Introduction
Introduction
Prescribed area-measure and curvature equations form a central part of the Brunn–Minkowski theory and fully nonlinear geometric analysis. In the closed setting, the classical Christoffel–Minkowski problem is naturally expressed through elementary symmetric functions of the principal radii; its fully nonlinear analytic theory is closely related to Hessian equations and the convexity principle [7 , 8 , 4 ] . The L p L_p L p extension is linked to the Brunn–Minkowski–Firey theory [18 ] . The Orlicz–Minkowski problem and the broader Orlicz–Brunn–Minkowski theory were developed in the works [6 , 5 , 26 ] . Curvature flows provide an effective approach to these problems: Ivaki [12 ] studied flows driven by σ k \sigma_k σ k of the principal radii and the support function, while Bryan–Ivaki–Scheuer developed broad parabolic approaches to curvature equations and Orlicz–Minkowski flows [2 , 1 ] . Their inverse-concavity and viscosity methods provide an important model for the present work [3 ] .
For capillary hypersurfaces in the Euclidean half-space, the support function is defined on a spherical cap and satisfies a Robin boundary condition. The capillary Minkowski problem was introduced and solved in [19 ] , with an L p L_p L p extension developed in [20 ] . The intermediate-order capillary Christoffel–Minkowski problem has been studied from complementary viewpoints in [22 , 10 ] , and related capillary L p L_p L p curvature and Christoffel–Minkowski problems were treated in [21 , 9 , 11 ] . Capillary Alexandrov–Fenchel and Minkowski-type inequalities have also been obtained by curvature-flow methods [27 , 28 ] . At the Monge–Ampère endpoint k = n k=n k = n , the capillary Orlicz–Minkowski problem was developed in [25 ] , while Li and Liu [24 ] studied the capillary Orlicz–Minkowski flow in the upper half-space.
The present paper concerns the intermediate orders 1 ≤ k ≤ n − 1 1 \le k \le n-1 1 ≤ k ≤ n − 1 . In this sense, it is complementary to the determinant-level flow of Li–Liu and is a dynamic companion to the elliptic capillary Orlicz–Christoffel–Minkowski problem studied by Zhu [23 ] . Let
R + n + 1 = { x ∈ R n + 1 : x n + 1 > 0 } , e = − E n + 1 , \mathbb{R}^{n+1}_{+}=\{x\in\mathbb{R}^{n+1}:x_{n+1}>0\},\qquad e=-E_{n+1}, R + n + 1 = { x ∈ R n + 1 : x n + 1 > 0 } , e = − E n + 1 ,
and let C θ C_\theta C θ be the unit capillary spherical cap with non-obtuse contact angle
0 < θ ≤ π 2 . 0<\theta\le\frac{\pi}{2}. 0 < θ ≤ 2 π .
Its support function is
ℓ ( ξ ) = sin 2 θ + cos θ ⟨ ξ , e ⟩ , \ell(\xi)=\sin^2\theta+\cos\theta\langle\xi,e\rangle, ℓ ( ξ ) = sin 2 θ + cos θ ⟨ ξ , e ⟩ ,
and
∇ 2 ℓ + ℓ σ = σ , ∇ μ ℓ = cot θ ℓ . (1) \nabla^2\ell+\ell\sigma=\sigma,\qquad\nabla_\mu\ell=\cot\theta\,\ell. \tag*{(1)} ∇ 2 ℓ + ℓ σ = σ , ∇ μ ℓ = cot θ ℓ . (1)
For 0 < θ ≤ π / 2 0<\theta\le\pi/2 0 < θ ≤ π /2 one has
0 < 1 − cos θ ≤ ℓ ≤ sin 2 θ , (2) 0<1-\cos\theta\le\ell\le\sin^2\theta, \tag*{(2)} 0 < 1 − cos θ ≤ ℓ ≤ sin 2 θ , (2)
with ℓ = 1 \ell=1 ℓ = 1 when θ = π / 2 \theta=\pi/2 θ = π /2 . For a positive capillary support function h h h we put
W [ h ] = ∇ 2 h + h σ . W[h]=\nabla^2h+h\sigma. W [ h ] = ∇ 2 h + hσ .
The eigenvalues of W [ h ] W[h] W [ h ] are the principal radii. We use the normalization
σ ‾ k ( A ) = ( n k ) − 1 σ k ( A ) . \overline{\sigma}_k(A)=\binom{n}{k}^{-1}\sigma_k(A). σ k ( A ) = ( k n ) − 1 σ k ( A ) .
The stationary Orlicz–Christoffel–Minkowski equation is
h φ ( ℓ / h ) σ ˉ k ( W [ h ] ) = f in C θ , ∇ μ h = cot θ h on ∂ C θ . (3) h\varphi(\ell/h)\bar{\sigma}_k(W[h])=f \quad\text{in } C_\theta,\qquad\nabla_\mu h=\cot\theta\,h \quad\text{on } \partial C_\theta. \tag*{(3)} h φ ( ℓ / h ) σ ˉ k ( W [ h ]) = f in C θ , ∇ μ h = cot θ h on ∂ C θ . (3)
We study the flow without a global normalization term
∂ t h = h 2 φ ( ℓ / h ) f σ ˉ k ( W [ h ] ) − h , ∇ μ h = cot θ h . (4) \partial_t h=\frac{h^2\varphi(\ell/h)}{f}\bar{\sigma}_k(W[h])-h,\qquad\nabla_\mu h=\cot\theta\,h. \tag*{(4)} ∂ t h = f h 2 φ ( ℓ / h ) σ ˉ k ( W [ h ]) − h , ∇ μ h = cot θ h . (4)
Its stationary points are exactly the solutions of (3) . In the power case φ ( s ) = s p \varphi(s)=s^p φ ( s ) = s p , the speed becomes
ℓ p f h 2 − p σ ˉ k ( W [ h ] ) − h , \frac{\ell^p}{f}h^{2-p}\bar{\sigma}_k(W[h])-h, f ℓ p h 2 − p σ ˉ k ( W [ h ]) − h ,
which is the capillary analogue of the closed L p L_p L p Christoffel–Minkowski flows in [12 , 1 ] .
Our first observation is variational. The functional
M k [ h ] = 1 k + 1 ∫ C θ h σ ˉ k ( W [ h ] ) d σ \mathcal{M}_k[h]=\frac{1}{k+1}\int_{C_\theta}h\bar{\sigma}_k(W[h])\,d\sigma M k [ h ] = k + 1 1 ∫ C θ h σ ˉ k ( W [ h ]) d σ
has first variation
D M k [ h ] ( η ) = ∫ C θ η σ ˉ k ( W [ h ] ) d σ D\mathcal{M}_k[h](\eta)=\int_{C_\theta}\eta\bar{\sigma}_k(W[h])\,d\sigma D M k [ h ] ( η ) = ∫ C θ η σ ˉ k ( W [ h ]) d σ
for Robin variations. If
J φ ( s ) = ∫ 1 s d τ τ φ ( 1 / τ ) , J_\varphi(s)=\int_1^s\frac{d\tau}{\tau\varphi(1/\tau)}, J φ ( s ) = ∫ 1 s τ φ ( 1/ τ ) d τ ,
then
E f , φ [ h ] = M k [ h ] − ∫ C θ f J φ ( h / ℓ ) d σ \mathcal{E}_{f,\varphi}[h]=\mathcal{M}_k[h]-\int_{C_\theta}fJ_\varphi(h/\ell)\,d\sigma E f , φ [ h ] = M k [ h ] − ∫ C θ f J φ ( h / ℓ ) d σ
has Euler–Lagrange equation (3) , and (4) is a gradient ascent flow for this functional with positive mobility.
The central scalar quantity is
Q = h φ ( ℓ / h ) σ ˉ k ( W [ h ] ) f . Q=\frac{h\varphi(\ell/h)\bar{\sigma}_k(W[h])}{f}. Q = f h φ ( ℓ / h ) σ ˉ k ( W [ h ]) .
We prove that Q Q Q satisfies a closed scalar parabolic equation with a homogeneous Neumann boundary condition. In the supercritical range
a φ ( s ) : = s φ ′ ( s ) φ ( s ) ≥ k + 1 + δ ( s > 0 ) (5) a_\varphi(s):=\frac{s\varphi'(s)}{\varphi(s)}\ge k+1+\delta\qquad(s>0) \tag*{(5)} a φ ( s ) := φ ( s ) s φ ′ ( s ) ≥ k + 1 + δ ( s > 0 ) (5)
this yields time-uniform upper and lower bounds for Q Q Q and, eventually, exponential convergence Q → 1 Q\to1 Q → 1 .
The main new difficulty is the principal-radius estimate. For the dynamic problem it is useful to introduce
H = f ℓ k − 1 , q 0 = log ℓ , K = log H . H=f\ell^{k-1},\qquad q_0=\log\ell,\qquad K=\log H. H = f ℓ k − 1 , q 0 = log ℓ , K = log H .
For s > 0 s>0 s > 0 put
b ( s ) = a φ ( s ) − 2 , d ( s ) = s a φ ′ ( s ) , b(s)=a_\varphi(s)-2,\qquad d(s)=sa_\varphi'(s), b ( s ) = a φ ( s ) − 2 , d ( s ) = s a φ ′ ( s ) ,
and
A φ ℓ ( s ) = b ( s ) + d ( s ) + b ( s ) 2 k + 1 . A_{\varphi}^{\ell}(s)=b(s)+d(s)+\frac{b(s)^2}{k+1}. A φ ℓ ( s ) = b ( s ) + d ( s ) + k + 1 b ( s ) 2 .
The dynamic structural tensor is
T φ , f ( s ) : = − ∇ 2 K − 2 d q 0 ⊙ d K + b ( s ) + d ( s ) ( k + 1 ) A φ ℓ ( s ) d K ⊗ d K + b ( s ) + k + 1 ℓ σ . (6) \mathcal{T}_{\varphi,f}(s):=-\nabla^2K-2\,dq_0\mathbin{\odot}dK+\frac{b(s)+d(s)}{(k+1)A_{\varphi}^{\ell}(s)}dK\otimes dK
+\frac{b(s)+k+1}{\ell}\sigma. \tag*{(6)} T φ , f ( s ) := − ∇ 2 K − 2 d q 0 ⊙ d K + ( k + 1 ) A φ ℓ ( s ) b ( s ) + d ( s ) d K ⊗ d K + ℓ b ( s ) + k + 1 σ . (6)
Here α ⊙ β = 1 2 ( α ⊗ β + β ⊗ α ) \alpha\mathbin{\odot}\beta=\frac{1}{2}(\alpha\otimes\beta+\beta\otimes\alpha) α ⊙ β = 2 1 ( α ⊗ β + β ⊗ α ) . In the power case φ ( s ) = s p \varphi(s)=s^p φ ( s ) = s p , (6) reduces to
T φ , f = ( p + k − 1 ) ∇ 2 ( ℓ H − 1 / ( p + k − 1 ) ) + ℓ H − 1 / ( p + k − 1 ) σ ℓ H − 1 / ( p + k − 1 ) . \mathcal{T}_{\varphi,f}=(p+k-1)\frac{\nabla^2(\ell H^{-1/(p+k-1)})+\ell H^{-1/(p+k-1)}\sigma}{\ell H^{-1/(p+k-1)}}. T φ , f = ( p + k − 1 ) ℓ H − 1/ ( p + k − 1 ) ∇ 2 ( ℓ H − 1/ ( p + k − 1 ) ) + ℓ H − 1/ ( p + k − 1 ) σ .
recovering the spherical convexity exponent in [12 ] .
We now state the quantitative theorem. The boundary assumption is not claimed to be optimal. The curvature estimate is an a priori statement on a smooth existence interval. Its passage to global smooth existence uses the classical parabolic estimates proved in Lemma 2.1 and the quantitative interior concave estimate in Lemma 5.1. The reflection argument below reduces the boundary regularity step to the latter estimate.
Theorem 1.1 (Radius bounds and global flow). Let
0 < θ ≤ π 2 , 1 ≤ k ≤ n − 1. 0<\theta\le\frac{\pi}{2},\qquad1\le k\le n-1. 0 < θ ≤ 2 π , 1 ≤ k ≤ n − 1. Let f ∈ C ∞ ( C θ ‾ ) f\in C^\infty(\overline{C_\theta}) f ∈ C ∞ ( C θ ) and φ ∈ C ∞ ( ( 0 , ∞ ) ) \varphi\in C^\infty((0,\infty)) φ ∈ C ∞ (( 0 , ∞ )) be positive. Assume (5),
s α φ ′ ( s ) ≥ 0 , (7) s\alpha'_{\varphi}(s)\ge0, \tag*{(7)} s α φ ′ ( s ) ≥ 0 , (7) and suppose that there is τ 0 > 0 \tau_0>0 τ 0 > 0 such that
T φ , f ( s ) ≥ τ 0 σ for every s > 0. (8) \mathcal{T}_{\varphi,f}(s)\ge\tau_0\sigma\qquad\text{for every }s>0. \tag*{(8)} T φ , f ( s ) ≥ τ 0 σ for every s > 0. (8) Assume also
f μ f ≥ n − k n cot θ on ∂ C θ . (9) \frac{f_\mu}{f}\ge\frac{n-k}{n}\cot\theta\qquad\text{on }\partial C_\theta. \tag*{(9)} f f μ ≥ n n − k cot θ on ∂ C θ . (9) Let h 0 h_0 h 0 be a smooth positive strictly convex capillary support function satisfying the compatibility hierarchy specified in Subsection 2.2. Every smooth positive solution of (4) issuing from h 0 h_0 h 0 remains strictly convex and satisfies
0 < c σ ≤ W [ h ( ⋅ , t ) ] ≤ C σ ( 0 ≤ t < T ) , 0<c\sigma\le W[h(\cdot,t)]\le C\sigma\qquad(0\le t<T), 0 < c σ ≤ W [ h ( ⋅ , t )] ≤ C σ ( 0 ≤ t < T ) , on its smooth existence interval [ 0 , T ) [0,T) [ 0 , T ) , with c , C c,C c , C independent of T T T . Moreover, (4) admits a unique smooth solution for all t ≥ 0 t\ge0 t ≥ 0 , and there exists a smooth strictly convex capillary support function h ∞ h_\infty h ∞ such that
h ( ⋅ , t ) ⟶ h ∞ in C ∞ ( C θ ‾ ) as t → ∞ . h(\cdot,t)\longrightarrow h_\infty\quad\text{in }C^\infty(\overline{C_\theta})\text{ as }t\to\infty. h ( ⋅ , t ) ⟶ h ∞ in C ∞ ( C θ ) as t → ∞. The convergence of Q Q Q to 1 is exponential, and h ∞ h_\infty h ∞ solves (3).
The natural capillary boundary condition arising in the static problem is the weaker condition
H μ ≥ 0 , H = f ℓ k − 1 , (10) H_{\mu} \ge0,\qquad H = f\ell^{k-1}, \tag*{(10)} H μ ≥ 0 , H = f ℓ k − 1 , (10)
or equivalently
f μ f + ( k − 1 ) cot θ ≥ 0. \frac{f_{\mu}}{f} + (k-1)\cot\theta\ge0. f f μ + ( k − 1 ) cot θ ≥ 0.
Although the argument below does not yield a time-uniform quantitative non-collapsing estimate from (10) alone, strict convexity is preserved.
Theorem 1.2 (Finite-time preservation under the natural boundary sign). Let 1 ≤ k ≤ n − 1 1 \le k \le n-1 1 ≤ k ≤ n − 1 and 0 < θ ≤ π / 2 0 < \theta\le\pi/2 0 < θ ≤ π /2 , and let f f f and φ \varphi φ be smooth and positive as above. Assume (5), (7), T φ , f ( s ) ≥ 0 T_{\varphi,f}(s) \ge0 T φ , f ( s ) ≥ 0 for every s > 0 s > 0 s > 0 , and (10). If a smooth positive solution of (4) exists on [ 0 , T ] [0,T] [ 0 , T ] and W [ h 0 ] > 0 W[h_0] > 0 W [ h 0 ] > 0 , then
W [ h ( ⋅ , t ) ] > 0 for every 0 ≤ t ≤ T . W[h(\cdot,t)] > 0 \qquad\text{for every } 0 \le t \le T. W [ h ( ⋅ , t )] > 0 for every 0 ≤ t ≤ T .
Theorem 1.3 (Full-angle scalar flow for k = 1 k=1 k = 1 ). Let 0 < θ < π 0 < \theta< \pi 0 < θ < π , let f ∈ C ∞ ( C θ ‾ ) f \in C^{\infty}(\overline{C_{\theta}}) f ∈ C ∞ ( C θ ) and φ ∈ C ∞ ( ( 0 , ∞ ) ) \varphi\in C^{\infty}((0,\infty)) φ ∈ C ∞ (( 0 , ∞ )) be positive, and assume
a φ ( s ) ≥ 2 + δ ( s > 0 ) a_{\varphi}(s) \ge2+\delta\qquad(s>0) a φ ( s ) ≥ 2 + δ ( s > 0 ) for some δ > 0 \delta> 0 δ > 0 . Let h 0 > 0 h_0 > 0 h 0 > 0 be smooth and satisfy the Robin compatibility conditions, and assume
σ ‾ 1 ( W [ h 0 ] ) > 0. \overline{\sigma}_1(W[h_0]) > 0. σ 1 ( W [ h 0 ]) > 0. Then the scalar flow
h t = h 2 φ ( ℓ / h ) f σ ‾ 1 ( W [ h ] ) − h , h μ = cot θ h , h_t = \frac{h^2\varphi(\ell/h)}{f}\overline{\sigma}_1(W[h])-h,\qquad h_{\mu}=\cot\theta\,h, h t = f h 2 φ ( ℓ / h ) σ 1 ( W [ h ]) − h , h μ = cot θ h , has a unique smooth positive solution for all t ≥ 0 t \ge0 t ≥ 0 . Its speed ratio satisfies
0 < c Q ≤ Q ( ξ , t ) ≤ C Q , ∥ Q ( ⋅ , t ) − 1 ∥ L ∞ ≤ C e − c t , 0 < c_Q \le Q(\xi,t) \le C_Q,\qquad\|Q(\cdot,t)-1\|_{L^\infty} \le Ce^{-ct}, 0 < c Q ≤ Q ( ξ , t ) ≤ C Q , ∥ Q ( ⋅ , t ) − 1 ∥ L ∞ ≤ C e − c t , and there is a unique positive stationary solution h ∞ h_{\infty} h ∞ with σ ‾ 1 ( W [ h ∞ ] ) > 0 \overline{\sigma}_1(W[h_{\infty}])>0 σ 1 ( W [ h ∞ ]) > 0 such that
h ( ⋅ , t ) ⟶ h ∞ in C ∞ ( C θ ‾ ) . h(\cdot,t) \longrightarrow h_{\infty}\quad\text{in }C^{\infty}(\overline{C_{\theta}}). h ( ⋅ , t ) ⟶ h ∞ in C ∞ ( C θ ) . For θ > π / 2 \theta> \pi/2 θ > π /2 this theorem is a scalar parabolic result; it does not assert that W [ h ( ⋅ , t ) ] W[h(\cdot,t)] W [ h ( ⋅ , t )] remains positive definite.
Remark 1.4 (Contact-angle range). The support-function identities (1), the Robin first variation, the C 0 C^0 C 0 and C 1 C^1 C 1 estimates, the Q Q Q -equation, and the interior Orlicz square-completion are meaningful for every fixed contact angle 0 < θ < π 0 < \theta< \pi 0 < θ < π because ℓ \ell ℓ remains positive. The present high-order global theorem is restricted to the non-obtuse range because a tangential minimum-radius support v = h W ( e , e ) v=hW(e,e) v = hW ( e , e ) satisfies
v μ = h cot θ λ n . v_{\mu}=h\cot\theta\,\lambda_n. v μ = h cot θ λ n .
For 0 < θ < π / 2 0 < \theta< \pi/2 0 < θ < π /2 this is positive, while at θ = π / 2 \theta=\pi/2 θ = π /2 it is zero; both signs are compatible with the boundary barrier used in the non-collapsing argument. Thus the right-angle case requires no new curvature mechanism. For π / 2 < θ < π \pi/2 < \theta< \pi π /2 < θ < π , however, cot θ < 0 \cot\theta< 0 cot θ < 0 and this boundary sign is reversed. The present argument therefore does not control the obtuse-angle boundary. This is the geometric obstruction in the present proof to extending Theorems 1.1 –1.2 to all contact angles. For k = 1 k = 1 k = 1 the transformed scalar equation is uniformly parabolic after the support bound, giving the separate full-angle result above. Preservation of geometric strict convexity in the obtuse range is not asserted here.
The paper is organized into five sections. § 3 develops the capillary support-function geometry, Newton-tensor identities, short-time flow, variational structure, and the speed-ratio equation. Section 3 establishes the C 0 C^{0} C 0 , C 1 C^{1} C 1 , and Q Q Q estimates and derives the dynamic Orlicz square completion. § 5 proves preservation of strict convexity and the quantitative non-collapsing and two-sided principal-radius estimates. § 6 contains the long-time existence and convergence arguments, the independent stationary uniqueness proof, the all-angle scalar result for k = 1 k = 1 k = 1 , and the discussion of the natural boundary condition and the remaining obtuse-angle problem.
Geometric and variational framework
Capillary support functions and Newton tensors
Consider
C θ = { ξ ∈ R + n + 1 : ∣ ξ − cos θ e ∣ = 1 } , e = − E n + 1 , C_{\theta} = \left\{ \xi\in\mathbb{R}_{+}^{n+1} : \left|\xi- \cos\theta e\right| = 1 \right\}, \qquad e = -E_{n+1}, C θ = { ξ ∈ R + n + 1 : ∣ ξ − cos θ e ∣ = 1 } , e = − E n + 1 ,
equipped with its round metric σ \sigma σ . We write μ \mu μ for the outward unit conormal of ∂ C θ ⊂ C θ \partial C_{\theta} \subset C_{\theta} ∂ C θ ⊂ C θ . The function ℓ \ell ℓ in (1) is positive for every fixed 0 < θ < π 0 < \theta< \pi 0 < θ < π . Indeed, if t = ⟨ ξ , e ⟩ t = \langle\xi,e\rangle t = ⟨ ξ , e ⟩ , then t ∈ [ cos θ − 1 , 0 ] t \in[\cos\theta- 1,0] t ∈ [ cos θ − 1 , 0 ] on C θ C_{\theta} C θ , and hence
ℓ ( ξ ) = sin 2 θ + cos θ t . \ell(\xi) = \sin^{2}\theta+ \cos\theta t. ℓ ( ξ ) = sin 2 θ + cos θ t .
Consequently
0 < ℓ − ( θ ) : = min { 1 − cos θ , sin 2 θ } ≤ ℓ ≤ max { 1 − cos θ , sin 2 θ } = : ℓ + ( θ ) . (11) 0 < \ell_{-}(\theta) := \min\left\{1 - \cos\theta,\sin^{2}\theta\right\} \leq\ell\leq\max\left\{1 - \cos\theta,\sin^{2}\theta\right\} =: \ell_{+}(\theta). \tag*{(11)} 0 < ℓ − ( θ ) := min { 1 − cos θ , sin 2 θ } ≤ ℓ ≤ max { 1 − cos θ , sin 2 θ } =: ℓ + ( θ ) . (11)
In the non-obtuse range used in the high-order main theorem this reduces to (2).
If a strictly convex capillary hypersurface is parametrized by its capillary Gauss map, its position vector can be reconstructed from the support function by
X ( ξ ) = ∇ h + h ( ξ − cos θ e ) . X(\xi) = \nabla h + h(\xi- \cos\theta e). X ( ξ ) = ∇ h + h ( ξ − cos θ e ) .
The capillary angle condition becomes
∇ μ h = cot θ h . (12) \nabla_{\mu}h = \cot\theta h. \tag*{(12)} ∇ μ h = cot θ h . (12)
The radius tensor is
W [ h ] = ∇ 2 h + h σ . W[h] = \nabla^{2}h + h\sigma. W [ h ] = ∇ 2 h + hσ .
Strict convexity is equivalent to W [ h ] > 0 W[h] > 0 W [ h ] > 0 .
For a symmetric matrix A A A , let σ j ( A ) \sigma_{j}(A) σ j ( A ) denote the j j j -th elementary symmetric function and set
σ ‾ j ( A ) = ( n j ) − 1 σ j ( A ) . \overline{\sigma}_{j}(A) = \binom{n}{j}^{-1}\sigma_{j}(A). σ j ( A ) = ( j n ) − 1 σ j ( A ) .
The Gårding cone is
Γ k = { λ : σ j ( λ ) > 0 , 1 ≤ j ≤ k } . \Gamma_k = \{\lambda: \sigma_j(\lambda) > 0,\ 1 \le j \le k\}. Γ k = { λ : σ j ( λ ) > 0 , 1 ≤ j ≤ k } .
We use
S i j ( A ) = ∂ σ ‾ k ( A ) ∂ A i j . S^{ij}(A) = \frac{\partial\overline{\sigma}_k(A)}{\partial A_{ij}}. S ij ( A ) = ∂ A ij ∂ σ k ( A ) .
When A = W [ h ] A = W[h] A = W [ h ] , the Codazzi property implies
∇ i S i j = 0. (13) \nabla_i S^{ij} = 0. \tag*{(13)} ∇ i S ij = 0. (13)
Euler homogeneity gives
S i j W i j = k σ ‾ k ( W ) . (14) S^{ij}W_{ij} = k\overline{\sigma}_k(W). \tag*{(14)} S ij W ij = k σ k ( W ) . (14)
The function σ ‾ k 1 / k \overline{\sigma}_k^{1/k} σ k 1/ k is concave on Γ k \Gamma_k Γ k and inverse-concave in the sense of [3 Definition 1.3] . Here the inverse-concavity test is applied to positive definite matrices: the map A ↦ − σ ‾ k ( A − 1 ) 1 / k A \mapsto-\overline{\sigma}_k(A^{-1})^{1/k} A ↦ − σ k ( A − 1 ) 1/ k is concave there. All uses of the corresponding differential inequality below take place while W [ h ] > 0 W[h] > 0 W [ h ] > 0 .
The Robin condition has two useful consequences. In a boundary orthonormal frame { e α } α < n \{e_\alpha\}_{\alpha<n} { e α } α < n , e n = μ e_n = \mu e n = μ , one has
W α n = 0 (15) W_{\alpha n} = 0 \tag*{(15)} W α n = 0 (15)
and
W α α ; μ = cot θ ( W n n − W α α ) . (16) W_{\alpha\alpha;\mu} = \cot\theta\left(W_{nn} - W_{\alpha\alpha}\right). \tag*{(16)} W α α ; μ = cot θ ( W nn − W α α ) . (16)
For completeness, let κ = cot θ \kappa= \cot\theta κ = cot θ . The outward conormal satisfies ∇ e α μ = κ e α \nabla_{e_\alpha}\mu= \kappa e_\alpha ∇ e α μ = κ e α . Differentiating h μ = κ h h_\mu= \kappa h h μ = κh tangentially therefore gives
0 = e α ( h μ − κ h ) = ∇ 2 h ( e α , μ ) , 0 = e_\alpha(h_\mu- \kappa h) = \nabla^2 h(e_\alpha,\mu), 0 = e α ( h μ − κh ) = ∇ 2 h ( e α , μ ) ,
and hence W α n = 0 W_{\alpha n} = 0 W α n = 0 . The tensor W = ∇ 2 h + h σ W = \nabla^2 h + h\sigma W = ∇ 2 h + hσ is Codazzi on the unit sphere. At a boundary point choose the tangential frame to be geodesic for the induced boundary metric. Then ∇ e α e α = − κ μ \nabla_{e_\alpha}e_\alpha= -\kappa\mu ∇ e α e α = − κ μ at that point, and
( ∇ μ W ) ( e α , e α ) = ( ∇ e α W ) ( μ , e α ) = κ W n n − κ W α α . (\nabla_\mu W)(e_\alpha,e_\alpha) = (\nabla_{e_\alpha}W)(\mu,e_\alpha) = \kappa W_{nn} - \kappa W_{\alpha\alpha}. ( ∇ μ W ) ( e α , e α ) = ( ∇ e α W ) ( μ , e α ) = κ W nn − κ W α α .
This proves both boundary identities, including the case κ = 0 \kappa= 0 κ = 0 .
It is often useful to divide by ℓ \ell ℓ :
u = h ℓ . u = \frac{h}{\ell}. u = ℓ h .
Then (1) and (12) imply
u μ = 0 , (17) u_\mu= 0, \tag*{(17)} u μ = 0 , (17)
and
W [ h ] = ℓ ∇ 2 u + ∇ ℓ ⊗ ∇ u + ∇ u ⊗ ∇ ℓ + u σ . (18) W[h] = \ell\nabla^2u + \nabla\ell\otimes\nabla u + \nabla u \otimes\nabla\ell+ u\sigma. \tag*{(18)} W [ h ] = ℓ ∇ 2 u + ∇ ℓ ⊗ ∇ u + ∇ u ⊗ ∇ ℓ + u σ . (18)
The flow, parabolic estimates, and the variational structure
We record and prove the parabolic results used in the continuation arguments, including their precise classical sources.
Lemma 2.1 (Parabolic estimates and local existence). On the fixed smooth cap with 0 < θ < π 0 < \theta< \pi 0 < θ < π , the following statements hold. The linear estimates use fixed uniform ellipticity constants.
A bounded smooth solution of a uniformly parabolic linear nondivergence equation with bounded measurable principal coefficients, bounded drift, bounded zeroth-order coefficient, and bounded source has a uniform parabolic Hölder estimate on smaller positive-time cylinders. The same estimate holds up to the boundary under homogeneous Neumann data.
Let v v v be a bounded smooth solution of a linear equation with homogeneous boundary data B v = 0 Bv = 0 B v = 0 . Let 0 < γ < 1 0 < \gamma< 1 0 < γ < 1 . If its coefficients and source are uniformly C γ , γ / 2 C^{\gamma,\gamma/2} C γ , γ /2 , and the fixed boundary operator is smooth and uniformly oblique of Neumann or Robin type, then the solution has uniform C 2 + γ , 1 + γ / 2 C^{2+\gamma,1+\gamma/2} C 2 + γ , 1 + γ /2 estimates on smaller positive-time strips, controlled by the source norm and ∥ v ∥ ∞ \|v\|_\infty ∥ v ∥ ∞ . For inhomogeneous data B v = Φ Bv = \Phi B v = Φ , add the boundary norm ∥ Φ ∥ C 1 + γ , ( 1 + γ ) / 2 \|\Phi\|_{C^{1+\gamma,(1+\gamma)/2}} ∥Φ ∥ C 1 + γ , ( 1 + γ ) /2 . Higher-order versions hold with the corresponding coefficient, source, and boundary-data norms.
A smooth fully nonlinear scalar speed whose Hessian derivative is positive definite near the initial two-jet, together with a fixed smooth linear uniformly oblique boundary operator B B B , the homogeneous condition B h = 0 Bh = 0 B h = 0 , and smooth compatible initial data, has a unique smooth short-time solution. Compatibility means all identities obtained by differentiating B h = 0 Bh = 0 B h = 0 in time and using the equation. A smooth compatible uniformly parabolic endpoint jet can be used to restart the solution.
The positive-time constants depend only on the indicated data norms, ellipticity, the fixed geometry and boundary operator, and strip lengths and separation. They are uniform under time translation.
Proof. For the first assertion, absorb any bounded zeroth-order term d v dv d v into the source. Apply Krylov–Safonov [14 Theorems 4.2–4.3, p. 164] ; bounded forcing controls the required L n + 1 L^{n+1} L n + 1 norm on each fixed cylinder. In Fermi coordinates a Neumann solution satisfies v r = v y r = 0 v_r = v_{yr} = 0 v r = v y r = 0 at the boundary, so its even extension is a strong W n + 1 2 , 1 W^{2,1}_{n+1} W n + 1 2 , 1 solution. Reflect the principal matrix by R A R RAR R A R and the drift by R b Rb R b , where R = diag ( 1 , … , 1 , − 1 ) R = \operatorname{diag}(1,\ldots,1,-1) R = diag ( 1 , … , 1 , − 1 ) . The extended coefficients remain bounded and uniformly elliptic; coefficient jumps are allowed in the cited theorem. Its backward parabolic distance gives the asserted positive-time estimates.
For the second assertion, use Ladyzhenskaya–Solonnikov–Ural’tseva [15 Chapter IV, Theorem 5.3, pp. 320–321] for problem (5.4). Stereographic projection from the omitted south pole identifies the fixed cap with a bounded Euclidean ball. Since θ < π \theta< \pi θ < π , the map, the metric coefficients, and their inverses are smooth on its closure. Uniform parabolicity and obliqueness are preserved with fixed constants, so the cited Euclidean-cylinder theorem applies. For l = γ ∈ ( 0 , 1 ) l = \gamma\in(0,1) l = γ ∈ ( 0 , 1 ) , the coefficients of the equation and its source are in H l , l / 2 H^{l,l/2} H l , l /2 , the boundary coefficients and inhomogeneous first-order boundary data are in H l + 1 , ( l + 1 ) / 2 H^{l+1,(l+1)/2} H l + 1 , ( l + 1 ) /2 , and the solution is in H l + 2 , l / 2 + 1 H^{l+2,l/2+1} H l + 2 , l /2 + 1 . The fixed cap and boundary operator satisfy the required smoothness and obliqueness conditions. A smooth time cutoff χ \chi χ vanishing near the earlier end of a strip removes initial compatibility issues. The cutoff equation has source χ r + χ ′ v \chi r+\chi'v χr + χ ′ v when L v = r Lv=r Lv = r , and boundary data χ Φ \chi\Phi χ Φ . In the applications below, the lower Hölder norm of v v v is already controlled. The estimate in terms of ∥ v ∥ ∞ \lVert v\rVert_\infty ∥ v ∥ ∞ alone follows by interpolation and nested-strip absorption: on nested strips J ′ ⋐ p J J' \Subset_p J J ′ ⋐ p J , with the same terminal time,
∥ v ∥ 2 + γ ; J ′ ≤ ε ∥ v ∥ 2 + γ ; J + C ε gap ( J ′ , J ) − q ( ∥ v ∥ ∞ + ∥ r ∥ γ ; J + ∥ Φ ∥ 1 + γ ; ∂ J ) \lVert v\rVert_{2+\gamma;J'} \le\varepsilon\lVert v\rVert_{2+\gamma;J}+C_\varepsilon\operatorname{gap}(J',J)^{-q}\left(\lVert v\rVert_\infty+\lVert r\rVert_{\gamma;J}+\lVert\Phi\rVert_{1+\gamma;\partial J}\right) ∥ v ∥ 2 + γ ; J ′ ≤ ε ∥ v ∥ 2 + γ ; J + C ε gap ( J ′ , J ) − q ( ∥ v ∥ ∞ + ∥ r ∥ γ ; J + ∥ Φ ∥ 1 + γ ; ∂ J )
where the displayed norms are parabolic Hölder norms, the boundary norm is on the lateral boundary, and q < ∞ q<\infty q < ∞ . Choose geometrically decreasing gaps and then a fixed ε < 2 − q \varepsilon<2^{-q} ε < 2 − q . Iteration gives a convergent geometric series; its final remainder tends to zero since the solution is smooth on the fixed outer strip. This proves the asserted bound without a prior estimate for the highest derivatives of the solution. Applying the same theorem with larger noninteger l l l proves the higher-order statements. The constants depend on the controlled norms and strip geometry, not on the location of the strips in time.
For local existence, fix 0 < α < 1 0<\alpha<1 0 < α < 1 , write the speed as N ( h ) \mathcal{N}(h) N ( h ) , and set h ∗ = h 0 + t N ( h 0 ) h^*=h_0+t\mathcal{N}(h_0) h ∗ = h 0 + t N ( h 0 ) . The first two compatibility identities give B h ∗ = 0 Bh^*=0 B h ∗ = 0 . On a sufficiently short interval the linearization L ∗ v = v t − D N ( h ∗ ) v L_*v=v_t-D\mathcal{N}(h^*)v L ∗ v = v t − D N ( h ∗ ) v is uniformly parabolic. Use the usual unscaled parabolic Hölder norms and the zero-initial spaces
X T = { v ∈ C 2 + α , 1 + α / 2 : B v = 0 , v ( 0 ) = v t ( 0 ) = 0 } , Y T = { f ∈ C α , α / 2 : f ( 0 ) = 0 } . X_T=\{v\in C^{2+\alpha,1+\alpha/2}:Bv=0,\ v(0)=v_t(0)=0\},\qquad Y_T=\{f\in C^{\alpha,\alpha/2}:f(0)=0\}. X T = { v ∈ C 2 + α , 1 + α /2 : B v = 0 , v ( 0 ) = v t ( 0 ) = 0 } , Y T = { f ∈ C α , α /2 : f ( 0 ) = 0 } .
The source’s zero-initial conditions on p. 321 require f ( 0 ) = 0 f(0)=0 f ( 0 ) = 0 for 0 < l = α < 1 0<l=\alpha<1 0 < l = α < 1 , and the zero boundary source satisfies its trace condition. Conversely the equation with v ( 0 ) = 0 v(0)=0 v ( 0 ) = 0 implies v t ( 0 ) = f ( 0 ) = 0 v_t(0)=f(0)=0 v t ( 0 ) = f ( 0 ) = 0 . Thus [15 Chapter IV, Theorem 5.4 and (5.14), p. 322] gives an inverse for L ∗ : X T → Y T L_*:X_T\to Y_T L ∗ : X T → Y T whose norm remains bounded as T ↓ 0 T\downarrow0 T ↓ 0 . The smooth coefficients of L ∗ L_* L ∗ are restrictions of a fixed family on a sufficiently small interval [ 0 , T 0 ] [0,T_0] [ 0 , T 0 ] . The residual R 0 = N ( h ∗ ) − h t ∗ R_0=\mathcal{N}(h^*)-h_t^* R 0 = N ( h ∗ ) − h t ∗ belongs to Y T Y_T Y T and ∥ R 0 ∥ Y T ≤ C T 1 − α / 2 → 0 \lVert R_0\rVert_{Y_T}\le CT^{1-\alpha/2}\to0 ∥ R 0 ∥ Y T ≤ C T 1 − α /2 → 0 : the residual is smooth and vanishes at t = 0 t=0 t = 0 . The nonlinear remainder
R ( v ) = N ( h ∗ + v ) − N ( h ∗ ) − D N ( h ∗ ) v R(v)=\mathcal{N}(h^*+v)-\mathcal{N}(h^*)-D\mathcal{N}(h^*)v R ( v ) = N ( h ∗ + v ) − N ( h ∗ ) − D N ( h ∗ ) v
vanishes initially because the entire spatial two-jet of v v v vanishes there, and satisfies
∥ R ( v ) − R ( w ) ∥ Y T ≤ C ( ∥ v ∥ X T + ∥ w ∥ X T ) ∥ v − w ∥ X T . \lVert R(v)-R(w)\rVert_{Y_T}\le C\left(\lVert v\rVert_{X_T}+\lVert w\rVert_{X_T}\right)\lVert v-w\rVert_{X_T}. ∥ R ( v ) − R ( w ) ∥ Y T ≤ C ( ∥ v ∥ X T + ∥ w ∥ X T ) ∥ v − w ∥ X T .
The Hölder product and composition constants in this estimate are independent of small T T T . Write C ∗ C_* C ∗ for the inverse bound and C R C_R C R for the remainder constant. First choose a fixed radius ρ \rho ρ within the uniformly parabolic neighborhood with 2 C ∗ C R ρ < 1 / 2 2C_*C_R\rho<1/2 2 C ∗ C R ρ < 1/2 , and then choose T T T so small that C ∗ ∥ R 0 ∥ Y T ≤ ρ / 2 C_*\lVert R_0\rVert_{Y_T}\le\rho/2 C ∗ ∥ R 0 ∥ Y T ≤ ρ /2 . The map v ↦ L ∗ − 1 ( R 0 + R ( v ) ) v\mapsto L_*^{-1}(R_0+R(v)) v ↦ L ∗ − 1 ( R 0 + R ( v )) preserves this ball and is a contraction. Its fixed point gives the local solution. The higher compatibility identities allow the time and tangential boundary bootstrap, initiated by difference quotients and the linear theorem, up to the initial time. Uniform ellipticity then recovers the missing normal derivatives from the equation. This gives smoothness. The difference of two local solutions satisfies a linear uniformly parabolic oblique problem with zero initial and boundary data, so linear uniqueness applies. The same construction at a smooth uniformly parabolic endpoint gives continuation.
We consider
h t = Θ S − h , Θ = h 2 φ ( ℓ / h ) f , S = σ ˉ k ( W [ h ] ) . h_t=\Theta S-h,\qquad\Theta=\frac{h^2\varphi(\ell/h)}{f},\qquad S=\bar{\sigma}_k(W[h]). h t = Θ S − h , Θ = f h 2 φ ( ℓ / h ) , S = σ ˉ k ( W [ h ]) .
Here and below, put B = ∂ μ − cot θ B = \partial_\mu- \cot\theta B = ∂ μ − cot θ and let H j = ∂ t j h ∣ t = 0 H_j = \partial_t^j h|_{t=0} H j = ∂ t j h ∣ t = 0 be determined recursively by the flow, starting with H 0 = h 0 H_0 = h_0 H 0 = h 0 . Compatibility means B H j = 0 BH_j = 0 B H j = 0 for every j ≥ 0 j \ge0 j ≥ 0 . The first nontrivial compatibility condition is
( Θ 0 σ ˉ k ( W [ h 0 ] ) − h 0 ) μ = cot θ ( Θ 0 σ ˉ k ( W [ h 0 ] ) − h 0 ) , \left(\Theta_0\bar{\sigma}_k(W[h_0]) - h_0\right)_\mu= \cot\theta\left(\Theta_0\bar{\sigma}_k(W[h_0]) - h_0\right), ( Θ 0 σ ˉ k ( W [ h 0 ]) − h 0 ) μ = cot θ ( Θ 0 σ ˉ k ( W [ h 0 ]) − h 0 ) ,
or equivalently ( Q 0 ) μ = 0 (Q_0)_\mu= 0 ( Q 0 ) μ = 0 .
We record the conditions for short-time parabolic existence explicitly. In a sufficiently small neighborhood of the initial two-jet, h > 0 h > 0 h > 0 and W [ h ] > 0 W[h] > 0 W [ h ] > 0 . Compactness and smoothness imply that
F i j : = ∂ ∂ h i j ( Θ σ ˉ k ( W [ h ] ) − h ) = Θ S i j F^{ij} := \frac{\partial}{\partial h_{ij}}\left(\Theta\bar{\sigma}_k(W[h]) - h\right) = \Theta S^{ij} F ij := ∂ h ij ∂ ( Θ σ ˉ k ( W [ h ]) − h ) = Θ S ij
has eigenvalues bounded above and below by positive constants. The boundary operator B ( x , z , p ) = ⟨ μ ( x ) , p ⟩ − cot θ z B(x,z,p) = \langle\mu(x),p\rangle- \cot\theta\,z B ( x , z , p ) = ⟨ μ ( x ) , p ⟩ − cot θ z is smooth and uniformly oblique, since B p ⋅ μ = 1 B_p \cdot\mu= 1 B p ⋅ μ = 1 . For example, at h 0 h_0 h 0 the spatial linearization of the speed is
v ⟼ Θ 0 S 0 i j v i j + [ ( 2 − a φ ( ℓ / h 0 ) ) Q 0 + Θ 0 tr ( S 0 i j ) − 1 ] v , v \longmapsto\Theta_0 S_0^{ij}v_{ij} + \left[(2-a_\varphi(\ell/h_0))Q_0+\Theta_0\operatorname{tr}(S_0^{ij})-1\right]v, v ⟼ Θ 0 S 0 ij v ij + [ ( 2 − a φ ( ℓ / h 0 )) Q 0 + Θ 0 tr ( S 0 ij ) − 1 ] v ,
with boundary condition v μ − cot θ v = 0 v_\mu-\cot\theta\,v=0 v μ − cot θ v = 0 . Thus the local-existence hypotheses of Lemma 2.1 are satisfied, giving smooth local existence and uniqueness for compatible initial data. Positivity of h h h and W [ h ] W[h] W [ h ] persists for a short time by continuity. At θ = π / 2 \theta=\pi/2 θ = π /2 the boundary operator is homogeneous Neumann.
First variation and monotone energy
Define
M k [ h ] = 1 k + 1 ∫ C θ h σ ˉ k ( W [ h ] ) d σ . (19) \mathcal{M}_k[h] = \frac{1}{k+1}\int_{\mathcal{C}_\theta}h\bar{\sigma}_k(W[h])\,d\sigma. \tag*{(19)} M k [ h ] = k + 1 1 ∫ C θ h σ ˉ k ( W [ h ]) d σ . (19)
Lemma 2.2 (Robin first variation). If η μ = cot θ η \eta_\mu=\cot\theta\,\eta η μ = cot θ η on ∂ C θ \partial\mathcal{C}_\theta ∂ C θ , then
D M k [ h ] ( η ) = ∫ C θ η σ ˉ k ( W [ h ] ) d σ . D\mathcal{M}_k[h](\eta)=\int_{\mathcal{C}_\theta}\eta\bar{\sigma}_k(W[h])\,d\sigma. D M k [ h ] ( η ) = ∫ C θ η σ ˉ k ( W [ h ]) d σ . Proof. Differentiating (19) produces
1 k + 1 ∫ C θ { η σ ˉ k ( W ) + h S i j ( η i j + η σ i j ) } d σ . \frac{1}{k+1}\int_{\mathcal{C}_\theta}\left\{\eta\bar{\sigma}_k(W)+hS^{ij}(\eta_{ij}+\eta\sigma_{ij})\right\}\,d\sigma. k + 1 1 ∫ C θ { η σ ˉ k ( W ) + h S ij ( η ij + η σ ij ) } d σ . Use (13) twice. The boundary term is
∫ ∂ C θ S μ μ ( h η μ − η h μ ) d s , \int_{\partial\mathcal{C}_\theta}S^{\mu\mu}(h\eta_\mu-\eta h_\mu)\,ds, ∫ ∂ C θ S μμ ( h η μ − η h μ ) d s , because (15) implies that the mixed Newton-tensor components vanish. The common Robin condition cancels this term. The interior term is
1 k + 1 ∫ C θ η S i j W i j d σ = k k + 1 ∫ C θ η σ ˉ k ( W ) d σ , \frac{1}{k+1}\int_{\mathcal{C}_\theta}\eta S^{ij}W_{ij}\,d\sigma
=
\frac{k}{k+1}\int_{\mathcal{C}_\theta}\eta\bar{\sigma}_k(W)\,d\sigma, k + 1 1 ∫ C θ η S ij W ij d σ = k + 1 k ∫ C θ η σ ˉ k ( W ) d σ , and the claim follows.
Let
J φ ( s ) = ∫ 1 s d τ τ φ ( 1 / τ ) J_{\varphi}(s)=\int_{1}^{s}\frac{d\tau}{\tau\varphi(1/\tau)} J φ ( s ) = ∫ 1 s τ φ ( 1/ τ ) d τ
and define
E f , φ [ h ] = M κ [ h ] − ∫ C θ f J φ ( h / ℓ ) d σ . \mathcal{E}_{f,\varphi}[h]=\mathcal{M}_{\kappa}[h]-\int_{\mathcal{C}_{\theta}}fJ_{\varphi}(h/\ell)\,d\sigma. E f , φ [ h ] = M κ [ h ] − ∫ C θ f J φ ( h / ℓ ) d σ .
Then
D E f , φ [ h ] ( η ) = ∫ C θ η ( S − f h φ ( ℓ / h ) ) d σ . D\mathcal{E}_{f,\varphi}[h](\eta)=\int_{\mathcal{C}_{\theta}}\eta\left(S-\frac{f}{h\varphi(\ell/h)}\right)\,d\sigma. D E f , φ [ h ] ( η ) = ∫ C θ η ( S − h φ ( ℓ / h ) f ) d σ .
Since
h t = Θ ( S − f h φ ( ℓ / h ) ) , h_t=\Theta\left(S-\frac{f}{h\varphi(\ell/h)}\right), h t = Θ ( S − h φ ( ℓ / h ) f ) ,
we have the exact identity
d d t E f , φ [ h ( t ) ] = ∫ C θ f h 2 φ ( ℓ / h ) h t 2 d σ ≥ 0. \frac{d}{dt}\mathcal{E}_{f,\varphi}[h(t)]=\int_{\mathcal{C}_{\theta}}\frac{f}{h^2\varphi(\ell/h)}h_t^2\,d\sigma\geq0. d t d E f , φ [ h ( t )] = ∫ C θ h 2 φ ( ℓ / h ) f h t 2 d σ ≥ 0.
The speed ratio
Set
Q = Θ S h = h φ ( ℓ / h ) S f . (20) Q=\frac{\Theta S}{h}=\frac{h\varphi(\ell/h)S}{f}. \tag*{(20)} Q = h Θ S = f h φ ( ℓ / h ) S . (20)
Then
h t = h ( Q − 1 ) . (21) h_t=h(Q-1). \tag*{(21)} h t = h ( Q − 1 ) . (21)
Differentiating (12) in time yields
Q μ = 0 on ∂ C θ . (22) Q_{\mu}=0\qquad\text{on }\partial\mathcal{C}_{\theta}. \tag*{(22)} Q μ = 0 on ∂ C θ . (22)
Lemma 2.3 (Evolution of the speed ratio). Let
L = Θ S i j ∇ i ∇ j . \mathcal{L}=\Theta S^{ij}\nabla_i\nabla_j. L = Θ S ij ∇ i ∇ j . Then
( ∂ t − L ) Q = 2 Θ S i j ( log h ) i Q j + ( k + 1 − a φ ( ℓ / h ) ) Q ( Q − 1 ) . (23) (\partial_t-\mathcal{L})Q=2\Theta S^{ij}(\log h)_iQ_j+(k+1-a_{\varphi}(\ell/h))Q(Q-1). \tag*{(23)} ( ∂ t − L ) Q = 2Θ S ij ( log h ) i Q j + ( k + 1 − a φ ( ℓ / h )) Q ( Q − 1 ) . (23)
Proof. From (21) ,
W t = ( Q − 1 ) W + h ∇ 2 Q + ∇ h ⊗ ∇ Q + ∇ Q ⊗ ∇ h . W_t=(Q-1)W+h\nabla^2Q+\nabla h\otimes\nabla Q+\nabla Q\otimes\nabla h. W t = ( Q − 1 ) W + h ∇ 2 Q + ∇ h ⊗ ∇ Q + ∇ Q ⊗ ∇ h .
Using (14) ,
S t = k S ( Q − 1 ) + h S i j Q i j + 2 S i j h i Q j . S_t=kS(Q-1)+hS^{ij}Q_{ij}+2S^{ij}h_iQ_j. S t = k S ( Q − 1 ) + h S ij Q ij + 2 S ij h i Q j .
On the other hand
∂ t log Θ h = ∂ t log h φ ( ℓ / h ) f = ( 1 − a φ ( ℓ / h ) ) ( Q − 1 ) . \partial_t\log\frac{\Theta}{h}=\partial_t\log\frac{h\varphi(\ell/h)}{f}=(1-a_{\varphi}(\ell/h))(Q-1). ∂ t log h Θ = ∂ t log f h φ ( ℓ / h ) = ( 1 − a φ ( ℓ / h )) ( Q − 1 ) .
Since Q = Θ S / h Q=\Theta S/h Q = Θ S / h , combining these identities gives (23) .
Uniform C 0 C^0 C 0 and C 1 C^1 C 1 estimates
Throughout this section, we assume that the solution remains strictly convex on the time interval under consideration.
C 0 C^0 C 0 estimate
Put u = h / ℓ u=h/\ell u = h / ℓ . By (18) ,
u t = ℓ u 2 φ ( 1 / u ) f σ ‾ k ( B θ [ u ] ) − u , u μ = 0 , u_t=\frac{\ell u^2\varphi(1/u)}{f}\,\overline{\sigma}_k(B_\theta[u])-u,\qquad u_\mu=0, u t = f ℓ u 2 φ ( 1/ u ) σ k ( B θ [ u ]) − u , u μ = 0 ,
where B θ [ u ] B_\theta[u] B θ [ u ] denotes the right-hand side of (18) .
Define
g ( s ) = s k + 1 φ ( 1 / s ) . g(s)=s^{k+1}\varphi(1/s). g ( s ) = s k + 1 φ ( 1/ s ) .
Then
g ′ ( s ) g ( s ) = k + 1 − α φ ( 1 / s ) s ≤ − δ s . \frac{g'(s)}{g(s)}=\frac{k+1-\alpha_\varphi(1/s)}{s}\leq-\frac{\delta}{s}. g ( s ) g ′ ( s ) = s k + 1 − α φ ( 1/ s ) ≤ − s δ .
Hence g g g is strictly decreasing,
g ( s ) ⟶ ∞ ( s ↓ 0 ) , g ( s ) ⟶ 0 ( s ↑ ∞ ) . (24) g(s)\longrightarrow\infty\quad(s\downarrow0),\qquad g(s)\longrightarrow0\quad(s\uparrow\infty). \tag*{(24)} g ( s ) ⟶ ∞ ( s ↓ 0 ) , g ( s ) ⟶ 0 ( s ↑ ∞ ) . (24)
Proposition 3.1 (Uniform support bound). There are constants 0 < c 0 < C 0 < ∞ 0<c_0<C_0<\infty 0 < c 0 < C 0 < ∞ , depending only on the fixed data and u 0 u_0 u 0 , such that
c 0 ≤ u ( ξ , t ) ≤ C 0 c_0\leq u(\xi,t)\leq C_0 c 0 ≤ u ( ξ , t ) ≤ C 0 on every smooth strictly convex existence interval.
Proof. Let M ( t ) = max C ‾ θ u ( ⋅ , t ) M(t)=\max_{\overline{C}_\theta}u(\cdot,t) M ( t ) = max C θ u ( ⋅ , t ) . At a maximum point, including a boundary maximum, the condition (17) and the one-sided second-derivative test give ∇ u = 0 \nabla u=0 ∇ u = 0 and ∇ 2 u ≤ 0 \nabla^2u\leq0 ∇ 2 u ≤ 0 . Thus
B θ [ u ] ≤ M ( t ) σ , σ ‾ k ( B θ [ u ] ) ≤ M ( t ) k . B_\theta[u]\leq M(t)\sigma,\qquad\overline{\sigma}_k(B_\theta[u])\leq M(t)^k. B θ [ u ] ≤ M ( t ) σ , σ k ( B θ [ u ]) ≤ M ( t ) k .
Therefore
D + M ≤ M ( ℓ max f min g ( M ) − 1 ) . D^+M\leq M\left(\frac{\ell_{\max}}{f_{\min}}g(M)-1\right). D + M ≤ M ( f m i n ℓ m a x g ( M ) − 1 ) .
By (24), the bracket is negative when M M M is sufficiently large. Similarly, at a minimum point,
B θ [ u ] ≥ m ( t ) σ , σ ‾ k ( B θ [ u ] ) ≥ m ( t ) k , B_\theta[u]\geq m(t)\sigma,\qquad\overline{\sigma}_k(B_\theta[u])\geq m(t)^k, B θ [ u ] ≥ m ( t ) σ , σ k ( B θ [ u ]) ≥ m ( t ) k ,
and
D + m ≥ m ( ℓ min f max g ( m ) − 1 ) , D_+m\geq m\left(\frac{\ell_{\min}}{f_{\max}}g(m)-1\right), D + m ≥ m ( f m a x ℓ m i n g ( m ) − 1 ) ,
which is positive when m m m is sufficiently small. This gives the claimed time-independent bounds. ∎
Since ℓ \ell ℓ has fixed positive upper and lower bounds,
0 < c ≤ h ≤ C . (25) 0<c\leq h\leq C. \tag*{(25)} 0 < c ≤ h ≤ C . (25)
C 1 C^1 C 1 estimate
The strict positivity of B θ [ u ] B_\theta[u] B θ [ u ] and (18) imply, for every unit vector v v v ,
∇ 2 u ( v , v ) ≥ − C ( 1 + ∣ ∇ u ∣ ) , \nabla^2u(v,v) \ge-C(1+|\nabla u|), ∇ 2 u ( v , v ) ≥ − C ( 1 + ∣∇ u ∣ ) ,
where C C C depends only on the fixed cap geometry and the C 0 C^0 C 0 bound.
Proposition 3.2 (Uniform gradient bound). There exists C 1 C_1 C 1 such that
∣ ∇ u ∣ ≤ C 1 |\nabla u| \le C_1 ∣∇ u ∣ ≤ C 1 on every smooth strictly convex existence interval.
Proof. Fix a time and suppose ∣ ∇ u ( x 0 ) ∣ |\nabla u(x_0)| ∣∇ u ( x 0 ) ∣ is large. On the noncritical set, consider the unit gradient curve
γ ′ ( s ) = ∇ u ∣ ∇ u ∣ . \gamma'(s)=\frac{\nabla u}{|\nabla u|}. γ ′ ( s ) = ∣∇ u ∣ ∇ u . The Neumann condition makes this vector field tangent to the boundary, so its integral curve remains in C θ ‾ \overline{C_\theta} C θ . Put p ( s ) = ∣ ∇ u ( γ ( s ) ) ∣ p(s)=|\nabla u(\gamma(s))| p ( s ) = ∣∇ u ( γ ( s )) ∣ . Then
p ′ ( s ) = ∇ 2 u ( γ ′ , γ ′ ) ≥ − C ( 1 + p ( s ) ) . p'(s)=\nabla^2u(\gamma',\gamma')\ge-C(1+p(s)). p ′ ( s ) = ∇ 2 u ( γ ′ , γ ′ ) ≥ − C ( 1 + p ( s )) . For a fixed sufficiently small L 0 > 0 L_0>0 L 0 > 0 , if ∣ ∇ u ( x 0 ) ∣ |\nabla u(x_0)| ∣∇ u ( x 0 ) ∣ is large then p ( s ) ≥ ∣ ∇ u ( x 0 ) ∣ / 4 p(s)\ge|\nabla u(x_0)|/4 p ( s ) ≥ ∣∇ u ( x 0 ) ∣/4 on [ 0 , L 0 ] [0,L_0] [ 0 , L 0 ] . Hence
osc C θ u ≥ u ( γ ( L 0 ) ) − u ( γ ( 0 ) ) = ∫ 0 L 0 p ( s ) d s ≥ L 0 4 ∣ ∇ u ( x 0 ) ∣ , \operatorname{osc}_{C_\theta}u\ge u(\gamma(L_0))-u(\gamma(0))=\int_0^{L_0}p(s)\,ds\ge\frac{L_0}{4}|\nabla u(x_0)|, osc C θ u ≥ u ( γ ( L 0 )) − u ( γ ( 0 )) = ∫ 0 L 0 p ( s ) d s ≥ 4 L 0 ∣∇ u ( x 0 ) ∣ , contradicting Proposition 3.1. Thus ∣ ∇ u ∣ |\nabla u| ∣∇ u ∣ is uniformly bounded.
Consequently
∥ h ( ⋅ , t ) ∥ C 1 ( C θ ‾ ) ≤ C . \|h(\cdot,t)\|_{C^1(\overline{C_\theta})}\le C. ∥ h ( ⋅ , t ) ∥ C 1 ( C θ ) ≤ C .
Uniform control of the speed ratio
Proposition 3.3 (Two-sided Q Q Q estimate). There exist positive constants c Q , C Q c_Q,C_Q c Q , C Q such that
0 < c Q ≤ Q ( ξ , t ) ≤ C Q 0<c_Q\le Q(\xi,t)\le C_Q 0 < c Q ≤ Q ( ξ , t ) ≤ C Q on every smooth strictly convex existence interval.
Proof. At a spatial maximum of Q Q Q , the drift in (23) vanishes. If Q > 1 Q>1 Q > 1 , then
D + Q max ≤ − δ Q max ( Q max − 1 ) . D^+Q_{\max}\le-\delta Q_{\max}(Q_{\max}-1). D + Q m a x ≤ − δ Q m a x ( Q m a x − 1 ) . At a spatial minimum with Q < 1 Q<1 Q < 1 ,
D + Q min ≥ δ Q min ( 1 − Q min ) . D^+Q_{\min}\ge\delta Q_{\min}(1-Q_{\min}). D + Q m i n ≥ δ Q m i n ( 1 − Q m i n ) . The boundary condition (22) permits the same maximum-principle argument at the boundary. Hence
min { 1 , min Q ( ⋅ , 0 ) } ≤ Q ( ⋅ , t ) ≤ max { 1 , max Q ( ⋅ , 0 ) } . \min\{1,\min Q(\cdot,0)\}\le Q(\cdot,t)\le\max\{1,\max Q(\cdot,0)\}. min { 1 , min Q ( ⋅ , 0 )} ≤ Q ( ⋅ , t ) ≤ max { 1 , max Q ( ⋅ , 0 )} .
By (20) , (25) , and the compact range of ℓ / h \ell/h ℓ / h ,
0 < c S ≤ σ ‾ k ( W [ h ] ) ≤ C S . (26) 0<c_S\le\overline{\sigma}_k(W[h])\le C_S. \tag*{(26)} 0 < c S ≤ σ k ( W [ h ]) ≤ C S . (26)
The dynamic Orlicz square completion
Let
r = ℓ h , ρ = log h ℓ , q 0 = log ℓ , H = f ℓ k − 1 , K = log H . r=\frac{\ell}{h},\qquad\rho=\log\frac{h}{\ell},\qquad q_0=\log\ell,\qquad H=f\ell^{k-1},\qquad K=\log H. r = h ℓ , ρ = log ℓ h , q 0 = log ℓ , H = f ℓ k − 1 , K = log H .
Write
a = a φ ( r ) , d = r a φ ′ ( r ) , b = a − 2 , a=a_\varphi(r),\qquad d=ra_\varphi'(r),\qquad b=a-2, a = a φ ( r ) , d = r a φ ′ ( r ) , b = a − 2 ,
and
A φ f l = b + d + b 2 k + 1 . A_\varphi^{\mathrm{fl}}=b+d+\frac{b^2}{k+1}. A φ fl = b + d + k + 1 b 2 .
Under (5)–(7), A φ f l > 0 A_\varphi^{\mathrm{fl}}>0 A φ fl > 0 .
Lemma 3.4 (Dynamic Orlicz square completion). Let e e e be a unit eigenvector of W [ h ] W[h] W [ h ] . Then
1 Θ [ ( k + 1 ) Θ + Θ e e − k k + 1 Θ e 2 Θ + ( 1 − k ) Θ W ( e , e ) h ] = A φ f l ( r ) [ ρ e + b ( r ) ( k + 1 ) A φ f l ( r ) K e ] 2 + T φ , f ( r ) ( e , e ) (27) \frac{1}{\Theta}\left[(k+1)\Theta+\Theta_{ee}-\frac{k}{k+1}\frac{\Theta_e^2}{\Theta}+(1-k)\Theta\frac{W(e,e)}{h}\right]
=
A_\varphi^{\mathrm{fl}}(r)\left[\rho_e+\frac{b(r)}{(k+1)A_\varphi^{\mathrm{fl}}(r)}K_e\right]^2+\mathcal{T}_{\varphi,f}(r)(e,e) \tag*{(27)} Θ 1 [ ( k + 1 ) Θ + Θ ee − k + 1 k Θ Θ e 2 + ( 1 − k ) Θ h W ( e , e ) ] = A φ fl ( r ) [ ρ e + ( k + 1 ) A φ fl ( r ) b ( r ) K e ] 2 + T φ , f ( r ) ( e , e ) (27) + ( 3 − k − a φ ( r ) ) W ( e , e ) h , \qquad+(3-k-a_\varphi(r))\frac{W(e,e)}{h}, + ( 3 − k − a φ ( r )) h W ( e , e ) , where T φ , f \mathcal{T}_{\varphi,f} T φ , f is defined in (6) .
Proof. We give the algebra in detail because the coefficient shifts relative to the static problem are important. Put L = log Θ L=\log\Theta L = log Θ . Since
log h = ρ + q 0 , log f = K − ( k − 1 ) q 0 , r = e − ρ , \log h=\rho+q_0,\qquad\log f=K-(k-1)q_0,\qquad r=e^{-\rho}, log h = ρ + q 0 , log f = K − ( k − 1 ) q 0 , r = e − ρ ,
we have
L = 2 ρ + ( k + 1 ) q 0 + log φ ( e − ρ ) − K . (28) L=2\rho+(k+1)q_0+\log\varphi(e^{-\rho})-K. \tag*{(28)} L = 2 ρ + ( k + 1 ) q 0 + log φ ( e − ρ ) − K . (28)
Differentiating (28) in the unit direction e e e gives
L e = − b ρ e + ( k + 1 ) ( q 0 ) e − K e . (29) L_e=-b\rho_e+(k+1)(q_0)_e-K_e. \tag*{(29)} L e = − b ρ e + ( k + 1 ) ( q 0 ) e − K e . (29)
Since r e = − r ρ e r_e=-r\rho_e r e = − r ρ e , one has
b e = a φ ′ ( r ) r e = − d ρ e , b_e=a_\varphi'(r)r_e=-d\rho_e, b e = a φ ′ ( r ) r e = − d ρ e ,
and therefore
L e e = d ρ e 2 − b ρ e e + ( k + 1 ) ( q 0 ) e e − K e e . L_{ee}=d\rho_e^2-b\rho_{ee}+(k+1)(q_0)_{ee}-K_{ee}. L ee = d ρ e 2 − b ρ ee + ( k + 1 ) ( q 0 ) ee − K ee .
We next express ρ e e \rho_{ee} ρ ee in geometric variables. From W [ h ] ( e , e ) = h e e + h W[h](e,e)=h_{ee}+h W [ h ] ( e , e ) = h ee + h and W [ ℓ ] ( e , e ) = 1 W[\ell](e,e)=1 W [ ℓ ] ( e , e ) = 1 ,
( log h ) e e = W ( e , e ) h − 1 − ( ρ e + ( q 0 ) e ) 2 , (\log h)_{ee}=\frac{W(e,e)}{h}-1-(\rho_e+(q_0)_e)^2, ( log h ) ee = h W ( e , e ) − 1 − ( ρ e + ( q 0 ) e ) 2 ,
while
( q 0 ) e e = ℓ e e ℓ − ( q 0 ) e 2 = 1 ℓ − 1 − ( q 0 ) e 2 . (q_0)_{ee}=\frac{\ell_{ee}}{\ell}-(q_0)_e^2=\frac{1}{\ell}-1-(q_0)_e^2. ( q 0 ) ee = ℓ ℓ ee − ( q 0 ) e 2 = ℓ 1 − 1 − ( q 0 ) e 2 .
Subtracting yields
ρ e e = W ( e , e ) h − 1 ℓ − ρ e 2 − 2 ( q 0 ) e ρ e . (30) \rho_{ee}=\frac{W(e,e)}{h}-\frac{1}{\ell}-\rho_e^2-2(q_0)_e\rho_e. \tag*{(30)} ρ ee = h W ( e , e ) − ℓ 1 − ρ e 2 − 2 ( q 0 ) e ρ e . (30)
Convexity and curvature estimates
Because Θ e e / Θ = L e e + L e 2 \Theta_{ee}/\Theta=L_{ee}+L_e^2 Θ ee /Θ = L ee + L e 2 , substituting (29)–(30) into the left-hand side of (27) and collecting terms gives
( k + 1 ) + L e e + 1 k + 1 L e 2 + ( 1 − k ) W ( e , e ) h = ( b + d + b 2 k + 1 ) ρ e 2 + 2 b k + 1 ρ e K e (k+1)+L_{ee}+\frac{1}{k+1}L_e^2+(1-k)\frac{W(e,e)}{h}
=
\left(b+d+\frac{b^2}{k+1}\right)\rho_e^2+\frac{2b}{k+1}\rho_eK_e ( k + 1 ) + L ee + k + 1 1 L e 2 + ( 1 − k ) h W ( e , e ) = ( b + d + k + 1 b 2 ) ρ e 2 + k + 1 2 b ρ e K e
− K e e − 2 ( q 0 ) e K e + 1 k + 1 K e 2 + b + k + 1 ℓ + ( 3 − k − a ) W ( e , e ) h . (31) {}-K_{ee}-2(q_0)_eK_e+\frac{1}{k+1}K_e^2+\frac{b+k+1}{\ell}+(3-k-a)\frac{W(e,e)}{h}. \tag*{(31)} − K ee − 2 ( q 0 ) e K e + k + 1 1 K e 2 + ℓ b + k + 1 + ( 3 − k − a ) h W ( e , e ) . (31)
The cancellation of the ρ e ( q 0 ) e \rho_e(q_0)_e ρ e ( q 0 ) e and ( q 0 ) e 2 (q_0)_e^2 ( q 0 ) e 2 terms is the reason for introducing H = f ℓ k − 1 H=f\ell^{k-1} H = f ℓ k − 1 . Completing the square in ρ e \rho_e ρ e gives
A φ f l ρ e 2 + 2 b k + 1 ρ e K e + 1 k + 1 K e 2 = A φ f l ( ρ e + b ( k + 1 ) A φ f l K e ) 2 + b + d ( k + 1 ) A φ f l K e 2 . (32) \begin{aligned}
A_\varphi^{\mathrm{fl}}\rho_e^2+\frac{2b}{k+1}\rho_eK_e+\frac{1}{k+1}K_e^2 \\
&=A_\varphi^{\mathrm{fl}}\left(\rho_e+\frac{b}{(k+1)A_\varphi^{\mathrm{fl}}}K_e\right)^2+\frac{b+d}{(k+1)A_\varphi^{\mathrm{fl}}}K_e^2. \tag*{(32)}
\end{aligned} A φ fl ρ e 2 + k + 1 2 b ρ e K e + k + 1 1 K e 2 = A φ fl ( ρ e + ( k + 1 ) A φ fl b K e ) 2 + ( k + 1 ) A φ fl b + d K e 2 . (32)
Combining (31) and (32) gives exactly (27) and the tensor formula (6). For later use we record an equivalent factorization. For a fixed parameter s > 0 s>0 s > 0 , set
c φ f l ( s ) = b ( s ) 2 b ( s ) + d ( s ) , α φ f l ( s ) = 1 k + 1 + c φ f l ( s ) . c_\varphi^{\mathrm{fl}}(s)=\frac{b(s)^2}{b(s)+d(s)},\qquad
\alpha_\varphi^{\mathrm{fl}}(s)=\frac{1}{k+1+c_\varphi^{\mathrm{fl}}(s)}. c φ fl ( s ) = b ( s ) + d ( s ) b ( s ) 2 , α φ fl ( s ) = k + 1 + c φ fl ( s ) 1 .
With s s s held fixed, α φ f l ( s ) \alpha_\varphi^{\mathrm{fl}}(s) α φ fl ( s ) is constant under spatial differentiation, and
∇ 2 ( ℓ H − α φ f l ( s ) ) + ℓ H − α φ f l ( s ) σ ℓ H − α φ f l ( s ) = 1 ℓ σ − α φ f l ( s ) ∇ 2 K − 2 α φ f l ( s ) d q 0 ⊙ d K + ( α φ f l ( s ) ) 2 d K ⊗ d K . \frac{\nabla^2(\ell H^{-\alpha_\varphi^{\mathrm{fl}}(s)})+\ell H^{-\alpha_\varphi^{\mathrm{fl}}(s)}\sigma}{\ell H^{-\alpha_\varphi^{\mathrm{fl}}(s)}}
=\frac{1}{\ell}\sigma-\alpha_\varphi^{\mathrm{fl}}(s)\nabla^2K-2\alpha_\varphi^{\mathrm{fl}}(s)dq_0\mathbin{\odot}dK+\left(\alpha_\varphi^{\mathrm{fl}}(s)\right)^2dK\otimes dK. ℓ H − α φ fl ( s ) ∇ 2 ( ℓ H − α φ fl ( s ) ) + ℓ H − α φ fl ( s ) σ = ℓ 1 σ − α φ fl ( s ) ∇ 2 K − 2 α φ fl ( s ) d q 0 ⊙ d K + ( α φ fl ( s ) ) 2 d K ⊗ d K .
Using
α φ f l ( s ) = b ( s ) + d ( s ) ( k + 1 ) ( b ( s ) + d ( s ) ) + b ( s ) 2 = b ( s ) + d ( s ) ( k + 1 ) A φ f l ( s ) \alpha_\varphi^{\mathrm{fl}}(s)=\frac{b(s)+d(s)}{(k+1)(b(s)+d(s))+b(s)^2}
=\frac{b(s)+d(s)}{(k+1)A_\varphi^{\mathrm{fl}}(s)} α φ fl ( s ) = ( k + 1 ) ( b ( s ) + d ( s )) + b ( s ) 2 b ( s ) + d ( s ) = ( k + 1 ) A φ fl ( s ) b ( s ) + d ( s )
and
b ( s ) − c φ f l ( s ) = b ( s ) d ( s ) b ( s ) + d ( s ) , b(s)-c_\varphi^{\mathrm{fl}}(s)=\frac{b(s)d(s)}{b(s)+d(s)}, b ( s ) − c φ fl ( s ) = b ( s ) + d ( s ) b ( s ) d ( s ) ,
we obtain
T φ , f ( s ) = ( k + 1 + c φ f l ( s ) ) ∇ 2 ( ℓ H − α φ f l ( s ) ) + ℓ H − α φ f l ( s ) σ ℓ H − α φ f l ( s ) + b ( s ) d ( s ) ℓ [ b ( s ) + d ( s ) ] σ . \mathcal{T}_{\varphi,f}(s)=\left(k+1+c_\varphi^{\mathrm{fl}}(s)\right)
\frac{\nabla^2(\ell H^{-\alpha_\varphi^{\mathrm{fl}}(s)})+\ell H^{-\alpha_\varphi^{\mathrm{fl}}(s)}\sigma}{\ell H^{-\alpha_\varphi^{\mathrm{fl}}(s)}}
+\frac{b(s)d(s)}{\ell[b(s)+d(s)]}\sigma. T φ , f ( s ) = ( k + 1 + c φ fl ( s ) ) ℓ H − α φ fl ( s ) ∇ 2 ( ℓ H − α φ fl ( s ) ) + ℓ H − α φ fl ( s ) σ + ℓ [ b ( s ) + d ( s )] b ( s ) d ( s ) σ .
After evaluating this identity at s = r = ℓ / h s=r=\ell/h s = r = ℓ / h , we see that the condition d ≥ 0 d\geq0 d ≥ 0 and a strict spherical-convexity lower bound for the first term imply (8).
Remark 3.5 (Power-law check). If φ ( s ) = s p \varphi(s)=s^p φ ( s ) = s p , then a φ = p a_\varphi=p a φ = p , d = 0 d=0 d = 0 , and
A φ f l = ( p − 2 ) ( p + k − 1 ) k + 1 . A_\varphi^{\mathrm{fl}}=\frac{(p-2)(p+k-1)}{k+1}. A φ fl = k + 1 ( p − 2 ) ( p + k − 1 ) .
A direct computation gives
T φ , f = ( p + k − 1 ) ∇ 2 ( ℓ H − 1 / ( p + k − 1 ) ) + ℓ H − 1 / ( p + k − 1 ) σ ℓ H − 1 / ( p + k − 1 ) . \mathcal{T}_{\varphi,f}=(p+k-1)\frac{\nabla^2(\ell H^{-1/(p+k-1)})+\ell H^{-1/(p+k-1)}\sigma}{\ell H^{-1/(p+k-1)}}. T φ , f = ( p + k − 1 ) ℓ H − 1/ ( p + k − 1 ) ∇ 2 ( ℓ H − 1/ ( p + k − 1 ) ) + ℓ H − 1/ ( p + k − 1 ) σ .
Thus the dynamic structural tensor recovers exactly the exponent 1 / ( p + k − 1 ) 1/(p+k-1) 1/ ( p + k − 1 ) appearing in the closed flow of Ivaki [12 ] .
Convexity and curvature estimates
Preservation of strict convexity
In this section the strict inequality (8) is not needed; it suffices that
T φ , f ( s ) ≥ 0 ( s > 0 ) . \mathcal{T}_{\varphi,f}(s) \ge0 \qquad(s > 0). T φ , f ( s ) ≥ 0 ( s > 0 ) .
Define
q = h λ min ( W [ h ] ) . q = h\lambda_{\min}(W[h]). q = h λ m i n ( W [ h ]) .
The smallest eigenvalue is only Lipschitz in general, so the correct formulation is in the viscosity sense, following [12 , 3 ] .
Lemma 4.1 (Viscosity inequality for the minimum radius). Let h h h be a smooth solution with h > 0 h > 0 h > 0 and W [ h ] > 0 W[h] > 0 W [ h ] > 0 near an interior space-time point ( x 0 , t 0 ) (x_0,t_0) ( x 0 , t 0 ) , where t 0 > 0 t_0 > 0 t 0 > 0 . Let ξ > 0 \xi> 0 ξ > 0 be a smooth parabolic lower support of q = h λ min ( W ) q = h\lambda_{\min}(W) q = h λ m i n ( W ) there: ξ ≤ q \xi\le q ξ ≤ q in a neighborhood of x 0 x_0 x 0 for times immediately preceding t 0 t_0 t 0 , with equality at ( x 0 , t 0 ) (x_0,t_0) ( x 0 , t 0 ) . If e e e is a unit minimum eigenvector at the contact point, then
( ∂ t − L ) ξ + 2 Θ S i j ξ i ∇ j log ξ h + [ ( a φ ( r ) + k − 3 ) Q + 2 ] ξ ≥ h 2 Q T φ , f ( r ) ( e , e ) . (33) (\partial_t-\mathcal{L})\xi+2\Theta S^{ij}\xi_i\nabla_j\log\frac{\xi}{h}
+\left[(a_\varphi(r)+k-3)Q+2\right]\xi
\ge h^2Q\mathcal{T}_{\varphi,f}(r)(e,e). \tag*{(33)} ( ∂ t − L ) ξ + 2Θ S ij ξ i ∇ j log h ξ + [ ( a φ ( r ) + k − 3 ) Q + 2 ] ξ ≥ h 2 Q T φ , f ( r ) ( e , e ) . (33)
Proof. All quantities below are evaluated at the interior contact point ( x 0 , t 0 ) (x_0,t_0) ( x 0 , t 0 ) . Choose an orthonormal frame which is geodesic at x 0 x_0 x 0 and diagonalizes W W W , and write
W i j = λ i δ i j , λ : = λ 1 = ⋯ = λ m < λ m + 1 ≤ ⋯ ≤ λ n , e = e 1 . W_{ij}=\lambda_i\delta_{ij},\qquad\lambda:=\lambda_1=\cdots=\lambda_m<\lambda_{m+1}\le\cdots\le\lambda_n,\qquad e=e_1. W ij = λ i δ ij , λ := λ 1 = ⋯ = λ m < λ m + 1 ≤ ⋯ ≤ λ n , e = e 1 .
Put
A i j = Θ S i j , D i = ξ i − λ h i h , η i j = ∇ 1 W i j . A^{ij}=\Theta S^{ij},\qquad D_i=\frac{\xi_i-\lambda h_i}{h},\qquad\eta_{ij}=\nabla_1W_{ij}. A ij = Θ S ij , D i = h ξ i − λ h i , η ij = ∇ 1 W ij .
For each unit vector in the minimum eigenspace, h W ( e , e ) − ξ hW(e,e)-\xi hW ( e , e ) − ξ has a spatial minimum of zero. Differentiating and polarizing within that eigenspace gives
∇ i W a b = δ a b D i ( 1 ≤ a , b ≤ m ) . (34) \nabla_iW_{ab}=\delta_{ab}D_i\qquad(1\le a,b\le m). \tag*{(34)} ∇ i W ab = δ ab D i ( 1 ≤ a , b ≤ m ) . (34)
The second-derivative test, allowing the test eigenvector to vary in the complementary eigenspaces, gives
ξ i i ≤ ( h W 11 ) ; i i − 2 h ∑ p > m ( ∇ i W 1 p ) 2 λ p − λ . (35) \xi_{ii}\le(hW_{11})_{;ii}-2h\sum_{p>m}\frac{(\nabla_iW_{1p})^2}{\lambda_p-\lambda}. \tag*{(35)} ξ ii ≤ ( h W 11 ) ; ii − 2 h p > m ∑ λ p − λ ( ∇ i W 1 p ) 2 . (35)
Indeed, one obtains this formula by choosing the p p p -component of the first derivative of the unit test vector to be − ∇ i W 1 p / ( λ p − λ ) -\nabla_iW_{1p}/(\lambda_p-\lambda) − ∇ i W 1 p / ( λ p − λ ) and completing the resulting quadratic expression. No division by a gap inside the minimum eigenspace is needed. The parabolic lower-support condition gives ξ t ≥ ∂ t ( h W 11 ) \xi_t\ge\partial_t(hW_{11}) ξ t ≥ ∂ t ( h W 11 ) : the difference is nonnegative for past times and vanishes at the terminal contact time.
Since W W W is Codazzi,
η i p = ∇ 1 W i p = ∇ i W 1 p , η i p = δ 1 p D i ( p ≤ m ) . \eta_{ip}=\nabla_1W_{ip}=\nabla_iW_{1p},\qquad\eta_{ip}=\delta_{1p}D_i\qquad(p\le m). η i p = ∇ 1 W i p = ∇ i W 1 p , η i p = δ 1 p D i ( p ≤ m ) .
At a diagonal matrix, write S i = S i i > 0 S_i=S^{ii}>0 S i = S ii > 0 . The flow, Euler homogeneity, and the unit-sphere commutator yield
( ∂ t − L ) h = ( 1 − k ) Θ S − h + Θ h ∑ i S i (\partial_t-\mathcal{L})h=(1-k)\Theta S-h+\Theta h\sum_i S_i ( ∂ t − L ) h = ( 1 − k ) Θ S − h + Θ h i ∑ S i
( ∂ t − L ) W 11 = S Θ 11 + 2 Θ 1 S 1 d i r + Θ S i j , r s η i j η r s + ( k + 1 ) Θ S − ( Θ ∑ i S i + 1 ) λ (\partial_t-\mathcal{L})W_{11}=S\Theta_{11}+2\Theta_1S^{\mathrm{dir}}_1+\Theta S^{ij,rs}\eta_{ij}\eta_{rs}+(k+1)\Theta S-\left(\Theta\sum_i S_i+1\right)\lambda ( ∂ t − L ) W 11 = S Θ 11 + 2 Θ 1 S 1 dir + Θ S ij , r s η ij η r s + ( k + 1 ) Θ S − ( Θ i ∑ S i + 1 ) λ
where S 1 d i r : = ∇ 1 S = S i j η i j S^{\mathrm{dir}}_1:=\nabla_1S=S^{ij}\eta_{ij} S 1 dir := ∇ 1 S = S ij η ij ; this notation distinguishes the directional derivative from S 1 = S 11 S_1=S^{11} S 1 = S 11 . Using (35) in the product rule, and using Θ S = h Q \Theta S=hQ Θ S = h Q and ξ = h λ \xi=h\lambda ξ = hλ , gives
( ∂ t − L ) ξ + 2 Θ S i j ξ i ∇ j log ξ h + [ ( k − 1 ) Q + 2 ] ξ ≥ h S [ ( k + 1 ) Θ + Θ 11 ] + 2 h Θ 1 S 1 d i r + h Θ [ S i j , r s η i j η r s + 2 ∑ i S i ( D i 2 λ + ∑ p > m η i p 2 λ p − λ ) ] . (36) \begin{aligned}
(\partial_t-\mathcal{L})\xi+2\Theta S^{ij}\xi_i\nabla_j\log\frac{\xi}{h}+\left[(k-1)Q+2\right]\xi
&\ge hS\left[(k+1)\Theta+\Theta_{11}\right]+2h\Theta_1S^{\mathrm{dir}}_1\\
&\quad+h\Theta\left[S^{ij,rs}\eta_{ij}\eta_{rs}+2\sum_iS_i\left(\frac{D_i^2}{\lambda}+\sum_{p>m}\frac{\eta_{ip}^2}{\lambda_p-\lambda}\right)\right]. \tag*{(36)}
\end{aligned} ( ∂ t − L ) ξ + 2Θ S ij ξ i ∇ j log h ξ + [ ( k − 1 ) Q + 2 ] ξ ≥ h S [ ( k + 1 ) Θ + Θ 11 ] + 2 h Θ 1 S 1 dir + h Θ [ S ij , r s η ij η r s + 2 i ∑ S i ( λ D i 2 + p > m ∑ λ p − λ η i p 2 ) ] . (36)
In particular, the drift is exact: since ∇ i log ( ξ / h ) = D i / λ \nabla_i\log(\xi/h)=D_i/\lambda ∇ i log ( ξ / h ) = D i / λ , adding it cancels the product term − 2 Θ S i j h i D j -2\Theta S^{ij}h_iD_j − 2Θ S ij h i D j and leaves 2 h Θ ∑ i S i D i 2 / λ 2h\Theta\sum_iS_iD_i^2/\lambda 2 h Θ ∑ i S i D i 2 / λ .
We record explicitly the inverse-concavity inequality used here. For W > 0 W>0 W > 0 and every symmetric matrix η \eta η ,
( S i j , r s + 2 S i r ( W − 1 ) j s ) η i j η r s ≥ k + 1 k S ( S i j η i j ) 2 . (37) \left(S^{ij,rs}+2S^{ir}(W^{-1})^{js}\right)\eta_{ij}\eta_{rs}\ge\frac{k+1}{kS}\left(S^{ij}\eta_{ij}\right)^2. \tag*{(37)} ( S ij , r s + 2 S i r ( W − 1 ) j s ) η ij η r s ≥ k S k + 1 ( S ij η ij ) 2 . (37)
Here is a direct verification, including its domain. Set F = S 1 / k F=S^{1/k} F = S 1/ k and F ∗ ( B ) = 1 / F ( B − 1 ) F_*(B)=1/F(B^{-1}) F ∗ ( B ) = 1/ F ( B − 1 ) on the positive definite cone. On positive coordinate vectors x x x , let I \mathcal{I} I be the collection of k k k -element subsets of { 1 , … , n } \{1,\ldots,n\} { 1 , … , n } and put
g I ( x ) = ( ∏ i ∈ I x i ) 1 / k , P ( y ) = ( ∑ I ∈ I y I − k ) − 1 / k . g_I(x)=\left(\prod_{i\in I}x_i\right)^{1/k},\qquad P(y)=\left(\sum_{I\in\mathcal{I}}y_I^{-k}\right)^{-1/k}. g I ( x ) = ( i ∈ I ∏ x i ) 1/ k , P ( y ) = ( I ∈ I ∑ y I − k ) − 1/ k .
Each g I g_I g I is positive and concave. The function P P P is increasing in each variable and concave: writing Z = ∑ I y I − k Z=\sum_I y_I^{-k} Z = ∑ I y I − k , one has
D 2 P ( y ) [ v , v ] = ( k + 1 ) Z − 1 / k − 2 [ ( ∑ I y I − k − 1 v I ) 2 − Z ∑ I y I − k − 2 v I 2 ] ≤ 0 D^2P(y)[v,v]=(k+1)Z^{-1/k-2}\left[\left(\sum_I y_I^{-k-1}v_I\right)^2-Z\sum_I y_I^{-k-2}v_I^2\right]\le0 D 2 P ( y ) [ v , v ] = ( k + 1 ) Z − 1/ k − 2 ( I ∑ y I − k − 1 v I ) 2 − Z I ∑ y I − k − 2 v I 2 ≤ 0
by Cauchy–Schwarz. Consequently
F ∗ ( x ) = ( n k ) 1 / k P ( ( g I ( x ) ) I ∈ I ) F_*(x)=\binom{n}{k}^{1/k}P\bigl((g_I(x))_{I\in\mathcal{I}}\bigr) F ∗ ( x ) = ( k n ) 1/ k P ( ( g I ( x ) ) I ∈ I )
is a symmetric concave function of the positive eigenvalues. It is also concave as a spectral function on positive definite matrices. For completeness, at distinct eigenvalues its second derivative is
D 2 F ∗ ( B ) [ Z , Z ] = ∑ i , j ( F ∗ ) i j Z i i Z j j + 2 ∑ i < j ( F ∗ ) i − ( F ∗ ) j x i − x j Z i j 2 ≤ 0 ; D^2F_*(B)[Z,Z]=\sum_{i,j}(F_*)_{ij}Z_{ii}Z_{jj}+2\sum_{i<j}\frac{(F_*)_i-(F_*)_j}{x_i-x_j}Z_{ij}^2\le0; D 2 F ∗ ( B ) [ Z , Z ] = i , j ∑ ( F ∗ ) ij Z ii Z j j + 2 i < j ∑ x i − x j ( F ∗ ) i − ( F ∗ ) j Z ij 2 ≤ 0 ;
the first term is nonpositive by scalar concavity, and the quotients are nonpositive by symmetry and concavity. The case of repeated eigenvalues follows by continuity.
Now differentiate F ∗ F_\ast F ∗ along B ( s ) = W − 1 − s W − 1 η W − 1 B(s)=W^{-1}-sW^{-1}\eta W^{-1} B ( s ) = W − 1 − s W − 1 η W − 1 , which remains positive definite for small ∣ s ∣ |s| ∣ s ∣ . If A ( s ) = B ( s ) − 1 A(s)=B(s)^{-1} A ( s ) = B ( s ) − 1 , then A ′ ( 0 ) = η A'(0)=\eta A ′ ( 0 ) = η and A ′ ′ ( 0 ) = 2 η W − 1 η A''(0)=2\eta W^{-1}\eta A ′′ ( 0 ) = 2 η W − 1 η . The inequality d 2 d s 2 F ∗ ( B ( s ) ) ∣ s = 0 ≤ 0 \left.\frac{d^2}{ds^2}F_\ast(B(s))\right|_{s=0}\leq0 d s 2 d 2 F ∗ ( B ( s )) s = 0 ≤ 0 is therefore
F i j , r s η i j η r s + 2 F i r ( W − 1 ) j s η i j η r s ≥ 2 F ( F i j η i j ) 2 F^{ij,rs}\eta_{ij}\eta_{rs}+2F^{ir}(W^{-1})^{js}\eta_{ij}\eta_{rs}\geq\frac{2}{F}(F^{ij}\eta_{ij})^2 F ij , r s η ij η r s + 2 F i r ( W − 1 ) j s η ij η r s ≥ F 2 ( F ij η ij ) 2
Substituting F = S 1 / k F=S^{1/k} F = S 1/ k and multiplying by k S 1 − 1 / k kS^{1-1/k} k S 1 − 1/ k gives (37) , with the coefficient 2 / k − ( 1 / k − 1 ) = ( k + 1 ) / k 2/k-(1/k-1)=(k+1)/k 2/ k − ( 1/ k − 1 ) = ( k + 1 ) / k .
The bracket in (36) is bounded below by the left-hand side of (37) . Indeed, its difference from that expression is exactly
2 ∑ i S i ∑ p > m ( 1 λ p − λ − 1 λ p ) η i p 2 ≥ 0 2\sum_i S_i\sum_{p>m}\left(\frac{1}{\lambda_p-\lambda}-\frac{1}{\lambda_p}\right)\eta_{ip}^2\geq0 2 i ∑ S i p > m ∑ ( λ p − λ 1 − λ p 1 ) η i p 2 ≥ 0
because (34) supplies all the terms with p ≤ m p\leq m p ≤ m and 0 < λ < λ p 0<\lambda<\lambda_p 0 < λ < λ p . Thus
h Θ k + 1 k S ( S 1 dir ) 2 + 2 h Θ 1 S 1 dir ≥ − h S k k + 1 Θ 1 2 Θ . h\Theta\frac{k+1}{kS}(S_1^{\operatorname{dir}})^2+2h\Theta_1S_1^{\operatorname{dir}}\geq-hS\frac{k}{k+1}\frac{\Theta_1^2}{\Theta}. h Θ k S k + 1 ( S 1 dir ) 2 + 2 h Θ 1 S 1 dir ≥ − h S k + 1 k Θ Θ 1 2 .
Inserting this into (36) proves
( ∂ t − L ) ξ + 2 Θ S i j ξ i ∇ j log ξ h + [ ( k − 1 ) Q + 2 ] ξ ≥ h S [ ( k + 1 ) Θ + Θ e e − k k + 1 Θ e 2 Θ ] . (38) (\partial_t-\mathcal{L})\xi+2\Theta S^{ij}\xi_i\nabla_j\log\frac{\xi}{h}+[(k-1)Q+2]\xi
\geq hS\left[(k+1)\Theta+\Theta_{ee}-\frac{k}{k+1}\frac{\Theta_e^2}{\Theta}\right]. \tag*{(38)} ( ∂ t − L ) ξ + 2Θ S ij ξ i ∇ j log h ξ + [( k − 1 ) Q + 2 ] ξ ≥ h S [ ( k + 1 ) Θ + Θ ee − k + 1 k Θ Θ e 2 ] . (38)
We now insert the Orlicz coefficient. Lemma 3.4 implies
1 Θ [ ( k + 1 ) Θ + Θ e e − k k + 1 Θ e 2 Θ ] = A φ f l [ ρ e + b ( k + 1 ) A φ f l K e ] 2 + T φ , f ( e , e ) + ( 2 − a φ ( r ) ) W ( e , e ) h . (39) \frac{1}{\Theta}\left[(k+1)\Theta+\Theta_{ee}-\frac{k}{k+1}\frac{\Theta_e^2}{\Theta}\right]
=A_\varphi^{\mathrm{fl}}\left[\rho_e+\frac{b}{(k+1)A_\varphi^{\mathrm{fl}}}K_e\right]^2+\mathcal{T}_{\varphi,f}(e,e)+(2-a_\varphi(r))\frac{W(e,e)}{h}. \tag*{(39)} Θ 1 [ ( k + 1 ) Θ + Θ ee − k + 1 k Θ Θ e 2 ] = A φ fl [ ρ e + ( k + 1 ) A φ fl b K e ] 2 + T φ , f ( e , e ) + ( 2 − a φ ( r )) h W ( e , e ) . (39)
Indeed, (39) is obtained from (27) by subtracting ( 1 − k ) W ( e , e ) / h (1-k)W(e,e)/h ( 1 − k ) W ( e , e ) / h from both sides. At the contact point
h W ( e , e ) = ξ , Θ S = h Q . hW(e,e)=\xi,\qquad\Theta S=hQ. hW ( e , e ) = ξ , Θ S = h Q .
Multiplying (39) by h S Θ = h 2 Q hS\Theta=h^2Q h S Θ = h 2 Q , discarding the nonnegative square, and moving
( 2 − a φ ( r ) ) Q ξ (2-a_\varphi(r))Q\xi ( 2 − a φ ( r )) Q ξ
to the left of (38) yields
( ∂ t − L ) ξ + 2 Θ S i j ξ i ∇ j log ξ h + [ ( a φ ( r ) + k − 3 ) Q + 2 ] ξ ≥ h 2 Q T φ , f ( r ) ( e , e ) , (\partial_t-\mathcal{L})\xi+2\Theta S^{ij}\xi_i\nabla_j\log\frac{\xi}{h}+[(a_\varphi(r)+k-3)Q+2]\xi\geq h^2Q\mathcal{T}_{\varphi,f}(r)(e,e), ( ∂ t − L ) ξ + 2Θ S ij ξ i ∇ j log h ξ + [( a φ ( r ) + k − 3 ) Q + 2 ] ξ ≥ h 2 Q T φ , f ( r ) ( e , e ) ,
which is (33) .
Lemma 4.2 (Boundary derivative of a minimum-radius support). Assume 0 < θ ≤ π 2 , W > 0 , H μ ≥ 0. 0 < \theta\le\frac{\pi}{2}, \qquad W > 0, \qquad H_{\mu} \ge0. 0 < θ ≤ 2 π , W > 0 , H μ ≥ 0.
At a boundary point, choose a unit minimum eigendirection adapted to the boundary splitting: if the minimum eigenspace contains a tangential direction, choose such a direction; otherwise choose e = μ e=\mu e = μ . Extend e e e smoothly near the point and set v = h W ( e , e ) v=hW(e,e) v = hW ( e , e ) . Then v v v is a smooth local upper support of q = h λ min ( W ) q=h\lambda_{\min}(W) q = h λ m i n ( W ) and
v μ + ( n − 1 ) cot θ v ≥ 0. v_{\mu}+(n-1)\cot\theta\,v\ge0. v μ + ( n − 1 ) cot θ v ≥ 0.
Proof. At the boundary choose an orthonormal frame e 1 , … , e n − 1 , e n = μ e_1,\ldots,e_{n-1},e_n=\mu e 1 , … , e n − 1 , e n = μ which diagonalizes the tangential block of W W W . The mixed components vanish by (15) .
If the minimum eigenvalue is tangential, say λ α \lambda_{\alpha} λ α , then
v μ = h μ λ α + h W α α ; μ = h cot θ λ n ≥ 0. v_{\mu}=h_{\mu}\lambda_{\alpha}+hW_{\alpha\alpha;\mu}=h\cot\theta\,\lambda_n\ge0. v μ = h μ λ α + h W α α ; μ = h cot θ λ n ≥ 0.
For 0 < θ < π / 2 0<\theta<\pi/2 0 < θ < π /2 the sign is strict, while at θ = π / 2 \theta=\pi/2 θ = π /2 it is zero. Suppose now that the minimum eigenvalue is the normal one λ n \lambda_n λ n . If k = 1 k=1 k = 1 , then Q μ = 0 Q_{\mu}=0 Q μ = 0 and ( ℓ / h ) μ = 0 (\ell/h)_{\mu}=0 ( ℓ / h ) μ = 0 give
( σ 1 ( W ) ) μ σ 1 ( W ) = f μ f − cot θ . \frac{\left(\sigma_1(W)\right)_{\mu}}{\sigma_1(W)}=\frac{f_{\mu}}{f}-\cot\theta. σ 1 ( W ) ( σ 1 ( W ) ) μ = f f μ − cot θ .
Using
σ 1 ( W ) = ∑ α < n λ α + λ n , ( λ α ) μ = cot θ ( λ n − λ α ) , \sigma_1(W)=\sum_{\alpha<n}\lambda_{\alpha}+\lambda_n,\qquad
(\lambda_{\alpha})_{\mu}=\cot\theta(\lambda_n-\lambda_{\alpha}), σ 1 ( W ) = α < n ∑ λ α + λ n , ( λ α ) μ = cot θ ( λ n − λ α ) ,
one obtains directly
v μ v = ( 1 + ∑ α < n λ α λ n ) f μ f − ( n − 1 ) cot θ . \frac{v_{\mu}}{v}=\left(1+\frac{\sum_{\alpha<n}\lambda_{\alpha}}{\lambda_n}\right)\frac{f_{\mu}}{f}-(n-1)\cot\theta. v v μ = ( 1 + λ n ∑ α < n λ α ) f f μ − ( n − 1 ) cot θ .
Here H = f H=f H = f , so H μ ≥ 0 H_{\mu}\ge0 H μ ≥ 0 gives
v μ v ≥ − ( n − 1 ) cot θ . \frac{v_{\mu}}{v}\ge-(n-1)\cot\theta. v v μ ≥ − ( n − 1 ) cot θ .
Assume next that 2 ≤ k ≤ n − 1 2\le k\le n-1 2 ≤ k ≤ n − 1 and write
s j = σ j ( λ 1 , … , λ n − 1 ) . s_j=\sigma_j(\lambda_1,\ldots,\lambda_{n-1}). s j = σ j ( λ 1 , … , λ n − 1 ) .
Since Q μ = 0 Q_{\mu}=0 Q μ = 0 and ( ℓ / h ) μ = 0 (\ell/h)_{\mu}=0 ( ℓ / h ) μ = 0 ,
( σ k ( W ) ) μ σ k ( W ) = f μ f − cot θ . \frac{\left(\sigma_k(W)\right)_{\mu}}{\sigma_k(W)}=\frac{f_{\mu}}{f}-\cot\theta. σ k ( W ) ( σ k ( W ) ) μ = f f μ − cot θ .
Using
σ k ( W ) = s k + λ n s k − 1 , \sigma_k(W)=s_k+\lambda_n s_{k-1}, σ k ( W ) = s k + λ n s k − 1 ,
the boundary identity (16) , and the identities
∑ α < n σ k − 1 ( λ ′ ∣ α ) = ( n − k ) s k − 1 , \sum_{\alpha<n}\sigma_{k-1}(\lambda'|\alpha)=(n-k)s_{k-1}, α < n ∑ σ k − 1 ( λ ′ ∣ α ) = ( n − k ) s k − 1 ,
∑ α < n λ α σ k − 1 ( λ ′ ∣ α ) = k s k , \sum_{\alpha<n}\lambda_{\alpha}\sigma_{k-1}(\lambda'|\alpha)=ks_k, α < n ∑ λ α σ k − 1 ( λ ′ ∣ α ) = k s k ,
∑ α < n σ k − 2 ( λ ′ ∣ α ) = ( n − k + 1 ) s k − 2 , \sum_{\alpha<n}\sigma_{k-2}(\lambda'|\alpha)=(n-k+1)s_{k-2}, α < n ∑ σ k − 2 ( λ ′ ∣ α ) = ( n − k + 1 ) s k − 2 ,
∑ α < n λ α σ k − 2 ( λ ′ ∣ α ) = ( k − 1 ) s k − 1 . \sum_{\alpha<n}\lambda_{\alpha}\sigma_{k-2}(\lambda'|\alpha)=(k-1)s_{k-1}. α < n ∑ λ α σ k − 2 ( λ ′ ∣ α ) = ( k − 1 ) s k − 1 .
we may calculate as follows. Put c θ = cot θ c_\theta=\cot\theta c θ = cot θ and r n = λ n r_n=\lambda_n r n = λ n . The tangential derivative formula gives
( s j ) μ = c θ [ r n ( n − j ) s j − 1 − j s j ] ( 1 ≤ j ≤ n − 1 ) , (s_j)_\mu=c_\theta\left[r_n(n-j)s_{j-1}-js_j\right]\qquad(1\le j\le n-1), ( s j ) μ = c θ [ r n ( n − j ) s j − 1 − j s j ] ( 1 ≤ j ≤ n − 1 ) ,
where s 0 = 1 s_0=1 s 0 = 1 . Hence
( σ k ( W ) ) μ = ( r n ) μ s k − 1 + c θ [ ( n − 2 k + 1 ) r n s k − 1 − k s k + ( n − k + 1 ) r n 2 s k − 2 ] . (\sigma_k(W))_\mu=(r_n)_\mu s_{k-1}+c_\theta\left[(n-2k+1)r_ns_{k-1}-ks_k+(n-k+1)r_n^2s_{k-2}\right]. ( σ k ( W ) ) μ = ( r n ) μ s k − 1 + c θ [ ( n − 2 k + 1 ) r n s k − 1 − k s k + ( n − k + 1 ) r n 2 s k − 2 ] .
Equating this with ( s k + r n s k − 1 ) ( f μ / f − c θ ) (s_k+r_ns_{k-1})(f_\mu/f-c_\theta) ( s k + r n s k − 1 ) ( f μ / f − c θ ) and using v μ / v = c θ + ( r n ) μ / r n v_\mu/v=c_\theta+(r_n)_\mu/r_n v μ / v = c θ + ( r n ) μ / r n gives
v μ v = ( 1 + s k r n s k − 1 ) f μ f + c θ [ ( k − 1 ) s k r n s k − 1 − ( n − 2 k + 1 ) − ( n − k + 1 ) r n s k − 2 s k − 1 ] . \frac{v_\mu}{v}=\left(1+\frac{s_k}{r_ns_{k-1}}\right)\frac{f_\mu}{f}+c_\theta\left[(k-1)\frac{s_k}{r_ns_{k-1}}-(n-2k+1)-(n-k+1)r_n\frac{s_{k-2}}{s_{k-1}}\right]. v v μ = ( 1 + r n s k − 1 s k ) f f μ + c θ [ ( k − 1 ) r n s k − 1 s k − ( n − 2 k + 1 ) − ( n − k + 1 ) r n s k − 1 s k − 2 ] .
Finally, substitute f μ / f = H μ / H − ( k − 1 ) c θ f_\mu/f=H_\mu/H-(k-1)c_\theta f μ / f = H μ / H − ( k − 1 ) c θ to obtain
v μ v = ( 1 + s k λ n s k − 1 ) H μ H − ( n − k + ( n − k + 1 ) λ n s k − 2 s k − 1 ) cot θ . (40) \frac{v_\mu}{v}=\left(1+\frac{s_k}{\lambda_ns_{k-1}}\right)\frac{H_\mu}{H}-\left(n-k+(n-k+1)\lambda_n\frac{s_{k-2}}{s_{k-1}}\right)\cot\theta. \tag*{(40)} v v μ = ( 1 + λ n s k − 1 s k ) H H μ − ( n − k + ( n − k + 1 ) λ n s k − 1 s k − 2 ) cot θ . (40)
Moreover,
( k − 1 ) s k − 1 = ∑ α < n λ α σ k − 2 ( λ ′ ∣ α ) ≥ λ n ( n − k + 1 ) s k − 2 , (k-1)s_{k-1}=\sum_{\alpha<n}\lambda_\alpha\sigma_{k-2}(\lambda'|\alpha)\ge\lambda_n(n-k+1)s_{k-2}, ( k − 1 ) s k − 1 = α < n ∑ λ α σ k − 2 ( λ ′ ∣ α ) ≥ λ n ( n − k + 1 ) s k − 2 ,
so
( n − k + 1 ) λ n s k − 2 s k − 1 ≤ k − 1. (n-k+1)\lambda_n\frac{s_{k-2}}{s_{k-1}}\le k-1. ( n − k + 1 ) λ n s k − 1 s k − 2 ≤ k − 1.
Hence H μ ≥ 0 H_\mu\ge0 H μ ≥ 0 and cot θ ≥ 0 \cot\theta\ge0 cot θ ≥ 0 imply
v μ v ≥ − ( n − 1 ) cot θ . \frac{v_\mu}{v}\ge-(n-1)\cot\theta. v v μ ≥ − ( n − 1 ) cot θ .
This proves the claim in all cases.
Proof of Theorem 1.2 . Fix T > 0 T>0 T > 0 up to which the solution is smooth. Choose a smooth function β \beta β on C θ C_\theta C θ such that β ≥ 0 \beta\ge0 β ≥ 0 , β = 0 \beta=0 β = 0 on the boundary, and β μ = − 1 \beta_\mu=-1 β μ = − 1 there. Fix ζ > 0 \zeta>0 ζ > 0 and set
ψ = − ( n − 1 ) log ℓ + ζ β , ξ ( x , t ) = ε e − M t + ψ ( x ) . \psi=-(n-1)\log\ell+\zeta\beta,\qquad\xi(x,t)=\varepsilon e^{-Mt+\psi(x)}. ψ = − ( n − 1 ) log ℓ + ζ β , ξ ( x , t ) = ε e − M t + ψ ( x ) .
Choose
0 < ε < min C θ ( q ( x , 0 ) e − ψ ( x ) ) . 0<\varepsilon<\min_{C_\theta}\left(q(x,0)e^{-\psi(x)}\right). 0 < ε < C θ min ( q ( x , 0 ) e − ψ ( x ) ) .
Thus q ( x , 0 ) > ξ ( x , 0 ) q(x,0)>\xi(x,0) q ( x , 0 ) > ξ ( x , 0 ) for every x ∈ C θ ‾ x\in\overline{C_\theta} x ∈ C θ .
We work up to a possible first loss of positive definiteness of W W W . On this interval, the support bounds and smoothness on [ 0 , T ] [0,T] [ 0 , T ] give upper bounds for A i j = Θ S i j A^{ij}=\Theta S^{ij} A ij = Θ S ij , ∇ h / h \nabla h/h ∇ h / h , and
b = ( a φ ( ℓ / h ) + k − 3 ) Q + 2. b=(a_\varphi(\ell/h)+k-3)Q+2. b = ( a φ ( ℓ / h ) + k − 3 ) Q + 2.
These bounds require no positive lower bound for the eigenvalues of W W W : the derivatives S i j S^{ij} S ij are polynomials in its entries. Since ψ \psi ψ is fixed, there is a finite constant C T C_T C T , independent of M M M and ε \varepsilon ε , such that
∣ A i j ( − ψ i j + ψ i ψ j − 2 ψ i h j / h ) + b ∣ ≤ C T \left|A^{ij}\left(-\psi_{ij}+\psi_i\psi_j-2\psi_i h_j/h\right)+\mathbf{b}\right|\le C_T A ij ( − ψ ij + ψ i ψ j − 2 ψ i h j / h ) + b ≤ C T
throughout the positive-definite interval. Fix M > C T M>C_T M > C T .
Suppose that q − ξ q-\xi q − ξ has a first zero at ( x ∗ , t ∗ ) (x_*,t_*) ( x ∗ , t ∗ ) with t ∗ > 0 t_*>0 t ∗ > 0 . At this contact point, q = ξ > 0 q=\xi>0 q = ξ > 0 , so W W W is still positive definite. If the contact is interior, ξ \xi ξ is a positive parabolic lower support and Lemma 4.1 applies. Direct differentiation of ξ \xi ξ gives
( ∂ t − L ) ξ + 2 Θ S i j ξ i ∇ j log ξ h + b ξ = ξ [ − M + A i j ( − ψ i j + ψ i ψ j − 2 ψ i h j / h ) + b ] < 0. (\partial_t-\mathcal{L})\xi+2\Theta S^{ij}\xi_i\nabla_j\log\frac{\xi}{h}+b\xi
=\xi\left[-M+A^{ij}\left(-\psi_{ij}+\psi_i\psi_j-2\psi_i h_j/h\right)+b\right]<0. ( ∂ t − L ) ξ + 2Θ S ij ξ i ∇ j log h ξ + b ξ = ξ [ − M + A ij ( − ψ ij + ψ i ψ j − 2 ψ i h j / h ) + b ] < 0.
This contradicts the nonnegative right-hand side of (33) , since T φ , f ≥ 0 \mathcal{T}_{\varphi,f}\ge0 T φ , f ≥ 0 .
Suppose instead that the first contact lies on the boundary. The identity W α n = 0 W_{\alpha n}=0 W α n = 0 splits W W W into its tangential block and its normal entry. If the minimum eigenspace contains a tangential unit vector, choose one; otherwise, that eigenspace is the normal line, and we choose e = μ e=\mu e = μ . Extend e e e to a smooth unit vector field near the point and put v = h W ( e , e ) v=hW(e,e) v = hW ( e , e ) . Then v ≥ q ≥ ξ v\ge q\ge\xi v ≥ q ≥ ξ , with equality at the contact point, and therefore ( v − ξ ) μ ≤ 0 (v-\xi)_\mu\le0 ( v − ξ ) μ ≤ 0 for the outward conormal. In both cases Lemma 4.2 gives
v μ + ( n − 1 ) cot θ v ≥ 0. v_\mu+(n-1)\cot\theta\,v\ge0. v μ + ( n − 1 ) cot θ v ≥ 0.
On the other hand, ℓ μ / ℓ = cot θ \ell_\mu/\ell=\cot\theta ℓ μ / ℓ = cot θ and β μ = − 1 \beta_\mu=-1 β μ = − 1 imply
ξ μ + ( n − 1 ) cot θ ξ = − ζ ξ < 0. \xi_\mu+(n-1)\cot\theta\,\xi=-\zeta\xi<0. ξ μ + ( n − 1 ) cot θ ξ = − ζ ξ < 0.
Since v = ξ v=\xi v = ξ at the contact point, these inequalities contradict ( v − ξ ) μ ≤ 0 (v-\xi)_\mu\le0 ( v − ξ ) μ ≤ 0 .
Thus no first contact is possible. In particular, q ≥ ε e − M T + min ψ > 0 q\ge\varepsilon e^{-MT+\min\psi}>0 q ≥ ε e − M T + m i n ψ > 0 up to any proposed first loss of positive definiteness. Continuity excludes such a loss in [ 0 , T ] [0,T] [ 0 , T ] . Hence W [ h ] > 0 W[h]>0 W [ h ] > 0 throughout [ 0 , T ] [0,T] [ 0 , T ] .
A time-uniform non-collapsing estimate
We now assume the strict dynamic condition (8) and the quantitative boundary condition (9).
Lemma 4.3 (Quantitative normal boundary sign). At a boundary point at which the normal radius λ n \lambda_n λ n is the minimum eigenvalue, the normal support v = h W ( μ , μ ) v=hW(\mu,\mu) v = hW ( μ , μ ) satisfies
Proof. If k = 1 k=1 k = 1 , then λ α ≥ λ n \lambda_\alpha\ge\lambda_n λ α ≥ λ n gives
1 + ∑ α < n λ α λ n ≥ n . 1+\frac{\sum_{\alpha<n}\lambda_\alpha}{\lambda_n}\ge n. 1 + λ n ∑ α < n λ α ≥ n . The boundary assumption (9) becomes
f μ f ≥ n − 1 n cot θ , \frac{f_\mu}{f}\ge\frac{n-1}{n}\cot\theta, f f μ ≥ n n − 1 cot θ ,
so the k = 1 k=1 k = 1 formula in Lemma 4.2 gives v μ ≥ 0 v_\mu\ge0 v μ ≥ 0 .
Assume 2 ≤ k ≤ n − 1 2 \le k \le n-1 2 ≤ k ≤ n − 1 and use s j s_j s j from Lemma 4.2. The identity
k s k = ∑ α < n λ α σ k − 1 ( χ ∣ α ) ks_k=\sum_{\alpha<n}\lambda_\alpha\sigma_{k-1}(\chi|\alpha) k s k = α < n ∑ λ α σ k − 1 ( χ ∣ α )
and λ α ≥ λ n \lambda_\alpha\ge\lambda_n λ α ≥ λ n give
s k λ n s k − 1 ≥ n − k k . \frac{s_k}{\lambda_n s_{k-1}}\ge\frac{n-k}{k}. λ n s k − 1 s k ≥ k n − k .
Together with
( n − k + 1 ) λ n s k − 2 s k − 1 ≤ k − 1 , (n-k+1)\lambda_n\frac{s_{k-2}}{s_{k-1}}\le k-1, ( n − k + 1 ) λ n s k − 1 s k − 2 ≤ k − 1 ,
this gives the two coefficient bounds needed in (40). Moreover,
H μ H = f μ f + ( k − 1 ) cot θ ≥ k ( n − 1 ) n cot θ \frac{H_\mu}{H}=\frac{f_\mu}{f}+(k-1)\cot\theta\ge\frac{k(n-1)}{n}\cot\theta H H μ = f f μ + ( k − 1 ) cot θ ≥ n k ( n − 1 ) cot θ
by (9). Substitution into (40) yields
v μ v ≥ 0. \frac{v_\mu}{v}\ge0. v v μ ≥ 0.
At θ = π / 2 \theta=\pi/2 θ = π /2 all cot θ \cot\theta cot θ terms vanish, so the same conclusion holds.
Proposition 4.4 (Uniform lower principal-radius bound). Under the geometric, data, and initial hypotheses of Theorem 1.1, there exists c > 0 c>0 c > 0 such that
h λ min ( W [ h ] ) ≥ c for all t ≥ 0 h\lambda_{\min}(W[h])\ge c \qquad\text{for all }t\ge0 h λ m i n ( W [ h ]) ≥ c for all t ≥ 0 as long as the smooth solution exists. This estimate does not require Lemmas 2.1 and 5.1.
Proof. The previously established estimates provide constants, independent of time on the smooth strictly convex existence interval, such that
0 < h − ≤ h ≤ h + , 0 < Q − ≤ Q ≤ Q + , S ≤ S + , Θ ≤ Θ + , ∣ ∇ h ∣ ≤ H 1 . 0<h_-\le h\le h_+,\qquad0<Q_-\le Q\le Q_+,\qquad S\le S_+,\qquad\Theta\le\Theta_+,\qquad|\nabla h|\le H_1. 0 < h − ≤ h ≤ h + , 0 < Q − ≤ Q ≤ Q + , S ≤ S + , Θ ≤ Θ + , ∣∇ h ∣ ≤ H 1 .
Put
c ∗ : = τ 0 h − 2 Q − > 0 , B ∗ : = sup ∣ ( a φ ( ℓ / h ) + k − 3 ) Q + 2 ∣ < ∞ . c_*:=\tau_0h_-^2Q_->0,\qquad B_*:=\sup\left|(a_\varphi(\ell/h)+k-3)Q+2\right|<\infty. c ∗ := τ 0 h − 2 Q − > 0 , B ∗ := sup ∣ ( a φ ( ℓ / h ) + k − 3 ) Q + 2 ∣ < ∞.
The compact range of ℓ / h \ell/h ℓ / h makes B ∗ B_* B ∗ finite. Thus the right-hand side of (33) is at least c ∗ c_* c ∗ . Let q 0 , − = min C θ ‾ q ( ⋅ , 0 ) > 0 q_{0,-}=\min_{\overline{C_\theta}}q(\cdot,0)>0 q 0 , − = min C θ q ( ⋅ , 0 ) > 0 and choose
0 < ε < min { q 0 , − 4 , c ∗ 8 ( B ∗ + 1 ) } . 0<\varepsilon<\min\left\{\frac{q_{0,-}}{4},\frac{c_*}{8(B_*+1)}\right\}. 0 < ε < min { 4 q 0 , − , 8 ( B ∗ + 1 ) c ∗ } .
Choose a fixed smooth function β ≥ 0 \beta\ge0 β ≥ 0 with β = 0 \beta=0 β = 0 and β μ = − 1 \beta_\mu=-1 β μ = − 1 on the boundary. We will use ξ ( x ) = ε e ζ β ( x ) \xi(x)=\varepsilon e^{\zeta\beta(x)} ξ ( x ) = ε e ζ β ( x ) , where ζ > 0 \zeta>0 ζ > 0 will be chosen after estimating the coefficients in terms of ε \varepsilon ε .
Before and at any first contact of q q q with ξ \xi ξ , the inequality q ≥ ξ ≥ ε q\ge\xi\ge\varepsilon q ≥ ξ ≥ ε implies
λ min ( W ) ≥ ε h + . \lambda_{\min}(W)\ge\frac{\varepsilon}{h_+}. λ m i n ( W ) ≥ h + ε .
A monomial of σ k ( W ) \sigma_k(W) σ k ( W ) containing λ max \lambda_{\max} λ m a x and any k − 1 k-1 k − 1 other eigenvalues therefore yields
λ max ( W ) ≤ L ε : = ( n k ) S + ( h + ε ) k − 1 . \lambda_{\max}(W) \le L_{\varepsilon} := \binom{n}{k} S_+ \left(\frac{h_+}{\varepsilon}\right)^{k-1}. λ m a x ( W ) ≤ L ε := ( k n ) S + ( ε h + ) k − 1 .
This also covers k = 1 k=1 k = 1 , with the empty product equal to one. Consequently the positive matrix A i j = Θ S i j A^{ij} = \Theta S^{ij} A ij = Θ S ij satisfies
∥ A ∥ o p ≤ Λ ε : = Θ + k n L ε k − 1 . \lVert A\rVert_{\mathrm{op}} \le\Lambda_{\varepsilon} := \Theta_+ \frac{k}{n} L_{\varepsilon}^{k-1}. ∥ A ∥ op ≤ Λ ε := Θ + n k L ε k − 1 .
This follows from S i i = ( n k ) − 1 σ k − 1 ( λ ∣ i ) S^{ii} = \binom{n}{k}^{-1}\sigma_{k-1}(\lambda|i) S ii = ( k n ) − 1 σ k − 1 ( λ ∣ i ) at a diagonal matrix. All these bounds are independent of the proposed contact time and of ζ \zeta ζ .
Write K 1 = ∥ ∇ β ∥ ∞ K_1 = \lVert\nabla\beta\rVert_{\infty} K 1 = ∥ ∇ β ∥ ∞ , K 2 = ∥ ∇ 2 β ∥ o p , ∞ K_2 = \lVert\nabla^2\beta\rVert_{\mathrm{op},\infty} K 2 = ∥ ∇ 2 β ∥ op , ∞ and K 0 = ∥ β ∥ ∞ K_0 = \lVert\beta\rVert_{\infty} K 0 = ∥ β ∥ ∞ , and set
D ε = n Λ ε ( K 2 + 2 K 1 H 1 / h − + K 1 2 ) . D_{\varepsilon} = n\Lambda_{\varepsilon}\left(K_2 + 2K_1H_1/h_- + K_1^2\right). D ε = n Λ ε ( K 2 + 2 K 1 H 1 / h − + K 1 2 ) .
Choose
0 < ζ ≤ min { 1 , log 2 K 0 + 1 , c ∗ 8 ε ( D ε + 1 ) } . 0 < \zeta\le\min\left\{1,\frac{\log2}{K_0+1},\frac{c_*}{8\varepsilon(D_{\varepsilon}+1)}\right\}. 0 < ζ ≤ min { 1 , K 0 + 1 log 2 , 8 ε ( D ε + 1 ) c ∗ } .
Then ε ≤ ξ ≤ 2 ε < q 0 , − \varepsilon\le\xi\le2\varepsilon< q_{0,-} ε ≤ ξ ≤ 2 ε < q 0 , − , so the initial comparison holds at every point of C θ ‾ \overline{C_\theta} C θ . At a proposed contact, direct differentiation gives
− L ξ + 2 Θ S i j ξ i ∇ j log ξ h = ξ ζ A i j ( − β i j + ζ β i β j − 2 β i h j / h ) . -\mathcal{L}\xi+ 2\Theta S^{ij}\xi_i\nabla_j\log\frac{\xi}{h}
= \xi\zeta A^{ij}\left(-\beta_{ij}+\zeta\beta_i\beta_j-2\beta_i h_j/h\right). − L ξ + 2Θ S ij ξ i ∇ j log h ξ = ξ ζ A ij ( − β ij + ζ β i β j − 2 β i h j / h ) .
Its absolute value is at most ξ ζ D ε ≤ c ∗ / 4 \xi\zeta D_{\varepsilon}\le c_*/4 ξ ζ D ε ≤ c ∗ /4 . The absolute value of the zeroth-order contribution is at most B ∗ ξ ≤ 2 B ∗ ε < c ∗ / 4 B_*\xi\le2B_*\varepsilon<c_*/4 B ∗ ξ ≤ 2 B ∗ ε < c ∗ /4 . Since ξ t = 0 \xi_t=0 ξ t = 0 , an interior first contact would therefore give c ∗ ≤ c ∗ / 2 c_*\le c_*/2 c ∗ ≤ c ∗ /2 in (33) , a contradiction.
At a first contact on the boundary, use the splitting W α n = 0 W_{\alpha n}=0 W α n = 0 to choose a tangential minimum eigenvector if one exists and e = μ e=\mu e = μ otherwise. Extend e e e as a smooth unit vector field and set v = h W ( e , e ) v=hW(e,e) v = hW ( e , e ) . Then v ≥ q ≥ ξ v\ge q\ge\xi v ≥ q ≥ ξ and v = ξ v=\xi v = ξ at contact, so ( v − ξ ) μ ≤ 0 (v-\xi)_\mu\le0 ( v − ξ ) μ ≤ 0 . For a tangential minimum, the boundary identity gives v μ = h cot θ λ n ≥ 0 v_\mu=h\cot\theta\,\lambda_n\ge0 v μ = h cot θ λ n ≥ 0 . For a normal minimum, Lemma 4.3 gives the same inequality. But
ξ μ = − ζ ξ < 0 , \xi_\mu=-\zeta\xi<0, ξ μ = − ζ ξ < 0 ,
which contradicts ( v − ξ ) μ ≤ 0 (v-\xi)_\mu\le0 ( v − ξ ) μ ≤ 0 in either case, including θ = π / 2 \theta=\pi/2 θ = π /2 .
There is therefore no first contact. Since all constants were chosen from the initial data and the preceding time-uniform estimates,
h λ min ( W ) = q ≥ ξ ≥ ε > 0 h\lambda_{\min}(W)=q\ge\xi\ge\varepsilon>0 h λ m i n ( W ) = q ≥ ξ ≥ ε > 0
on the entire smooth existence interval.
Combining Proposition 4.4 with (25) gives
λ min ( W ) ≥ c > 0. (41) \lambda_{\min}(W)\ge c>0. \tag*{(41)} λ m i n ( W ) ≥ c > 0. (41)
Proposition 4.5 (Two-sided radius estimate). There exist 0 < c < C < ∞ 0 < c < C < \infty 0 < c < C < ∞ such that
c σ ≤ W [ h ( ⋅ , t ) ] ≤ C σ c\sigma\le W[h(\cdot,t)] \le C\sigma c σ ≤ W [ h ( ⋅ , t )] ≤ C σ on the whole existence interval.
Proof. The lower bound is (41) . If λ n = λ max ( W ) \lambda_n = \lambda_{\max}(W) λ n = λ m a x ( W ) , then one term in σ k ( W ) \sigma_k(W) σ k ( W ) contains λ n \lambda_n λ n times k − 1 k-1 k − 1 other eigenvalues, each bounded below by c c c . Hence
σ k ( W ) ≥ c k − 1 λ n . \sigma_k(W) \ge c^{k-1}\lambda_n. σ k ( W ) ≥ c k − 1 λ n .
The upper bound in (26) therefore yields λ n ≤ C \lambda_n \le C λ n ≤ C .
Long-time behavior and contact-angle issues
Long-time existence
Lemma 5.1 (Quantitative interior concave estimate). Let n ≥ 2 n \ge2 n ≥ 2 and let F : B 1 × Sym ( n ) → R \mathfrak{F}: B_1 \times\operatorname{Sym}(n) \to\mathbb{R} F : B 1 × Sym ( n ) → R be continuous, concave in its matrix variable, and uniformly elliptic:
λ tr N ≤ F ( x , M + N ) − F ( x , M ) ≤ Λ tr N ( N ≥ 0 ) . \lambda\operatorname{tr} N \le\mathfrak{F}(x,M+N)-\mathfrak{F}(x,M) \le\Lambda\operatorname{tr} N \qquad(N \ge0). λ tr N ≤ F ( x , M + N ) − F ( x , M ) ≤ Λ tr N ( N ≥ 0 ) . Suppose
∣ F ( x , M ) − F ( y , M ) ∣ ≤ L ( 1 + ∥ M ∥ ) ∣ x − y ∣ , ∣ F ( x , 0 ) ∣ ≤ L . |\mathfrak{F}(x,M)-\mathfrak{F}(y,M)| \le L(1+\lVert M\rVert)|x-y|,\qquad|\mathfrak{F}(x,0)| \le L. ∣ F ( x , M ) − F ( y , M ) ∣ ≤ L ( 1 + ∥ M ∥) ∣ x − y ∣ , ∣ F ( x , 0 ) ∣ ≤ L . If g ∈ C α ( B 1 ) g \in C^\alpha(B_1) g ∈ C α ( B 1 ) , 0 < α < 1 0 < \alpha< 1 0 < α < 1 , and a bounded classical solution satisfies F ( x , D 2 v ) = g \mathfrak{F}(x,D^2v)=g F ( x , D 2 v ) = g , then
∥ v ∥ C 2 , β ( B 1 / 2 ) ≤ C ( ∥ v ∥ L ∞ ( B 1 ) + 1 + ∥ g ∥ C α ( B 1 ) ) \lVert v\rVert_{C^{2,\beta}(B_{1/2})} \le C\bigl(\lVert v\rVert_{L^\infty(B_1)}+1+\lVert g\rVert_{C^\alpha(B_1)}\bigr) ∥ v ∥ C 2 , β ( B 1/2 ) ≤ C ( ∥ v ∥ L ∞ ( B 1 ) + 1 + ∥ g ∥ C α ( B 1 ) ) for some 0 < β < α 0 < \beta< \alpha 0 < β < α . The constants depend only on n , λ , Λ , L , α n,\lambda,\Lambda,L,\alpha n , λ , Λ , L , α and the chosen admissible β \beta β . No differentiability in the matrix variable is required.
Proof. Set U = − v U=-v U = − v and G ( x , z , p , M ) = − F ( x , − M ) + g ( x ) G(x,z,p,M)=-\mathfrak{F}(x,-M)+g(x) G ( x , z , p , M ) = − F ( x , − M ) + g ( x ) . Then G G G is convex in M M M , independent of ( z , p ) (z,p) ( z , p ) , and satisfies the rank-one ellipticity condition with ν = min { 1 , λ , Λ − 1 } \nu=\min\{1,\lambda,\Lambda^{-1}\} ν = min { 1 , λ , Λ − 1 } . Choose 0 < β < min { α , α ˉ ( n , ν ) } 0 < \beta< \min\{\alpha,\bar{\alpha}(n,\nu)\} 0 < β < min { α , α ˉ ( n , ν )} . On a ball of diameter at most two,
[ G ( ⋅ , z , p , M ) ] C β ≤ 2 1 − β L ( 1 + ∥ M ∥ ) + 2 α − β [ g ] C α . [G(\cdot,z,p,M)]_{C^\beta} \le2^{1-\beta}L(1+\lVert M\rVert)+2^{\alpha-\beta}[g]_{C^\alpha}. [ G ( ⋅ , z , p , M ) ] C β ≤ 2 1 − β L ( 1 + ∥ M ∥) + 2 α − β [ g ] C α .
Together with ∣ G ( x , 0 , 0 , 0 ) ∣ ≤ L + ∥ g ∥ ∞ |G(x,0,0,0)| \le L+\lVert g\rVert_\infty ∣ G ( x , 0 , 0 , 0 ) ∣ ≤ L + ∥ g ∥ ∞ , this verifies all of Safonov’s conditions (F0)–(F4) [13 Assumptions 3.1, p. 9] , including the required linear growth of the spatial Hölder modulus in ∑ i , j ∣ M i j ∣ \sum_{i,j}|M_{ij}| ∑ i , j ∣ M ij ∣ . Apply his Theorem 3.1 and estimate (3.6), p. 10, on a ball compactly contained in B 1 B_1 B 1 , with boundary values taken from U U U . The resulting interior C 2 , β C^{2,\beta} C 2 , β solution, continuous up to the boundary of the ball, equals U U U by classical comparison [13 Corollary 3.1, p. 12] . The claim follows by restricting the weighted estimate to a smaller ball and using a finite covering of B 1 / 2 B_{1/2} B 1/2 .
Remark 5.2 (Choice of the interior estimate). Theorem 6.2 of Lieberman–Trudinger [17 ] , under its condition (6.1), requires a spatial derivative bound for the operator. Applied with right-hand side g = ( f Q / ( h φ ( ℓ / h ) ) ) 1 / k g=(fQ/(h\varphi(\ell/h)))^{1/k} g = ( f Q / ( h φ ( ℓ / h )) ) 1/ k , that bound would require an unavailable uniform bound for D g Dg D g . A spatial C α C^\alpha C α bound for g g g alone does not verify that application. We instead use Lemma 5.1, which requires only the stated spatial Hölder norm of the right-hand side.
Lemma 5.3 (Boundary second-derivative Hölder estimate). Fix n ≥ 2 n \ge2 n ≥ 2 , 0 < θ < π 0 < \theta< \pi 0 < θ < π , 1 ≤ k ≤ n 1 \le k \le n 1 ≤ k ≤ n , 0 < c < C 0 < c < C 0 < c < C , M 0 < ∞ M_0 < \infty M 0 < ∞ , and 0 < α < 1 0 < \alpha< 1 0 < α < 1 . Let h h h be smooth on the closed spherical cap C θ C_\theta C θ and suppose
h μ = cot θ h , c σ ≤ W [ h ] : = ∇ 2 h + h σ ≤ C σ , ∥ h ∥ C 2 ≤ M 0 . h_\mu= \cot\theta\, h,\qquad c\sigma\le W[h] := \nabla^2 h + h\sigma\le C\sigma,\qquad\lVert h\rVert_{C^2} \le M_0. h μ = cot θ h , c σ ≤ W [ h ] := ∇ 2 h + hσ ≤ C σ , ∥ h ∥ C 2 ≤ M 0 . If
σ ˉ k ( W [ h ] ) 1 / k = g , ∥ g ∥ C α ( C θ ‾ ) ≤ M 1 , \bar{\sigma}_k(W[h])^{1/k} = g,\qquad\lVert g\rVert_{C^\alpha(\overline{C_\theta})} \le M_1, σ ˉ k ( W [ h ] ) 1/ k = g , ∥ g ∥ C α ( C θ ) ≤ M 1 , then, for some β > 0 \beta> 0 β > 0 depending only on the indicated fixed structural data,
∥ h ∥ C 2 , β ( C θ ‾ ) ≤ C ∗ . \lVert h\rVert_{C^{2,\beta}(\overline{C_\theta})} \le C_*. ∥ h ∥ C 2 , β ( C θ ) ≤ C ∗ . Here C ∗ C_* C ∗ depends on n n n , k k k , θ \theta θ , c c c , C C C , M 0 M_0 M 0 , M 1 M_1 M 1 , α \alpha α and the constants in Lemma 5.1, and not on higher derivatives of h h h or g g g .
Proof. Set u = h / ℓ u = h/\ell u = h / ℓ . The fixed positive function ℓ \ell ℓ satisfies ∇ 2 ℓ + ℓ σ = σ \nabla^2\ell+ \ell\sigma= \sigma ∇ 2 ℓ + ℓ σ = σ and ℓ μ = cot θ ℓ \ell_\mu= \cot\theta\,\ell ℓ μ = cot θ ℓ , so
u μ = 0 , W [ h ] = ℓ ∇ 2 u + d ℓ ⊗ d u + d u ⊗ d ℓ + u σ . u_\mu= 0,\qquad W[h] = \ell\nabla^2u + d\ell\otimes du + du \otimes d\ell+ u\sigma. u μ = 0 , W [ h ] = ℓ ∇ 2 u + d ℓ ⊗ d u + d u ⊗ d ℓ + u σ .
In particular, ∥ u ∥ C 2 \lVert u\rVert_{C^2} ∥ u ∥ C 2 is bounded by fixed data. The second identity, including the term u σ u\sigma u σ , is used below.
Choose a collar of the boundary, let r ≥ 0 r \ge0 r ≥ 0 be the inward geodesic distance, and use angular coordinates y y y on S n − 1 S^{n-1} S n − 1 . The round metric and ℓ \ell ℓ are
σ = d r 2 + ψ ( r ) 2 γ , ψ ( r ) = sin ( θ − r ) , ℓ ( r ) = 1 − cos θ cos ( θ − r ) , \sigma= dr^2 + \psi(r)^2\gamma,\qquad\psi(r) = \sin(\theta-r),\qquad\ell(r) = 1 - \cos\theta\cos(\theta-r), σ = d r 2 + ψ ( r ) 2 γ , ψ ( r ) = sin ( θ − r ) , ℓ ( r ) = 1 − cos θ cos ( θ − r ) ,
where γ \gamma γ is the unit round metric on S n − 1 S^{n-1} S n − 1 . We choose and fix the collar small enough that ψ > 0 \psi> 0 ψ > 0 . Let E a = E a α ( y ) ∂ y α E_a = E_a^\alpha(y)\partial_{y^\alpha} E a = E a α ( y ) ∂ y α , a = 1 , … , n − 1 a = 1,\ldots,n-1 a = 1 , … , n − 1 , be a smooth local γ \gamma γ -orthonormal frame. An orthonormal frame for σ \sigma σ is e a = ψ − 1 E a e_a = \psi^{-1}E_a e a = ψ − 1 E a and e n = ∂ r e_n = \partial_r e n = ∂ r . In this frame direct computation gives
W a b = ℓ ψ 2 E a α E b β ( u α β − Γ α β ( γ ) δ u δ ) + ℓ ψ ′ ψ u r δ a b + u δ a b , W_{ab} = \frac{\ell}{\psi^2}E_a^\alpha E_b^\beta\left(u_{\alpha\beta} - \Gamma^{(\gamma)\delta}_{\alpha\beta}u_\delta\right) + \ell\frac{\psi'}{\psi}u_r\delta_{ab} + u\delta_{ab}, W ab = ψ 2 ℓ E a α E b β ( u α β − Γ α β ( γ ) δ u δ ) + ℓ ψ ψ ′ u r δ ab + u δ ab ,
W a n = ℓ ψ E a α u α r + 1 ψ E a α ( ℓ ′ − ℓ ψ ′ ψ ) u α , W_{an} = \frac{\ell}{\psi}E_a^\alpha u_{\alpha r} + \frac{1}{\psi}E_a^\alpha\left(\ell' - \ell\frac{\psi'}{\psi}\right)u_\alpha, W an = ψ ℓ E a α u α r + ψ 1 E a α ( ℓ ′ − ℓ ψ ψ ′ ) u α ,
W n n = ℓ u r r + 2 ℓ ′ u r + u . W_{nn} = \ell u_{rr} + 2\ell'u_r + u. W nn = ℓ u r r + 2 ℓ ′ u r + u .
Thus, in ordinary coordinate Hessians,
W = P D 2 u P T + L , P = ℓ ( ψ − 1 E 0 0 1 ) . W = P D^2uP^{T} + L,\qquad P = \sqrt{\ell}\begin{pmatrix}\psi^{-1}E & 0 \\ 0 & 1\end{pmatrix}. W = P D 2 u P T + L , P = ℓ ( ψ − 1 E 0 0 1 ) .
where E = ( E a α ) E = (E_a^\alpha) E = ( E a α ) . The displayed formulas specify every entry of L L L . The matrix P P P is uniformly invertible and smooth on each fixed coordinate collar. Because L L L depends smoothly on the fixed geometry, u u u , and D u Du D u , its Lipschitz norm on the half-collar is bounded by the already available C 2 C^2 C 2 norm of u u u .
The needed boundary cancellation is exact:
ℓ ( 0 ) = sin 2 θ , ℓ ′ ( 0 ) = − cos θ sin θ , ψ ′ ( 0 ) ψ ( 0 ) = − cot θ , ℓ ′ ( 0 ) − ℓ ( 0 ) ψ ′ ( 0 ) ψ ( 0 ) = 0. \ell(0) = \sin^2\theta,\qquad\ell'(0) = -\cos\theta\sin\theta,\qquad\frac{\psi'(0)}{\psi(0)} = -\cot\theta,\qquad\ell'(0) - \ell(0)\frac{\psi'(0)}{\psi(0)} = 0. ℓ ( 0 ) = sin 2 θ , ℓ ′ ( 0 ) = − cos θ sin θ , ψ ( 0 ) ψ ′ ( 0 ) = − cot θ , ℓ ′ ( 0 ) − ℓ ( 0 ) ψ ( 0 ) ψ ′ ( 0 ) = 0.
Consequently L a n ( y , 0 ) = 0 L_{an}(y,0)=0 L an ( y , 0 ) = 0 . Also u r ( y , 0 ) = 0 u_r(y,0)=0 u r ( y , 0 ) = 0 , and differentiation along the boundary gives u α r ( y , 0 ) = 0 u_{\alpha r}(y,0)=0 u α r ( y , 0 ) = 0 .
Write R = diag ( 1 , … , 1 , − 1 ) R=\operatorname{diag}(1,\ldots,1,-1) R = diag ( 1 , … , 1 , − 1 ) . Extend u u u evenly by u ~ ( y , r ) = u ( y , ∣ r ∣ ) \widetilde{u}(y,r)=u(y,|r|) u ( y , r ) = u ( y , ∣ r ∣ ) . The last two identities imply that this extension is C 2 C^2 C 2 across r = 0 r=0 r = 0 . For r < 0 r<0 r < 0 define
P ~ ( y , r ) = P ( y , − r ) , L ~ ( y , r ) = R L ( y , − r ) R , g ~ ( y , r ) = g ( y , − r ) . \widetilde{P}(y,r)=P(y,-r),\qquad\widetilde{L}(y,r)=RL(y,-r)R,\qquad\widetilde{g}(y,r)=g(y,-r). P ( y , r ) = P ( y , − r ) , L ( y , r ) = R L ( y , − r ) R , g ( y , r ) = g ( y , − r ) .
and keep the original values for r ≥ 0 r\geq0 r ≥ 0 . Since P P P is block diagonal, it commutes with R R R . The tangential and normal diagonal blocks of L L L are reflected evenly, while its mixed block is reflected oddly. Since L a n ( y , 0 ) = 0 L_{an}(y,0)=0 L an ( y , 0 ) = 0 , L ~ \widetilde{L} L is Lipschitz across the interface, with its Lipschitz constant bounded by the same fixed data. The even extension P ~ \widetilde{P} P is also Lipschitz and uniformly invertible, and g ~ \widetilde{g} g is uniformly C α C^\alpha C α . For r < 0 r<0 r < 0 ,
P ~ D 2 u ~ P ~ T + L ~ = R W ( y , − r ) R . \widetilde{P}D^2\widetilde{u}\widetilde{P}^{T}+\widetilde{L}=RW(y,-r)R. P D 2 u P T + L = R W ( y , − r ) R .
Thus this reflected matrix has eigenvalues in [ c , C ] [c,C] [ c , C ] everywhere. To put the operator into a globally uniformly elliptic class, let F ( A ) = σ ‾ k ( A ) 1 / k F(A)=\overline{\sigma}_k(A)^{1/k} F ( A ) = σ k ( A ) 1/ k and K = { A : ( c / 2 ) I ≤ A ≤ 2 C I } \mathcal{K}=\{A:(c/2)I\leq A\leq2CI\} K = { A : ( c /2 ) I ≤ A ≤ 2 C I } . Define
F ^ ( A ) = inf B ∈ K { F ( B ) + D F ( B ) : ( A − B ) } , A ∈ Sym ( n ) . \widehat{F}(A)=\inf_{B\in\mathcal{K}}\left\{F(B)+DF(B):(A-B)\right\},\qquad A\in\operatorname{Sym}(n). F ( A ) = B ∈ K inf { F ( B ) + D F ( B ) : ( A − B ) } , A ∈ Sym ( n ) .
Concavity of F F F on the positive cone shows that F ^ = F \widehat{F}=F F = F on K \mathcal{K} K . Since D F ( B ) DF(B) D F ( B ) has positive eigenvalues bounded above and below uniformly for B ∈ K B\in\mathcal{K} B ∈ K , there are 0 < λ 0 ≤ Λ 0 0<\lambda_0\leq\Lambda_0 0 < λ 0 ≤ Λ 0 such that, for every N ≥ 0 N\geq0 N ≥ 0 ,
λ 0 tr N ≤ F ^ ( A + N ) − F ^ ( A ) ≤ Λ 0 tr N . \lambda_0\operatorname{tr}N\leq\widehat{F}(A+N)-\widehat{F}(A)\leq\Lambda_0\operatorname{tr}N. λ 0 tr N ≤ F ( A + N ) − F ( A ) ≤ Λ 0 tr N .
The operator F ^ \widehat{F} F is concave, globally Lipschitz, and invariant under orthogonal conjugation: this last fact follows because both F F F and K \mathcal{K} K are invariant under such conjugation. Now set
F ~ ( x , M ) = F ^ ( P ~ ( x ) M P ~ ( x ) T + L ~ ( x ) ) . \widetilde{\mathcal{F}}(x,M)=\widehat{F}\left(\widetilde{P}(x)M\widetilde{P}(x)^T+\widetilde{L}(x)\right). F ( x , M ) = F ( P ( x ) M P ( x ) T + L ( x ) ) .
Uniform invertibility of P ~ \widetilde{P} P gives uniform ellipticity of F ~ \widetilde{\mathcal{F}} F ; concavity is preserved under this affine substitution. Its Lipschitz dependence on x x x satisfies
∣ F ~ ( x , M ) − F ~ ( y , M ) ∣ ≤ C 0 ( 1 + ∥ M ∥ ) ∣ x − y ∣ , |\widetilde{\mathcal{F}}(x,M)-\widetilde{\mathcal{F}}(y,M)|\leq C_0(1+\lVert M\rVert)|x-y|, ∣ F ( x , M ) − F ( y , M ) ∣ ≤ C 0 ( 1 + ∥ M ∥) ∣ x − y ∣ ,
where C 0 C_0 C 0 depends only on the existing C 2 C^2 C 2 bound and the fixed data. Moreover ∣ F ~ ( x , 0 ) ∣ |\widetilde{\mathcal{F}}(x,0)| ∣ F ( x , 0 ) ∣ is bounded by the same data. Orthogonal invariance and agreement with F F F on K \mathcal{K} K show that
F ~ ( x , D 2 u ~ ) = g ~ ( x ) \widetilde{\mathcal{F}}(x,D^2\widetilde{u})=\widetilde{g}(x) F ( x , D 2 u ) = g ( x )
on the doubled collar; at r = 0 r=0 r = 0 the equality follows by continuity. Application of Lemma 5.1 on balls in this doubled collar therefore yields a uniform C 2 , β C^{2,\beta} C 2 , β estimate up to the original boundary. The same construction without reflection gives the interior estimates. A finite fixed coordinate covering, followed by multiplication by ℓ \ell ℓ , proves the asserted estimate for h h h .
The proof uses only the existing C 2 C^2 C 2 bound to control the Lipschitz coefficients of the reflected operator. Thus the reduction to Lemma 5.1 does not presuppose the second-derivative Hölder estimate that it seeks to establish.
Proof of the long-time existence part of Theorem 1.1. Let [ 0 , T max ) [0,T_{\max}) [ 0 , T m a x ) be the maximal smooth strictly convex existence interval. By Proposition 4.5,
c σ ≤ W [ h ] ≤ C σ on C θ × [ 0 , T max ) , c\sigma\le W[h] \le C\sigma\qquad\text{on } C_\theta\times[0,T_{\max}), c σ ≤ W [ h ] ≤ C σ on C θ × [ 0 , T m a x ) ,
and by Proposition 3.1,
0 < c ≤ h ≤ C . 0 < c \le h \le C. 0 < c ≤ h ≤ C .
Since ℓ / h \ell/h ℓ / h stays in a compact subset of ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) and f , φ f,\varphi f , φ are fixed positive smooth data, the mobility Θ \Theta Θ also satisfies
0 < c Θ ≤ Θ ≤ C Θ . 0 < c_\Theta\le\Theta\le C_\Theta. 0 < c Θ ≤ Θ ≤ C Θ .
On the compact matrix set
{ A : c I ≤ A ≤ C I } , \{A:cI\le A\le CI\}, { A : c I ≤ A ≤ C I } ,
the derivative matrix S i j ( A ) S^{ij}(A) S ij ( A ) is positive definite with eigenvalues bounded above and below. Consequently Θ S i j \Theta S^{ij} Θ S ij is uniformly elliptic on the entire interval [ 0 , T max ) [0,T_{\max}) [ 0 , T m a x ) . The boundary operator ∇ μ h − cot θ h \nabla_\mu h-\cot\theta\,h ∇ μ h − cot θ h has gradient derivative μ \mu μ , so it is uniformly oblique; when θ = π / 2 \theta=\pi/2 θ = π /2 it is the homogeneous Neumann condition.
We now explain the regularity bootstrap. The speed ratio solves the linear uniformly parabolic equation with drift
Q t − Θ S i j Q i j − 2 Θ S i j ( log h ) i Q j = ( k + 1 − a φ ( ℓ / h ) ) Q ( Q − 1 ) , Q μ = 0. Q_t-\Theta S^{ij}Q_{ij}-2\Theta S^{ij}(\log h)_iQ_j=(k+1-a_\varphi(\ell/h))Q(Q-1),\qquad Q_\mu=0. Q t − Θ S ij Q ij − 2Θ S ij ( log h ) i Q j = ( k + 1 − a φ ( ℓ / h )) Q ( Q − 1 ) , Q μ = 0.
The coefficients are bounded on every positive-time strip and the right-hand side is bounded. The first assertion of Lemma 2.1, applied with homogeneous Neumann data, therefore gives, for each τ > 0 \tau>0 τ > 0 ,
∥ Q ∥ C α , α / 2 ( C θ ‾ × [ τ , T max ) ) ≤ C τ . \|Q\|_{C^{\alpha,\alpha/2}(\overline{C_\theta}\times[\tau,T_{\max}))}\le C_\tau. ∥ Q ∥ C α , α /2 ( C θ × [ τ , T m a x )) ≤ C τ .
At a fixed time the equation
σ ‾ k ( W [ h ] ) 1 / k = ( f Q h φ ( ℓ / h ) ) 1 / k \overline{\sigma}_k(W[h])^{1/k}=\left(\frac{fQ}{h\varphi(\ell/h)}\right)^{1/k} σ k ( W [ h ] ) 1/ k = ( h φ ( ℓ / h ) f Q ) 1/ k
is used in the regularity argument below.
Application to the flow. On each fixed positive-time strip, the previously established estimates give ∥ h ∥ C 2 ≤ M 0 \|h\|_{C^2}\le M_0 ∥ h ∥ C 2 ≤ M 0 , c σ ≤ W [ h ] ≤ C σ c\sigma\le W[h]\le C\sigma c σ ≤ W [ h ] ≤ C σ , and
g ( x , t ) = ( f ( x ) Q ( x , t ) h ( x , t ) φ ( ℓ ( x ) / h ( x , t ) ) ) 1 / k . g(x,t)=\left(\frac{f(x)Q(x,t)}{h(x,t)\varphi(\ell(x)/h(x,t))}\right)^{1/k}. g ( x , t ) = ( h ( x , t ) φ ( ℓ ( x ) / h ( x , t )) f ( x ) Q ( x , t ) ) 1/ k .
The bounds for h h h and Q Q Q , including their positive lower bounds, together with the spatial C 1 C^1 C 1 bound for h h h and the spatial C α C^\alpha C α bound for Q Q Q , give sup t ∥ g ( ⋅ , t ) ∥ C α ≤ M 1 \sup_t\|g(\cdot,t)\|_{C^\alpha}\le M_1 sup t ∥ g ( ⋅ , t ) ∥ C α ≤ M 1 . The boundary second-derivative estimate therefore gives a uniform spatial C 2 , γ C^{2,\gamma} C 2 , γ bound for some 0 < γ < α 0<\gamma<\alpha 0 < γ < α , where α \alpha α denotes the preceding parabolic Hölder exponent for Q Q Q . No spatial derivative of Q Q Q is required in this step.
Since h t = h ( Q − 1 ) h_t=h(Q-1) h t = h ( Q − 1 ) , we have ∥ h t ∥ ∞ ≤ C \|h_t\|_\infty\le C ∥ h t ∥ ∞ ≤ C . Fix 0 < τ < T max 0<\tau<T_{\max} 0 < τ < T m a x and use the spatial estimate on [ τ / 2 , T max ) [\tau/2,T_{\max}) [ τ /2 , T m a x ) . For s , t s,t s , t in this interval, interpolation on the fixed smooth cap gives
∥ D 2 h ( t ) − D 2 h ( s ) ∥ ∞ ≤ C ∥ h ( t ) − h ( s ) ∥ ∞ γ / ( 2 + γ ) ∥ h ( t ) − h ( s ) ∥ C 2 , γ 2 / ( 2 + γ ) ≤ C τ ∣ t − s ∣ γ / ( 2 + γ ) . \|D^2h(t)-D^2h(s)\|_\infty\le C\|h(t)-h(s)\|_\infty^{\gamma/(2+\gamma)}\|h(t)-h(s)\|_{C^{2,\gamma}}^{2/(2+\gamma)}\le C_\tau|t-s|^{\gamma/(2+\gamma)}. ∥ D 2 h ( t ) − D 2 h ( s ) ∥ ∞ ≤ C ∥ h ( t ) − h ( s ) ∥ ∞ γ / ( 2 + γ ) ∥ h ( t ) − h ( s ) ∥ C 2 , γ 2/ ( 2 + γ ) ≤ C τ ∣ t − s ∣ γ / ( 2 + γ ) .
Choose
0 < β ≤ min { α , 2 γ / ( 2 + γ ) } . 0 < \beta\le\min\{\alpha, 2\gamma/(2+\gamma)\}. 0 < β ≤ min { α , 2 γ / ( 2 + γ )} .
The coefficients Θ S i j \Theta^{Sij} Θ S ij and 2 Θ S i j ( log h ) i 2\Theta^{Sij}(\log h)_i 2 Θ S ij ( log h ) i , and the source in the Q Q Q -equation, are uniformly C β , β / 2 C^{\beta,\beta/2} C β , β /2 on positive-time strips. Lemma 2.1(2) with homogeneous Neumann data then gives uniform C 2 + β , 1 + β / 2 C^{2+\beta,1+\beta/2} C 2 + β , 1 + β /2 estimates for Q Q Q on smaller strips. All these estimates are applied on backward strips of a fixed length depending on τ \tau τ , with nested earlier strips contained in [ τ / 2 , T max ) [\tau/2,T_{\max}) [ τ /2 , T m a x ) . Their constants are independent of the strip’s location and of T max T_{\max} T m a x . The already smooth local solution covers the initial layer.
Also h t = h ( Q − 1 ) h_t=h(Q-1) h t = h ( Q − 1 ) is uniformly C β , β / 2 C^{\beta,\beta/2} C β , β /2 . Together with the spatial and time bounds for D 2 h D^2h D 2 h , this yields h ∈ C 2 + β , 1 + β / 2 h\in C^{2+\beta,1+\beta/2} h ∈ C 2 + β , 1 + β /2 on the smaller strips. We estimate the time modulus of D h Dh D h using centered spatial differences, which have an error O ( r 1 + β ) O(r^{1+\beta}) O ( r 1 + β ) , while their change between times s , t s,t s , t is bounded by C ∣ t − s ∣ r β − 1 C|t-s|r^{\beta-1} C ∣ t − s ∣ r β − 1 using the spatial C β C^\beta C β bound for h t h_t h t . Taking r = ∣ t − s ∣ 1 / 2 r=|t-s|^{1/2} r = ∣ t − s ∣ 1/2 gives the required ∣ t − s ∣ ( 1 + β ) / 2 |t-s|^{(1+\beta)/2} ∣ t − s ∣ ( 1 + β ) /2 bound. A fixed smooth-domain extension operator gives the same estimate in boundary charts.
For clarity, the boundary bootstrap can be carried out with time and tangential derivatives only. Put w = h t w=h_t w = h t . Then
w t = Θ S i j ∇ i ∇ j w + ( Θ tr σ S + ( 2 − a φ ( ℓ / h ) ) Q − 1 ) w , w μ = cot θ w . w_t=\Theta^{Sij}\nabla_i\nabla_jw+\left(\Theta\operatorname{tr}_{\sigma}S+\left(2-a_\varphi(\ell/h)\right)Q-1\right)w,\qquad w_\mu=\cot\theta\,w. w t = Θ S ij ∇ i ∇ j w + ( Θ tr σ S + ( 2 − a φ ( ℓ / h ) ) Q − 1 ) w , w μ = cot θ w .
Its coefficients are uniformly Hölder, and w w w is uniformly bounded; linear parabolic oblique Schauder estimates apply on a smaller strip. In boundary Fermi coordinates ( y , r ) (y,r) ( y , r ) , with r r r inward distance, the boundary condition is h r + cot θ h = 0 h_r+\cot\theta\,h=0 h r + cot θ h = 0 . Consequently each tangential coordinate derivative v α = ∂ y α h v_\alpha=\partial_{y^\alpha}h v α = ∂ y α h satisfies the same homogeneous Robin condition. Differentiating the interior equation yields a linear uniformly parabolic equation for v α v_\alpha v α , whose source consists only of derivatives of the coefficients determined by the fixed geometry and data, together with derivatives of h h h of order at most two; these are uniformly Hölder at the first step. Thus the same linear estimate controls all third spatial derivatives having a tangential index. Uniform ellipticity allows the original equation to be solved locally for h r r h_{rr} h r r ; differentiating this identity, using the estimates for h t h_t h t , recovers the missing pure normal derivatives. Repeat this argument with time and tangential derivatives, on nested strips, to obtain all orders. At each fixed order the coefficients and differentiated boundary data required by the linear Schauder estimates are controlled by the preceding steps. Arbitrary normal derivatives are never assigned the original Robin boundary condition. This yields sup τ ≤ t < T max ∥ h ( t ) ∥ C m ≤ C m , τ \sup_{\tau\le t<T_{\max}}\|h(t)\|_{C^m}\le C_{m,\tau} sup τ ≤ t < T m a x ∥ h ( t ) ∥ C m ≤ C m , τ . If T max < ∞ T_{\max}<\infty T m a x < ∞ , the same bounds and the equation give uniform C m C^m C m bounds for h t h_t h t . Thus h ( t ) h(t) h ( t ) is Cauchy in every C m C^m C m as t ↑ T max t\uparrow T_{\max} t ↑ T m a x . Repeatedly differentiating the equation also passes every compatibility identity to the limiting jet.
Under this supposition, the limiting support function is smooth, positive, and strictly convex by the uniform bounds. The local existence assertion in Lemma 2.1 extends the solution through T max T_{\max} T m a x , contradicting maximality. Hence T max = ∞ T_{\max}=\infty T m a x = ∞ .
Uniqueness of the stationary solution
The supercritical growth condition also gives a simple comparison proof of uniqueness.
Proposition 5.4 (Stationary uniqueness). Assume
a φ ( s ) ≥ k + 1 + δ ( s > 0 ) . a_\varphi(s)\ge k+1+\delta\qquad(s>0). a φ ( s ) ≥ k + 1 + δ ( s > 0 ) .
Then (3) has at most one positive strictly convex solution. For k = 1 k=1 k = 1 , the same conclusion holds for positive solutions satisfying σ ˉ 1 ( W [ h ] ) > 0 \bar{\sigma}_1(W[h])>0 σ ˉ 1 ( W [ h ]) > 0 .
Proof. Let h 1 , h 2 h_1,h_2 h 1 , h 2 be two solutions and set
c = max C θ h 1 h 2 . c=\max_{\mathcal C_\theta}\frac{h_1}{h_2}. c = C θ max h 2 h 1 .
Suppose c > 1 c>1 c > 1 and let x 0 x_0 x 0 be a maximum point. Then w = h 1 − c h 2 w=h_1-ch_2 w = h 1 − c h 2 has maximum value zero at x 0 x_0 x 0 . At an interior maximum, the usual second-derivative test gives D w = 0 Dw=0 D w = 0 and D 2 w ≤ 0 D^2w\leq0 D 2 w ≤ 0 . At a boundary maximum of w = h 1 − c h 2 w=h_1-ch_2 w = h 1 − c h 2 , w = 0 w=0 w = 0 and the Robin condition gives ∇ μ w = 0 \nabla_{\mu}w=0 ∇ μ w = 0 . All tangential first derivatives also vanish, so D w = 0 Dw=0 D w = 0 . Taylor expansion along every inward-pointing direction gives D 2 w ( v , v ) ≤ 0 D^2w(v,v)\leq0 D 2 w ( v , v ) ≤ 0 . Since a quadratic form is even, the same inequality holds for outward-pointing directions; the inequality in tangential directions follows by continuity. Hence D 2 w ≤ 0 D^2w\leq0 D 2 w ≤ 0 on the whole tangent space and W [ h 1 ] ≤ c W [ h 2 ] W[h_1]\leq cW[h_2] W [ h 1 ] ≤ c W [ h 2 ] at the maximum point.
W [ h 1 ] ( x 0 ) ≤ c W [ h 2 ] ( x 0 ) . W[h_1](x_0)\leq cW[h_2](x_0). W [ h 1 ] ( x 0 ) ≤ c W [ h 2 ] ( x 0 ) .
For strictly convex matrices, monotonicity and k k k -homogeneity imply
σ ˉ k ( W [ h 1 ] ( x 0 ) ) ≤ c k σ ˉ k ( W [ h 2 ] ( x 0 ) ) . \bar{\sigma}_k(W[h_1](x_0))\leq c^k\bar{\sigma}_k(W[h_2](x_0)). σ ˉ k ( W [ h 1 ] ( x 0 )) ≤ c k σ ˉ k ( W [ h 2 ] ( x 0 )) .
For k = 1 k=1 k = 1 the same conclusion follows simply by taking traces. Since h 1 ( x 0 ) = c h 2 ( x 0 ) h_1(x_0)=ch_2(x_0) h 1 ( x 0 ) = c h 2 ( x 0 ) , the two stationary equations give
1 = h 1 φ ( ℓ / h 1 ) σ ˉ k ( W [ h 1 ] ) h 2 φ ( ℓ / h 2 ) σ ˉ k ( W [ h 2 ] ) ( x 0 ) ≤ c k + 1 φ ( ℓ ( x 0 ) / ( c h 2 ( x 0 ) ) ) φ ( ℓ ( x 0 ) / h 2 ( x 0 ) ) . \begin{aligned}
1=\frac{h_1\varphi(\ell/h_1)\bar{\sigma}_k(W[h_1])}{h_2\varphi(\ell/h_2)\bar{\sigma}_k(W[h_2])}(x_0) \\
&\leq c^{k+1}\frac{\varphi(\ell(x_0)/(ch_2(x_0)))}{\varphi(\ell(x_0)/h_2(x_0))}.
\end{aligned} 1 = h 2 φ ( ℓ / h 2 ) σ ˉ k ( W [ h 2 ]) h 1 φ ( ℓ / h 1 ) σ ˉ k ( W [ h 1 ]) ( x 0 ) ≤ c k + 1 φ ( ℓ ( x 0 ) / h 2 ( x 0 )) φ ( ℓ ( x 0 ) / ( c h 2 ( x 0 ))) .
The supercritical hypothesis implies
log φ ( ℓ ( x 0 ) / h 2 ( x 0 ) ) φ ( ℓ ( x 0 ) / ( c h 2 ( x 0 ) ) ) = ∫ ℓ ( x 0 ) / ( c h 2 ( x 0 ) ) ℓ ( x 0 ) / h 2 ( x 0 ) a φ ( s ) s d s ≥ ( k + 1 + δ ) log c . \log\frac{\varphi(\ell(x_0)/h_2(x_0))}{\varphi(\ell(x_0)/(ch_2(x_0)))}=
\int_{\ell(x_0)/(ch_2(x_0))}^{\ell(x_0)/h_2(x_0)}
\frac{a_\varphi(s)}{s}\,ds\geq(k+1+\delta)\log c. log φ ( ℓ ( x 0 ) / ( c h 2 ( x 0 ))) φ ( ℓ ( x 0 ) / h 2 ( x 0 )) = ∫ ℓ ( x 0 ) / ( c h 2 ( x 0 )) ℓ ( x 0 ) / h 2 ( x 0 ) s a φ ( s ) d s ≥ ( k + 1 + δ ) log c .
Hence
φ ( ℓ ( x 0 ) / ( c h 2 ( x 0 ) ) ) φ ( ℓ ( x 0 ) / h 2 ( x 0 ) ) ≤ c − ( k + 1 + δ ) \frac{\varphi(\ell(x_0)/(ch_2(x_0)))}{\varphi(\ell(x_0)/h_2(x_0))}\leq c^{-(k+1+\delta)} φ ( ℓ ( x 0 ) / h 2 ( x 0 )) φ ( ℓ ( x 0 ) / ( c h 2 ( x 0 ))) ≤ c − ( k + 1 + δ )
and consequently
1 ≤ c − δ < 1 , 1\leq c^{-\delta}<1, 1 ≤ c − δ < 1 ,
a contradiction. Thus h 1 ≤ h 2 h_1\leq h_2 h 1 ≤ h 2 . Interchanging the two solutions gives h 2 ≤ h 1 h_2\leq h_1 h 2 ≤ h 1 , and therefore h 1 = h 2 h_1=h_2 h 1 = h 2 .
Exponential convergence
Let
M ( t ) = max C θ Q ( ⋅ , t ) , m ( t ) = min C θ Q ( ⋅ , t ) . M(t)=\max_{\mathcal C_\theta}Q(\cdot,t),\qquad m(t)=\min_{\mathcal C_\theta}Q(\cdot,t). M ( t ) = C θ max Q ( ⋅ , t ) , m ( t ) = C θ min Q ( ⋅ , t ) .
From (23) , if M ( t ) > 1 M(t)>1 M ( t ) > 1 ,
D + ( M − 1 ) ≤ − δ M ( M − 1 ) ≤ − δ ( M − 1 ) . D^+(M-1)\leq-\delta M(M-1)\leq-\delta(M-1). D + ( M − 1 ) ≤ − δ M ( M − 1 ) ≤ − δ ( M − 1 ) .
If m ( t ) < 1 m(t)<1 m ( t ) < 1 , Proposition 3.3 gives m ( t ) ≥ c Q > 0 m(t)\ge c_Q>0 m ( t ) ≥ c Q > 0 and
D + ( 1 − m ) ≤ − δ c Q ( 1 − m ) . D^{+}(1-m)\le-\delta c_Q(1-m). D + ( 1 − m ) ≤ − δ c Q ( 1 − m ) .
Hence there are C , c > 0 C,c>0 C , c > 0 such that
∥ Q ( ⋅ , t ) − 1 ∥ L ∞ ( C θ ) ≤ C e − c t . (42) \|Q(\cdot,t)-1\|_{L^\infty(\mathcal{C}_\theta)}\le Ce^{-ct}. \tag*{(42)} ∥ Q ( ⋅ , t ) − 1 ∥ L ∞ ( C θ ) ≤ C e − c t . (42)
Since
∂ t log h = Q − 1 , \partial_t\log h=Q-1, ∂ t log h = Q − 1 ,
(42) implies
∫ 0 ∞ ∥ ∂ t log h ( ⋅ , t ) ∥ L ∞ d t < ∞ . \int_0^\infty\|\partial_t\log h(\cdot,t)\|_{L^\infty}\,dt<\infty. ∫ 0 ∞ ∥ ∂ t log h ( ⋅ , t ) ∥ L ∞ d t < ∞.
Therefore h ( ⋅ , t ) h(\cdot,t) h ( ⋅ , t ) converges uniformly to a positive function h ∞ h_\infty h ∞ . The time-uniform higher-order estimates from Subsection 5.1 imply precompactness in every C m C^m C m , and the uniqueness of the C 0 C^0 C 0 limit upgrades the convergence to
h ( ⋅ , t ) → h ∞ in C ∞ ( C ‾ θ ) . h(\cdot,t)\to h_\infty\quad\text{in }C^\infty(\overline{\mathcal{C}}_\theta). h ( ⋅ , t ) → h ∞ in C ∞ ( C θ ) .
Passing to the limit in Q → 1 Q\to1 Q → 1 gives
h ∞ φ ( ℓ / h ∞ ) σ ˉ k ( W [ h ∞ ] ) = f . h_\infty\varphi(\ell/h_\infty)\bar{\sigma}_k(W[h_\infty])=f. h ∞ φ ( ℓ / h ∞ ) σ ˉ k ( W [ h ∞ ]) = f .
The Robin condition passes to the limit as well. Proposition 5.4 shows that the limiting stationary solution is unique. This proves the convergence part of Theorem 1.1.
The all-angle scalar flow when k = 1 k=1 k = 1
We now prove Theorem 1.3. No convexity estimate is used in this section. For k = 1 k=1 k = 1 ,
σ ˉ 1 ( W [ h ] ) = 1 n ( Δ h + n h ) . \bar{\sigma}_1(W[h])=\frac{1}{n}(\Delta h+nh). σ ˉ 1 ( W [ h ]) = n 1 ( Δ h + nh ) .
Set u = h / ℓ u=h/\ell u = h / ℓ . By (18),
σ ˉ 1 ( W [ h ] ) = 1 n ( ℓ Δ u + 2 ⟨ ∇ ℓ , ∇ u ⟩ + n u ) . \bar{\sigma}_1(W[h])=\frac{1}{n}\left(\ell\Delta u+2\langle\nabla\ell,\nabla u\rangle+nu\right). σ ˉ 1 ( W [ h ]) = n 1 ( ℓ Δ u + 2 ⟨ ∇ ℓ , ∇ u ⟩ + n u ) .
Thus the flow becomes
u t = a ( x , u ) Δ u + b ( x , u ) ⋅ ∇ u + c ( x , u ) , u μ = 0 , (43) \begin{aligned}
u_t&=a(x,u)\Delta u+b(x,u)\cdot\nabla u+c(x,u),\\
u_\mu&=0, \tag*{(43)}
\end{aligned} u t u μ = a ( x , u ) Δ u + b ( x , u ) ⋅ ∇ u + c ( x , u ) , = 0 , (43)
where
a ( x , u ) = ℓ 2 u 2 φ ( 1 / u ) n f , (44) a(x,u)=\frac{\ell^2u^2\varphi(1/u)}{nf}, \tag*{(44)} a ( x , u ) = n f ℓ 2 u 2 φ ( 1/ u ) , (44)
b ( x , u ) = 2 ℓ u 2 φ ( 1 / u ) n f ∇ ℓ , b(x,u)=\frac{2\ell u^2\varphi(1/u)}{nf}\nabla\ell, b ( x , u ) = n f 2 ℓ u 2 φ ( 1/ u ) ∇ ℓ ,
c ( x , u ) = ℓ u 3 φ ( 1 / u ) f − u . c(x,u)=\frac{\ell u^3\varphi(1/u)}{f}-u. c ( x , u ) = f ℓ u 3 φ ( 1/ u ) − u .
For every fixed 0 < θ < π 0<\theta<\pi 0 < θ < π , (11) gives ℓ − > 0 \ell_->0 ℓ − > 0 .
The speed ratio and convergence
Uniform scalar bounds and global parabolicity
Define
g 1 ( s ) = s 2 φ ( 1 / s ) . g_1(s)=s^2\varphi(1/s). g 1 ( s ) = s 2 φ ( 1/ s ) .
Then
g 1 ′ ( s ) g 1 ( s ) = 2 − a φ ( 1 / s ) s ≤ − δ s . \frac{g_1'(s)}{g_1(s)}=\frac{2-a_\varphi(1/s)}{s}\leq-\frac{\delta}{s}. g 1 ( s ) g 1 ′ ( s ) = s 2 − a φ ( 1/ s ) ≤ − s δ .
Hence g 1 ( s ) → ∞ g_1(s)\to\infty g 1 ( s ) → ∞ as s ↓ 0 s\downarrow0 s ↓ 0 and g 1 ( s ) → 0 g_1(s)\to0 g 1 ( s ) → 0 as s ↑ ∞ s\uparrow\infty s ↑ ∞ .
At a spatial maximum of u u u , the drift term in (43) vanishes and Δ u ≤ 0 \Delta u\leq0 Δ u ≤ 0 . Therefore
D + u max ≤ u max [ ℓ max f min g 1 ( u max ) − 1 ] . D^+u_{\max}\leq u_{\max}\left[\frac{\ell_{\max}}{f_{\min}}g_1(u_{\max})-1\right]. D + u m a x ≤ u m a x [ f m i n ℓ m a x g 1 ( u m a x ) − 1 ] .
At a spatial minimum,
D + u min ≥ u min [ ℓ min f max g 1 ( u min ) − 1 ] . D_+u_{\min}\geq u_{\min}\left[\frac{\ell_{\min}}{f_{\max}}g_1(u_{\min})-1\right]. D + u m i n ≥ u m i n [ f m a x ℓ m i n g 1 ( u m i n ) − 1 ] .
The same argument applies at the boundary because u μ = 0 u_\mu=0 u μ = 0 . Consequently
0 < c ≤ u ≤ C for all times of existence. (45) 0<c\leq u\leq C\qquad\text{for all times of existence.} \tag*{(45)} 0 < c ≤ u ≤ C for all times of existence. (45)
By (44), (11), and (45),
0 < a 0 ≤ a ( x , u ) ≤ a 1 < ∞ . 0<a_0\leq a(x,u)\leq a_1<\infty. 0 < a 0 ≤ a ( x , u ) ≤ a 1 < ∞.
We now give the regularity argument for the scalar equation. Along the given solution, it can be written as the linear equation
u t − a ( x , u ) Δ u − b ( x , u ) ⋅ ∇ u = c ( x , u ) , u μ = 0. u_t-a(x,u)\Delta u-b(x,u)\cdot\nabla u=c(x,u),\qquad u_\mu=0. u t − a ( x , u ) Δ u − b ( x , u ) ⋅ ∇ u = c ( x , u ) , u μ = 0.
The bounds c 0 ≤ u ≤ C 0 c_0\leq u\leq C_0 c 0 ≤ u ≤ C 0 imply uniform ellipticity of a ( x , u ) a(x,u) a ( x , u ) , and uniform bounds for a ( x , u ) , b ( x , u ) , c ( x , u ) a(x,u),b(x,u),c(x,u) a ( x , u ) , b ( x , u ) , c ( x , u ) . At this stage, derivatives of these composed coefficients are not assumed bounded: only the derivatives of the functions a , b , c a,b,c a , b , c in ( x , z ) ∈ C ‾ θ × [ c 0 , C 0 ] (x,z)\in\overline{C}_\theta\times[c_0,C_0] ( x , z ) ∈ C θ × [ c 0 , C 0 ] are bounded. In Fermi coordinates near the boundary, extend u u u evenly across x n = 0 x_n=0 x n = 0 . Since u x n = 0 u_{x_n}=0 u x n = 0 , this is a classical C 2 C^2 C 2 extension. Reflect the principal matrix by A ↦ R A R A\mapsto RAR A ↦ R A R , the drift by b ↦ R b b\mapsto Rb b ↦ R b , and the source evenly, where R = diag ( 1 , … , 1 , − 1 ) R=\operatorname{diag}(1,\ldots,1,-1) R = diag ( 1 , … , 1 , − 1 ) . Uniform ellipticity and all required coefficient bounds are preserved. Interior parabolic Krylov–Safonov estimates for bounded measurable principal coefficients, bounded drift, and bounded source now yield a uniform C α , α / 2 C^{\alpha,\alpha/2} C α , α /2 bound for u u u on positive-time strips. No preliminary gradient estimate is required. Consequently the composed coefficients are uniformly C α , α / 2 C^{\alpha,\alpha/2} C α , α /2 . The linear parabolic Schauder estimate with smooth homogeneous Neumann boundary data yields a uniform C 2 + α , 1 + α / 2 C^{2+\alpha,1+\alpha/2} C 2 + α , 1 + α /2 bound on a smaller strip. Higher-order Neumann Schauder estimates, with the smooth dependence of a , b , c a,b,c a , b , c on ( x , u ) (x,u) ( x , u ) , give bounds for all spatial and time derivatives by induction. For each fixed τ > 0 \tau>0 τ > 0 , use nested backward strips of length bounded below by a positive number depending on τ \tau τ (and on the desired derivative order); this makes the constants independent of the position of the strip and of the maximal time. Thus, on every maximal smooth interval [ 0 , T max ) [0,T_{\max}) [ 0 , T m a x ) ,
sup t ∈ [ τ , T max ) ∥ u ( t ) ∥ C m ≤ C m , τ . (46) \sup_{t\in[\tau,T_{\max})}\|u(t)\|_{C^m}\leq C_{m,\tau}. \tag*{(46)} t ∈ [ τ , T m a x ) sup ∥ u ( t ) ∥ C m ≤ C m , τ . (46)
If T max < ∞ T_{\max}<\infty T m a x < ∞ , these estimates and the equation produce a smooth positive limiting jet at T max T_{\max} T m a x , preserving the compatibility hierarchy, and the short-time theorem extends the solution. This proves global existence. All constants may depend on the fixed angle; no uniformity as θ ↓ 0 \theta\downarrow0 θ ↓ 0 or θ ↑ π \theta\uparrow\pi θ ↑ π is asserted.
The speed ratio and convergence
For k = 1 k=1 k = 1 , Lemma 2.3 becomes
Q t − a ( x , h ) Δ Q = 2 a ( x , h ) ⟨ ∇ log h , ∇ Q ⟩ + ( 2 − a φ ( ℓ / h ) ) Q ( Q − 1 ) , Q_t-a(x,h)\Delta Q=2a(x,h)\langle\nabla\log h,\nabla Q\rangle+\left(2-a_{\varphi}(\ell/h)\right)Q(Q-1), Q t − a ( x , h ) Δ Q = 2 a ( x , h ) ⟨ ∇ log h , ∇ Q ⟩ + ( 2 − a φ ( ℓ / h ) ) Q ( Q − 1 ) ,
where
a ( x , h ) = Θ n > 0 , Q μ = 0. a(x,h)=\frac{\Theta}{n}>0,\qquad Q_{\mu}=0. a ( x , h ) = n Θ > 0 , Q μ = 0.
The initial assumption σ ˉ 1 ( W [ h 0 ] ) > 0 \bar{\sigma}_1(W[h_0])>0 σ ˉ 1 ( W [ h 0 ]) > 0 is exactly Q ( ⋅ , 0 ) > 0 Q(\cdot,0)>0 Q ( ⋅ , 0 ) > 0 . The maximum principle therefore gives
0 < c Q ≤ Q ≤ C Q . 0<c_Q\leq Q\leq C_Q. 0 < c Q ≤ Q ≤ C Q .
Since 2 − a φ ≤ − δ 2-a_{\varphi}\leq-\delta 2 − a φ ≤ − δ , the same logistic comparison as in Subsection 5.3 gives
∥ Q ( ⋅ , t ) − 1 ∥ L ∞ ≤ C e − c t . \lVert Q(\cdot,t)-1\rVert_{L^\infty}\leq Ce^{-ct}. ∥ Q ( ⋅ , t ) − 1 ∥ L ∞ ≤ C e − c t .
Because ∂ t log h = Q − 1 \partial_t\log h=Q-1 ∂ t log h = Q − 1 , log h ( ⋅ , t ) \log h(\cdot,t) log h ( ⋅ , t ) is Cauchy in C 0 C^0 C 0 and converges exponentially. The uniform estimates (46) and interpolation upgrade the convergence to C ∞ C^\infty C ∞ . Passing to the limit in the equation gives the stationary k = 1 k=1 k = 1 equation.
Finally, Proposition 5.4 gives uniqueness of the positive stationary solution with σ ˉ 1 ( W ) > 0 \bar{\sigma}_1(W)>0 σ ˉ 1 ( W ) > 0 . This proves Theorem 1.3 . Notice that nowhere in the scalar continuation argument did we use the sign of cot θ \cot\theta cot θ . In the obtuse range, however, this scalar continuation argument does not establish the separate geometric assertion that the full tensor W [ h ] W[h] W [ h ] stays positive definite.
The natural boundary condition and the contact-angle problem
The quantitative hypothesis (9) is a convenient sufficient condition for a time-uniform non-collapsing estimate, but it is stronger than the natural static condition
Under this weaker condition, Theorem 1.2 shows that strict convexity cannot be lost at any finite smooth time in the non-obtuse range. For a normal minimum-radius support v = h W ( μ , μ ) v=hW(\mu,\mu) v = hW ( μ , μ ) , the formulas in Lemma 4.2 , with k = 1 k=1 k = 1 treated separately, give
v μ + ( n − 1 ) cot θ v ≥ 0. v_{\mu}+(n-1)\cot\theta\,v\geq0. v μ + ( n − 1 ) cot θ v ≥ 0.
Thus the remaining issue under the natural boundary sign is quantitative rather than qualitative: one seeks a time-independent positive lower bound for q q q under this Robin sign alone.
The angle restriction is independent of this optimization. At the right angle, ℓ ≡ 1 \ell\equiv1 ℓ ≡ 1 and cot θ = 0 \cot\theta=0 cot θ = 0 , so all boundary signs used above become Neumann signs and the global theorem remains valid. In contrast, for π / 2 < θ < π \pi/2<\theta<\pi π /2 < θ < π , a tangential minimum-radius support satisfies
v μ = h cot θ λ n < 0. v_{\mu}=h\cot\theta\,\lambda_n<0. v μ = h cot θ λ n < 0.
This has the wrong sign for the boundary barriers used in the finite-time preservation and quantitative non-collapsing arguments. Hence neither proof closes by the present boundary mechanism. The obtuse-angle problem therefore requires a different boundary curvature estimate.
For k = 1 k = 1 k = 1 the scalar equation is uniformly parabolic after the C 0 C^0 C 0 estimate, and Theorem 1.3 gives an all-angle scalar flow for 0 < θ < π 0 < \theta< \pi 0 < θ < π . What remains open in the obtuse range is the geometric strengthening: preservation of the full matrix inequality W > 0 W > 0 W > 0 and hence an interpretation as a strictly convex capillary hypersurface for all times.
Acknowledgments: Hua Zhu was supported by the National Natural Science Foundation of China (Grant No. 12501273, 12661041) and the Research Foundation of Southwest University of Science and Technology (Grant No. 25zx7153, 25zx7154). Peihe Wang is supported by National Natural Science Foundation of Shandong Province(ZR2026MS0039).
Declaration on generative AI assistance
During the preparation of this manuscript, generative AI tools (GPT) were used to assist with language editing, literature searches, and the checking of mathematical arguments. The authors are responsible for the correctness of the mathematical arguments.
References