det A ≥ 0 , ρ = 0. (3)
We also use A A A for its associated self-adjoint endomorphism with respect to σ \sigma σ . The associated support map is
F f ( q ) = ∇ σ f ( q ) + f ( q ) q , d F f = A . F_{f}(q) = \nabla_{\sigma} f(q) + f(q)q,\qquad dF_{f} = A. F f ( q ) = ∇ σ f ( q ) + f ( q ) q , d F f = A .
The condition ρ ≠ 0 \rho\ne0 ρ = 0 follows already from A ≠ 0 A \ne0 A = 0 . Indeed, if ρ = 0 \rho= 0 ρ = 0 , then A A A has maximal rank one. Lemma 2.3 , with m = 2 m = 2 m = 2 and r = 1 r = 1 r = 1 , gives a complete great circle on which F f F_{f} F f is constant, say z z z . The function f − ⟨ z , q ⟩ f - \langle z,q\rangle f − ⟨ z , q ⟩ vanishes on that circle, so Lemma 2.4 would give A = 0 A = 0 A = 0 . Panina’s construction gives a smooth function and hence a smooth map F f F_{f} F f ; the map need not be an immersion, and its image need not be a smooth surface. We use F f F_{f} F f to organize auxiliary Euclidean data. The metric on H H H will be produced by a collar construction and an interior correction that control its sectional curvature globally. In particular, no realization as the boundary of a Euclidean convex body is assumed. The tensor A A A is Codazzi by Lemma 2.2 . For the proposed boundary data h ε = ε A h_{\varepsilon} = \varepsilon A h ε = ε A , the tangential Gauss equation requires
sec g ε ( T ∂ H ) = 1 − ε 2 det σ A = 1 + ε 2 ρ ≥ 1. \operatorname{sec}_{g_{\varepsilon}}(T\partial H) = 1 - \varepsilon^{2}\det_{\sigma} A = 1 + \varepsilon^{2}\rho\ge1. sec g ε ( T ∂ H ) = 1 − ε 2 σ det A = 1 + ε 2 ρ ≥ 1.
The task is to realize these data while controlling all sectional curvatures in the interior.
For a metric g g g on H H H , we use the convention
h g ( X , Y ) = g ( ∇ X g ν g , Y ) , X , Y ∈ T ∂ H , h_{g}(X,Y) = g(\nabla^{g}_{X}\nu_{g},Y),\qquad X,Y \in T\partial H, h g ( X , Y ) = g ( ∇ X g ν g , Y ) , X , Y ∈ T ∂ H ,
where ν g \nu_{g} ν g is the outward unit normal. We shall prove the following.
Theorem 3.1. For every f ∈ C ∞ ( S 2 ) f \in C^{\infty}(\mathbb{S}^{2}) f ∈ C ∞ ( S 2 ) satisfying (3) , there exist ε 0 > 0 \varepsilon_{0} > 0 ε 0 > 0 and a smooth family of Riemannian metrics { g ε } ∣ ε ∣ < ε 0 \{g_{\varepsilon}\}_{|\varepsilon|<\varepsilon_{0}} { g ε } ∣ ε ∣ < ε 0 on H H H such that
g ε = 0 = g 0 , sec g ε ≥ 1 , g ε ∣ T ∂ H = σ , h g ε = ε A . g_{\varepsilon=0} = g_{0},\qquad\operatorname{sec}_{g_{\varepsilon}} \ge1,\qquad g_{\varepsilon}|_{T\partial H} = \sigma,\qquad h_{g_{\varepsilon}} = \varepsilon A. g ε = 0 = g 0 , sec g ε ≥ 1 , g ε ∣ T ∂ H = σ , h g ε = ε A . In particular, the boundary second fundamental form is Codazzi for every ∣ ε ∣ < ε 0 |\varepsilon| < \varepsilon_{0} ∣ ε ∣ < ε 0 , and ( H , g ε ) (H,g_{\varepsilon}) ( H , g ε ) is not isometric to the standard hemisphere whenever 0 < ∣ ε ∣ < ε 0 0 < |\varepsilon| < \varepsilon_{0} 0 < ∣ ε ∣ < ε 0 .
Away from the north pole, we write
X = ( x , z ) = ( cos t q , sin t ) ∈ S 3 ⊂ R 3 × R , g 0 = d t 2 + cos 2 t σ , X=(x,z)=(\cos t\,q,\sin t)\in S^3\subset\mathbb{R}^3\times\mathbb{R},\qquad g_0=dt^2+\cos^2t\,\sigma, X = ( x , z ) = ( cos t q , sin t ) ∈ S 3 ⊂ R 3 × R , g 0 = d t 2 + cos 2 t σ ,
where q ∈ S 2 q\in S^2 q ∈ S 2 and t ∈ [ 0 , π / 2 ) t\in[0,\pi/2) t ∈ [ 0 , π /2 ) is the inward distance from the equator. Near the equator, we also use the signed extension of t t t to the ambient sphere, with t < 0 t<0 t < 0 on the southern side. On the open northern hemisphere, the gnomonic coordinates are y = x z ∈ R 3 y=\frac{x}{z}\in\mathbb{R}^3 y = z x ∈ R 3 . For 0 < t < π / 2 0<t<\pi/2 0 < t < π /2 , this gives y = cot t q y=\cot t\,q y = cot t q , while the north pole corresponds to y = 0 y=0 y = 0 . In these coordinates,
g 0 = ∑ i ( d y i ) 2 1 + ∣ y ∣ 2 − ( ∑ i y i d y i ) 2 ( 1 + ∣ y ∣ 2 ) 2 . g_0=\frac{\sum_i(dy^i)^2}{1+|y|^2}-\frac{\left(\sum_i y^i\,dy^i\right)^2}{(1+|y|^2)^2}. g 0 = 1 + ∣ y ∣ 2 ∑ i ( d y i ) 2 − ( 1 + ∣ y ∣ 2 ) 2 ( ∑ i y i d y i ) 2 .
Thus, for 0 < c < π / 2 0<c<\pi/2 0 < c < π /2 , the interior cap { t ≥ c } \{t\geq c\} { t ≥ c } corresponds to the closed Euclidean ball { ∣ y ∣ ≤ cot c } \{|y|\leq\cot c\} { ∣ y ∣ ≤ cot c } .
In dimension three, the sectional curvature bound can be expressed as the nonnegativity of a symmetric two-tensor. For a Riemannian metric g g g , define
T ( g ) = ( Scal g 2 − 1 ) g − Ric g . T(g)=\left(\frac{\operatorname{Scal}_g}{2}-1\right)g-\operatorname{Ric}_g. T ( g ) = ( 2 Scal g − 1 ) g − Ric g .
If n n n is a g g g -unit vector, then
T ( g ) ( n , n ) = sec g ( n ⊥ ) − 1 , T(g)(n,n)=\sec_g(n^\perp)-1, T ( g ) ( n , n ) = sec g ( n ⊥ ) − 1 ,
where n ⊥ n^\perp n ⊥ is the two-plane orthogonal to n n n with respect to g g g . Consequently,
sec g ≥ 1 ⟺ T ( g ) ≥ 0. \sec_g\geq1\qquad\Longleftrightarrow\qquad T(g)\geq0. sec g ≥ 1 ⟺ T ( g ) ≥ 0.
More generally, T ( g ) ≥ δ g T(g)\geq\delta g T ( g ) ≥ δ g is equivalent to sec g ≥ 1 + δ \sec_g\geq1+\delta sec g ≥ 1 + δ . Thus T ( g ) T(g) T ( g ) records the sectional curvature excess over 1. Moreover, the contracted Bianchi identity gives
div g T ( g ) = 0 , T ( g 0 ) = 0. \operatorname{div}_gT(g)=0,\qquad T(g_0)=0. div g T ( g ) = 0 , T ( g 0 ) = 0.
The construction has three steps. Choose the parameters in the order described below, and then fix them independently of ε \varepsilon ε :
0 < t 0 < a < b < t 1 < t 2 < π 4 . 0<t_0<a<b<t_1<t_2<\frac{\pi}{4}. 0 < t 0 < a < b < t 1 < t 2 < 4 π .
First, using gnomonic coordinates, we construct a smooth symmetric tensor D D D on an open neighborhood of the closed hemisphere H H H such that
D ≥ 0 , div g 0 D = 0. D\geq0,\qquad\operatorname{div}_{g_0}D=0. D ≥ 0 , div g 0 D = 0.
This tensor is intended to be the coefficient of ε 2 \varepsilon^2 ε 2 in T ( g ε ) T(g_\varepsilon) T ( g ε ) . Its normal components at the boundary are prescribed by the Gauss and Codazzi equations:
D ( ∂ t , ∂ t ) = ρ , D ( ∂ t , X ) = 0 for X ∈ T ∂ H . D(\partial_t,\partial_t)=\rho,\qquad D(\partial_t,X)=0\quad\text{for }X\in T\partial H. D ( ∂ t , ∂ t ) = ρ , D ( ∂ t , X ) = 0 for X ∈ T ∂ H .
We also arrange that D D D is uniformly positive definite on the interior cap { t ≥ t 0 } \{t \ge t_0\} { t ≥ t 0 } . On a larger, two-sided collar [ − τ , t 2 ] × S 2 [-\tau,t_2]\times S^2 [ − τ , t 2 ] × S 2 , where τ > 0 \tau>0 τ > 0 , the tensor D D D admits a nonnegative geodesic integral representation, which provides the data for the next step.
Second, we construct a smooth family of metrics G ε G_\varepsilon G ε on the fixed boundary collar { 0 ≤ t ≤ t 1 } \{0\le t\le t_1\} { 0 ≤ t ≤ t 1 } , with induced boundary metric σ \sigma σ and second fundamental form ε A \varepsilon A ε A , such that
T ( G ε ) ≥ 0 , T ( G ε ) = ε 2 D + O ( ε 3 ) . T(G_\varepsilon)\ge0,\qquad T(G_\varepsilon)=\varepsilon^2D+O(\varepsilon^3). T ( G ε ) ≥ 0 , T ( G ε ) = ε 2 D + O ( ε 3 ) .
The first inequality is exact, including at points where D D D is degenerate. On the interior slab { t 0 ≤ t ≤ t 1 } \{t_0\le t\le t_1\} { t 0 ≤ t ≤ t 1 } , the positivity of D D D gives a strictly positive curvature margin of order ε 2 \varepsilon^2 ε 2 .
Finally, we extend the collar metrics smoothly to H H H , retaining G ε G_\varepsilon G ε on { t ≤ a } \{t\le a\} { t ≤ a } and placing the cutoff transition inside the slab { a < t < b } \{a<t<b\} { a < t < b } . The extension is arranged so that its first variation is a Lie derivative of g 0 g_0 g 0 . Since T ( g 0 ) = 0 T(g_0)=0 T ( g 0 ) = 0 and T T T is natural under diffeomorphisms, its curvature excess has no linear term in ε \varepsilon ε . The discrepancy between its quadratic term and D D D is divergence-free and supported in the closed cap { t ≥ a } \{t\ge a\} { t ≥ a } . A supported solution of the linearized equation for T T T removes this discrepancy by a metric correction of order ε 2 \varepsilon^2 ε 2 , giving
T ( g ε ) = ε 2 D + O ( ε 3 ) on H . T(g_\varepsilon)=\varepsilon^2D+O(\varepsilon^3)\quad\text{on }H. T ( g ε ) = ε 2 D + O ( ε 3 ) on H .
On { t ≥ a } \{t\ge a\} { t ≥ a } , the uniform positivity of D D D absorbs the remaining cubic error for sufficiently small ∣ ε ∣ |\varepsilon| ∣ ε ∣ . On { t ≤ a } \{t\le a\} { t ≤ a } , the metric remains equal to the exact collar metric, so the curvature bound and the prescribed boundary data are preserved. All remainder estimates are understood in fixed C k C^k C k norms on the indicated regions.
Curvature data and geodesic transport
We first extend the boundary curvature excess to a nonnegative divergence-free tensor. Its Euclidean description gives both the required positivity in the interior and a representation by round geodesics, which will persist under the collar deformation.
Lemma 3.2. There exists a smooth symmetric two-tensor D D D on H H H , smooth up to the boundary, satisfying
D ≥ 0 , div g 0 D = 0. D\ge0,\qquad\operatorname{div}_{g_0}D=0. D ≥ 0 , div g 0 D = 0. At the equator,
D ( ∂ t , ∂ t ) = ρ , D ( ∂ t , Y ) = 0 for every Y ∈ T ∂ H . D(\partial_t,\partial_t)=\rho,\qquad D(\partial_t,Y)=0\quad\text{for every }Y\in T\partial H. D ( ∂ t , ∂ t ) = ρ , D ( ∂ t , Y ) = 0 for every Y ∈ T ∂ H . Moreover, there exist constants c 0 > 0 c_0>0 c 0 > 0 and 0 < t 0 < π / 4 0<t_0<\pi/4 0 < t 0 < π /4 such that
D ≥ c 0 g 0 on { t ≥ t 0 } . D\ge c_0g_0\quad\text{on }\{t\ge t_0\}. D ≥ c 0 g 0 on { t ≥ t 0 } .
Proof. We first establish the balance identity
∫ S 2 ρ ( q ) q d μ σ ( q ) = 0. (4) \int_{\mathbb{S}^{2}} \rho(q)q\,d\mu_{\sigma}(q)=0. \tag*{(4)} ∫ S 2 ρ ( q ) q d μ σ ( q ) = 0. (4)
The identity states that the total oriented area vector of F f F_f F f vanishes. We can apply Stokes’ theorem on the parameter sphere even where F f F_f F f is singular. Orient S 2 \mathbb{S}^{2} S 2 by its outward normal q q q . For every cyclic permutation ( i , j , k ) (i,j,k) ( i , j , k ) of ( 1 , 2 , 3 ) (1,2,3) ( 1 , 2 , 3 ) , the identity d F f = A dF_f=A d F f = A gives
d F f j ∧ d F f k = ( det σ A ) q i d μ σ . dF_f^{j}\wedge dF_f^{k}=(\det_{\sigma} A)q^{i}\,d\mu_{\sigma}. d F f j ∧ d F f k = ( σ det A ) q i d μ σ .
Indeed, on a positively oriented orthonormal frame, the cross product of the two images under A A A is ( det σ A ) q (\det_{\sigma} A)q ( det σ A ) q . Stokes’ theorem now yields
∫ S 2 ( det σ A ) q i d μ σ = ∫ S 2 d F f j ∧ d F f k = ∫ S 2 d ( F f j d F f k ) = 0. \int_{\mathbb{S}^{2}}(\det_{\sigma} A)q^{i}\,d\mu_{\sigma}
=\int_{\mathbb{S}^{2}}dF_f^{j}\wedge dF_f^{k}
=\int_{\mathbb{S}^{2}}d(F_f^{j}dF_f^{k})=0. ∫ S 2 ( σ det A ) q i d μ σ = ∫ S 2 d F f j ∧ d F f k = ∫ S 2 d ( F f j d F f k ) = 0.
Since ρ = − det σ A \rho=-\det_{\sigma} A ρ = − det σ A , this proves (4) .
We next construct a tensor in Euclidean space. For y ≠ 0 y\ne0 y = 0 , define the symmetric contravariant tensor
Q i j ( y ) = ρ ( y ∣ y ∣ ) y i y j ∣ y ∣ 4 . (5) Q^{ij}(y)=\rho\left(\frac{y}{|y|}\right)\frac{y^{i}y^{j}}{|y|^{4}}. \tag*{(5)} Q ij ( y ) = ρ ( ∣ y ∣ y ) ∣ y ∣ 4 y i y j . (5)
It is nonnegative, locally integrable, and satisfies ∂ i Q i j = 0 \partial_iQ^{ij}=0 ∂ i Q ij = 0 away from the origin. Its distributional divergence also vanishes at the origin. Indeed, for a compactly supported smooth vector field φ \varphi φ on R 3 \mathbb{R}^{3} R 3 , integration by parts outside a ball of radius r r r gives
⟨ div Q , φ ⟩ = lim r ↓ 0 ∫ S 2 ρ ( q ) ⟨ q , φ ( r q ) ⟩ d μ σ ( q ) = ⟨ ∫ S 2 ρ ( q ) q d μ σ ( q ) , φ ( 0 ) ⟩ = 0. \langle\operatorname{div}Q,\varphi\rangle
=\lim_{r\downarrow0}\int_{\mathbb{S}^{2}}\rho(q)\langle q,\varphi(rq)\rangle\,d\mu_{\sigma}(q)
=\left\langle\int_{\mathbb{S}^{2}}\rho(q)q\,d\mu_{\sigma}(q),\varphi(0)\right\rangle=0. ⟨ div Q , φ ⟩ = r ↓ 0 lim ∫ S 2 ρ ( q ) ⟨ q , φ ( r q )⟩ d μ σ ( q ) = ⟨ ∫ S 2 ρ ( q ) q d μ σ ( q ) , φ ( 0 ) ⟩ = 0.
Fix a smooth radial function η ≥ 0 \eta\ge0 η ≥ 0 on R 3 \mathbb{R}^{3} R 3 such that
supp η ⊂ B 1 ( 0 ) , η > 0 on B 1 ( 0 ) , ∫ R 3 η ( a ) d a = 1. \operatorname{supp}\eta\subset B_{1}(0),\qquad\eta>0\text{ on }B_{1}(0),\qquad\int_{\mathbb{R}^{3}}\eta(a)\,da=1. supp η ⊂ B 1 ( 0 ) , η > 0 on B 1 ( 0 ) , ∫ R 3 η ( a ) d a = 1.
The convolution
D ^ i j ( y ) = ( Q i j ∗ η ) ( y ) = ∫ R 3 ρ ( y − a ∣ y − a ∣ ) ( y − a ) i ( y − a ) j ∣ y − a ∣ 4 η ( a ) d a \widehat{D}^{ij}(y)=(Q^{ij}*\eta)(y)=\int_{\mathbb{R}^{3}}\rho\left(\frac{y-a}{|y-a|}\right)\frac{(y-a)^{i}(y-a)^{j}}{|y-a|^{4}}\eta(a)\,da D ij ( y ) = ( Q ij ∗ η ) ( y ) = ∫ R 3 ρ ( ∣ y − a ∣ y − a ) ∣ y − a ∣ 4 ( y − a ) i ( y − a ) j η ( a ) d a
is smooth, nonnegative, and divergence-free with respect to the Euclidean metric.
We now transfer this tensor to the open northern hemisphere using the gnomonic coordinates from the setup. Write s = 1 + ∣ y ∣ 2 s=1+|y|^{2} s = 1 + ∣ y ∣ 2 , and define D D D by prescribing its contravariant components:
D i j = s 3 D ^ i j . (6) D^{ij}=s^{3}\widehat{D}^{ij}. \tag*{(6)} D ij = s 3 D ij . (6)
Here the indices of D D D are raised and lowered using g 0 g_0 g 0 . In particular, its covariant components are
D i j = s 3 ( g 0 ) i k ( g 0 ) j ℓ D ^ k ℓ . D_{ij}=s^3(g_0)_{ik}(g_0)_{j\ell}\widehat{D}^{k\ell}. D ij = s 3 ( g 0 ) ik ( g 0 ) j ℓ D k ℓ .
so D ≥ 0 D\geq0 D ≥ 0 . Writing y i = y i y_i=y^i y i = y i for the Euclidean coordinate components, the Christoffel symbols are
Γ i j k = − y i δ j k + y j δ i k s . \Gamma^k_{ij}=-\frac{y_i\delta^k_j+y_j\delta^k_i}{s}. Γ ij k = − s y i δ j k + y j δ i k .
Thus, for a symmetric contravariant tensor,
∇ i D i j = ∂ i D i j − 6 y i s D i j = s 3 ∂ i ( s − 3 D i j ) . \nabla_iD^{ij}=\partial_iD^{ij}-\frac{6y_i}{s}D^{ij}=s^3\partial_i\left(s^{-3}D^{ij}\right). ∇ i D ij = ∂ i D ij − s 6 y i D ij = s 3 ∂ i ( s − 3 D ij ) .
Together with (6) , this proves div g 0 D = 0 \operatorname{div}_{g_0}D=0 div g 0 D = 0 in the interior of H H H .
To verify smoothness at the equator, we rewrite the construction in ambient spherical coordinates. For a ∈ R 3 a\in\mathbb{R}^3 a ∈ R 3 , set
w a = x − z a , W a = pr T X S 3 ( a , 1 ) . w_a=x-za,\qquad W_a=\operatorname{pr}_{T_XS^3}(a,1). w a = x − z a , W a = pr T X S 3 ( a , 1 ) .
The tensor defined above has the equivalent expression
D X = ∫ R 3 ρ ( w a ∣ w a ∣ ) W a ♭ ⊗ W a ♭ ∣ w a ∣ 4 η ( a ) d a . (7) D_X=\int_{\mathbb{R}^3}\rho\left(\frac{w_a}{|w_a|}\right)\frac{W_a^\flat\otimes W_a^\flat}{|w_a|^4}\eta(a)\,da. \tag*{(7)} D X = ∫ R 3 ρ ( ∣ w a ∣ w a ) ∣ w a ∣ 4 W a ♭ ⊗ W a ♭ η ( a ) d a . (7)
where ♭ \flat ♭ denotes the dual with respect to g 0 g_0 g 0 . Indeed, in the northern gnomonic coordinates,
w a = y − a s , W a i = s ( a i − y i ) , w_a=\frac{y-a}{\sqrt{s}},\qquad W_a^i=\sqrt{s}(a^i-y^i), w a = s y − a , W a i = s ( a i − y i ) ,
which gives precisely (6) .
For a ∈ supp η a\in\operatorname{supp}\eta a ∈ supp η and ∣ t ∣ < π / 4 |t|<\pi/4 ∣ t ∣ < π /4 ,
∣ w a ∣ ≥ cos t − ∣ sin t ∣ > 0. |w_a|\geq\cos t-|\sin t|>0. ∣ w a ∣ ≥ cos t − ∣ sin t ∣ > 0.
Hence (7) gives a smooth, nonnegative extension of D D D across the equator. In particular, D D D is smooth up to ∂ H \partial H ∂ H , and its divergence vanishes there by continuity.
At X = ( q , 0 ) X=(q,0) X = ( q , 0 ) , we have
w a = q , W a = ( a − ⟨ a , q ⟩ q , 1 ) , ∂ t = ( 0 , 1 ) . w_a=q,\qquad W_a=(a-\langle a,q\rangle q,1),\qquad\partial_t=(0,1). w a = q , W a = ( a − ⟨ a , q ⟩ q , 1 ) , ∂ t = ( 0 , 1 ) .
It follows that
D ( ∂ t , ∂ t ) = ρ ( q ) ∫ R 3 η ( a ) d a = ρ ( q ) . D(\partial_t,\partial_t)=\rho(q)\int_{\mathbb{R}^3}\eta(a)\,da=\rho(q). D ( ∂ t , ∂ t ) = ρ ( q ) ∫ R 3 η ( a ) d a = ρ ( q ) .
For Y ∈ T q S 2 Y\in T_qS^2 Y ∈ T q S 2 , radial symmetry of η \eta η gives
D ( ∂ t , Y ) = ρ ( q ) ∫ R 3 ⟨ a , Y ⟩ η ( a ) d a = 0. D(\partial_t,Y)=\rho(q)\int_{\mathbb{R}^3}\langle a,Y\rangle\eta(a)\,da=0. D ( ∂ t , Y ) = ρ ( q ) ∫ R 3 ⟨ a , Y ⟩ η ( a ) d a = 0.
It remains to prove strict positivity on an interior cap. We claim that D ^ ( y ) \widehat{D}(y) D ( y ) is positive definite for every ∣ y ∣ ≤ 1 |y| \le1 ∣ y ∣ ≤ 1 . Fix such a point y y y and a nonzero Euclidean covector ξ \xi ξ . If ∣ y ∣ < 1 |y| < 1 ∣ y ∣ < 1 , choose q ∈ S 2 q \in\mathbb{S}^{2} q ∈ S 2 such that
ρ ( q ) > 0 , ⟨ ξ , q ⟩ ≠ 0. \rho(q)>0,\qquad\langle\xi,q\rangle\ne0. ρ ( q ) > 0 , ⟨ ξ , q ⟩ = 0.
Such a choice is possible because { ρ > 0 } \{\rho>0\} { ρ > 0 } is a nonempty open set. For sufficiently small r > 0 r>0 r > 0 , the point a = y − r q a=y-rq a = y − r q lies in B 1 ( 0 ) B_{1}(0) B 1 ( 0 ) .
If ∣ y ∣ = 1 |y|=1 ∣ y ∣ = 1 , the balance identity (4) implies that the open set
{ q ∈ S 2 : ρ ( q ) > 0 , ⟨ y , q ⟩ > 0 } \{q\in\mathbb{S}^{2}:\rho(q)>0,\ \langle y,q\rangle>0\} { q ∈ S 2 : ρ ( q ) > 0 , ⟨ y , q ⟩ > 0 }
is nonempty. Indeed, otherwise the nonnegative smooth density ρ \rho ρ would be supported in a closed hemisphere, contradicting (4) . We may therefore choose q q q in this set with ⟨ ξ , q ⟩ ≠ 0 \langle\xi,q\rangle\ne0 ⟨ ξ , q ⟩ = 0 . Again, a = y − r q a=y-rq a = y − r q belongs to B 1 ( 0 ) B_{1}(0) B 1 ( 0 ) for sufficiently small r > 0 r>0 r > 0 .
In either case, the integrand in
D ^ i j ( y ) ξ i ξ j = ∫ R 3 ρ ( y − a ∣ y − a ∣ ) ⟨ ξ , y − a ⟩ 2 ∣ y − a ∣ 4 η ( a ) d a \widehat{D}^{ij}(y)\xi_i\xi_j=\int_{\mathbb{R}^{3}}\rho\left(\frac{y-a}{|y-a|}\right)\frac{\langle\xi,y-a\rangle^{2}}{|y-a|^{4}}\eta(a)\,da D ij ( y ) ξ i ξ j = ∫ R 3 ρ ( ∣ y − a ∣ y − a ) ∣ y − a ∣ 4 ⟨ ξ , y − a ⟩ 2 η ( a ) d a
is strictly positive on a nonempty open set. Thus D ^ ( y ) \widehat{D}(y) D ( y ) is positive definite on B 1 ( 0 ) ‾ \overline{B_{1}(0)} B 1 ( 0 ) . By smoothness and compactness, it remains positive definite on B R ( 0 ) ‾ \overline{B_{R}(0)} B R ( 0 ) for some R > 1 R>1 R > 1 . Since ∣ y ∣ = cot t |y|=\cot t ∣ y ∣ = cot t , this ball corresponds to the cap { t ≥ t 0 } \{t\ge t_{0}\} { t ≥ t 0 } , where t_{0}=\arccot R<\frac{\pi}{4} . The asserted uniform lower bound follows from compactness of this cap. □
We next show that the extension (7) has a representation adapted to the round geodesic flow. This additional structure will be used in the construction of the collar metric.
Lemma 3.3. Let D D D be the tensor constructed in Lemma 3.2 , extended across the equator by (7) . Choose t 0 < t 1 < t 2 < π 4 t_{0}<t_{1}<t_{2}<\frac{\pi}{4} t 0 < t 1 < t 2 < 4 π . There exist τ > 0 \tau>0 τ > 0 , κ > 0 \kappa>0 κ > 0 , and a smooth nonnegative function F r d F_{\mathrm{rd}} F rd on the round unit tangent bundle over a neighborhood of C c = [ − τ , t 2 ] × S 2 C_{c}=[-\tau,t_{2}]\times\mathbb{S}^{2} C c = [ − τ , t 2 ] × S 2 such that
D X = ∫ S X 3 F r d ( X , v ) v ♭ ⊗ v ♭ d μ X ( v ) . (8) D_{X}=\int_{\mathbb{S}_{X}^{3}}F_{\mathrm{rd}}(X,v)\,v^{\flat}\otimes v^{\flat}\,d\mu_{X}(v). \tag*{(8)} D X = ∫ S X 3 F rd ( X , v ) v ♭ ⊗ v ♭ d μ X ( v ) . (8) The function F r d F_{\mathrm{rd}} F rd is invariant under the round geodesic flow, as long as the base geodesic remains in this neighborhood, and
supp F r d ∣ C c ⊂ { ( X , v ) : v 0 ≥ κ } , v 0 = ⟨ v , ∂ t ⟩ g 0 . \left.\operatorname{supp}F_{\mathrm{rd}}\right|_{C_{c}}\subset\{(X,v):v^{0}\ge\kappa\},\qquad v^{0}=\langle v,\partial_{t}\rangle_{g_{0}}. supp F rd ∣ C c ⊂ {( X , v ) : v 0 ≥ κ } , v 0 = ⟨ v , ∂ t ⟩ g 0 . At the equator,
F r d ( 0 , q , v ) = ρ ( q ) Φ ( v ) . (9) F_{\mathrm{rd}}(0,q,v)=\rho(q)\Phi(v). \tag*{(9)} F rd ( 0 , q , v ) = ρ ( q ) Φ ( v ) . (9)
where Φ \Phi Φ is smooth, nonnegative, invariant under rotations of the tangential components, and satisfies
∫ S 2 Φ ( v ) ( v 0 ) 2 d μ ( v ) = 1 , ∫ S 2 Φ ( v ) v 0 v α d μ ( v ) = 0 , α = 1 , 2. (10) \int_{\mathbb{S}^{2}} \Phi(v)(v^{0})^{2}\,d\mu(v)=1,\qquad\int_{\mathbb{S}^{2}} \Phi(v)v^{0}v^{\alpha}\,d\mu(v)=0,\quad\alpha=1,2. \tag*{(10)} ∫ S 2 Φ ( v ) ( v 0 ) 2 d μ ( v ) = 1 , ∫ S 2 Φ ( v ) v 0 v α d μ ( v ) = 0 , α = 1 , 2. (10)
Here v α v^{\alpha} v α are the components in any orthonormal tangential frame, and the integrals are over the sphere of unit tangent directions.
Proof. We work on an open collar U U U containing C c C_c C c , with U ‾ ⊂ { ∣ t ∣ < π / 4 } \overline{U}\subset\{|t|<\pi/4\} U ⊂ { ∣ t ∣ < π /4 } . For a ∈ supp η a\in\operatorname{supp}\eta a ∈ supp η , retain the notation w a , W a w_a,W_a w a , W a from (7) , and set
ℓ a ( X ) = ⟨ X , ( a , 1 ) ⟩ , r a = ∣ W a ∣ g 0 , v a = W a r a . \ell_a(X)=\langle X,(a,1)\rangle,\qquad r_a=|W_a|_{g_0},\qquad v_a=\frac{W_a}{r_a}. ℓ a ( X ) = ⟨ X , ( a , 1 )⟩ , r a = ∣ W a ∣ g 0 , v a = r a W a .
and
λ a = ρ ( w a ∣ w a ∣ ) r a 2 ∣ w a ∣ 4 . \lambda_a=\rho\left(\frac{w_a}{|w_a|}\right)\frac{r_a^2}{|w_a|^4}. λ a = ρ ( ∣ w a ∣ w a ) ∣ w a ∣ 4 r a 2 .
These quantities are smooth on U U U . Indeed,
∣ w a ∣ ≥ cos t − ∣ sin t ∣ , ⟨ W a , ∂ t ⟩ = cos t − sin t ⟨ q , a ⟩ ≥ cos t − ∣ sin t ∣ . |w_a|\geq\cos t-|\sin t|,\qquad\langle W_a,\partial_t\rangle=\cos t-\sin t\langle q,a\rangle\geq\cos t-|\sin t|. ∣ w a ∣ ≥ cos t − ∣ sin t ∣ , ⟨ W a , ∂ t ⟩ = cos t − sin t ⟨ q , a ⟩ ≥ cos t − ∣ sin t ∣.
In particular, w a w_a w a and W a W_a W a are nonzero. Since r a ≤ ∣ ( a , 1 ) ∣ ≤ 2 r_a\leq|(a,1)|\leq\sqrt{2} r a ≤ ∣ ( a , 1 ) ∣ ≤ 2 , we also obtain a uniform lower bound
⟨ v a , ∂ t ⟩ ≥ κ > 0 on C c . \langle v_a,\partial_t\rangle\geq\kappa>0\qquad\text{on }C_c. ⟨ v a , ∂ t ⟩ ≥ κ > 0 on C c .
We first verify the geometric properties of each vector field v a v_a v a . Since W a = ∇ g 0 ℓ a W_a=\nabla^{g_0}\ell_a W a = ∇ g 0 ℓ a and ℓ a \ell_a ℓ a is the restriction of a linear function to the unit sphere,
∇ Z W a = − ℓ a Z for every Z ∈ T S 3 . \nabla_ZW_a=-\ell_a Z\qquad\text{for every }Z\in TS^3. ∇ Z W a = − ℓ a Z for every Z ∈ T S 3 .
Direct differentiation also gives
W a ( w a ) = − ℓ a w a , W a ( r a ) = − ℓ a r a . W_a(w_a)=-\ell_a w_a,\qquad W_a(r_a)=-\ell_a r_a. W a ( w a ) = − ℓ a w a , W a ( r a ) = − ℓ a r a .
It follows that
∇ v a v a = 0 , div g 0 v a = − 2 ℓ a r a , v a ( λ a ) = 2 ℓ a r a λ a . \nabla_{v_a}v_a=0,\qquad\operatorname{div}_{g_0}v_a=-\frac{2\ell_a}{r_a},\qquad v_a(\lambda_a)=\frac{2\ell_a}{r_a}\lambda_a. ∇ v a v a = 0 , div g 0 v a = − r a 2 ℓ a , v a ( λ a ) = r a 2 ℓ a λ a .
Thus the integral curves of v a v_a v a are unit-speed round geodesics, and their weighted flux satisfies
div g 0 ( λ a v a ) = 0. (11) \operatorname{div}_{g_0}(\lambda_a v_a)=0. \tag*{(11)} div g 0 ( λ a v a ) = 0. (11)
These identities also imply
div g 0 ( λ a v a ♭ ⊗ v a ♭ ) = 0. \operatorname{div}_{g_0}\left(\lambda_a v_a^\flat\otimes v_a^\flat\right)=0. div g 0 ( λ a v a ♭ ⊗ v a ♭ ) = 0.
Consequently, the extension of D D D is divergence-free on the whole two-sided collar.
For X ∈ U X \in U X ∈ U , push the measure λ a ( X ) η ( a ) d a \lambda_a(X)\eta(a)\,\mathrm{d}a λ a ( X ) η ( a ) d a forward by a ↦ v a ( X ) a \mapsto v_a(X) a ↦ v a ( X ) to the unit tangent sphere. Denote the resulting density by F r d F_{\mathrm{rd}} F rd :
∫ S X S 3 ψ ( v ) F r d ( X , v ) d μ X ( v ) = ∫ R 3 ψ ( v a ( X ) ) λ a ( X ) η ( a ) d a . (12) \int_{S_XS^3} \psi(v)F_{\mathrm{rd}}(X,v)\,\mathrm{d}\mu_X(v)=\int_{\mathbb{R}^3}\psi(v_a(X))\lambda_a(X)\eta(a)\,\mathrm{d}a. \tag*{(12)} ∫ S X S 3 ψ ( v ) F rd ( X , v ) d μ X ( v ) = ∫ R 3 ψ ( v a ( X )) λ a ( X ) η ( a ) d a . (12)
To justify that this density is smooth, write
a = u + b q , u ∈ T q S 2 . a=u+bq,\qquad u\in T_qS^2. a = u + b q , u ∈ T q S 2 .
Relative to the normal and tangential splitting at X X X , the components of W a W_a W a are
W a = ( cos t − b sin t , u ) . W_a=(\cos t-b\sin t,u). W a = ( cos t − b sin t , u ) .
On a neighborhood of supp η \operatorname{supp}\eta supp η , its first component is positive. For fixed b b b , the map
u ⟼ ( cos t − b sin t , u ) ( cos t − b sin t ) 2 + ∣ u ∣ 2 u\longmapsto\frac{(\cos t-b\sin t,u)}{\sqrt{(\cos t-b\sin t)^2+\lvert u\rvert^2}} u ⟼ ( cos t − b sin t ) 2 + ∣ u ∣ 2 ( cos t − b sin t , u )
has rank two. Hence a ↦ v a ( X ) a\mapsto v_a(X) a ↦ v a ( X ) is a submersion on a neighborhood of the support of the integrand. Integration along its fibers therefore gives a smooth density, depending smoothly on X X X . This density is nonnegative, and its support satisfies the stated transversality condition. Moreover, since
λ a v a b ⊗ v a b = ρ ( w a ∣ w a ∣ ) W a b ⊗ W a b ∣ w a ∣ 4 , \lambda_a v_a^b\otimes v_a^b=\rho\left(\frac{w_a}{\lvert w_a\rvert}\right)\frac{W_a^b\otimes W_a^b}{\lvert w_a\rvert^4}, λ a v a b ⊗ v a b = ρ ( ∣ w a ∣ w a ) ∣ w a ∣ 4 W a b ⊗ W a b ,
formula (12) gives (8) .
We now prove invariance under the round geodesic flow. Let X 0 X_0 X 0 be its generating vector field on the unit tangent bundle S U SU S U . For every ψ ∈ C ∞ ( S U ) \psi\in C^\infty(SU) ψ ∈ C ∞ ( S U ) , the identity ∇ v a v a = 0 \nabla_{v_a}v_a=0 ∇ v a v a = 0 gives
( X 0 ψ ) ( X , v a ( X ) ) = v a [ ψ ( X , v a ( X ) ) ] . (X_0\psi)(X,v_a(X))=v_a[\psi(X,v_a(X))]. ( X 0 ψ ) ( X , v a ( X )) = v a [ ψ ( X , v a ( X ))] .
Using (12) and (11) , we obtain
∫ S U F r d X 0 ψ d vol 0 ( X ) d μ X ( v ) = ∫ R 3 η ( a ) ∫ U λ a v a [ ψ ( X , v a ( X ) ) ] d vol 0 ( X ) d a = 0. \begin{aligned}
\int_{SU}F_{\mathrm{rd}}X_0\psi\,\mathrm{d}\operatorname{vol}_0(X)\,\mathrm{d}\mu_X(v)
&=\int_{\mathbb{R}^3}\eta(a)\int_U\lambda_a v_a[\psi(X,v_a(X))]\,\mathrm{d}\operatorname{vol}_0(X)\,\mathrm{d}a\\
&=0.
\end{aligned} ∫ S U F rd X 0 ψ d vol 0 ( X ) d μ X ( v ) = ∫ R 3 η ( a ) ∫ U λ a v a [ ψ ( X , v a ( X ))] d vol 0 ( X ) d a = 0.
The round geodesic flow preserves the Liouville measure d vol 0 ( X ) d μ X ( v ) \mathrm{d}\operatorname{vol}_0(X)\,\mathrm{d}\mu_X(v) d vol 0 ( X ) d μ X ( v ) . Integration by parts therefore yields X 0 F r d = 0 X_0F_{\mathrm{rd}}=0 X 0 F rd = 0 . Notice that this invariance concerns the density F r d F_{\mathrm{rd}} F rd on the unit tangent bundle; the individual weights λ a \lambda_a λ a satisfy the flux conservation law (11) .
Finally, we compute the density at the equator. Writing again a = u + b q a=u+bq a = u + b q , we have, in normal and tangential components,
W a = ( 1 , u ) , v a = ( 1 , u ) 1 + ∣ u ∣ 2 , λ a = ρ ( q ) ( 1 + ∣ u ∣ 2 ) . W_a=(1,u),\qquad v_a=\frac{(1,u)}{\sqrt{1+\lvert u\rvert^2}},\qquad\lambda_a=\rho(q)(1+\lvert u\rvert^2). W a = ( 1 , u ) , v a = 1 + ∣ u ∣ 2 ( 1 , u ) , λ a = ρ ( q ) ( 1 + ∣ u ∣ 2 ) .
Define
η t a n ( u ) = ∫ R η ( u + b q ) d b . \eta_{\mathrm{tan}}(u)=\int_{\mathbb{R}}\eta(u+bq)\,db. η tan ( u ) = ∫ R η ( u + b q ) d b .
Radial symmetry of η \eta η implies that η t a n \eta_{\mathrm{tan}} η tan depends only on ∣ u ∣ \lvert u\rvert ∣ u ∣ and is independent of q q q after an orthonormal identification of T q S 2 T_qS^2 T q S 2 with R 2 \mathbb{R}^2 R 2 . Under the parametrization
v = ( 1 , u ) 1 + ∣ u ∣ 2 , v=\frac{(1,u)}{\sqrt{1+\lvert u\rvert^2}}, v = 1 + ∣ u ∣ 2 ( 1 , u ) ,
the area element on the unit tangent sphere is
d μ ( v ) = ( 1 + ∣ u ∣ 2 ) − 3 / 2 d u . d\mu(v)=(1+\lvert u\rvert^2)^{-3/2}\,du. d μ ( v ) = ( 1 + ∣ u ∣ 2 ) − 3/2 d u .
Consequently, (9) holds with
Φ ( ( 1 , u ) 1 + ∣ u ∣ 2 ) = ( 1 + ∣ u ∣ 2 ) 5 / 2 η t a n ( u ) . \Phi\left(\frac{(1,u)}{\sqrt{1+\lvert u\rvert^2}}\right)=(1+\lvert u\rvert^2)^{5/2}\eta_{\mathrm{tan}}(u). Φ ( 1 + ∣ u ∣ 2 ( 1 , u ) ) = ( 1 + ∣ u ∣ 2 ) 5/2 η tan ( u ) .
Since η t a n \eta_{\mathrm{tan}} η tan is smooth and supported in { ∣ u ∣ ≤ 1 } \{\lvert u\rvert\leq1\} {∣ u ∣ ≤ 1 } , this formula extends by zero to a smooth function on the entire unit sphere, supported where v 0 ≥ 1 / 2 v^0\geq1/\sqrt{2} v 0 ≥ 1/ 2 . Its rotational invariance follows from that of η t a n \eta_{\mathrm{tan}} η tan . Finally,
∫ S 2 Φ ( v ) ( v 0 ) 2 d μ ( v ) = ∫ R 2 η t a n ( u ) d u = 1 , \int_{S^2}\Phi(v)(v^0)^2\,d\mu(v)=\int_{\mathbb{R}^2}\eta_{\mathrm{tan}}(u)\,du=1, ∫ S 2 Φ ( v ) ( v 0 ) 2 d μ ( v ) = ∫ R 2 η tan ( u ) d u = 1 ,
and
∫ S 2 Φ ( v ) v 0 v α d μ ( v ) = ∫ R 2 u α η t a n ( u ) d u = 0. \int_{S^2}\Phi(v)v^0v^\alpha\,d\mu(v)=\int_{\mathbb{R}^2}u^\alpha\eta_{\mathrm{tan}}(u)\,du=0. ∫ S 2 Φ ( v ) v 0 v α d μ ( v ) = ∫ R 2 u α η tan ( u ) d u = 0.
This proves (10) . □
Realization on an exact boundary collar
We now realize the tensor D D D as the leading curvature excess of a collar metric with the prescribed boundary data. The geodesic representation from Lemma 3.3 is essential near the equator: transporting its nonnegative density will give an exact curvature inequality even where D D D vanishes. We write L V g \mathcal{L}_Vg L V g for the Lie derivative of a metric g g g along a vector field V V V .
Lemma 3.4. There exist ε 0 > 0 \varepsilon_0>0 ε 0 > 0 and a fixed open collar U ⊃ [ 0 , t 1 ] × S 2 U\supset[0,t_1]\times S^2 U ⊃ [ 0 , t 1 ] × S 2 such that, for ∣ ε ∣ < ε 0 \lvert\varepsilon\rvert<\varepsilon_0 ∣ ε ∣ < ε 0 , there is a smooth family of metrics G ε = d t 2 + γ ε ( t ) G_\varepsilon=dt^2+\gamma_\varepsilon(t) G ε = d t 2 + γ ε ( t ) on U U U satisfying
G ε = 0 = g 0 , sec G ε ≥ 1 , γ ε ( 0 ) = σ , γ ˙ ε ( 0 ) = − 2 ε A . G_{\varepsilon=0}=g_0,\qquad\sec_{G_\varepsilon}\geq1,\qquad\gamma_\varepsilon(0)=\sigma,\qquad\dot{\gamma}_\varepsilon(0)=-2\varepsilon A. G ε = 0 = g 0 , sec G ε ≥ 1 , γ ε ( 0 ) = σ , γ ˙ ε ( 0 ) = − 2 ε A . :::`
Moreover,
T ( G ε ) = ε 2 D + O ( ∣ ε ∣ 3 ) , G ε = g 0 + ε L V g 0 + O ( ε 2 ) , (13) T(G_{\varepsilon})=\varepsilon^{2}D+O(|\varepsilon|^{3}),\qquad G_{\varepsilon}=g_{0}+\varepsilon\mathcal{L}_{V}g_{0}+O(\varepsilon^{2}), \tag*{(13)} T ( G ε ) = ε 2 D + O ( ∣ ε ∣ 3 ) , G ε = g 0 + ε L V g 0 + O ( ε 2 ) , (13)
where V = f ∂ t − tan t ∇ σ f V=f\partial_{t}-\tan t\nabla_{\sigma}f V = f ∂ t − tan t ∇ σ f . In particular, there exists c > 0 c>0 c > 0 such that
sec G ε ≥ 1 + c ε 2 on [ t 0 , t 1 ] × S 2 \sec G_{\varepsilon}\ge1+c\varepsilon^{2}\qquad\text{on }[t_{0},t_{1}]\times\mathbb{S}^{2} sec G ε ≥ 1 + c ε 2 on [ t 0 , t 1 ] × S 2
for all sufficiently small ∣ ε ∣ |\varepsilon| ∣ ε ∣ . The expansions hold in every fixed C k C^{k} C k norm on the collar.
Proof. We construct the metric and an auxiliary metric connection simultaneously. The evolution equations will impose the curvature components involving the normal direction. We will then verify, using a conservation law and the Bianchi identities, that the remaining curvature equations also hold and that the auxiliary connection is the Levi–Civita connection.
Step 1: the unknown metric and the evolution equations.
Choose a local oriented σ \sigma σ -orthonormal coframe { ω σ 1 , ω σ 2 } \{\omega_{\sigma}^{1},\omega_{\sigma}^{2}\} { ω σ 1 , ω σ 2 } , pulled back to the collar. We seek spatial one-forms
ω α = ( C ε ) β α ω σ β , α , β ∈ { 1 , 2 } , \omega^{\alpha}=(C_{\varepsilon})^{\alpha}_{\ \beta}\omega_{\sigma}^{\beta},\qquad\alpha,\beta\in\{1,2\}, ω α = ( C ε ) β α ω σ β , α , β ∈ { 1 , 2 } ,
and set
ω 0 = d t , G ε = ∑ a = 0 2 ω a ⊗ ω a = d t 2 + σ ( C ε ⋅ , C ε ⋅ ) . \omega^{0}=dt,\qquad G_{\varepsilon}=\sum_{a=0}^{2}\omega^{a}\otimes\omega^{a}=dt^{2}+\sigma(C_{\varepsilon}\cdot,C_{\varepsilon}\cdot). ω 0 = d t , G ε = a = 0 ∑ 2 ω a ⊗ ω a = d t 2 + σ ( C ε ⋅ , C ε ⋅ ) .
Thus the matrix C ε C_{\varepsilon} C ε determines the metric on each slice. It suffices that C ε C_{\varepsilon} C ε remain invertible. All time derivatives of spatial forms are taken after pulling them back to the fixed sphere S 2 \mathbb{S}^{2} S 2 .
Let { e a } a = 0 2 \{e_{a}\}_{a=0}^{2} { e a } a = 0 2 be the dual frame, so that e 0 = ∂ t e_{0}=\partial_{t} e 0 = ∂ t . We also seek a nonnegative function F = F ε ( t , q , v ) F=F_{\varepsilon}(t,q,v) F = F ε ( t , q , v ) on unit tangent directions. Writing
v = v a e a , ∑ a = 0 2 ( v a ) 2 = 1 , v=v^{a}e_{a},\qquad\sum_{a=0}^{2}(v^{a})^{2}=1, v = v a e a , a = 0 ∑ 2 ( v a ) 2 = 1 ,
define
M a b ( t , q ) = ∫ S v 2 F ( t , q , v ) v a v b d μ ( v ) . (14) M_{ab}(t,q)=\int_{\mathbb{S}^{2}_{v}}F(t,q,v)v^{a}v^{b}\,d\mu(v). \tag*{(14)} M ab ( t , q ) = ∫ S v 2 F ( t , q , v ) v a v b d μ ( v ) . (14)
Here S v 2 \mathbb{S}^{2}_{v} S v 2 is the sphere of velocity components, and the integration takes place at the fixed base point ( t , q ) (t,q) ( t , q ) . The matrix M M M is symmetric and nonnegative, since
M a b ξ a ξ b = ∫ S v 2 F ( t , q , v ) ( ∑ a ξ a v a ) 2 d μ ( v ) ≥ 0. M_{ab}\xi^{a}\xi^{b}=\int_{\mathbb{S}^{2}_{v}}F(t,q,v)\left(\sum_{a}\xi^{a}v^{a}\right)^{2}\,d\mu(v)\ge0. M ab ξ a ξ b = ∫ S v 2 F ( t , q , v ) ( a ∑ ξ a v a ) 2 d μ ( v ) ≥ 0.
Our objective is to construct G ε G_{\varepsilon} G ε so that
T ( G ε ) = ε 2 M a b ω a ⊗ ω b . (15) T(G_{\varepsilon})=\varepsilon^{2}M_{ab}\,\omega^{a}\otimes\omega^{b}. \tag*{(15)} T ( G ε ) = ε 2 M ab ω a ⊗ ω b . (15)
Indeed, for a G ε G_\varepsilon G ε -unit vector n = n a e a n=n^a e_a n = n a e a , this identity would give
sec G ε ( n ⊥ ) − 1 = ε 2 ∫ S v 2 F ( t , q , v ) ( ∑ a n v v a ) 2 d μ ( v ) ≥ 0. \sec_{G_\varepsilon}(n^\perp)-1=\varepsilon^2\int_{S_v^2}F(t,q,v)\left(\sum_a n^v v^a\right)^2\,d\mu(v)\geq0. sec G ε ( n ⊥ ) − 1 = ε 2 ∫ S v 2 F ( t , q , v ) ( a ∑ n v v a ) 2 d μ ( v ) ≥ 0.
To obtain a closed evolution system, we introduce an auxiliary metric connection with connection forms Γ b a \Gamma^a_b Γ b a . At this stage it is not assumed to be torsion-free. We impose
Γ 0 α = ω ˙ α , Γ b a = − Γ a b , Γ b a ( ∂ t ) = 0 , \Gamma^\alpha_0=\dot{\omega}^\alpha,\qquad\Gamma^a_b=-\Gamma^b_a,\qquad\Gamma^a_b(\partial_t)=0, Γ 0 α = ω ˙ α , Γ b a = − Γ a b , Γ b a ( ∂ t ) = 0 ,
and regard the remaining spatial form Γ 2 1 \Gamma^1_2 Γ 2 1 as an independent unknown. We use the coefficient convention
Γ c b = Γ c a b ω a , ∇ e a Γ e c = Γ c a b e b . \Gamma^b_c=\Gamma^b_{ca}\omega^a,\qquad\nabla^\Gamma_{e_a}e_c=\Gamma^b_{ca}e_b. Γ c b = Γ c a b ω a , ∇ e a Γ e c = Γ c a b e b .
The curvature forms corresponding to the desired curvature operator are prescribed by
\begin{pmatrix}
K^1_2\\
K^2_0\\
K^0_1
\end{pmatrix}
=(I+\varepsilon^2M)
\begin{pmatrix}
\omega^1\wedge\omega^2\\
\omega^2\wedge\omega^0\\
\omega^0\wedge\omega^1
\end{pmatrix}. \tag*{(16)}\end{pmatrix}
The three rows correspond to the planes with normals e 0 , e 1 , e 2 e_0,e_1,e_2 e 0 , e 1 , e 2 , respectively. Once these are the actual curvature forms of G ε G_\varepsilon G ε , this prescription is equivalent to (15) . First require that
d Γ b a + Γ c a ∧ Γ b c d\Gamma^a_b+\Gamma^a_c\wedge\Gamma^c_b d Γ b a + Γ c a ∧ Γ b c
agree with K b a K^a_b K b a on pairs containing ∂ t \partial_t ∂ t . Together with transport of F F F along the geodesics of the auxiliary connection, this gives
ω ¨ α + ω α = − ε 2 ( ( M 11 + M 22 ) ω α − M α β ω β ) , Γ ˙ 2 1 = ε 2 ( M 02 ω 1 − M 01 ω 2 ) , v a e a ( F ) − Γ c a b v a v c ∂ F ∂ v b = 0. (17) \begin{aligned}
\ddot{\omega}^\alpha+\omega^\alpha&=-\varepsilon^2\left((M_{11}+M_{22})\omega^\alpha-M_{\alpha\beta}\omega^\beta\right),\\
\dot{\Gamma}^1_2&=\varepsilon^2(M_{02}\omega^1-M_{01}\omega^2),\\
v^a e_a(F)-\Gamma^b_{ca}v^a v^c\frac{\partial F}{\partial v_b}&=0. \tag*{(17)}
\end{aligned} ω ¨ α + ω α Γ ˙ 2 1 v a e a ( F ) − Γ c a b v a v c ∂ v b ∂ F = − ε 2 ( ( M 11 + M 22 ) ω α − M α β ω β ) , = ε 2 ( M 02 ω 1 − M 01 ω 2 ) , = 0. (17)
In the last equation, e a ( F ) e_a(F) e a ( F ) differentiates the base point while holding the velocity components fixed. The second term accounts for the change of those components along an auxiliary geodesic. Skew-symmetry of Γ \Gamma Γ makes this vertical vector field tangent to S v 2 S_v^2 S v 2 , so the equation is well defined on unit directions. We prescribe the initial data
ω α ( 0 ) = ω σ α , ω ˙ α ( 0 ) = − ε A α β ω σ β , \omega^\alpha(0)=\omega^\alpha_\sigma,\qquad\dot{\omega}^\alpha(0)=-\varepsilon A^\alpha{}_\beta\omega^\beta_\sigma, ω α ( 0 ) = ω σ α , ω ˙ α ( 0 ) = − ε A α β ω σ β ,
Γ 2 1 ( 0 ) = Γ 2 , σ 1 , F ( 0 , q , v ) = ρ ( q ) Φ ( v ) . (18) \Gamma^1_2(0)=\Gamma^1_{2,\sigma},\qquad F(0,q,v)=\rho(q)\Phi(v). \tag*{(18)} Γ 2 1 ( 0 ) = Γ 2 , σ 1 , F ( 0 , q , v ) = ρ ( q ) Φ ( v ) . (18)
where Γ 2 , σ 1 \Gamma^1_{2,\sigma} Γ 2 , σ 1 is the round connection form and Φ \Phi Φ is supplied by Lemma 3.3 .
The first equation in (17) is a perturbed Jacobi equation. More explicitly, set
M ^ = ( M 22 − M 12 − M 12 M 11 ) . \widehat{M}=\begin{pmatrix} M_{22} & -M_{12} \\ -M_{12} & M_{11} \end{pmatrix}. M = ( M 22 − M 12 − M 12 M 11 ) .
Then
C ¨ ε + C ε = − ε 2 M ^ C ε , C ε ( 0 ) = I , C ˙ ε ( 0 ) = − ε A . \ddot{C}_{\varepsilon}+C_{\varepsilon}=-\varepsilon^{2}\widehat{M}C_{\varepsilon},\qquad C_{\varepsilon}(0)=I,\qquad\dot{C}_{\varepsilon}(0)=-\varepsilon A. C ¨ ε + C ε = − ε 2 M C ε , C ε ( 0 ) = I , C ˙ ε ( 0 ) = − ε A .
Consequently,
C ε ( t ) = cos t I − ε sin t A − ε 2 ∫ 0 t sin ( t − s ) M ^ ( s ) C ε ( s ) d s . (19) C_{\varepsilon}(t)=\cos t I-\varepsilon\sin t A-\varepsilon^{2}\int_{0}^{t}\sin(t-s)\widehat{M}(s)C_{\varepsilon}(s)\,\mathrm{d}s. \tag*{(19)} C ε ( t ) = cos t I − ε sin t A − ε 2 ∫ 0 t sin ( t − s ) M ( s ) C ε ( s ) d s . (19)
This exhibits the desired metric as a correction of order ε 2 \varepsilon^{2} ε 2 to the cosine collar. It is an integral equation: the matrix M ^ \widehat{M} M still depends on the unknown density F F F .
Step 2: existence on a fixed collar.
We give the existence argument before checking the remaining geometric constraints. Its useful feature is that the auxiliary connection is an independent unknown. Thus the coefficients in the transport equation involve the coframe and the connection, but no spatial derivatives of these unknowns.
Choose a compact interval
J = [ t − , t + ] ⊂ ( − τ , t 2 ) , t − < 0 < t 1 < t + . J=[t_{-},t_{+}]\subset(-\tau,t_{2}),\qquad t_{-}<0<t_{1}<t_{+}. J = [ t − , t + ] ⊂ ( − τ , t 2 ) , t − < 0 < t 1 < t + .
The round density F r d F_{\mathrm{rd}} F rd from Lemma 3.3 is defined on a neighborhood of this interval, and its support satisfies v 0 ≥ κ v^{0}\geq\kappa v 0 ≥ κ for some κ > 0 \kappa>0 κ > 0 .
For the analytic argument, regard the velocity variable as belonging to the fixed sphere bundle P = S ( R ⊕ T S 2 ) P=S(\mathbb{R}\oplus TS^{2}) P = S ( R ⊕ T S 2 ) , with the norm induced by 1 ⊕ σ 1\oplus\sigma 1 ⊕ σ . This is a compact four-dimensional manifold. At ( t , q ) (t,q) ( t , q ) , we identify its fiber with the G ε G_{\varepsilon} G ε -unit tangent sphere by
( v 0 , v ⊤ ) ⟼ v 0 ∂ t + C ε − 1 v ⊤ , v ⊤ ∈ T q S 2 , (v^{0},v^{\top})\longmapsto v^{0}\partial_{t}+C_{\varepsilon}^{-1}v^{\top},\qquad v^{\top}\in T_{q}S^{2}, ( v 0 , v ⊤ ) ⟼ v 0 ∂ t + C ε − 1 v ⊤ , v ⊤ ∈ T q S 2 ,
where the second term is viewed as tangent to the slice { t } × S 2 \{t\}\times S^{2} { t } × S 2 . We use the round case C 0 ( t ) = cos t I C_{0}(t)=\cos t I C 0 ( t ) = cos t I to regard F r d F_{\mathrm{rd}} F rd as a function on J × P J\times P J × P . Under time-independent changes of the local oriented orthonormal coframe, C ε C_{\varepsilon} C ε and C ˙ ε \dot{C}_{\varepsilon} C ˙ ε transform as bundle endomorphisms, while ζ = Γ 2 1 − Γ 2 , σ 1 \zeta=\Gamma^{1}_{2}-\Gamma^{1}_{2,\sigma} ζ = Γ 2 1 − Γ 2 , σ 1 is a globally defined one-form. We write Z = ( C ε , C ˙ ε , ζ ) Z=(C_{\varepsilon},\dot{C}_{\varepsilon},\zeta) Z = ( C ε , C ˙ ε , ζ ) .
On the transverse region, division of the transport equation by v 0 v^{0} v 0 puts it in evolution form. To solve an equation on all of P P P , choose a smooth function ϑ ( v 0 ) \vartheta(v^{0}) ϑ ( v 0 ) such that
ϑ ( v 0 ) = 1 v 0 for v 0 ≥ κ / 2 , ϑ ( v 0 ) = 0 for v 0 ≤ κ / 4. \vartheta(v^{0})=\frac{1}{v^{0}}\quad\text{for }v^{0}\geq\kappa/2,\qquad\vartheta(v^{0})=0\quad\text{for }v^{0}\leq\kappa/4. ϑ ( v 0 ) = v 0 1 for v 0 ≥ κ /2 , ϑ ( v 0 ) = 0 for v 0 ≤ κ /4.
Initially replace division by v 0 v^{0} v 0 by multiplication by ϑ ( v 0 ) \vartheta(v^{0}) ϑ ( v 0 ) . The resulting equation is
∂ t F + B ( Z ) F = 0 , \partial_{t}F+B(Z)F=0, ∂ t F + B ( Z ) F = 0 ,
where B ( Z ) B(Z) B ( Z ) is a smooth vector field on P P P . As long as C ε C_\varepsilon C ε is invertible, its coefficients depend smoothly on Z Z Z , C ε − 1 C_\varepsilon^{-1} C ε − 1 , and the fixed background data, without involving spatial derivatives of Z Z Z .
For an integer m ≥ 5 m \ge5 m ≥ 5 , use H m ( S 2 ) H^m(S^2) H m ( S 2 ) for Z Z Z and H m ( P ) H^m(P) H m ( P ) for F F F . Fiber integration gives a bounded map
F ⟼ M , H m ( P ) ⟶ H m ( S 2 ) . F \longmapsto M,\qquad H^m(P) \longrightarrow H^m(S^2). F ⟼ M , H m ( P ) ⟶ H m ( S 2 ) .
The transport energy estimate has the form
d d t ∥ F ∥ H m ≤ C ( ∥ B ( Z ) ∥ W 1 , ∞ ∥ F ∥ H m + ∥ B ( Z ) ∥ H m ∥ F ∥ W 1 , ∞ ) . \frac{d}{dt}\|F\|_{H^m} \le C\left(\|B(Z)\|_{W^{1,\infty}}\|F\|_{H^m}+\|B(Z)\|_{H^m}\|F\|_{W^{1,\infty}}\right). d t d ∥ F ∥ H m ≤ C ( ∥ B ( Z ) ∥ W 1 , ∞ ∥ F ∥ H m + ∥ B ( Z ) ∥ H m ∥ F ∥ W 1 , ∞ ) .
The Sobolev embedding theorem controls the first derivatives in L ∞ L^\infty L ∞ . Since the coefficients contain no spatial derivatives of Z Z Z , this estimate closes together with the evolution estimates for Z Z Z . Picard iteration, with uniform bounds in H m H^m H m and convergence in H m − 1 H^{m-1} H m − 1 , therefore gives a unique local solution.
To extend this solution over the fixed interval J J J , compare it with the round solution
C 0 ( t ) = cos t I , C ˙ 0 ( t ) = − sin t I , ζ 0 ( t ) = 0 , F 0 = F r d . C_0(t)=\cos t\,I,\qquad\dot C_0(t)=-\sin t\,I,\qquad\zeta_0(t)=0,\qquad F_0=F_{\mathrm{rd}}. C 0 ( t ) = cos t I , C ˙ 0 ( t ) = − sin t I , ζ 0 ( t ) = 0 , F 0 = F rd .
The round coframe is uniformly invertible on J J J . Moreover, this reference solution satisfies the modified transport equation, because the modification is inactive on its support.
Set
E m ( t ) = ∥ Z ε ( t ) − Z 0 ( t ) ∥ H m ( S 2 ) + ∥ F ε ( t ) − F r d ( t ) ∥ H m ( P ) . E_m(t)=\|Z_\varepsilon(t)-Z_0(t)\|_{H^m(S^2)}+\|F_\varepsilon(t)-F_{\mathrm{rd}}(t)\|_{H^m(P)}. E m ( t ) = ∥ Z ε ( t ) − Z 0 ( t ) ∥ H m ( S 2 ) + ∥ F ε ( t ) − F rd ( t ) ∥ H m ( P ) .
The initial data give E m ( 0 ) ≤ C m ∣ ε ∣ E_m(0)\le C_m|\varepsilon| E m ( 0 ) ≤ C m ∣ ε ∣ . For the transport difference, we use
∂ t ( F ε − F r d ) + B ( Z ε ) ( F ε − F r d ) = − ( B ( Z ε ) − B ( Z 0 ) ) F r d . \partial_t(F_\varepsilon-F_{\mathrm{rd}})+B(Z_\varepsilon)(F_\varepsilon-F_{\mathrm{rd}})=-\left(B(Z_\varepsilon)-B(Z_0)\right)F_{\mathrm{rd}}. ∂ t ( F ε − F rd ) + B ( Z ε ) ( F ε − F rd ) = − ( B ( Z ε ) − B ( Z 0 ) ) F rd .
The derivative on the right-hand side falls on the known smooth function F r d F_{\mathrm{rd}} F rd . On a fixed small neighborhood of the round solution in the H 5 H^5 H 5 norms, the tame energy estimates give
d d t E m ≤ C m ( E m + ε 2 ) . \frac{d}{dt}E_m\le C_m(E_m+\varepsilon^2). d t d E m ≤ C m ( E m + ε 2 ) .
The same estimate holds after reversing time. Gronwall’s inequality and a continuation argument yield
sup t ∈ J E m ( t ) ≤ C m ∣ ε ∣ . \sup_{t\in J}E_m(t)\le C_m|\varepsilon|. t ∈ J sup E m ( t ) ≤ C m ∣ ε ∣.
First apply this estimate at m = 5 m=5 m = 5 to choose a single sufficiently small parameter interval. It preserves the invertibility of C ε C_\varepsilon C ε and gives a solution throughout J J J . Higher-order tame estimates are linear in the higher norms, with constants depending on m m m , the fixed smooth background, and the controlled low-order norms. They therefore propagate every Sobolev norm on the same interval, without further shrinking the parameter interval.
The characteristic vector fields are uniformly close to the round characteristic field. Hence the characteristics issuing from the support of ρ Φ \rho\Phi ρ Φ remain close to their round counterparts on J J J . After reducing ε 0 \varepsilon_0 ε 0 , their velocity components satisfy v 0 ≥ κ / 2 v^0 \ge\kappa/2 v 0 ≥ κ /2 . Thus the modification is inactive on the support of F ε F_\varepsilon F ε , and the solution satisfies the original transport equation in (17) . Constancy along characteristics also gives F ε ≥ 0 F_\varepsilon\ge0 F ε ≥ 0 .
Differentiating the equations with respect to ε \varepsilon ε , using higher spatial regularity at each differentiation, gives smooth dependence on ε \varepsilon ε . Finally, covariance under time-independent frame changes and uniqueness identify the local solutions on overlaps. We therefore obtain globally defined objects on the fixed open collar U = ( t − , t + ) × S 2 U=(t_-,t_+)\times S^2 U = ( t − , t + ) × S 2 .
Step 3: the conservation law for the moment tensor.
The transport equation implies a conservation law for M M M that will preserve the curvature constraints. This calculation uses only the skew-symmetry of the auxiliary connection.
At a fixed base point, the vertical transport vector field on the velocity sphere has components
B b ( v ) = − Γ c a b v a v c . B^b(v)=-\Gamma^b_{ca}v^av^c. B b ( v ) = − Γ c a b v a v c .
It is tangent to S v 2 S^2_v S v 2 and satisfies
div S v 2 B = − Γ c a a v c . \operatorname{div}_{S^2_v}B=-\Gamma^a_{ca}v^c. div S v 2 B = − Γ c a a v c .
Multiplying the transport equation by v b v^b v b and integrating over S v 2 S^2_v S v 2 , we obtain
e a M a b + Γ c a a M c b + Γ c a b M a c = 0. (20) e_aM^{ab}+\Gamma^a_{ca}M^{cb}+\Gamma^b_{ca}M^{ac}=0. \tag*{(20)} e a M ab + Γ c a a M c b + Γ c a b M a c = 0. (20)
This is precisely the covariant divergence equation
M a b ; a = 0 M^{ab}{}_{;a}=0 M ab ; a = 0
for the auxiliary connection. In particular, it holds before that connection has been shown to be Levi–Civita.
Step 4: verification of the geometric constraints.
Define the torsion and curvature errors by
Θ a = d ω a + Γ b a ∧ ω b , E b a = d Γ b a + Γ c a ∧ Γ b c − K b a . \Theta^a=d\omega^a+\Gamma^a_b\wedge\omega^b,\qquad E^a_b=d\Gamma^a_b+\Gamma^a_c\wedge\Gamma^c_b-K^a_b. Θ a = d ω a + Γ b a ∧ ω b , E b a = d Γ b a + Γ c a ∧ Γ b c − K b a .
The definition Γ 0 a = ω ˙ a \Gamma^a_0=\dot{\omega}^a Γ 0 a = ω ˙ a implies ι ∂ t Θ a = 0 \iota_{\partial_t}\Theta^a=0 ι ∂ t Θ a = 0 . Likewise, the first two evolution equations in (17) imply ι ∂ t E b a = 0 \iota_{\partial_t}E^a_b=0 ι ∂ t E b a = 0 . Thus both errors are spatial two-forms, which we view as time-dependent two-forms on the fixed sphere S 2 S^2 S 2 .
Symmetry of M M M implies the algebraic Bianchi identity
K b a ∧ ω b = 0. K^a_b\wedge\omega^b=0. K b a ∧ ω b = 0.
The Cartan identities therefore give
d Γ Θ a = E b a ∧ ω b , d Γ E = − d Γ K . (21) d^{\Gamma}\Theta^{a}=E^{a}_{\ b}\wedge\omega^{b},\qquad d^{\Gamma}E=-d^{\Gamma}K. \tag*{(21)} d Γ Θ a = E b a ∧ ω b , d Γ E = − d Γ K . (21)
We explain how the conservation law removes the derivatives of M M M from the second identity. Introduce the cyclic vector of two-forms
β = ( ω 1 ∧ ω 2 ω 2 ∧ ω 0 ω 0 ∧ ω 1 ) . \beta=\begin{pmatrix}
\omega^{1}\wedge\omega^{2}\\
\omega^{2}\wedge\omega^{0}\\
\omega^{0}\wedge\omega^{1}
\end{pmatrix}. β = ω 1 ∧ ω 2 ω 2 ∧ ω 0 ω 0 ∧ ω 1 .
Because the torsion forms are spatial,
d Γ β = d t ∧ ( 0 Θ 2 − Θ 1 ) . d^{\Gamma}\beta=dt\wedge\begin{pmatrix}
0\\
\Theta^{2}\\
-\Theta^{1}
\end{pmatrix}. d Γ β = d t ∧ 0 Θ 2 − Θ 1 .
Write the prescribed curvature as the cyclic vector
K = ( K 2 1 K 0 2 K 1 0 ) = ( I + ε 2 M ) β , K=\begin{pmatrix}
K^{1}_{\ 2}\\
K^{2}_{\ 0}\\
K^{0}_{\ 1}
\end{pmatrix}=(I+\varepsilon^{2}M)\beta, K = K 2 1 K 0 2 K 1 0 = ( I + ε 2 M ) β ,
so the covariant product rule gives
d Γ K = ε 2 ( d Γ M ) ∧ β + ( I + ε 2 M ) d Γ β . d^{\Gamma}K=\varepsilon^{2}(d^{\Gamma}M)\wedge\beta+(I+\varepsilon^{2}M)d^{\Gamma}\beta. d Γ K = ε 2 ( d Γ M ) ∧ β + ( I + ε 2 M ) d Γ β .
By symmetry of M M M and (20) ,
[ ( d Γ M ) ∧ β ] a = M ; b a b ω 0 ∧ ω 1 ∧ ω 2 = 0. \left[(d^{\Gamma}M)\wedge\beta\right]_{a}=M^{ab}_{\ ;b}\omega^{0}\wedge\omega^{1}\wedge\omega^{2}=0. [ ( d Γ M ) ∧ β ] a = M ; b ab ω 0 ∧ ω 1 ∧ ω 2 = 0.
Thus only the terms linear in the torsion Θ \Theta Θ remain. Taking the components containing d t dt d t in (21) now gives
Θ ˙ a = E 0 a , \dot{\Theta}^{a}=E^{a}_{\ 0}, Θ ˙ a = E 0 a ,
E ˙ 0 α = − Θ α − ε 2 ( ( M 11 + M 22 ) Θ α − M α β Θ β ) , \dot{E}^{\alpha}_{\ 0}=-\Theta^{\alpha}-\varepsilon^{2}\left((M_{11}+M_{22})\Theta^{\alpha}-M_{\alpha\beta}\Theta^{\beta}\right), E ˙ 0 α = − Θ α − ε 2 ( ( M 11 + M 22 ) Θ α − M α β Θ β ) ,
E ˙ 2 1 = ε 2 ( M 02 Θ 1 − M 01 Θ 2 ) . (22) \dot{E}^{1}_{\ 2}=\varepsilon^{2}(M_{02}\Theta^{1}-M_{01}\Theta^{2}). \tag*{(22)} E ˙ 2 1 = ε 2 ( M 02 Θ 1 − M 01 Θ 2 ) . (22)
For each fixed q q q , these form a homogeneous linear system of ordinary differential equations for the error forms.
We verify that their initial values vanish. The round surface connection and the symmetry of A A A give Θ a ( 0 ) = 0 \Theta^{a}(0)=0 Θ a ( 0 ) = 0 . The normalization of Φ \Phi Φ gives
M 00 ( 0 ) = ρ , M 0 α ( 0 ) = 0. M_{00}(0)=\rho,\qquad M_{0\alpha}(0)=0. M 00 ( 0 ) = ρ , M 0 α ( 0 ) = 0.
The Codazzi equation for A A A then yields
E 0 α ( 0 ) = 0. E^{\alpha}_{\ 0}(0)=0. E 0 α ( 0 ) = 0.
For the tangential curvature component, the initial connection has
( d Γ 2 1 + Γ 0 1 ∧ Γ 2 0 ) ∣ t = 0 = ( 1 − ε 2 det A ) d μ σ , \left.\left(d\Gamma^{1}_{2}+\Gamma^{1}_{0}\wedge\Gamma^{0}_{2}\right)\right|_{t=0}=\left(1-\varepsilon^{2}\det A\right)d\mu_{\sigma}, ( d Γ 2 1 + Γ 0 1 ∧ Γ 2 0 ) t = 0 = ( 1 − ε 2 det A ) d μ σ ,
whereas the prescribed curvature form restricts to
K 2 1 ∣ t = 0 = ( 1 + ε 2 ρ ) d μ σ . \left.K^{1}_{\ 2}\right|_{t=0}=\left(1+\varepsilon^{2}\rho\right)d\mu_{\sigma}. K 2 1 t = 0 = ( 1 + ε 2 ρ ) d μ σ .
Consequently,
E 2 1 ( 0 ) = − ε 2 ( det A + ρ ) d μ σ = 0. E^{1}_{2}(0)=-\varepsilon^{2}\left(\det A+\rho\right)d\mu_{\sigma}=0. E 2 1 ( 0 ) = − ε 2 ( det A + ρ ) d μ σ = 0.
Uniqueness for (22) therefore gives
Θ = 0 , E = 0 throughout U . \Theta=0,\qquad E=0\qquad\text{throughout }U. Θ = 0 , E = 0 throughout U .
The auxiliary connection is thus metric-compatible and torsion-free, hence is the Levi–Civita connection of G ε G_{\varepsilon} G ε . Its curvature forms are exactly (16) . This proves the exact identity
T ( G ε ) = ε 2 M a b ω a ⊗ ω b ≥ 0 , T(G_{\varepsilon})=\varepsilon^{2}M_{ab}\,\omega^{a}\otimes\omega^{b}\geq0, T ( G ε ) = ε 2 M ab ω a ⊗ ω b ≥ 0 ,
and therefore sec G ε ≥ 1 \operatorname{sec}_{G_{\varepsilon}}\geq1 sec G ε ≥ 1 on U U U .
Step 5: boundary data and the asymptotic expansions.
The initial coframe data immediately give
γ ε ( 0 ) = σ , γ ˙ ε ( 0 ) = − 2 ε A . \gamma_{\varepsilon}(0)=\sigma,\qquad\dot{\gamma}_{\varepsilon}(0)=-2\varepsilon A. γ ε ( 0 ) = σ , γ ˙ ε ( 0 ) = − 2 ε A .
Since the outward unit normal at t = 0 t=0 t = 0 is − ∂ t -\partial_t − ∂ t , the boundary second fundamental form is
h G ε = − 1 2 γ ˙ ε ( 0 ) = ε A . h_{G_{\varepsilon}}=-\frac{1}{2}\dot{\gamma}_{\varepsilon}(0)=\varepsilon A. h G ε = − 2 1 γ ˙ ε ( 0 ) = ε A .
At ε = 0 \varepsilon=0 ε = 0 , the metric and connection are round, and uniqueness for the transport equation gives F 0 = F r d F_{0}=F_{\mathrm{rd}} F 0 = F rd . Hence Lemma 3.3 implies
M a b ω a ⊗ ω b ∣ ε = 0 = D . \left.M_{ab}\omega^{a}\otimes\omega^{b}\right|_{\varepsilon=0}=D. M ab ω a ⊗ ω b ε = 0 = D .
Smooth dependence on ε \varepsilon ε yields
T ( G ε ) = ε 2 D + O ( ∣ ε ∣ 3 ) . T(G_{\varepsilon})=\varepsilon^{2}D+O\left(|\varepsilon|^{3}\right). T ( G ε ) = ε 2 D + O ( ∣ ε ∣ 3 ) .
On the slab [ t 0 , t 1 ] × S 2 [t_{0},t_{1}]\times S^{2} [ t 0 , t 1 ] × S 2 , we have D ≥ c 0 g 0 D\geq c_{0}g_{0} D ≥ c 0 g 0 . Thus, for sufficiently small ∣ ε ∣ |\varepsilon| ∣ ε ∣ ,
T ( G ε ) ≥ c 0 2 ε 2 g 0 ≥ c ε 2 G ε T(G_{\varepsilon})\geq\frac{c_{0}}{2}\varepsilon^{2}g_{0}\geq c\varepsilon^{2}G_{\varepsilon} T ( G ε ) ≥ 2 c 0 ε 2 g 0 ≥ c ε 2 G ε
for some c > 0 c>0 c > 0 . This is equivalent to the asserted strict sectional curvature bound. Finally, differentiating (19) at ε = 0 \varepsilon=0 ε = 0 gives
∂ ε C ε ∣ ε = 0 = − sin t A . \left.\partial_{\varepsilon}C_{\varepsilon}\right|_{\varepsilon=0}=-\sin t\,A. ∂ ε C ε ∣ ε = 0 = − sin t A .
Therefore the first metric variation has no normal or mixed components, and its tangential part is
∂ ε G ε ∣ ε = 0 = − 2 sin t cos t A . \left.\partial_{\varepsilon}G_{\varepsilon}\right|_{\varepsilon=0}=-2\sin t\cos t\,A. ∂ ε G ε ∣ ε = 0 = − 2 sin t cos t A .
The vector field V = f ∂ t − tan t ∇ σ f V=f\partial_t-\tan t\nabla_{\sigma}f V = f ∂ t − tan t ∇ σ f has the following expression along X ( t , q ) = ( cos t q , sin t ) X(t,q)=(\cos t\,q,\sin t) X ( t , q ) = ( cos t q , sin t ) in R 3 × R \mathbb{R}^{3}\times\mathbb{R} R 3 × R :
V ( t , q ) = ( − sin t F f ( q ) , cos t f ( q ) ) . V(t,q)=(-\sin t\,F_f(q),\cos t\,f(q)). V ( t , q ) = ( − sin t F f ( q ) , cos t f ( q )) .
For Y ∈ T q S 2 Y\in T_qS^2 Y ∈ T q S 2 , Euclidean differentiation and d F f = A dF_f=A d F f = A give
∂ t V = ( − cos t F f , − sin t f ) , d q V ( Y ) = ( − sin t A ( Y ) , cos t Y f ) . \partial_tV=(-\cos t\,F_f,-\sin t\,f),\qquad d_qV(Y)=(-\sin t\,A(Y),\cos t\,Yf). ∂ t V = ( − cos t F f , − sin t f ) , d q V ( Y ) = ( − sin t A ( Y ) , cos t Y f ) .
Pairing these derivatives with ∂ t X = ( − sin t q , cos t ) \partial_tX=(-\sin t\,q,\cos t) ∂ t X = ( − sin t q , cos t ) and d q X ( Y ) = ( cos t Y , 0 ) d_qX(Y)=(\cos t\,Y,0) d q X ( Y ) = ( cos t Y , 0 ) yields
( Lie V g 0 ) ( ∂ t , ∂ t ) = 0 , ( Lie V g 0 ) ( ∂ t , Y ) = 0 , (\operatorname{Lie}_Vg_0)(\partial_t,\partial_t)=0,\qquad(\operatorname{Lie}_Vg_0)(\partial_t,Y)=0, ( Lie V g 0 ) ( ∂ t , ∂ t ) = 0 , ( Lie V g 0 ) ( ∂ t , Y ) = 0 ,
( Lie V g 0 ) ∣ T S 2 = − 2 sin t cos t A . (\operatorname{Lie}_Vg_0)|_{TS^2}=-2\sin t\cos t\,A. ( Lie V g 0 ) ∣ T S 2 = − 2 sin t cos t A .
This proves the second expansion in (13) . □
The curvature correction with prescribed support
To extend the collar metric, we need to prescribe its quadratic curvature term without changing the metric near the boundary. The compatibility condition is the linearized Bianchi identity; gnomonic coordinates will turn it into an ordinary Euclidean divergence equation. Let
L k = d d λ ∣ λ = 0 T ( g 0 + λ k ) Lk=\left.\frac{d}{d\lambda}\right|_{\lambda=0}T(g_0+\lambda k) L k = d λ d λ = 0 T ( g 0 + λk )
be the linearization of T T T at the round metric. If a smooth extension g ‾ ε \overline{g}_{\varepsilon} g ε satisfies
T ( g ‾ ε ) = ε 2 E + O ( ∣ ε ∣ 3 ) , T(\overline{g}_{\varepsilon})=\varepsilon^2E+O(|\varepsilon|^3), T ( g ε ) = ε 2 E + O ( ∣ ε ∣ 3 ) ,
then adding a metric correction ε 2 k \varepsilon^2k ε 2 k gives
T ( g ‾ ε + ε 2 k ) = ε 2 ( E + L k ) + O ( ∣ ε ∣ 3 ) . T(\overline{g}_{\varepsilon}+\varepsilon^2k)=\varepsilon^2(E+Lk)+O(|\varepsilon|^3). T ( g ε + ε 2 k ) = ε 2 ( E + L k ) + O ( ∣ ε ∣ 3 ) .
Thus, to replace the quadratic curvature term E E E by D D D , we will solve L k = D − E Lk=D-E L k = D − E . The correction must be supported in the interior cap so that the exact collar metric is preserved.
Naturality of T T T and the identity T ( g 0 ) = 0 T(g_0)=0 T ( g 0 ) = 0 imply
L ( Lie W g 0 ) = 0 , div g 0 ( L k ) = 0. (23) L(\operatorname{Lie}_Wg_0)=0,\qquad\operatorname{div}_{g_0}(Lk)=0. \tag*{(23)} L ( Lie W g 0 ) = 0 , div g 0 ( L k ) = 0. (23)
The second identity follows by differentiating the contracted Bianchi identity at g 0 g_0 g 0 . It gives the necessary compatibility condition for the linearized equation. All divergences in this linearized problem are taken with respect to the fixed round metric g 0 g_0 g 0 .
Lemma 3.5. Let Ω = { t > a } \Omega=\{t>a\} Ω = { t > a } , and 0 < a < π 2 0<a<\frac{\pi}{2} 0 < a < 2 π . Suppose that S S S is a smooth symmetric two-tensor on H H H satisfying
div g 0 S = 0 , supp S ⊂ Ω ‾ . \operatorname{div}_{g_0} S=0,\qquad\operatorname{supp} S\subset\overline{\Omega}. div g 0 S = 0 , supp S ⊂ Ω . Then there exists a smooth symmetric two-tensor k k k on H H H such that
L k = S , supp k ⊂ Ω ‾ . Lk=S,\qquad\operatorname{supp} k\subset\overline{\Omega}. L k = S , supp k ⊂ Ω . In particular, k k k is smooth across ∂ Ω \partial\Omega ∂ Ω and vanishes to all orders there. The tensor k k k can be chosen linearly in S S S .
The proof uses the gnomonic coordinates to turn the linearized curvature equation into a Euclidean double-curl equation. We construct its potential by integrating along outward rays from a point inside the ball. This direction of integration preserves support: a ray starting outside the ball and moving away from an interior point never encounters the prescribed tensor. The potential may be singular at the chosen center, but averaging over the centers removes this singularity while preserving both the equation and the support.
Proof. In gnomonic coordinates, Ω \Omega Ω corresponds to the Euclidean ball B R ( 0 ) B_R(0) B R ( 0 ) , where R = cot a R=\cot a R = cot a . Write s = 1 + ∣ y ∣ 2 s=1+|y|^2 s = 1 + ∣ y ∣ 2 . For a Euclidean matrix field p p p , we use row-wise divergence and curl:
( Div p ) i = ∂ j p i j , ( Curl p ) i j = ϵ j k ℓ ∂ k p i ℓ . (\operatorname{Div}p)_i=\partial_jp_{ij},\qquad(\operatorname{Curl}p)_{ij}=\epsilon_{jk\ell}\partial_kp_{i\ell}. ( Div p ) i = ∂ j p ij , ( Curl p ) ij = ϵ j k ℓ ∂ k p i ℓ .
where ϵ i j k \epsilon_{ijk} ϵ ij k is the Euclidean alternating symbol. For brevity, set
Inc p = Curl ( ( Curl p ) T ) . \operatorname{Inc}p=\operatorname{Curl}\bigl((\operatorname{Curl}p)^{T}\bigr). Inc p = Curl ( ( Curl p ) T ) .
Euclidean indices are raised and lowered using the standard Euclidean metric.
Step 1: reduction to the Euclidean double-curl equation.
We claim that, for every smooth symmetric matrix field p p p ,
[ L ( s − 1 p ) ] i j = − 1 2 s 3 ( Inc p ) i j (24) \left[L(s^{-1}p)\right]^{ij}=-\frac{1}{2}s^3(\operatorname{Inc}p)^{ij} \tag*{(24)} [ L ( s − 1 p ) ] ij = − 2 1 s 3 ( Inc p ) ij (24)
Here s − 1 p s^{-1}p s − 1 p denotes the covariant tensor with components s − 1 p i j s^{-1}p_{ij} s − 1 p ij , while the indices on the left-hand side are raised using g 0 g_0 g 0 .
We first compare the second-order terms. The linearization formulas for scalar and Ricci curvature give
( L h ) i j = − 1 2 det g 0 ϵ i m n ϵ j k ℓ ∂ m ∂ k h n ℓ + terms involving at most first derivatives of h . (Lh)^{ij}=-\frac{1}{2\det g_0}\epsilon_{imn}\epsilon_{jk\ell}\partial_m\partial_kh_{n\ell}+\text{terms involving at most first derivatives of }h. ( L h ) ij = − 2 det g 0 1 ϵ imn ϵ j k ℓ ∂ m ∂ k h n ℓ + terms involving at most first derivatives of h .
Since det g 0 = s − 4 \det g_0=s^{-4} det g 0 = s − 4 , the second-order part of L ( s − 1 p ) L(s^{-1}p) L ( s − 1 p ) is − 1 2 s 3 Inc p -\frac{1}{2}s^3\operatorname{Inc}p − 2 1 s 3 Inc p . Hence
R 0 ( p ) i j = [ L ( s − 1 p ) ] i j + 1 2 s 3 ( Inc p ) i j . R_0(p)^{ij}=\left[L(s^{-1}p)\right]^{ij}+\frac{1}{2}s^3(\operatorname{Inc}p)^{ij}. R 0 ( p ) ij = [ L ( s − 1 p ) ] ij + 2 1 s 3 ( Inc p ) ij .
is a differential operator of order at most one. We show that this remaining operator vanishes by using diffeomorphism invariance. The gnomonic Christoffel symbols give, for every one-form α \alpha α ,
∇ i α j + ∇ j α i = s − 1 [ ∂ i ( s α j ) + ∂ j ( s α i ) ] . \nabla_i\alpha_j+\nabla_j\alpha_i=s^{-1}\left[\partial_i(s\alpha_j)+\partial_j(s\alpha_i)\right]. ∇ i α j + ∇ j α i = s − 1 [ ∂ i ( s α j ) + ∂ j ( s α i ) ] .
Thus, if β \beta β is a Euclidean one-form and α = s − 1 β \alpha=s^{-1}\beta α = s − 1 β , then
Lie W g 0 = s − 1 sym ∂ β , W = α ♯ g 0 , \operatorname{Lie}_W g_0=s^{-1}\operatorname{sym}\partial\beta,\qquad W=\alpha^{\sharp g_0}, Lie W g 0 = s − 1 sym ∂ β , W = α ♯ g 0 ,
where
( sym ∂ β ) i j = ∂ i β j + ∂ j β i . (\operatorname{sym}\partial\beta)_{ij}=\partial_i\beta_j+\partial_j\beta_i. ( sym ∂ β ) ij = ∂ i β j + ∂ j β i .
By (23) ,
L ( s − 1 sym ∂ β ) = 0. L\left(s^{-1}\operatorname{sym}\partial\beta\right)=0. L ( s − 1 sym ∂ β ) = 0.
Commutativity of Euclidean partial derivatives also gives
Inc ( sym ∂ β ) = 0. \operatorname{Inc}(\operatorname{sym}\partial\beta)=0. Inc ( sym ∂ β ) = 0.
Therefore R 0 R_0 R 0 annihilates every Euclidean symmetric gradient.
The value and first derivatives of a symmetric gradient can be prescribed arbitrarily at a point. Indeed, for any prescribed symmetric first jet ( p i j , p i j , k ) (p_{ij},p_{ij,k}) ( p ij , p ij , k ) , choose
∂ j β i = 1 2 p i j , ∂ j ∂ k β i = 1 2 ( p i j , k + p i k , j − p j k , i ) \partial_j\beta_i=\frac{1}{2}p_{ij},\qquad\partial_j\partial_k\beta_i=\frac{1}{2}(p_{ij,k}+p_{ik,j}-p_{jk,i}) ∂ j β i = 2 1 p ij , ∂ j ∂ k β i = 2 1 ( p ij , k + p ik , j − p j k , i )
at that point. These choices give
( sym ∂ β ) i j = p i j , ∂ k ( sym ∂ β ) i j = p i j , k . (\operatorname{sym}\partial\beta)_{ij}=p_{ij},\qquad\partial_k(\operatorname{sym}\partial\beta)_{ij}=p_{ij,k}. ( sym ∂ β ) ij = p ij , ∂ k ( sym ∂ β ) ij = p ij , k .
Since R 0 R_0 R 0 has order at most one and annihilates all such jets, it is identically zero. This proves (24) , including its lower-order terms.
Now define the Euclidean symmetric tensor
U i j = s − 3 S i j , (25) U^{ij}=s^{-3}S^{ij}, \tag*{(25)} U ij = s − 3 S ij , (25)
where the indices on S S S are raised using g 0 g_0 g 0 . The divergence identity in Lemma 3.2 gives ∂ i U i j = 0 \partial_iU^{ij}=0 ∂ i U ij = 0 . Since U U U is symmetric, this is equivalent to Div U = 0 \operatorname{Div}U=0 Div U = 0 . Moreover, U U U is smooth on R 3 \mathbb{R}^3 R 3 and supported in B R ( 0 ) B_R(0) B R ( 0 ) .
By (24) , it remains to construct a smooth symmetric matrix field p p p , supported in B R ( 0 ) B_R(0) B R ( 0 ) , such that Inc p = U \operatorname{Inc}p=U Inc p = U . We will then take k = − 2 s − 1 p k=-2s^{-1}p k = − 2 s − 1 p .
Step 2: a vector potential obtained from outward rays.
Fix b ∈ B R ( 0 ) b\in B_R(0) b ∈ B R ( 0 ) . For a smooth, compactly supported, divergence-free Euclidean vector field v v v , define, for y ≠ b y\ne b y = b ,
K b v ( y ) = ∫ 1 ∞ u ( y − b ) × v ( b + u ( y − b ) ) d u . (26) K_bv(y)=\int_1^\infty u\,(y-b)\times v\bigl(b+u(y-b)\bigr)\,du. \tag*{(26)} K b v ( y ) = ∫ 1 ∞ u ( y − b ) × v ( b + u ( y − b ) ) d u . (26)
This integral samples v v v along the ray starting at y y y and continuing away from $b.
Put x = y − b x=y-b x = y − b . A direct calculation using div v = 0 \operatorname{div}v=0 div v = 0 gives
curl y [ u x × v ( b + u x ) ] = − d d u [ u 2 v ( b + u x ) ] . \operatorname{curl}_{y}\left[u\,x\times v(b+ux)\right]=-\frac{d}{du}\left[u^{2}v(b+ux)\right]. curl y [ u x × v ( b + ux ) ] = − d u d [ u 2 v ( b + ux ) ] .
Since v v v is compactly supported, integration in u u u yields
curl K b v = v on R 3 ∖ { b } . \operatorname{curl}K_{b}v=v\quad\text{on }\mathbb{R}^{3}\setminus\{b\}. curl K b v = v on R 3 ∖ { b } .
This construction also preserves support in the ball. Indeed, if v v v is supported in B R ( 0 ) B_{R}(0) B R ( 0 ) and y ∉ B R ( 0 ) y\notin B_{R}(0) y ∈ / B R ( 0 ) , then
b + u ( y − b ) ∉ B R ( 0 ) for every u ≥ 1 , b+u(y-b)\notin B_{R}(0)\quad\text{for every }u\ge1, b + u ( y − b ) ∈ / B R ( 0 ) for every u ≥ 1 ,
by convexity of the ball. Hence K b v ( y ) = 0 K_{b}v(y)=0 K b v ( y ) = 0 .
Step 3: a symmetric double-curl potential.
Apply K b K_{b} K b to each row of U U U , and write Q b = K b U Q_{b}=K_{b}U Q b = K b U . Then
Curl Q b = U away from b . \operatorname{Curl}Q_{b}=U\quad\text{away from }b. Curl Q b = U away from b .
For x = y − b x=y-b x = y − b , let J x J_{x} J x denote the skew-symmetric matrix J x w = x × w J_{x}w=x\times w J x w = x × w . In matrix notation,
Q b ( y ) = ∫ 1 ∞ u U ( b + u x ) J x T d u . Q_{b}(y)=\int_{1}^{\infty}u\,U(b+ux)J_{x}^{T}\,du. Q b ( y ) = ∫ 1 ∞ u U ( b + ux ) J x T d u .
Since U U U is symmetric and J x J_{x} J x is skew-symmetric, tr Q b = 0 \operatorname{tr}Q_{b}=0 tr Q b = 0 .
The symmetry of Curl Q b = U \operatorname{Curl}Q_{b}=U Curl Q b = U gives another useful identity:
∂ j ( Q b ) j i = ∂ i tr Q b = 0. \partial_{j}(Q_{b})_{ji}=\partial_{i}\operatorname{tr}Q_{b}=0. ∂ j ( Q b ) j i = ∂ i tr Q b = 0.
Thus every row of Q b T Q_{b}^{T} Q b T is divergence-free, and we may apply the vector potential construction a second time. Set
P b U = K b ( Q b T ) . P_{b}U=K_{b}(Q_{b}^{T}). P b U = K b ( Q b T ) .
The tensors involved are smooth away from b b b and vanish outside the closed ball, so the outward-ray integrals are well defined there. We obtain
Curl ( P b U ) = Q b T , Inc ( P b U ) = U away from b . \operatorname{Curl}(P_{b}U)=Q_{b}^{T},\qquad\operatorname{Inc}(P_{b}U)=U\quad\text{away from }b. Curl ( P b U ) = Q b T , Inc ( P b U ) = U away from b .
Combining the two radial integrals gives the explicit formula
P b U ( y ) = ∫ 1 ∞ u ( u − 1 ) J y − b U ( b + u ( y − b ) ) J y − b T d u . (27) P_{b}U(y)=\int_{1}^{\infty}u(u-1)J_{y-b}U\bigl(b+u(y-b)\bigr)J_{y-b}^{T}\,du. \tag*{(27)} P b U ( y ) = ∫ 1 ∞ u ( u − 1 ) J y − b U ( b + u ( y − b ) ) J y − b T d u . (27)
Indeed, the two radial parameters r , λ ≥ 1 r,\lambda\ge1 r , λ ≥ 1 combine as u = r λ u=r\lambda u = r λ ; the factor r 2 λ d λ d r r^{2}\lambda\,d\lambda\,dr r 2 λ d λ d r becomes u d u d r u\,du\,dr u d u d r , with 1 ≤ r ≤ u 1\le r\le u 1 ≤ r ≤ u . Integration in r r r gives u ( u − 1 ) u(u-1) u ( u − 1 ) . Formula (27) shows directly that P b U P_{b}U P b U is symmetric.
We must also check the equations at the center b b b . For y y y near b b b , compact support of U U U restricts the radial parameter to an interval of length O ( ∣ y − b ∣ − 1 ) O(|y-b|^{-1}) O ( ∣ y − b ∣ − 1 ) . The preceding formulas give
Q b ( y ) = O ( ∣ y − b ∣ − 1 ) , P b U ( y ) = O ( ∣ y − b ∣ − 1 ) . Q_b(y)=O(|y-b|^{-1}), \qquad P_bU(y)=O(|y-b|^{-1}). Q b ( y ) = O ( ∣ y − b ∣ − 1 ) , P b U ( y ) = O ( ∣ y − b ∣ − 1 ) .
These singularities are locally integrable. When either first-curl identity is tested by integration by parts outside B δ ( b ) B_\delta(b) B δ ( b ) , the boundary term is bounded by
C δ 2 δ − 1 = C δ , C\delta^2\delta^{-1}=C\delta, C δ 2 δ − 1 = C δ ,
and therefore tends to zero as δ ↓ 0 \delta\downarrow0 δ ↓ 0 . Consequently,
Curl Q b = U , Curl ( P b U ) = Q b T \operatorname{Curl}Q_b=U, \qquad\operatorname{Curl}(P_bU)=Q_b^{T} Curl Q b = U , Curl ( P b U ) = Q b T
hold distributionally on all of R 3 \mathbb{R}^3 R 3 . In particular,
Inc ( P b U ) = U \operatorname{Inc}(P_bU)=U Inc ( P b U ) = U
holds distributionally across the center as well.
Step 4: averaging the centers to obtain a smooth potential.
Choose a nonnegative function
θ ∈ C c ∞ ( B R ( 0 ) ) , ∫ R 3 θ ( b ) d b = 1 , \theta\in C_c^\infty(B_R(0)), \qquad\int_{\mathbb{R}^3}\theta(b)\,db=1, θ ∈ C c ∞ ( B R ( 0 )) , ∫ R 3 θ ( b ) d b = 1 ,
and define
P U ( y ) = ∫ R 3 θ ( b ) P b U ( y ) d b . (28) PU(y)=\int_{\mathbb{R}^3}\theta(b)P_bU(y)\,db. \tag*{(28)} P U ( y ) = ∫ R 3 θ ( b ) P b U ( y ) d b . (28)
The singularity ∣ y − b ∣ − 1 |y-b|^{-1} ∣ y − b ∣ − 1 is integrable in b b b , so this average is well defined. It is symmetric, and the distributional identities above give
Inc ( P U ) = U . \operatorname{Inc}(PU)=U. Inc ( P U ) = U .
Each P b U P_bU P b U is supported in B R ( 0 ) B_R(0) B R ( 0 ) , and therefore so is P U PU P U .
We now verify that the averaging removes the singularity completely. In (28) and (27) , make the changes of variables
w = ( u − 1 ) ( y − b ) , r = ( u − 1 ) − 1 . w=(u-1)(y-b), \qquad r=(u-1)^{-1}. w = ( u − 1 ) ( y − b ) , r = ( u − 1 ) − 1 .
They give
P U ( y ) = ∫ R 3 J w U ( y + w ) J w T [ ∫ 0 ∞ ( r + r 2 ) θ ( y − r w ) d r ] d w . (29) PU(y)=\int_{\mathbb{R}^3}J_wU(y+w)J_w^T\left[\int_0^\infty(r+r^2)\theta(y-rw)\,dr\right]dw. \tag*{(29)} P U ( y ) = ∫ R 3 J w U ( y + w ) J w T [ ∫ 0 ∞ ( r + r 2 ) θ ( y − r w ) d r ] d w . (29)
Fix a compact set of y y y -values. On the support of the integrand, both w w w and r ∣ w ∣ r|w| r ∣ w ∣ are bounded, because U U U and θ \theta θ have compact support. For w ≠ 0 w\ne0 w = 0 , this gives
∣ w ∣ 2 ∫ 0 C / ∣ w ∣ ( r + r 2 ) d r ≤ C ′ ( 1 + ∣ w ∣ − 1 ) . |w|^2\int_0^{C/|w|}(r+r^2)\,dr\le C'(1+|w|^{-1}). ∣ w ∣ 2 ∫ 0 C /∣ w ∣ ( r + r 2 ) d r ≤ C ′ ( 1 + ∣ w ∣ − 1 ) .
The right-hand side is locally integrable in R 3 \mathbb{R}^{3} R 3 .
With w w w fixed, every derivative in y y y in (29) falls only on U U U or θ \theta θ . In particular, differentiating θ ( y − r w ) \theta(y-rw) θ ( y − r w ) in y y y introduces no additional factor of r r r . Thus every y y y -derivative of the integrand satisfies the same type of integrable bound, with a constant depending on the derivative order. Differentiation under the integral proves
P U ∈ C ∞ ( R 3 ) . PU \in C^{\infty}(\mathbb{R}^{3}). P U ∈ C ∞ ( R 3 ) .
We have therefore constructed a smooth symmetric tensor satisfying
Inc ( P U ) = U , supp ( P U ) ⊂ B R ( 0 ) ‾ . \operatorname{Inc}(PU)=U,\qquad\operatorname{supp}(PU)\subset\overline{B_{R}(0)}. Inc ( P U ) = U , supp ( P U ) ⊂ B R ( 0 ) .
Since it is globally smooth and vanishes outside the closed ball, it vanishes to all orders on the boundary sphere.
Step 5: return to the spherical metric.
Define the covariant tensor k k k in gnomonic coordinates by
k i j = − 2 s − 1 ( P U ) i j . k_{ij}=-2s^{-1}(PU)_{ij}. k ij = − 2 s − 1 ( P U ) ij .
Using (24) and (25) , we obtain
( L k ) i j = s 3 ( Inc ( P U ) ) i j = s 3 U i j = S i j . (Lk)^{ij}=s^{3}\bigl(\operatorname{Inc}(PU)\bigr)^{ij}=s^{3}U^{ij}=S^{ij}. ( L k ) ij = s 3 ( Inc ( P U ) ) ij = s 3 U ij = S ij .
Thus L k = S Lk=S L k = S .
The support of k k k is contained in the closed cap Ω ‾ \overline{\Omega} Ω . Its smoothness across ∂ Ω \partial\Omega ∂ Ω follows from the global smoothness of P U PU P U . Extending it by zero toward the equator gives a smooth tensor on H H H . Finally, with θ \theta θ fixed, every operation in the construction is linear in S S S , so the resulting choice of k k k is linear. □
Gluing and the global curvature bound
The exact collar controls the boundary region, while the strict positivity of D D D absorbs the remaining error on the interior cap. The supported correction joins these two parts without changing the prescribed boundary geometry.
Figure 1 summarizes the regions used in the proof of Theorem 3.1 .
Figure 1. A schematic figure for the construction
Proof of Theorem 3.1 . Let D D D be the tensor constructed in Lemma 3.2 , and let G ε G_{\varepsilon} G ε be the collar metrics supplied by Lemma 3.4 . We extend these metrics to H H H and then correct their quadratic curvature term inside the cap, keeping the metric unchanged near the boundary.
Recall that t 0 < a < b < t 1 t_{0}<a<b<t_{1} t 0 < a < b < t 1 . Choose a smooth cutoff function χ \chi χ on H H H such that
0 ≤ χ ≤ 1 , χ = 1 on { t ≤ a } , χ = 0 on { t ≥ b } . 0\leq\chi\leq1,\qquad\chi=1\ \text{on }\{t\leq a\},\qquad\chi=0\ \text{on }\{t\geq b\}. 0 ≤ χ ≤ 1 , χ = 1 on { t ≤ a } , χ = 0 on { t ≥ b } .
Extend the vector field V = f ∂ t − tan t ∇ σ f V=f\partial_{t}-\tan t\nabla_{\sigma}f V = f ∂ t − tan t ∇ σ f smoothly to H H H , retaining its original values on a neighborhood of { t ≤ b } \{t\leq b\} { t ≤ b } and making it vanish near the north pole. We denote the extension also by V V V .
Define
g ˉ ε = g 0 + ε L V g 0 + χ ( G ε − g 0 − ε L V g 0 ) (30) \bar{g}_{\varepsilon}=g_{0}+\varepsilon\mathcal{L}_{V}g_{0}+\chi\left(G_{\varepsilon}-g_{0}-\varepsilon\mathcal{L}_{V}g_{0}\right) \tag*{(30)} g ˉ ε = g 0 + ε L V g 0 + χ ( G ε − g 0 − ε L V g 0 ) (30)
where the term multiplied by χ \chi χ is extended by zero outside the collar. This defines a smooth family of symmetric tensors on H H H . It agrees with G ε G_{\varepsilon} G ε on { t ≤ a } \{t\leq a\} { t ≤ a } and with g 0 g_{0} g 0 near the north pole.
The first-variation formula in Lemma 3.4 shows that
G ε − g 0 − ε L V g 0 = O ( ε 2 ) G_{\varepsilon}-g_{0}-\varepsilon\mathcal{L}_{V}g_{0}=O(\varepsilon^{2}) G ε − g 0 − ε L V g 0 = O ( ε 2 )
on the support of χ \chi χ . Consequently,
g ˉ ε = g 0 + ε L V g 0 + O ( ε 2 ) \bar{g}_{\varepsilon}=g_{0}+\varepsilon\mathcal{L}_{V}g_{0}+O(\varepsilon^{2}) g ˉ ε = g 0 + ε L V g 0 + O ( ε 2 )
on H H H . In particular, g ˉ ε \bar{g}_{\varepsilon} g ˉ ε is positive definite for sufficiently small ∣ ε ∣ |\varepsilon| ∣ ε ∣ .
Since T ( g 0 ) = 0 T(g_{0})=0 T ( g 0 ) = 0 and L ( L V g 0 ) = 0 L(\mathcal{L}_{V}g_{0})=0 L ( L V g 0 ) = 0 , the curvature expansion of g ˉ ε \bar{g}_{\varepsilon} g ˉ ε starts at second order:
T ( g ˉ ε ) = ε 2 E + O ( ∣ ε ∣ 3 ) (31) T(\bar{g}_{\varepsilon})=\varepsilon^{2}E+O(|\varepsilon|^{3}) \tag*{(31)} T ( g ˉ ε ) = ε 2 E + O ( ∣ ε ∣ 3 ) (31)
where E E E is a smooth symmetric tensor on H H H , independent of ε \varepsilon ε . The exact agreement with G ε G_{\varepsilon} G ε on { t ≤ a } \{t\leq a\} { t ≤ a } and the expansion (13) imply
E = D on { t ≤ a } . E=D\quad\text{on }\{t\leq a\}. E = D on { t ≤ a } .
We next verify the compatibility condition for the interior correction. The contracted Bianchi identity gives div g ˉ ε T ( g ˉ ε ) = 0 \operatorname{div}_{\bar{g}_{\varepsilon}}T(\bar{g}_{\varepsilon})=0 div g ˉ ε T ( g ˉ ε ) = 0 . Using (31) and g ˉ ε = g 0 + O ( ε ) \bar{g}_{\varepsilon}=g_0+O(\varepsilon) g ˉ ε = g 0 + O ( ε ) , we obtain
0 = ε 2 div g 0 E + O ( ∣ ε ∣ 3 ) . 0=\varepsilon^2\operatorname{div}_{g_0}E+O(|\varepsilon|^3). 0 = ε 2 div g 0 E + O ( ∣ ε ∣ 3 ) .
Dividing by ε 2 \varepsilon^2 ε 2 and letting ε → 0 \varepsilon\to0 ε → 0 yields div g 0 E = 0 \operatorname{div}_{g_0}E=0 div g 0 E = 0 . Therefore S : = D − E S:=D-E S := D − E is a smooth symmetric tensor satisfying
div g 0 S = 0 , supp S ⊂ { t ≥ a } . \operatorname{div}_{g_0}S=0,\qquad\operatorname{supp}S\subset\{t\geq a\}. div g 0 S = 0 , supp S ⊂ { t ≥ a } .
Apply Lemma 3.5 with Ω = { t > a } \Omega=\{t>a\} Ω = { t > a } . It gives a smooth symmetric tensor k k k on H H H such that
L k = D − E , supp k ⊂ { t ≥ a } . Lk=D-E,\qquad\operatorname{supp}k\subset\{t\geq a\}. L k = D − E , supp k ⊂ { t ≥ a } .
The tensor k k k is fixed, independent of ε \varepsilon ε . Define the final metrics by g ε = g ˉ ε + ε 2 k g_{\varepsilon}=\bar{g}_{\varepsilon}+\varepsilon^2k g ε = g ˉ ε + ε 2 k . They form a smooth family and converge to g 0 g_0 g 0 as ε → 0 \varepsilon\to0 ε → 0 , so they are positive definite for sufficiently small ∣ ε ∣ |\varepsilon| ∣ ε ∣ . Moreover,
g ε = 0 = g 0 , g ε = G ε on { t ≤ a } . g_{\varepsilon=0}=g_0,\qquad g_{\varepsilon}=G_{\varepsilon}\quad\text{on }\{t\leq a\}. g ε = 0 = g 0 , g ε = G ε on { t ≤ a } .
Taylor expansion at the round metric gives
T ( g ε ) = T ( g ˉ ε ) + ε 2 L k + O ( ∣ ε ∣ 3 ) . T(g_{\varepsilon})=T(\bar{g}_{\varepsilon})+\varepsilon^2Lk+O(|\varepsilon|^3). T ( g ε ) = T ( g ˉ ε ) + ε 2 L k + O ( ∣ ε ∣ 3 ) .
Indeed, the linearization at g ˉ ε \bar{g}_{\varepsilon} g ˉ ε differs from L L L by O ( ∣ ε ∣ ) O(|\varepsilon|) O ( ∣ ε ∣ ) , since g ˉ ε = g 0 + O ( ε ) \bar{g}_{\varepsilon}=g_0+O(\varepsilon) g ˉ ε = g 0 + O ( ε ) . Hence
T ( g ε ) = ε 2 ( E + L k ) + O ( ∣ ε ∣ 3 ) = ε 2 D + O ( ∣ ε ∣ 3 ) . (32) \begin{aligned}
T(g_{\varepsilon})&=\varepsilon^2(E+Lk)+O(|\varepsilon|^3)\\
&=\varepsilon^2D+O(|\varepsilon|^3). \tag*{(32)}
\end{aligned} T ( g ε ) = ε 2 ( E + L k ) + O ( ∣ ε ∣ 3 ) = ε 2 D + O ( ∣ ε ∣ 3 ) . (32)
We verify the curvature bound separately on the retained collar and the interior cap. On { t ≤ a } \{t\leq a\} { t ≤ a } , the metric is exactly G ε G_{\varepsilon} G ε , and Lemma 3.4 gives sec g ε ≥ 1 \operatorname{sec}_{g_{\varepsilon}}\geq1 sec g ε ≥ 1 . On { t ≥ a } \{t\geq a\} { t ≥ a } , the inequality a > t 0 a>t_0 a > t 0 and Lemma 3.2 give D ≥ c 0 g 0 D\geq c_0g_0 D ≥ c 0 g 0 . The remainder in (32) is uniform on this compact cap. Thus, as quadratic forms,
T ( g ε ) ≥ ε 2 ( c 0 − C ∣ ε ∣ ) g 0 ≥ 0 T(g_{\varepsilon})\geq\varepsilon^2(c_0-C|\varepsilon|)g_0\geq0 T ( g ε ) ≥ ε 2 ( c 0 − C ∣ ε ∣ ) g 0 ≥ 0
for sufficiently small ∣ ε ∣ |\varepsilon| ∣ ε ∣ . By the three-dimensional characterization of the sectional curvature bound through T T T , this proves
sec g ε ≥ 1 on all of H . \operatorname{sec}_{g_{\varepsilon}}\geq1\quad\text{on all of }H. sec g ε ≥ 1 on all of H .
Finally, the metric is unchanged from G ε G_{\varepsilon} G ε near the boundary. Therefore g ε ∣ ∂ H = σ g_{\varepsilon}|_{\partial H}=\sigma g ε ∣ ∂ H = σ . The outward unit normal is − ∂ t -\partial_t − ∂ t , so
h g ε = − 1 2 γ ˙ ε ( 0 ) = ε A . h_{g_{\varepsilon}}=-\frac{1}{2}\dot{\gamma}_{\varepsilon}(0)=\varepsilon A. h g ε = − 2 1 γ ˙ ε ( 0 ) = ε A .
Since A A A is Codazzi with respect to σ \sigma σ , the boundary second fundamental form is Codazzi for every ε \varepsilon ε . Moreover, ρ ≢ 0 \rho\not\equiv0 ρ ≡ 0 implies A ≢ 0 A\not\equiv0 A ≡ 0 . Thus the boundary is not totally geodesic when ε ≠ 0 \varepsilon\neq0 ε = 0 , and ( H , g ε ) (H,g_{\varepsilon}) ( H , g ε ) cannot be isometric to the standard hemisphere. □
An inradius estimate in dimension three
The equality case of Theorem C again reduces to a collapsed great circle. Distance comparison produces a nonnegative boundary excess whose zero set is the contact set. We isolate the resulting spherical argument in Proposition 4.5. A contact barrier and the Codazzi equation turn a suitable regular level curve of a scalar projection of the support map into a complete great circle on which the map is constant. The Gálvez–Mira theorem supplies two distinct support values so that this level curve can be chosen; Lemma 2.4 then forces the tensor to vanish. Recall that
r i n ( M ) : = max p ∈ M d g ( p , ∂ M ) . r_{\mathrm{in}}(M) := \max_{p\in M} d_g(p,\partial M). r in ( M ) := p ∈ M max d g ( p , ∂ M ) .
Identify ∂ M \partial M ∂ M with the unit round sphere S 2 S^2 S 2 , and write g ˉ = g ∣ T ∂ M \bar{g}=g|_{T\partial M} g ˉ = g ∣ T ∂ M . Denote its Levi-Civita connection by ∇ \nabla ∇ and the ambient connection by ∇ M \nabla^M ∇ M . With the outward unit normal ν \nu ν , set
h ( X , Y ) = g ( ∇ X M ν , Y ) , h ( X , Y ) = g ˉ ( A X , Y ) . h(X,Y)=g(\nabla^M_X\nu,Y),\qquad h(X,Y)=\bar{g}(AX,Y). h ( X , Y ) = g ( ∇ X M ν , Y ) , h ( X , Y ) = g ˉ ( A X , Y ) .
By hypothesis, A A A is a self-adjoint Codazzi tensor. The Gauss equation gives
det A = 1 − sec M ( T ∂ M ) ≤ 0. (33) \det A=1-\operatorname{sec}_M(T\partial M)\le0. \tag*{(33)} det A = 1 − sec M ( T ∂ M ) ≤ 0. (33)
Distance comparison at the boundary
The following comparison holds in every dimension. Besides bounding the second fundamental form at a nearest boundary point, it ensures that the distance function is smooth there. This smoothness will allow us to use the distance excess as a local barrier.
Lemma 4.1. Let ( M n , g ) (M^n,g) ( M n , g ) , n ≥ 2 n\ge2 n ≥ 2 , be a compact connected Riemannian manifold with smooth nonempty boundary and sec g ≥ 1 \operatorname{sec}_g\ge1 sec g ≥ 1 . For p ∈ int M p\in\operatorname{int} M p ∈ int M , set
R = d g ( p , ∂ M ) , C p = { q ∈ ∂ M : d g ( p , q ) = R } . R=d_g(p,\partial M),\qquad C_p=\{q\in\partial M:d_g(p,q)=R\}. R = d g ( p , ∂ M ) , C p = { q ∈ ∂ M : d g ( p , q ) = R } . Then R < π R<\pi R < π . Moreover, the distance function r = d g ( p , ⋅ ) r=d_g(p,\mathord\cdot) r = d g ( p , ⋅ ) is smooth in a neighborhood of C p C_p C p and
∇ M r ( q ) = ν ( q ) , h q ≤ cot R g ˉ q for every q ∈ C p . \nabla^M r(q)=\nu(q),\qquad h_q\le\cot R\,\bar{g}_q\qquad\text{for every }q\in C_p. ∇ M r ( q ) = ν ( q ) , h q ≤ cot R g ˉ q for every q ∈ C p . Proof. Fix q ∈ C p q\in C_p q ∈ C p , and let α : [ 0 , R ] ⟶ M \alpha:[0,R]\longrightarrow M α : [ 0 , R ] ⟶ M be a minimizing unit-speed geodesic from p p p to q q q . Its interior lies in int M \operatorname{int} M int M , and the first variation formula gives α ′ ( R ) = ν ( q ) \alpha'(R)=\nu(q) α ′ ( R ) = ν ( q ) .
For a vector field V V V along α \alpha α perpendicular to α ′ ( s ) \alpha'(s) α ′ ( s ) satisfying
V ( 0 ) = 0 , V ( R ) ∈ T q ∂ M , V(0)=0,\qquad V(R)\in T_q\partial M, V ( 0 ) = 0 , V ( R ) ∈ T q ∂ M , the second variation formula gives
Q ( V , V ) : = I α ( V , V ) − h q ( V ( R ) , V ( R ) ) ≥ 0 , Q(V,V):=I_\alpha(V,V)-h_q(V(R),V(R))\ge0, Q ( V , V ) := I α ( V , V ) − h q ( V ( R ) , V ( R )) ≥ 0 ,
where
I α ( V , W ) = ∫ 0 R ( ⟨ D s V , D s W ⟩ − ⟨ R M ( V , α ′ ) α ′ , W ⟩ ) d s . I_{\alpha}(V,W)=\int_{0}^{R}\left(\langle D_sV,D_sW\rangle-\langle R^M(V,\alpha')\alpha',W\rangle\right)\,ds. I α ( V , W ) = ∫ 0 R ( ⟨ D s V , D s W ⟩ − ⟨ R M ( V , α ′ ) α ′ , W ⟩ ) d s .
We first show that q q q is not conjugate to p p p along α \alpha α . Otherwise, there would exist a nonzero perpendicular Jacobi field J J J satisfying J ( 0 ) = J ( R ) = 0 J(0)=J(R)=0 J ( 0 ) = J ( R ) = 0 and Q ( J , J ) = 0 Q(J,J)=0 Q ( J , J ) = 0 . Since Q Q Q is positive semidefinite, Q ( J , V ) = 0 Q(J,V)=0 Q ( J , V ) = 0 for every admissible V V V . Integration by parts gives
0 = Q ( J , V ) = ⟨ D s J ( R ) , V ( R ) ⟩ . 0=Q(J,V)=\langle D_sJ(R),V(R)\rangle. 0 = Q ( J , V ) = ⟨ D s J ( R ) , V ( R )⟩ .
Since D s J ( R ) D_sJ(R) D s J ( R ) is perpendicular to ν ( q ) \nu(q) ν ( q ) , it belongs to T q ∂ M T_q\partial M T q ∂ M . The arbitrariness of V ( R ) V(R) V ( R ) therefore implies D s J ( R ) = 0 D_sJ(R)=0 D s J ( R ) = 0 . Together with J ( R ) = 0 J(R)=0 J ( R ) = 0 , this forces J ≡ 0 J\equiv0 J ≡ 0 , a contradiction.
There are no conjugate points in ( 0 , R ) (0,R) ( 0 , R ) , since α \alpha α is minimizing. Conjugate point comparison under sec g ≥ 1 \sec_g\ge1 sec g ≥ 1 therefore yields R < π R<\pi R < π .
Every minimizing geodesic from p p p to q q q has terminal velocity ν ( q ) \nu(q) ν ( q ) . Uniqueness for the geodesic equation thus implies that α \alpha α is the unique minimizing geodesic from p p p to q q q .
To justify smoothness of the distance at boundary points, take a smooth complete extension of ( M , g ) (M,g) ( M , g ) . Any curve from p p p to ∂ M \partial M ∂ M must first reach the boundary after a segment contained in M M M of length at least R R R . Consequently, the distance from p p p to q q q in the extension is also R R R , and every minimizing geodesic to q q q remains in M M M until its endpoint. Thus q q q is joined to p p p by a unique minimizing geodesic and is not conjugate to p p p in the extension. The ambient distance function is therefore smooth near q q q .
By transversality at q q q , the nearby minimizing geodesics to points of M M M also remain in M M M . Hence the ambient distance agrees locally with d g ( p , ⋅ ) d_g(p,\cdot) d g ( p , ⋅ ) on M M M . This proves smoothness near C p C_p C p , up to the boundary, and gives ∇ M r ( q ) = ν ( q ) \nabla^M r(q)=\nu(q) ∇ M r ( q ) = ν ( q ) .
Finally, let X ∈ T q ∂ M X\in T_q\partial M X ∈ T q ∂ M , and let P ( s ) P(s) P ( s ) be its parallel transport along α \alpha α . Taking V ( s ) = sin s sin R P ( s ) V(s)=\frac{\sin s}{\sin R}P(s) V ( s ) = s i n R s i n s P ( s ) , the second variation inequality and sec g ≥ 1 \sec_g\ge1 sec g ≥ 1 then imply
h q ( X , X ) ≤ I α ( V , V ) ≤ ∣ X ∣ 2 sin 2 R ∫ 0 R ( cos 2 s − sin 2 s ) d s = cot R ∣ X ∣ 2 . \begin{aligned}
h_q(X,X)\le I_\alpha(V,V) \\
&\le\frac{|X|^2}{\sin^2 R}\int_0^R(\cos^2s-\sin^2s)\,ds \\
&=\cot R|X|^2.
\end{aligned} h q ( X , X ) ≤ I α ( V , V ) ≤ sin 2 R ∣ X ∣ 2 ∫ 0 R ( cos 2 s − sin 2 s ) d s = cot R ∣ X ∣ 2 .
The claimed tensor inequality follows. □
Contact barriers and kernel propagation
We need two consequences of the distance inequality. The first prevents a separated part of the contact set from avoiding a closed geodesic semicircle. A subset of C C C is called relatively clopen if it is both open and closed in C C C .
Lemma 4.2. Let U ⊂ S 2 U \subset S^{2} U ⊂ S 2 be open, and let u ∈ C ∞ ( U ) u \in C^{\infty}(U) u ∈ C ∞ ( U ) be nonnegative. Suppose that
C = { x ∈ U : u ( x ) = 0 } C = \{x \in U : u(x) = 0\} C = { x ∈ U : u ( x ) = 0 } is nonempty and compact, and that
λ min ( ∇ 2 u ) ≤ 0 on U . (34) \lambda_{\min}(\nabla^{2}u) \le0 \qquad\text{on } U. \tag*{(34)} λ m i n ( ∇ 2 u ) ≤ 0 on U . (34) Then every nonempty relatively clopen subset of C C C intersects every closed geodesic semicircle of S 2 S^{2} S 2 .
Proof. Let K ⊂ C K \subset C K ⊂ C be nonempty and relatively clopen, and suppose that K K K is disjoint from a closed geodesic semicircle Γ \Gamma Γ .
By the convex supporting theorem of Jost–Xin–Yang [2 Theorem 2.1] , the complement S 2 ∖ Γ S^{2} \setminus\Gamma S 2 ∖ Γ admits a smooth strictly convex function on a neighborhood of any compact subset. Thus there exist an open neighborhood W W W of K K K and a smooth function ψ \psi ψ on W W W such that
∇ 2 ψ > 0. \nabla^{2}\psi> 0. ∇ 2 ψ > 0. Since K K K is relatively clopen in the compact set C C C , we can choose an open set V V V satisfying
K ⊂ V , V ‾ ⊂ U ∩ W , V ‾ ∩ C = K . K \subset V,\qquad\overline{V} \subset U \cap W,\qquad\overline{V} \cap C = K. K ⊂ V , V ⊂ U ∩ W , V ∩ C = K . In particular, u > 0 u > 0 u > 0 on ∂ V \partial V ∂ V . Set m = max K ψ m = \max_{K}\psi m = max K ψ . For sufficiently small ε > 0 \varepsilon> 0 ε > 0 , the function
w = u − ε ( ψ − m ) w = u - \varepsilon(\psi- m) w = u − ε ( ψ − m ) is positive on ∂ V \partial V ∂ V , while w = 0 w = 0 w = 0 at a point of K K K where ψ = m \psi= m ψ = m . Therefore w w w attains a nonpositive minimum at an interior point x ∈ V x \in V x ∈ V . At this point,
∇ 2 u ( x ) ≥ ε ∇ 2 ψ ( x ) > 0 , \nabla^{2}u(x) \ge\varepsilon\nabla^{2}\psi(x) > 0, ∇ 2 u ( x ) ≥ ε ∇ 2 ψ ( x ) > 0 , contradicting (34) . □ \square □
The second consequence is local propagation along the kernel at a rank-one contact point. A kernel vector alone need not extend to a curve in a support-map fiber when rank-two points occur nearby. The barrier inequality below supplies this missing propagation.
Lemma 4.3. Let h h h be a smooth symmetric ( 0 , 2 ) (0,2) ( 0 , 2 ) Codazzi tensor on an open subset U ⊂ S 2 U \subset S^{2} U ⊂ S 2 , and let A A A be its associated self-adjoint endomorphism. Assume that det A ≤ 0 \det A \le0 det A ≤ 0 on U U U .
Suppose that u , b ∈ C ∞ ( U ) u,b \in C^{\infty}(U) u , b ∈ C ∞ ( U ) satisfy
u ≥ 0 , b > 0 , ∇ 2 u + b h ≤ 0. u \ge0,\qquad b > 0,\qquad\nabla^{2}u + bh \le0. u ≥ 0 , b > 0 , ∇ 2 u + bh ≤ 0.
If u ( q ) = 0 u(q)=0 u ( q ) = 0 and A ( q ) ≠ 0 A(q)\ne0 A ( q ) = 0 , then there exists a unit-speed geodesic segment σ \sigma σ through q q q such that
u ( σ ( t ) ) = 0 , A ( σ ( t ) ) σ ˙ ( t ) = 0 u(\sigma(t))=0,\qquad A(\sigma(t))\dot{\sigma}(t)=0 u ( σ ( t )) = 0 , A ( σ ( t )) σ ˙ ( t ) = 0
for all sufficiently small ∣ t ∣ |t| ∣ t ∣ .
Proof. Since q q q is a minimum point of u u u ,
∇ 2 u ( q ) ≥ 0. \nabla^2u(q)\ge0. ∇ 2 u ( q ) ≥ 0.
Thus h q ≤ 0 h_q\le0 h q ≤ 0 . Together with det A ( q ) ≤ 0 \det A(q)\le0 det A ( q ) ≤ 0 and A ( q ) ≠ 0 A(q)\ne0 A ( q ) = 0 , this shows that A ( q ) A(q) A ( q ) has eigenvalues 0 0 0 and a strictly negative number.
After shrinking the neighborhood of q q q , we can choose smooth orthonormal eigenvector fields T , E T,E T , E such that
A T = μ T , A E = − η E , μ ≥ 0 , η > 0. AT=\mu T,\qquad AE=-\eta E,\qquad\mu\ge0,\qquad\eta>0. A T = μ T , A E = − η E , μ ≥ 0 , η > 0.
The nonnegativity of μ \mu μ follows from det A ≤ 0 \det A\le0 det A ≤ 0 . The Codazzi equation gives
∇ T T = E μ μ + η E . (35) \nabla_TT=\frac{E\mu}{\mu+\eta}E. \tag*{(35)} ∇ T T = μ + η E μ E . (35)
Indeed, comparing the T T T -components of ( ∇ E A ) T (\nabla_EA)T ( ∇ E A ) T and ( ∇ T A ) E (\nabla_TA)E ( ∇ T A ) E yields this identity.
Let σ \sigma σ be an integral curve of T T T with σ ( 0 ) = q \sigma(0)=q σ ( 0 ) = q , and put y ( t ) = u ( σ ( t ) ) y(t)=u(\sigma(t)) y ( t ) = u ( σ ( t )) . Using (35) , we obtain
y ′ ′ = ∇ 2 u ( T , T ) + d u ( ∇ T T ) ≤ − b μ + E μ μ + η E u . \begin{aligned}
y''=\nabla^2u(T,T)+du(\nabla_TT) \\
&\le-b\mu+\frac{E\mu}{\mu+\eta}Eu.
\end{aligned} y ′′ = ∇ 2 u ( T , T ) + d u ( ∇ T T ) ≤ − b μ + μ + η E μ E u .
On a sufficiently small relatively compact neighborhood, there are constants b 0 , η 0 > 0 b_0,\eta_0>0 b 0 , η 0 > 0 such that
b ≥ b 0 , η ≥ η 0 . b\ge b_0,\qquad\eta\ge\eta_0. b ≥ b 0 , η ≥ η 0 .
The elementary local gradient estimate for nonnegative smooth functions, obtained from Taylor’s formula, gives
∣ ∇ u ∣ 2 ≤ c u , ∣ ∇ μ ∣ 2 ≤ c μ . |\nabla u|^2\le cu,\qquad|\nabla\mu|^2\le c\mu. ∣∇ u ∣ 2 ≤ c u , ∣∇ μ ∣ 2 ≤ c μ .
Consequently, Young’s inequality implies
y ′ ′ ≤ − b 0 μ + c 1 y μ ≤ c 2 y . y''\le-b_0\mu+c_1\sqrt{y\mu}\le c_2y. y ′′ ≤ − b 0 μ + c 1 y μ ≤ c 2 y .
Moreover,
y ≥ 0 , y ( 0 ) = y ′ ( 0 ) = 0. y\ge0,\qquad y(0)=y'(0)=0. y ≥ 0 , y ( 0 ) = y ′ ( 0 ) = 0.
Integrating twice gives, for t ≥ 0 t\ge0 t ≥ 0 sufficiently small,
0 ≤ y ( t ) ≤ c 2 ∫ 0 t ( t − s ) y ( s ) d s . 0\le y(t)\le c_2\int_0^t(t-s)y(s)\,ds. 0 ≤ y ( t ) ≤ c 2 ∫ 0 t ( t − s ) y ( s ) d s .
Taking the maximum on a sufficiently short interval shows that y ≡ 0 y \equiv0 y ≡ 0 . The same argument applied to y ( − t ) y(-t) y ( − t ) gives the conclusion for negative t t t .
Along this segment, d u = 0 du = 0 d u = 0 , and hence
0 = y ′ ′ = ∇ 2 u ( T , T ) ≤ − b μ . 0 = y'' = \nabla^2 u(T,T) \le-b\mu. 0 = y ′′ = ∇ 2 u ( T , T ) ≤ − b μ .
Thus μ = 0 \mu= 0 μ = 0 along σ \sigma σ . Since μ ≥ 0 \mu\ge0 μ ≥ 0 in the neighborhood, its gradient also vanishes along σ \sigma σ . Equation (35) now gives ∇ T T = 0 \nabla_T T = 0 ∇ T T = 0 . Therefore σ \sigma σ is a unit-speed geodesic and A σ ˙ = 0 A\dot{\sigma} = 0 A σ ˙ = 0 along it. □
Rigidity from support-map fibers
To choose a regular level meeting the contact set, we use the following consequence of Gálvez–Mira’s global result on S 2 \mathbb{S}^2 S 2 .
Theorem 4.4 (Consequence of Gálvez–Mira [GM26, Theorem 1.4 and Section 3]). Let f ∈ C 2 ( S 2 ) f \in C^2(\mathbb{S}^2) f ∈ C 2 ( S 2 ) satisfy det A f ≤ 0 \det A_f \le0 det A f ≤ 0 and A f ≢ 0 A_f \not\equiv0 A f ≡ 0 . Then there are two closed geodesic semicircles Γ 1 , Γ 2 ⊂ S 2 \Gamma_1,\Gamma_2 \subset\mathbb{S}^2 Γ 1 , Γ 2 ⊂ S 2 and distinct vectors z 1 , z 2 ∈ R 3 z_1,z_2 \in\mathbb{R}^3 z 1 , z 2 ∈ R 3 such that the support map satisfies
F f ( x ) = ∇ f ( x ) + f ( x ) x = z i for every x ∈ Γ i , i = 1 , 2. F_f(x) = \nabla f(x) + f(x)x = z_i \qquad\text{for every } x \in\Gamma_i,\quad i=1,2. F f ( x ) = ∇ f ( x ) + f ( x ) x = z i for every x ∈ Γ i , i = 1 , 2. The distinct values follow from the gradient-map statement in [GM26, Section 3], applied to the homogeneous extension f ~ ( r x ) = r f ( x ) \widetilde{f}(rx) = rf(x) f ( r x ) = r f ( x ) , for which ∇ f ~ ∣ S 2 = F f \left.\nabla\widetilde{f}\right|_{\mathbb{S}^2} = F_f ∇ f S 2 = F f .
Proposition 4.5. Let A A A be a smooth self-adjoint Codazzi tensor on S 2 \mathbb{S}^2 S 2 with det A ≤ 0 \det A \le0 det A ≤ 0 , and put h ( X , Y ) = ⟨ A X , Y ⟩ h(X,Y) = \langle AX,Y\rangle h ( X , Y ) = ⟨ A X , Y ⟩ . Suppose that an open set U ⊂ S 2 U \subset\mathbb{S}^2 U ⊂ S 2 and functions u , b ∈ C ∞ ( U ) u,b \in C^\infty(U) u , b ∈ C ∞ ( U ) satisfy
u ≥ 0 , b > 0 , ∇ 2 u + b h ≤ 0 on U , u \ge0,\qquad b > 0,\qquad\nabla^2u + bh \le0 \qquad\text{on } U, u ≥ 0 , b > 0 , ∇ 2 u + bh ≤ 0 on U , and that C = { x ∈ U : u ( x ) = 0 } C = \{x \in U : u(x) = 0\} C = { x ∈ U : u ( x ) = 0 } is nonempty and compact. Then A ≡ 0 A \equiv0 A ≡ 0 on S 2 \mathbb{S}^2 S 2 .
Proof. At each point of U U U , the condition det A ≤ 0 \det A \le0 det A ≤ 0 gives a nonnegative eigenvalue of A A A . Evaluating the tensor inequality on a corresponding unit eigenvector shows that
λ min ( ∇ 2 u ) ≤ 0. \lambda_{\min}(\nabla^2u) \le0. λ m i n ( ∇ 2 u ) ≤ 0.
Lemma 4.2 therefore applies to C C C . In particular, every nonempty relatively clopen subset of C C C meets every closed geodesic semicircle.
At every q ∈ C q \in C q ∈ C , the minimum property of u u u also gives
h q ≤ 0 , det A ( q ) = 0 , rank A ( q ) ≤ 1. (36) h_q \le0,\qquad\det A(q) = 0,\qquad\operatorname{rank} A(q) \le1. \tag*{(36)} h q ≤ 0 , det A ( q ) = 0 , rank A ( q ) ≤ 1. (36)
Suppose that A ≢ 0 A \not\equiv0 A ≡ 0 , and let F = ∇ f + f x F = \nabla f + fx F = ∇ f + f x be the support map from Lemma 2.2 , so that
d F = A . (37) dF = A. \tag*{(37)} d F = A . (37)
By Theorem 4.4 , we may choose two closed geodesic semicircles Γ 1 , Γ 2 \Gamma_1,\Gamma_2 Γ 1 , Γ 2 with F ∣ Γ i = z i F|_{\Gamma_i} = z_i F ∣ Γ i = z i , z 1 ≠ z 2 z_1 \ne z_2 z 1 = z 2 .
F ∣ Γ i = z i , z 1 ≠ z 2 . F|_{\Gamma_i} = z_i,\qquad z_1 \ne z_2. F ∣ Γ i = z i , z 1 = z 2 .
Choose a linear functional L : R 3 → R L:\mathbb{R}^{3}\to\mathbb{R} L : R 3 → R such that a 1 : = L ( z 1 ) < L ( z 2 ) = : a 2 a_{1}:=L(z_{1})<L(z_{2})=:a_{2} a 1 := L ( z 1 ) < L ( z 2 ) =: a 2 , and put
G = L ∘ F : S 2 ⟶ R . G=L\circ F:S^{2}\longrightarrow\mathbb{R}. G = L ∘ F : S 2 ⟶ R .
By Lemma 4.2 , each Γ i \Gamma_i Γ i meets C C C , so a i ∈ G ( C ) a_i\in G(C) a i ∈ G ( C ) . In fact,
[ a 1 , a 2 ] ⊂ G ( C ) . [a_{1},a_{2}]\subset G(C). [ a 1 , a 2 ] ⊂ G ( C ) .
Otherwise, for some s ∈ ( a 1 , a 2 ) ∖ G ( C ) s\in(a_{1},a_{2})\setminus G(C) s ∈ ( a 1 , a 2 ) ∖ G ( C ) , the set
K = C ∩ { G < s } K=C\cap\{G<s\} K = C ∩ { G < s }
would be a nonempty relatively clopen subset of C C C disjoint from Γ 2 \Gamma_2 Γ 2 , contradicting the same lemma.
The contact inequality will now provide kernel propagation, even though A A A need not have constant rank near C C C . By Sard’s theorem, choose a regular value s ∈ ( a 1 , a 2 ) s\in(a_{1},a_{2}) s ∈ ( a 1 , a 2 ) of G G G . Let Σ \Sigma Σ be a connected component of G − 1 ( s ) G^{-1}(s) G − 1 ( s ) which meets C C C . Then Σ \Sigma Σ is a compact connected embedded curve without boundary.
For every x ∈ Σ ∩ C x\in\Sigma\cap C x ∈ Σ ∩ C , regularity gives
0 ≠ d G x = L ∘ A ( x ) . 0\ne dG_x=L\circ A(x). 0 = d G x = L ∘ A ( x ) .
Together with (36) , this implies
rank A ( x ) = 1 , T x Σ = ker d G x = ker A ( x ) . \operatorname{rank}A(x)=1,\qquad T_x\Sigma=\ker dG_x=\ker A(x). rank A ( x ) = 1 , T x Σ = ker d G x = ker A ( x ) .
Lemma 4.3 provides a geodesic segment through x x x contained in C C C and tangent to ker A \ker A ker A along the segment. By (37) , F F F , and hence G G G , is constant along this segment. It therefore lies in Σ \Sigma Σ and forms a neighborhood of x x x in Σ \Sigma Σ .
Thus Σ ∩ C \Sigma\cap C Σ ∩ C is relatively open in Σ \Sigma Σ . It is also closed and nonempty, so connectedness gives Σ ⊂ C \Sigma\subset C Σ ⊂ C . The same propagation argument at every point of Σ \Sigma Σ shows that Σ \Sigma Σ is totally geodesic. Since it is compact, it is complete, and therefore it is a full great circle of S 2 S^{2} S 2 .
Moreover, T Σ = ker A T\Sigma=\ker A T Σ = ker A and d F = A dF=A d F = A , so F F F is constant on Σ \Sigma Σ , say F = z F=z F = z . Thus f ( x ) − ⟨ z , x ⟩ = 0 f(x)-\langle z,x\rangle=0 f ( x ) − ⟨ z , x ⟩ = 0 on Σ \Sigma Σ . Lemma 2.4 implies that this function is linear on S 2 S^{2} S 2 . Its spherical Hessian is A A A , so A ≡ 0 A\equiv0 A ≡ 0 , a contradiction. □
Proof of Theorem C. Choose p ∈ int M p\in\operatorname{int}M p ∈ int M such that R = d g ( p , ∂ M ) = r i n ( M ) R=d_g(p,\partial M)=r_{\mathrm{in}}(M) R = d g ( p , ∂ M ) = r in ( M ) . At any nearest boundary point q q q , Lemma 4.1 gives
h q ≤ cot R g ˉ q , R < π . h_q\le\cot R\,\bar g_q,\qquad R<\pi. h q ≤ cot R g ˉ q , R < π .
If R > π / 2 R>\pi/2 R > π /2 , then h q h_q h q is negative definite, so det A ( q ) > 0 \det A(q)>0 det A ( q ) > 0 . This contradicts (33) . Therefore r i n ( M ) ≤ π 2 r_{\mathrm{in}}(M)\le\frac{\pi}{2} r in ( M ) ≤ 2 π .
Assume now that R = π / 2 R=\pi/2 R = π /2 , and define the contact set
C = { q ∈ ∂ M : d g ( p , q ) = π 2 } . C=\left\{q\in\partial M:d_g(p,q)=\frac{\pi}{2}\right\}. C = { q ∈ ∂ M : d g ( p , q ) = 2 π } .
This is a nonempty compact subset of S 2 S^2 S 2 .
By Lemma 4.1 , the function r = d g ( p , ⋅ ) r=d_g(p,\cdot) r = d g ( p , ⋅ ) is smooth near C C C . On a sufficiently small boundary neighborhood U U U of C C C , set
u = r ∣ ∂ M − π 2 , b = ⟨ ∇ M r , ν ⟩ . u=r|_{\partial M}-\frac{\pi}{2},\qquad b=\langle\nabla^M r,\nu\rangle. u = r ∣ ∂ M − 2 π , b = ⟨ ∇ M r , ν ⟩ .
We can arrange that
u ≥ 0 , C = { u = 0 } , b > 0 , π 2 ≤ r < π u\geq0,\qquad C=\{u=0\},\qquad b>0,\qquad\frac{\pi}{2}\leq r<\pi u ≥ 0 , C = { u = 0 } , b > 0 , 2 π ≤ r < π
on U U U .
The nearby minimizing geodesics lie in M M M , as in Lemma 4.1 , so the boundary Hessian formula and Hessian comparison under sec g ≥ 1 \operatorname{sec}_g\geq1 sec g ≥ 1 give
∇ 2 u + b h = ∇ M 2 r ∣ T ∂ M ≤ cot r ( g ˉ − d u ⊗ d u ) ≤ 0. \nabla^2u+bh=\left.\nabla^2_Mr\right|_{T\partial M}
\leq\cot r\,(\bar g-du\otimes du)\leq0. ∇ 2 u + bh = ∇ M 2 r T ∂ M ≤ cot r ( g ˉ − d u ⊗ d u ) ≤ 0.
Here ∣ d u ∣ g ˉ ≤ 1 |du|_{\bar g}\leq1 ∣ d u ∣ g ˉ ≤ 1 , so g ˉ − d u ⊗ d u \bar g-du\otimes du g ˉ − d u ⊗ d u is positive semidefinite, whereas cot r ≤ 0 \cot r\leq0 cot r ≤ 0 . Proposition 4.5 now gives A ≡ 0 A\equiv0 A ≡ 0 , so ∂ M \partial M ∂ M is totally geodesic. Since sec g ≥ 1 \operatorname{sec}_g\geq1 sec g ≥ 1 implies Ric g ≥ 2 g \operatorname{Ric}_g\geq2g Ric g ≥ 2 g , the hemisphere rigidity theorem of Hang–Wang [1 ] gives ( M 3 , g ) ≅ S + 3 (M^3,g)\cong S^3_+ ( M 3 , g ) ≅ S + 3 . Conversely, the standard upper hemisphere has inradius π / 2 \pi/2 π /2 . This completes the proof. □
Declaration of AI use. We used ChatGPT for proof exploration and editing. We take responsibility for verification and exposition.
References
School of Mathematical Sciences, Laboratory of Mathematics and Complex Systems, Beijing Normal University, Beijing 100875, P. R. China Email address: jge@bnu.edu.cn
School of Mathematical Sciences, Shanghai Jiao Tong University, Shanghai 200240, China Email address: laimijia@sjtu.edu.cn
School of Mathematical Sciences, Shanghai Jiao Tong University, Shanghai 200240, China Email address: fxy012805@sjtu.edu.cn