Introduction
We consider nonnegative weak solutions to the homogeneous doubly nonlinear equation
∂ t u p − 1 − div ( ∣ ∇ u ∣ p − 2 ∇ u ) = 0 in Ω T , (1) \partial_t u^{p-1}-\operatorname{div}\left(\lvert\nabla u\rvert^{p-2}\nabla u\right)=0 \quad\text{in }\Omega_T, \tag*{(1)} ∂ t u p − 1 − div ( ∣ ∇ u ∣ p − 2 ∇ u ) = 0 in Ω T , (1)
also known as Trudinger’s equation. Here p > 1 p>1 p > 1 , Ω ⊂ R N \Omega\subset\mathbb{R}^N Ω ⊂ R N is a bounded open set, T > 0 T>0 T > 0 , and Ω T : = Ω × ( 0 , T ] \Omega_T:=\Omega\times(0,T] Ω T := Ω × ( 0 , T ] .
Equation (1) is a benchmark in the nonlinear parabolic regularity theory: although nonlinear in both the solution and its spatial gradient, it is invariant under multiplication of the solution by positive constants. For p ≠ 2 p\ne2 p = 2 , however, invariance under the addition of constants is lost, creating a basic difficulty in the regularity analysis. While Hölder continuity is well established (see, e.g., [8 , 21 , 22 ] ), a central question is whether spatial gradient estimates remain uniform as the solution approaches zero.
Our first main result answers this question by establishing local Lipschitz continuity in space. The estimates do not depend on a positive lower bound for the solution and therefore apply also across its zero set.
Theorem 1.1 (Lipschitz estimates). Let u u u be a locally bounded, nonnegative, local weak solution to (1) in Ω T \Omega_T Ω T . Then, we have
∇ u ∈ L l o c ∞ ( Ω T , R N ) . \nabla u\in L^\infty_{\mathrm{loc}}(\Omega_T,\mathbb{R}^N). ∇ u ∈ L loc ∞ ( Ω T , R N ) .
Moreover, there is a constant C = C ( N , p ) C = C(N,p) C = C ( N , p ) such that for any cylinder Q 2 R ( z o ) ⋐ Ω T Q_{2R}(z_o) \Subset\Omega_T Q 2 R ( z o ) ⋐ Ω T there holds
∥ ∇ u ∥ L ∞ ( Q R ( z o ) ) ≤ C R ∥ u ∥ L ∞ ( Q 2 R ( z o ) ) . \|\nabla u\|_{L^\infty(Q_R(z_o))} \leq\frac{C}{R}\|u\|_{L^\infty(Q_{2R}(z_o))}. ∥∇ u ∥ L ∞ ( Q R ( z o )) ≤ R C ∥ u ∥ L ∞ ( Q 2 R ( z o )) .
We refer to Definition 2.1 below for the precise notion of a weak solution. The boundedness assumption in Theorem 1.1 is supported by the known regularity theory; see [18 , 29 ] . Next, we show that the spatial gradient is not only locally bounded but also locally Hölder continuous.
Theorem 1.2 (Gradient Hölder estimate). Under the hypotheses of Theorem 1.1 , the continuous representative of u u u is spatially differentiable at every point of Ω T \Omega_T Ω T and the spatial gradient is locally Hölder continuous for some Hölder exponent α = α ( N , p ) ∈ ( 0 , 1 ) \alpha= \alpha(N,p) \in(0,1) α = α ( N , p ) ∈ ( 0 , 1 ) . Moreover, there is a constant C = C ( N , p ) > 1 C = C(N,p) > 1 C = C ( N , p ) > 1 such that
∣ ∇ u ( x , t ) − ∇ u ( y , s ) ∣ ≤ C R ∥ u ∥ L ∞ ( Q 2 R ( z o ) ) ( ∣ x − y ∣ + ∣ t − s ∣ 1 p R ) α (2) |\nabla u(x,t)-\nabla u(y,s)| \leq\frac{C}{R}\|u\|_{L^\infty(Q_{2R}(z_o))}\left(\frac{|x-y|+|t-s|^{\frac{1}{p}}}{R}\right)^\alpha \tag*{(2)} ∣∇ u ( x , t ) − ∇ u ( y , s ) ∣ ≤ R C ∥ u ∥ L ∞ ( Q 2 R ( z o )) ( R ∣ x − y ∣ + ∣ t − s ∣ p 1 ) α (2) for all ( x , t ) , ( y , s ) ∈ Q R ( z o ) (x,t),(y,s) \in Q_R(z_o) ( x , t ) , ( y , s ) ∈ Q R ( z o ) . In addition, we have ∇ u = 0 \nabla u = 0 ∇ u = 0 on { u = 0 } \{u=0\} { u = 0 } .
The significance of these results lies in the uniformity of the gradient estimates as the solution approaches zero. For strictly positive solutions, regularity estimates may depend on a positive lower bound for the solution and therefore provide no uniform control when that bound vanishes. Removing this dependence addresses a structural difficulty: the coefficient ( p − 1 ) u p − 2 (p-1)u^{p-2} ( p − 1 ) u p − 2 of the time derivative degenerates or becomes singular at zero, while the lack of invariance under the addition of constants prevents a direct reduction to the strictly positive case.
Theorems 1.1 and 1.2 establish quantitative first-order spatial regularity throughout the all nonnegative solutions, with estimates that remain valid near the zero set. They treat the homogeneous endpoint q = p − 1 q=p-1 q = p − 1 of the general equation,
∂ t u q = Δ p u , q > 0 , (3) \partial_t u^q = \Delta_p u,\qquad q>0, \tag*{(3)} ∂ t u q = Δ p u , q > 0 , (3)
which lies outside the gradient regularity theory developed in [5 , 6 ] for q > p − 1 q>p-1 q > p − 1 . Thus the results extend the regularity theory of Trudinger’s equation from Hölder continuity of the solution to the gradient level, without a quantitative positivity assumption.
Previous regularity results
The regularity theory for (1) originates in the work of Trudinger [29 ] , who established a parabolic Harnack inequality. Alternative approaches to Harnack’s inequality for nonnegative weak solutions were subsequently developed by Gianazza and Vespri [13 ] and by Kinnunen and Kuusi [18 ] .
Hölder continuity for the broader family of doubly nonlinear parabolic equations (3) was studied in various ranges of p p p and q q q in the early 1990s by Chen, Ivanov, Porzio, Vespri, and Zhou [10 , 15 , 16 , 17 , 25 , 3 ] . Such theory is still fragmented. For Trudinger’s equation, Kuusi, Laleoglu, Siljander, and Urbano [21 , 22 ] established local Hölder continuity of nonnegative weak solutions using Harnack’s inequality. Interior and boundary Hölder continuity for sign-changing solutions was later obtained by Bögel̈ein, Duzaar, and Liao [8 ] .
The regularity of the spatial gradient presents further difficulties. An earlier approach was proposed in [28 ] ; however, the proof of [28 Lemma 3.3] contains an oversight in the treatment of the time-derivative term, already in the case of Lebesgue measure.
For the broader equation (3) , Bögel̈ein, Duzaar, Gianazza, Liao, and Scheven [5 , 6 ] established local Hölder continuity of the spatial gradient in a range with q > p − 1 q>p-1 q > p − 1 . Their argument relies on a time-insensitive Harnack inequality and does not extend to the borderline case q = p − 1 q=p-1 q = p − 1 corresponding to Trudinger’s equation.
For Trudinger’s equation itself, Lindqvist, Parviainen, and Siltakoski [24 ] proved local Lipschitz regularity when solutions are strictly positive by combining the Ishii–Lions method with the equivalence of positive weak and viscosity solutions. The resulting Lipschitz estimate depends on a positive lower bound for the solution and deteriorates as that bound tends to zero; see [24 Corollary 3.2] . Lipschitz regularity under such a positivity assumption also holds for the more general equation (3) ; see [6 Theorem 6.8] .
A related line of research concerns higher integrability of the spatial gradient. Bögelein, Duzaar, Kinnunen, and Scheven [7 ] established a Gehring-type self-improvement in the range max { 1 , 2 N N + 2 } < p < 2 N ( N − 2 ) ( N − 2 ) + \max\left\{1,\frac{2N}{N+2}\right\}<p<\frac{2N(N-2)}{(N-2)_{+}} max { 1 , N + 2 2 N } < p < ( N − 2 ) + 2 N ( N − 2 ) . Their result applies to systems and uses an intrinsic geometry depending on both the solution and its spatial gradient. For nonnegative scalar solutions, Saari and Schwarzacher [26 ] subsequently removed the upper restriction on p p p , obtaining higher integrability throughout the range p ≥ 2 p\ge2 p ≥ 2 .
The central issue addressed here is pointwise gradient regularity as the solution approaches zero. Our main results, Theorems 1.1 and 1.2 , establish gradient estimates for nonnegative weak solutions that remain uniform near the zero set, without assuming a positive lower bound for the solution.
Ideas of the proof
The argument is developed within the weak formulation and yields quantitative estimates for the spatial gradient. The proof builds on the structure of the positivity set: by the expansion of positivity for Trudinger’s equation, a nonnegative solution may vanish identically up to a time t ∗ t_* t ∗ , while it is strictly positive thereafter. Moreover, comparison with an exponential barrier shows that, for a short time after t ∗ t_* t ∗ , the solution remains exponentially small in the interior; see § 6.1 .
The key ingredient is a pointwise gradient estimate for a suitable power of the solution on its positivity set. Writing u = v k u=v^k u = v k with k > 1 k>1 k > 1 sufficiently large, we employ a localized Bernstein argument to control ∇ v \nabla v ∇ v independently of any positive lower bound for u u u . To give a glimpse into the machinery, consider the linear case p = 2 p=2 p = 2 , w = ∣ ∇ v ∣ 2 w=|\nabla v|^2 w = ∣∇ v ∣ 2 will satisfy
( ∂ t − Δ − 2 ( k − 1 ) ∇ v v ⋅ ∇ ) w = − 2 ∣ ∇ 2 v ∣ 2 − 2 ( k − 1 ) w 2 v 2 . \left(\partial_t-\Delta-2(k-1)\frac{\nabla v}{v}\cdot\nabla\right)w
=-2|\nabla^2v|^2-2(k-1)\frac{w^2}{v^2}. ( ∂ t − Δ − 2 ( k − 1 ) v ∇ v ⋅ ∇ ) w = − 2∣ ∇ 2 v ∣ 2 − 2 ( k − 1 ) v 2 w 2 .
At an interior maximum of w w w , the drift term disappears, while the negative zeroth-order term prevents w w w from becoming too large, suggesting an estimate of the form w ≲ v 2 w\lesssim v^2 w ≲ v 2 . Notice that this mechanism already requires k > 1 k>1 k > 1 for the heat equation.
The same idea extends, in a considerably more delicate form, to Trudinger’s equation. Its homogeneity is essential, since it allows the power k k k to be chosen so that the favorable gradient contribution absorbs the terms generated by differentiating the diffusion coefficient. Indeed, for the general equation (3) the transformation u = v k u=v^k u = v k gives
∂ t v = k p − 2 q v σ Δ p v + k p − 2 q ( p − 1 ) ( k − 1 ) v σ − 1 ∣ ∇ v ∣ p , σ = 2 − p + k ( p − 1 − q ) . \partial_t v=\frac{k^{p-2}}{q}v^\sigma\Delta_pv+\frac{k^{p-2}}{q}(p-1)(k-1)v^{\sigma-1}|\nabla v|^p,\qquad\sigma=2-p+k(p-1-q). ∂ t v = q k p − 2 v σ Δ p v + q k p − 2 ( p − 1 ) ( k − 1 ) v σ − 1 ∣∇ v ∣ p , σ = 2 − p + k ( p − 1 − q ) .
Precisely in the Trudinger case q = p − 1 q=p-1 q = p − 1 , the exponent σ \sigma σ is independent of k k k .
Since the Bernstein computation involves third-order derivatives, we carry it out for a suitable regularization and then pass to the limit. For k k k sufficiently large and on every cylinder B 2 r ( x 0 ) × ( s − 2 h , s ] B_{2r}(x_0)\times(s-2h,s] B 2 r ( x 0 ) × ( s − 2 h , s ] with 0 < m ≤ u ≤ M 0<m\le u\le M 0 < m ≤ u ≤ M , this yields the gradient bound
∥ ∇ v ∥ L ∞ ( B r ( x 0 ) × ( s − h , s ] ) ≲ M 1 k ( 1 r + 1 h 1 / p ) . (4) \lVert\nabla v\rVert_{L^\infty(B_r(x_0)\times(s-h,s])}\lesssim M^{\frac{1}{k}}\left(\frac{1}{r}+\frac{1}{h^{1/p}}\right). \tag*{(4)} ∥ ∇ v ∥ L ∞ ( B r ( x 0 ) × ( s − h , s ]) ≲ M k 1 ( r 1 + h 1/ p 1 ) . (4)
with a constant independent of the lower bound m m m . Returning to the original variables gives local spatial Lipschitz continuity together with quantitative decay of ∇ u \nabla u ∇ u as u u u approaches zero. Combined with the exponential decay near the zero set, this allows the estimates on the positivity set and the zero set to be patched together into the local gradient bound stated in Theorem 1.1 .
The estimate in (4) is in fact stronger than the gradient bound asserted in Theorem 1.1 . As stated in Theorem 6.3 , it yields, since u = v k u=v^k u = v k with k > 1 k>1 k > 1 , the quantitative decay
∣ ∇ u ∣ ≲ u 1 − 1 k |\nabla u|\lesssim u^{1-\frac{1}{k}} ∣∇ u ∣ ≲ u 1 − k 1
as u u u approaches zero. Combining this decay near the zero set with the local gradient estimates in the positivity set provided by the Schauder theory in [6 ] , we obtain the Hölder continuity of the gradient stated in Theorem 1.2 .
To the best of our knowledge, this approach is new. Related ideas, however, appear in earlier work. Kotschwar and Ni [20 ] use a logarithmic transformation, a nonlinear Bochner identity, and the maximum principle for Trudinger’s equation, but their estimates involve different quantities, in particular combinations of spatial gradients and time derivatives. For a related p p p -Laplace equation with gradient absorption, Bartier and Laurençot [2 ] employ a power transformation and a Bernstein-type argument in which the Hessian terms are controlled by completing squares.
Notation and preliminaries
For a point z o = ( x o , t o ) ∈ R N × R z_o=(x_o,t_o)\in\mathbb{R}^{N}\times\mathbb{R} z o = ( x o , t o ) ∈ R N × R in space-time and a radius ϱ > 0 \varrho>0 ϱ > 0 we consider parabolic cylinders of the form
Q ϱ ( z o ) : = B ϱ ( x o ) × ( t o − ϱ p , t o ] . Q_{\varrho}(z_o):=B_{\varrho}(x_o)\times(t_o-\varrho^p,t_o]. Q ϱ ( z o ) := B ϱ ( x o ) × ( t o − ϱ p , t o ] .
where B ϱ ( x o ) ⊂ R N B_{\varrho}(x_o)\subset\mathbb{R}^{N} B ϱ ( x o ) ⊂ R N denotes the open ball with radius ϱ > 0 \varrho>0 ϱ > 0 and center x o x_o x o . Occasionally, we also use more general cylinders of the type
Q r , s ( z o ) : = B r ( x o ) × ( t o − s , t o ] Q_{r,s}(z_o):=B_r(x_o)\times(t_o-s,t_o] Q r , s ( z o ) := B r ( x o ) × ( t o − s , t o ]
with r , s > 0 r,s>0 r , s > 0 . The parabolic boundary of a cylindrical domain Ω T : = Ω × ( 0 , T ] \Omega_T:=\Omega\times(0,T] Ω T := Ω × ( 0 , T ] is given by
∂ p a r Ω T : = ( Ω ‾ × { 0 } ) ∪ ( ∂ Ω × ( 0 , T ] ) . \partial_{\mathrm{par}}\Omega_T:=\bigl(\overline{\Omega}\times\{0\}\bigr)\cup\bigl(\partial\Omega\times(0,T]\bigr). ∂ par Ω T := ( Ω × { 0 } ) ∪ ( ∂ Ω × ( 0 , T ] ) .
For two points z 1 = ( x 1 , t 1 ) z_1=(x_1,t_1) z 1 = ( x 1 , t 1 ) , z 2 = ( x 2 , t 2 ) ∈ R N + 1 z_2=(x_2,t_2)\in\mathbb{R}^{N+1} z 2 = ( x 2 , t 2 ) ∈ R N + 1 their parabolic distance is defined as
dist p a r ( z 1 , z 2 ) : = ∣ x 1 − x 2 ∣ + ∣ t 1 − t 2 ∣ 1 p . \operatorname{dist}_{\mathrm{par}}(z_1,z_2):=|x_1-x_2|+|t_1-t_2|^{\frac{1}{p}}. dist par ( z 1 , z 2 ) := ∣ x 1 − x 2 ∣ + ∣ t 1 − t 2 ∣ p 1 .
The spatial gradient of u ( x , t ) u(x,t) u ( x , t ) is denoted by ∇ u \nabla u ∇ u , whereas the Hessian matrix is ∇ 2 u \nabla^2u ∇ 2 u . We also use some operators of matrices:
A ⊗ B : = ( a i b j ) N × N for A = ( a 1 , … , a N ) , B = ( b 1 , … , b N ) A\otimes B:=(a_i b_j)_{N\times N}\quad\text{for }A=(a_1,\ldots,a_N),\ B=(b_1,\ldots,b_N) A ⊗ B := ( a i b j ) N × N for A = ( a 1 , … , a N ) , B = ( b 1 , … , b N )
and
A ∙ B : = ∑ 1 ≤ i , j ≤ N a i j b i j for A = ( a i j ) N × N , B = ( b i j ) N × N . A\mathbin{\bullet}B:=\sum_{1\leq i,j\leq N}a_{ij}b_{ij}\quad\text{for }A=(a_{ij})_{N\times N},\ B=(b_{ij})_{N\times N}. A ∙ B := 1 ≤ i , j ≤ N ∑ a ij b ij for A = ( a ij ) N × N , B = ( b ij ) N × N .
Definition 2.1. A function
u ∈ C ( 0 , T ; L l o c p ( E ) ) ∩ L l o c p ( 0 , T ; W l o c 1 , p ( E ) ) u\in C(0,T;L^p_{\mathrm{loc}}(E))\cap L^p_{\mathrm{loc}}(0,T;W^{1,p}_{\mathrm{loc}}(E)) u ∈ C ( 0 , T ; L loc p ( E )) ∩ L loc p ( 0 , T ; W loc 1 , p ( E )) is a local, weak solution to (1.1), if the identity
∬ Ω T [ u p − 1 φ t − ∣ ∇ u ∣ p − 2 ∇ u ⋅ ∇ φ ] d x d t = 0 \iint_{\Omega_T}\left[u^{p-1}\varphi_t-|\nabla u|^{p-2}\nabla u\cdot\nabla\varphi\right]\,dx\,dt=0 ∬ Ω T [ u p − 1 φ t − ∣∇ u ∣ p − 2 ∇ u ⋅ ∇ φ ] d x d t = 0 holds, for any testing function φ ∈ C 0 ∞ ( Ω T ) \varphi\in C^\infty_0(\Omega_T) φ ∈ C 0 ∞ ( Ω T ) .
Lemma 2.2 (Scaling invariance). Let ϱ , M > 0 \varrho,M>0 ϱ , M > 0 and a < b a<b a < b . Suppose that u u u is a nonnegative weak solution to (1.1) in
B ϱ ( x o ) × ( t ∗ + a ϱ p , t ∗ + b ϱ p ] B_{\varrho}(x_o)\times(t_*+a\varrho^p,t_*+b\varrho^p] B ϱ ( x o ) × ( t ∗ + a ϱ p , t ∗ + b ϱ p ] and satisfies 0 ≤ u ≤ M 0\leq u\leq M 0 ≤ u ≤ M there. Then
v ( y , τ ) : = 1 M u ( x o + ϱ y , t ∗ + ϱ p τ ) v(y,\tau):=\frac{1}{M}u(x_o+\varrho y,t_*+\varrho^p\tau) v ( y , τ ) := M 1 u ( x o + ϱ y , t ∗ + ϱ p τ ) is a nonnegative weak solution to (1.1) in B 1 × ( a , b ) B_1\times(a,b) B 1 × ( a , b ) and satisfies 0 ≤ v ≤ 1 0\leq v\leq1 0 ≤ v ≤ 1 .
The boundedness and Hölder continuity of weak solutions was proved in [8 , 21 , 22 ] .
Proposition 2.3. Let u u u be a nonnegative weak solution to (1.1) in Q 2 R ( z o ) Q_{2R}(z_o) Q 2 R ( z o ) . Then, u ∈ C l o c η ( Q 2 R ( z o ) ) u \in C^\eta_{\mathrm{loc}}(Q_{2R}(z_o)) u ∈ C loc η ( Q 2 R ( z o )) for some η = η ( N , p ) ∈ ( 0 , 1 ) \eta=\eta(N,p)\in(0,1) η = η ( N , p ) ∈ ( 0 , 1 ) , and there exists C = C ( N , p ) > 1 C=C(N,p)>1 C = C ( N , p ) > 1 , such that for all ϱ ∈ ( 0 , R ) \varrho\in(0,R) ϱ ∈ ( 0 , R ) ,
osc Q ϱ ( z o ) u ≤ C ( ϱ R ) η sup Q R ( z o ) u . \underset{Q_{\varrho}(z_o)}{\operatorname{osc}} u\leq C\left(\frac{\varrho}{R}\right)^\eta\sup_{Q_R(z_o)}u. Q ϱ ( z o ) osc u ≤ C ( R ϱ ) η Q R ( z o ) sup u .
Proposition 2.4 (Strong maximum principle). Let Ω ⊂ R N \Omega\subset\mathbb{R}^{N} Ω ⊂ R N be open and connected, and let u u u be the continuous representative of a nonnegative weak solution to (1.1) in Ω T : = Ω × ( 0 , T ] \Omega_T:=\Omega\times(0,T] Ω T := Ω × ( 0 , T ] . Then there exists t ∗ ∈ [ 0 , T ] t_*\in[0,T] t ∗ ∈ [ 0 , T ] such that
u ( x , t ) = 0 for all ( x , t ) ∈ Ω × ( 0 , t ∗ ] , u(x,t)=0\qquad\text{for all }(x,t)\in\Omega\times(0,t_*], u ( x , t ) = 0 for all ( x , t ) ∈ Ω × ( 0 , t ∗ ] , and
u ( x , t ) > 0 for all ( x , t ) ∈ Ω × ( t ∗ , T ] . u(x,t)>0\qquad\text{for all }(x,t)\in\Omega\times(t_*,T]. u ( x , t ) > 0 for all ( x , t ) ∈ Ω × ( t ∗ , T ] . In particular, t ∗ = 0 t_*=0 t ∗ = 0 corresponds to u > 0 u>0 u > 0 throughout Ω T \Omega_T Ω T , whereas t ∗ = T t_*=T t ∗ = T corresponds to u ≡ 0 u\equiv0 u ≡ 0 throughout Ω T \Omega_T Ω T .
Proof. We extend u u u up to time 2 T 2T 2 T by solving (1.1) on Ω × ( T , 2 T ] \Omega\times(T,2T] Ω × ( T , 2 T ] with initial datum u ( ⋅ , T ) u(\cdot,T) u ( ⋅ , T ) and zero lateral boundary data. Pick a point x ∗ ∈ Ω x_*\in\Omega x ∗ ∈ Ω and let t ∗ t_* t ∗ be the supremum of { 0 } ∪ { t ∈ ( 0 , T ] : u ( x ∗ , t ) = 0 } \{0\}\cup\{t\in(0,T]:u(x_*,t)=0\} { 0 } ∪ { t ∈ ( 0 , T ] : u ( x ∗ , t ) = 0 } . It suffices to show that if u ( x ∗ , t ∗ ) = 0 u(x_*,t_*)=0 u ( x ∗ , t ∗ ) = 0 , then
u ≡ 0 in Ω × ( 0 , t ∗ ] . u\equiv0\qquad\text{in }\Omega\times(0,t_*]. u ≡ 0 in Ω × ( 0 , t ∗ ] .
Indeed, suppose that u ( x ~ , t ~ ) = 0 u(\widetilde{x},\widetilde{t})=0 u ( x , t ) = 0 for some x ~ ∈ Ω \widetilde{x}\in\Omega x ∈ Ω with x ~ ≠ x ∗ \widetilde{x}\ne x_* x = x ∗ and some t ~ > t ∗ \widetilde{t}>t_* t > t ∗ . Then, by the argument below, u ( x ∗ , t ) = 0 u(x_*,t)=0 u ( x ∗ , t ) = 0 for all t ∗ < t ≤ t ~ t_*<t\leq\widetilde{t} t ∗ < t ≤ t , contradicting the assumption that t ∗ t_* t ∗ is the supremum of the times at which u ( x ∗ , ⋅ ) u(x_*,\cdot) u ( x ∗ , ⋅ ) vanishes. We use the following expansion of positivity, cf. [8 Proposition B.1] . There exist constants η ∈ ( 0 , 1 ) \eta\in(0,1) η ∈ ( 0 , 1 ) , θ > 0 \theta>0 θ > 0 , and Λ > max { 2 , θ } \Lambda>\max\{2,\theta\} Λ > max { 2 , θ } , depending only on N N N and p p p , such that the following holds. If r > 0 r>0 r > 0 and
B Λ r ( y ) ‾ × [ s , s + Λ r p ] ⊂ Ω T , \overline{B_{\Lambda r}(y)}\times[s,s+\Lambda r^p]\subset\Omega_T, B Λ r ( y ) × [ s , s + Λ r p ] ⊂ Ω T ,
and if, for some ℓ > 0 \ell>0 ℓ > 0 ,
u ( x , s ) ≥ ℓ for every x ∈ B r ( y ) , u(x,s)\geq\ell\qquad\text{for every }x\in B_r(y), u ( x , s ) ≥ ℓ for every x ∈ B r ( y ) ,
then
u ( x , s + θ r p ) ≥ η ℓ for every x ∈ B 2 r ( y ) . (5) u(x,s+\theta r^p)\geq\eta\ell\qquad\text{for every }x\in B_{2r}(y). \tag*{(5)} u ( x , s + θ r p ) ≥ η ℓ for every x ∈ B 2 r ( y ) . (5)
Suppose, towards a contradiction, that u ( y , s ) > 0 u(y,s)>0 u ( y , s ) > 0 for some y ∈ Ω y\in\Omega y ∈ Ω and 0 < s < t ∗ 0<s<t_* 0 < s < t ∗ . We will propagate this positivity along a chain of balls to obtain u ( x ∗ , t ∗ ) > 0 u(x_*,t_*)>0 u ( x ∗ , t ∗ ) > 0 , contrary to our assumption. By continuity, there exist ℓ > 0 \ell>0 ℓ > 0 and ϱ > 0 \varrho>0 ϱ > 0 , with B ϱ ( y ) ‾ ⊂ Ω \overline{B_\varrho(y)}\subset\Omega B ϱ ( y ) ⊂ Ω , such that
u ( x , s ) ≥ ℓ for every x ∈ B ϱ ( y ) . (6) u(x,s)\geq\ell\qquad\text{for every }x\in B_\varrho(y). \tag*{(6)} u ( x , s ) ≥ ℓ for every x ∈ B ϱ ( y ) . (6)
No quantitative bounds for ℓ \ell ℓ or ϱ \varrho ϱ are needed. Since Ω \Omega Ω is open and connected, there exists a polygonal path γ : [ 0 , 1 ] → Ω \gamma:[0,1]\to\Omega γ : [ 0 , 1 ] → Ω joining y y y to x ∗ x_* x ∗ . If y ≠ x ∗ y\ne x_* y = x ∗ , we parametrize γ \gamma γ proportionally to arc length, so that
γ ( 0 ) = y , γ ( 1 ) = x ∗ , ∣ γ ( a ) − γ ( b ) ∣ ≤ L ∣ a − b ∣ for a , b ∈ [ 0 , 1 ] , \gamma(0)=y,\qquad\gamma(1)=x_*,\qquad|\gamma(a)-\gamma(b)|\leq L|a-b|\qquad\text{for }a,b\in[0,1], γ ( 0 ) = y , γ ( 1 ) = x ∗ , ∣ γ ( a ) − γ ( b ) ∣ ≤ L ∣ a − b ∣ for a , b ∈ [ 0 , 1 ] ,
where L L L denotes the length of the path. If y = x ∗ y=x_* y = x ∗ , we take the constant path and set L = 0 L=0 L = 0 . The image γ ( [ 0 , 1 ] ) \gamma([0,1]) γ ([ 0 , 1 ]) is a compact subset of Ω \Omega Ω . Hence, we can choose d > 0 d>0 d > 0 such that its closed d d d -neighborhood satisfies
{ x ∈ R N : dist ( x , γ ( [ 0 , 1 ] ) ) ≤ d } ⊂ Ω . \{x\in\mathbb{R}^{N}:\operatorname{dist}(x,\gamma([0,1]))\leq d\}\subset\Omega. { x ∈ R N : dist ( x , γ ([ 0 , 1 ])) ≤ d } ⊂ Ω.
Set Δ : = t ∗ − s > 0 \Delta:=t_*-s>0 Δ := t ∗ − s > 0 . For an integer m ≥ 1 m\geq1 m ≥ 1 , to be chosen below, define
r m : = ( Δ θ m ) 1 p , τ j : = s + j Δ m , z j : = γ ( j m ) , j = 0 , … , m . r_m:=\left(\frac{\Delta}{\theta m}\right)^{\frac{1}{p}},\qquad\tau_j:=s+\frac{j\Delta}{m},\qquad z_j:=\gamma\left(\frac{j}{m}\right),\qquad j=0,\ldots,m. r m := ( θ m Δ ) p 1 , τ j := s + m j Δ , z j := γ ( m j ) , j = 0 , … , m .
Thus, for j = 0 , … , m − 1 j=0,\ldots,m-1 j = 0 , … , m − 1 ,
τ j + 1 − τ j = θ r m p , ∣ z j + 1 − z j ∣ ≤ L m . \tau_{j+1}-\tau_j=\theta r_m^p,\qquad|z_{j+1}-z_j|\le\frac{L}{m}. τ j + 1 − τ j = θ r m p , ∣ z j + 1 − z j ∣ ≤ m L .
Since p > 1 p>1 p > 1 , we have
L / m r m = L ( θ Δ ) 1 p m 1 p − 1 ⟶ 0 as m → ∞ . \frac{L/m}{r_m}=L\left(\frac{\theta}{\Delta}\right)^{\frac{1}{p}}m^{\frac{1}{p}-1}\longrightarrow0\qquad\text{as }m\to\infty. r m L / m = L ( Δ θ ) p 1 m p 1 − 1 ⟶ 0 as m → ∞.
Moreover, also r m → 0 r_m\to0 r m → 0 as m → ∞ m\to\infty m → ∞ . Therefore, we can choose a finite integer m m m sufficiently large such that
L m ≤ r m , r m < ϱ , 2 Λ r m < d , Λ r m p < 2 T − t ∗ . (7) \frac{L}{m}\le r_m,\qquad r_m<\varrho,\qquad2\Lambda r_m<d,\qquad\Lambda r_m^p<2T-t_*. \tag*{(7)} m L ≤ r m , r m < ϱ , 2Λ r m < d , Λ r m p < 2 T − t ∗ . (7)
Fix such an m m m and write r : = r m r:=r_m r := r m . These choices ensure that every enlarged cylinder
B Λ r ( z j ) ‾ × [ τ j , τ j + Λ r p ] , j = 0 , … , m − 1 , \overline{B_{\Lambda r}(z_j)}\times[\tau_j,\tau_j+\Lambda r^p],\qquad j=0,\ldots,m-1, B Λ r ( z j ) × [ τ j , τ j + Λ r p ] , j = 0 , … , m − 1 ,
is contained in Ω 2 T \Omega_{2T} Ω 2 T . Indeed, each z j z_j z j lies on the path and Λ r < d \Lambda r<d Λ r < d , so B Λ r ( z j ) ‾ ⊂ Ω \overline{B_{\Lambda r}(z_j)}\subset\Omega B Λ r ( z j ) ⊂ Ω . Moreover,
0 < s ≤ τ j , τ j + Λ r p ≤ t ∗ + Λ r p < 2 T . 0<s\le\tau_j,\qquad\tau_j+\Lambda r^p\le t_*+\Lambda r^p<2T. 0 < s ≤ τ j , τ j + Λ r p ≤ t ∗ + Λ r p < 2 T .
The last application of the expansion of positivity starts at τ m − 1 = t ∗ − θ r p \tau_{m-1}=t_*-\theta r^p τ m − 1 = t ∗ − θ r p and yields positivity at t ∗ t_* t ∗ . However, it requires the solution to be defined up to τ m − 1 + Λ r p = t ∗ + ( Λ − θ ) r p > t ∗ \tau_{m-1}+\Lambda r^p=t_*+(\Lambda-\theta)r^p>t_* τ m − 1 + Λ r p = t ∗ + ( Λ − θ ) r p > t ∗ . The last inequality in (7) ensures that this time is still less than 2 T 2T 2 T .
We now prove by induction that
u ( x , τ j ) ≥ η j ℓ for every x ∈ B r ( z j ) , j = 0 , … , m . (8) u(x,\tau_j)\ge\eta^j\ell\qquad\text{for every }x\in B_r(z_j),\quad j=0,\ldots,m. \tag*{(8)} u ( x , τ j ) ≥ η j ℓ for every x ∈ B r ( z j ) , j = 0 , … , m . (8)
For j = 0 j=0 j = 0 , we have z 0 = y z_0=y z 0 = y and τ 0 = s \tau_0=s τ 0 = s , so the assertion follows from (6) and r < ϱ r<\varrho r < ϱ .
Suppose that (8) holds for some j ∈ { 0 , … , m − 1 } j\in\{0,\ldots,m-1\} j ∈ { 0 , … , m − 1 } . Applying (5) with center z j z_j z j , initial time τ j \tau_j τ j , and lower bound η j ℓ \eta^j\ell η j ℓ , we obtain
u ( x , τ j + 1 ) ≥ η j + 1 ℓ for every x ∈ B 2 r ( z j ) , u(x,\tau_{j+1})\ge\eta^{j+1}\ell\qquad\text{for every }x\in B_{2r}(z_j), u ( x , τ j + 1 ) ≥ η j + 1 ℓ for every x ∈ B 2 r ( z j ) ,
where we used τ j + 1 = τ j + θ r p \tau_{j+1}=\tau_j+\theta r^p τ j + 1 = τ j + θ r p . Moreover, (7) gives ∣ z j + 1 − z j ∣ ≤ r |z_{j+1}-z_j|\le r ∣ z j + 1 − z j ∣ ≤ r , and hence B r ( z j + 1 ) ⊂ B 2 r ( z j ) ⊂ Ω B_r(z_{j+1})\subset B_{2r}(z_j)\subset\Omega B r ( z j + 1 ) ⊂ B 2 r ( z j ) ⊂ Ω . Thus, (8) holds for j + 1 j+1 j + 1 , which completes the induction.
Since z m = x ∗ z_m=x_* z m = x ∗ and τ m = t ∗ \tau_m=t_* τ m = t ∗ , we conclude that
u ( x ∗ , t ∗ ) ≥ η m ℓ > 0 , u(x_*,t_*)\ge\eta^m\ell>0, u ( x ∗ , t ∗ ) ≥ η m ℓ > 0 ,
contrary to u ( x ∗ , t ∗ ) = 0 u(x_*,t_*)=0 u ( x ∗ , t ∗ ) = 0 . The chain contains only finitely many steps, so the lower bound η m ℓ \eta^m\ell η m ℓ remains strictly positive; no passage to the limit as m → ∞ m\to\infty m → ∞ is involved. It follows that u ≡ 0 u\equiv0 u ≡ 0 in Ω × ( 0 , t ∗ ) \Omega\times(0,t_*) Ω × ( 0 , t ∗ ) . By continuity, u ( ⋅ , t ∗ ) = 0 u(\cdot,t_*)=0 u ( ⋅ , t ∗ ) = 0 as well. □
Gradient bounds for smooth solutions
This section treats smooth positive solutions u u u to (1) and establishes an upper bound for the gradient of u 1 k u^{\frac{1}{k}} u k 1 , for suitable k k k . For the reader’s convenience, we outline the four steps of the argument:
(a) the power transformation u = v k u=v^k u = v k turns (1) into
v t = a ( v ) Δ p v + b ( v ) ∣ ∇ v ∣ p ; v_t=a(v)\Delta_pv+b(v)|\nabla v|^p; v t = a ( v ) Δ p v + b ( v ) ∣∇ v ∣ p ;
where a a a and b b b are defined in (11) ;
(b) differentiating in x x x gives a Bochner-type identity
L w = − 2 a w p − 2 2 tr ( P ( ∇ 2 v ) 2 ) + 2 a ′ w p 2 P ⋅ ∇ 2 v + 2 b ′ w p + 2 2 \mathcal{L}w=-2a w^{\frac{p-2}{2}}\operatorname{tr}\left(\mathbf{P}(\nabla^2v)^2\right)+2a'w^{\frac{p}{2}}\mathbf{P}\cdot\nabla^2v+2b'w^{\frac{p+2}{2}} L w = − 2 a w 2 p − 2 tr ( P ( ∇ 2 v ) 2 ) + 2 a ′ w 2 p P ⋅ ∇ 2 v + 2 b ′ w 2 p + 2
for w = ∣ ∇ v ∣ 2 w=|\nabla v|^2 w = ∣∇ v ∣ 2 , where the operator L \mathcal{L} L is defined below in (19) . The first-order term, with b ′ < 0 b'<0 b ′ < 0 , is the only source of coercivity;
Figure 1. Schematic space-time representation of the positivity chain for m = 5 m=5 m = 5 . The horizontal disks represent spatial balls at successive times, with τ j + 1 − τ j = θ r p \tau_{j+1}-\tau_j=\theta r^p τ j + 1 − τ j = θ r p . The dotted lines project their centers onto the spatial path. For clarity, the d d d -neighborhood and the enlarged comparison cylinders are omitted.
(c) Cauchy-Schwarz and Young’s inequalities remove the mixed Hessian term at the price of the term ( N + p − 2 ) ( a ′ ) 2 / a (N+p-2)(a')^2/a ( N + p − 2 ) ( a ′ ) 2 / a , which is compensated by 2 b ′ 2b' 2 b ′ as soon as k ≥ k o k\ge k_o k ≥ k o ;
(d) a cutoff function η \eta η and the maximum principle at a maximum point of η 2 w \eta^2w η 2 w localize the estimate.
A more detailed discussion is given at the end of the section. We consider equation (1) in Ω T \Omega_T Ω T , and we set
A ∗ = ( N + p − 2 ) ( p − 2 ) 2 , k o = 2 + A ∗ ( p − 1 ) 2 , k ≥ k o . (9) A_*=(N+p-2)(p-2)^2,\qquad k_o=2+\frac{A_*}{(p-1)^2},\qquad k\ge k_o. \tag*{(9)} A ∗ = ( N + p − 2 ) ( p − 2 ) 2 , k o = 2 + ( p − 1 ) 2 A ∗ , k ≥ k o . (9)
The following calculation is carried out locally in a region where u > 0 u>0 u > 0 and ∇ v ≠ 0 \nabla v\ne0 ∇ v = 0 , with v = u 1 / k v=u^{1/k} v = u 1/ k . We assume sufficient smoothness for all derivatives below to exist and for the relevant mixed derivatives to commute. For weak solutions, this differentiation requires a separate justification. This is the reason why in § 4 we consider solutions to a regularized equation. The specific lower bound on k k k is needed later, but not for the identities derived below.
The transformation u = v k u=v^k u = v k
Let u u u be a smooth nonnegative solution to (1) . On the open set where u > 0 u>0 u > 0 , we define
The transformation is smooth on this set. In deriving the pointwise differentiated identity below, we work at points where ∇ v ≠ 0 \nabla v\ne0 ∇ v = 0 . This is a local restriction for the calculation, not an additional assumption on the solution. The resulting identity is therefore established on
{ ( x , t ) ∈ Ω T : u ( x , t ) > 0 , ∣ ∇ v ( x , t ) ∣ > 0 } . \{(x,t)\in\Omega_T:u(x,t)>0,\ |\nabla v(x,t)|>0\}. {( x , t ) ∈ Ω T : u ( x , t ) > 0 , ∣∇ v ( x , t ) ∣ > 0 } .
Its use at points where u = 0 u=0 u = 0 or ∇ v = 0 \nabla v=0 ∇ v = 0 requires a separate justification. Using v = u 1 / k v=u^{1/k} v = u 1/ k and applying the chain rule, we obtain
v t = ∂ t ( ( u p − 1 ) 1 k ( p − 1 ) ) = 1 k ( p − 1 ) u 1 k − ( p − 1 ) ∂ t ( u p − 1 ) = 1 k ( p − 1 ) v 1 − k ( p − 1 ) div ( ∣ ∇ ( v k ) ∣ p − 2 ∇ ( v k ) ) . \begin{aligned}
v_t=\partial_t\left((u^{p-1})^{\frac{1}{k(p-1)}}\right) \\
&=\frac{1}{k(p-1)}u^{\frac{1}{k}-(p-1)}\partial_t(u^{p-1}) \\
&=\frac{1}{k(p-1)}v^{1-k(p-1)}\operatorname{div}\left(\lvert\nabla(v^k)\rvert^{p-2}\nabla(v^k)\right).
\end{aligned} v t = ∂ t ( ( u p − 1 ) k ( p − 1 ) 1 ) = k ( p − 1 ) 1 u k 1 − ( p − 1 ) ∂ t ( u p − 1 ) = k ( p − 1 ) 1 v 1 − k ( p − 1 ) div ( ∣ ∇ ( v k ) ∣ p − 2 ∇ ( v k ) ) .
Since
∇ ( v k ) = k v k − 1 ∇ v , \nabla(v^k)=kv^{k-1}\nabla v, ∇ ( v k ) = k v k − 1 ∇ v ,
we have
∣ ∇ ( v k ) ∣ p − 2 ∇ ( v k ) = k p − 1 v ( k − 1 ) ( p − 1 ) ∣ ∇ v ∣ p − 2 ∇ v . \lvert\nabla(v^k)\rvert^{p-2}\nabla(v^k)=k^{p-1}v^{(k-1)(p-1)}\lvert\nabla v\rvert^{p-2}\nabla v. ∣ ∇ ( v k ) ∣ p − 2 ∇ ( v k ) = k p − 1 v ( k − 1 ) ( p − 1 ) ∣ ∇ v ∣ p − 2 ∇ v .
Consequently, the product rule gives
v t = k p − 2 p − 1 v 1 − k ( p − 1 ) div [ v ( k − 1 ) ( p − 1 ) ∣ ∇ v ∣ p − 2 ∇ v ] = k p − 2 p − 1 v 1 − k ( p − 1 ) [ v ( k − 1 ) ( p − 1 ) div ( ∣ ∇ v ∣ p − 2 ∇ v ) + ∇ ( v ( k − 1 ) ( p − 1 ) ) ⋅ ( ∣ ∇ v ∣ p − 2 ∇ v ) ] = k p − 2 p − 1 v 1 − k ( p − 1 ) [ v ( k − 1 ) ( p − 1 ) Δ p v + ( k − 1 ) ( p − 1 ) v ( k − 1 ) ( p − 1 ) − 1 ∣ ∇ v ∣ p − 2 ∇ v ⋅ ∇ v ⏟ = ∣ ∇ v ∣ 2 ] = k p − 2 p − 1 v 1 − k ( p − 1 ) [ v ( k − 1 ) ( p − 1 ) Δ p v + ( k − 1 ) ( p − 1 ) v ( k − 1 ) ( p − 1 ) − 1 ∣ ∇ v ∣ p ] . \begin{aligned}
v_t=\frac{k^{p-2}}{p-1}v^{1-k(p-1)}\operatorname{div}\left[v^{(k-1)(p-1)}\lvert\nabla v\rvert^{p-2}\nabla v\right] \\
&=\frac{k^{p-2}}{p-1}v^{1-k(p-1)}\left[v^{(k-1)(p-1)}\operatorname{div}\left(\lvert\nabla v\rvert^{p-2}\nabla v\right)+\nabla\left(v^{(k-1)(p-1)}\right)\cdot\left(\lvert\nabla v\rvert^{p-2}\nabla v\right)\right] \\
&=\frac{k^{p-2}}{p-1}v^{1-k(p-1)}\left[v^{(k-1)(p-1)}\Delta_pv+(k-1)(p-1)v^{(k-1)(p-1)-1}\lvert\nabla v\rvert^{p-2}\underbrace{\nabla v\cdot\nabla v}_{=\lvert\nabla v\rvert^2}\right] \\
&=\frac{k^{p-2}}{p-1}v^{1-k(p-1)}\left[v^{(k-1)(p-1)}\Delta_pv+(k-1)(p-1)v^{(k-1)(p-1)-1}\lvert\nabla v\rvert^p\right].
\end{aligned} v t = p − 1 k p − 2 v 1 − k ( p − 1 ) div [ v ( k − 1 ) ( p − 1 ) ∣ ∇ v ∣ p − 2 ∇ v ] = p − 1 k p − 2 v 1 − k ( p − 1 ) [ v ( k − 1 ) ( p − 1 ) div ( ∣ ∇ v ∣ p − 2 ∇ v ) + ∇ ( v ( k − 1 ) ( p − 1 ) ) ⋅ ( ∣ ∇ v ∣ p − 2 ∇ v ) ] = p − 1 k p − 2 v 1 − k ( p − 1 ) v ( k − 1 ) ( p − 1 ) Δ p v + ( k − 1 ) ( p − 1 ) v ( k − 1 ) ( p − 1 ) − 1 ∣ ∇ v ∣ p − 2 = ∣ ∇ v ∣ 2 ∇ v ⋅ ∇ v = p − 1 k p − 2 v 1 − k ( p − 1 ) [ v ( k − 1 ) ( p − 1 ) Δ p v + ( k − 1 ) ( p − 1 ) v ( k − 1 ) ( p − 1 ) − 1 ∣ ∇ v ∣ p ] .
The exponents simplify according to
1 − k ( p − 1 ) + ( k − 1 ) ( p − 1 ) = 2 − p 1-k(p-1)+(k-1)(p-1)=2-p 1 − k ( p − 1 ) + ( k − 1 ) ( p − 1 ) = 2 − p
and
1 − k ( p − 1 ) + ( k − 1 ) ( p − 1 ) − 1 = 1 − p . 1-k(p-1)+(k-1)(p-1)-1=1-p. 1 − k ( p − 1 ) + ( k − 1 ) ( p − 1 ) − 1 = 1 − p .
Combining the powers of v v v , we obtain
v t = a ( v ) Δ p v + b ( v ) ∣ ∇ v ∣ p , (10) v_t=a(v)\Delta_pv+b(v)\lvert\nabla v\rvert^p, \tag*{(10)} v t = a ( v ) Δ p v + b ( v ) ∣ ∇ v ∣ p , (10)
where
a ( v ) = c v 2 − p , b ( v ) = c ( p − 1 ) ( k − 1 ) v 1 − p , c = k p − 2 p − 1 . (11) a(v)=cv^{2-p},\qquad b(v)=c(p-1)(k-1)v^{1-p},\qquad c=\frac{k^{p-2}}{p-1}. \tag*{(11)} a ( v ) = c v 2 − p , b ( v ) = c ( p − 1 ) ( k − 1 ) v 1 − p , c = p − 1 k p − 2 . (11)
In particular, a , b > 0 a,b>0 a , b > 0 , and
a ′ ( v ) = c ( 2 − p ) v 1 − p , b ′ ( v ) = − c ( p − 1 ) 2 ( k − 1 ) v − p . a'(v)=c(2-p)v^{1-p},\qquad b'(v)=-c(p-1)^2(k-1)v^{-p}. a ′ ( v ) = c ( 2 − p ) v 1 − p , b ′ ( v ) = − c ( p − 1 ) 2 ( k − 1 ) v − p .
We point out that the only effect of the power transformation is the new first-order term b ( v ) ∣ ∇ v ∣ p b(v)\lvert\nabla v\rvert^p b ( v ) ∣ ∇ v ∣ p (for k = 1 k=1 k = 1 one has b ≡ 0 b\equiv0 b ≡ 0 ). The sign b > 0 b>0 b > 0 , which requires k > 1 k>1 k > 1 , is used in the drift estimate below, and b ′ < 0 b'<0 b ′ < 0 will provide the coercivity.
Expanding the p p p -Laplace operator
At points where ∇ v ≠ 0 \nabla v \ne0 ∇ v = 0 , the product rule gives
Δ p v = div ( ∣ ∇ v ∣ p − 2 ∇ v ) = ∣ ∇ v ∣ p − 2 Δ v + ∇ ( ∣ ∇ v ∣ p − 2 ) ⋅ ∇ v . \Delta_p v = \operatorname{div}\left(\lvert\nabla v\rvert^{p-2}\nabla v\right) = \lvert\nabla v\rvert^{p-2}\Delta v + \nabla\left(\lvert\nabla v\rvert^{p-2}\right)\cdot\nabla v. Δ p v = div ( ∣ ∇ v ∣ p − 2 ∇ v ) = ∣ ∇ v ∣ p − 2 Δ v + ∇ ( ∣ ∇ v ∣ p − 2 ) ⋅ ∇ v .
To compute the second term, we use the chain rule:
∇ ( ∣ ∇ v ∣ p − 2 ) = ∇ ( ( ∣ ∇ v ∣ 2 ) ( p − 2 ) / 2 ) = p − 2 2 ∣ ∇ v ∣ p − 4 ∇ ( ∣ ∇ v ∣ 2 ) = ( p − 2 ) ∣ ∇ v ∣ p − 4 ∇ 2 v ∇ v . \begin{aligned}
\nabla\left(\lvert\nabla v\rvert^{p-2}\right)
&= \nabla\left(\left(\lvert\nabla v\rvert^2\right)^{(p-2)/2}\right) \\
&= \frac{p-2}{2}\lvert\nabla v\rvert^{p-4}\nabla\left(\lvert\nabla v\rvert^2\right) \\
&= (p-2)\lvert\nabla v\rvert^{p-4}\nabla^2v\nabla v.
\end{aligned} ∇ ( ∣ ∇ v ∣ p − 2 ) = ∇ ( ( ∣ ∇ v ∣ 2 ) ( p − 2 ) /2 ) = 2 p − 2 ∣ ∇ v ∣ p − 4 ∇ ( ∣ ∇ v ∣ 2 ) = ( p − 2 ) ∣ ∇ v ∣ p − 4 ∇ 2 v ∇ v .
Here ∇ 2 v \nabla^2v ∇ 2 v denotes the Hessian matrix of v v v , and we have used
∇ ( ∣ ∇ v ∣ 2 ) = 2 ∇ 2 v ∇ v . \nabla\left(\lvert\nabla v\rvert^2\right)=2\nabla^2v\nabla v. ∇ ( ∣ ∇ v ∣ 2 ) = 2 ∇ 2 v ∇ v .
Substituting this expression yields
Δ p v = ∣ ∇ v ∣ p − 2 Δ v + ( p − 2 ) ∣ ∇ v ∣ p − 4 ( ∇ 2 v ∇ v ) ⋅ ∇ v . (12) \Delta_p v = \lvert\nabla v\rvert^{p-2}\Delta v + (p-2)\lvert\nabla v\rvert^{p-4}(\nabla^2v\nabla v)\cdot\nabla v. \tag*{(12)} Δ p v = ∣ ∇ v ∣ p − 2 Δ v + ( p − 2 ) ∣ ∇ v ∣ p − 4 ( ∇ 2 v ∇ v ) ⋅ ∇ v . (12)
Consequently, equation (10) takes the form
v t = a ( v ) ∣ ∇ v ∣ p − 2 Δ v + ( p − 2 ) a ( v ) ∣ ∇ v ∣ p − 4 ( ∇ 2 v ∇ v ) ⋅ ∇ v + b ( v ) ∣ ∇ v ∣ p . v_t = a(v)\lvert\nabla v\rvert^{p-2}\Delta v + (p-2)a(v)\lvert\nabla v\rvert^{p-4}(\nabla^2v\nabla v)\cdot\nabla v + b(v)\lvert\nabla v\rvert^p. v t = a ( v ) ∣ ∇ v ∣ p − 2 Δ v + ( p − 2 ) a ( v ) ∣ ∇ v ∣ p − 4 ( ∇ 2 v ∇ v ) ⋅ ∇ v + b ( v ) ∣ ∇ v ∣ p .
At points where ∇ v ≠ 0 \nabla v\ne0 ∇ v = 0 , we can factor out a ( v ) ∣ ∇ v ∣ p − 2 a(v)\lvert\nabla v\rvert^{p-2} a ( v ) ∣ ∇ v ∣ p − 2 and write
v t = a ( v ) ∣ ∇ v ∣ p − 2 [ Δ v + ( p − 2 ) ∇ 2 v ⋅ ( ∇ v ∣ ∇ v ∣ ⊗ ∇ v ∣ ∇ v ∣ ) ] + b ( v ) ∣ ∇ v ∣ p . v_t = a(v)\lvert\nabla v\rvert^{p-2}\left[\Delta v+(p-2)\nabla^2v\cdot\left(\frac{\nabla v}{\lvert\nabla v\rvert}\otimes\frac{\nabla v}{\lvert\nabla v\rvert}\right)\right]+b(v)\lvert\nabla v\rvert^p. v t = a ( v ) ∣ ∇ v ∣ p − 2 [ Δ v + ( p − 2 ) ∇ 2 v ⋅ ( ∣ ∇ v ∣ ∇ v ⊗ ∣ ∇ v ∣ ∇ v ) ] + b ( v ) ∣ ∇ v ∣ p .
Here the bold dot denotes the Frobenius inner product of matrices, so that
∇ 2 v ⋅ ( ∇ v ∣ ∇ v ∣ ⊗ ∇ v ∣ ∇ v ∣ ) = ( ∇ 2 v ∇ v ) ⋅ ∇ v ∣ ∇ v ∣ 2 . \nabla^2v\cdot\left(\frac{\nabla v}{\lvert\nabla v\rvert}\otimes\frac{\nabla v}{\lvert\nabla v\rvert}\right)=\frac{(\nabla^2v\nabla v)\cdot\nabla v}{\lvert\nabla v\rvert^2}. ∇ 2 v ⋅ ( ∣ ∇ v ∣ ∇ v ⊗ ∣ ∇ v ∣ ∇ v ) = ∣ ∇ v ∣ 2 ( ∇ 2 v ∇ v ) ⋅ ∇ v .
Differentiating the equation
We take the spatial gradient of (10) and then take the scalar product with 2 ∇ v 2\nabla v 2∇ v . The product and chain rules give
∂ t ∣ ∇ v ∣ 2 = 2 ∇ v ⋅ ∇ ∂ t v = 2 ∇ v ⋅ ∇ ( a ( v ) Δ p v + b ( v ) ∣ ∇ v ∣ p ) = 2 a ′ ( v ) ∣ ∇ v ∣ 2 Δ p v + 2 a ( v ) ∇ v ⋅ ∇ Δ p v + 2 b ′ ( v ) ∣ ∇ v ∣ p + 2 + 2 b ( v ) ∇ v ⋅ ∇ ( ∣ ∇ v ∣ p ) . \begin{aligned}
\partial_t\lvert\nabla v\rvert^2
&=2\nabla v\cdot\nabla\partial_t v
=2\nabla v\cdot\nabla\left(a(v)\Delta_pv+b(v)\lvert\nabla v\rvert^p\right)\\
&=2a'(v)\lvert\nabla v\rvert^2\Delta_pv+2a(v)\nabla v\cdot\nabla\Delta_pv\\
&\quad+2b'(v)\lvert\nabla v\rvert^{p+2}+2b(v)\nabla v\cdot\nabla\left(\lvert\nabla v\rvert^p\right).
\end{aligned} ∂ t ∣ ∇ v ∣ 2 = 2∇ v ⋅ ∇ ∂ t v = 2∇ v ⋅ ∇ ( a ( v ) Δ p v + b ( v ) ∣ ∇ v ∣ p ) = 2 a ′ ( v ) ∣ ∇ v ∣ 2 Δ p v + 2 a ( v ) ∇ v ⋅ ∇ Δ p v + 2 b ′ ( v ) ∣ ∇ v ∣ p + 2 + 2 b ( v ) ∇ v ⋅ ∇ ( ∣ ∇ v ∣ p ) .
Since
∇ ( ∣ ∇ v ∣ p ) = 1 2 p ∣ ∇ v ∣ p − 2 ∇ ( ∣ ∇ v ∣ 2 ) , \nabla\left(\lvert\nabla v\rvert^p\right)=\frac{1}{2}p\lvert\nabla v\rvert^{p-2}\nabla\left(\lvert\nabla v\rvert^2\right), ∇ ( ∣ ∇ v ∣ p ) = 2 1 p ∣ ∇ v ∣ p − 2 ∇ ( ∣ ∇ v ∣ 2 ) ,
we obtain
∂ t ∣ ∇ v ∣ 2 = 2 a ′ ( v ) ∣ ∇ v ∣ 2 Δ p v + 2 a ( v ) ∇ v ⋅ ∇ Δ p v + 2 b ′ ( v ) ∣ ∇ v ∣ p + 2 + p b ( v ) ∣ ∇ v ∣ p − 2 ∇ v ⋅ ∇ ( ∣ ∇ v ∣ 2 ) . (13) \partial_t\lvert\nabla v\rvert^2=2a'(v)\lvert\nabla v\rvert^2\Delta_pv+2a(v)\nabla v\cdot\nabla\Delta_pv+2b'(v)\lvert\nabla v\rvert^{p+2}+pb(v)\lvert\nabla v\rvert^{p-2}\nabla v\cdot\nabla\left(\lvert\nabla v\rvert^2\right). \tag*{(13)} ∂ t ∣ ∇ v ∣ 2 = 2 a ′ ( v ) ∣ ∇ v ∣ 2 Δ p v + 2 a ( v ) ∇ v ⋅ ∇ Δ p v + 2 b ′ ( v ) ∣ ∇ v ∣ p + 2 + p b ( v ) ∣ ∇ v ∣ p − 2 ∇ v ⋅ ∇ ( ∣ ∇ v ∣ 2 ) . (13)
We next rewrite the term containing ∇ Δ p v \nabla\Delta_pv ∇ Δ p v . For each m ∈ { 1 , … , N } m\in\{1,\ldots,N\} m ∈ { 1 , … , N } , differentiating the flux gives
∂ x m Δ p v = div ( ∂ x m ( ∣ ∇ v ∣ p − 2 ∇ v ) ) = div ( ∣ ∇ v ∣ p − 2 ∇ ( ∂ x m v ) + ( p − 2 ) ∣ ∇ v ∣ p − 4 ( ∇ v ⋅ ∇ ( ∂ x m v ) ) ∇ v ) . \begin{aligned}
\partial_{x_m}\Delta_pv=\operatorname{div}\left(\partial_{x_m}\left(\lvert\nabla v\rvert^{p-2}\nabla v\right)\right) \\
&=\operatorname{div}\left(\lvert\nabla v\rvert^{p-2}\nabla(\partial_{x_m}v)+(p-2)\lvert\nabla v\rvert^{p-4}\left(\nabla v\cdot\nabla(\partial_{x_m}v)\right)\nabla v\right).
\end{aligned} ∂ x m Δ p v = div ( ∂ x m ( ∣ ∇ v ∣ p − 2 ∇ v ) ) = div ( ∣ ∇ v ∣ p − 2 ∇ ( ∂ x m v ) + ( p − 2 ) ∣ ∇ v ∣ p − 4 ( ∇ v ⋅ ∇ ( ∂ x m v ) ) ∇ v ) .
Multiplying by ∂ x m v \partial_{x_m}v ∂ x m v and summing over m m m , we obtain
∇ v ⋅ ∇ Δ p v = ∑ m = 1 N ( ∂ x m v ) div ( ∣ ∇ v ∣ p − 2 ∇ ( ∂ x m v ) + ( p − 2 ) ∣ ∇ v ∣ p − 4 ( ∇ v ⋅ ∇ ( ∂ x m v ) ) ∇ v ) . \nabla v\cdot\nabla\Delta_pv
=\sum_{m=1}^{N}(\partial_{x_m}v)\operatorname{div}\left(\lvert\nabla v\rvert^{p-2}\nabla(\partial_{x_m}v)+(p-2)\lvert\nabla v\rvert^{p-4}\left(\nabla v\cdot\nabla(\partial_{x_m}v)\right)\nabla v\right). ∇ v ⋅ ∇ Δ p v = m = 1 ∑ N ( ∂ x m v ) div ( ∣ ∇ v ∣ p − 2 ∇ ( ∂ x m v ) + ( p − 2 ) ∣ ∇ v ∣ p − 4 ( ∇ v ⋅ ∇ ( ∂ x m v ) ) ∇ v ) .
We now apply the product rule
( ∂ x m v ) div X = div ( ( ∂ x m v ) X ) − ∇ ( ∂ x m v ) ⋅ X (\partial_{x_m}v)\operatorname{div}X=\operatorname{div}\left((\partial_{x_m}v)X\right)-\nabla(\partial_{x_m}v)\cdot X ( ∂ x m v ) div X = div ( ( ∂ x m v ) X ) − ∇ ( ∂ x m v ) ⋅ X
to each summand, with X X X denoting either of the two vector fields inside the divergence in the preceding identity. This gives
∇ v ⋅ ∇ Δ p v = div ( ∣ ∇ v ∣ p − 2 ∑ m = 1 N ( ∂ x m v ) ∇ ( ∂ x m v ) + ( p − 2 ) ∣ ∇ v ∣ p − 4 [ ∑ m = 1 N ( ∂ x m v ) ( ∇ v ⋅ ∇ ( ∂ x m v ) ) ] ∇ v ) − ∣ ∇ v ∣ p − 2 ∑ m = 1 N ∣ ∇ ( ∂ x m v ) ∣ 2 − ( p − 2 ) ∣ ∇ v ∣ p − 4 ∑ m = 1 N ( ∇ v ⋅ ∇ ( ∂ x m v ) ) 2 . \begin{aligned}
\nabla v \cdot\nabla\Delta_p v ={}& \operatorname{div}\left(|\nabla v|^{p-2}\sum_{m=1}^{N}(\partial_{x_m}v)\nabla(\partial_{x_m}v)\right.\\
&\left.{}+(p-2)|\nabla v|^{p-4}\left[\sum_{m=1}^{N}(\partial_{x_m}v)(\nabla v\cdot\nabla(\partial_{x_m}v))\right]\nabla v\right)\\
&{}-|\nabla v|^{p-2}\sum_{m=1}^{N}|\nabla(\partial_{x_m}v)|^2-(p-2)|\nabla v|^{p-4}\sum_{m=1}^{N}(\nabla v\cdot\nabla(\partial_{x_m}v))^2.
\end{aligned} ∇ v ⋅ ∇ Δ p v = div ( ∣∇ v ∣ p − 2 m = 1 ∑ N ( ∂ x m v ) ∇ ( ∂ x m v ) + ( p − 2 ) ∣∇ v ∣ p − 4 [ m = 1 ∑ N ( ∂ x m v ) ( ∇ v ⋅ ∇ ( ∂ x m v )) ] ∇ v ) − ∣∇ v ∣ p − 2 m = 1 ∑ N ∣∇ ( ∂ x m v ) ∣ 2 − ( p − 2 ) ∣∇ v ∣ p − 4 m = 1 ∑ N ( ∇ v ⋅ ∇ ( ∂ x m v ) ) 2 .
The two sums inside the divergence on the right-hand side satisfy
∑ m = 1 N ( ∂ x m v ) ∇ ( ∂ x m v ) = 1 2 ∇ ( ∣ ∇ v ∣ 2 ) . \sum_{m=1}^{N}(\partial_{x_m}v)\nabla(\partial_{x_m}v)=\frac{1}{2}\nabla(|\nabla v|^2). m = 1 ∑ N ( ∂ x m v ) ∇ ( ∂ x m v ) = 2 1 ∇ ( ∣∇ v ∣ 2 ) .
and
∑ m = 1 N ( ∂ x m v ) ( ∇ v ⋅ ∇ ( ∂ x m v ) ) = 1 2 ∇ v ⋅ ∇ ( ∣ ∇ v ∣ 2 ) . \sum_{m=1}^{N}(\partial_{x_m}v)(\nabla v\cdot\nabla(\partial_{x_m}v))=\frac{1}{2}\nabla v\cdot\nabla(|\nabla v|^2). m = 1 ∑ N ( ∂ x m v ) ( ∇ v ⋅ ∇ ( ∂ x m v )) = 2 1 ∇ v ⋅ ∇ ( ∣∇ v ∣ 2 ) .
By the definition of the Frobenius norm,
∑ m = 1 N ∣ ∇ ( ∂ x m v ) ∣ 2 = ∑ m = 1 N ∑ j = 1 N ∣ ∂ x j ∂ x m v ∣ 2 = ∣ ∇ 2 v ∣ 2 . \sum_{m=1}^{N}|\nabla(\partial_{x_m}v)|^2=\sum_{m=1}^{N}\sum_{j=1}^{N}|\partial_{x_j}\partial_{x_m}v|^2=|\nabla^2v|^2. m = 1 ∑ N ∣∇ ( ∂ x m v ) ∣ 2 = m = 1 ∑ N j = 1 ∑ N ∣ ∂ x j ∂ x m v ∣ 2 = ∣ ∇ 2 v ∣ 2 .
Moreover, since the mixed second derivatives commute,
∇ v ⋅ ∇ ( ∂ x m v ) = ∑ j = 1 N ( ∂ x j v ) ( ∂ x j ∂ x m v ) = ∑ j = 1 N ( ∂ x m ∂ x j v ) ( ∂ x j v ) = ( ∇ 2 v ∇ v ) m . \nabla v\cdot\nabla(\partial_{x_m}v)=\sum_{j=1}^{N}(\partial_{x_j}v)(\partial_{x_j}\partial_{x_m}v)=\sum_{j=1}^{N}(\partial_{x_m}\partial_{x_j}v)(\partial_{x_j}v)=(\nabla^2v\nabla v)_m. ∇ v ⋅ ∇ ( ∂ x m v ) = j = 1 ∑ N ( ∂ x j v ) ( ∂ x j ∂ x m v ) = j = 1 ∑ N ( ∂ x m ∂ x j v ) ( ∂ x j v ) = ( ∇ 2 v ∇ v ) m .
Squaring and summing over m m m therefore gives
∑ m = 1 N ( ∇ v ⋅ ∇ ( ∂ x m v ) ) 2 = ∑ m = 1 N ( ∇ 2 v ∇ v ) m 2 = ∣ ∇ 2 v ∇ v ∣ 2 . \sum_{m=1}^{N}(\nabla v\cdot\nabla(\partial_{x_m}v))^2=\sum_{m=1}^{N}(\nabla^2v\nabla v)_m^2=|\nabla^2v\nabla v|^2. m = 1 ∑ N ( ∇ v ⋅ ∇ ( ∂ x m v ) ) 2 = m = 1 ∑ N ( ∇ 2 v ∇ v ) m 2 = ∣ ∇ 2 v ∇ v ∣ 2 .
Therefore,
2 ∇ v ⋅ ∇ Δ p v = div ( ∣ ∇ v ∣ p − 2 ∇ ( ∣ ∇ v ∣ 2 ) + ( p − 2 ) ∣ ∇ v ∣ p − 4 ( ∇ v ⋅ ∇ ( ∣ ∇ v ∣ 2 ) ) ∇ v ) − 2 ∣ ∇ v ∣ p − 2 ∣ ∇ 2 v ∣ 2 − 2 ( p − 2 ) ∣ ∇ v ∣ p − 4 ∣ ∇ 2 v ∇ v ∣ 2 . (14) \begin{aligned}
2\nabla v\cdot\nabla\Delta_pv={}&\operatorname{div}\left(|\nabla v|^{p-2}\nabla(|\nabla v|^2)+(p-2)|\nabla v|^{p-4}(\nabla v\cdot\nabla(|\nabla v|^2))\nabla v\right)\\
&{}-2|\nabla v|^{p-2}|\nabla^2v|^2-2(p-2)|\nabla v|^{p-4}|\nabla^2v\nabla v|^2. \tag*{(14)}
\end{aligned} 2∇ v ⋅ ∇ Δ p v = div ( ∣∇ v ∣ p − 2 ∇ ( ∣∇ v ∣ 2 ) + ( p − 2 ) ∣∇ v ∣ p − 4 ( ∇ v ⋅ ∇ ( ∣∇ v ∣ 2 )) ∇ v ) − 2∣∇ v ∣ p − 2 ∣ ∇ 2 v ∣ 2 − 2 ( p − 2 ) ∣∇ v ∣ p − 4 ∣ ∇ 2 v ∇ v ∣ 2 . (14)
We first substitute (14) into (13) , leaving the first term on the right-hand side, i.e. 2 a ′ ( v ) ∣ ∇ v ∣ 2 Δ p v 2a'(v)|\nabla v|^2\Delta_pv 2 a ′ ( v ) ∣∇ v ∣ 2 Δ p v , unchanged. This gives
∂ t ∣ ∇ v ∣ 2 = 2 a ′ ( v ) ∣ ∇ v ∣ 2 Δ p v + a ( v ) div ( ∣ ∇ v ∣ p − 2 ∇ ( ∣ ∇ v ∣ 2 ) + ( p − 2 ) ∣ ∇ v ∣ p − 4 ( ∇ v ⋅ ∇ ( ∣ ∇ v ∣ 2 ) ) ∇ v ) − 2 a ( v ) ∣ ∇ v ∣ p − 2 ∣ ∇ 2 v ∣ 2 − 2 ( p − 2 ) a ( v ) ∣ ∇ v ∣ p − 4 ∣ ∇ 2 v ∇ v ∣ 2 + 2 b ′ ( v ) ∣ ∇ v ∣ p + 2 + p b ( v ) ∣ ∇ v ∣ p − 2 ∇ v ⋅ ∇ ( ∣ ∇ v ∣ 2 ) . \begin{aligned}
\partial_t|\nabla v|^2={}&2a'(v)|\nabla v|^2\Delta_pv\\
&{}+a(v)\operatorname{div}\left(|\nabla v|^{p-2}\nabla(|\nabla v|^2)+(p-2)|\nabla v|^{p-4}(\nabla v\cdot\nabla(|\nabla v|^2))\nabla v\right)\\
&{}-2a(v)|\nabla v|^{p-2}|\nabla^2v|^2-2(p-2)a(v)|\nabla v|^{p-4}|\nabla^2v\nabla v|^2\\
&{}+2b'(v)|\nabla v|^{p+2}+pb(v)|\nabla v|^{p-2}\nabla v\cdot\nabla(|\nabla v|^2).
\end{aligned} ∂ t ∣∇ v ∣ 2 = 2 a ′ ( v ) ∣∇ v ∣ 2 Δ p v + a ( v ) div ( ∣∇ v ∣ p − 2 ∇ ( ∣∇ v ∣ 2 ) + ( p − 2 ) ∣∇ v ∣ p − 4 ( ∇ v ⋅ ∇ ( ∣∇ v ∣ 2 )) ∇ v ) − 2 a ( v ) ∣∇ v ∣ p − 2 ∣ ∇ 2 v ∣ 2 − 2 ( p − 2 ) a ( v ) ∣∇ v ∣ p − 4 ∣ ∇ 2 v ∇ v ∣ 2 + 2 b ′ ( v ) ∣∇ v ∣ p + 2 + p b ( v ) ∣∇ v ∣ p − 2 ∇ v ⋅ ∇ ( ∣∇ v ∣ 2 ) .
The terms appearing in the second and third lines on the right-hand side together equal 2 a ( v ) ∇ v ⋅ ∇ Δ p v 2a(v)\nabla v\cdot\nabla\Delta_pv 2 a ( v ) ∇ v ⋅ ∇ Δ p v . Next, we expand the first term on the right-hand side using (12) :
2 a ′ ( v ) ∣ ∇ v ∣ 2 Δ p v = 2 a ′ ( v ) ∣ ∇ v ∣ 2 [ ∣ ∇ v ∣ p − 2 Δ v + ( p − 2 ) ∣ ∇ v ∣ p − 4 ( ∇ 2 v ∇ v ) ⋅ ∇ v ] = 2 a ′ ( v ) ∣ ∇ v ∣ p Δ v + 2 ( p − 2 ) a ′ ( v ) ∣ ∇ v ∣ p − 2 ( ∇ 2 v ∇ v ) ⋅ ∇ v . \begin{aligned}
2a'(v)|\nabla v|^2\Delta_pv={}&2a'(v)|\nabla v|^2\left[|\nabla v|^{p-2}\Delta v+(p-2)|\nabla v|^{p-4}(\nabla^2v\nabla v)\cdot\nabla v\right]\\
={}&2a'(v)|\nabla v|^p\Delta v+2(p-2)a'(v)|\nabla v|^{p-2}(\nabla^2v\nabla v)\cdot\nabla v.
\end{aligned} 2 a ′ ( v ) ∣∇ v ∣ 2 Δ p v = = 2 a ′ ( v ) ∣∇ v ∣ 2 [ ∣∇ v ∣ p − 2 Δ v + ( p − 2 ) ∣∇ v ∣ p − 4 ( ∇ 2 v ∇ v ) ⋅ ∇ v ] 2 a ′ ( v ) ∣∇ v ∣ p Δ v + 2 ( p − 2 ) a ′ ( v ) ∣∇ v ∣ p − 2 ( ∇ 2 v ∇ v ) ⋅ ∇ v .
Substituting this expression into the preceding identity yields
∂ t ∣ ∇ v ∣ 2 = a ( v ) div ( ∣ ∇ v ∣ p − 2 ∇ ( ∣ ∇ v ∣ 2 ) + ( p − 2 ) ∣ ∇ v ∣ p − 4 ( ∇ v ⋅ ∇ ( ∣ ∇ v ∣ 2 ) ) ∇ v ) + p b ( v ) ∣ ∇ v ∣ p − 2 ∇ v ⋅ ∇ ( ∣ ∇ v ∣ 2 ) − 2 a ( v ) ( ∣ ∇ v ∣ p − 2 ∣ ∇ 2 v ∣ 2 + ( p − 2 ) ∣ ∇ v ∣ p − 4 ∣ ∇ 2 v ∇ v ∣ 2 ) + 2 a ′ ( v ) ( ∣ ∇ v ∣ p Δ v + ( p − 2 ) ∣ ∇ v ∣ p − 2 ( ∇ 2 v ∇ v ) ⋅ ∇ v ) + 2 b ′ ( v ) ∣ ∇ v ∣ p + 2 . \begin{aligned}
\partial_t |\nabla v|^2 ={}& a(v)\operatorname{div}\left(|\nabla v|^{p-2}\nabla(|\nabla v|^2) +(p-2)|\nabla v|^{p-4}(\nabla v\cdot\nabla(|\nabla v|^2))\nabla v\right)\\
&+pb(v)|\nabla v|^{p-2}\nabla v\cdot\nabla(|\nabla v|^2)\\
&-2a(v)\left(|\nabla v|^{p-2}|\nabla^2v|^2 +(p-2)|\nabla v|^{p-4}|\nabla^2v\nabla v|^2\right)\\
&+2a'(v)\left(|\nabla v|^p\Delta v +(p-2)|\nabla v|^{p-2}(\nabla^2v\nabla v)\cdot\nabla v\right)\\
&+2b'(v)|\nabla v|^{p+2}.
\end{aligned} ∂ t ∣∇ v ∣ 2 = a ( v ) div ( ∣∇ v ∣ p − 2 ∇ ( ∣∇ v ∣ 2 ) + ( p − 2 ) ∣∇ v ∣ p − 4 ( ∇ v ⋅ ∇ ( ∣∇ v ∣ 2 )) ∇ v ) + p b ( v ) ∣∇ v ∣ p − 2 ∇ v ⋅ ∇ ( ∣∇ v ∣ 2 ) − 2 a ( v ) ( ∣∇ v ∣ p − 2 ∣ ∇ 2 v ∣ 2 + ( p − 2 ) ∣∇ v ∣ p − 4 ∣ ∇ 2 v ∇ v ∣ 2 ) + 2 a ′ ( v ) ( ∣∇ v ∣ p Δ v + ( p − 2 ) ∣∇ v ∣ p − 2 ( ∇ 2 v ∇ v ) ⋅ ∇ v ) + 2 b ′ ( v ) ∣∇ v ∣ p + 2 .
Finally, we move the divergence term and the term containing b ( v ) b(v) b ( v ) to the left-hand side. We obtain the identity
∂ t ∣ ∇ v ∣ 2 − a ( v ) div ( ∣ ∇ v ∣ p − 2 ∇ ( ∣ ∇ v ∣ 2 ) + ( p − 2 ) ∣ ∇ v ∣ p − 4 ( ∇ v ⋅ ∇ ( ∣ ∇ v ∣ 2 ) ) ∇ v ) − p b ( v ) ∣ ∇ v ∣ p − 2 ∇ v ⋅ ∇ ( ∣ ∇ v ∣ 2 ) ⏟ 1st-order term = − 2 a ( v ) ( ∣ ∇ v ∣ p − 2 ∣ ∇ 2 v ∣ 2 + ( p − 2 ) ∣ ∇ v ∣ p − 4 ∣ ∇ 2 v ∇ v ∣ 2 ) + 2 a ′ ( v ) ( ∣ ∇ v ∣ p Δ v + ( p − 2 ) ∣ ∇ v ∣ p − 2 ( ∇ 2 v ∇ v ) ⋅ ∇ v ) + 2 b ′ ( v ) ∣ ∇ v ∣ p + 2 . (15) \begin{aligned}
\partial_t|\nabla v|^2
&-a(v)\operatorname{div}\left(|\nabla v|^{p-2}\nabla(|\nabla v|^2)+(p-2)|\nabla v|^{p-4}(\nabla v\cdot\nabla(|\nabla v|^2))\nabla v\right)\\
&\underbrace{-pb(v)|\nabla v|^{p-2}\nabla v\cdot\nabla(|\nabla v|^2)}_{\text{1st-order term}}\\
={}&-2a(v)\left(|\nabla v|^{p-2}|\nabla^2v|^2+(p-2)|\nabla v|^{p-4}|\nabla^2v\nabla v|^2\right)\\
&+2a'(v)\left(|\nabla v|^p\Delta v+(p-2)|\nabla v|^{p-2}(\nabla^2v\nabla v)\cdot\nabla v\right)\\
&+2b'(v)|\nabla v|^{p+2}. \tag*{(15)}
\end{aligned} ∂ t ∣∇ v ∣ 2 = − a ( v ) div ( ∣∇ v ∣ p − 2 ∇ ( ∣∇ v ∣ 2 ) + ( p − 2 ) ∣∇ v ∣ p − 4 ( ∇ v ⋅ ∇ ( ∣∇ v ∣ 2 )) ∇ v ) 1st-order term − p b ( v ) ∣∇ v ∣ p − 2 ∇ v ⋅ ∇ ( ∣∇ v ∣ 2 ) − 2 a ( v ) ( ∣∇ v ∣ p − 2 ∣ ∇ 2 v ∣ 2 + ( p − 2 ) ∣∇ v ∣ p − 4 ∣ ∇ 2 v ∇ v ∣ 2 ) + 2 a ′ ( v ) ( ∣∇ v ∣ p Δ v + ( p − 2 ) ∣∇ v ∣ p − 2 ( ∇ 2 v ∇ v ) ⋅ ∇ v ) + 2 b ′ ( v ) ∣∇ v ∣ p + 2 . (15)
The coefficient a ( v ) a(v) a ( v ) remains outside the divergence throughout the calculation.
Identifying the drift
We work at points where ∇ v ≠ 0 \nabla v\ne0 ∇ v = 0 . The explicit first-order term on the left side has been marked in (15) . To identify the first-order terms hidden in the divergence structure on the left-hand side of (15) , we expand the two contributions to the divergence separately. For the first contribution, the product and chain rules give
div ( ∣ ∇ v ∣ p − 2 ∇ ( ∣ ∇ v ∣ 2 ) ) = ∣ ∇ v ∣ p − 2 Δ ( ∣ ∇ v ∣ 2 ) + ∇ ( ∣ ∇ v ∣ p − 2 ) ⋅ ∇ ( ∣ ∇ v ∣ 2 ) = ∣ ∇ v ∣ p − 2 Δ ( ∣ ∇ v ∣ 2 ) + ( p − 2 ) ∣ ∇ v ∣ p − 4 ( ∇ 2 v ∇ v ) ⋅ ∇ ( ∣ ∇ v ∣ 2 ) . \begin{aligned}
\operatorname{div}\left(|\nabla v|^{p-2}\nabla(|\nabla v|^2)\right)
&=|\nabla v|^{p-2}\Delta(|\nabla v|^2)+\nabla(|\nabla v|^{p-2})\cdot\nabla(|\nabla v|^2)\\
&=|\nabla v|^{p-2}\Delta(|\nabla v|^2)+(p-2)|\nabla v|^{p-4}(\nabla^2v\nabla v)\cdot\nabla(|\nabla v|^2).
\end{aligned} div ( ∣∇ v ∣ p − 2 ∇ ( ∣∇ v ∣ 2 ) ) = ∣∇ v ∣ p − 2 Δ ( ∣∇ v ∣ 2 ) + ∇ ( ∣∇ v ∣ p − 2 ) ⋅ ∇ ( ∣∇ v ∣ 2 ) = ∣∇ v ∣ p − 2 Δ ( ∣∇ v ∣ 2 ) + ( p − 2 ) ∣∇ v ∣ p − 4 ( ∇ 2 v ∇ v ) ⋅ ∇ ( ∣∇ v ∣ 2 ) .
For the second contribution, we first leave aside the constant factor ( p − 2 ) (p-2) ( p − 2 ) and compute
div ( ∣ ∇ v ∣ p − 4 ( ∇ v ⋅ ∇ ( ∣ ∇ v ∣ 2 ) ) ∇ v ) = ( ∇ ( ∣ ∇ v ∣ p − 4 ) ⋅ ∇ v ) ( ∇ v ⋅ ∇ ( ∣ ∇ v ∣ 2 ) ) + ∣ ∇ v ∣ p − 4 ∇ ( ∇ v ⋅ ∇ ( ∣ ∇ v ∣ 2 ) ) ⋅ ∇ v + ∣ ∇ v ∣ p − 4 ( ∇ v ⋅ ∇ ( ∣ ∇ v ∣ 2 ) ) Δ v . \begin{aligned}
\operatorname{div}\left(|\nabla v|^{p-4}(\nabla v\cdot\nabla(|\nabla v|^2))\nabla v\right)
={}&\left(\nabla(|\nabla v|^{p-4})\cdot\nabla v\right)\left(\nabla v\cdot\nabla(|\nabla v|^2)\right)\\
&+|\nabla v|^{p-4}\nabla(\nabla v\cdot\nabla(|\nabla v|^2))\cdot\nabla v\\
&+|\nabla v|^{p-4}(\nabla v\cdot\nabla(|\nabla v|^2))\Delta v.
\end{aligned} div ( ∣∇ v ∣ p − 4 ( ∇ v ⋅ ∇ ( ∣∇ v ∣ 2 )) ∇ v ) = ( ∇ ( ∣∇ v ∣ p − 4 ) ⋅ ∇ v ) ( ∇ v ⋅ ∇ ( ∣∇ v ∣ 2 ) ) + ∣∇ v ∣ p − 4 ∇ ( ∇ v ⋅ ∇ ( ∣∇ v ∣ 2 )) ⋅ ∇ v + ∣∇ v ∣ p − 4 ( ∇ v ⋅ ∇ ( ∣∇ v ∣ 2 )) Δ v .
Using
∇ ( ∣ ∇ v ∣ p − 4 ) = ( p − 4 ) ∣ ∇ v ∣ p − 6 ∇ 2 v ∇ v \nabla(|\nabla v|^{p-4})=(p-4)|\nabla v|^{p-6}\nabla^2v\nabla v ∇ ( ∣∇ v ∣ p − 4 ) = ( p − 4 ) ∣∇ v ∣ p − 6 ∇ 2 v ∇ v
and
∇ ( ∇ v ⋅ ∇ ( ∣ ∇ v ∣ 2 ) ) = ∇ 2 v ∇ ( ∣ ∇ v ∣ 2 ) + ∇ 2 ( ∣ ∇ v ∣ 2 ) ∇ v , \nabla(\nabla v\cdot\nabla(|\nabla v|^2))=\nabla^2v\nabla(|\nabla v|^2)+\nabla^2(|\nabla v|^2)\nabla v, ∇ ( ∇ v ⋅ ∇ ( ∣∇ v ∣ 2 )) = ∇ 2 v ∇ ( ∣∇ v ∣ 2 ) + ∇ 2 ( ∣∇ v ∣ 2 ) ∇ v ,
we can expand the derivatives in the preceding identity. Since ∇ 2 v \nabla^2v ∇ 2 v is symmetric, we also have
( ∇ 2 v ∇ ( ∣ ∇ v ∣ 2 ) ) ⋅ ∇ v = ( ∇ 2 v ∇ v ) ⋅ ∇ ( ∣ ∇ v ∣ 2 ) . (\nabla^2v\nabla(|\nabla v|^2))\cdot\nabla v=(\nabla^2v\nabla v)\cdot\nabla(|\nabla v|^2). ( ∇ 2 v ∇ ( ∣∇ v ∣ 2 )) ⋅ ∇ v = ( ∇ 2 v ∇ v ) ⋅ ∇ ( ∣∇ v ∣ 2 ) .
Consequently, apart from the factor ( p − 2 ) (p-2) ( p − 2 ) the second contribution is
div ( ∣ ∇ v ∣ p − 4 ( ∇ v ⋅ ∇ ( ∣ ∇ v ∣ 2 ) ) ∇ v ) = ( p − 4 ) ∣ ∇ v ∣ p − 6 ( ( ∇ 2 v ∇ v ) ⋅ ∇ v ) ( ∇ v ⋅ ∇ ( ∣ ∇ v ∣ 2 ) ) + ∣ ∇ v ∣ p − 4 ( ∇ 2 v ∇ v ) ⋅ ∇ ( ∣ ∇ v ∣ 2 ) + ∣ ∇ v ∣ p − 4 ( ∇ 2 ( ∣ ∇ v ∣ 2 ) ∇ v ) ⋅ ∇ v . + ∣ ∇ v ∣ p − 4 Δ v ∇ v ⋅ ∇ ( ∣ ∇ v ∣ 2 ) . \begin{aligned}
\operatorname{div}\left(|\nabla v|^{p-4}(\nabla v\cdot\nabla(|\nabla v|^2))\nabla v\right)
={}&(p-4)|\nabla v|^{p-6}\left((\nabla^2v\nabla v)\cdot\nabla v\right)\left(\nabla v\cdot\nabla(|\nabla v|^2)\right)\\
&+|\nabla v|^{p-4}(\nabla^2v\nabla v)\cdot\nabla(|\nabla v|^2)\\
&+|\nabla v|^{p-4}\left(\nabla^2(|\nabla v|^2)\nabla v\right)\cdot\nabla v. \\
&+ |\nabla v|^{p-4}\Delta v\,\nabla v \cdot\nabla(|\nabla v|^2).
\end{aligned} div ( ∣∇ v ∣ p − 4 ( ∇ v ⋅ ∇ ( ∣∇ v ∣ 2 )) ∇ v ) = ( p − 4 ) ∣∇ v ∣ p − 6 ( ( ∇ 2 v ∇ v ) ⋅ ∇ v ) ( ∇ v ⋅ ∇ ( ∣∇ v ∣ 2 ) ) + ∣∇ v ∣ p − 4 ( ∇ 2 v ∇ v ) ⋅ ∇ ( ∣∇ v ∣ 2 ) + ∣∇ v ∣ p − 4 ( ∇ 2 ( ∣∇ v ∣ 2 ) ∇ v ) ⋅ ∇ v . + ∣∇ v ∣ p − 4 Δ v ∇ v ⋅ ∇ ( ∣∇ v ∣ 2 ) .
Multiplying this identity by ( p − 2 ) (p-2) ( p − 2 ) and adding the expansion of the first contribution, we find
div ( ∣ ∇ v ∣ p − 2 ∇ ( ∣ ∇ v ∣ 2 ) + ( p − 2 ) ∣ ∇ v ∣ p − 4 ( ∇ v ⋅ ∇ ( ∣ ∇ v ∣ 2 ) ) ∇ v ) = ∣ ∇ v ∣ p − 2 Δ ( ∣ ∇ v ∣ 2 ) + ( p − 2 ) ∣ ∇ v ∣ p − 4 ( ∇ 2 ( ∣ ∇ v ∣ 2 ) ∇ v ) ⋅ ∇ v + 2 ( p − 2 ) ∣ ∇ v ∣ p − 4 ( ∇ 2 v ∇ v ) ⋅ ∇ ( ∣ ∇ v ∣ 2 ) + ( p − 2 ) ∣ ∇ v ∣ p − 4 Δ v ∇ v ⋅ ∇ ( ∣ ∇ v ∣ 2 ) + ( p − 2 ) ( p − 4 ) ∣ ∇ v ∣ p − 6 ( ( ∇ 2 v ∇ v ) ⋅ ∇ v ) ∇ v ⋅ ∇ ( ∣ ∇ v ∣ 2 ) = ∣ ∇ v ∣ p − 2 Δ ( ∣ ∇ v ∣ 2 ) + ( p − 2 ) ∣ ∇ v ∣ p − 4 ( ∇ 2 ( ∣ ∇ v ∣ 2 ) ∇ v ) ⋅ ∇ v + ( p − 2 ) ∣ ∇ v ∣ p − 4 ( 2 ∇ 2 v ∇ v + Δ v ∇ v ) ⋅ ∇ ( ∣ ∇ v ∣ 2 ) + ( p − 2 ) ( p − 4 ) ∣ ∇ v ∣ p − 6 ( ( ∇ 2 v ∇ v ) ⋅ ∇ v ) ∇ v ⋅ ∇ ( ∣ ∇ v ∣ 2 ) . \begin{aligned}
\operatorname{div}\Bigl(&|\nabla v|^{p-2}\nabla(|\nabla v|^2) +(p-2)|\nabla v|^{p-4}(\nabla v\cdot\nabla(|\nabla v|^2))\nabla v\Bigr)\\
={}&|\nabla v|^{p-2}\Delta(|\nabla v|^2)+(p-2)|\nabla v|^{p-4}(\nabla^2(|\nabla v|^2)\nabla v)\cdot\nabla v\\
&+2(p-2)|\nabla v|^{p-4}(\nabla^2v\nabla v)\cdot\nabla(|\nabla v|^2)\\
&+(p-2)|\nabla v|^{p-4}\Delta v\,\nabla v\cdot\nabla(|\nabla v|^2)\\
&+(p-2)(p-4)|\nabla v|^{p-6}((\nabla^2v\nabla v)\cdot\nabla v)\nabla v\cdot\nabla(|\nabla v|^2)\\
={}&|\nabla v|^{p-2}\Delta(|\nabla v|^2)+(p-2)|\nabla v|^{p-4}(\nabla^2(|\nabla v|^2)\nabla v)\cdot\nabla v\\
&+(p-2)|\nabla v|^{p-4}(2\nabla^2v\nabla v+\Delta v\nabla v)\cdot\nabla(|\nabla v|^2)\\
&+(p-2)(p-4)|\nabla v|^{p-6}((\nabla^2v\nabla v)\cdot\nabla v)\nabla v\cdot\nabla(|\nabla v|^2).
\end{aligned} div ( = = ∣∇ v ∣ p − 2 ∇ ( ∣∇ v ∣ 2 ) + ( p − 2 ) ∣∇ v ∣ p − 4 ( ∇ v ⋅ ∇ ( ∣∇ v ∣ 2 )) ∇ v ) ∣∇ v ∣ p − 2 Δ ( ∣∇ v ∣ 2 ) + ( p − 2 ) ∣∇ v ∣ p − 4 ( ∇ 2 ( ∣∇ v ∣ 2 ) ∇ v ) ⋅ ∇ v + 2 ( p − 2 ) ∣∇ v ∣ p − 4 ( ∇ 2 v ∇ v ) ⋅ ∇ ( ∣∇ v ∣ 2 ) + ( p − 2 ) ∣∇ v ∣ p − 4 Δ v ∇ v ⋅ ∇ ( ∣∇ v ∣ 2 ) + ( p − 2 ) ( p − 4 ) ∣∇ v ∣ p − 6 (( ∇ 2 v ∇ v ) ⋅ ∇ v ) ∇ v ⋅ ∇ ( ∣∇ v ∣ 2 ) ∣∇ v ∣ p − 2 Δ ( ∣∇ v ∣ 2 ) + ( p − 2 ) ∣∇ v ∣ p − 4 ( ∇ 2 ( ∣∇ v ∣ 2 ) ∇ v ) ⋅ ∇ v + ( p − 2 ) ∣∇ v ∣ p − 4 ( 2 ∇ 2 v ∇ v + Δ v ∇ v ) ⋅ ∇ ( ∣∇ v ∣ 2 ) + ( p − 2 ) ( p − 4 ) ∣∇ v ∣ p − 6 (( ∇ 2 v ∇ v ) ⋅ ∇ v ) ∇ v ⋅ ∇ ( ∣∇ v ∣ 2 ) .
We have identified the first-order terms hidden in the divergence. The factor 2 multiplying ( ∇ 2 v ∇ v ) ⋅ ∇ ( ∣ ∇ v ∣ 2 ) (\nabla^2v\nabla v)\cdot\nabla(|\nabla v|^2) ( ∇ 2 v ∇ v ) ⋅ ∇ ( ∣∇ v ∣ 2 ) after the first equality arises from combining the identical contributions from the two expansions, after multiplying the second expansion by ( p − 2 ) (p-2) ( p − 2 ) .
Now we can identify the first order terms on the left side of (15) . Indeed, multiplying the last display by − a ( v ) -a(v) − a ( v ) , as prescribed by (15) , and combining them with the explicit first-order term
− p b ( v ) ∣ ∇ v ∣ p − 2 ∇ v ⋅ ∇ ( ∣ ∇ v ∣ 2 ) , -pb(v)|\nabla v|^{p-2}\nabla v\cdot\nabla(|\nabla v|^2), − p b ( v ) ∣∇ v ∣ p − 2 ∇ v ⋅ ∇ ( ∣∇ v ∣ 2 ) ,
we can write the first-order terms on the left-hand side as
− B ⋅ ∇ ( ∣ ∇ v ∣ 2 ) , -\mathbf{B}\cdot\nabla(|\nabla v|^2), − B ⋅ ∇ ( ∣∇ v ∣ 2 ) ,
where
B = ( p − 2 ) a ( v ) ∣ ∇ v ∣ p − 4 ( 2 ∇ 2 v ∇ v + Δ v ∇ v ) + ( p − 2 ) ( p − 4 ) a ( v ) ∣ ∇ v ∣ p − 6 ( ( ∇ 2 v ∇ v ) ⋅ ∇ v ) ∇ v + p b ( v ) ∣ ∇ v ∣ p − 2 ∇ v . (16) \mathbf{B}=(p-2)a(v)|\nabla v|^{p-4}(2\nabla^2v\nabla v+\Delta v\nabla v)
+(p-2)(p-4)a(v)|\nabla v|^{p-6}((\nabla^2v\nabla v)\cdot\nabla v)\nabla v
+pb(v)|\nabla v|^{p-2}\nabla v. \tag*{(16)} B = ( p − 2 ) a ( v ) ∣∇ v ∣ p − 4 ( 2 ∇ 2 v ∇ v + Δ v ∇ v ) + ( p − 2 ) ( p − 4 ) a ( v ) ∣∇ v ∣ p − 6 (( ∇ 2 v ∇ v ) ⋅ ∇ v ) ∇ v + p b ( v ) ∣∇ v ∣ p − 2 ∇ v . (16)
The terms containing a ( v ) a(v) a ( v ) remain on the right-hand side. With this definition, equation (15) becomes
∂ t ∣ ∇ v ∣ 2 − a ( v ) ∣ ∇ v ∣ p − 2 [ Δ ( ∣ ∇ v ∣ 2 ) + ( p − 2 ) ( ∇ 2 ( ∣ ∇ v ∣ 2 ) ∇ v ) ⋅ ∇ v ∣ ∇ v ∣ 2 ] − B ⋅ ∇ ( ∣ ∇ v ∣ 2 ) = − 2 a ( v ) ∣ ∇ v ∣ p − 2 [ ∣ ∇ 2 v ∣ 2 + ( p − 2 ) ∣ ∇ 2 v ∇ v ∣ 2 ∣ ∇ v ∣ 2 ] + 2 a ′ ( v ) ∣ ∇ v ∣ p [ Δ v + ( p − 2 ) ( ∇ 2 v ∇ v ) ⋅ ∇ v ∣ ∇ v ∣ 2 ] + 2 b ′ ( v ) ∣ ∇ v ∣ p + 2 . (17) \begin{aligned}
\partial_t|\nabla v|^2-a(v)|\nabla v|^{p-2}\left[\Delta(|\nabla v|^2)+(p-2)\frac{(\nabla^2(|\nabla v|^2)\nabla v)\cdot\nabla v}{|\nabla v|^2}\right]-\mathbf{B}\cdot\nabla(|\nabla v|^2)\\
=-2a(v)|\nabla v|^{p-2}\left[|\nabla^2v|^2+(p-2)\frac{|\nabla^2v\nabla v|^2}{|\nabla v|^2}\right]\\
\quad+2a'(v)|\nabla v|^p\left[\Delta v+(p-2)\frac{(\nabla^2v\nabla v)\cdot\nabla v}{|\nabla v|^2}\right]\\
\quad+2b'(v)|\nabla v|^{p+2}. \tag*{(17)}
\end{aligned} ∂ t ∣∇ v ∣ 2 − a ( v ) ∣∇ v ∣ p − 2 [ Δ ( ∣∇ v ∣ 2 ) + ( p − 2 ) ∣∇ v ∣ 2 ( ∇ 2 ( ∣∇ v ∣ 2 ) ∇ v ) ⋅ ∇ v ] − B ⋅ ∇ ( ∣∇ v ∣ 2 ) = − 2 a ( v ) ∣∇ v ∣ p − 2 [ ∣ ∇ 2 v ∣ 2 + ( p − 2 ) ∣∇ v ∣ 2 ∣ ∇ 2 v ∇ v ∣ 2 ] + 2 a ′ ( v ) ∣∇ v ∣ p [ Δ v + ( p − 2 ) ∣∇ v ∣ 2 ( ∇ 2 v ∇ v ) ⋅ ∇ v ] + 2 b ′ ( v ) ∣∇ v ∣ p + 2 . (17)
Estimating the drift
We estimate the three contributions to B \mathbf{B} B in (16) . Observe the elementary inequalities
∣ Δ v ∣ ≤ N ∣ ∇ 2 v ∣ , ∣ ∇ 2 v ∇ v ∣ ≤ ∣ ∇ 2 v ∣ ∣ ∇ v ∣ ; |\Delta v|\le\sqrt{N}|\nabla^2v|,\qquad|\nabla^2v\nabla v|\le|\nabla^2v||\nabla v|; ∣Δ v ∣ ≤ N ∣ ∇ 2 v ∣ , ∣ ∇ 2 v ∇ v ∣ ≤ ∣ ∇ 2 v ∣∣∇ v ∣ ;
the latter implies
∣ ( ∇ 2 v ∇ v ) ⋅ ∇ v ∣ ≤ ∣ ∇ 2 v ∣ ∣ ∇ v ∣ 2 . |(\nabla^2v\nabla v)\cdot\nabla v|\le|\nabla^2v||\nabla v|^2. ∣ ( ∇ 2 v ∇ v ) ⋅ ∇ v ∣ ≤ ∣ ∇ 2 v ∣∣∇ v ∣ 2 .
Here ∣ ∇ 2 v ∣ |\nabla^2v| ∣ ∇ 2 v ∣ denotes the Frobenius norm.
For the first contribution, we obtain
∣ ( p − 2 ) a ( v ) ∣ ∇ v ∣ p − 4 ( 2 ∇ 2 v ∇ v + Δ v ∇ v ) ∣ ≤ ∣ p − 2 ∣ a ( v ) ∣ ∇ v ∣ p − 4 ( 2 ∣ ∇ 2 v ∇ v ∣ + ∣ Δ v ∣ ∣ ∇ v ∣ ) ≤ ∣ p − 2 ∣ ( 2 + N ) a ( v ) ∣ ∇ v ∣ p − 3 ∣ ∇ 2 v ∣ . \left|(p-2)a(v)|\nabla v|^{p-4}\left(2\nabla^{2}v\nabla v+\Delta v\nabla v\right)\right|
\le|p-2|a(v)|\nabla v|^{p-4}\left(2|\nabla^{2}v\nabla v|+|\Delta v||\nabla v|\right)
\le|p-2|(2+\sqrt{N})a(v)|\nabla v|^{p-3}|\nabla^{2}v|. ( p − 2 ) a ( v ) ∣∇ v ∣ p − 4 ( 2 ∇ 2 v ∇ v + Δ v ∇ v ) ≤ ∣ p − 2∣ a ( v ) ∣∇ v ∣ p − 4 ( 2∣ ∇ 2 v ∇ v ∣ + ∣Δ v ∣∣∇ v ∣ ) ≤ ∣ p − 2∣ ( 2 + N ) a ( v ) ∣∇ v ∣ p − 3 ∣ ∇ 2 v ∣.
For the second contribution,
∣ ( p − 2 ) ( p − 4 ) a ( v ) ∣ ∇ v ∣ p − 6 ( ( ∇ 2 v ∇ v ) ⋅ ∇ v ) ∇ v ∣ ≤ ∣ p − 2 ∣ ∣ p − 4 ∣ a ( v ) ∣ ∇ v ∣ p − 6 ∣ ∇ 2 v ∣ ∣ ∇ v ∣ 3 = ∣ p − 2 ∣ ∣ p − 4 ∣ a ( v ) ∣ ∇ v ∣ p − 3 ∣ ∇ 2 v ∣ . \left|(p-2)(p-4)a(v)|\nabla v|^{p-6}\left((\nabla^{2}v\nabla v)\cdot\nabla v\right)\nabla v\right|
\le|p-2||p-4|a(v)|\nabla v|^{p-6}|\nabla^{2}v||\nabla v|^{3}
=|p-2||p-4|a(v)|\nabla v|^{p-3}|\nabla^{2}v|. ( p − 2 ) ( p − 4 ) a ( v ) ∣∇ v ∣ p − 6 ( ( ∇ 2 v ∇ v ) ⋅ ∇ v ) ∇ v ≤ ∣ p − 2∣∣ p − 4∣ a ( v ) ∣∇ v ∣ p − 6 ∣ ∇ 2 v ∣∣∇ v ∣ 3 = ∣ p − 2∣∣ p − 4∣ a ( v ) ∣∇ v ∣ p − 3 ∣ ∇ 2 v ∣.
Finally, since b ( v ) > 0 b(v)>0 b ( v ) > 0 ,
∣ p b ( v ) ∣ ∇ v ∣ p − 2 ∇ v ∣ = p b ( v ) ∣ ∇ v ∣ p − 1 . |pb(v)|\nabla v|^{p-2}\nabla v|=pb(v)|\nabla v|^{p-1}. ∣ p b ( v ) ∣∇ v ∣ p − 2 ∇ v ∣ = p b ( v ) ∣∇ v ∣ p − 1 .
Combining these estimates gives
∣ B ∣ ≤ ∣ p − 2 ∣ ( 2 + N + ∣ p − 4 ∣ ) a ( v ) ∣ ∇ v ∣ p − 3 ∣ ∇ 2 v ∣ + p b ( v ) ∣ ∇ v ∣ p − 1 . |\mathbf{B}|\le|p-2|(2+\sqrt{N}+|p-4|)a(v)|\nabla v|^{p-3}|\nabla^{2}v|
+pb(v)|\nabla v|^{p-1}. ∣ B ∣ ≤ ∣ p − 2∣ ( 2 + N + ∣ p − 4∣ ) a ( v ) ∣∇ v ∣ p − 3 ∣ ∇ 2 v ∣ + p b ( v ) ∣∇ v ∣ p − 1 .
Thus,
∣ B ∣ ≤ C ( N , p ) ( a ( v ) ∣ ∇ v ∣ p − 3 ∣ ∇ 2 v ∣ + b ( v ) ∣ ∇ v ∣ p − 1 ) , (18) |\mathbf{B}|\le C(N,p)\left(a(v)|\nabla v|^{p-3}|\nabla^{2}v|+b(v)|\nabla v|^{p-1}\right), \tag*{(18)} ∣ B ∣ ≤ C ( N , p ) ( a ( v ) ∣∇ v ∣ p − 3 ∣ ∇ 2 v ∣ + b ( v ) ∣∇ v ∣ p − 1 ) , (18)
where C ( N , p ) C(N,p) C ( N , p ) depends only on N N N and p p p .
Simplifying the notation
We set
w : = ∣ ∇ v ∣ 2 , P : = I + ( p − 2 ) ∇ v ∣ ∇ v ∣ ⊗ ∇ v ∣ ∇ v ∣ . w:=|\nabla v|^{2},\qquad\mathbf{P}:=I+(p-2)\frac{\nabla v}{|\nabla v|}\otimes\frac{\nabla v}{|\nabla v|}. w := ∣∇ v ∣ 2 , P := I + ( p − 2 ) ∣∇ v ∣ ∇ v ⊗ ∣∇ v ∣ ∇ v .
where I I I denotes the identity matrix. With the convention C ⋅ D = tr ( C T D ) C\mathbin{\cdot}D=\operatorname{tr}(C^{T}D) C ⋅ D = tr ( C T D ) , we have
P ⋅ ∇ 2 w = Δ ( ∣ ∇ v ∣ 2 ) + ( p − 2 ) ( ∇ 2 ( ∣ ∇ v ∣ 2 ) ∇ v ) ⋅ ∇ v ∣ ∇ v ∣ 2 , \mathbf{P}\mathbin{\cdot}\nabla^{2}w
=\Delta(|\nabla v|^{2})+(p-2)\frac{\left(\nabla^{2}(|\nabla v|^{2})\nabla v\right)\mathbin{\cdot}\nabla v}{|\nabla v|^{2}}, P ⋅ ∇ 2 w = Δ ( ∣∇ v ∣ 2 ) + ( p − 2 ) ∣∇ v ∣ 2 ( ∇ 2 ( ∣∇ v ∣ 2 ) ∇ v ) ⋅ ∇ v ,
and
P ⋅ ∇ 2 v = Δ v + ( p − 2 ) ( ∇ 2 v ∇ v ) ⋅ ∇ v ∣ ∇ v ∣ 2 . \mathbf{P}\mathbin{\cdot}\nabla^{2}v
=\Delta v+(p-2)\frac{(\nabla^{2}v\nabla v)\mathbin{\cdot}\nabla v}{|\nabla v|^{2}}. P ⋅ ∇ 2 v = Δ v + ( p − 2 ) ∣∇ v ∣ 2 ( ∇ 2 v ∇ v ) ⋅ ∇ v .
Moreover,
tr ( P ( ∇ 2 v ) 2 ) = ∣ ∇ 2 v ∣ 2 + ( p − 2 ) ∣ ∇ 2 v ∇ v ∣ 2 ∣ ∇ v ∣ 2 . \operatorname{tr}\left(\mathbf{P}(\nabla^{2}v)^{2}\right)
=|\nabla^{2}v|^{2}+(p-2)\frac{|\nabla^{2}v\nabla v|^{2}}{|\nabla v|^{2}}. tr ( P ( ∇ 2 v ) 2 ) = ∣ ∇ 2 v ∣ 2 + ( p − 2 ) ∣∇ v ∣ 2 ∣ ∇ 2 v ∇ v ∣ 2 .
To verify the last identity, we use the definition of P \mathbf{P} P and the linearity of the trace:
tr ( P ( ∇ 2 v ) 2 ) = tr ( ( ∇ 2 v ) 2 ) + p − 2 ∣ ∇ v ∣ 2 tr ( ( ∇ v ⊗ ∇ v ) ( ∇ 2 v ) 2 ) . \operatorname{tr}\left(\mathbf{P}(\nabla^{2}v)^{2}\right)
=\operatorname{tr}\left((\nabla^{2}v)^{2}\right)+\frac{p-2}{|\nabla v|^{2}}\operatorname{tr}\left((\nabla v\otimes\nabla v)(\nabla^{2}v)^{2}\right). tr ( P ( ∇ 2 v ) 2 ) = tr ( ( ∇ 2 v ) 2 ) + ∣∇ v ∣ 2 p − 2 tr ( ( ∇ v ⊗ ∇ v ) ( ∇ 2 v ) 2 ) .
Since ∇ 2 v \nabla^{2}v ∇ 2 v is symmetric,
tr ( ( ∇ 2 v ) 2 ) = tr ( ( ∇ 2 v ) T ∇ 2 v ) = ∣ ∇ 2 v ∣ 2 . \operatorname{tr}\left((\nabla^{2}v)^{2}\right)
=\operatorname{tr}\left((\nabla^{2}v)^{T}\nabla^{2}v\right)
=|\nabla^{2}v|^{2}. tr ( ( ∇ 2 v ) 2 ) = tr ( ( ∇ 2 v ) T ∇ 2 v ) = ∣ ∇ 2 v ∣ 2 .
Furthermore, the identity tr ( ( z ⊗ z ) A ) = z ⋅ A z \operatorname{tr}((z\otimes z)A)=z\mathbin{\cdot}Az tr (( z ⊗ z ) A ) = z ⋅ A z gives
tr ( ( ∇ v ⊗ ∇ v ) ( ∇ 2 v ) 2 ) = ∇ v ⋅ ( ∇ 2 v ) 2 ∇ v = ( ∇ 2 v ∇ v ) ⋅ ( ∇ 2 v ∇ v ) = ∣ ∇ 2 v ∇ v ∣ 2 , \operatorname{tr}\left((\nabla v\otimes\nabla v)(\nabla^{2}v)^{2}\right)
=\nabla v\mathbin{\cdot}(\nabla^{2}v)^{2}\nabla v
=(\nabla^{2}v\nabla v)\mathbin{\cdot}(\nabla^{2}v\nabla v)
=|\nabla^{2}v\nabla v|^{2}, tr ( ( ∇ v ⊗ ∇ v ) ( ∇ 2 v ) 2 ) = ∇ v ⋅ ( ∇ 2 v ) 2 ∇ v = ( ∇ 2 v ∇ v ) ⋅ ( ∇ 2 v ∇ v ) = ∣ ∇ 2 v ∇ v ∣ 2 ,
where we again used the symmetry of ∇ 2 v \nabla^{2}v ∇ 2 v . Combining these identities proves the claimed formula.
We define the operator L \mathcal{L} L by
L φ : = ∂ t φ − a ( v ) w p − 2 2 P ⋅ ∇ 2 φ − B ⋅ ∇ φ (19) \mathcal{L}\varphi:=\partial_{t}\varphi-a(v)w^{\frac{p-2}{2}}\mathbf{P}\mathbin{\cdot}\nabla^{2}\varphi-\mathbf{B}\mathbin{\cdot}\nabla\varphi \tag*{(19)} L φ := ∂ t φ − a ( v ) w 2 p − 2 P ⋅ ∇ 2 φ − B ⋅ ∇ φ (19)
for smooth scalar functions φ \varphi φ .
Two remarks are in order here. First, B \mathbf{B} B depends on ∇ 2 v \nabla^{2}v ∇ 2 v , so L \mathcal{L} L is not an operator with a priori controlled coefficients, and no maximum principle for L \mathcal{L} L on the whole cylinder is available. We use L \mathcal{L} L only pointwise , at a maximum point of η 2 w \eta^{2}w η 2 w where ∣ ∇ v ∣ > 0 |\nabla v|>0 ∣∇ v ∣ > 0 . This is why half of the negative Hessian term must be retained in the Young step below: it absorbs the ∣ ∇ 2 v ∣ |\nabla^{2}v| ∣ ∇ 2 v ∣ coming from ∣ B ∣ |\mathbf{B}| ∣ B ∣ . Second, w p − 3 2 w^{\frac{p-3}{2}} w 2 p − 3 is singular at ∇ v = 0 \nabla v=0 ∇ v = 0 for p < 3 p<3 p < 3 ; in the localization step it always appears multiplied by w w w , giving w p − 1 2 w^{\frac{p-1}{2}} w 2 p − 1 with a positive exponent, so no singularity survives.
With the above notation, equation (17) becomes
L w = − 2 a ( v ) w p − 2 2 tr ( P ( ∇ 2 v ) 2 ) + 2 a ′ ( v ) w p 2 ( P ⋅ ∇ 2 v ) + 2 b ′ ( v ) w p + 2 2 (20) \mathcal{L}w=-2a(v)w^{\frac{p-2}{2}}\operatorname{tr}\left(\mathbf{P}(\nabla^{2}v)^{2}\right)+2a'(v)w^{\frac{p}{2}}\left(\mathbf{P}\mathbin{\vcenter{\text{\(\cdot\)}}}\nabla^{2}v\right)+2b'(v)w^{\frac{p+2}{2}} \tag*{(20)} L w = − 2 a ( v ) w 2 p − 2 tr ( P ( ∇ 2 v ) 2 ) + 2 a ′ ( v ) w 2 p ( P ⋅ ∇ 2 v ) + 2 b ′ ( v ) w 2 p + 2 (20)
In terms of w w w , the drift estimate (18) reads
∣ B ∣ ≤ C ( N , p ) ( a ( v ) w p − 3 2 ∣ ∇ 2 v ∣ + b ( v ) w p − 1 2 ) . |\mathbf{B}|\le C(N,p)\left(a(v)w^{\frac{p-3}{2}}|\nabla^{2}v|+b(v)w^{\frac{p-1}{2}}\right). ∣ B ∣ ≤ C ( N , p ) ( a ( v ) w 2 p − 3 ∣ ∇ 2 v ∣ + b ( v ) w 2 p − 1 ) .
For later use, we record the following properties of P \mathbf{P} P . For every ξ ∈ R N \xi\in\mathbb{R}^{N} ξ ∈ R N ,
ξ ⋅ P ξ = ∣ ξ ∣ 2 + ( p − 2 ) ( ξ ⋅ ∇ v ∣ ∇ v ∣ ) 2 . \xi\mathbin{\vcenter{\text{\(\cdot\)}}}\mathbf{P}\xi=|\xi|^{2}+(p-2)\left(\xi\mathbin{\vcenter{\text{\(\cdot\)}}}\frac{\nabla v}{|\nabla v|}\right)^{2}. ξ ⋅ P ξ = ∣ ξ ∣ 2 + ( p − 2 ) ( ξ ⋅ ∣∇ v ∣ ∇ v ) 2 .
Consequently,
min { 1 , p − 1 } ∣ ξ ∣ 2 ≤ ξ ⋅ P ξ ≤ max { 1 , p − 1 } ∣ ξ ∣ 2 . (21) \min\{1,p-1\}|\xi|^{2}\le\xi\mathbin{\vcenter{\text{\(\cdot\)}}}\mathbf{P}\xi\le\max\{1,p-1\}|\xi|^{2}. \tag*{(21)} min { 1 , p − 1 } ∣ ξ ∣ 2 ≤ ξ ⋅ P ξ ≤ max { 1 , p − 1 } ∣ ξ ∣ 2 . (21)
In particular, P \mathbf{P} P is positive definite. Its eigenvalue in the direction of ∇ v \nabla v ∇ v is p − 1 p-1 p − 1 , and its N − 1 N-1 N − 1 eigenvalues in the orthogonal directions are 1 1 1 . Hence
tr P = N + p − 2. \operatorname{tr}\mathbf{P}=N+p-2. tr P = N + p − 2.
Since ∇ 2 v \nabla^{2}v ∇ 2 v is symmetric, cyclicity of the trace gives
tr ( P ( ∇ 2 v ) 2 ) = tr ( ∇ 2 v P ∇ 2 v ) = ∑ j = 1 N ( ∇ 2 v e j ) ⋅ P ( ∇ 2 v e j ) , \operatorname{tr}\left(\mathbf{P}(\nabla^{2}v)^{2}\right)=\operatorname{tr}\left(\nabla^{2}v\mathbf{P}\nabla^{2}v\right)=\sum_{j=1}^{N}(\nabla^{2}v e_{j})\mathbin{\vcenter{\text{\(\cdot\)}}}\mathbf{P}(\nabla^{2}v e_{j}), tr ( P ( ∇ 2 v ) 2 ) = tr ( ∇ 2 v P ∇ 2 v ) = j = 1 ∑ N ( ∇ 2 v e j ) ⋅ P ( ∇ 2 v e j ) ,
and applying (21) to the vectors ∇ 2 v e j \nabla^{2}v e_{j} ∇ 2 v e j yields
min { 1 , p − 1 } ∣ ∇ 2 v ∣ 2 ≤ tr ( P ( ∇ 2 v ) 2 ) ≤ max { 1 , p − 1 } ∣ ∇ 2 v ∣ 2 , \min\{1,p-1\}|\nabla^{2}v|^{2}\le\operatorname{tr}\left(\mathbf{P}(\nabla^{2}v)^{2}\right)\le\max\{1,p-1\}|\nabla^{2}v|^{2}, min { 1 , p − 1 } ∣ ∇ 2 v ∣ 2 ≤ tr ( P ( ∇ 2 v ) 2 ) ≤ max { 1 , p − 1 } ∣ ∇ 2 v ∣ 2 ,
because ∑ j ∣ ∇ 2 v e j ∣ 2 = ∣ ∇ 2 v ∣ 2 \sum_{j}|\nabla^{2}v e_{j}|^{2}=|\nabla^{2}v|^{2} ∑ j ∣ ∇ 2 v e j ∣ 2 = ∣ ∇ 2 v ∣ 2 .
The negative coefficient and the Hessian term
Applying the Cauchy–Schwarz inequality for the Frobenius inner product, we obtain
( P ⋅ ∇ 2 v ) 2 ≤ ( tr P ) tr ( P ( ∇ 2 v ) 2 ) . \left(\mathbf{P}\mathbin{\vcenter{\text{\(\cdot\)}}}\nabla^{2}v\right)^{2}\le(\operatorname{tr}\mathbf{P})\operatorname{tr}\left(\mathbf{P}(\nabla^{2}v)^{2}\right). ( P ⋅ ∇ 2 v ) 2 ≤ ( tr P ) tr ( P ( ∇ 2 v ) 2 ) .
Indeed, writing P 1 2 \mathbf{P}^{\frac{1}{2}} P 2 1 for the symmetric positive definite square root of P \mathbf{P} P , we have
P ⋅ ∇ 2 v = P 1 2 ⋅ ( P 1 2 ∇ 2 v ) . \mathbf{P}\mathbin{\vcenter{\text{\(\cdot\)}}}\nabla^{2}v=\mathbf{P}^{\frac{1}{2}}\mathbin{\vcenter{\text{\(\cdot\)}}}\left(\mathbf{P}^{\frac{1}{2}}\nabla^{2}v\right). P ⋅ ∇ 2 v = P 2 1 ⋅ ( P 2 1 ∇ 2 v ) .
The Frobenius Cauchy–Schwarz inequality therefore yields
( P ⋅ ∇ 2 v ) 2 ≤ ∣ P 1 2 ∣ 2 ∣ P 1 2 ∇ 2 v ∣ 2 = ( tr P ) tr ( P ( ∇ 2 v ) 2 ) . \left(\mathbf{P}\mathbin{\vcenter{\text{\(\cdot\)}}}\nabla^{2}v\right)^{2}\le\left|\mathbf{P}^{\frac{1}{2}}\right|^{2}\left|\mathbf{P}^{\frac{1}{2}}\nabla^{2}v\right|^{2}=(\operatorname{tr}\mathbf{P})\operatorname{tr}\left(\mathbf{P}(\nabla^{2}v)^{2}\right). ( P ⋅ ∇ 2 v ) 2 ≤ P 2 1 2 P 2 1 ∇ 2 v 2 = ( tr P ) tr ( P ( ∇ 2 v ) 2 ) .
Using tr P = N + p − 2 \operatorname{tr}\mathbf{P}=N+p-2 tr P = N + p − 2 , we can estimate the mixed term in (20) as follows:
2 a ′ ( v ) w p 2 ( P ⋅ ∇ 2 v ) 2a'(v)w^{\frac{p}{2}}\left(\mathbf{P}\mathbin{\vcenter{\text{\(\cdot\)}}}\nabla^{2}v\right) 2 a ′ ( v ) w 2 p ( P ⋅ ∇ 2 v )
≤ 2 ∣ a ′ ( v ) ∣ w p 2 N + p − 2 tr ( P ( ∇ 2 v ) 2 ) = 2 [ a ( v ) w p − 2 4 tr ( P ( ∇ 2 v ) 2 ) ] [ N + p − 2 ∣ a ′ ( v ) ∣ a ( v ) w p + 2 4 ] ≤ a ( v ) w p − 2 2 tr ( P ( ∇ 2 v ) 2 ) + ( N + p − 2 ) ( a ′ ( v ) ) 2 a ( v ) w p + 2 2 , \begin{aligned}
\le2|a'(v)|w^{\frac{p}{2}}\sqrt{N+p-2}\sqrt{\operatorname{tr}\left(\mathbf{P}(\nabla^{2}v)^{2}\right)} \\
&=2\left[\sqrt{a(v)}w^{\frac{p-2}{4}}\sqrt{\operatorname{tr}\left(\mathbf{P}(\nabla^{2}v)^{2}\right)}\right]\left[\sqrt{N+p-2}\frac{|a'(v)|}{\sqrt{a(v)}}w^{\frac{p+2}{4}}\right] \\
&\le a(v)w^{\frac{p-2}{2}}\operatorname{tr}\left(\mathbf{P}(\nabla^{2}v)^{2}\right)+(N+p-2)\frac{(a'(v))^{2}}{a(v)}w^{\frac{p+2}{2}},
\end{aligned} ≤ 2∣ a ′ ( v ) ∣ w 2 p N + p − 2 tr ( P ( ∇ 2 v ) 2 ) = 2 [ a ( v ) w 4 p − 2 tr ( P ( ∇ 2 v ) 2 ) ] [ N + p − 2 a ( v ) ∣ a ′ ( v ) ∣ w 4 p + 2 ] ≤ a ( v ) w 2 p − 2 tr ( P ( ∇ 2 v ) 2 ) + ( N + p − 2 ) a ( v ) ( a ′ ( v ) ) 2 w 2 p + 2 ,
where the last step follows from Young’s inequality. Substituting this estimate into (20) gives
L w ≤ − a ( v ) w p − 2 2 tr ( P ( ∇ 2 v ) 2 ) + [ 2 b ′ ( v ) + ( N + p − 2 ) ( a ′ ( v ) ) 2 a ( v ) ] w p + 2 2 . (22) \mathcal{L}w \leq-a(v)w^{\frac{p-2}{2}}\operatorname{tr}\left(\mathbf{P}(\nabla^{2}v)^2\right)+\left[2b'(v)+(N+p-2)\frac{(a'(v))^2}{a(v)}\right]w^{\frac{p+2}{2}}. \tag*{(22)} L w ≤ − a ( v ) w 2 p − 2 tr ( P ( ∇ 2 v ) 2 ) + [ 2 b ′ ( v ) + ( N + p − 2 ) a ( v ) ( a ′ ( v ) ) 2 ] w 2 p + 2 . (22)
We now compute the coefficient in parentheses. Since
a ( v ) = c v 2 − p , a ′ ( v ) = c ( 2 − p ) v 1 − p , a(v)=cv^{2-p},\qquad a'(v)=c(2-p)v^{1-p}, a ( v ) = c v 2 − p , a ′ ( v ) = c ( 2 − p ) v 1 − p ,
we have
( a ′ ( v ) ) 2 a ( v ) = c 2 ( p − 2 ) 2 v 2 − 2 p c v 2 − p = c ( p − 2 ) 2 v − p . \frac{(a'(v))^2}{a(v)}=\frac{c^2(p-2)^2v^{2-2p}}{cv^{2-p}}=c(p-2)^2v^{-p}. a ( v ) ( a ′ ( v ) ) 2 = c v 2 − p c 2 ( p − 2 ) 2 v 2 − 2 p = c ( p − 2 ) 2 v − p .
Furthermore,
b ′ ( v ) = − c ( p − 1 ) 2 ( k − 1 ) v − p . b'(v)=-c(p-1)^2(k-1)v^{-p}. b ′ ( v ) = − c ( p − 1 ) 2 ( k − 1 ) v − p .
Recalling A ∗ = ( N + p − 2 ) ( p − 2 ) 2 A_*=(N+p-2)(p-2)^2 A ∗ = ( N + p − 2 ) ( p − 2 ) 2 , we obtain
2 b ′ ( v ) + ( N + p − 2 ) ( a ′ ( v ) ) 2 a ( v ) = c v − p [ A ∗ − 2 ( p − 1 ) 2 ( k − 1 ) ] ≤ − c [ A ∗ + 2 ( p − 1 ) 2 ] v − p . (23) \begin{aligned}
2b'(v)+(N+p-2)\frac{(a'(v))^2}{a(v)}
=cv^{-p}\left[A_*-2(p-1)^2(k-1)\right] \\
&\leq-c\left[A_*+2(p-1)^2\right]v^{-p}. \tag*{(23)}
\end{aligned} 2 b ′ ( v ) + ( N + p − 2 ) a ( v ) ( a ′ ( v ) ) 2 = c v − p [ A ∗ − 2 ( p − 1 ) 2 ( k − 1 ) ] ≤ − c [ A ∗ + 2 ( p − 1 ) 2 ] v − p . (23)
Here we used
k ≥ 2 + A ∗ ( p − 1 ) 2 , k\geq2+\frac{A_*}{(p-1)^2}, k ≥ 2 + ( p − 1 ) 2 A ∗ ,
which implies
A ∗ − 2 ( p − 1 ) 2 ( k − 1 ) ≤ − A ∗ − 2 ( p − 1 ) 2 . A_*-2(p-1)^2(k-1)\leq-A_*-2(p-1)^2. A ∗ − 2 ( p − 1 ) 2 ( k − 1 ) ≤ − A ∗ − 2 ( p − 1 ) 2 .
Recalling that c = k p − 2 p − 1 / ( p − 1 ) c=k^{\frac{p-2}{p-1}}/(p-1) c = k p − 1 p − 2 / ( p − 1 ) , we set
κ : = k p − 2 p − 1 p − 1 [ A ∗ + 2 ( p − 1 ) 2 ] > 0. \kappa:=\frac{k^{\frac{p-2}{p-1}}}{p-1}\left[A_*+2(p-1)^2\right]>0. κ := p − 1 k p − 1 p − 2 [ A ∗ + 2 ( p − 1 ) 2 ] > 0.
Thus, κ \kappa κ depends only on N , p , k N,p,k N , p , k . We conclude from (22) that
L w ≤ − a ( v ) w p − 2 2 tr ( P ( ∇ 2 v ) 2 ) − κ v − p w p + 2 2 . (24) \mathcal{L}w\leq-a(v)w^{\frac{p-2}{2}}\operatorname{tr}\left(\mathbf{P}(\nabla^{2}v)^2\right)-\kappa v^{-p}w^{\frac{p+2}{2}}. \tag*{(24)} L w ≤ − a ( v ) w 2 p − 2 tr ( P ( ∇ 2 v ) 2 ) − κ v − p w 2 p + 2 . (24)
Remark 3.1. Young’s inequality separates the negative Hessian contribution. The remaining term
− a ( v ) w p − 2 2 tr ( P ( ∇ 2 v ) 2 ) -a(v)w^{\frac{p-2}{2}}\operatorname{tr}\left(\mathbf{P}(\nabla^{2}v)^2\right) − a ( v ) w 2 p − 2 tr ( P ( ∇ 2 v ) 2 )
will control the terms involving ∣ ∇ 2 v ∣ |\nabla^2v| ∣ ∇ 2 v ∣ that arise from the drift when a cutoff function is introduced.
Remark 3.2. The proof of (23) only needs the coefficient A ∗ − 2 ( p − 1 ) 2 ( k − 1 ) A_*-2(p-1)^2(k-1) A ∗ − 2 ( p − 1 ) 2 ( k − 1 ) to be strictly negative, i.e., k > 1 + A ∗ 2 ( p − 1 ) 2 k>1+\frac{A_*}{2(p-1)^2} k > 1 + 2 ( p − 1 ) 2 A ∗ ; then (24) would hold with κ = k p − 2 p − 1 [ 2 ( p − 1 ) 2 ( k − 1 ) − A ∗ ] \kappa=\frac{k^{p-2}}{p-1}\left[2(p-1)^2(k-1)-A_*\right] κ = p − 1 k p − 2 [ 2 ( p − 1 ) 2 ( k − 1 ) − A ∗ ] . The condition k ≥ k o = 2 + A ∗ ( p − 1 ) 2 k\geq k_o=2+\frac{A_*}{(p-1)^2} k ≥ k o = 2 + ( p − 1 ) 2 A ∗ is a convenient sufficient choice. The estimate for u u u itself is finally obtained with the fixed value k = k o ( N , p ) k=k_o(N,p) k = k o ( N , p ) , so the dependence of C C C on k k k is harmless.
A local gradient estimate via Bernstein’s method
Lemma 3.3. Let p > 1 p>1 p > 1 and let k ≥ k o k\geq k_o k ≥ k o , where k o k_o k o is defined in (9) . Suppose that u u u is a smooth nonnegative solution to (1) in Ω T \Omega_T Ω T . Let r , h > 0 r,h>0 r , h > 0 , and let
Q : = B 2 r ( x o ) × ( s − 2 h , s ] Q:=B_{2r}(x_o)\times(s-2h,s] Q := B 2 r ( x o ) × ( s − 2 h , s ] satisfy
Q ‾ ⊂ { ( x , t ) ∈ Ω T : u ( x , t ) > 0 } . \overline{Q}\subset\{(x,t)\in\Omega_T:u(x,t)>0\}. Q ⊂ {( x , t ) ∈ Ω T : u ( x , t ) > 0 } .
Assume that u ≤ M u \le M u ≤ M in Q Q Q and set v = u 1 k v = u^{\frac{1}{k}} v = u k 1 . Then
∥ ∇ v ∥ L ∞ ( B r ( x o ) × ( s − h , s ) ) ≤ C M 1 k [ 1 r + 1 h 1 p ] , \|\nabla v\|_{L^\infty(B_r(x_o)\times(s-h,s))} \le CM^{\frac{1}{k}}\left[\frac{1}{r}+\frac{1}{h^{\frac{1}{p}}}\right], ∥∇ v ∥ L ∞ ( B r ( x o ) × ( s − h , s )) ≤ C M k 1 [ r 1 + h p 1 1 ] ,
where C > 0 C>0 C > 0 depends only on N , p , k N,p,k N , p , k . In particular, C C C is independent of any positive lower bound for u u u in Q Q Q .
Proof. Throughout the proof, C C C denotes a positive constant depending only on N , p , k N,p,k N , p , k , whose value may change from line to line. By assumption, u > 0 u>0 u > 0 on Q ‾ \overline{Q} Q , so v = u 1 k v=u^{\frac{1}{k}} v = u k 1 is smooth in a neighborhood of Q ‾ \overline{Q} Q . We do not require ∣ ∇ v ∣ > 0 |\nabla v|>0 ∣∇ v ∣ > 0 throughout Q Q Q : this condition holds automatically at any positive maximum of W = η 2 ∣ ∇ v ∣ 2 W=\eta^2|\nabla v|^2 W = η 2 ∣∇ v ∣ 2 considered below. Choose a smooth cutoff function 0 ≤ η ≤ 1 0\le\eta\le1 0 ≤ η ≤ 1 which vanishes on the parabolic boundary of Q Q Q , equals one on B r ( x o ) × ( s − h , s ] B_r(x_o)\times(s-h,s] B r ( x o ) × ( s − h , s ] , and satisfies
∣ ∇ η ∣ ≤ C r , ∣ ∇ 2 η ∣ ≤ C r 2 , ∣ ∂ t η ∣ ≤ C h . (25) |\nabla\eta|\le\frac{C}{r},\qquad|\nabla^2\eta|\le\frac{C}{r^2},\qquad|\partial_t\eta|\le\frac{C}{h}. \tag*{(25)} ∣∇ η ∣ ≤ r C , ∣ ∇ 2 η ∣ ≤ r 2 C , ∣ ∂ t η ∣ ≤ h C . (25)
In (25) , the constants in the gradient, Hessian, and time-derivative bounds can be chosen as 4 4 4 , 32 N 32\sqrt{N} 32 N , and 2 2 2 , respectively.
Recall that w = ∣ ∇ v ∣ 2 w=|\nabla v|^2 w = ∣∇ v ∣ 2 and consider
W : = η 2 w = η 2 ∣ ∇ v ∣ 2 on B 2 r ‾ × [ s − 2 h , s ] . W:=\eta^2w=\eta^2|\nabla v|^2\qquad\text{on }\overline{B_{2r}}\times[s-2h,s]. W := η 2 w = η 2 ∣∇ v ∣ 2 on B 2 r × [ s − 2 h , s ] .
Figure 2. Schematic spatial cross-section of Q = B 2 r ( x o ) × ( s − 2 h , s ] Q=B_{2r}(x_o)\times(s-2h,s] Q = B 2 r ( x o ) × ( s − 2 h , s ] . The blue rectangle represents B 2 r ( x o ) × [ s − 2 h , s ] B_{2r}(x_o)\times[s-2h,s] B 2 r ( x o ) × [ s − 2 h , s ] . The green region represents B r ( x o ) × [ s − h , s ] B_r(x_o)\times[s-h,s] B r ( x o ) × [ s − h , s ] , where η = 1 \eta=1 η = 1 .
If the maximum of W W W is zero, then W ≡ 0 W\equiv0 W ≡ 0 . Since η = 1 \eta=1 η = 1 on B r ( x o ) × ( s − h , s ] B_r(x_o)\times(s-h,s] B r ( x o ) × ( s − h , s ] , this implies ∇ v = 0 \nabla v=0 ∇ v = 0 there, and the gradient estimate holds trivially. Otherwise, let ( x ∗ , t ∗ ) (x_*,t_*) ( x ∗ , t ∗ ) be a point where W W W attains a positive maximum. Since η \eta η vanishes on the parabolic boundary of Q Q Q , we have
x ∗ ∈ B 2 r ( x o ) , t ∗ ∈ ( s − 2 h , s ] , η ( x ∗ , t ∗ ) > 0 , w ( x ∗ , t ∗ ) > 0. x_*\in B_{2r}(x_o),\qquad t_*\in(s-2h,s],\qquad\eta(x_*,t_*)>0,\qquad w(x_*,t_*)>0. x ∗ ∈ B 2 r ( x o ) , t ∗ ∈ ( s − 2 h , s ] , η ( x ∗ , t ∗ ) > 0 , w ( x ∗ , t ∗ ) > 0.
The spatial maximum is attained at an interior point, so ∇ W = 0 \nabla W=0 ∇ W = 0 and ∇ 2 W ≤ 0 \nabla^2W\le0 ∇ 2 W ≤ 0 at ( x ∗ , t ∗ ) (x_*,t_*) ( x ∗ , t ∗ ) . If t ∗ < s t_*<s t ∗ < s , then ∂ t W = 0 \partial_tW=0 ∂ t W = 0 . If t ∗ = s t_*=s t ∗ = s , the maximality of W W W gives
∂ t W ( x ∗ , s ) = lim τ → 0 W ( x ∗ , s ) − W ( x ∗ , s − τ ) τ ≥ 0. \partial_tW(x_*,s)=\lim_{\tau\to0}\frac{W(x_*,s)-W(x_*,s-\tau)}{\tau}\ge0. ∂ t W ( x ∗ , s ) = τ → 0 lim τ W ( x ∗ , s ) − W ( x ∗ , s − τ ) ≥ 0.
Thus, ∂ t W ≥ 0 \partial_t W \ge0 ∂ t W ≥ 0 in either case. Since P \mathbf{P} P is positive definite and ∇ 2 W \nabla^2 W ∇ 2 W is negative semidefinite at ( x ∗ , t ∗ ) (x_*,t_*) ( x ∗ , t ∗ ) , we have
P ⋅ ∇ 2 W = tr ( P 1 / 2 ∇ 2 W P 1 / 2 ) ≤ 0. \mathbf{P}\mathbin{\cdot}\nabla^2 W=\operatorname{tr}\left(\mathbf{P}^{1/2}\nabla^2 W\mathbf{P}^{1/2}\right)\le0. P ⋅ ∇ 2 W = tr ( P 1/2 ∇ 2 W P 1/2 ) ≤ 0.
Since w ( x ∗ , t ∗ ) > 0 w(x_*,t_*)>0 w ( x ∗ , t ∗ ) > 0 and ∇ v \nabla v ∇ v is continuous, we have ∣ ∇ v ∣ > 0 \lvert\nabla v\rvert>0 ∣ ∇ v ∣ > 0 in a neighborhood of ( x ∗ , t ∗ ) (x_*,t_*) ( x ∗ , t ∗ ) . The identities and inequalities derived above therefore apply in this neighborhood. Together with ∇ W = 0 \nabla W=0 ∇ W = 0 , ∂ t W ≥ 0 \partial_t W\ge0 ∂ t W ≥ 0 , and a ( v ) w p − 2 2 > 0 a(v)w^{\frac{p-2}{2}}>0 a ( v ) w 2 p − 2 > 0 , the inequality P ⋅ ∇ 2 W ≤ 0 \mathbf{P}\mathbin{\cdot}\nabla^2W\le0 P ⋅ ∇ 2 W ≤ 0 yields
L W = ∂ t W − a ( v ) w p − 2 2 P ⋅ ∇ 2 W − B ⋅ ∇ W ≥ 0 at ( x ∗ , t ∗ ) . (26) \mathcal{L}W=\partial_tW-a(v)w^{\frac{p-2}{2}}\mathbf{P}\mathbin{\cdot}\nabla^2W-\mathbf{B}\mathbin{\cdot}\nabla W\ge0\quad\text{at }(x_*,t_*). \tag*{(26)} L W = ∂ t W − a ( v ) w 2 p − 2 P ⋅ ∇ 2 W − B ⋅ ∇ W ≥ 0 at ( x ∗ , t ∗ ) . (26)
Moreover, at ( x ∗ , t ∗ ) (x_*,t_*) ( x ∗ , t ∗ ) we have
0 = ∇ W = ∇ ( η 2 w ) = 2 η w ∇ η + η 2 ∇ w . 0=\nabla W=\nabla(\eta^2w)=2\eta w\nabla\eta+\eta^2\nabla w. 0 = ∇ W = ∇ ( η 2 w ) = 2 η w ∇ η + η 2 ∇ w .
Since η ( x ∗ , t ∗ ) > 0 \eta(x_*,t_*)>0 η ( x ∗ , t ∗ ) > 0 , we obtain
∇ w = − 2 w ∇ η η at ( x ∗ , t ∗ ) . (27) \nabla w=-2w\frac{\nabla\eta}{\eta}\quad\text{at }(x_*,t_*). \tag*{(27)} ∇ w = − 2 w η ∇ η at ( x ∗ , t ∗ ) . (27)
Notice that where ∇ v = 0 \nabla v=0 ∇ v = 0 the principal coefficient a w p − 2 2 P aw^{\frac{p-2}{2}}\mathbf{P} a w 2 p − 2 P vanishes (p > 2 p>2 p > 2 ) or blows up (p < 2 p<2 p < 2 ), and the differentiated identity is not available; the positivity of W ( x ∗ , t ∗ ) W(x_*,t_*) W ( x ∗ , t ∗ ) is what keeps the argument away from this set.
To expand L ( η 2 w ) \mathcal{L}(\eta^2w) L ( η 2 w ) , we first note that, for smooth scalar functions f f f and g g g ,
∇ 2 ( f g ) = f ∇ 2 g + g ∇ 2 f + ∇ f ⊗ ∇ g + ∇ g ⊗ ∇ f . \nabla^2(fg)=f\nabla^2g+g\nabla^2f+\nabla f\otimes\nabla g+\nabla g\otimes\nabla f. ∇ 2 ( f g ) = f ∇ 2 g + g ∇ 2 f + ∇ f ⊗ ∇ g + ∇ g ⊗ ∇ f .
Since P \mathbf{P} P is symmetric, we obtain
P ⋅ ∇ 2 ( f g ) = f P ⋅ ∇ 2 g + g P ⋅ ∇ 2 f + 2 ( P ∇ f ) ⋅ ∇ g . \mathbf{P}\mathbin{\cdot}\nabla^2(fg)=f\mathbf{P}\mathbin{\cdot}\nabla^2g+g\mathbf{P}\mathbin{\cdot}\nabla^2f+2(\mathbf{P}\nabla f)\mathbin{\cdot}\nabla g. P ⋅ ∇ 2 ( f g ) = f P ⋅ ∇ 2 g + g P ⋅ ∇ 2 f + 2 ( P ∇ f ) ⋅ ∇ g .
The time derivative and the drift term satisfy the usual product rule. Consequently,
L ( f g ) = f L g + g L f − 2 a ( v ) w p − 2 2 ( P ∇ f ) ⋅ ∇ g . \mathcal{L}(fg)=f\mathcal{L}g+g\mathcal{L}f-2a(v)w^{\frac{p-2}{2}}(\mathbf{P}\nabla f)\mathbin{\cdot}\nabla g. L ( f g ) = f L g + g L f − 2 a ( v ) w 2 p − 2 ( P ∇ f ) ⋅ ∇ g .
Applying this identity with f = η 2 f=\eta^2 f = η 2 and g = w g=w g = w , we obtain
L ( η 2 w ) = η 2 L w + w L ( η 2 ) − 2 a ( v ) w p − 2 2 ( P ∇ ( η 2 ) ) ⋅ ∇ w . \mathcal{L}(\eta^2w)=\eta^2\mathcal{L}w+w\mathcal{L}(\eta^2)-2a(v)w^{\frac{p-2}{2}}(\mathbf{P}\nabla(\eta^2))\mathbin{\cdot}\nabla w. L ( η 2 w ) = η 2 L w + w L ( η 2 ) − 2 a ( v ) w 2 p − 2 ( P ∇ ( η 2 )) ⋅ ∇ w .
Applying the same product rule to η 2 = η η \eta^2=\eta\eta η 2 = η η yields
L ( η 2 ) = 2 η L η − 2 a ( v ) w p − 2 2 ( P ∇ η ) ⋅ ∇ η . \mathcal{L}(\eta^2)=2\eta\mathcal{L}\eta-2a(v)w^{\frac{p-2}{2}}(\mathbf{P}\nabla\eta)\mathbin{\cdot}\nabla\eta. L ( η 2 ) = 2 η L η − 2 a ( v ) w 2 p − 2 ( P ∇ η ) ⋅ ∇ η .
Substituting this expression and using ∇ ( η 2 ) = 2 η ∇ η \nabla(\eta^2)=2\eta\nabla\eta ∇ ( η 2 ) = 2 η ∇ η , we find
L ( η 2 w ) = η 2 L w + 2 η w L η − 2 a ( v ) w p − 2 2 ( P ∇ η ) ⋅ ∇ η − 4 a ( v ) η w p 2 ( P ∇ η ) ⋅ ∇ w . \mathcal{L}(\eta^2w)=\eta^2\mathcal{L}w+2\eta w\mathcal{L}\eta-2a(v)w^{\frac{p-2}{2}}(\mathbf{P}\nabla\eta)\mathbin{\cdot}\nabla\eta-4a(v)\eta w^{\frac{p}{2}}(\mathbf{P}\nabla\eta)\mathbin{\cdot}\nabla w. L ( η 2 w ) = η 2 L w + 2 η w L η − 2 a ( v ) w 2 p − 2 ( P ∇ η ) ⋅ ∇ η − 4 a ( v ) η w 2 p ( P ∇ η ) ⋅ ∇ w .
We now evaluate this identity at ( x ∗ , t ∗ ) (x_*,t_*) ( x ∗ , t ∗ ) and omit the evaluation point for readability. By (27) , we have
− 4 a ( v ) η w p 2 ( P ∇ η ) ⋅ ∇ w = 8 a ( v ) w p 2 ( P ∇ η ) ⋅ ∇ η . -4a(v)\eta w^{\frac{p}{2}}(\mathbf{P}\nabla\eta)\mathbin{\cdot}\nabla w=8a(v)w^{\frac{p}{2}}(\mathbf{P}\nabla\eta)\mathbin{\cdot}\nabla\eta. − 4 a ( v ) η w 2 p ( P ∇ η ) ⋅ ∇ w = 8 a ( v ) w 2 p ( P ∇ η ) ⋅ ∇ η .
Combining this with the preceding identity and using (26) , we obtain
0 ≤ L W = η 2 L w + 2 η w ( ∂ t η − a ( v ) w p − 2 2 P ⋅ ∇ 2 η − B ⋅ ∇ η ) + 6 a ( v ) w p 2 ( P ∇ η ) ⋅ ∇ η . (28) \begin{aligned}
0\le\mathcal{L}W
&=\eta^2\mathcal{L}w+2\eta w\left(\partial_t\eta-a(v)w^{\frac{p-2}{2}}\mathbf{P}\mathbin{\cdot}\nabla^2\eta-\mathbf{B}\mathbin{\cdot}\nabla\eta\right)\\
&\qquad+6a(v)w^{\frac{p}{2}}(\mathbf{P}\nabla\eta)\mathbin{\cdot}\nabla\eta. \tag*{(28)}
\end{aligned} 0 ≤ L W = η 2 L w + 2 η w ( ∂ t η − a ( v ) w 2 p − 2 P ⋅ ∇ 2 η − B ⋅ ∇ η ) + 6 a ( v ) w 2 p ( P ∇ η ) ⋅ ∇ η . (28)
We apply (24) to the term η 2 L w \eta^2\mathcal{L}w η 2 L w :
η 2 L w ≤ − a ( v ) η 2 w p − 2 2 tr ( P ( ∇ 2 v ) 2 ) − κ η 2 v − p w p + 2 2 . \eta^2\mathcal{L}w\le-a(v)\eta^2w^{\frac{p-2}{2}}\operatorname{tr}\left(\mathbf{P}(\nabla^2v)^2\right)-\kappa\eta^2v^{-p}w^{\frac{p+2}{2}}. η 2 L w ≤ − a ( v ) η 2 w 2 p − 2 tr ( P ( ∇ 2 v ) 2 ) − κ η 2 v − p w 2 p + 2 .
We estimate the remaining terms separately. Since η \eta η and w w w are nonnegative, the term containing the time derivative satisfies
2 η w ∂ t η ≤ 2 η w ∣ ∂ t η ∣ . 2\eta w\partial_t\eta\le2\eta w\lvert\partial_t\eta\rvert. 2 η w ∂ t η ≤ 2 η w ∣ ∂ t η ∣ .
For the term containing ∇ 2 η \nabla^2\eta ∇ 2 η , we use the Cauchy–Schwarz inequality for the Frobenius inner product: ∣ P ⋅ ∇ 2 η ∣ ≤ ∣ P ∣ ∣ ∇ 2 η ∣ |\mathbf{P}\cdot\nabla^2\eta|\le|\mathbf{P}||\nabla^2\eta| ∣ P ⋅ ∇ 2 η ∣ ≤ ∣ P ∣∣ ∇ 2 η ∣ . Since P \mathbf{P} P is symmetric, its squared Frobenius norm equals the sum of the squares of its eigenvalues. Thus,
∣ P ∣ 2 = tr ( P T P ) = tr ( P 2 ) = ( N − 1 ) + ( p − 1 ) 2 , |\mathbf{P}|^2=\operatorname{tr}(\mathbf{P}^{T}\mathbf{P})=\operatorname{tr}(\mathbf{P}^2)=(N-1)+(p-1)^2, ∣ P ∣ 2 = tr ( P T P ) = tr ( P 2 ) = ( N − 1 ) + ( p − 1 ) 2 ,
and therefore
− 2 a ( v ) w p 2 P ⋅ ∇ 2 η ≤ 2 N − 1 + ( p − 1 ) 2 a ( v ) w p 2 ∣ ∇ 2 η ∣ . -2a(v)w^{\frac{p}{2}}\mathbf{P}\cdot\nabla^2\eta\le2\sqrt{N-1+(p-1)^2}\,a(v)w^{\frac{p}{2}}|\nabla^2\eta|. − 2 a ( v ) w 2 p P ⋅ ∇ 2 η ≤ 2 N − 1 + ( p − 1 ) 2 a ( v ) w 2 p ∣ ∇ 2 η ∣.
The constant in this estimate depends only on N N N and p p p . For the drift term, the usual Cauchy–Schwarz inequality gives
− 2 η w B ⋅ ∇ η ≤ 2 η w ∣ B ∣ ∣ ∇ η ∣ . -2\eta w\mathbf{B}\cdot\nabla\eta\le2\eta w|\mathbf{B}||\nabla\eta|. − 2 η w B ⋅ ∇ η ≤ 2 η w ∣ B ∣∣∇ η ∣.
We leave ∣ B ∣ |\mathbf{B}| ∣ B ∣ unchanged at this stage and estimate it later using (18) . Finally, the upper bound in (21) yields
( P ∇ η ) ⋅ ∇ η ≤ max { 1 , p − 1 } ∣ ∇ η ∣ 2 . (\mathbf{P}\nabla\eta)\cdot\nabla\eta\le\max\{1,p-1\}|\nabla\eta|^2. ( P ∇ η ) ⋅ ∇ η ≤ max { 1 , p − 1 } ∣∇ η ∣ 2 .
Consequently,
6 a ( v ) w p 2 ( P ∇ η ) ⋅ ∇ η ≤ 6 max { 1 , p − 1 } a ( v ) w p 2 ∣ ∇ η ∣ 2 . 6a(v)w^{\frac{p}{2}}(\mathbf{P}\nabla\eta)\cdot\nabla\eta\le6\max\{1,p-1\}a(v)w^{\frac{p}{2}}|\nabla\eta|^2. 6 a ( v ) w 2 p ( P ∇ η ) ⋅ ∇ η ≤ 6 max { 1 , p − 1 } a ( v ) w 2 p ∣∇ η ∣ 2 .
Here, the constant depends only on p p p . Substituting these estimates into the preceding inequality (28) and moving the two negative terms to the left-hand side, we obtain
a ( v ) η 2 w p − 2 2 tr ( P ( ∇ 2 v ) 2 ) + κ η 2 v − p w p + 2 2 ≤ C η ∣ ∂ t η ∣ w + C a ( v ) w p 2 ( ∣ ∇ η ∣ 2 + η ∣ ∇ 2 η ∣ ) + 2 η w ∣ B ∣ ∣ ∇ η ∣ . (29) \begin{aligned}
a(v)\eta^2w^{\frac{p-2}{2}}\operatorname{tr}\left(\mathbf{P}(\nabla^2v)^2\right)+\kappa\eta^2v^{-p}w^{\frac{p+2}{2}}
&\le C\eta|\partial_t\eta|w+Ca(v)w^{\frac{p}{2}}\left(|\nabla\eta|^2+\eta|\nabla^2\eta|\right)\\
&\quad+2\eta w|\mathbf{B}||\nabla\eta|. \tag*{(29)}
\end{aligned} a ( v ) η 2 w 2 p − 2 tr ( P ( ∇ 2 v ) 2 ) + κ η 2 v − p w 2 p + 2 ≤ C η ∣ ∂ t η ∣ w + C a ( v ) w 2 p ( ∣∇ η ∣ 2 + η ∣ ∇ 2 η ∣ ) + 2 η w ∣ B ∣∣∇ η ∣. (29)
Here C C C depends only on N N N and p p p . By the drift estimate (18) , there exists a constant C = C ( N , p ) > 0 C=C(N,p)>0 C = C ( N , p ) > 0 such that
2 η w ∣ B ∣ ∣ ∇ η ∣ ≤ C a ( v ) η w p − 1 2 ∣ ∇ 2 v ∣ ∣ ∇ η ∣ + C b ( v ) η w p + 1 2 ∣ ∇ η ∣ . 2\eta w|\mathbf{B}||\nabla\eta|\le Ca(v)\eta w^{\frac{p-1}{2}}|\nabla^2v||\nabla\eta|+Cb(v)\eta w^{\frac{p+1}{2}}|\nabla\eta|. 2 η w ∣ B ∣∣∇ η ∣ ≤ C a ( v ) η w 2 p − 1 ∣ ∇ 2 v ∣∣∇ η ∣ + C b ( v ) η w 2 p + 1 ∣∇ η ∣.
After substituting the drift estimate into (29) , we absorb the contribution involving ∣ ∇ 2 v ∣ |\nabla^2v| ∣ ∇ 2 v ∣ into the Hessian term on the left-hand side. Indeed, Young’s inequality yields
C a ( v ) η w p − 1 2 ∣ ∇ 2 v ∣ ∣ ∇ η ∣ = ( a ( v ) η w p − 2 4 ∣ ∇ 2 v ∣ ) ( C a ( v ) w p 4 ∣ ∇ η ∣ ) ≤ ε a ( v ) η 2 w p − 2 2 ∣ ∇ 2 v ∣ 2 + C 2 4 ε a ( v ) w p 2 ∣ ∇ η ∣ 2 ≤ ε a ( v ) η 2 w p − 2 2 min { 1 , p − 1 } tr ( P ( ∇ 2 v ) 2 ) + C 2 4 ε a ( v ) w p 2 ∣ ∇ η ∣ 2 , \begin{aligned}
Ca(v)\eta w^{\frac{p-1}{2}}|\nabla^2v||\nabla\eta|
&=\left(\sqrt{a(v)}\eta w^{\frac{p-2}{4}}|\nabla^2v|\right)\left(C\sqrt{a(v)}w^{\frac{p}{4}}|\nabla\eta|\right)\\
&\le\varepsilon a(v)\eta^2w^{\frac{p-2}{2}}|\nabla^2v|^2+\frac{C^2}{4\varepsilon}a(v)w^{\frac{p}{2}}|\nabla\eta|^2\\
&\le\frac{\varepsilon a(v)\eta^2w^{\frac{p-2}{2}}}{\min\{1,p-1\}}\operatorname{tr}\left(\mathbf{P}(\nabla^2v)^2\right)+\frac{C^2}{4\varepsilon}a(v)w^{\frac{p}{2}}|\nabla\eta|^2,
\end{aligned} C a ( v ) η w 2 p − 1 ∣ ∇ 2 v ∣∣∇ η ∣ = ( a ( v ) η w 4 p − 2 ∣ ∇ 2 v ∣ ) ( C a ( v ) w 4 p ∣∇ η ∣ ) ≤ ε a ( v ) η 2 w 2 p − 2 ∣ ∇ 2 v ∣ 2 + 4 ε C 2 a ( v ) w 2 p ∣∇ η ∣ 2 ≤ min { 1 , p − 1 } ε a ( v ) η 2 w 2 p − 2 tr ( P ( ∇ 2 v ) 2 ) + 4 ε C 2 a ( v ) w 2 p ∣∇ η ∣ 2 ,
where we used
tr ( P ( ∇ 2 v ) 2 ) ≥ min { 1 , p − 1 } ∣ ∇ 2 v ∣ 2 . \operatorname{tr}\left(\mathbf{P}(\nabla^2v)^2\right)\ge\min\{1,p-1\}|\nabla^2v|^2. tr ( P ( ∇ 2 v ) 2 ) ≥ min { 1 , p − 1 } ∣ ∇ 2 v ∣ 2 .
Choosing
ε = 1 2 min { 1 , p − 1 } , \varepsilon=\frac{1}{2}\min\{1,p-1\}, ε = 2 1 min { 1 , p − 1 } ,
we obtain
C a ( v ) η w p − 1 2 ∣ ∇ 2 v ∣ ∣ ∇ η ∣ ≤ 1 2 a ( v ) η 2 w p − 2 2 tr ( P ( ∇ 2 v ) 2 ) + C 2 2 min { 1 , p − 1 } a ( v ) w p 2 ∣ ∇ η ∣ 2 . \begin{aligned}
Ca(v)\eta w^{\frac{p-1}{2}}|\nabla^2v||\nabla\eta|
&\le\frac{1}{2}a(v)\eta^2w^{\frac{p-2}{2}}\operatorname{tr}\left(\mathbf{P}(\nabla^2v)^2\right)\\
&\quad+\frac{C^2}{2\min\{1,p-1\}}a(v)w^{\frac{p}{2}}|\nabla\eta|^2.
\end{aligned} C a ( v ) η w 2 p − 1 ∣ ∇ 2 v ∣∣∇ η ∣ ≤ 2 1 a ( v ) η 2 w 2 p − 2 tr ( P ( ∇ 2 v ) 2 ) + 2 min { 1 , p − 1 } C 2 a ( v ) w 2 p ∣∇ η ∣ 2 .
The first term is absorbed into the left-hand side of (29) . The constant multiplying the second term on the right-hand side depends only on N N N and p p p . Substituting the estimates for the terms on the right-hand side of (29) , absorbing the Hessian contribution into the left-hand side, and then dropping the remaining nonnegative Hessian term, we obtain
κ η 2 v − p w p + 2 2 ≤ C [ η ∣ ∂ t η ∣ w + a ( v ) w p 2 ( ∣ ∇ η ∣ 2 + η ∣ ∇ 2 η ∣ ) + b ( v ) η w p + 1 2 ∣ ∇ η ∣ ] . (30) \kappa\eta^2v^{-p}w^{\frac{p+2}{2}}\le C\left[\eta|\partial_t\eta|w+a(v)w^{\frac{p}{2}}\left(|\nabla\eta|^2+\eta|\nabla^2\eta|\right)+b(v)\eta w^{\frac{p+1}{2}}|\nabla\eta|\right]. \tag*{(30)} κ η 2 v − p w 2 p + 2 ≤ C [ η ∣ ∂ t η ∣ w + a ( v ) w 2 p ( ∣∇ η ∣ 2 + η ∣ ∇ 2 η ∣ ) + b ( v ) η w 2 p + 1 ∣∇ η ∣ ] . (30)
Here C C C depends only on N N N and p p p . To estimate the maximum of W W W , we set
Y : = η w = W . \mathbf{Y}:=\eta\sqrt{w}=\sqrt{W}. Y := η w = W .
Multiplying (30) by v p η p v^{p}\eta^{p} v p η p and recalling Y = η w \mathbf{Y}=\eta\sqrt{w} Y = η w , we obtain
κ η p + 2 w p + 2 2 ≤ C [ v p ∣ ∂ t η ∣ η p + 1 w + a ( v ) v p η p 2 w p 2 ( ∣ ∇ η ∣ 2 + η ∣ ∇ 2 η ∣ ) + b ( v ) v p η p + 1 w p + 1 2 ∣ ∇ η ∣ ] . \kappa\eta^{p+2}w^{\frac{p+2}{2}} \leq C\left[v^{p}|\partial_t\eta|\eta^{p+1}w+a(v)v^{p}\eta^{\frac{p}{2}}w^{\frac{p}{2}}\left(|\nabla\eta|^{2}+\eta|\nabla^{2}\eta|\right)+b(v)v^{p}\eta^{p+1}w^{\frac{p+1}{2}}|\nabla\eta|\right]. κ η p + 2 w 2 p + 2 ≤ C [ v p ∣ ∂ t η ∣ η p + 1 w + a ( v ) v p η 2 p w 2 p ( ∣∇ η ∣ 2 + η ∣ ∇ 2 η ∣ ) + b ( v ) v p η p + 1 w 2 p + 1 ∣∇ η ∣ ] .
Thus,
κ Y p + 2 ≤ C [ v p η p − 1 ∣ ∂ t η ∣ Y 2 + a ( v ) v p Y p ( ∣ ∇ η ∣ 2 + η ∣ ∇ 2 η ∣ ) + b ( v ) v p Y p + 1 ∣ ∇ η ∣ ] . \kappa\mathbf{Y}^{p+2}\leq C\left[v^{p}\eta^{p-1}|\partial_t\eta|\mathbf{Y}^{2}+a(v)v^{p}\mathbf{Y}^{p}\left(|\nabla\eta|^{2}+\eta|\nabla^{2}\eta|\right)+b(v)v^{p}\mathbf{Y}^{p+1}|\nabla\eta|\right]. κ Y p + 2 ≤ C [ v p η p − 1 ∣ ∂ t η ∣ Y 2 + a ( v ) v p Y p ( ∣∇ η ∣ 2 + η ∣ ∇ 2 η ∣ ) + b ( v ) v p Y p + 1 ∣∇ η ∣ ] .
Since p > 1 p>1 p > 1 and 0 ≤ η ≤ 1 0\leq\eta\leq1 0 ≤ η ≤ 1 , we have η p − 1 ≤ 1 \eta^{p-1}\leq1 η p − 1 ≤ 1 . Recalling that c = k p − 2 / ( p − 1 ) c=k^{p-2}/(p-1) c = k p − 2 / ( p − 1 ) , we substitute a ( v ) v p = c v 2 a(v)v^{p}=cv^{2} a ( v ) v p = c v 2 and b ( v ) v p = c ( p − 1 ) ( k − 1 ) v b(v)v^{p}=c(p-1)(k-1)v b ( v ) v p = c ( p − 1 ) ( k − 1 ) v . Using v ≤ M 1 k v\leq M^{\frac{1}{k}} v ≤ M k 1 and the derivative bounds in (25) , and dividing by κ \kappa κ , we obtain
Y p + 2 ≤ C [ M 2 k h Y 2 + M 2 k r 2 Y p + M 1 k r Y p + 1 ] . (31) \mathbf{Y}^{p+2}\leq C\left[\frac{M^{\frac{2}{k}}}{h}\mathbf{Y}^{2}+\frac{M^{\frac{2}{k}}}{r^{2}}\mathbf{Y}^{p}+\frac{M^{\frac{1}{k}}}{r}\mathbf{Y}^{p+1}\right]. \tag*{(31)} Y p + 2 ≤ C [ h M k 2 Y 2 + r 2 M k 2 Y p + r M k 1 Y p + 1 ] . (31)
where C = C ( N , p , k ) > 0 C=C(N,p,k)>0 C = C ( N , p , k ) > 0 is a new constant. Young’s inequality gives, for every ε > 0 \varepsilon>0 ε > 0 ,
M p k h Y 2 ≤ ε Y p + 2 + C ( p ) ε − p p M p + 2 k h p + 2 p , \frac{M^{\frac{p}{k}}}{h}\mathbf{Y}^{2}\leq\varepsilon\mathbf{Y}^{p+2}+C(p)\varepsilon^{-\frac{p}{p}} \frac{M^{\frac{p+2}{k}}}{h^{\frac{p+2}{p}}}, h M k p Y 2 ≤ ε Y p + 2 + C ( p ) ε − p p h p p + 2 M k p + 2 ,
M 2 k r 2 Y p ≤ ε Y p + 2 + C ( p ) ε − p 2 M p + 2 k r p + 2 , \frac{M^{\frac{2}{k}}}{r^{2}}\mathbf{Y}^{p}\leq\varepsilon\mathbf{Y}^{p+2}+C(p)\varepsilon^{-\frac{p}{2}}\frac{M^{\frac{p+2}{k}}}{r^{p+2}}, r 2 M k 2 Y p ≤ ε Y p + 2 + C ( p ) ε − 2 p r p + 2 M k p + 2 ,
and
M 1 k r Y p + 1 ≤ ε Y p + 2 + C ( p ) ε − ( p + 1 ) M p + 2 k r p + 2 . \frac{M^{\frac{1}{k}}}{r}\mathbf{Y}^{p+1}\leq\varepsilon\mathbf{Y}^{p+2}+C(p)\varepsilon^{-(p+1)}\frac{M^{\frac{p+2}{k}}}{r^{p+2}}. r M k 1 Y p + 1 ≤ ε Y p + 2 + C ( p ) ε − ( p + 1 ) r p + 2 M k p + 2 .
Choosing ε = 1 6 C \varepsilon=\frac{1}{6C} ε = 6 C 1 , where C C C is the constant in (31) , we can absorb the three contributions C ε Y p + 2 C\varepsilon\mathbf{Y}^{p+2} C ε Y p + 2 into the left-hand side. We obtain
Y p + 2 ≤ C M p + 2 k [ 1 h p + 2 p + 1 r p + 2 ] , \mathbf{Y}^{p+2}\leq CM^{\frac{p+2}{k}}\left[\frac{1}{h^{\frac{p+2}{p}}}+\frac{1}{r^{p+2}}\right], Y p + 2 ≤ C M k p + 2 [ h p p + 2 1 + r p + 2 1 ] ,
where C = C ( N , p , k ) > 0 C=C(N,p,k)>0 C = C ( N , p , k ) > 0 . Taking the ( p + 2 ) (p+2) ( p + 2 ) -th root, we conclude that
Y ( x ∗ , t ∗ ) ≤ C M 1 k [ 1 r + 1 h 1 p ] . \mathbf{Y}(x_{*},t_{*})\leq CM^{\frac{1}{k}}\left[\frac{1}{r}+\frac{1}{h^{\frac{1}{p}}}\right]. Y ( x ∗ , t ∗ ) ≤ C M k 1 [ r 1 + h p 1 1 ] .
Since ( x ∗ , t ∗ ) (x_{*},t_{*}) ( x ∗ , t ∗ ) is a maximum point of W = Y 2 W=\mathbf{Y}^{2} W = Y 2 on B 2 r ( x o ) ‾ × [ s − 2 h , s ] \overline{B_{2r}(x_{o})}\times[s-2h,s] B 2 r ( x o ) × [ s − 2 h , s ] and Y ≥ 0 \mathbf{Y}\geq0 Y ≥ 0 , we have
Y ( x , t ) ≤ Y ( x ∗ , t ∗ ) for all ( x , t ) ∈ B 2 r ( x o ) ‾ × [ s − 2 h , s ] . \mathbf{Y}(x,t)\leq\mathbf{Y}(x_{*},t_{*})\qquad\text{for all }(x,t)\in\overline{B_{2r}(x_{o})}\times[s-2h,s]. Y ( x , t ) ≤ Y ( x ∗ , t ∗ ) for all ( x , t ) ∈ B 2 r ( x o ) × [ s − 2 h , s ] .
In particular, this inequality holds on B r ( x o ) × ( s − h , s ) B_{r}(x_{o})\times(s-h,s) B r ( x o ) × ( s − h , s ) , where η = 1 \eta=1 η = 1 . Thus, for every ( x , t ) (x,t) ( x , t ) in this smaller cylinder,
∣ ∇ v ( x , t ) ∣ = η ( x , t ) w ( x , t ) = Y ( x , t ) ≤ Y ( x ∗ , t ∗ ) ≤ C M 1 k [ 1 r + 1 h 1 p ] . |\nabla v(x,t)|=\eta(x,t)\sqrt{w(x,t)}=\mathbf{Y}(x,t)\leq\mathbf{Y}(x_{*},t_{*})\leq CM^{\frac{1}{k}}\left[\frac{1}{r}+\frac{1}{h^{\frac{1}{p}}}\right]. ∣∇ v ( x , t ) ∣ = η ( x , t ) w ( x , t ) = Y ( x , t ) ≤ Y ( x ∗ , t ∗ ) ≤ C M k 1 [ r 1 + h p 1 1 ] .
Taking the supremum over the smaller cylinder yields
∥ ∇ v ∥ L ∞ ( B r ( x o ) × ( s − h , s ) ) ≤ C M 1 k [ 1 r + 1 h 1 p ] , \|\nabla v\|_{L^{\infty}(B_{r}(x_{o})\times(s-h,s))}\leq CM^{\frac{1}{k}}\left[\frac{1}{r}+\frac{1}{h^{\frac{1}{p}}}\right], ∥∇ v ∥ L ∞ ( B r ( x o ) × ( s − h , s )) ≤ C M k 1 [ r 1 + h p 1 1 ] ,
where the constant C = C ( N , p , k ) C=C(N,p,k) C = C ( N , p , k ) is independent of s s s . No positive lower bound for v v v enters the constant in this estimate.
Remark 3.4 (The idea of the gradient estimate). The argument begins with the transformation u = v k u=v^k u = v k , which gives
v t = a ( v ) Δ p v + b ( v ) ∣ ∇ v ∣ p . v_t=a(v)\Delta_p v+b(v)|\nabla v|^p. v t = a ( v ) Δ p v + b ( v ) ∣∇ v ∣ p .
The exponent k k k is chosen so that, after differentiating the equation and applying Young’s inequality, the coefficient of the highest power of ∣ ∇ v ∣ |\nabla v| ∣∇ v ∣ is strictly negative. More precisely, for w = ∣ ∇ v ∣ 2 w=|\nabla v|^2 w = ∣∇ v ∣ 2 we obtain
L w ≤ − a ( v ) w p − 2 2 tr ( ( ∇ 2 v ) 2 ) − κ v − p w p + 2 2 . \mathcal{L}w\le-a(v)w^{\frac{p-2}{2}}\operatorname{tr}\left((\nabla^2v)^2\right)-\kappa v^{-p}w^{\frac{p+2}{2}}. L w ≤ − a ( v ) w 2 p − 2 tr ( ( ∇ 2 v ) 2 ) − κ v − p w 2 p + 2 .
The two negative terms serve different purposes. The first controls the Hessian, whereas the second provides the term needed to bound the gradient. We then localize the argument by introducing W = η 2 w W=\eta^2w W = η 2 w . At a positive maximum of W W W , we have
L W ≥ 0 , ∇ w = − 2 w ∇ η η . \mathcal{L}W\ge0,\qquad\nabla w=-2w\frac{\nabla\eta}{\eta}. L W ≥ 0 , ∇ w = − 2 w η ∇ η .
These relations allow us to control the additional terms arising from the cutoff. In particular, the negative Hessian term absorbs the Hessian-dependent contribution of the drift. After this absorption, the bound v ≤ M 1 k v\le M^{\frac{1}{k}} v ≤ M k 1 and the cutoff estimates reduce the differential inequality to the algebraic inequality
Y p + 2 ≤ C [ M p k h Y 2 + M p k r 2 Y p + M 1 k r Y p + 1 ] . \mathbf{Y}^{p+2}\le C\left[\frac{M^{\frac{p}{k}}}{h}\mathbf{Y}^2+\frac{M^{\frac{p}{k}}}{r^2}\mathbf{Y}^p+\frac{M^{\frac{1}{k}}}{r}\mathbf{Y}^{p+1}\right]. Y p + 2 ≤ C [ h M k p Y 2 + r 2 M k p Y p + r M k 1 Y p + 1 ] .
at the maximum point, where Y = η w \mathbf{Y}=\eta\sqrt{w} Y = η w . All powers of Y \mathbf{Y} Y on the right-hand side are strictly smaller than p + 2 p+2 p + 2 . Therefore, Young’s inequality gives a bound for Y \mathbf{Y} Y , and the fact that η = 1 \eta=1 η = 1 on the smaller cylinder yields
∥ ∇ v ∥ L ∞ ( B r ( x o ) × ( s − h , s ) ) ≤ C M 1 k [ 1 r + 1 h 1 p ] . \|\nabla v\|_{L^\infty(B_r(x_o)\times(s-h,s))}\le CM^{\frac{1}{k}}\left[\frac{1}{r}+\frac{1}{h^{\frac{1}{p}}}\right]. ∥∇ v ∥ L ∞ ( B r ( x o ) × ( s − h , s )) ≤ C M k 1 [ r 1 + h p 1 1 ] .
This is a localized Bernstein argument (see [3 , 4 ] ), in the sense that a differential inequality for the squared gradient is combined with the maximum principle to obtain a gradient bound. The homogeneity of the equation accounts for the linear dependence on M 1 k M^{\frac{1}{k}} M k 1 . The constant C C C depends only on N N N , p p p , k k k , and not on a positive lower bound for v v v .
Gradient estimate for regularized equation
For ε > 0 \varepsilon>0 ε > 0 , we consider the regularized equation
∂ t ( u ε p − 1 ) = div ( ( ε 2 + ∣ ∇ u ε ∣ 2 ) p − 2 2 ∇ u ε ) . (32) \partial_t\left(u_\varepsilon^{p-1}\right)=\operatorname{div}\left(\left(\varepsilon^2+|\nabla u_\varepsilon|^2\right)^{\frac{p-2}{2}}\nabla u_\varepsilon\right). \tag*{(32)} ∂ t ( u ε p − 1 ) = div ( ( ε 2 + ∣∇ u ε ∣ 2 ) 2 p − 2 ∇ u ε ) . (32)
We work with smooth solutions on a cylinder Q = B 2 r ( x o ) × ( s − 2 h , s ] Q=B_{2r}(x_o)\times(s-2h,s] Q = B 2 r ( x o ) × ( s − 2 h , s ] with Q ‾ ⊂ Ω T \overline{Q}\subset\Omega_T Q ⊂ Ω T , satisfying 0 < m ≤ u ε ≤ M 0<m\le u_\varepsilon\le M 0 < m ≤ u ε ≤ M in Q Q Q , where m , M m,M m , M do not depend on ε \varepsilon ε . All constants below are independent of ε , m , M , r , h \varepsilon,m,M,r,h ε , m , M , r , h . The treatment of points where the limiting solution vanishes requires a separate step. Throughout this section, k k k satisfies (3.1).
The transformed equation
The argument follows the same strategy as in the previous section for smooth positive solutions. We set
v ε : = u ε 1 k , w ε : = ∣ ∇ v ε ∣ 2 , v_\varepsilon:=u_\varepsilon^{\frac{1}{k}},\qquad w_\varepsilon:=|\nabla v_\varepsilon|^2, v ε := u ε k 1 , w ε := ∣∇ v ε ∣ 2 ,
and recall
c : = k p − 2 p − 1 , a ( v ) : = c v 2 − p . c:=\frac{k^{p-2}}{p-1},\qquad a(v):=cv^{2-p}. c := p − 1 k p − 2 , a ( v ) := c v 2 − p .
The regularization introduces an additional dependence on v ε v_\varepsilon v ε in the transformed coefficients, which requires a separate calculation. Since u ε = v ε k u_\varepsilon=v_\varepsilon^k u ε = v ε k , we have
∇ u ε = k v ε k − 1 ∇ v ε \nabla u_\varepsilon=kv_\varepsilon^{k-1}\nabla v_\varepsilon ∇ u ε = k v ε k − 1 ∇ v ε
and
∇ 2 u ε = k v ε k − 1 ∇ 2 v ε + k ( k − 1 ) v ε k − 2 ∇ v ε ⊗ ∇ v ε . \nabla^2u_\varepsilon=kv_\varepsilon^{k-1}\nabla^2v_\varepsilon+k(k-1)v_\varepsilon^{k-2}\nabla v_\varepsilon\otimes\nabla v_\varepsilon. ∇ 2 u ε = k v ε k − 1 ∇ 2 v ε + k ( k − 1 ) v ε k − 2 ∇ v ε ⊗ ∇ v ε .
Moreover,
∂ t ( u ε k ( p − 1 ) ) = k ( p − 1 ) v ε k ( p − 1 ) − 1 ∂ t v ε . \partial_t(u_\varepsilon^{k(p-1)})=k(p-1)v_\varepsilon^{k(p-1)-1}\partial_tv_\varepsilon. ∂ t ( u ε k ( p − 1 ) ) = k ( p − 1 ) v ε k ( p − 1 ) − 1 ∂ t v ε .
To expand the regularized divergence, we use
div ( ( ε 2 + ∣ ∇ u ε ∣ 2 ) p − 2 2 ∇ u ε ) = ( ε 2 + ∣ ∇ u ε ∣ 2 ) p − 2 2 ( I + ( p − 2 ) ∇ u ε ⊗ ∇ u ε ε 2 + ∣ ∇ u ε ∣ 2 ) ⋅ ∇ 2 u ε . \operatorname{div}\left((\varepsilon^2+|\nabla u_\varepsilon|^2)^{\frac{p-2}{2}}\nabla u_\varepsilon\right)
=(\varepsilon^2+|\nabla u_\varepsilon|^2)^{\frac{p-2}{2}}\left(I+(p-2)\frac{\nabla u_\varepsilon\otimes\nabla u_\varepsilon}{\varepsilon^2+|\nabla u_\varepsilon|^2}\right)\cdot\nabla^2u_\varepsilon. div ( ( ε 2 + ∣∇ u ε ∣ 2 ) 2 p − 2 ∇ u ε ) = ( ε 2 + ∣∇ u ε ∣ 2 ) 2 p − 2 ( I + ( p − 2 ) ε 2 + ∣∇ u ε ∣ 2 ∇ u ε ⊗ ∇ u ε ) ⋅ ∇ 2 u ε .
We introduce
w ε : = w ε + ε 2 k 2 v ε 2 ( k − 1 ) , θ ε : = w ε w ε . \mathbf{w}_\varepsilon:=w_\varepsilon+\frac{\varepsilon^2}{k^2v_\varepsilon^{2(k-1)}},\qquad\theta_\varepsilon:=\frac{w_\varepsilon}{\mathbf{w}_\varepsilon}. w ε := w ε + k 2 v ε 2 ( k − 1 ) ε 2 , θ ε := w ε w ε .
With this notation,
ε 2 + ∣ ∇ u ε ∣ 2 = k 2 v ε 2 ( k − 1 ) w ε \varepsilon^2+|\nabla u_\varepsilon|^2=k^2v_\varepsilon^{2(k-1)}\mathbf{w}_\varepsilon ε 2 + ∣∇ u ε ∣ 2 = k 2 v ε 2 ( k − 1 ) w ε
and
∇ u ε ⊗ ∇ u ε ε 2 + ∣ ∇ u ε ∣ 2 = ∇ v ε ⊗ ∇ v ε w ε . \frac{\nabla u_\varepsilon\otimes\nabla u_\varepsilon}{\varepsilon^2+|\nabla u_\varepsilon|^2}=\frac{\nabla v_\varepsilon\otimes\nabla v_\varepsilon}{\mathbf{w}_\varepsilon}. ε 2 + ∣∇ u ε ∣ 2 ∇ u ε ⊗ ∇ u ε = w ε ∇ v ε ⊗ ∇ v ε .
Substituting these expressions into (32) and dividing by k ( p − 1 ) v ε k ( p − 1 ) − 1 k(p-1)v_\varepsilon^{k(p-1)-1} k ( p − 1 ) v ε k ( p − 1 ) − 1 yields
∂ t v ε = a ( v ε ) w ε p − 2 2 ( I + ( p − 2 ) ∇ v ε ⊗ ∇ v ε w ε ) ⋅ ∇ 2 v ε + a ( v ε ) ( k − 1 ) v ε w ε p − 2 2 ( I + ( p − 2 ) ∇ v ε ⊗ ∇ v ε w ε ) ⋅ ( ∇ v ε ⊗ ∇ v ε ) . \partial_tv_\varepsilon=a(v_\varepsilon)\mathbf{w}_\varepsilon^{\frac{p-2}{2}}\left(I+(p-2)\frac{\nabla v_\varepsilon\otimes\nabla v_\varepsilon}{\mathbf{w}_\varepsilon}\right)\cdot\nabla^2v_\varepsilon
+\frac{a(v_\varepsilon)(k-1)}{v_\varepsilon}\mathbf{w}_\varepsilon^{\frac{p-2}{2}}\left(I+(p-2)\frac{\nabla v_\varepsilon\otimes\nabla v_\varepsilon}{\mathbf{w}_\varepsilon}\right)\cdot(\nabla v_\varepsilon\otimes\nabla v_\varepsilon). ∂ t v ε = a ( v ε ) w ε 2 p − 2 ( I + ( p − 2 ) w ε ∇ v ε ⊗ ∇ v ε ) ⋅ ∇ 2 v ε + v ε a ( v ε ) ( k − 1 ) w ε 2 p − 2 ( I + ( p − 2 ) w ε ∇ v ε ⊗ ∇ v ε ) ⋅ ( ∇ v ε ⊗ ∇ v ε ) .
The contraction in the last term satisfies
( I + ( p − 2 ) ∇ v ε ⊗ ∇ v ε w ε ) ⋅ ( ∇ v ε ⊗ ∇ v ε ) = w ε + ( p − 2 ) w ε 2 w ε = w ε ( 1 + ( p − 2 ) θ ε ) . \left(I+(p-2)\frac{\nabla v_\varepsilon\otimes\nabla v_\varepsilon}{\mathbf{w}_\varepsilon}\right)\cdot(\nabla v_\varepsilon\otimes\nabla v_\varepsilon)=w_\varepsilon+(p-2)\frac{w_\varepsilon^2}{\mathbf{w}_\varepsilon}
=w_\varepsilon(1+(p-2)\theta_\varepsilon). ( I + ( p − 2 ) w ε ∇ v ε ⊗ ∇ v ε ) ⋅ ( ∇ v ε ⊗ ∇ v ε ) = w ε + ( p − 2 ) w ε w ε 2 = w ε ( 1 + ( p − 2 ) θ ε ) .
Consequently,
∂ t v ε = a ( v ε ) w ε p − 2 2 ( I + ( p − 2 ) ∇ v ε ⊗ ∇ v ε w ε ) ⋅ ∇ 2 v ε + a ( v ε ) ( k − 1 ) v ε w ε p − 2 2 w ε ( 1 + ( p − 2 ) θ ε ) . (33) \partial_tv_\varepsilon=a(v_\varepsilon)\mathbf{w}_\varepsilon^{\frac{p-2}{2}}\left(I+(p-2)\frac{\nabla v_\varepsilon\otimes\nabla v_\varepsilon}{\mathbf{w}_\varepsilon}\right)\cdot\nabla^2v_\varepsilon
+\frac{a(v_\varepsilon)(k-1)}{v_\varepsilon}\mathbf{w}_\varepsilon^{\frac{p-2}{2}}w_\varepsilon(1+(p-2)\theta_\varepsilon). \tag*{(33)} ∂ t v ε = a ( v ε ) w ε 2 p − 2 ( I + ( p − 2 ) w ε ∇ v ε ⊗ ∇ v ε ) ⋅ ∇ 2 v ε + v ε a ( v ε ) ( k − 1 ) w ε 2 p − 2 w ε ( 1 + ( p − 2 ) θ ε ) . (33)
Notice that the regularization contributes the term
ε 2 k 2 v ε 2 ( k − 1 ) \frac{\varepsilon^2}{k^2v_\varepsilon^{2(k-1)}} k 2 v ε 2 ( k − 1 ) ε 2
to w ε w_\varepsilon w ε . Since this term depends on v ε v_\varepsilon v ε , differentiating the coefficients in (33) produces additional terms that we compute below.
The coefficients and their derivatives
We now work at points where w ε > 0 w_\varepsilon> 0 w ε > 0 . Then 0 < θ ε < 1 0 < \theta_\varepsilon< 1 0 < θ ε < 1 and w ε = w ε / θ ε w_\varepsilon= w_\varepsilon/\theta_\varepsilon w ε = w ε / θ ε . To keep the dependence on the gradient direction explicit, for θ ∈ ( 0 , 1 ] \theta\in(0,1] θ ∈ ( 0 , 1 ] and ξ ∈ R N ∖ { 0 } \xi\in\mathbb{R}^{N} \setminus\{0\} ξ ∈ R N ∖ { 0 } we define
C θ ( ξ ) : = θ − p − 2 2 ( I + ( p − 2 ) ξ ⊗ ξ ∣ ξ ∣ 2 ) C_\theta(\xi) := \theta^{-\frac{p-2}{2}}\left(I+(p-2)\frac{\xi\otimes\xi}{|\xi|^2}\right) C θ ( ξ ) := θ − 2 p − 2 ( I + ( p − 2 ) ∣ ξ ∣ 2 ξ ⊗ ξ )
and
ȷ ( θ ) : = θ − p − 2 2 ( 1 + ( p − 2 ) θ ) . \jmath(\theta) := \theta^{-\frac{p-2}{2}}(1+(p-2)\theta). ( θ ) := θ − 2 p − 2 ( 1 + ( p − 2 ) θ ) .
When differentiating C θ ( ξ ) C_\theta(\xi) C θ ( ξ ) with respect to θ \theta θ , we keep ξ \xi ξ fixed. For independent variables v > 0 v>0 v > 0 and ξ ∈ R N \xi\in\mathbb{R}^{N} ξ ∈ R N , we define the principal coefficient matrix by
A ε ( v , ξ ) : = a ( v ) ( ∣ ξ ∣ 2 + ε 2 k 2 v 2 ( k − 1 ) ) p − 2 2 ( I + ( p − 2 ) ξ ⊗ ξ ∣ ξ ∣ 2 + ε 2 k 2 v 2 ( k − 1 ) ) , A_\varepsilon(v,\xi):=a(v)\left(|\xi|^2+\frac{\varepsilon^2}{k^2v^{2(k-1)}}\right)^{\frac{p-2}{2}}\left(I+(p-2)\frac{\xi\otimes\xi}{|\xi|^2+\frac{\varepsilon^2}{k^2v^{2(k-1)}}}\right), A ε ( v , ξ ) := a ( v ) ( ∣ ξ ∣ 2 + k 2 v 2 ( k − 1 ) ε 2 ) 2 p − 2 ( I + ( p − 2 ) ∣ ξ ∣ 2 + k 2 v 2 ( k − 1 ) ε 2 ξ ⊗ ξ ) ,
and the scalar term by
F ε ( v , ξ ) : = a ( v ) ( k − 1 ) v ( ∣ ξ ∣ 2 + ε 2 k 2 v 2 ( k − 1 ) ) p 2 ∣ ξ ∣ 2 ( 1 + ( p − 2 ) ∣ ξ ∣ 2 ∣ ξ ∣ 2 + ε 2 k 2 v 2 ( k − 1 ) ) . F_\varepsilon(v,\xi):=\frac{a(v)(k-1)}{v}\left(|\xi|^2+\frac{\varepsilon^2}{k^2v^{2(k-1)}}\right)^{\frac{p}{2}}|\xi|^2\left(1+(p-2)\frac{|\xi|^2}{|\xi|^2+\frac{\varepsilon^2}{k^2v^{2(k-1)}}}\right). F ε ( v , ξ ) := v a ( v ) ( k − 1 ) ( ∣ ξ ∣ 2 + k 2 v 2 ( k − 1 ) ε 2 ) 2 p ∣ ξ ∣ 2 ( 1 + ( p − 2 ) ∣ ξ ∣ 2 + k 2 v 2 ( k − 1 ) ε 2 ∣ ξ ∣ 2 ) .
Here v v v and ξ \xi ξ are independent arguments of the coefficient functions. They do not replace the notation v ε v_\varepsilon v ε for the transformed solution. At points where ∇ v ε ≠ 0 \nabla v_\varepsilon\ne0 ∇ v ε = 0 , evaluating at ( v , ξ ) = ( v ε , ∇ v ε ) (v,\xi)=(v_\varepsilon,\nabla v_\varepsilon) ( v , ξ ) = ( v ε , ∇ v ε ) gives the principal coefficient matrix in the form
A ε ( v ε , ∇ v ε ) = a ( v ε ) w ε p − 2 2 C θ ε ( ∇ v ε ) , A_\varepsilon(v_\varepsilon,\nabla v_\varepsilon)=a(v_\varepsilon)w_\varepsilon^{\frac{p-2}{2}}C_{\theta_\varepsilon}(\nabla v_\varepsilon), A ε ( v ε , ∇ v ε ) = a ( v ε ) w ε 2 p − 2 C θ ε ( ∇ v ε ) ,
while the scalar term is given by
F ε ( v ε , ∇ v ε ) = a ( v ε ) ( k − 1 ) v ε w ε p 2 ȷ ( θ ε ) . F_\varepsilon(v_\varepsilon,\nabla v_\varepsilon)=\frac{a(v_\varepsilon)(k-1)}{v_\varepsilon}w_\varepsilon^{\frac{p}{2}}\jmath(\theta_\varepsilon). F ε ( v ε , ∇ v ε ) = v ε a ( v ε ) ( k − 1 ) w ε 2 p ( θ ε ) .
Therefore, equation (33) can be re-written as
∂ t v ε = A ε ( v ε , ∇ v ε ) ⋅ ∇ 2 v ε + F ε ( v ε , ∇ v ε ) . (34) \partial_t v_\varepsilon=A_\varepsilon(v_\varepsilon,\nabla v_\varepsilon)\cdot\nabla^2v_\varepsilon+F_\varepsilon(v_\varepsilon,\nabla v_\varepsilon). \tag*{(34)} ∂ t v ε = A ε ( v ε , ∇ v ε ) ⋅ ∇ 2 v ε + F ε ( v ε , ∇ v ε ) . (34)
To carry out the Bernstein argument, we derive an equation for w ε = ∣ ∇ v ε ∣ 2 w_\varepsilon=|\nabla v_\varepsilon|^2 w ε = ∣∇ v ε ∣ 2 by differentiating (34) with respect to the spatial variables. The chain rule then produces derivatives of A ε A_\varepsilon A ε and F ε F_\varepsilon F ε with respect to both their scalar and vector arguments. We first compute the partial derivatives with respect to the scalar argument, keeping the vector argument fixed. For this purpose, we introduce the function
Θ ε ( v , ξ ) : = ∣ ξ ∣ 2 ∣ ξ ∣ 2 + ε 2 k 2 v 2 ( k − 1 ) , v > 0 , ξ ∈ R N . \Theta_\varepsilon(v,\xi):=\frac{|\xi|^2}{|\xi|^2+\frac{\varepsilon^2}{k^2v^{2(k-1)}}},\qquad v>0,\quad\xi\in\mathbb{R}^{N}. Θ ε ( v , ξ ) := ∣ ξ ∣ 2 + k 2 v 2 ( k − 1 ) ε 2 ∣ ξ ∣ 2 , v > 0 , ξ ∈ R N .
The previously defined quantity θ ε \theta_\varepsilon θ ε is then given by
θ ε = Θ ε ( v ε , ∇ v ε ) . \theta_\varepsilon=\Theta_\varepsilon(v_\varepsilon,\nabla v_\varepsilon). θ ε = Θ ε ( v ε , ∇ v ε ) .
Since
∂ v ( ∣ ξ ∣ 2 + ε 2 k 2 v 2 ( k − 1 ) ) = − 2 ( k − 1 ) v ε 2 k 2 v 2 ( k − 1 ) , \partial_v\left(|\xi|^2+\frac{\varepsilon^2}{k^2v^{2(k-1)}}\right)=-\frac{2(k-1)}{v}\frac{\varepsilon^2}{k^2v^{2(k-1)}}, ∂ v ( ∣ ξ ∣ 2 + k 2 v 2 ( k − 1 ) ε 2 ) = − v 2 ( k − 1 ) k 2 v 2 ( k − 1 ) ε 2 ,
we obtain
∂ v Θ ε ( v , ξ ) = 2 ( k − 1 ) v ∣ ξ ∣ 2 ∣ ξ ∣ 2 + ε 2 k 2 v 2 ( k − 1 ) ε 2 k 2 v 2 ( k − 1 ) ∣ ξ ∣ 2 + ε 2 k 2 v 2 ( k − 1 ) = 2 ( k − 1 ) v Θ ε ( v , ξ ) ( 1 − Θ ε ( v , ξ ) ) . (35) \begin{aligned}
\partial_v\Theta_\varepsilon(v,\xi)
&=\frac{2(k-1)}{v}\frac{|\xi|^2}{|\xi|^2+\frac{\varepsilon^2}{k^2v^{2(k-1)}}}\frac{\frac{\varepsilon^2}{k^2v^{2(k-1)}}}{|\xi|^2+\frac{\varepsilon^2}{k^2v^{2(k-1)}}}\\
&=\frac{2(k-1)}{v}\Theta_\varepsilon(v,\xi)(1-\Theta_\varepsilon(v,\xi)). \tag*{(35)}
\end{aligned} ∂ v Θ ε ( v , ξ ) = v 2 ( k − 1 ) ∣ ξ ∣ 2 + k 2 v 2 ( k − 1 ) ε 2 ∣ ξ ∣ 2 ∣ ξ ∣ 2 + k 2 v 2 ( k − 1 ) ε 2 k 2 v 2 ( k − 1 ) ε 2 = v 2 ( k − 1 ) Θ ε ( v , ξ ) ( 1 − Θ ε ( v , ξ )) . (35)
We also compute the derivative of C θ ( ξ ) C_\theta(\xi) C θ ( ξ ) with respect to θ \theta θ , keeping ξ ≠ 0 \xi\ne0 ξ = 0 fixed. From its definition, the product rule gives
∂ θ C θ ( ξ ) = − 1 2 ( p − 2 ) θ − p 2 ( I + ( p − 2 ) θ ξ ⊗ ξ ∣ ξ ∣ 2 ) + ( p − 2 ) θ − p − 2 2 ξ ⊗ ξ ∣ ξ ∣ 2 = − 1 2 ( p − 2 ) θ − p 2 I + 1 2 ( p − 2 ) ( 4 − p ) θ − p − 2 2 ξ ⊗ ξ ∣ ξ ∣ 2 . (36) \begin{aligned}
\partial_\theta C_\theta(\xi)
&= -\frac{1}{2}(p-2)\theta^{-\frac{p}{2}}\left(I+(p-2)\theta\frac{\xi\otimes\xi}{|\xi|^2}\right)+(p-2)\theta^{-\frac{p-2}{2}}\frac{\xi\otimes\xi}{|\xi|^2}\\
&= -\frac{1}{2}(p-2)\theta^{-\frac{p}{2}}I+\frac{1}{2}(p-2)(4-p)\theta^{-\frac{p-2}{2}}\frac{\xi\otimes\xi}{|\xi|^2}. \tag*{(36)}
\end{aligned} ∂ θ C θ ( ξ ) = − 2 1 ( p − 2 ) θ − 2 p ( I + ( p − 2 ) θ ∣ ξ ∣ 2 ξ ⊗ ξ ) + ( p − 2 ) θ − 2 p − 2 ∣ ξ ∣ 2 ξ ⊗ ξ = − 2 1 ( p − 2 ) θ − 2 p I + 2 1 ( p − 2 ) ( 4 − p ) θ − 2 p − 2 ∣ ξ ∣ 2 ξ ⊗ ξ . (36)
Similarly, differentiating j ( θ ) \mathbf{j}(\theta) j ( θ ) , we obtain
j ′ ( θ ) = − 1 2 ( p − 2 ) θ − p 2 + 1 2 ( p − 2 ) ( 4 − p ) θ − p − 2 2 . (37) \mathbf{j}'(\theta)=-\frac{1}{2}(p-2)\theta^{-\frac{p}{2}}+\frac{1}{2}(p-2)(4-p)\theta^{-\frac{p-2}{2}}. \tag*{(37)} j ′ ( θ ) = − 2 1 ( p − 2 ) θ − 2 p + 2 1 ( p − 2 ) ( 4 − p ) θ − 2 p − 2 . (37)
With these derivatives at hand, we first compute ∂ v A ε ( v , ξ ) \partial_v\mathbf{A}_\varepsilon(v,\xi) ∂ v A ε ( v , ξ ) , keeping the vector argument ξ \xi ξ fixed. For v > 0 v>0 v > 0 and ξ ≠ 0 \xi\ne0 ξ = 0 , we have
A ε ( v , ξ ) = a ( v ) ∣ ξ ∣ p − 2 C Θ ε ( v , ξ ) ( ξ ) . \mathbf{A}_\varepsilon(v,\xi)=a(v)|\xi|^{p-2}\mathbf{C}_{\Theta_\varepsilon(v,\xi)}(\xi). A ε ( v , ξ ) = a ( v ) ∣ ξ ∣ p − 2 C Θ ε ( v , ξ ) ( ξ ) .
The product and chain rules give
∂ v A ε ( v , ξ ) = a ′ ( v ) ∣ ξ ∣ p − 2 C Θ ε ( v , ξ ) ( ξ ) + a ( v ) ∣ ξ ∣ p − 2 ∂ θ C θ ( ξ ) ∣ θ = Θ ε ( v , ξ ) ∂ v Θ ε ( v , ξ ) . \partial_v\mathbf{A}_\varepsilon(v,\xi)=a'(v)|\xi|^{p-2}\mathbf{C}_{\Theta_\varepsilon(v,\xi)}(\xi)+a(v)|\xi|^{p-2}\left.\partial_\theta\mathbf{C}_\theta(\xi)\right|_{\theta=\Theta_\varepsilon(v,\xi)}\partial_v\Theta_\varepsilon(v,\xi). ∂ v A ε ( v , ξ ) = a ′ ( v ) ∣ ξ ∣ p − 2 C Θ ε ( v , ξ ) ( ξ ) + a ( v ) ∣ ξ ∣ p − 2 ∂ θ C θ ( ξ ) ∣ θ = Θ ε ( v , ξ ) ∂ v Θ ε ( v , ξ ) .
Using a ′ ( v ) = ( 2 − p ) a ( v ) / v a'(v)=(2-p)a(v)/v a ′ ( v ) = ( 2 − p ) a ( v ) / v and (35) , and factoring out a ( v ) v ∣ ξ ∣ p − 2 \frac{a(v)}{v}|\xi|^{p-2} v a ( v ) ∣ ξ ∣ p − 2 , we obtain
∂ v A ε ( v , ξ ) = a ( v ) v ∣ ξ ∣ p − 2 [ ( 2 − p ) C Θ ε ( v , ξ ) ( ξ ) + 2 ( k − 1 ) Θ ε ( v , ξ ) ( 1 − Θ ε ( v , ξ ) ) ∂ θ C θ ( ξ ) ∣ θ = Θ ε ( v , ξ ) ] . \partial_v\mathbf{A}_\varepsilon(v,\xi)=\frac{a(v)}{v}|\xi|^{p-2}\left[(2-p)\mathbf{C}_{\Theta_\varepsilon(v,\xi)}(\xi)+2(k-1)\Theta_\varepsilon(v,\xi)(1-\Theta_\varepsilon(v,\xi))\left.\partial_\theta\mathbf{C}_\theta(\xi)\right|_{\theta=\Theta_\varepsilon(v,\xi)}\right]. ∂ v A ε ( v , ξ ) = v a ( v ) ∣ ξ ∣ p − 2 [ ( 2 − p ) C Θ ε ( v , ξ ) ( ξ ) + 2 ( k − 1 ) Θ ε ( v , ξ ) ( 1 − Θ ε ( v , ξ )) ∂ θ C θ ( ξ ) ∣ θ = Θ ε ( v , ξ ) ] .
The preceding calculation treats v v v and ξ \xi ξ as independent variables. To use the resulting coefficient derivative in the equation for v ε v_\varepsilon v ε , we now evaluate it at ( v , ξ ) = ( v ε , ∇ v ε ) (v,\xi)=(v_\varepsilon,\nabla v_\varepsilon) ( v , ξ ) = ( v ε , ∇ v ε ) . Here ( ∂ v A ε ) ( v ε , ∇ v ε ) (\partial_v\mathbf{A}_\varepsilon)(v_\varepsilon,\nabla v_\varepsilon) ( ∂ v A ε ) ( v ε , ∇ v ε ) denotes the partial derivative with respect to the scalar argument, computed with ξ \xi ξ fixed and then evaluated along the solution. Since Θ ε ( v ε , ∇ v ε ) = θ ε \Theta_\varepsilon(v_\varepsilon,\nabla v_\varepsilon)=\theta_\varepsilon Θ ε ( v ε , ∇ v ε ) = θ ε and ∣ ∇ v ε ∣ p − 2 = w ε p − 2 2 |\nabla v_\varepsilon|^{p-2}=w_\varepsilon^{\frac{p-2}{2}} ∣∇ v ε ∣ p − 2 = w ε 2 p − 2 , we obtain
( ∂ v A ε ) ( v ε , ∇ v ε ) = a ( v ε ) v ε w ε p − 2 2 [ ( 2 − p ) C θ ε ( ∇ v ε ) + 2 ( k − 1 ) θ ε ( 1 − θ ε ) ∂ θ C θ ( ∇ v ε ) ∣ θ = θ ε ] . \begin{aligned}
(\partial_v\mathbf{A}_\varepsilon)(v_\varepsilon,\nabla v_\varepsilon)
&=\frac{a(v_\varepsilon)}{v_\varepsilon}w_\varepsilon^{\frac{p-2}{2}}\left[(2-p)\mathbf{C}_{\theta_\varepsilon}(\nabla v_\varepsilon)\right.\\
&\qquad\left.+2(k-1)\theta_\varepsilon(1-\theta_\varepsilon)\left.\partial_\theta\mathbf{C}_\theta(\nabla v_\varepsilon)\right|_{\theta=\theta_\varepsilon}\right].
\end{aligned} ( ∂ v A ε ) ( v ε , ∇ v ε ) = v ε a ( v ε ) w ε 2 p − 2 [ ( 2 − p ) C θ ε ( ∇ v ε ) + 2 ( k − 1 ) θ ε ( 1 − θ ε ) ∂ θ C θ ( ∇ v ε ) ∣ θ = θ ε ] .
We next differentiate the scalar term
F ε ( v , ξ ) = ( k − 1 ) a ( v ) v ∣ ξ ∣ p j ( Θ ε ( v , ξ ) ) F_\varepsilon(v,\xi)=(k-1)\frac{a(v)}{v}|\xi|^p\mathbf{j}(\Theta_\varepsilon(v,\xi)) F ε ( v , ξ ) = ( k − 1 ) v a ( v ) ∣ ξ ∣ p j ( Θ ε ( v , ξ ))
with respect to v v v , keeping ξ ≠ 0 \xi\ne0 ξ = 0 fixed. First, since a ′ ( v ) = ( 2 − p ) a ( v ) / v a'(v)=(2-p)a(v)/v a ′ ( v ) = ( 2 − p ) a ( v ) / v , we have
d d v ( a ( v ) v ) = ( 1 − p ) a ( v ) v 2 . \frac{\mathrm{d}}{\mathrm{d}v}\left(\frac{a(v)}{v}\right)=(1-p)\frac{a(v)}{v^2}. d v d ( v a ( v ) ) = ( 1 − p ) v 2 a ( v ) .
The product and chain rules give
∂ v F ε ( v , ξ ) = ( k − 1 ) ( 1 − p ) a ( v ) v 2 ∣ ξ ∣ p j ( Θ ε ( v , ξ ) ) + ( k − 1 ) a ( v ) v ∣ ξ ∣ p j ′ ( Θ ε ( v , ξ ) ) ∂ v Θ ε ( v , ξ ) = a ( v ) ( k − 1 ) v 2 ∣ ξ ∣ p [ ( 1 − p ) j ( Θ ε ( v , ξ ) ) + v j ′ ( Θ ε ( v , ξ ) ) ∂ v Θ ε ( v , ξ ) ] . \begin{aligned}
\partial_vF_\varepsilon(v,\xi)
&=(k-1)(1-p)\frac{a(v)}{v^2}|\xi|^p\mathbf{j}(\Theta_\varepsilon(v,\xi))+(k-1)\frac{a(v)}{v}|\xi|^p\mathbf{j}'(\Theta_\varepsilon(v,\xi))\partial_v\Theta_\varepsilon(v,\xi)\\
&=\frac{a(v)(k-1)}{v^2}|\xi|^p\left[(1-p)\mathbf{j}(\Theta_\varepsilon(v,\xi))+v\mathbf{j}'(\Theta_\varepsilon(v,\xi))\partial_v\Theta_\varepsilon(v,\xi)\right].
\end{aligned} ∂ v F ε ( v , ξ ) = ( k − 1 ) ( 1 − p ) v 2 a ( v ) ∣ ξ ∣ p j ( Θ ε ( v , ξ )) + ( k − 1 ) v a ( v ) ∣ ξ ∣ p j ′ ( Θ ε ( v , ξ )) ∂ v Θ ε ( v , ξ ) = v 2 a ( v ) ( k − 1 ) ∣ ξ ∣ p [ ( 1 − p ) j ( Θ ε ( v , ξ )) + v j ′ ( Θ ε ( v , ξ )) ∂ v Θ ε ( v , ξ ) ] .
Substituting (35) , we obtain
∂ v F ε ( v , ξ ) = a ( v ) ( k − 1 ) v 2 ∣ ξ ∣ p [ ( 1 − p ) j ( Θ ε ( v , ξ ) ) + 2 ( k − 1 ) Θ ε ( v , ξ ) ( 1 − Θ ε ( v , ξ ) ) j ′ ( Θ ε ( v , ξ ) ) ] . \partial_vF_\varepsilon(v,\xi)=\frac{a(v)(k-1)}{v^2}|\xi|^p\left[(1-p)\mathbf{j}(\Theta_\varepsilon(v,\xi))+2(k-1)\Theta_\varepsilon(v,\xi)(1-\Theta_\varepsilon(v,\xi))\mathbf{j}'(\Theta_\varepsilon(v,\xi))\right]. ∂ v F ε ( v , ξ ) = v 2 a ( v ) ( k − 1 ) ∣ ξ ∣ p [ ( 1 − p ) j ( Θ ε ( v , ξ )) + 2 ( k − 1 ) Θ ε ( v , ξ ) ( 1 − Θ ε ( v , ξ )) j ′ ( Θ ε ( v , ξ )) ] .
Evaluating along the solution as before, we obtain
( ∂ v F ε ) ( v ε , ∇ v ε ) = a ( v ε ) ( k − 1 ) v ε 2 w ε p 2 [ ( 1 − p ) j ( θ ε ) + 2 ( k − 1 ) θ ε ( 1 − θ ε ) j ′ ( θ ε ) ] . (\partial_v F_\varepsilon)(v_\varepsilon,\nabla v_\varepsilon)=\frac{a(v_\varepsilon)(k-1)}{v_\varepsilon^2}w_\varepsilon^{\frac{p}{2}}\left[(1-p)j(\theta_\varepsilon)+2(k-1)\theta_\varepsilon(1-\theta_\varepsilon)j'(\theta_\varepsilon)\right]. ( ∂ v F ε ) ( v ε , ∇ v ε ) = v ε 2 a ( v ε ) ( k − 1 ) w ε 2 p [ ( 1 − p ) j ( θ ε ) + 2 ( k − 1 ) θ ε ( 1 − θ ε ) j ′ ( θ ε ) ] .
To write the preceding formulas for ( ∂ v A ε ) ( v ε , ∇ v ε ) (\partial_v A_\varepsilon)(v_\varepsilon,\nabla v_\varepsilon) ( ∂ v A ε ) ( v ε , ∇ v ε ) and ( ∂ v F ε ) ( v ε , ∇ v ε ) (\partial_v F_\varepsilon)(v_\varepsilon,\nabla v_\varepsilon) ( ∂ v F ε ) ( v ε , ∇ v ε ) more compactly, we introduce, for θ ∈ ( 0 , 1 ] \theta\in(0,1] θ ∈ ( 0 , 1 ] and ξ ≠ 0 \xi\ne0 ξ = 0 ,
D θ ( ξ ) : = ( 2 − p ) C θ ( ξ ) + 2 ( k − 1 ) θ ( 1 − θ ) ∂ θ C θ ( ξ ) , (38) D_\theta(\xi):=(2-p)C_\theta(\xi)+2(k-1)\theta(1-\theta)\partial_\theta C_\theta(\xi), \tag*{(38)} D θ ( ξ ) := ( 2 − p ) C θ ( ξ ) + 2 ( k − 1 ) θ ( 1 − θ ) ∂ θ C θ ( ξ ) , (38)
and
J ( θ ) : = ( 1 − p ) j ( θ ) + 2 ( k − 1 ) θ ( 1 − θ ) j ′ ( θ ) . J(\theta):=(1-p)j(\theta)+2(k-1)\theta(1-\theta)j'(\theta). J ( θ ) := ( 1 − p ) j ( θ ) + 2 ( k − 1 ) θ ( 1 − θ ) j ′ ( θ ) .
Thus,
( ∂ v A ε ) ( v ε , ∇ v ε ) = a ( v ε ) v ε w ε p − 2 2 D θ ε ( ∇ v ε ) , (39) (\partial_v A_\varepsilon)(v_\varepsilon,\nabla v_\varepsilon)=\frac{a(v_\varepsilon)}{v_\varepsilon}w_\varepsilon^{\frac{p-2}{2}}D_{\theta_\varepsilon}(\nabla v_\varepsilon), \tag*{(39)} ( ∂ v A ε ) ( v ε , ∇ v ε ) = v ε a ( v ε ) w ε 2 p − 2 D θ ε ( ∇ v ε ) , (39)
and
( ∂ v F ε ) ( v ε , ∇ v ε ) = a ( v ε ) ( k − 1 ) v ε 2 w ε p 2 J ( θ ε ) . (40) (\partial_v F_\varepsilon)(v_\varepsilon,\nabla v_\varepsilon)=\frac{a(v_\varepsilon)(k-1)}{v_\varepsilon^2}w_\varepsilon^{\frac{p}{2}}J(\theta_\varepsilon). \tag*{(40)} ( ∂ v F ε ) ( v ε , ∇ v ε ) = v ε 2 a ( v ε ) ( k − 1 ) w ε 2 p J ( θ ε ) . (40)
The differentiated equation
For ℓ = 1 , … , N \ell=1,\ldots,N ℓ = 1 , … , N , we define the components of the drift by
( B ε ) ℓ : = ( ∂ ξ ℓ A ε ) ( v ε , ∇ v ε ) ⋅ ∇ 2 v ε + ( ∂ ξ ℓ F ε ) ( v ε , ∇ v ε ) , (41) (B_\varepsilon)_\ell:=(\partial_{\xi_\ell}A_\varepsilon)(v_\varepsilon,\nabla v_\varepsilon)\cdot\nabla^2v_\varepsilon+(\partial_{\xi_\ell}F_\varepsilon)(v_\varepsilon,\nabla v_\varepsilon), \tag*{(41)} ( B ε ) ℓ := ( ∂ ξ ℓ A ε ) ( v ε , ∇ v ε ) ⋅ ∇ 2 v ε + ( ∂ ξ ℓ F ε ) ( v ε , ∇ v ε ) , (41)
and introduce the operator
L ε φ : = ∂ t φ − A ε ( v ε , ∇ v ε ) ⋅ ∇ 2 φ − B ε ⋅ ∇ φ (42) \mathcal{L}_\varepsilon\varphi:=\partial_t\varphi-A_\varepsilon(v_\varepsilon,\nabla v_\varepsilon)\cdot\nabla^2\varphi-B_\varepsilon\cdot\nabla\varphi \tag*{(42)} L ε φ := ∂ t φ − A ε ( v ε , ∇ v ε ) ⋅ ∇ 2 φ − B ε ⋅ ∇ φ (42)
for smooth scalar functions φ \varphi φ . We differentiate (34) with respect to x ℓ x_\ell x ℓ , multiply by 2 ∂ x ℓ v ε 2\partial_{x_\ell}v_\varepsilon 2 ∂ x ℓ v ε , and sum over ℓ = 1 , … , N \ell=1,\ldots,N ℓ = 1 , … , N . In the following calculation, all coefficient functions and their partial derivatives are evaluated at ( v ε , ∇ v ε ) (v_\varepsilon,\nabla v_\varepsilon) ( v ε , ∇ v ε ) . The chain rule, the definition of B ε B_\varepsilon B ε , and the commutation of mixed spatial derivatives give
∂ t ( ∂ x ℓ v ε ) = A ε ⋅ ∇ 2 ( ∂ x ℓ v ε ) + ( ∂ v A ε ) ⋅ ∇ 2 v ε ∂ x ℓ v ε + ( ∂ v F ε ) ∂ x ℓ v ε + ∑ j = 1 N ( ∂ ξ j A ε ) ⋅ ∇ 2 v ε ∂ x ℓ ∂ x j v ε + ∑ j = 1 N ( ∂ ξ j F ε ) ∂ x ℓ ∂ x j v ε = A ε ⋅ ∇ 2 ( ∂ x ℓ v ε ) + [ ( ∂ v A ε ) ⋅ ∇ 2 v ε + ( ∂ v F ε ) ] ∂ x ℓ v ε + ∑ j = 1 N [ ( ∂ ξ j A ε ) ⋅ ∇ 2 v ε + ( ∂ ξ j F ε ) ] ∂ x ℓ ∂ x j v ε = A ε ⋅ ∇ 2 ( ∂ x ℓ v ε ) + B ε ⋅ ∇ ( ∂ x ℓ v ε ) + [ ( ∂ v A ε ) ⋅ ∇ 2 v ε + ∂ v F ε ] ∂ x ℓ v ε . \begin{aligned}
\partial_t(\partial_{x_\ell}v_\varepsilon)
={}&A_\varepsilon\cdot\nabla^2(\partial_{x_\ell}v_\varepsilon)+(\partial_vA_\varepsilon)\cdot\nabla^2v_\varepsilon\,\partial_{x_\ell}v_\varepsilon+(\partial_vF_\varepsilon)\partial_{x_\ell}v_\varepsilon\\
&+\sum_{j=1}^{N}(\partial_{\xi_j}A_\varepsilon)\cdot\nabla^2v_\varepsilon\,\partial_{x_\ell}\partial_{x_j}v_\varepsilon+\sum_{j=1}^{N}(\partial_{\xi_j}F_\varepsilon)\partial_{x_\ell}\partial_{x_j}v_\varepsilon\\
={}&A_\varepsilon\cdot\nabla^2(\partial_{x_\ell}v_\varepsilon)+\left[(\partial_vA_\varepsilon)\cdot\nabla^2v_\varepsilon+(\partial_vF_\varepsilon)\right]\partial_{x_\ell}v_\varepsilon\\
&+\sum_{j=1}^{N}\left[(\partial_{\xi_j}A_\varepsilon)\cdot\nabla^2v_\varepsilon+(\partial_{\xi_j}F_\varepsilon)\right]\partial_{x_\ell}\partial_{x_j}v_\varepsilon\\
={}&A_\varepsilon\cdot\nabla^2(\partial_{x_\ell}v_\varepsilon)+B_\varepsilon\cdot\nabla(\partial_{x_\ell}v_\varepsilon)+\left[(\partial_vA_\varepsilon)\cdot\nabla^2v_\varepsilon+\partial_vF_\varepsilon\right]\partial_{x_\ell}v_\varepsilon.
\end{aligned} ∂ t ( ∂ x ℓ v ε ) = = = A ε ⋅ ∇ 2 ( ∂ x ℓ v ε ) + ( ∂ v A ε ) ⋅ ∇ 2 v ε ∂ x ℓ v ε + ( ∂ v F ε ) ∂ x ℓ v ε + j = 1 ∑ N ( ∂ ξ j A ε ) ⋅ ∇ 2 v ε ∂ x ℓ ∂ x j v ε + j = 1 ∑ N ( ∂ ξ j F ε ) ∂ x ℓ ∂ x j v ε A ε ⋅ ∇ 2 ( ∂ x ℓ v ε ) + [ ( ∂ v A ε ) ⋅ ∇ 2 v ε + ( ∂ v F ε ) ] ∂ x ℓ v ε + j = 1 ∑ N [ ( ∂ ξ j A ε ) ⋅ ∇ 2 v ε + ( ∂ ξ j F ε ) ] ∂ x ℓ ∂ x j v ε A ε ⋅ ∇ 2 ( ∂ x ℓ v ε ) + B ε ⋅ ∇ ( ∂ x ℓ v ε ) + [ ( ∂ v A ε ) ⋅ ∇ 2 v ε + ∂ v F ε ] ∂ x ℓ v ε .
Multiplying by 2 ∂ x ℓ v ε 2\partial_{x_\ell}v_\varepsilon 2 ∂ x ℓ v ε and summing over ℓ \ell ℓ , we obtain
∂ t w ε = 2 ∑ ℓ = 1 N ( ∂ x ℓ v ε ) A ε ⋅ ∇ 2 ( ∂ x ℓ v ε ) + B ε ⋅ ∇ w ε + 2 w ε ( ( ∂ v A ε ) ⋅ ∇ 2 v ε + ∂ v F ε ) . \partial_tw_\varepsilon=2\sum_{\ell=1}^{N}(\partial_{x_\ell}v_\varepsilon)A_\varepsilon\cdot\nabla^2(\partial_{x_\ell}v_\varepsilon)+B_\varepsilon\cdot\nabla w_\varepsilon+2w_\varepsilon\left((\partial_vA_\varepsilon)\cdot\nabla^2v_\varepsilon+\partial_vF_\varepsilon\right). ∂ t w ε = 2 ℓ = 1 ∑ N ( ∂ x ℓ v ε ) A ε ⋅ ∇ 2 ( ∂ x ℓ v ε ) + B ε ⋅ ∇ w ε + 2 w ε ( ( ∂ v A ε ) ⋅ ∇ 2 v ε + ∂ v F ε ) .
To rewrite the first term, we use
∇ 2 w ε = 2 ∑ ℓ = 1 N ∇ ( ∂ x ℓ v ε ) ⊗ ∇ ( ∂ x ℓ v ε ) + 2 ∑ ℓ = 1 N ( ∂ x ℓ v ε ) ∇ 2 ( ∂ x ℓ v ε ) . \nabla^2w_\varepsilon=2\sum_{\ell=1}^{N}\nabla(\partial_{x_\ell}v_\varepsilon)\otimes\nabla(\partial_{x_\ell}v_\varepsilon)+2\sum_{\ell=1}^{N}(\partial_{x_\ell}v_\varepsilon)\nabla^2(\partial_{x_\ell}v_\varepsilon). ∇ 2 w ε = 2 ℓ = 1 ∑ N ∇ ( ∂ x ℓ v ε ) ⊗ ∇ ( ∂ x ℓ v ε ) + 2 ℓ = 1 ∑ N ( ∂ x ℓ v ε ) ∇ 2 ( ∂ x ℓ v ε ) .
Let e 1 , … , e N e_1,\ldots,e_N e 1 , … , e N denote the standard basis of R N \mathbb{R}^N R N . Expanding the gradients in this basis and rearranging the sums, we obtain
∑ ℓ = 1 N ∇ ( ∂ x ℓ v ε ) ⊗ ∇ ( ∂ x ℓ v ε ) = ∑ ℓ = 1 N ( ∑ i = 1 N ∂ x i ∂ x ℓ v ε e i ) ⊗ ( ∑ j = 1 N ∂ x j ∂ x ℓ v ε e j ) = ∑ i , j = 1 N ( ∑ ℓ = 1 N ( ∂ x i ∂ x ℓ v ε ) ( ∂ x j ∂ x ℓ v ε ) ) e i ⊗ e j = ∑ i , j = 1 N ( ∑ ℓ = 1 N ( ∂ x i ∂ x ℓ v ε ) ( ∂ x ℓ ∂ x j v ε ) ) e i ⊗ e j = ( ∇ 2 v ε ) 2 . \begin{aligned}
\sum_{\ell=1}^{N}\nabla(\partial_{x_\ell}v_\varepsilon)\otimes\nabla(\partial_{x_\ell}v_\varepsilon)
&=\sum_{\ell=1}^{N}\left(\sum_{i=1}^{N}\partial_{x_i}\partial_{x_\ell}v_\varepsilon e_i\right)\otimes\left(\sum_{j=1}^{N}\partial_{x_j}\partial_{x_\ell}v_\varepsilon e_j\right)\\
&=\sum_{i,j=1}^{N}\left(\sum_{\ell=1}^{N}(\partial_{x_i}\partial_{x_\ell}v_\varepsilon)(\partial_{x_j}\partial_{x_\ell}v_\varepsilon)\right)e_i\otimes e_j\\
&=\sum_{i,j=1}^{N}\left(\sum_{\ell=1}^{N}(\partial_{x_i}\partial_{x_\ell}v_\varepsilon)(\partial_{x_\ell}\partial_{x_j}v_\varepsilon)\right)e_i\otimes e_j\\
&=(\nabla^2v_\varepsilon)^2.
\end{aligned} ℓ = 1 ∑ N ∇ ( ∂ x ℓ v ε ) ⊗ ∇ ( ∂ x ℓ v ε ) = ℓ = 1 ∑ N ( i = 1 ∑ N ∂ x i ∂ x ℓ v ε e i ) ⊗ ( j = 1 ∑ N ∂ x j ∂ x ℓ v ε e j ) = i , j = 1 ∑ N ( ℓ = 1 ∑ N ( ∂ x i ∂ x ℓ v ε ) ( ∂ x j ∂ x ℓ v ε ) ) e i ⊗ e j = i , j = 1 ∑ N ( ℓ = 1 ∑ N ( ∂ x i ∂ x ℓ v ε ) ( ∂ x ℓ ∂ x j v ε ) ) e i ⊗ e j = ( ∇ 2 v ε ) 2 .
In the penultimate step, we used the symmetry of ∇ 2 v ε \nabla^2v_\varepsilon ∇ 2 v ε . Here ( ∇ 2 v ε ) 2 (\nabla^2v_\varepsilon)^2 ( ∇ 2 v ε ) 2 denotes the matrix product of the Hessian with itself. Rearranging the identity for ∇ 2 w ε \nabla^2w_\varepsilon ∇ 2 w ε , we obtain
2 ∑ ℓ = 1 N ( ∂ x ℓ v ε ) ∇ 2 ( ∂ x ℓ v ε ) = ∇ 2 w ε − 2 ( ∇ 2 v ε ) 2 . 2\sum_{\ell=1}^{N}(\partial_{x_\ell}v_\varepsilon)\nabla^2(\partial_{x_\ell}v_\varepsilon)=\nabla^2w_\varepsilon-2(\nabla^2v_\varepsilon)^2. 2 ℓ = 1 ∑ N ( ∂ x ℓ v ε ) ∇ 2 ( ∂ x ℓ v ε ) = ∇ 2 w ε − 2 ( ∇ 2 v ε ) 2 .
Taking the Frobenius inner product with A ε A_\varepsilon A ε and using its linearity gives
2 ∑ ℓ = 1 N ( ∂ x ℓ v ε ) A ε ∙ ∇ 2 ( ∂ x ℓ v ε ) = A ε ∙ ∇ 2 w ε − 2 A ε ∙ ( ∇ 2 v ε ) 2 = A ε ∙ ∇ 2 w ε − 2 tr ( A ε ( ∇ 2 v ε ) 2 ) . \begin{aligned}
2\sum_{\ell=1}^{N}(\partial_{x_\ell}v_\varepsilon)A_\varepsilon\mathbin{\bullet}\nabla^2(\partial_{x_\ell}v_\varepsilon)
&=A_\varepsilon\mathbin{\bullet}\nabla^2w_\varepsilon-2A_\varepsilon\mathbin{\bullet}(\nabla^2v_\varepsilon)^2\\
&=A_\varepsilon\mathbin{\bullet}\nabla^2w_\varepsilon-2\operatorname{tr}\left(A_\varepsilon(\nabla^2v_\varepsilon)^2\right).
\end{aligned} 2 ℓ = 1 ∑ N ( ∂ x ℓ v ε ) A ε ∙ ∇ 2 ( ∂ x ℓ v ε ) = A ε ∙ ∇ 2 w ε − 2 A ε ∙ ( ∇ 2 v ε ) 2 = A ε ∙ ∇ 2 w ε − 2 tr ( A ε ( ∇ 2 v ε ) 2 ) .
In the last step, we used the definition of the Frobenius inner product and the symmetry of A ε A_\varepsilon A ε . Substituting the preceding identity into the equation for ∂ t w ε \partial_tw_\varepsilon ∂ t w ε and subtracting A ε ∙ ∇ 2 w ε A_\varepsilon\mathbin{\bullet}\nabla^2w_\varepsilon A ε ∙ ∇ 2 w ε and B ε ⋅ ∇ w ε B_\varepsilon\cdot\nabla w_\varepsilon B ε ⋅ ∇ w ε from both sides gives
∂ t w ε − A ε ∙ ∇ 2 w ε − B ε ⋅ ∇ w ε ⏟ = L ε w ε = − 2 tr ( A ε ( ∇ 2 v ε ) 2 ) + 2 w ε ( ∂ v A ε ) ∙ ∇ 2 v ε + 2 w ε ∂ v F ε . \begin{aligned}
\underbrace{\partial_tw_\varepsilon-A_\varepsilon\mathbin{\bullet}\nabla^2w_\varepsilon-B_\varepsilon\cdot\nabla w_\varepsilon}_{=\mathcal{L}_\varepsilon w_\varepsilon}
&=-2\operatorname{tr}\left(A_\varepsilon(\nabla^2v_\varepsilon)^2\right)\\
&\quad+2w_\varepsilon(\partial_vA_\varepsilon)\mathbin{\bullet}\nabla^2v_\varepsilon+2w_\varepsilon\partial_vF_\varepsilon.
\end{aligned} = L ε w ε ∂ t w ε − A ε ∙ ∇ 2 w ε − B ε ⋅ ∇ w ε = − 2 tr ( A ε ( ∇ 2 v ε ) 2 ) + 2 w ε ( ∂ v A ε ) ∙ ∇ 2 v ε + 2 w ε ∂ v F ε .
Here A ε A_\varepsilon A ε , ∂ v A ε \partial_vA_\varepsilon ∂ v A ε , and ∂ v F ε \partial_vF_\varepsilon ∂ v F ε are evaluated at ( v ε , ∇ v ε ) (v_\varepsilon,\nabla v_\varepsilon) ( v ε , ∇ v ε ) , while B ε B_\varepsilon B ε is the drift defined in (41) . All quantities are evaluated at the same space-time point. The partial derivatives with respect to v v v are taken with the gradient argument fixed. We now substitute
A ε ( v ε , ∇ v ε ) = a ( v ε ) w ε p − 2 2 C θ ε ( ∇ v ε ) , A_\varepsilon(v_\varepsilon,\nabla v_\varepsilon)=a(v_\varepsilon)w_\varepsilon^{\frac{p-2}{2}}C_{\theta_\varepsilon}(\nabla v_\varepsilon), A ε ( v ε , ∇ v ε ) = a ( v ε ) w ε 2 p − 2 C θ ε ( ∇ v ε ) ,
together with the expressions for ∂ v A ε \partial_vA_\varepsilon ∂ v A ε and ∂ v F ε \partial_vF_\varepsilon ∂ v F ε from (4.8) and (4.9)`. Combining the powers of w ε w_\varepsilon w ε , we obtain
L ε w ε = − 2 a ( v ε ) w ε p − 2 2 tr ( C θ ε ( ∇ v ε ) ( ∇ 2 v ε ) 2 ) + 2 a ( v ε ) v ε w ε p 2 ( D θ ε ( ∇ v ε ) ∙ ∇ 2 v ε ) + 2 a ( v ε ) ( k − 1 ) v ε 2 w ε p + 2 2 J ( θ ε ) . (43) \begin{aligned}
\mathcal{L}_\varepsilon w_\varepsilon
={}&-2a(v_\varepsilon)w_\varepsilon^{\frac{p-2}{2}}\operatorname{tr}\left(C_{\theta_\varepsilon}(\nabla v_\varepsilon)(\nabla^2v_\varepsilon)^2\right)\\
&+2\frac{a(v_\varepsilon)}{v_\varepsilon}w_\varepsilon^{\frac{p}{2}}\left(D_{\theta_\varepsilon}(\nabla v_\varepsilon)\mathbin{\bullet}\nabla^2v_\varepsilon\right)\\
&+2\frac{a(v_\varepsilon)(k-1)}{v_\varepsilon^2}w_\varepsilon^{\frac{p+2}{2}}J(\theta_\varepsilon). \tag*{(43)}
\end{aligned} L ε w ε = − 2 a ( v ε ) w ε 2 p − 2 tr ( C θ ε ( ∇ v ε ) ( ∇ 2 v ε ) 2 ) + 2 v ε a ( v ε ) w ε 2 p ( D θ ε ( ∇ v ε ) ∙ ∇ 2 v ε ) + 2 v ε 2 a ( v ε ) ( k − 1 ) w ε 2 p + 2 J ( θ ε ) . (43)
For ε > 0 \varepsilon>0 ε > 0 , A ε A_\varepsilon A ε and F ε F_\varepsilon F ε are smooth in ( v , ξ ) (v,\xi) ( v , ξ ) even at ξ = 0 \xi=0 ξ = 0 , so the differentiated equation holds at every point (with u ε ≥ m > 0 u_\varepsilon\ge m>0 u ε ≥ m > 0 ), not only where ∇ v ε ≠ 0 \nabla v_\varepsilon\ne0 ∇ v ε = 0 ; only the representation through C θ , D θ , J C_\theta,D_\theta,J C θ , D θ , J requires w ε > 0 w_\varepsilon>0 w ε > 0 . This is the difference from the previous section, where the coefficients degenerate or blow up at ∇ v = 0 \nabla v=0 ∇ v = 0 .
On the other hand, when ε = 0 \varepsilon=0 ε = 0 , we have θ = 1 \theta=1 θ = 1 ; in this case C 1 = P C_1=P C 1 = P , D 1 = ( 2 − p ) P D_1=(2-p)P D 1 = ( 2 − p ) P , and J ( 1 ) = − ( p − 1 ) 2 J(1)=-(p-1)^2 J ( 1 ) = − ( p − 1 ) 2 . Since a ′ = ( 2 − p ) a / v a'=(2-p)a/v a ′ = ( 2 − p ) a / v and b ′ = − c ( p − 1 ) 2 ( k − 1 ) v − p b'=-c(p-1)^2(k-1)v^{-p} b ′ = − c ( p − 1 ) 2 ( k − 1 ) v − p , the middle term of (43) becomes 2 a ′ v p 2 P ∙ ∇ 2 v 2a'v^{\frac{p}{2}}\mathbf{P}\mathbin{\bullet}\nabla^2v 2 a ′ v 2 p P ∙ ∇ 2 v and the last one 2 b ′ v p + 2 2 2b'v^{\frac{p+2}{2}} 2 b ′ v 2 p + 2 . Therefore, (43) reduces to (20) .
Estimating the term linear in the Hessian
We next estimate the second term on the right-hand side of (43) , which is linear in ∇ 2 v ε \nabla^{2}v_{\varepsilon} ∇ 2 v ε and has no prescribed sign. Using the Cauchy–Schwarz and Young inequalities, we absorb its Hessian-dependent contribution into the negative first term, retaining half of that term for the localization argument. Recall that, for θ ∈ ( 0 , 1 ] \theta\in(0,1] θ ∈ ( 0 , 1 ] and ξ ≠ 0 \xi\ne0 ξ = 0 ,
C θ ( ξ ) = θ − p − 2 2 ( I + ( p − 2 ) θ ξ ⊗ ξ ∣ ξ ∣ 2 ) . \mathbf{C}_{\theta}(\xi)=\theta^{-\frac{p-2}{2}}\left(I+(p-2)\theta\frac{\xi\otimes\xi}{|\xi|^{2}}\right). C θ ( ξ ) = θ − 2 p − 2 ( I + ( p − 2 ) θ ∣ ξ ∣ 2 ξ ⊗ ξ ) .
If ξ ⊥ ξ \xi\perp\xi ξ ⊥ ξ , then C θ ( ξ ) ξ = θ − p − 2 2 ξ \mathbf{C}_{\theta}(\xi)\xi=\theta^{-\frac{p-2}{2}}\xi C θ ( ξ ) ξ = θ − 2 p − 2 ξ , whereas
C θ ( ξ ) ξ ∣ ξ ∣ = θ − p − 2 2 ( 1 + ( p − 2 ) θ ) ξ ∣ ξ ∣ . \mathbf{C}_{\theta}(\xi)\frac{\xi}{|\xi|}=\theta^{-\frac{p-2}{2}}\left(1+(p-2)\theta\right)\frac{\xi}{|\xi|}. C θ ( ξ ) ∣ ξ ∣ ξ = θ − 2 p − 2 ( 1 + ( p − 2 ) θ ) ∣ ξ ∣ ξ .
Thus, its eigenvalues are θ − p − 2 2 \theta^{-\frac{p-2}{2}} θ − 2 p − 2 in directions orthogonal to ξ \xi ξ and θ − p − 2 2 ( 1 + ( p − 2 ) θ ) \theta^{-\frac{p-2}{2}}\left(1+(p-2)\theta\right) θ − 2 p − 2 ( 1 + ( p − 2 ) θ ) in the direction of ξ \xi ξ . Since
1 + ( p − 2 ) θ ≥ min { 1 , p − 1 } > 0 , 1+(p-2)\theta\ge\min\{1,p-1\}>0, 1 + ( p − 2 ) θ ≥ min { 1 , p − 1 } > 0 ,
the symmetric matrix C θ ( ξ ) \mathbf{C}_{\theta}(\xi) C θ ( ξ ) is positive definite and hence invertible. Its positive definite square root C θ ( ξ ) 1 2 \mathbf{C}_{\theta}(\xi)^{\frac{1}{2}} C θ ( ξ ) 2 1 and the inverse C θ ( ξ ) − 1 2 \mathbf{C}_{\theta}(\xi)^{-\frac{1}{2}} C θ ( ξ ) − 2 1 are therefore well defined. For every symmetric matrix H ∈ R N × N H\in\mathbb{R}^{N\times N} H ∈ R N × N , inserting these factors and applying the Cauchy–Schwarz inequality for the Frobenius inner product gives
( D θ ( ξ ) ⋅ H ) 2 = [ ( C θ ( ξ ) − 1 2 D θ ( ξ ) ) ⋅ ( C θ ( ξ ) 1 2 H ) ] 2 ≤ ∣ C θ ( ξ ) − 1 2 D θ ( ξ ) ∣ 2 ∣ C θ ( ξ ) 1 2 H ∣ 2 = tr ( C θ ( ξ ) − 1 D θ ( ξ ) 2 ) tr ( C θ ( ξ ) H 2 ) . (44) \begin{aligned}
\left(\mathbf{D}_{\theta}(\xi)\cdot H\right)^{2}
=\left[\left(\mathbf{C}_{\theta}(\xi)^{-\frac{1}{2}}\mathbf{D}_{\theta}(\xi)\right)\cdot\left(\mathbf{C}_{\theta}(\xi)^{\frac{1}{2}}H\right)\right]^{2} \\
&\le\left|\mathbf{C}_{\theta}(\xi)^{-\frac{1}{2}}\mathbf{D}_{\theta}(\xi)\right|^{2}\left|\mathbf{C}_{\theta}(\xi)^{\frac{1}{2}}H\right|^{2} \\
&=\operatorname{tr}\left(\mathbf{C}_{\theta}(\xi)^{-1}\mathbf{D}_{\theta}(\xi)^{2}\right)\operatorname{tr}\left(\mathbf{C}_{\theta}(\xi)H^{2}\right). \tag*{(44)}
\end{aligned} ( D θ ( ξ ) ⋅ H ) 2 = [ ( C θ ( ξ ) − 2 1 D θ ( ξ ) ) ⋅ ( C θ ( ξ ) 2 1 H ) ] 2 ≤ C θ ( ξ ) − 2 1 D θ ( ξ ) 2 C θ ( ξ ) 2 1 H 2 = tr ( C θ ( ξ ) − 1 D θ ( ξ ) 2 ) tr ( C θ ( ξ ) H 2 ) . (44)
To justify the last equality, we use the definition of the Frobenius norm, the symmetry of the matrices, and the cyclic property of the trace.1 For the first factor, these properties give
∣ C θ ( ξ ) − 1 2 D θ ( ξ ) ∣ 2 = tr ( ( C θ ( ξ ) − 1 2 D θ ( ξ ) ) T C θ ( ξ ) − 1 2 D θ ( ξ ) ) = tr ( D θ ( ξ ) C θ ( ξ ) − 1 D θ ( ξ ) ) = tr ( C θ ( ξ ) − 1 D θ ( ξ ) 2 ) . \begin{aligned}
\left|\mathbf{C}_{\theta}(\xi)^{-\frac{1}{2}}\mathbf{D}_{\theta}(\xi)\right|^{2}
=\operatorname{tr}\left(\left(\mathbf{C}_{\theta}(\xi)^{-\frac{1}{2}}\mathbf{D}_{\theta}(\xi)\right)^{\mathsf{T}}\mathbf{C}_{\theta}(\xi)^{-\frac{1}{2}}\mathbf{D}_{\theta}(\xi)\right) \\
&=\operatorname{tr}\left(\mathbf{D}_{\theta}(\xi)\mathbf{C}_{\theta}(\xi)^{-1}\mathbf{D}_{\theta}(\xi)\right) \\
&=\operatorname{tr}\left(\mathbf{C}_{\theta}(\xi)^{-1}\mathbf{D}_{\theta}(\xi)^{2}\right).
\end{aligned} C θ ( ξ ) − 2 1 D θ ( ξ ) 2 = tr ( ( C θ ( ξ ) − 2 1 D θ ( ξ ) ) T C θ ( ξ ) − 2 1 D θ ( ξ ) ) = tr ( D θ ( ξ ) C θ ( ξ ) − 1 D θ ( ξ ) ) = tr ( C θ ( ξ ) − 1 D θ ( ξ ) 2 ) .
Similarly, for the second factor,
∣ C θ ( ξ ) 1 2 H ∣ 2 = tr ( ( C θ ( ξ ) 1 2 H ) T C θ ( ξ ) 1 2 H ) = tr ( H C θ ( ξ ) H ) = tr ( C θ ( ξ ) H 2 ) . \begin{aligned}
\left|\mathbf{C}_{\theta}(\xi)^{\frac{1}{2}}H\right|^{2}
=\operatorname{tr}\left(\left(\mathbf{C}_{\theta}(\xi)^{\frac{1}{2}}H\right)^{\mathsf{T}}\mathbf{C}_{\theta}(\xi)^{\frac{1}{2}}H\right) \\
&=\operatorname{tr}\left(H\mathbf{C}_{\theta}(\xi)H\right) \\
&=\operatorname{tr}\left(\mathbf{C}_{\theta}(\xi)H^{2}\right).
\end{aligned} C θ ( ξ ) 2 1 H 2 = tr ( ( C θ ( ξ ) 2 1 H ) T C θ ( ξ ) 2 1 H ) = tr ( H C θ ( ξ ) H ) = tr ( C θ ( ξ ) H 2 ) .
Applying (44) with θ = θ ε \theta=\theta_{\varepsilon} θ = θ ε , ξ = ∇ v ε \xi=\nabla v_{\varepsilon} ξ = ∇ v ε , and H = ∇ 2 v ε H=\nabla^{2}v_{\varepsilon} H = ∇ 2 v ε , followed by Young’s inequality, yields
2 a ( v ε ) v ε w ε p 2 ( D θ ε ( ∇ v ε ) ⋅ ∇ 2 v ε ) 2\frac{a(v_{\varepsilon})}{v_{\varepsilon}}w_{\varepsilon}^{\frac{p}{2}}\left(\mathbf{D}_{\theta_{\varepsilon}}(\nabla v_{\varepsilon})\cdot\nabla^{2}v_{\varepsilon}\right) 2 v ε a ( v ε ) w ε 2 p ( D θ ε ( ∇ v ε ) ⋅ ∇ 2 v ε )
≤ 2 a ( v ε ) v ε w ε p 2 tr ( C θ ε ( ∇ v ε ) − 1 D θ ε ( ∇ v ε ) 2 ) tr ( C θ ε ( ∇ v ε ) ( ∇ 2 v ε ) 2 ) ≤ a ( v ε ) w ε p − 2 2 tr ( C θ ε ( ∇ v ε ) ( ∇ 2 v ε ) 2 ) + a ( v ε ) v ε 2 w ε p + 2 2 tr ( C θ ε ( ∇ v ε ) − 1 D θ ε ( ∇ v ε ) 2 ) . \begin{aligned}
\le2\frac{a(v_{\varepsilon})}{v_{\varepsilon}}w_{\varepsilon}^{\frac{p}{2}}\sqrt{\operatorname{tr}\left(\mathbf{C}_{\theta_{\varepsilon}}(\nabla v_{\varepsilon})^{-1}\mathbf{D}_{\theta_{\varepsilon}}(\nabla v_{\varepsilon})^{2}\right)}\sqrt{\operatorname{tr}\left(\mathbf{C}_{\theta_{\varepsilon}}(\nabla v_{\varepsilon})(\nabla^{2}v_{\varepsilon})^{2}\right)} \\
&\le a(v_{\varepsilon})w_{\varepsilon}^{\frac{p-2}{2}}\operatorname{tr}\left(\mathbf{C}_{\theta_{\varepsilon}}(\nabla v_{\varepsilon})(\nabla^{2}v_{\varepsilon})^{2}\right)+\frac{a(v_{\varepsilon})}{v_{\varepsilon}^{2}}w_{\varepsilon}^{\frac{p+2}{2}}\operatorname{tr}\left(\mathbf{C}_{\theta_{\varepsilon}}(\nabla v_{\varepsilon})^{-1}\mathbf{D}_{\theta_{\varepsilon}}(\nabla v_{\varepsilon})^{2}\right).
\end{aligned} ≤ 2 v ε a ( v ε ) w ε 2 p tr ( C θ ε ( ∇ v ε ) − 1 D θ ε ( ∇ v ε ) 2 ) tr ( C θ ε ( ∇ v ε ) ( ∇ 2 v ε ) 2 ) ≤ a ( v ε ) w ε 2 p − 2 tr ( C θ ε ( ∇ v ε ) ( ∇ 2 v ε ) 2 ) + v ε 2 a ( v ε ) w ε 2 p + 2 tr ( C θ ε ( ∇ v ε ) − 1 D θ ε ( ∇ v ε ) 2 ) .
Consequently,
L ε w ε ≤ − a ( v ε ) w ε p − 2 2 tr ( C θ ε ( ∇ v ε ) ( ∇ 2 v ε ) 2 ) + a ( v ε ) v ε 2 w ε p + 2 2 [ tr ( ( C θ ε ( ∇ v ε ) ) − 1 D θ ε ( ∇ v ε ) 2 ) + 2 ( k − 1 ) J ( θ ε ) ] . (45) \mathcal{L}_{\varepsilon}w_{\varepsilon}\leq-a(v_{\varepsilon})w_{\varepsilon}^{\frac{p-2}{2}}\operatorname{tr}\left(\mathbf{C}_{\theta_{\varepsilon}}(\nabla v_{\varepsilon})(\nabla^{2}v_{\varepsilon})^{2}\right)
+\frac{a(v_{\varepsilon})}{v_{\varepsilon}^{2}}w_{\varepsilon}^{\frac{p+2}{2}}\left[\operatorname{tr}\left((\mathbf{C}_{\theta_{\varepsilon}}(\nabla v_{\varepsilon}))^{-1}\mathbf{D}_{\theta_{\varepsilon}}(\nabla v_{\varepsilon})^{2}\right)+2(k-1)J(\theta_{\varepsilon})\right]. \tag*{(45)} L ε w ε ≤ − a ( v ε ) w ε 2 p − 2 tr ( C θ ε ( ∇ v ε ) ( ∇ 2 v ε ) 2 ) + v ε 2 a ( v ε ) w ε 2 p + 2 [ tr ( ( C θ ε ( ∇ v ε ) ) − 1 D θ ε ( ∇ v ε ) 2 ) + 2 ( k − 1 ) J ( θ ε ) ] . (45)
A coercive inequality for large gradients
For each ξ ≠ 0 \xi\ne0 ξ = 0 , the matrix C θ ( ξ ) \mathbf{C}_{\theta}(\xi) C θ ( ξ ) is a linear combination of I I I and ξ ⊗ ξ ∣ ξ ∣ 2 \frac{\xi\otimes\xi}{|\xi|^{2}} ∣ ξ ∣ 2 ξ ⊗ ξ . By (36) , its derivative ∂ θ C θ ( ξ ) \partial_{\theta}\mathbf{C}_{\theta}(\xi) ∂ θ C θ ( ξ ) has the same structure. Hence, by (38) , so does D θ ( ξ ) \mathbf{D}_{\theta}(\xi) D θ ( ξ ) . Their eigenvalues depend only on θ \theta θ , p p p , k k k , and not on the direction of ξ \xi ξ . Since C θ ( ξ ) \mathbf{C}_{\theta}(\xi) C θ ( ξ ) and D θ ( ξ ) \mathbf{D}_{\theta}(\xi) D θ ( ξ ) have a common orthonormal eigenbasis, the trace of C θ ( ξ ) − 1 D θ ( ξ ) 2 \mathbf{C}_{\theta}(\xi)^{-1}\mathbf{D}_{\theta}(\xi)^{2} C θ ( ξ ) − 1 D θ ( ξ ) 2 is the sum of the squared eigenvalues of D θ ( ξ ) \mathbf{D}_{\theta}(\xi) D θ ( ξ ) divided by the corresponding eigenvalues of C θ ( ξ ) \mathbf{C}_{\theta}(\xi) C θ ( ξ ) . Consequently,
tr ( C θ ( ξ ) − 1 D θ ( ξ ) 2 ) + 2 ( k − 1 ) J ( θ ) (46) \operatorname{tr}\left(\mathbf{C}_{\theta}(\xi)^{-1}\mathbf{D}_{\theta}(\xi)^{2}\right)+2(k-1)J(\theta) \tag*{(46)} tr ( C θ ( ξ ) − 1 D θ ( ξ ) 2 ) + 2 ( k − 1 ) J ( θ ) (46)
depends only on θ \theta θ , N N N , p p p , k k k . The dependence on N N N arises from the multiplicity N − 1 N-1 N − 1 of the eigenvalues corresponding to directions orthogonal to ξ \xi ξ .
For each fixed ξ ≠ 0 \xi\ne0 ξ = 0 , the functions C θ ( ξ ) \mathbf{C}_{\theta}(\xi) C θ ( ξ ) , D θ ( ξ ) \mathbf{D}_{\theta}(\xi) D θ ( ξ ) , j ( θ ) j(\theta) j ( θ ) , and J ( θ ) J(\theta) J ( θ ) are continuous at θ = 1 \theta=1 θ = 1 . At this endpoint, we have
C 1 ( ξ ) = I + ( p − 2 ) ξ ∣ ξ ∣ ⊗ ξ ∣ ξ ∣ , D 1 ( ξ ) = ( 2 − p ) C 1 ( ξ ) , \mathbf{C}_{1}(\xi)=I+(p-2)\frac{\xi}{|\xi|}\otimes\frac{\xi}{|\xi|},\qquad\mathbf{D}_{1}(\xi)=(2-p)\mathbf{C}_{1}(\xi), C 1 ( ξ ) = I + ( p − 2 ) ∣ ξ ∣ ξ ⊗ ∣ ξ ∣ ξ , D 1 ( ξ ) = ( 2 − p ) C 1 ( ξ ) ,
and
j ( 1 ) = p − 1 , J ( 1 ) = − ( p − 1 ) 2 . j(1)=p-1,\qquad J(1)=-(p-1)^{2}. j ( 1 ) = p − 1 , J ( 1 ) = − ( p − 1 ) 2 .
Since D 1 ( ξ ) = ( 2 − p ) C 1 ( ξ ) \mathbf{D}_{1}(\xi)=(2-p)\mathbf{C}_{1}(\xi) D 1 ( ξ ) = ( 2 − p ) C 1 ( ξ ) , we obtain
tr ( C 1 ( ξ ) − 1 D 1 ( ξ ) 2 ) = ( 2 − p ) 2 tr ( C 1 ( ξ ) − 1 C 1 ( ξ ) 2 ) = ( p − 2 ) 2 tr C 1 ( ξ ) = ( p − 2 ) 2 ( ( N − 1 ) + ( p − 1 ) ) = ( N + p − 2 ) ( p − 2 ) 2 = A ∗ . \begin{aligned}
\operatorname{tr}\left(\mathbf{C}_{1}(\xi)^{-1}\mathbf{D}_{1}(\xi)^{2}\right)
&=(2-p)^{2}\operatorname{tr}\left(\mathbf{C}_{1}(\xi)^{-1}\mathbf{C}_{1}(\xi)^{2}\right)\\
&=(p-2)^{2}\operatorname{tr}\mathbf{C}_{1}(\xi)\\
&=(p-2)^{2}\left((N-1)+(p-1)\right)\\
&=(N+p-2)(p-2)^{2}=A_{*}.
\end{aligned} tr ( C 1 ( ξ ) − 1 D 1 ( ξ ) 2 ) = ( 2 − p ) 2 tr ( C 1 ( ξ ) − 1 C 1 ( ξ ) 2 ) = ( p − 2 ) 2 tr C 1 ( ξ ) = ( p − 2 ) 2 ( ( N − 1 ) + ( p − 1 ) ) = ( N + p − 2 ) ( p − 2 ) 2 = A ∗ .
Here we used that the trace of C 1 ( ξ ) \mathbf{C}_{1}(\xi) C 1 ( ξ ) is the sum of its eigenvalues, namely p − 1 p-1 p − 1 in the direction of ξ \xi ξ and 1 1 1 in each of the N − 1 N-1 N − 1 orthogonal directions. Using J ( 1 ) = − ( p − 1 ) 2 J(1)=-(p-1)^{2} J ( 1 ) = − ( p − 1 ) 2 and the preceding trace identity, we obtain
tr ( C 1 ( ξ ) − 1 D 1 ( ξ ) 2 ) + 2 ( k − 1 ) J ( 1 ) = A ∗ − 2 ( k − 1 ) ( p − 1 ) 2 ≤ A ∗ − 2 ( 1 + A ∗ ( p − 1 ) 2 ) ( p − 1 ) 2 = − A ∗ − 2 ( p − 1 ) 2 < 0. (47) \begin{aligned}
\operatorname{tr}\left(\mathbf{C}_{1}(\xi)^{-1}\mathbf{D}_{1}(\xi)^{2}\right)+2(k-1)J(1)
&=A_{*}-2(k-1)(p-1)^{2}\\
&\leq A_{*}-2\left(1+\frac{A_{*}}{(p-1)^{2}}\right)(p-1)^{2}\\
&=-A_{*}-2(p-1)^{2}<0. \tag*{(47)}
\end{aligned} tr ( C 1 ( ξ ) − 1 D 1 ( ξ ) 2 ) + 2 ( k − 1 ) J ( 1 ) = A ∗ − 2 ( k − 1 ) ( p − 1 ) 2 ≤ A ∗ − 2 ( 1 + ( p − 1 ) 2 A ∗ ) ( p − 1 ) 2 = − A ∗ − 2 ( p − 1 ) 2 < 0. (47)
Here we used the choice of k k k in (9) . The last inequality is strict because A ∗ ≥ 0 A_{*}\geq0 A ∗ ≥ 0 and p > 1 p>1 p > 1 . In fact, the stated choice of k k k is not sharp, but this is irrelevant for the argument, since only the existence of some admissible k = k ( N , p ) k=k(N,p) k = k ( N , p ) matters. The expression on the left-hand side of (46) is a continuous function of θ \theta θ depending only on N N N , p p p , k k k ; in particular, it is independent of ξ \xi ξ . By (47) , its value at θ = 1 \theta=1 θ = 1 is at most − [ A ∗ + 2 ( p − 1 ) 2 ] -[A_{*}+2(p-1)^{2}] − [ A ∗ + 2 ( p − 1 ) 2 ] . Continuity therefore allows us to choose θ o = θ o ( N , p , k ) ∈ ( 0 , 1 ) \theta_{o}=\theta_{o}(N,p,k)\in(0,1) θ o = θ o ( N , p , k ) ∈ ( 0 , 1 ) such that
tr ( C θ ( ξ ) − 1 D θ ( ξ ) 2 ) + 2 ( k − 1 ) J ( θ ) ≤ − 1 2 [ A ∗ + 2 ( p − 1 ) 2 ] ∀ θ ∈ [ θ o , 1 ] , ξ ≠ 0. \operatorname{tr}\left(\mathbf{C}_{\theta}(\xi)^{-1}\mathbf{D}_{\theta}(\xi)^{2}\right)+2(k-1)J(\theta)\leq-\frac{1}{2}\left[A_{*}+2(p-1)^{2}\right]\qquad\forall\ \theta\in[\theta_{o},1],\ \xi\ne0. tr ( C θ ( ξ ) − 1 D θ ( ξ ) 2 ) + 2 ( k − 1 ) J ( θ ) ≤ − 2 1 [ A ∗ + 2 ( p − 1 ) 2 ] ∀ θ ∈ [ θ o , 1 ] , ξ = 0.
The choice of θ o \theta_{o} θ o is independent of ε \varepsilon ε and the lower bound m \mathbf{m} m . We next obtain a uniform positive lower bound for the eigenvalues of C θ ( ξ ) \mathbf{C}_{\theta}(\xi) C θ ( ξ ) on [ θ o , 1 ] [\theta_{o},1] [ θ o , 1 ] . Recall that these eigenvalues are θ p − 2 2 \theta^{\frac{p-2}{2}} θ 2 p − 2 in directions orthogonal to ξ \xi ξ , and θ p − 2 2 ( 1 + ( p − 2 ) θ ) \theta^{\frac{p-2}{2}}\left(1+(p-2)\theta\right) θ 2 p − 2 ( 1 + ( p − 2 ) θ ) in the direction of ξ \xi ξ . For θ ∈ [ θ o , 1 ] \theta\in[\theta_o,1] θ ∈ [ θ o , 1 ] , we have
θ − p − 2 2 ≥ min { 1 , θ o 2 − p 2 } , 1 + ( p − 2 ) θ ≥ min { 1 , p − 1 } . \theta^{-\frac{p-2}{2}}\ge\min\left\{1,\theta_o^{\frac{2-p}{2}}\right\},\qquad1+(p-2)\theta\ge\min\{1,p-1\}. θ − 2 p − 2 ≥ min { 1 , θ o 2 2 − p } , 1 + ( p − 2 ) θ ≥ min { 1 , p − 1 } .
For the eigenvalues in directions orthogonal to ξ \xi ξ , it follows that
θ − p − 2 2 ≥ min { 1 , θ o 2 − p 2 } ≥ min { 1 , p − 1 } min { 1 , θ o 2 − p 2 } , \theta^{-\frac{p-2}{2}}\ge\min\left\{1,\theta_o^{\frac{2-p}{2}}\right\}\ge\min\{1,p-1\}\min\left\{1,\theta_o^{\frac{2-p}{2}}\right\}, θ − 2 p − 2 ≥ min { 1 , θ o 2 2 − p } ≥ min { 1 , p − 1 } min { 1 , θ o 2 2 − p } ,
where the last inequality uses 0 < min { 1 , p − 1 } ≤ 1 0<\min\{1,p-1\}\le1 0 < min { 1 , p − 1 } ≤ 1 . For the eigenvalue in the direction of ξ \xi ξ , multiplying the two positive lower bounds gives
θ − p − 2 2 ( 1 + ( p − 2 ) θ ) ≥ min { 1 , p − 1 } min { 1 , θ o 2 − p 2 } . \theta^{-\frac{p-2}{2}}\left(1+(p-2)\theta\right)\ge\min\{1,p-1\}\min\left\{1,\theta_o^{\frac{2-p}{2}}\right\}. θ − 2 p − 2 ( 1 + ( p − 2 ) θ ) ≥ min { 1 , p − 1 } min { 1 , θ o 2 2 − p } .
Thus, both eigenvalues are bounded below by
c 1 : = min { 1 , p − 1 } min { 1 , θ o 2 − p 2 } > 0. c_1:=\min\{1,p-1\}\min\left\{1,\theta_o^{\frac{2-p}{2}}\right\}>0. c 1 := min { 1 , p − 1 } min { 1 , θ o 2 2 − p } > 0.
Since θ o \theta_o θ o depends only on N , p , k N,p,k N , p , k , the same is true of c 1 c_1 c 1 . In particular, c 1 c_1 c 1 is independent of ε \varepsilon ε , m \mathbf{m} m , and ξ \xi ξ . Consequently, for every ζ ∈ R N \zeta\in\mathbb{R}^N ζ ∈ R N ,
ζ ⋅ C θ ( ξ ) ζ ≥ c 1 ∣ ζ ∣ 2 for θ ∈ [ θ o , 1 ] , ξ ≠ 0. \zeta\cdot C_\theta(\xi)\zeta\ge c_1|\zeta|^2\qquad\text{for }\theta\in[\theta_o,1],\ \xi\ne0. ζ ⋅ C θ ( ξ ) ζ ≥ c 1 ∣ ζ ∣ 2 for θ ∈ [ θ o , 1 ] , ξ = 0.
Suppose that θ ε ≥ θ o \theta_\varepsilon\ge\theta_o θ ε ≥ θ o . The j j j -th column of ∇ 2 v ε \nabla^2v_\varepsilon ∇ 2 v ε is ∇ ( ∂ x j v ε ) \nabla(\partial_{x_j}v_\varepsilon) ∇ ( ∂ x j v ε ) . Applying the preceding inequality to this vector gives
∇ ( ∂ x j v ε ) ⋅ C θ ε ( ∇ v ε ) ∇ ( ∂ x j v ε ) ≥ c 1 ∣ ∇ ( ∂ x j v ε ) ∣ 2 . \nabla(\partial_{x_j}v_\varepsilon)\cdot C_{\theta_\varepsilon}(\nabla v_\varepsilon)\nabla(\partial_{x_j}v_\varepsilon)\ge c_1|\nabla(\partial_{x_j}v_\varepsilon)|^2. ∇ ( ∂ x j v ε ) ⋅ C θ ε ( ∇ v ε ) ∇ ( ∂ x j v ε ) ≥ c 1 ∣∇ ( ∂ x j v ε ) ∣ 2 .
Using the symmetry of ∇ 2 v ε \nabla^2v_\varepsilon ∇ 2 v ε and the cyclic property of the trace, we obtain
tr ( C θ ε ( ∇ v ε ) ( ∇ 2 v ε ) 2 ) = tr ( ( ∇ 2 v ε ) T C θ ε ( ∇ v ε ) ∇ 2 v ε ) = ∑ j = 1 N ∇ ( ∂ x j v ε ) ⋅ C θ ε ( ∇ v ε ) ∇ ( ∂ x j v ε ) ≥ c 1 ∑ j = 1 N ∣ ∇ ( ∂ x j v ε ) ∣ 2 = c 1 ∣ ∇ 2 v ε ∣ 2 . \begin{aligned}
\operatorname{tr}\left(C_{\theta_\varepsilon}(\nabla v_\varepsilon)(\nabla^2v_\varepsilon)^2\right)
&=\operatorname{tr}\left((\nabla^2v_\varepsilon)^T C_{\theta_\varepsilon}(\nabla v_\varepsilon)\nabla^2v_\varepsilon\right)\\
&=\sum_{j=1}^{N}\nabla(\partial_{x_j}v_\varepsilon)\cdot C_{\theta_\varepsilon}(\nabla v_\varepsilon)\nabla(\partial_{x_j}v_\varepsilon)\\
&\ge c_1\sum_{j=1}^{N}|\nabla(\partial_{x_j}v_\varepsilon)|^2\\
&=c_1|\nabla^2v_\varepsilon|^2.
\end{aligned} tr ( C θ ε ( ∇ v ε ) ( ∇ 2 v ε ) 2 ) = tr ( ( ∇ 2 v ε ) T C θ ε ( ∇ v ε ) ∇ 2 v ε ) = j = 1 ∑ N ∇ ( ∂ x j v ε ) ⋅ C θ ε ( ∇ v ε ) ∇ ( ∂ x j v ε ) ≥ c 1 j = 1 ∑ N ∣∇ ( ∂ x j v ε ) ∣ 2 = c 1 ∣ ∇ 2 v ε ∣ 2 .
We now substitute the lower bound for the Hessian term and the negative upper bound for the coefficient in square brackets into (45) . For θ ε ≥ θ o \theta_\varepsilon\ge\theta_o θ ε ≥ θ o , this gives
L ε w ε ≤ − c 1 a ( v ε ) w ε p − 2 2 ∣ ∇ 2 v ε ∣ 2 − 1 2 [ A ∗ + 2 ( p − 1 ) 2 ] a ( v ε ) v ε 2 w ε p + 2 2 . \mathcal{L}_\varepsilon w_\varepsilon\le-c_1a(v_\varepsilon)w_\varepsilon^{\frac{p-2}{2}}|\nabla^2v_\varepsilon|^2-\frac{1}{2}\left[A_*+2(p-1)^2\right]\frac{a(v_\varepsilon)}{v_\varepsilon^2}w_\varepsilon^{\frac{p+2}{2}}. L ε w ε ≤ − c 1 a ( v ε ) w ε 2 p − 2 ∣ ∇ 2 v ε ∣ 2 − 2 1 [ A ∗ + 2 ( p − 1 ) 2 ] v ε 2 a ( v ε ) w ε 2 p + 2 .
Here the lower bound for the trace gives an upper bound for its negative multiple, since a ( v ε ) w ε p − 2 2 > 0 a(v_\varepsilon)w_\varepsilon^{\frac{p-2}{2}}>0 a ( v ε ) w ε 2 p − 2 > 0 . Recalling a ( v ) = c v 2 − p a(v)=cv^{2-p} a ( v ) = c v 2 − p , we have a ( v ε ) / v ε 2 = c v ε − p a(v_\varepsilon)/v_\varepsilon^2=cv_\varepsilon^{-p} a ( v ε ) / v ε 2 = c v ε − p . Consequently,
L ε w ε ≤ − c 1 a ( v ε ) w ε p − 2 2 ∣ ∇ 2 v ε ∣ 2 − 1 2 c [ A ∗ + 2 ( p − 1 ) 2 ] v ε − p w ε p + 2 2 . \mathcal{L}_\varepsilon w_\varepsilon\le-c_1a(v_\varepsilon)w_\varepsilon^{\frac{p-2}{2}}|\nabla^2v_\varepsilon|^2-\frac{1}{2}c\left[A_*+2(p-1)^2\right]v_\varepsilon^{-p}w_\varepsilon^{\frac{p+2}{2}}. L ε w ε ≤ − c 1 a ( v ε ) w ε 2 p − 2 ∣ ∇ 2 v ε ∣ 2 − 2 1 c [ A ∗ + 2 ( p − 1 ) 2 ] v ε − p w ε 2 p + 2 .
Setting
κ o : = 1 2 c [ A ∗ + 2 ( p − 1 ) 2 ] = k p − 2 2 ( p − 1 ) [ A ∗ + 2 ( p − 1 ) 2 ] > 0 , \kappa_o:=\frac{1}{2}c\left[A_*+2(p-1)^2\right]=\frac{k^{p-2}}{2(p-1)}\left[A_*+2(p-1)^2\right]>0, κ o := 2 1 c [ A ∗ + 2 ( p − 1 ) 2 ] = 2 ( p − 1 ) k p − 2 [ A ∗ + 2 ( p − 1 ) 2 ] > 0 ,
we obtain
L ε w ε ≤ − c 1 a ( v ε ) w ε p − 2 2 ∣ ∇ 2 v ε ∣ 2 − κ o v ε − p w ε p + 2 2 if θ ε ≥ θ o . (48) \mathcal{L}_\varepsilon w_\varepsilon\le-c_1a(v_\varepsilon)w_\varepsilon^{\frac{p-2}{2}}|\nabla^2v_\varepsilon|^2-\kappa_o v_\varepsilon^{-p}w_\varepsilon^{\frac{p+2}{2}}\qquad\text{if }\theta_\varepsilon\ge\theta_o. \tag*{(48)} L ε w ε ≤ − c 1 a ( v ε ) w ε 2 p − 2 ∣ ∇ 2 v ε ∣ 2 − κ o v ε − p w ε 2 p + 2 if θ ε ≥ θ o . (48)
The constants c 1 c_1 c 1 and κ o \kappa_o κ o depend only on N , p , k N,p,k N , p , k and are independent of ε \varepsilon ε and the lower bound m \mathbf{m} m . Recalling the definitions of w ε w_\varepsilon w ε and θ ε \theta_\varepsilon θ ε , the condition θ ε ≥ θ o \theta_\varepsilon\ge\theta_o θ ε ≥ θ o under which (48) holds is equivalent to
w ε = ∣ ∇ v ε ∣ 2 ≥ θ o 1 − θ o ε 2 k 2 v ε 2 ( k − 1 ) . w_\varepsilon=|\nabla v_\varepsilon|^2\ge\frac{\theta_o}{1-\theta_o}\frac{\varepsilon^2}{k^2v_\varepsilon^{2(k-1)}}. w ε = ∣∇ v ε ∣ 2 ≥ 1 − θ o θ o k 2 v ε 2 ( k − 1 ) ε 2 .
Remark 4.1 (The gradient threshold and the lower bound). Since
∣ ∇ u ε ∣ 2 = k 2 v ε 2 ( k − 1 ) w ε , |\nabla u_\varepsilon|^2 = k^2 v_\varepsilon^{2(k-1)}w_\varepsilon, ∣∇ u ε ∣ 2 = k 2 v ε 2 ( k − 1 ) w ε ,
the condition θ ε ≥ θ o \theta_\varepsilon\ge\theta_o θ ε ≥ θ o can equivalently be expressed as
∣ ∇ u ε ∣ ≥ θ o 1 − θ o ε . |\nabla u_\varepsilon| \ge\sqrt{\frac{\theta_o}{1-\theta_o}}\,\varepsilon. ∣∇ u ε ∣ ≥ 1 − θ o θ o ε .
Thus, (48) applies where the magnitude of ∇ u ε \nabla u_\varepsilon ∇ u ε is at least a fixed multiple of ε \varepsilon ε . This multiple depends only on N , p , k N,p,k N , p , k and is independent of the lower bound m .
The lower bound enters when estimating the transformed gradient in the complementary region. Indeed, if θ ε < θ o \theta_\varepsilon< \theta_o θ ε < θ o , then
∣ ∇ v ε ∣ < θ o 1 − θ o ε k v ε k − 1 ≤ θ o 1 − θ o ε k m k − 1 k , |\nabla v_\varepsilon| < \sqrt{\frac{\theta_o}{1-\theta_o}}\frac{\varepsilon}{k v_\varepsilon^{k-1}} \le\sqrt{\frac{\theta_o}{1-\theta_o}}\frac{\varepsilon}{k\mathbf{m}^{\frac{k-1}{k}}}, ∣∇ v ε ∣ < 1 − θ o θ o k v ε k − 1 ε ≤ 1 − θ o θ o k m k k − 1 ε ,
where we used v ε = u ε 1 k ≥ m 1 k v_\varepsilon= u_\varepsilon^{\frac{1}{k}} \ge\mathbf{m}^{\frac{1}{k}} v ε = u ε k 1 ≥ m k 1 . For fixed m > 0 \mathbf{m}>0 m > 0 , this bound tends to zero as ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 .
Bounds for the principal matrix and the drift
We continue to assume θ ε ≥ θ o \theta_\varepsilon\ge\theta_o θ ε ≥ θ o . For θ ∈ [ θ o , 1 ] \theta\in[\theta_o,1] θ ∈ [ θ o , 1 ] , we have
θ − p − 2 2 ≤ max { 1 , θ o − p − 2 2 } , 1 + ( p − 2 ) θ ≤ max { 1 , p − 1 } . \theta^{-\frac{p-2}{2}} \le\max\left\{1,\theta_o^{-\frac{p-2}{2}}\right\}, \qquad1+(p-2)\theta\le\max\{1,p-1\}. θ − 2 p − 2 ≤ max { 1 , θ o − 2 p − 2 } , 1 + ( p − 2 ) θ ≤ max { 1 , p − 1 } .
Thus, both eigenvalues of C θ ( ξ ) C_\theta(\xi) C θ ( ξ ) are bounded above by
max { 1 , p − 1 } max { 1 , θ o − p − 2 2 } , \max\{1,p-1\}\max\left\{1,\theta_o^{-\frac{p-2}{2}}\right\}, max { 1 , p − 1 } max { 1 , θ o − 2 p − 2 } ,
uniformly in θ ∈ [ θ o , 1 ] \theta\in[\theta_o,1] θ ∈ [ θ o , 1 ] and ξ ≠ 0 \xi\ne0 ξ = 0 . Since θ o \theta_o θ o depends only on N , p , k N,p,k N , p , k , this upper bound has the same dependence. Since the squared Frobenius norm of a symmetric matrix equals the sum of its squared eigenvalues, the preceding upper bound gives
∣ C θ ( ξ ) ∣ ≤ N max { 1 , p − 1 } max { 1 , θ o − p − 2 2 } |C_\theta(\xi)| \le\sqrt{N}\max\{1,p-1\}\max\left\{1,\theta_o^{-\frac{p-2}{2}}\right\} ∣ C θ ( ξ ) ∣ ≤ N max { 1 , p − 1 } max { 1 , θ o − 2 p − 2 }
for θ ∈ [ θ o , 1 ] \theta\in[\theta_o,1] θ ∈ [ θ o , 1 ] , ξ ≠ 0 \xi\ne0 ξ = 0 .
Using the representation of A ε A_\varepsilon A ε and the homogeneity of the Frobenius norm, we therefore obtain
∣ A ε ( v ε , ∇ v ε ) ∣ = a ( v ε ) w ε p − 2 2 ∣ C θ ε ( ∇ v ε ) ∣ ≤ C a ( v ε ) w ε p − 2 2 , |A_\varepsilon(v_\varepsilon,\nabla v_\varepsilon)| = a(v_\varepsilon)w_\varepsilon^{\frac{p-2}{2}}|C_{\theta_\varepsilon}(\nabla v_\varepsilon)| \le C a(v_\varepsilon)w_\varepsilon^{\frac{p-2}{2}}, ∣ A ε ( v ε , ∇ v ε ) ∣ = a ( v ε ) w ε 2 p − 2 ∣ C θ ε ( ∇ v ε ) ∣ ≤ C a ( v ε ) w ε 2 p − 2 ,
provided that θ ε ≥ θ o \theta_\varepsilon\ge\theta_o θ ε ≥ θ o . Here we may take
C = N max { 1 , p − 1 } max { 1 , θ o − p − 2 2 } , C=\sqrt{N}\max\{1,p-1\}\max\left\{1,\theta_o^{-\frac{p-2}{2}}\right\}, C = N max { 1 , p − 1 } max { 1 , θ o − 2 p − 2 } ,
which depends only on N , p , k N,p,k N , p , k .
To estimate B ε B_\varepsilon B ε , we compute the partial derivatives of A ε ( v , ξ ) A_\varepsilon(v,\xi) A ε ( v , ξ ) and F ε ( v , ξ ) F_\varepsilon(v,\xi) F ε ( v , ξ ) with respect to the gradient argument ξ \xi ξ , keeping v > 0 v>0 v > 0 fixed. These are precisely the derivatives appearing in the definition of B ε B_\varepsilon B ε . For ξ ≠ 0 \xi\ne0 ξ = 0 , differentiation gives
∇ ξ Θ ε ( v , ξ ) = 2 Θ ε ( v , ξ ) ( 1 − Θ ε ( v , ξ ) ) ξ ∣ ξ ∣ 2 . \nabla_\xi\Theta_\varepsilon(v,\xi)=2\Theta_\varepsilon(v,\xi)(1-\Theta_\varepsilon(v,\xi))\frac{\xi}{|\xi|^2}. ∇ ξ Θ ε ( v , ξ ) = 2 Θ ε ( v , ξ ) ( 1 − Θ ε ( v , ξ )) ∣ ξ ∣ 2 ξ .
Since 0 < Θ ε ( v , ξ ) < 1 0<\Theta_\varepsilon(v,\xi)<1 0 < Θ ε ( v , ξ ) < 1 , we have Θ ε ( v , ξ ) ( 1 − Θ ε ( v , ξ ) ) ≤ 1 4 \Theta_\varepsilon(v,\xi)(1-\Theta_\varepsilon(v,\xi))\le\frac14 Θ ε ( v , ξ ) ( 1 − Θ ε ( v , ξ )) ≤ 4 1 , and we obtain
∣ ∂ ξ ℓ Θ ε ( v , ξ ) ∣ ≤ ∣ ∇ ξ Θ ε ( v , ξ ) ∣ ≤ 1 2 ∣ ξ ∣ . |\partial_{\xi_\ell}\Theta_\varepsilon(v,\xi)|\le|\nabla_\xi\Theta_\varepsilon(v,\xi)|\le\frac{1}{2|\xi|}. ∣ ∂ ξ ℓ Θ ε ( v , ξ ) ∣ ≤ ∣ ∇ ξ Θ ε ( v , ξ ) ∣ ≤ 2∣ ξ ∣ 1 .
We next consider the explicit dependence of C θ ( ξ ) C_\theta(\xi) C θ ( ξ ) on the direction of ξ \xi ξ . For ℓ = 1 , … , N \ell=1,\ldots,N ℓ = 1 , … , N ,
∂ ξ ℓ ( ξ ∣ ξ ∣ ) = e ℓ ∣ ξ ∣ − ξ ℓ ξ ∣ ξ ∣ 3 , \partial_{\xi_\ell}\left(\frac{\xi}{|\xi|}\right)=\frac{e_\ell}{|\xi|}-\frac{\xi_\ell\xi}{|\xi|^3}, ∂ ξ ℓ ( ∣ ξ ∣ ξ ) = ∣ ξ ∣ e ℓ − ∣ ξ ∣ 3 ξ ℓ ξ ,
where e ℓ e_\ell e ℓ is the ℓ \ell ℓ -th standard basis vector. In particular,
∣ ∂ ξ ℓ ( ξ ∣ ξ ∣ ) ∣ ≤ 1 ∣ ξ ∣ . \left|\partial_{\xi_\ell}\left(\frac{\xi}{|\xi|}\right)\right|\le\frac{1}{|\xi|}. ∂ ξ ℓ ( ∣ ξ ∣ ξ ) ≤ ∣ ξ ∣ 1 .
Keeping θ \theta θ fixed and applying the product rule to the tensor product in C θ ( ξ ) \mathbf C_\theta(\xi) C θ ( ξ ) therefore yields
∣ ∂ ξ ℓ C θ ( ξ ) ∣ ≤ 2 ∣ p − 2 ∣ ∣ ξ ∣ θ − p 2 θ ≤ C 1 ∣ ξ ∣ for θ ∈ [ θ o , 1 ] . \left|\partial_{\xi_\ell}\mathbf C_\theta(\xi)\right|
\le\frac{2|p-2|}{|\xi|}\theta^{-\frac p2}\theta
\le\frac{C_1}{|\xi|}
\qquad\text{for }\theta\in[\theta_o,1]. ∣ ∂ ξ ℓ C θ ( ξ ) ∣ ≤ ∣ ξ ∣ 2∣ p − 2∣ θ − 2 p θ ≤ ∣ ξ ∣ C 1 for θ ∈ [ θ o , 1 ] .
To account also for the dependence through Θ ε ( v , ξ ) \mathbf\Theta_\varepsilon(v,\xi) Θ ε ( v , ξ ) , we use (36) and (37) . These formulas imply
∣ ∂ θ C θ ( ξ ) ∣ ≤ C 2 , ∣ j ′ ( θ ) ∣ ≤ C 3 for θ ∈ [ θ o , 1 ] , ξ ≠ 0. |\partial_\theta\mathbf C_\theta(\xi)|\le C_2,\qquad
|\mathbf j'(\theta)|\le C_3
\qquad\text{for }\theta\in[\theta_o,1],\ \xi\ne0. ∣ ∂ θ C θ ( ξ ) ∣ ≤ C 2 , ∣ j ′ ( θ ) ∣ ≤ C 3 for θ ∈ [ θ o , 1 ] , ξ = 0.
The constants C 1 , C 2 , C 3 C_1,C_2,C_3 C 1 , C 2 , C 3 depend only on N , p , k N,p,k N , p , k . By their definitions, C θ ( ξ ) \mathbf C_\theta(\xi) C θ ( ξ ) and j ( θ ) \mathbf j(\theta) j ( θ ) are bounded on the same interval, with the same parameter dependence.
Suppose now that Θ ε ( v , ξ ) ≥ θ o \mathbf\Theta_\varepsilon(v,\xi)\ge\theta_o Θ ε ( v , ξ ) ≥ θ o . The chain rule gives
∂ ξ ℓ ( C Θ ε ( v , ξ ) ( ξ ) ) = ∂ ξ ℓ C θ ( ξ ) ∣ θ = Θ ε ( v , ξ ) + ∂ θ C θ ( ξ ) ∣ θ = Θ ε ( v , ξ ) ∂ ξ ℓ Θ ε ( v , ξ ) . \partial_{\xi_\ell}\left(\mathbf C_{\mathbf\Theta_\varepsilon(v,\xi)}(\xi)\right)
=
\left.\partial_{\xi_\ell}\mathbf C_\theta(\xi)\right|_{\theta=\mathbf\Theta_\varepsilon(v,\xi)}
+
\left.\partial_\theta\mathbf C_\theta(\xi)\right|_{\theta=\mathbf\Theta_\varepsilon(v,\xi)}
\partial_{\xi_\ell}\mathbf\Theta_\varepsilon(v,\xi). ∂ ξ ℓ ( C Θ ε ( v , ξ ) ( ξ ) ) = ∂ ξ ℓ C θ ( ξ ) ∣ θ = Θ ε ( v , ξ ) + ∂ θ C θ ( ξ ) ∣ θ = Θ ε ( v , ξ ) ∂ ξ ℓ Θ ε ( v , ξ ) .
Here the first term accounts for the explicit dependence on ξ \xi ξ , with θ \theta θ held fixed, and the second for the dependence through Θ ε ( v , ξ ) \mathbf\Theta_\varepsilon(v,\xi) Θ ε ( v , ξ ) . Combining the preceding bounds, we obtain
∣ ∂ ξ ℓ ( C Θ ε ( v , ξ ) ( ξ ) ) ∣ ≤ C 1 + 2 C 2 ∣ ξ ∣ . \left|\partial_{\xi_\ell}\left(\mathbf C_{\mathbf\Theta_\varepsilon(v,\xi)}(\xi)\right)\right|
\le\frac{C_1+2C_2}{|\xi|}. ∂ ξ ℓ ( C Θ ε ( v , ξ ) ( ξ ) ) ≤ ∣ ξ ∣ C 1 + 2 C 2 .
The function j \mathbf j j depends on ξ \xi ξ only through Θ ε ( v , ξ ) \mathbf\Theta_\varepsilon(v,\xi) Θ ε ( v , ξ ) . Hence the chain rule gives
∇ ξ ( j ( Θ ε ( v , ξ ) ) ) = j ′ ( Θ ε ( v , ξ ) ) ∇ ξ Θ ε ( v , ξ ) . \nabla_\xi\left(\mathbf j(\mathbf\Theta_\varepsilon(v,\xi))\right)
=
\mathbf j'(\mathbf\Theta_\varepsilon(v,\xi))
\nabla_\xi\mathbf\Theta_\varepsilon(v,\xi). ∇ ξ ( j ( Θ ε ( v , ξ )) ) = j ′ ( Θ ε ( v , ξ )) ∇ ξ Θ ε ( v , ξ ) .
Since ∣ j ′ ( θ ) ∣ ≤ C 3 |\mathbf j'(\theta)|\le C_3 ∣ j ′ ( θ ) ∣ ≤ C 3 on [ θ o , 1 ] [\theta_o,1] [ θ o , 1 ] , with C 3 C_3 C 3 depending only on N , p , k N,p,k N , p , k , we obtain
∣ ∇ ξ ( j ( Θ ε ( v , ξ ) ) ) ∣ = ∣ j ′ ( Θ ε ( v , ξ ) ) ∣ ∣ ∇ ξ Θ ε ( v , ξ ) ∣ ≤ 2 C 3 ∣ ξ ∣ . \left|\nabla_\xi\left(\mathbf j(\mathbf\Theta_\varepsilon(v,\xi))\right)\right|
=
|\mathbf j'(\mathbf\Theta_\varepsilon(v,\xi))|
|\nabla_\xi\mathbf\Theta_\varepsilon(v,\xi)|
\le\frac{2C_3}{|\xi|}. ∣ ∇ ξ ( j ( Θ ε ( v , ξ )) ) ∣ = ∣ j ′ ( Θ ε ( v , ξ )) ∣∣ ∇ ξ Θ ε ( v , ξ ) ∣ ≤ ∣ ξ ∣ 2 C 3 .
Taking C = max { C 1 + 2 C 2 , 2 C 3 } C=\max\{C_1+2C_2,2C_3\} C = max { C 1 + 2 C 2 , 2 C 3 } proves both estimates. This constant depends only on N , p , k N,p,k N , p , k and is independent of ε \varepsilon ε and v v v . The first estimate includes both the dependence through Θ ε ( v , ξ ) \mathbf\Theta_\varepsilon(v,\xi) Θ ε ( v , ξ ) and the explicit dependence on the direction of ξ \xi ξ .
We now differentiate the representations of A ε \mathbf A_\varepsilon A ε and F ε \mathbf F_\varepsilon F ε . Using
∇ ξ ∣ ξ ∣ p − 2 = ( p − 2 ) ∣ ξ ∣ p − 4 ξ , ∇ ξ ∣ ξ ∣ p = p ∣ ξ ∣ p − 2 ξ , \nabla_\xi|\xi|^{p-2}=(p-2)|\xi|^{p-4}\xi,\qquad
\nabla_\xi|\xi|^p=p|\xi|^{p-2}\xi, ∇ ξ ∣ ξ ∣ p − 2 = ( p − 2 ) ∣ ξ ∣ p − 4 ξ , ∇ ξ ∣ ξ ∣ p = p ∣ ξ ∣ p − 2 ξ ,
the product rule and the preceding bounds give
∣ ∂ ξ ℓ A ε ( v , ξ ) ∣ ≤ C a ( v ) ∣ ξ ∣ p − 3 , ∣ ∇ ξ F ε ( v , ξ ) ∣ ≤ C a ( v ) v ∣ ξ ∣ p − 1 for Θ ε ( v , ξ ) ≥ θ o . |\partial_{\xi_\ell}\mathbf A_\varepsilon(v,\xi)|
\le C a(v)|\xi|^{p-3},\qquad
|\nabla_\xi\mathbf F_\varepsilon(v,\xi)|
\le C\frac{a(v)}{v}|\xi|^{p-1}
\qquad\text{for }\mathbf\Theta_\varepsilon(v,\xi)\ge\theta_o. ∣ ∂ ξ ℓ A ε ( v , ξ ) ∣ ≤ C a ( v ) ∣ ξ ∣ p − 3 , ∣ ∇ ξ F ε ( v , ξ ) ∣ ≤ C v a ( v ) ∣ ξ ∣ p − 1 for Θ ε ( v , ξ ) ≥ θ o .
Here and below, C C C depends only on N , p , k N,p,k N , p , k . Evaluating at ( v , ξ ) = ( v ε , ∇ v ε ) (v,\xi)=(v_\varepsilon,\nabla v_\varepsilon) ( v , ξ ) = ( v ε , ∇ v ε ) and applying the Frobenius Cauchy–Schwarz inequality in the definition of B ε \mathbf B_\varepsilon B ε , we obtain
∣ B ε ∣ ≤ ( ∑ ℓ = 1 N ∣ ( ∂ ξ ℓ A ε ) ( v ε , ∇ v ε ) ∣ 2 ) 1 2 ∣ ∇ 2 v ε ∣ + ∣ ( ∇ ξ F ε ) ( v ε , ∇ v ε ) ∣ . |\mathbf B_\varepsilon|
\le
\left(\sum_{\ell=1}^{N}
\left|(\partial_{\xi_\ell}\mathbf A_\varepsilon)(v_\varepsilon,\nabla v_\varepsilon)\right|^2
\right)^{\frac12}
|\nabla^2v_\varepsilon|
+
|(\nabla_\xi\mathbf F_\varepsilon)(v_\varepsilon,\nabla v_\varepsilon)|. ∣ B ε ∣ ≤ ( ℓ = 1 ∑ N ∣ ( ∂ ξ ℓ A ε ) ( v ε , ∇ v ε ) ∣ 2 ) 2 1 ∣ ∇ 2 v ε ∣ + ∣ ( ∇ ξ F ε ) ( v ε , ∇ v ε ) ∣.
Consequently,
∣ B ε ∣ ≤ C ( a ( v ε ) w ε p − 3 2 ∣ ∇ 2 v ε ∣ + a ( v ε ) v ε w ε p − 1 2 ) for θ ε ≥ θ o . (49) |\mathbf B_\varepsilon|
\le C\left(
a(v_\varepsilon)w_\varepsilon^{\frac{p-3}{2}}|\nabla^2v_\varepsilon|
+
\frac{a(v_\varepsilon)}{v_\varepsilon}w_\varepsilon^{\frac{p-1}{2}}
\right)
\qquad\text{for }\theta_\varepsilon\ge\theta_o. \tag*{(49)} ∣ B ε ∣ ≤ C ( a ( v ε ) w ε 2 p − 3 ∣ ∇ 2 v ε ∣ + v ε a ( v ε ) w ε 2 p − 1 ) for θ ε ≥ θ o . (49)
The constant is independent of ε \varepsilon ε and the lower bound m \mathbf m m .
Two cases at the maximum
Let
Q : = B 2 r ( x o ) × ( s − 2 h , s ] , Q ‾ ⊂ Ω × I , Q := B_{2r}(x_o) \times(s - 2h, s], \qquad\overline{Q} \subset\Omega\times I, Q := B 2 r ( x o ) × ( s − 2 h , s ] , Q ⊂ Ω × I ,
and suppose that
0 < m ≤ u ε ≤ M in Q , 0 < m \le u_\varepsilon\le M \qquad\text{in } Q, 0 < m ≤ u ε ≤ M in Q ,
where m m m and M M M are independent of ε \varepsilon ε . Since v ε = u ε 1 k v_\varepsilon= u_\varepsilon^{\frac{1}{k}} v ε = u ε k 1 , we have
m 1 k ≤ v ε ≤ M 1 k . m^{\frac{1}{k}} \le v_\varepsilon\le M^{\frac{1}{k}}. m k 1 ≤ v ε ≤ M k 1 .
Choose a smooth cutoff function η \eta η as in § 3.8 . More precisely, we require 0 ≤ η ≤ 1 0 \le\eta\le1 0 ≤ η ≤ 1 , with η = 0 \eta= 0 η = 0 on the parabolic boundary of Q Q Q and η = 1 \eta= 1 η = 1 on B r ( x o ) × ( s − h , s ] B_r(x_o) \times(s-h,s] B r ( x o ) × ( s − h , s ] , such that
∣ ∇ η ∣ ≤ C r , ∣ ∇ 2 η ∣ ≤ C r 2 , ∣ ∂ t η ∣ ≤ C h , |\nabla\eta| \le\frac{C}{r}, \qquad|\nabla^2\eta| \le\frac{C}{r^2}, \qquad|\partial_t\eta| \le\frac{C}{h}, ∣∇ η ∣ ≤ r C , ∣ ∇ 2 η ∣ ≤ r 2 C , ∣ ∂ t η ∣ ≤ h C ,
where C C C depends only on N N N . Consider
W ε : = η 2 w ε on B 2 r ( x o ) ‾ × [ s − 2 h , s ] . W_\varepsilon:= \eta^2w_\varepsilon\qquad\text{on } \overline{B_{2r}(x_o)} \times[s-2h,s]. W ε := η 2 w ε on B 2 r ( x o ) × [ s − 2 h , s ] .
If its maximum is zero, then ∇ v ε = 0 \nabla v_\varepsilon= 0 ∇ v ε = 0 wherever η = 1 \eta= 1 η = 1 , and the desired estimate follows immediately. Otherwise, let ( x ∗ , t ∗ ) (x_*,t_*) ( x ∗ , t ∗ ) be a point where W ε W_\varepsilon W ε attains a positive maximum. In particular,
η ( x ∗ , t ∗ ) > 0 , w ε ( x ∗ , t ∗ ) > 0. \eta(x_*,t_*) > 0, \qquad w_\varepsilon(x_*,t_*) > 0. η ( x ∗ , t ∗ ) > 0 , w ε ( x ∗ , t ∗ ) > 0.
Throughout the following calculation, all quantities are evaluated at ( x ∗ , t ∗ ) (x_*,t_*) ( x ∗ , t ∗ ) . We first derive an inequality valid for every value of θ ε ( x ∗ , t ∗ ) \theta_\varepsilon(x_*,t_*) θ ε ( x ∗ , t ∗ ) , and only afterwards distinguish between two cases.
Since w ε ( x ∗ , t ∗ ) > 0 w_\varepsilon(x_*,t_*) > 0 w ε ( x ∗ , t ∗ ) > 0 , the gradient ∇ v ε \nabla v_\varepsilon ∇ v ε is nonzero in a neighborhood of this point, so the differentiated equation applies there. At the maximum point, we have
∇ W ε = 0 , ∇ 2 W ε ≤ 0 , ∂ t W ε ≥ 0 , \nabla W_\varepsilon= 0, \qquad\nabla^2W_\varepsilon\le0, \qquad\partial_tW_\varepsilon\ge0, ∇ W ε = 0 , ∇ 2 W ε ≤ 0 , ∂ t W ε ≥ 0 ,
where the time derivative is understood from the left if t ∗ = s t_* = s t ∗ = s . Since the principal matrix A ε A_\varepsilon A ε is positive definite, these relations and the definition (42) imply
L ε W ε = ∂ t W ε − A ε ( v ε , ∇ v ε ) ⋅ ∇ 2 W ε − B ε ⋅ ∇ W ε ≥ 0 \mathcal{L}_\varepsilon W_\varepsilon= \partial_tW_\varepsilon- A_\varepsilon(v_\varepsilon,\nabla v_\varepsilon) \mathbin{\cdot} \nabla^2W_\varepsilon- B_\varepsilon\mathbin{\cdot} \nabla W_\varepsilon\ge0 L ε W ε = ∂ t W ε − A ε ( v ε , ∇ v ε ) ⋅ ∇ 2 W ε − B ε ⋅ ∇ W ε ≥ 0
at ( x ∗ , t ∗ ) (x_*,t_*) ( x ∗ , t ∗ ) .
Moreover,
0 = ∇ W ε = 2 η w ε ∇ η + η 2 ∇ w ε . 0 = \nabla W_\varepsilon= 2\eta w_\varepsilon\nabla\eta+ \eta^2\nabla w_\varepsilon. 0 = ∇ W ε = 2 η w ε ∇ η + η 2 ∇ w ε .
As η > 0 \eta> 0 η > 0 at the maximum point ( x ∗ , t ∗ ) (x_*,t_*) ( x ∗ , t ∗ ) , we obtain
∇ w ε = − 2 w ε ∇ η η . (50) \nabla w_\varepsilon= -2w_\varepsilon\frac{\nabla\eta}{\eta}. \tag*{(50)} ∇ w ε = − 2 w ε η ∇ η . (50)
In the following formulas, we write A ε A_\varepsilon A ε for A ε ( v ε , ∇ v ε ) A_\varepsilon(v_\varepsilon,\nabla v_\varepsilon) A ε ( v ε , ∇ v ε ) . The product rule for L ε \mathcal{L}_\varepsilon L ε gives
L ε W ε = η 2 L ε w ε + 2 η w ε L ε η − 2 w ε ( A ε ∇ η ) ⋅ ∇ η − 4 η ( A ε ∇ η ) ⋅ ∇ w ε . \mathcal{L}_\varepsilon W_\varepsilon= \eta^2\mathcal{L}_\varepsilon w_\varepsilon+ 2\eta w_\varepsilon\mathcal{L}_\varepsilon\eta- 2w_\varepsilon(A_\varepsilon\nabla\eta)\mathbin{\cdot}\nabla\eta- 4\eta(A_\varepsilon\nabla\eta)\mathbin{\cdot}\nabla w_\varepsilon. L ε W ε = η 2 L ε w ε + 2 η w ε L ε η − 2 w ε ( A ε ∇ η ) ⋅ ∇ η − 4 η ( A ε ∇ η ) ⋅ ∇ w ε .
Using (50) , the last term becomes
− 4 η ( A ε ∇ η ) ⋅ ∇ w ε = 8 w ε ( A ε ∇ η ) ⋅ ∇ η . -4\eta(A_\varepsilon\nabla\eta)\mathbin{\cdot}\nabla w_\varepsilon= 8w_\varepsilon(A_\varepsilon\nabla\eta)\mathbin{\cdot}\nabla\eta. − 4 η ( A ε ∇ η ) ⋅ ∇ w ε = 8 w ε ( A ε ∇ η ) ⋅ ∇ η .
Consequently,
0 ≤ L ε W ε = η 2 L ε w ε + 2 η w ε ( ∂ t η − A ε ⋅ ∇ 2 η − B ε ⋅ ∇ η ) + 6 w ε ( A ε ∇ η ) ⋅ ∇ η . 0 \le\mathcal{L}_\varepsilon W_\varepsilon= \eta^2\mathcal{L}_\varepsilon w_\varepsilon+ 2\eta w_\varepsilon\left(\partial_t\eta- A_\varepsilon\mathbin{\cdot}\nabla^2\eta- B_\varepsilon\mathbin{\cdot}\nabla\eta\right) + 6w_\varepsilon(A_\varepsilon\nabla\eta)\mathbin{\cdot}\nabla\eta. 0 ≤ L ε W ε = η 2 L ε w ε + 2 η w ε ( ∂ t η − A ε ⋅ ∇ 2 η − B ε ⋅ ∇ η ) + 6 w ε ( A ε ∇ η ) ⋅ ∇ η .
Now we distinguish two cases according to whether θ ε ( x ∗ , t ∗ ) ≥ θ o \theta_\varepsilon(x_*,t_*) \ge\theta_o θ ε ( x ∗ , t ∗ ) ≥ θ o or θ ε ( x ∗ , t ∗ ) < θ o \theta_\varepsilon(x_*,t_*) < \theta_o θ ε ( x ∗ , t ∗ ) < θ o . Suppose first that θ ε ( x ∗ , t ∗ ) ≥ θ o \theta_\varepsilon(x_*,t_*) \ge\theta_o θ ε ( x ∗ , t ∗ ) ≥ θ o . Then the coercive inequality (48) , the bound for the principal matrix, and the drift estimate (49) allow us to carry out the localization argument from § 3.8 . In particular, (48) yields
η 2 L ε w ε ≤ − c 1 a ( v ε ) η 2 w ε p − 2 2 ∣ ∇ 2 v ε ∣ 2 − κ o η 2 v ε − p w ε p + 2 2 . \eta^2\mathcal{L}_\varepsilon w_\varepsilon\le-c_1a(v_\varepsilon)\eta^2w_\varepsilon^{\frac{p-2}{2}}|\nabla^2v_\varepsilon|^2-\kappa_o\eta^2v_\varepsilon^{-p}w_\varepsilon^{\frac{p+2}{2}}. η 2 L ε w ε ≤ − c 1 a ( v ε ) η 2 w ε 2 p − 2 ∣ ∇ 2 v ε ∣ 2 − κ o η 2 v ε − p w ε 2 p + 2 .
For the terms containing derivatives of η \eta η , the bound
∣ A ε ∣ ≤ C a ( v ε ) w ε p − 2 2 |A_\varepsilon| \le Ca(v_\varepsilon)w_\varepsilon^{\frac{p-2}{2}} ∣ A ε ∣ ≤ C a ( v ε ) w ε 2 p − 2
and the Cauchy–Schwarz inequality give
− 2 η w ε A ε ⋅ ∇ 2 η ≤ C a ( v ε ) η w ε p 2 ∣ ∇ 2 η ∣ -2\eta w_\varepsilon A_\varepsilon\cdot\nabla^2\eta\leq Ca(v_\varepsilon)\eta w_\varepsilon^{\frac{p}{2}}\lvert\nabla^2\eta\rvert − 2 η w ε A ε ⋅ ∇ 2 η ≤ C a ( v ε ) η w ε 2 p ∣ ∇ 2 η ∣
and
6 w ε ( A ε ∇ η ) ⋅ ∇ η ≤ C a ( v ε ) w ε p 2 ∣ ∇ η ∣ 2 . 6w_\varepsilon(A_\varepsilon\nabla\eta)\cdot\nabla\eta\leq Ca(v_\varepsilon)w_\varepsilon^{\frac{p}{2}}\lvert\nabla\eta\rvert^2. 6 w ε ( A ε ∇ η ) ⋅ ∇ η ≤ C a ( v ε ) w ε 2 p ∣ ∇ η ∣ 2 .
Estimating the time derivative and the drift term by their absolute values, and moving the two negative terms to the left-hand side, we obtain
c 1 a ( v ε ) η 2 w ε p − 2 2 ∣ ∇ 2 v ε ∣ 2 + κ o η 2 v ε − p w ε p + 2 2 ≤ 2 η ∣ ∂ t η ∣ w ε + C a ( v ε ) w ε p 2 ( ∣ ∇ η ∣ 2 + η ∣ ∇ 2 η ∣ ) + 2 η w ε ∣ B ε ∣ ∣ ∇ η ∣ . (51) \begin{aligned}
c_1a(v_\varepsilon)\eta^2w_\varepsilon^{\frac{p-2}{2}}\lvert\nabla^2v_\varepsilon\rvert^2+\kappa_o\eta^2v_\varepsilon^{-p}w_\varepsilon^{\frac{p+2}{2}}
&\leq2\eta\lvert\partial_t\eta\rvert w_\varepsilon+Ca(v_\varepsilon)w_\varepsilon^{\frac{p}{2}}\left(\lvert\nabla\eta\rvert^2+\eta\lvert\nabla^2\eta\rvert\right)\\
&\quad+2\eta w_\varepsilon\lvert B_\varepsilon\rvert\lvert\nabla\eta\rvert. \tag*{(51)}
\end{aligned} c 1 a ( v ε ) η 2 w ε 2 p − 2 ∣ ∇ 2 v ε ∣ 2 + κ o η 2 v ε − p w ε 2 p + 2 ≤ 2 η ∣ ∂ t η ∣ w ε + C a ( v ε ) w ε 2 p ( ∣ ∇ η ∣ 2 + η ∣ ∇ 2 η ∣ ) + 2 η w ε ∣ B ε ∣ ∣ ∇ η ∣ . (51)
Here and below, C C C denotes a positive constant depending only on N , p , k N,p,k N , p , k , whose value may change from line to line. We next estimate the drift term 2 η w ε ∣ B ε ∣ ∣ ∇ η ∣ 2\eta w_\varepsilon\lvert B_\varepsilon\rvert\lvert\nabla\eta\rvert 2 η w ε ∣ B ε ∣ ∣ ∇ η ∣ using (49) . This gives
2 η w ε ∣ B ε ∣ ∣ ∇ η ∣ ≤ C d a ( v ε ) η w ε p − 1 2 ∣ ∇ 2 v ε ∣ ∣ ∇ η ∣ + C d a ( v ε ) v ε w ε p + 1 2 ∣ ∇ η ∣ . 2\eta w_\varepsilon\lvert B_\varepsilon\rvert\lvert\nabla\eta\rvert\leq C_da(v_\varepsilon)\eta w_\varepsilon^{\frac{p-1}{2}}\lvert\nabla^2v_\varepsilon\rvert\lvert\nabla\eta\rvert+C_d\frac{a(v_\varepsilon)}{v_\varepsilon}w_\varepsilon^{\frac{p+1}{2}}\lvert\nabla\eta\rvert. 2 η w ε ∣ B ε ∣ ∣ ∇ η ∣ ≤ C d a ( v ε ) η w ε 2 p − 1 ∣ ∇ 2 v ε ∣ ∣ ∇ η ∣ + C d v ε a ( v ε ) w ε 2 p + 1 ∣ ∇ η ∣ .
Here C d C_d C d depends only on N , p , k N,p,k N , p , k . We denote this constant explicitly to track its contribution when applying Young’s inequality to absorb the Hessian-dependent term into the Hessian term on the left-hand side of (51) . Indeed, Young’s inequality gives
C d a ( v ε ) η w ε p − 1 2 ∣ ∇ 2 v ε ∣ ∣ ∇ η ∣ = ( a ( v ε ) η w ε p − 2 4 ∣ ∇ 2 v ε ∣ ) ( C d a ( v ε ) w ε p 4 ∣ ∇ η ∣ ) ≤ 1 2 c 1 a ( v ε ) η 2 w ε p − 2 2 ∣ ∇ 2 v ε ∣ 2 + C d 2 2 c 1 a ( v ε ) w ε p 2 ∣ ∇ η ∣ 2 . \begin{aligned}
C_da(v_\varepsilon)\eta w_\varepsilon^{\frac{p-1}{2}}\lvert\nabla^2v_\varepsilon\rvert\lvert\nabla\eta\rvert
&=\left(\sqrt{a(v_\varepsilon)}\eta w_\varepsilon^{\frac{p-2}{4}}\lvert\nabla^2v_\varepsilon\rvert\right)\left(C_d\sqrt{a(v_\varepsilon)}w_\varepsilon^{\frac{p}{4}}\lvert\nabla\eta\rvert\right)\\
&\leq\frac{1}{2}c_1a(v_\varepsilon)\eta^2w_\varepsilon^{\frac{p-2}{2}}\lvert\nabla^2v_\varepsilon\rvert^2+\frac{C_d^2}{2c_1}a(v_\varepsilon)w_\varepsilon^{\frac{p}{2}}\lvert\nabla\eta\rvert^2.
\end{aligned} C d a ( v ε ) η w ε 2 p − 1 ∣ ∇ 2 v ε ∣ ∣ ∇ η ∣ = ( a ( v ε ) η w ε 4 p − 2 ∣ ∇ 2 v ε ∣ ) ( C d a ( v ε ) w ε 4 p ∣ ∇ η ∣ ) ≤ 2 1 c 1 a ( v ε ) η 2 w ε 2 p − 2 ∣ ∇ 2 v ε ∣ 2 + 2 c 1 C d 2 a ( v ε ) w ε 2 p ∣ ∇ η ∣ 2 .
Since c 1 c_1 c 1 and C d C_d C d depend only on N , p , k N,p,k N , p , k , so does the constant in front of the second term on the right-hand side. Substituting these estimates into (51) , we absorb the Hessian-dependent term from the right-hand side into half of the Hessian term on the left-hand side. Dropping the remaining nonnegative Hessian term on the left-hand side, we obtain
κ o η 2 v ε − p w ε p + 2 2 ≤ C [ η ∣ ∂ t η ∣ w ε + a ( v ε ) w ε p 2 ( ∣ ∇ η ∣ 2 + η ∣ ∇ 2 η ∣ ) + a ( v ε ) v ε η w ε p + 1 2 ∣ ∇ η ∣ ] . (52) \kappa_o\eta^2v_\varepsilon^{-p}w_\varepsilon^{\frac{p+2}{2}}
\leq C\left[\eta\lvert\partial_t\eta\rvert w_\varepsilon+a(v_\varepsilon)w_\varepsilon^{\frac{p}{2}}\left(\lvert\nabla\eta\rvert^2+\eta\lvert\nabla^2\eta\rvert\right)+\frac{a(v_\varepsilon)}{v_\varepsilon}\eta w_\varepsilon^{\frac{p+1}{2}}\lvert\nabla\eta\rvert\right]. \tag*{(52)} κ o η 2 v ε − p w ε 2 p + 2 ≤ C [ η ∣ ∂ t η ∣ w ε + a ( v ε ) w ε 2 p ( ∣ ∇ η ∣ 2 + η ∣ ∇ 2 η ∣ ) + v ε a ( v ε ) η w ε 2 p + 1 ∣ ∇ η ∣ ] . (52)
We now set
Y ε : = η v ε w ε = W ε . Y_\varepsilon:=\eta v_\varepsilon w_\varepsilon=\sqrt{W_\varepsilon}. Y ε := η v ε w ε = W ε .
Multiplying (52) by v ε p n p v_\varepsilon^pn^p v ε p n p yields
κ o Y ε p + 2 ≤ C [ v ε p n p − 1 ∣ ∂ t η ∣ Y ε 2 + a ( v ε ) v ε p Y ε p ( ∣ ∇ η ∣ 2 + η ∣ ∇ 2 η ∣ ) + a ( v ε ) v ε p − 1 Y ε p + 1 ∣ ∇ η ∣ ] . \kappa_oY_\varepsilon^{p+2}\leq C\left[v_\varepsilon^pn^{p-1}\lvert\partial_t\eta\rvert Y_\varepsilon^2+a(v_\varepsilon)v_\varepsilon^pY_\varepsilon^p\left(\lvert\nabla\eta\rvert^2+\eta\lvert\nabla^2\eta\rvert\right)+a(v_\varepsilon)v_\varepsilon^{p-1}Y_\varepsilon^{p+1}\lvert\nabla\eta\rvert\right]. κ o Y ε p + 2 ≤ C [ v ε p n p − 1 ∣ ∂ t η ∣ Y ε 2 + a ( v ε ) v ε p Y ε p ( ∣ ∇ η ∣ 2 + η ∣ ∇ 2 η ∣ ) + a ( v ε ) v ε p − 1 Y ε p + 1 ∣ ∇ η ∣ ] .
Since a ( v ) = c v 2 − p a(v)=cv^{2-p} a ( v ) = c v 2 − p , where c = k p − 2 / ( p − 1 ) c=k^{p-2}/(p-1) c = k p − 2 / ( p − 1 ) , we have
a ( v ε ) v ε p = c v ε 2 , a ( v ε ) v ε p − 1 = c v ε . a(v_\varepsilon)v_\varepsilon^p=cv_\varepsilon^2,\qquad a(v_\varepsilon)v_\varepsilon^{p-1}=cv_\varepsilon. a ( v ε ) v ε p = c v ε 2 , a ( v ε ) v ε p − 1 = c v ε .
Using p > 1 p>1 p > 1 , 0 ≤ η ≤ 1 0\leq\eta\leq1 0 ≤ η ≤ 1 , v ε ≤ M 1 k v_\varepsilon\leq M^{\frac{1}{k}} v ε ≤ M k 1 , and the bounds for the derivatives of η \eta η , we conclude that
Y ε p + 2 ≤ C ( M p k h Y ε 2 + M p k r 2 Y ε p + M 1 k r Y ε p + 1 ) . (53) Y_\varepsilon^{p+2}\leq C\left(\frac{M^{\frac{p}{k}}}{h}Y_\varepsilon^2+\frac{M^{\frac{p}{k}}}{r^2}Y_\varepsilon^p+\frac{M^{\frac{1}{k}}}{r}Y_\varepsilon^{p+1}\right). \tag*{(53)} Y ε p + 2 ≤ C ( h M k p Y ε 2 + r 2 M k p Y ε p + r M k 1 Y ε p + 1 ) . (53)
For every δ > 0 \delta>0 δ > 0 , Young’s inequality gives
M p k h Y ε 2 ≤ δ Y ε p + 2 + C ( p ) δ − 2 p M p + 2 k h p + 2 p , \frac{M^{\frac{p}{k}}}{h}Y_\varepsilon^2\leq\delta Y_\varepsilon^{p+2}+C(p)\delta^{-\frac{2}{p}}\frac{M^{\frac{p+2}{k}}}{h^{\frac{p+2}{p}}}, h M k p Y ε 2 ≤ δ Y ε p + 2 + C ( p ) δ − p 2 h p p + 2 M k p + 2 ,
as well as
M 2 k r 2 Y ε p ≤ δ Y ε p + 2 + C ( p ) δ − p 2 M p + 2 k r p + 2 , \frac{M^{\frac{2}{k}}}{r^2}\mathcal{Y}_{\varepsilon}^{p} \le\delta\mathcal{Y}_{\varepsilon}^{p+2}+C(p)\delta^{-\frac{p}{2}}\frac{M^{\frac{p+2}{k}}}{r^{p+2}}, r 2 M k 2 Y ε p ≤ δ Y ε p + 2 + C ( p ) δ − 2 p r p + 2 M k p + 2 ,
and
M 1 k r Y ε p + 1 ≤ δ Y ε p + 2 + C ( p ) δ − ( p + 1 ) M p + 2 k r p + 2 . \frac{M^{\frac{1}{k}}}{r}\mathcal{Y}_{\varepsilon}^{p+1} \le\delta\mathcal{Y}_{\varepsilon}^{p+2}+C(p)\delta^{-(p+1)}\frac{M^{\frac{p+2}{k}}}{r^{p+2}}. r M k 1 Y ε p + 1 ≤ δ Y ε p + 2 + C ( p ) δ − ( p + 1 ) r p + 2 M k p + 2 .
Choosing δ = 1 / ( 6 C ) \delta=1/(6C) δ = 1/ ( 6 C ) , with C C C as in (53) , allows us to absorb the three terms containing Y ε p + 2 \mathcal{Y}_{\varepsilon}^{p+2} Y ε p + 2 on the right-hand side into the left-hand side. Combining the remaining constants, we obtain
Y ε p + 2 ≤ C M p + 2 k ( 1 h p + 2 p + 1 r p + 2 ) . \mathcal{Y}_{\varepsilon}^{p+2}\le CM^{\frac{p+2}{k}}\left(\frac{1}{h^{\frac{p+2}{p}}}+\frac{1}{r^{p+2}}\right). Y ε p + 2 ≤ C M k p + 2 ( h p p + 2 1 + r p + 2 1 ) .
Taking the ( p + 2 ) (p+2) ( p + 2 ) -th root gives
W ε ( x ∗ , t ∗ ) ≤ C M 1 k ( 1 r + 1 h 1 p ) , \sqrt{W_{\varepsilon}(x_*,t_*)}\le CM^{\frac{1}{k}}\left(\frac{1}{r}+\frac{1}{h^{\frac{1}{p}}}\right), W ε ( x ∗ , t ∗ ) ≤ C M k 1 ( r 1 + h p 1 1 ) ,
where C C C depends only on N , p , k N,p,k N , p , k . This completes the first case.
On the other hand, if θ ε ( x ∗ , t ∗ ) < θ o \theta_{\varepsilon}(x_*,t_*)<\theta_o θ ε ( x ∗ , t ∗ ) < θ o , its definition gives
w ε ( x ∗ , t ∗ ) < θ o 1 − θ o ε 2 k 2 v ε ( x ∗ , t ∗ ) 2 ( k − 1 ) . w_{\varepsilon}(x_*,t_*)<\frac{\theta_o}{1-\theta_o}\frac{\varepsilon^2}{k^2v_{\varepsilon}(x_*,t_*)^{2(k-1)}}. w ε ( x ∗ , t ∗ ) < 1 − θ o θ o k 2 v ε ( x ∗ , t ∗ ) 2 ( k − 1 ) ε 2 .
Since η ≤ 1 \eta\le1 η ≤ 1 and v ε ≥ m 1 k v_{\varepsilon}\ge\mathbf{m}^{\frac{1}{k}} v ε ≥ m k 1 , it follows that
W ε ( x ∗ , t ∗ ) ≤ w ε ( x ∗ , t ∗ ) ≤ 1 k θ o 1 − θ o ε m − k − 1 k . \sqrt{W_{\varepsilon}(x_*,t_*)}\le\sqrt{w_{\varepsilon}(x_*,t_*)}\le\frac{1}{k}\sqrt{\frac{\theta_o}{1-\theta_o}}\,\varepsilon\mathbf{m}^{-\frac{k-1}{k}}. W ε ( x ∗ , t ∗ ) ≤ w ε ( x ∗ , t ∗ ) ≤ k 1 1 − θ o θ o ε m − k k − 1 .
Notice that the second case uses no differential inequality, only the definition of θ ε \theta_{\varepsilon} θ ε and v ε ≥ m 1 k v_{\varepsilon}\ge\mathbf{m}^{\frac{1}{k}} v ε ≥ m k 1 ; this is the only place where m \mathbf{m} m enters quantitatively.
Combining the two cases, we find
W ε ( x ∗ , t ∗ ) ≤ C M 1 k ( 1 r + 1 h 1 p ) + C ε m − k − 1 k . \sqrt{W_{\varepsilon}(x_*,t_*)}\le CM^{\frac{1}{k}}\left(\frac{1}{r}+\frac{1}{h^{\frac{1}{p}}}\right)+C\varepsilon\mathbf{m}^{-\frac{k-1}{k}}. W ε ( x ∗ , t ∗ ) ≤ C M k 1 ( r 1 + h p 1 1 ) + C ε m − k k − 1 .
For every ( x , t ) ∈ B r ( x o ) × ( s − h , s ] (x,t)\in B_r(x_o)\times(s-h,s] ( x , t ) ∈ B r ( x o ) × ( s − h , s ] , we have η ( x , t ) = 1 \eta(x,t)=1 η ( x , t ) = 1 and hence
∣ ∇ v ε ( x , t ) ∣ = W ε ( x , t ) ≤ W ε ( x ∗ , t ∗ ) ≤ C M 1 k ( 1 r + 1 h 1 p ) + C ε m − k − 1 k . |\nabla v_{\varepsilon}(x,t)|=\sqrt{W_{\varepsilon}(x,t)}\le\sqrt{W_{\varepsilon}(x_*,t_*)}\le CM^{\frac{1}{k}}\left(\frac{1}{r}+\frac{1}{h^{\frac{1}{p}}}\right)+C\varepsilon\mathbf{m}^{-\frac{k-1}{k}}. ∣∇ v ε ( x , t ) ∣ = W ε ( x , t ) ≤ W ε ( x ∗ , t ∗ ) ≤ C M k 1 ( r 1 + h p 1 1 ) + C ε m − k k − 1 .
Recalling v ε = u ε 1 k v_{\varepsilon}=u_{\varepsilon}^{\frac{1}{k}} v ε = u ε k 1 , we conclude that
∥ ∇ ( u ε 1 k ) ∥ L ∞ ( B r ( x o ) × ( s − h , s ) ) ≤ C M 1 k ( 1 r + 1 h 1 p ) + C ε m − k − 1 k , (54) \left\|\nabla\left(u_{\varepsilon}^{\frac{1}{k}}\right)\right\|_{L^\infty(B_r(x_o)\times(s-h,s))}\le CM^{\frac{1}{k}}\left(\frac{1}{r}+\frac{1}{h^{\frac{1}{p}}}\right)+C\varepsilon\mathbf{m}^{-\frac{k-1}{k}}, \tag*{(54)} ∇ ( u ε k 1 ) L ∞ ( B r ( x o ) × ( s − h , s )) ≤ C M k 1 ( r 1 + h p 1 1 ) + C ε m − k k − 1 , (54)
where C C C depends only on N , p , k N,p,k N , p , k and is independent of ε , m , M , r , h \varepsilon,\mathbf{m},M,r,h ε , m , M , r , h .
We summarize the estimate just established in the following proposition.
Proposition 4.2 (Gradient estimate for regularized solutions). Let p > 1 p>1 p > 1 and
k ≥ 2 + ( N + p − 2 ) ( p − 2 ) 2 ( p − 1 ) 2 . k\ge2+\frac{(N+p-2)(p-2)^2}{(p-1)^2}. k ≥ 2 + ( p − 1 ) 2 ( N + p − 2 ) ( p − 2 ) 2 . Let
Q : = B 2 r ( x o ) × ( s − 2 h , s ] , r , h > 0 , Q ‾ ⊂ Ω × I . Q:=B_{2r}(x_o)\times(s-2h,s],\qquad r,h>0,\qquad\overline{Q}\subset\Omega\times I. Q := B 2 r ( x o ) × ( s − 2 h , s ] , r , h > 0 , Q ⊂ Ω × I . For ε > 0 \varepsilon>0 ε > 0 , let u ε u_{\varepsilon} u ε be a smooth positive solution to (32) in a neighborhood of Q ‾ \overline{Q} Q . Suppose that
0 < m ≤ u ε ≤ M in Q . 0<\mathbf{m}\le u_{\varepsilon}\le M\qquad\text{in }Q. 0 < m ≤ u ε ≤ M in Q . Then
∥ ∇ ( u ε 1 k ) ∥ L ∞ ( B r ( x o ) × ( s − h , s ) ) ≤ C M 1 k ( 1 r + 1 h 1 p ) + C ε m − k − 1 k , \left\|\nabla\left(u_{\varepsilon}^{\frac{1}{k}}\right)\right\|_{L^\infty(B_r(x_o)\times(s-h,s))}\le CM^{\frac{1}{k}}\left(\frac{1}{r}+\frac{1}{h^{\frac{1}{p}}}\right)+C\varepsilon\mathbf{m}^{-\frac{k-1}{k}}, ∇ ( u ε k 1 ) L ∞ ( B r ( x o ) × ( s − h , s )) ≤ C M k 1 ( r 1 + h p 1 1 ) + C ε m − k k − 1 ,
where C C C depends only on N N N , p p p , k k k and is independent of ε \varepsilon ε , m \mathbf{m} m , M \mathbf{M} M , r r r , h h h . In particular, if m \mathbf{m} m and M \mathbf{M} M are independent of ε \varepsilon ε , the second term tends to zero as ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 .
Remark 4.3. For ε = 0 \varepsilon= 0 ε = 0 one has θ ≡ 1 \theta\equiv1 θ ≡ 1 , the second case discussed above is void, the last term is absent, and Proposition 4.2 becomes Lemma 3.3.
Remark 4.4 (The mechanism behind the gradient estimate). The guiding idea is that a large gradient produces a negative contribution in the differentiated equation which limits its growth. To reveal this mechanism, we pass from u ε u_\varepsilon u ε to v ε = u ε 1 / k v_\varepsilon= u_\varepsilon^{1/k} v ε = u ε 1/ k and study w ε = ∣ ∇ v ε ∣ 2 w_\varepsilon= \lvert\nabla v_\varepsilon\rvert^2 w ε = ∣ ∇ v ε ∣ 2 . The power transformation is essential: the choice of k k k ensures that, when the gradient is sufficiently large relative to the regularization scale, the differentiated equation satisfies
L ε w ε ≤ − c 1 a ( v ε ) w ε p − 2 2 ∣ ∇ 2 v ε ∣ 2 − κ 0 v ε − p w ε p + 2 2 . \mathcal{L}_\varepsilon w_\varepsilon\leq-c_1a(v_\varepsilon)w_\varepsilon^{\frac{p-2}{2}}\lvert\nabla^2v_\varepsilon\rvert^2-\kappa_0v_\varepsilon^{-p}w_\varepsilon^{\frac{p+2}{2}}. L ε w ε ≤ − c 1 a ( v ε ) w ε 2 p − 2 ∣ ∇ 2 v ε ∣ 2 − κ 0 v ε − p w ε 2 p + 2 .
The last term becomes increasingly negative as the gradient grows and thus provides the main control on its size.
To use this control locally without prescribing the gradient on the boundary, we introduce W ε = η 2 w ε W_\varepsilon=\eta^2w_\varepsilon W ε = η 2 w ε . The cutoff function removes the influence of boundary values of the gradient, but its derivatives produce additional terms in the equation. These terms represent the cost of localization: their bounds increase as the spatial radius or the time length of the cylinder decreases.
At a positive maximum of W ε W_\varepsilon W ε , the maximum principle gives L ε W ε ≥ 0 \mathcal{L}_\varepsilon W_\varepsilon\geq0 L ε W ε ≥ 0 . On the other hand, the differentiated equation provides the negative contributions displayed above. Consequently, the terms generated by the cutoff function must balance these contributions. The retained negative Hessian term is used to absorb the Hessian-dependent part of the drift. After this absorption, setting Y ε = η w ε Y_\varepsilon=\eta\sqrt{w_\varepsilon} Y ε = η w ε and using v ε ≤ M 1 / k v_\varepsilon\leq\mathbf{M}^{1/k} v ε ≤ M 1/ k leads to
Y ε p + 2 ≤ C ( M p / k h Y ε 2 + M 2 / k r 2 Y ε p + M 1 / k r Y ε p + 1 ) . Y_\varepsilon^{p+2}\leq C\left(\frac{\mathbf{M}^{p/k}}{h}Y_\varepsilon^2+\frac{\mathbf{M}^{2/k}}{r^2}Y_\varepsilon^p+\frac{\mathbf{M}^{1/k}}{r}Y_\varepsilon^{p+1}\right). Y ε p + 2 ≤ C ( h M p / k Y ε 2 + r 2 M 2/ k Y ε p + r M 1/ k Y ε p + 1 ) .
The decisive point is that all powers on the right-hand side are strictly smaller than p + 2 p+2 p + 2 . Therefore, the cutoff terms cannot balance the left-hand side for arbitrarily large Y ε Y_\varepsilon Y ε . This bounds the maximum of W ε W_\varepsilon W ε and hence the gradient on the inner cylinder, where η = 1 \eta=1 η = 1 .
The regularization requires a distinction between two regimes at the maximum point. If the gradient is sufficiently large relative to the regularization scale, the coercive inequality above applies. In the complementary regime, the gradient is already controlled directly by that scale:
∣ ∇ u ε ∣ < θ 0 1 − θ 0 ε . \lvert\nabla u_\varepsilon\rvert<\sqrt{\frac{\theta_0}{1-\theta_0}}\,\varepsilon. ∣ ∇ u ε ∣ < 1 − θ 0 θ 0 ε .
The lower bound u ε ≥ m > 0 u_\varepsilon\geq\mathbf{m}>0 u ε ≥ m > 0 then prevents the power transformation from amplifying this small gradient uncontrollably. Indeed,
∣ ∇ v ε ∣ = 1 k u ε − k − 1 k ∣ ∇ u ε ∣ ≤ C ε m − k − 1 k . \lvert\nabla v_\varepsilon\rvert=\frac{1}{k}u_\varepsilon^{-\frac{k-1}{k}}\lvert\nabla u_\varepsilon\rvert\leq C\varepsilon\mathbf{m}^{-\frac{k-1}{k}}. ∣ ∇ v ε ∣ = k 1 u ε − k k − 1 ∣ ∇ u ε ∣ ≤ C ε m − k k − 1 .
If m > 0 \mathbf{m}>0 m > 0 and M \mathbf{M} M are independent of ε \varepsilon ε , this additional term vanishes as ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 , while the remaining bound is uniform in ε \varepsilon ε .
Thus, the coercive differential inequality is needed only in the regime where the gradient might be large. In the complementary regime, the condition defining that regime already provides the required control.
Approximation of positive solutions
The goal of this section is the proof of
Lemma 5.1. Let u u u be a continuous weak solution to (1) in Ω T \Omega_T Ω T , and let Q R , S ⊆ Ω T Q_{R,S}\subseteq\Omega_T Q R , S ⊆ Ω T be a cylinder such that
0 < m ≤ u ≤ M in Q R , S . 0<\mathbf{m}\leq u\leq\mathbf{M}\qquad\text{in }Q_{R,S}. 0 < m ≤ u ≤ M in Q R , S .
Then there exist a sequence ε j ↓ 0 \varepsilon_j \downarrow0 ε j ↓ 0 and functions u j ∈ C ∞ ( Q R , S ) ∩ C 0 ( Q R , S ‾ ) u_j \in C^\infty(Q_{R,S}) \cap C^0(\overline{Q_{R,S}}) u j ∈ C ∞ ( Q R , S ) ∩ C 0 ( Q R , S ) solving
∂ t ( u j p − 1 ) = div ( a ε j ( ∇ u j ) ) in Q R , S , (55) \partial_t(u_j^{p-1})=\operatorname{div}(a_{\varepsilon_j}(\nabla u_j)) \quad\text{in } Q_{R,S}, \tag*{(55)} ∂ t ( u j p − 1 ) = div ( a ε j ( ∇ u j )) in Q R , S , (55)
where
a ε ( ξ ) : = ( ε 2 + ∣ ξ ∣ 2 ) p − 2 2 ξ , a_\varepsilon(\xi):=(\varepsilon^2+|\xi|^2)^{\frac{p-2}{2}}\xi, a ε ( ξ ) := ( ε 2 + ∣ ξ ∣ 2 ) 2 p − 2 ξ ,
such that
1 2 m ≤ u j ≤ 2 M in Q R , S ‾ , \frac{1}{2}m\leq u_j\leq2M \quad\text{in } \overline{Q_{R,S}}, 2 1 m ≤ u j ≤ 2 M in Q R , S ,
and, as j → ∞ j\to\infty j → ∞ ,
u j → u locally uniformly in Q R , S , u_j\to u \quad\text{locally uniformly in } Q_{R,S}, u j → u locally uniformly in Q R , S ,
and
u j → u uniformly on ∂ p a r Q R , S . u_j\to u \quad\text{uniformly on } \partial_{\mathrm{par}}Q_{R,S}. u j → u uniformly on ∂ par Q R , S .
Proof. By translating the spatial and temporal coordinates, we may assume that Q R , S = B R ( 0 ) × ( − S , 0 ] Q_{R,S}=B_R(0)\times(-S,0] Q R , S = B R ( 0 ) × ( − S , 0 ] . We divide the proof into several steps.
Step 1: Construction of solutions to the regularized problem. By mollifying u u u in a neighborhood of Q R , S Q_{R,S} Q R , S , we obtain a sequence g j ∈ C ∞ ( Q R , S ‾ ) g_j\in C^\infty(\overline{Q_{R,S}}) g j ∈ C ∞ ( Q R , S ) such that
1 2 m ≤ g j ≤ 2 M in Q R , S ‾ , \frac{1}{2}m\leq g_j\leq2M \quad\text{in } \overline{Q_{R,S}}, 2 1 m ≤ g j ≤ 2 M in Q R , S ,
and
g j → u uniformly on Q R , S ‾ as j → ∞ . g_j\to u \quad\text{uniformly on } \overline{Q_{R,S}} \quad\text{as } j\to\infty. g j → u uniformly on Q R , S as j → ∞.
In particular,
g j ∣ ∂ p a r Q R , S → u ∣ ∂ p a r Q R , S uniformly on ∂ p a r Q R , S . g_j|_{\partial_{\mathrm{par}}Q_{R,S}}\to u|_{\partial_{\mathrm{par}}Q_{R,S}} \quad\text{uniformly on } \partial_{\mathrm{par}}Q_{R,S}. g j ∣ ∂ par Q R , S → u ∣ ∂ par Q R , S uniformly on ∂ par Q R , S .
Our first goal is to construct positive smooth solutions to (55) , with ε j ∈ ( 0 , 1 ) \varepsilon_j\in(0,1) ε j ∈ ( 0 , 1 ) , satisfying u j = g j u_j=g_j u j = g j on ∂ p a r Q R , S \partial_{\mathrm{par}}Q_{R,S} ∂ par Q R , S . For positive smooth functions, the identities
∂ t ( u j p − 1 ) = ( p − 1 ) u j p − 2 ∂ t u j \partial_t(u_j^{p-1})=(p-1)u_j^{p-2}\partial_tu_j ∂ t ( u j p − 1 ) = ( p − 1 ) u j p − 2 ∂ t u j
and
div ( a ε j ( ∇ u j ) ) = D ξ a ε j ( ∇ u j ) ⋅ ∇ 2 u j \operatorname{div}(a_{\varepsilon_j}(\nabla u_j))=D_\xi a_{\varepsilon_j}(\nabla u_j)\cdot\nabla^2u_j div ( a ε j ( ∇ u j )) = D ξ a ε j ( ∇ u j ) ⋅ ∇ 2 u j
allow us to rewrite (55) as
∂ t u j − u j 2 − p p − 1 D ξ a ε j ( ∇ u j ) ⋅ ∇ 2 u j = 0 in Q R , S . (56) \partial_tu_j-\frac{u_j^{2-p}}{p-1}D_\xi a_{\varepsilon_j}(\nabla u_j)\cdot\nabla^2u_j=0 \quad\text{in } Q_{R,S}. \tag*{(56)} ∂ t u j − p − 1 u j 2 − p D ξ a ε j ( ∇ u j ) ⋅ ∇ 2 u j = 0 in Q R , S . (56)
Here D ξ a ε j D_\xi a_{\varepsilon_j} D ξ a ε j denotes the Jacobian matrix of a ε j a_{\varepsilon_j} a ε j with respect to ξ \xi ξ . To construct solutions to (56) , we first modify its coefficients to obtain a uniformly parabolic auxiliary problem. We begin with the scalar factor u 2 − p p − 1 \frac{u^{2-p}}{p-1} p − 1 u 2 − p , extending it smoothly to all real arguments while preserving its values on the intended range of the solutions. More precisely, we choose a smooth function a o : R → ( 0 , ∞ ) a_o:\mathbb{R}\to(0,\infty) a o : R → ( 0 , ∞ ) such that
a o ( u ) = u 2 − p p − 1 for u ∈ [ 1 2 m , 2 M ] , a_o(u)=\frac{u^{2-p}}{p-1}\quad\text{for }u\in\left[\frac{1}{2}m,2M\right], a o ( u ) = p − 1 u 2 − p for u ∈ [ 2 1 m , 2 M ] ,
and
0 < c − ≤ a o ( u ) ≤ c + < ∞ for every u ∈ R , ∥ a o ′ ∥ L ∞ ( R ) ≤ C . 0<c_-\leq a_o(u)\leq c_+<\infty\quad\text{for every }u\in\mathbb{R},\qquad\|a_o'\|_{L^\infty(\mathbb{R})}\leq C. 0 < c − ≤ a o ( u ) ≤ c + < ∞ for every u ∈ R , ∥ a o ′ ∥ L ∞ ( R ) ≤ C .
Here u u u denotes an independent scalar variable, and c − c_- c − , c + c_+ c + and C C C depend only on p p p , m m m , M M M . The function a o a_o a o should not be confused with the coefficient a ( v ) a(v) a ( v ) introduced earlier. This modification ensures that the scalar factor is bounded and bounded away from zero. To obtain uniform parabolicity, we next truncate the dependence on the gradient. For fixed j j j , choose
K ∗ > ∥ ∇ g j ∥ L ∞ ( Q R , S ) , K_*> \|\nabla g_j\|_{L^\infty(Q_{R,S})}, K ∗ > ∥∇ g j ∥ L ∞ ( Q R , S ) ,
to be specified sufficiently large later. Let χ K ∗ ∈ C 0 ∞ ( R N ) \chi_{K_*}\in C^\infty_0(\mathbb{R}^N) χ K ∗ ∈ C 0 ∞ ( R N ) satisfy
0 ≤ χ K ∗ ≤ 1 , χ K ∗ ( ξ ) = 1 if ∣ ξ ∣ ≤ K ∗ , χ K ∗ ( ξ ) = 0 if ∣ ξ ∣ ≥ 2 K ∗ . 0\leq\chi_{K_*}\leq1,\qquad\chi_{K_*}(\xi)=1\ \text{if }|\xi|\leq K_*,\qquad\chi_{K_*}(\xi)=0\ \text{if }|\xi|\geq2K_*. 0 ≤ χ K ∗ ≤ 1 , χ K ∗ ( ξ ) = 1 if ∣ ξ ∣ ≤ K ∗ , χ K ∗ ( ξ ) = 0 if ∣ ξ ∣ ≥ 2 K ∗ .
For independent variables u ∈ R u \in\mathbb{R} u ∈ R and ξ ∈ R N \xi\in\mathbb{R}^{N} ξ ∈ R N , we define
E j , K ∗ ( u , ξ ) : = χ K ∗ ( ξ ) a o ( u ) D ξ a ε j ( ξ ) + ( 1 − χ K ∗ ( ξ ) ) I . \mathbf{E}_{j,K_*}(u,\xi):=\chi_{K_*}(\xi)a_o(u)D_\xi a_{\varepsilon_j}(\xi)+(1-\chi_{K_*}(\xi))I. E j , K ∗ ( u , ξ ) := χ K ∗ ( ξ ) a o ( u ) D ξ a ε j ( ξ ) + ( 1 − χ K ∗ ( ξ )) I .
Thus,
E j , K ∗ ( u , ξ ) = u 2 − p p − 1 D ξ a ε j ( ξ ) if 1 2 m ≤ u ≤ 2 M and ∣ ξ ∣ ≤ K ∗ . \mathbf{E}_{j,K_*}(u,\xi)=\frac{u^{2-p}}{p-1}D_\xi\mathbf{a}_{\varepsilon_j}(\xi)\qquad\text{if }\frac{1}{2}\mathbf{m}\leq u\leq2\mathbf{M}\text{ and }|\xi|\leq K_*. E j , K ∗ ( u , ξ ) = p − 1 u 2 − p D ξ a ε j ( ξ ) if 2 1 m ≤ u ≤ 2 M and ∣ ξ ∣ ≤ K ∗ .
For fixed j j j and K ∗ K_* K ∗ , the matrix E j , K ∗ \mathbf{E}_{j,K_*} E j , K ∗ is uniformly positive definite and bounded on R × R N \mathbb{R}\times\mathbb{R}^{N} R × R N . The corresponding ellipticity bounds may depend on p p p , m \mathbf{m} m , M \mathbf{M} M , ε j \varepsilon_j ε j , K ∗ K_* K ∗ . Classical existence theory (see, for example, [23 Theorem 12.16] ) yields a classical solution u j , K ∗ ∈ C ∞ ( Q R , S ) ∩ C ( Q R , S ‾ ) u_{j,K_*}\in C^\infty(Q_{R,S})\cap C(\overline{Q_{R,S}}) u j , K ∗ ∈ C ∞ ( Q R , S ) ∩ C ( Q R , S ) of the Cauchy–Dirichlet problem
{ ∂ t u j , K ∗ = E j , K ∗ ( u j , K ∗ , ∇ u j , K ∗ ) ∙ ∇ 2 u j , K ∗ in Q R , S , u j , K ∗ = g j on ∂ p a r Q R , S . \begin{cases}
\partial_tu_{j,K_*}=\mathbf{E}_{j,K_*}(u_{j,K_*},\nabla u_{j,K_*})\bullet\nabla^2u_{j,K_*}&\text{in }Q_{R,S},\\
u_{j,K_*}=g_j&\text{on }\partial_{\mathrm{par}}Q_{R,S}.
\end{cases} { ∂ t u j , K ∗ = E j , K ∗ ( u j , K ∗ , ∇ u j , K ∗ ) ∙ ∇ 2 u j , K ∗ u j , K ∗ = g j in Q R , S , on ∂ par Q R , S .
Since 1 2 m ≤ g j ≤ 2 M \frac{1}{2}\mathbf{m}\leq g_j\leq2\mathbf{M} 2 1 m ≤ g j ≤ 2 M on ∂ p a r Q R , S \partial_{\mathrm{par}}Q_{R,S} ∂ par Q R , S , the maximum principle yields
1 2 m ≤ u j , K ∗ ≤ 2 M in Q R , S ‾ . \frac{1}{2}\mathbf{m}\leq u_{j,K_*}\leq2\mathbf{M}\qquad\text{in }\overline{Q_{R,S}}. 2 1 m ≤ u j , K ∗ ≤ 2 M in Q R , S .
Consequently, the definition of a o a_o a o gives
a o ( u j , K ∗ ) = u j , K ∗ 2 − p p − 1 in Q R , S . a_o(u_{j,K_*})=\frac{u_{j,K_*}^{2-p}}{p-1}\qquad\text{in }Q_{R,S}. a o ( u j , K ∗ ) = p − 1 u j , K ∗ 2 − p in Q R , S .
We claim that, for K ∗ K_* K ∗ sufficiently large,
∣ ∇ u j , K ∗ ∣ < K ∗ throughout Q R , S . |\nabla u_{j,K_*}|<K_*\qquad\text{throughout }Q_{R,S}. ∣∇ u j , K ∗ ∣ < K ∗ throughout Q R , S .
Then χ K ∗ ( ∇ u j , K ∗ ) = 1 \chi_{K_*}(\nabla u_{j,K_*})=1 χ K ∗ ( ∇ u j , K ∗ ) = 1 , so, by the preceding identity for a o ( u j , K ∗ ) a_o(u_{j,K_*}) a o ( u j , K ∗ ) , the function u j , K ∗ u_{j,K_*} u j , K ∗ solves the untruncated regularized equation (55) . Suppose, towards a contradiction, that ∣ ∇ u j , K ∗ ∣ ≥ K ∗ |\nabla u_{j,K_*}|\geq K_* ∣∇ u j , K ∗ ∣ ≥ K ∗ at some point of Q R , S Q_{R,S} Q R , S . Since ∇ u j , K ∗ \nabla u_{j,K_*} ∇ u j , K ∗ is continuous on Q R , S ‾ \overline{Q_{R,S}} Q R , S , there is a first time t ∗ ∈ ( − S , 0 ) t_*\in(-S,0) t ∗ ∈ ( − S , 0 ) such that
max x ∈ B R ∣ ∇ u j , K ∗ ( x , t ∗ ) ∣ = K ∗ . \max_{x\in B_R}|\nabla u_{j,K_*}(x,t_*)|=K_*. x ∈ B R max ∣∇ u j , K ∗ ( x , t ∗ ) ∣ = K ∗ .
Here t ∗ > − S t_*>-S t ∗ > − S because u j , K ∗ ( ⋅ , − S ) = g j ( ⋅ , − S ) u_{j,K_*}(\cdot,-S)=g_j(\cdot,-S) u j , K ∗ ( ⋅ , − S ) = g j ( ⋅ , − S ) and
max x ∈ B R ∣ ∇ g j ( x , − S ) ∣ < K ∗ . \max_{x\in B_R}|\nabla g_j(x,-S)|<K_*. x ∈ B R max ∣∇ g j ( x , − S ) ∣ < K ∗ .
By the definition of t ∗ t_* t ∗ ,
∣ ∇ u j , K ∗ ∣ < K ∗ in Q ∗ : = B R × ( − S , t ∗ ) . |\nabla u_{j,K_*}|<K_*\qquad\text{in }Q_*:=B_R\times(-S,t_*). ∣∇ u j , K ∗ ∣ < K ∗ in Q ∗ := B R × ( − S , t ∗ ) .
Thus, u j , K ∗ u_{j,K_*} u j , K ∗ solves the untruncated regularized equation (55) in Q ∗ Q_* Q ∗ . Set v j : = u j , K ∗ 1 k v_j:=u_{j,K_*}^{\frac{1}{k}} v j := u j , K ∗ k 1 , with k ≥ k o k\geq k_o k ≥ k o as in (9) . We claim that
max Q ∗ ‾ ∣ ∇ v j ∣ ≤ max { max ∂ p a r Q ∗ ∣ ∇ v j ∣ , C ε j m − k − 1 } . (57) \max_{\overline{Q_*}}|\nabla v_j|\leq\max\left\{\max_{\partial_{\mathrm{par}}Q_*}|\nabla v_j|,C\varepsilon_j^{m-k^{-1}}\right\}. \tag*{(57)} Q ∗ max ∣∇ v j ∣ ≤ max { ∂ par Q ∗ max ∣∇ v j ∣ , C ε j m − k − 1 } . (57)
where C = C ( N , p , k ) C=C(N,p,k) C = C ( N , p , k ) is independent of j j j and K ∗ K_* K ∗ . Let
w j : = ∣ ∇ v j ∣ 2 , θ j : = w j w j + ε j 2 k 2 v j 2 ( k − 1 ) . w_j:=|\nabla v_j|^2,\qquad\theta_j:=\frac{w_j}{w_j+\dfrac{\varepsilon_j^2}{k^2v_j^{2(k-1)}}}. w j := ∣∇ v j ∣ 2 , θ j := w j + k 2 v j 2 ( k − 1 ) ε j 2 w j .
We apply the maximum argument from Section 4 directly to w j w_j w j , without introducing a cutoff function. We show that the constant in (57) can be chosen as
C ∗ : = 2 k − 1 k k θ o 1 − θ o . C_*:=\frac{2^{\frac{k-1}{k}}}{k}\sqrt{\frac{\theta_o}{1-\theta_o}}. C ∗ := k 2 k k − 1 1 − θ o θ o .
Suppose, towards a contradiction, that (57) fails with C = C ∗ C=C_* C = C ∗ . Let ( x ^ , t ^ ) (\hat{x},\hat{t}) ( x ^ , t ^ ) be a point where w j w_j w j attains its maximum on Q ∗ ‾ \overline{Q_*} Q ∗ . Since this maximum strictly exceeds the maximum on the parabolic boundary, we have x ^ ∈ B R \hat{x}\in B_R x ^ ∈ B R and t ^ ∈ ( − S , t ∗ ] \hat{t}\in(-S,t_*] t ^ ∈ ( − S , t ∗ ] . At this point, we have $\nabla w_j=0, ∇ 2 w j ≤ 0 \nabla^2 w_j \le0 ∇ 2 w j ≤ 0 , and ∂ t w j ≥ 0 \partial_t w_j \ge0 ∂ t w j ≥ 0 , where the time derivative is understood from the left if t ^ = t ∗ \hat{t}=t_* t ^ = t ∗ . Consequently, the positive definiteness of the principal matrix gives
L ε j w j ( x ^ , t ^ ) ≥ 0. \mathcal{L}_{\varepsilon_j}w_j(\hat{x},\hat{t}) \ge0. L ε j w j ( x ^ , t ^ ) ≥ 0.
On the other hand, the assumed failure of (57) implies
w j ( x ^ , t ^ ) > C ∗ 2 ε j 2 m − 2 ( k − 1 ) k . w_j(\hat{x},\hat{t}) > C_*^2\varepsilon_j^2\mathbf{m}^{-\frac{2(k-1)}{k}}. w j ( x ^ , t ^ ) > C ∗ 2 ε j 2 m − k 2 ( k − 1 ) .
Using the definition of C ∗ C_* C ∗ and the lower bound v j ≥ ( 1 2 m ) 1 k v_j \ge\left(\frac{1}{2}\mathbf{m}\right)^{\frac{1}{k}} v j ≥ ( 2 1 m ) k 1 , we obtain
w j ( x ^ , t ^ ) > θ o 1 − θ o ε j 2 k 2 ( 1 2 m ) 2 ( k − 1 ) k ≥ θ o 1 − θ o ε j 2 k 2 v j ( x ^ , t ^ ) 2 ( k − 1 ) . w_j(\hat{x},\hat{t}) >
\frac{\theta_o}{1-\theta_o}
\frac{\varepsilon_j^2}{k^2\left(\frac{1}{2}\mathbf{m}\right)^{\frac{2(k-1)}{k}}}
\ge
\frac{\theta_o}{1-\theta_o}
\frac{\varepsilon_j^2}{k^2v_j(\hat{x},\hat{t})^{2(k-1)}}. w j ( x ^ , t ^ ) > 1 − θ o θ o k 2 ( 2 1 m ) k 2 ( k − 1 ) ε j 2 ≥ 1 − θ o θ o k 2 v j ( x ^ , t ^ ) 2 ( k − 1 ) ε j 2 .
Equivalently, θ j ( x ^ , t ^ ) > θ o \theta_j(\hat{x},\hat{t})>\theta_o θ j ( x ^ , t ^ ) > θ o . The function u j , K ∗ u_{j,K_*} u j , K ∗ satisfies the untruncated regularized equation (55) in Q ∗ Q_* Q ∗ . If t ^ = t ∗ \hat{t}=t_* t ^ = t ∗ , the differentiated identity for w j w_j w j also holds at ( x ^ , t ∗ ) (\hat{x},t_*) ( x ^ , t ∗ ) by smoothness and passage to the limit from earlier times. Thus, (48) applies and yields, at ( x ^ , t ^ ) (\hat{x},\hat{t}) ( x ^ , t ^ ) ,
L ε j w j ≤ − c 1 a ( v j ) w j p − 2 2 ∣ ∇ 2 v j ∣ 2 − κ o v j − p w j p + 2 2 < 0 , \mathcal{L}_{\varepsilon_j}w_j \le-c_1a(v_j)w_j^{\frac{p-2}{2}}|\nabla^2v_j|^2-\kappa_o v_j^{-p}w_j^{\frac{p+2}{2}}<0, L ε j w j ≤ − c 1 a ( v j ) w j 2 p − 2 ∣ ∇ 2 v j ∣ 2 − κ o v j − p w j 2 p + 2 < 0 ,
because w j ( x ^ , t ^ ) > 0 w_j(\hat{x},\hat{t})>0 w j ( x ^ , t ^ ) > 0 . This contradicts L ε j w j ( x ^ , t ^ ) ≥ 0 \mathcal{L}_{\varepsilon_j}w_j(\hat{x},\hat{t})\ge0 L ε j w j ( x ^ , t ^ ) ≥ 0 and proves (57) . The constant C ∗ C_* C ∗ depends only on N N N , p p p , k k k .
To control the boundary term in (57) , we first estimate the normal derivatives of u j , K ∗ u_{j,K_*} u j , K ∗ on the lateral boundary of Q ∗ Q_* Q ∗ . The corresponding bounds for v j = u j , K ∗ 1 k v_j=u_{j,K_*}^{\frac{1}{k}} v j = u j , K ∗ k 1 then follow by the chain rule. Together with the prescribed initial values, these boundary estimates allow us to apply (57) on the whole cylinder. For the boundary estimate, we introduce
w ± ( x , t ) : = g j ( x , t ) ± K f ( x ) , f ( x ) : = 1 − e ∣ x ∣ − R , w_\pm(x,t):=g_j(x,t)\pm Kf(x),\qquad f(x):=1-e^{|x|-R}, w ± ( x , t ) := g j ( x , t ) ± K f ( x ) , f ( x ) := 1 − e ∣ x ∣ − R ,
where K > 1 K>1 K > 1 will be chosen sufficiently large. We compare these functions with u j , K ∗ u_{j,K_*} u j , K ∗ in the annular cylinder ( B R ∖ B R / 2 ) × ( − S , t ∗ ) (B_R\setminus B_{R/2})\times(-S,t_*) ( B R ∖ B R /2 ) × ( − S , t ∗ ) , where f f f is smooth. Since f ≥ 0 f\ge0 f ≥ 0 in B R B_R B R and f = 0 f=0 f = 0 on ∂ B R \partial B_R ∂ B R , we have w − ≤ g j ≤ w + w_-\le g_j\le w_+ w − ≤ g j ≤ w + in Q ∗ Q_* Q ∗ and w ± = g j = u j , K ∗ w_\pm=g_j=u_{j,K_*} w ± = g j = u j , K ∗ on ∂ B R × ( − S , t ∗ ) \partial B_R\times(-S,t_*) ∂ B R × ( − S , t ∗ ) . On the inner lateral boundary, we have f = 1 − e − R / 2 > 0 f=1-e^{-R/2}>0 f = 1 − e − R /2 > 0 . We therefore impose the additional condition
K ( 1 − e − R / 2 ) ≥ 2 M − 1 2 m . K(1-e^{-R/2})\ge2M-\frac{1}{2}\mathbf{m}. K ( 1 − e − R /2 ) ≥ 2 M − 2 1 m .
Using the bounds 1 2 m ≤ g j , u j , K ∗ ≤ 2 M \frac{1}{2}\mathbf{m}\le g_j,u_{j,K_*}\le2M 2 1 m ≤ g j , u j , K ∗ ≤ 2 M , we obtain
w − ≤ 2 M − K ( 1 − e − R / 2 ) ≤ 1 2 m ≤ u j , K ∗ w_-\le2M-K(1-e^{-R/2})\le\frac{1}{2}\mathbf{m}\le u_{j,K_*} w − ≤ 2 M − K ( 1 − e − R /2 ) ≤ 2 1 m ≤ u j , K ∗
and
w + ≥ 1 2 m + K ( 1 − e − R / 2 ) ≥ 2 M ≥ u j , K ∗ w_+\ge\frac{1}{2}\mathbf{m}+K(1-e^{-R/2})\ge2M\ge u_{j,K_*} w + ≥ 2 1 m + K ( 1 − e − R /2 ) ≥ 2 M ≥ u j , K ∗
on ∂ B R / 2 × ( − S , t ∗ ) \partial B_{R/2}\times(-S,t_*) ∂ B R /2 × ( − S , t ∗ ) . At the initial time, the same ordering follows from u j , K ∗ = g j u_{j,K_*}=g_j u j , K ∗ = g j and f ≥ 0 f\ge0 f ≥ 0 . Thus,
w − ≤ u j , K ∗ ≤ w + w_-\le u_{j,K_*}\le w_+ w − ≤ u j , K ∗ ≤ w +
on the entire parabolic boundary of the annular cylinder.
For 0 < ∣ x ∣ < R 0<|x|<R 0 < ∣ x ∣ < R , we have e − R ≤ ∣ ∇ f ( x ) ∣ ≤ 1 e^{-R}\le|\nabla f(x)|\le1 e − R ≤ ∣∇ f ( x ) ∣ ≤ 1 . Thus, choosing
K > ( e R + 1 ) ∥ ∇ g j ∥ L ∞ ( Q ∗ ) , K>(e^R+1)\|\nabla g_j\|_{L^\infty(Q_*)}, K > ( e R + 1 ) ∥∇ g j ∥ L ∞ ( Q ∗ ) ,
the triangle inequality yields
K e R ( e R + 1 ) ≤ ∣ ∇ w ± ∣ ≤ 2 K \frac{K}{e^R(e^R+1)}\le|\nabla w_\pm|\le2K e R ( e R + 1 ) K ≤ ∣∇ w ± ∣ ≤ 2 K
throughout the annular cylinder.
Differentiating the regularized flux gives
D ξ a ε j ( ξ ) = ( ε j 2 + ∣ ξ ∣ 2 ) p − 2 2 [ I + ( p − 2 ) ξ ⊗ ξ ε j 2 + ∣ ξ ∣ 2 ] . D_\xi a_{\varepsilon_j}(\xi)=\left(\varepsilon_j^2+|\xi|^2\right)^{\frac{p-2}{2}}
\left[I+(p-2)\frac{\xi\otimes\xi}{\varepsilon_j^2+|\xi|^2}\right]. D ξ a ε j ( ξ ) = ( ε j 2 + ∣ ξ ∣ 2 ) 2 p − 2 [ I + ( p − 2 ) ε j 2 + ∣ ξ ∣ 2 ξ ⊗ ξ ] .
For ξ ≠ 0 \xi\ne0 ξ = 0 , the matrix in brackets has eigenvalue 1 1 1 in directions orthogonal to ξ \xi ξ and eigenvalue
1 + ( p − 2 ) ∣ ξ ∣ 2 ε j 2 + ∣ ξ ∣ 2 ≥ min { 1 , p − 1 } . 1+(p-2)\frac{|\xi|^2}{\varepsilon_j^2+|\xi|^2}\ge\min\{1,p-1\}. 1 + ( p − 2 ) ε j 2 + ∣ ξ ∣ 2 ∣ ξ ∣ 2 ≥ min { 1 , p − 1 } .
in the direction of ξ \xi ξ . Evaluating at ξ = ∇ w ± \xi=\nabla w_\pm ξ = ∇ w ± , we conclude that the eigenvalues of D ξ a ε j ( ∇ w ± ) D_\xi a_{\varepsilon_j}(\nabla w_\pm) D ξ a ε j ( ∇ w ± ) are bounded below by
min { 1 , p − 1 } ( ε j 2 + ∣ ∇ w ± ∣ 2 ) p − 2 2 ≥ c ( p , R ) K p − 2 . \min\{1,p-1\}\left(\varepsilon_j^2+|\nabla w_\pm|^2\right)^{\frac{p-2}{2}}\ge c(p,R)K^{p-2}. min { 1 , p − 1 } ( ε j 2 + ∣∇ w ± ∣ 2 ) 2 p − 2 ≥ c ( p , R ) K p − 2 .
For p ≥ 2 p\ge2 p ≥ 2 , the last inequality follows from the lower bound for ∣ ∇ w ± ∣ |\nabla w_\pm| ∣∇ w ± ∣ . For 1 < p < 2 1<p<2 1 < p < 2 , we instead use ε j 2 + ∣ ∇ w ± ∣ 2 ≤ 1 + 4 K 2 ≤ 5 K 2 \varepsilon_j^2+|\nabla w_\pm|^2\le1+4K^2\le5K^2 ε j 2 + ∣∇ w ± ∣ 2 ≤ 1 + 4 K 2 ≤ 5 K 2 , since 0 < ε j < 1 0<\varepsilon_j<1 0 < ε j < 1 and K > 1 K>1 K > 1 , and obtain
( ε j 2 + ∣ ∇ w ± ∣ 2 ) p − 2 2 ≥ 5 p − 2 2 K p − 2 . \left(\varepsilon_j^2+|\nabla w_\pm|^2\right)^{\frac{p-2}{2}}\ge5^{\frac{p-2}{2}}K^{p-2}. ( ε j 2 + ∣∇ w ± ∣ 2 ) 2 p − 2 ≥ 5 2 p − 2 K p − 2 .
Thus, the lower bound for the eigenvalues of D ξ a ε j ( ∇ w ± ) D_\xi a_{\varepsilon_j}(\nabla w_\pm) D ξ a ε j ( ∇ w ± ) holds for every p > 1 p>1 p > 1 , with c ( p , R ) > 0 c(p,R)>0 c ( p , R ) > 0 independent of j j j , K K K , and K ∗ K_* K ∗ . Moreover, for 0 < ∣ x ∣ < R 0<|x|<R 0 < ∣ x ∣ < R , we have
∇ 2 f ( x ) = − e ∣ x ∣ − R [ x ⊗ x ∣ x ∣ 2 + 1 ∣ x ∣ ( I − x ⊗ x ∣ x ∣ 2 ) ] . \nabla^2 f(x)=-e^{|x|-R}\left[\frac{x\otimes x}{|x|^2}+\frac{1}{|x|}\left(I-\frac{x\otimes x}{|x|^2}\right)\right]. ∇ 2 f ( x ) = − e ∣ x ∣ − R [ ∣ x ∣ 2 x ⊗ x + ∣ x ∣ 1 ( I − ∣ x ∣ 2 x ⊗ x ) ] .
Its eigenvalues are − e ∣ x ∣ − R -e^{|x|-R} − e ∣ x ∣ − R in the radial direction and − e ∣ x ∣ − R / ∣ x ∣ -e^{|x|-R}/|x| − e ∣ x ∣ − R /∣ x ∣ in tangential directions. Since ∇ 2 w ± = ∇ 2 g j ± K ∇ 2 f \nabla^2w_\pm=\nabla^2g_j\pm K\nabla^2f ∇ 2 w ± = ∇ 2 g j ± K ∇ 2 f , every eigenvalue of ∇ 2 w + \nabla^2w_+ ∇ 2 w + is bounded above by
∥ ∇ 2 g j ∥ L ∞ ( Q ∗ ) − K e − R min { 1 , 1 R } , \|\nabla^2g_j\|_{L^\infty(Q_*)}-Ke^{-R}\min\left\{1,\frac{1}{R}\right\}, ∥ ∇ 2 g j ∥ L ∞ ( Q ∗ ) − K e − R min { 1 , R 1 } ,
whereas every eigenvalue of ∇ 2 w − \nabla^2w_- ∇ 2 w − is bounded below by
− ∥ ∇ 2 g j ∥ L ∞ ( Q ∗ ) + K e − R min { 1 , 1 R } . -\|\nabla^2g_j\|_{L^\infty(Q_*)}+Ke^{-R}\min\left\{1,\frac{1}{R}\right\}. − ∥ ∇ 2 g j ∥ L ∞ ( Q ∗ ) + K e − R min { 1 , R 1 } .
Choosing K K K so large that
K e − R min { 1 , 1 R } > ∥ ∇ 2 g j ∥ L ∞ ( Q R , S ) , Ke^{-R}\min\left\{1,\frac{1}{R}\right\}>\|\nabla^2g_j\|_{L^\infty(Q_{R,S})}, K e − R min { 1 , R 1 } > ∥ ∇ 2 g j ∥ L ∞ ( Q R , S ) ,
we conclude that ∇ 2 w + \nabla^2w_+ ∇ 2 w + is negative definite and ∇ 2 w − \nabla^2w_- ∇ 2 w − is positive definite throughout the annular cylinder. Combining the lower eigenvalue bound for D ξ a ε j ( ∇ w ± ) D_\xi a_{\varepsilon_j}(\nabla w_\pm) D ξ a ε j ( ∇ w ± ) with the preceding Hessian estimates, we obtain
− D ξ a ε j ( ∇ w + ) ⋅ ∇ 2 w + ≥ c ( p , R ) K p − 2 [ c ( R ) K − ∥ ∇ 2 g j ∥ L ∞ ( Q ∗ ) ] > 0 -D_\xi a_{\varepsilon_j}(\nabla w_+)\cdot\nabla^2w_+\ge c(p,R)K^{p-2}\left[c(R)K-\|\nabla^2g_j\|_{L^\infty(Q_*)}\right]>0 − D ξ a ε j ( ∇ w + ) ⋅ ∇ 2 w + ≥ c ( p , R ) K p − 2 [ c ( R ) K − ∥ ∇ 2 g j ∥ L ∞ ( Q ∗ ) ] > 0
and
D ξ a ε j ( ∇ w − ) ⋅ ∇ 2 w − ≥ c ( p , R ) K p − 2 [ c ( R ) K − ∥ ∇ 2 g j ∥ L ∞ ( Q ∗ ) ] > 0 D_\xi a_{\varepsilon_j}(\nabla w_-)\cdot\nabla^2w_-\ge c(p,R)K^{p-2}\left[c(R)K-\|\nabla^2g_j\|_{L^\infty(Q_*)}\right]>0 D ξ a ε j ( ∇ w − ) ⋅ ∇ 2 w − ≥ c ( p , R ) K p − 2 [ c ( R ) K − ∥ ∇ 2 g j ∥ L ∞ ( Q ∗ ) ] > 0
throughout the annular cylinder, where
c ( R ) : = e − R min { 1 , 1 R } . c(R):=e^{-R}\min\left\{1,\frac{1}{R}\right\}. c ( R ) := e − R min { 1 , R 1 } .
The strict positivity follows from our choice of K K K . Since f f f is independent of time, we have ∂ t w ± = ∂ t g j \partial_t w_\pm=\partial_t g_j ∂ t w ± = ∂ t g j , and therefore ∣ ∂ t w ± ∣ ≤ ∥ ∂ t g j ∥ L ∞ ( Q R , S ) |\partial_t w_\pm|\le\|\partial_tg_j\|_{L^\infty(Q_{R,S})} ∣ ∂ t w ± ∣ ≤ ∥ ∂ t g j ∥ L ∞ ( Q R , S ) . In particular, this bound is independent of K K K . For the upper barrier, the preceding estimate gives
∂ t w + − w + 2 − p p − 1 D ξ a ε j ( ∇ w + ) ⋅ ∇ 2 w + ≥ − ∥ ∂ t g j ∥ L ∞ ( Q R , S ) + w + 2 − p p − 1 c ( p , R ) K p − 2 [ c ( R ) K − ∥ ∇ 2 g j ∥ L ∞ ( Q R , S ) ] . \begin{aligned}
\partial_t w_+-\frac{w_+^{2-p}}{p-1}D_\xi a_{\varepsilon_j}(\nabla w_+)\cdot\nabla^2w_+ \\
&\ge-\|\partial_tg_j\|_{L^\infty(Q_{R,S})}+\frac{w_+^{2-p}}{p-1}c(p,R)K^{p-2}\left[c(R)K-\|\nabla^2g_j\|_{L^\infty(Q_{R,S})}\right].
\end{aligned} ∂ t w + − p − 1 w + 2 − p D ξ a ε j ( ∇ w + ) ⋅ ∇ 2 w + ≥ − ∥ ∂ t g j ∥ L ∞ ( Q R , S ) + p − 1 w + 2 − p c ( p , R ) K p − 2 [ c ( R ) K − ∥ ∇ 2 g j ∥ L ∞ ( Q R , S ) ] .
By increasing K \mathbf{K} K if necessary, we ensure that the expression in brackets is at least 1 2 c ( R ) K \frac{1}{2}c(R)\mathbf{K} 2 1 c ( R ) K . If 1 < p ≤ 2 1 < p \le2 1 < p ≤ 2 , then w + ≥ g j ≥ 1 2 m w_{+} \ge g_j \ge\frac{1}{2}\mathbf{m} w + ≥ g j ≥ 2 1 m implies
w + 2 − p p − 1 ≥ ( 1 2 m ) 2 − p p − 1 > 0. \frac{w_{+}^{2-p}}{p-1} \ge\frac{\left(\frac{1}{2}\mathbf{m}\right)^{2-p}}{p-1} > 0. p − 1 w + 2 − p ≥ p − 1 ( 2 1 m ) 2 − p > 0.
Since K p − 1 → ∞ \mathbf{K}^{p-1} \to\infty K p − 1 → ∞ as K → ∞ \mathbf{K} \to\infty K → ∞ , the positive term dominates the time derivative for sufficiently large K \mathbf{K} K . Thus, w + w_{+} w + is a supersolution to (56) throughout the annular cylinder. If p > 2 p > 2 p > 2 , we instead use
w + 2 − p p − 1 ≥ ( 2 M ) 2 − p p − 1 > 0 \frac{w_{+}^{2-p}}{p-1} \ge\frac{(2\mathbf{M})^{2-p}}{p-1} > 0 p − 1 w + 2 − p ≥ p − 1 ( 2 M ) 2 − p > 0
wherever w + ≤ 2 M w_{+} \le2\mathbf{M} w + ≤ 2 M .
The same argument shows that, for sufficiently large K \mathbf{K} K , the supersolution inequality holds at every point of the annular cylinder where w + ≤ 2 M w_{+} \le2\mathbf{M} w + ≤ 2 M . Consequently, by [8 Lemma A.1] the truncated function min { w + , 2 M } \min\{w_{+},2\mathbf{M}\} min { w + , 2 M } is a weak supersolution to (55) in this case. Similarly, for sufficiently large K \mathbf{K} K , the function w − w_{-} w − satisfies the subsolution inequality wherever w − > 0 w_{-}>0 w − > 0 if p ≥ 2 p\ge2 p ≥ 2 , and wherever w − > 1 2 m w_{-}>\frac{1}{2}\mathbf{m} w − > 2 1 m if 1 < p < 2 1<p<2 1 < p < 2 . Therefore, in the same way, the truncated function max { w − , 1 2 m } \max\{w_{-},\frac{1}{2}\mathbf{m}\} max { w − , 2 1 m } is a weak subsolution to (55) in the annular cylinder for every p > 1 p>1 p > 1 . The functions max { w − , 1 2 m } \max\{w_{-},\frac{1}{2}\mathbf{m}\} max { w − , 2 1 m } , u j , K ∗ u_{j,\mathbf{K}_*} u j , K ∗ , and min { w + , 2 M } \min\{w_{+},2\mathbf{M}\} min { w + , 2 M } are bounded and bounded away from zero. Together with the ordering on the parabolic boundary of the annular cylinder, the comparison principle from [9 ] therefore yields
max { w − , 1 2 m } ≤ u j , K ∗ ≤ min { w + , 2 M } in ( B R ∖ B R / 2 ) × ( − S , t ∗ ) . \max\left\{w_{-},\frac{1}{2}\mathbf{m}\right\}\le u_{j,\mathbf{K}_*}\le\min\{w_{+},2\mathbf{M}\}\quad\text{in }(B_R\setminus B_{R/2})\times(-S,t_*). max { w − , 2 1 m } ≤ u j , K ∗ ≤ min { w + , 2 M } in ( B R ∖ B R /2 ) × ( − S , t ∗ ) .
For ∣ x ∣ ≤ 1 2 R |x|\le\frac{1}{2}R ∣ x ∣ ≤ 2 1 R , we have f ( x ) ≥ 1 − e − R / 2 f(x)\ge1-e^{-R/2} f ( x ) ≥ 1 − e − R /2 . Recalling that K \mathbf{K} K was chosen so that
K ( 1 − e − R / 2 ) ≥ 2 M − 1 2 m , \mathbf{K}(1-e^{-R/2})\ge2\mathbf{M}-\frac{1}{2}\mathbf{m}, K ( 1 − e − R /2 ) ≥ 2 M − 2 1 m ,
we also obtain
g j − K f ≤ 1 2 m ≤ u j , K ∗ ≤ 2 M ≤ g j + K f g_j-\mathbf{K}f\le\frac{1}{2}\mathbf{m}\le u_{j,\mathbf{K}_*}\le2\mathbf{M}\le g_j+\mathbf{K}f g j − K f ≤ 2 1 m ≤ u j , K ∗ ≤ 2 M ≤ g j + K f
in this inner region. Consequently,
g j ( x , t ) − K f ( x ) ≤ u j , K ∗ ( x , t ) ≤ g j ( x , t ) + K f ( x ) for every ( x , t ) ∈ Q ∗ . g_j(x,t)-\mathbf{K}f(x)\le u_{j,\mathbf{K}_*}(x,t)\le g_j(x,t)+\mathbf{K}f(x)\quad\text{for every }(x,t)\in Q_*. g j ( x , t ) − K f ( x ) ≤ u j , K ∗ ( x , t ) ≤ g j ( x , t ) + K f ( x ) for every ( x , t ) ∈ Q ∗ .
At every point ( x , t ) ∈ ∂ B R × ( − S , t ∗ ) (x,t)\in\partial B_R\times(-S,t_*) ( x , t ) ∈ ∂ B R × ( − S , t ∗ ) , the boundary condition and f = 0 f=0 f = 0 give
w − ( x , t ) = u j , K ∗ ( x , t ) = w + ( x , t ) = g j ( x , t ) . w_{-}(x,t)=u_{j,\mathbf{K}_*}(x,t)=w_{+}(x,t)=g_j(x,t). w − ( x , t ) = u j , K ∗ ( x , t ) = w + ( x , t ) = g j ( x , t ) .
Fix an arbitrary point ( x , t ) ∈ ∂ B R × ( − S , t ∗ ) (x,t)\in\partial B_R\times(-S,t_*) ( x , t ) ∈ ∂ B R × ( − S , t ∗ ) . Since both barriers agree with u j , K ∗ u_{j,\mathbf{K}_*} u j , K ∗ at this boundary point, taking radial difference quotients from inside B R B_R B R in the preceding comparison inequality gives
∂ r w + ( x , t ) ≤ ∂ r u j , K ∗ ( x , t ) ≤ ∂ r w − ( x , t ) . \partial_r w_{+}(x,t)\le\partial_r u_{j,\mathbf{K}_*}(x,t)\le\partial_r w_{-}(x,t). ∂ r w + ( x , t ) ≤ ∂ r u j , K ∗ ( x , t ) ≤ ∂ r w − ( x , t ) .
Here ∂ r \partial_r ∂ r denotes differentiation in the outward radial direction. Using ∂ r f = − 1 \partial_r f=-1 ∂ r f = − 1 on ∂ B R \partial B_R ∂ B R , we obtain
∂ r g j ( x , t ) − K ≤ ∂ r u j , K ∗ ( x , t ) ≤ ∂ r g j ( x , t ) + K . \partial_r g_j(x,t)-\mathbf{K}\le\partial_r u_{j,\mathbf{K}_*}(x,t)\le\partial_r g_j(x,t)+\mathbf{K}. ∂ r g j ( x , t ) − K ≤ ∂ r u j , K ∗ ( x , t ) ≤ ∂ r g j ( x , t ) + K .
Consequently,
∣ ∂ r u j , K ∗ ( x , t ) ∣ ≤ ∣ ∂ r g j ( x , t ) ∣ + K ≤ ∥ ∇ g j ∥ L ∞ ( Q R , S ) + K . |\partial_r u_{j,\mathbf{K}_*}(x,t)|\le|\partial_r g_j(x,t)|+\mathbf{K}\le\|\nabla g_j\|_{L^\infty(Q_{R,S})}+\mathbf{K}. ∣ ∂ r u j , K ∗ ( x , t ) ∣ ≤ ∣ ∂ r g j ( x , t ) ∣ + K ≤ ∥∇ g j ∥ L ∞ ( Q R , S ) + K .
Since the boundary point was arbitrary, this estimate holds on ∂ B R × ( − S , t ∗ ) \partial B_R\times(-S,t_*) ∂ B R × ( − S , t ∗ ) and, by continuity, also at t = t ∗ t=t_* t = t ∗ . Since u j , K ∗ = g j u_{j,\mathbf{K}_*}=g_j u j , K ∗ = g j on the lateral boundary, their tangential derivatives agree there. Hence the difference ∇ u j , K ∗ − ∇ g j \nabla u_{j,\mathbf{K}_*}-\nabla g_j ∇ u j , K ∗ − ∇ g j is normal to ∂ B R \partial B_R ∂ B R . The preceding estimate therefore gives
∣ ∇ u j , K ∗ − ∇ g j ∣ = ∣ ∂ r u j , K ∗ − ∂ r g j ∣ ≤ K on ∂ B R × ( − S , t ∗ ] . |\nabla u_{j,\mathbf{K}_*}-\nabla g_j|=|\partial_r u_{j,\mathbf{K}_*}-\partial_r g_j|\le\mathbf{K}\quad\text{on }\partial B_R\times(-S,t_*]. ∣∇ u j , K ∗ − ∇ g j ∣ = ∣ ∂ r u j , K ∗ − ∂ r g j ∣ ≤ K on ∂ B R × ( − S , t ∗ ] .
On the initial boundary, we have
∇ u j , K ∗ ( ⋅ , − S ) = ∇ g j ( ⋅ , − S ) . \nabla u_{j,\mathbf{K}_*}(\cdot,-S)=\nabla g_j(\cdot,-S). ∇ u j , K ∗ ( ⋅ , − S ) = ∇ g j ( ⋅ , − S ) .
Consequently,
max ∂ p a r Q ∗ ∣ ∇ u j , K ∗ ∣ ≤ ∥ ∇ g j ∥ L ∞ ( Q R , S ) + K . \max_{\partial_{\mathrm{par}}Q_*}|\nabla u_{j,K_*}| \le\|\nabla g_j\|_{L^\infty(Q_{R,S})}+\mathbf{K}. ∂ par Q ∗ max ∣∇ u j , K ∗ ∣ ≤ ∥∇ g j ∥ L ∞ ( Q R , S ) + K .
Since v j = u j , K ∗ 1 k v_j=u_{j,K_*}^{\frac{1}{k}} v j = u j , K ∗ k 1 and u j , K ∗ ≥ 1 2 m u_{j,K_*}\ge\frac{1}{2}\mathbf{m} u j , K ∗ ≥ 2 1 m , the chain rule also yields
max ∂ p a r Q ∗ ∣ ∇ v j ∣ ≤ 1 k ( 2 m ) k − 1 k ( ∥ ∇ g j ∥ L ∞ ( Q R , S ) + K ) . \max_{\partial_{\mathrm{par}}Q_*}|\nabla v_j|\le\frac{1}{k}\left(\frac{2}{\mathbf{m}}\right)^{\frac{k-1}{k}}\left(\|\nabla g_j\|_{L^\infty(Q_{R,S})}+\mathbf{K}\right). ∂ par Q ∗ max ∣∇ v j ∣ ≤ k 1 ( m 2 ) k k − 1 ( ∥∇ g j ∥ L ∞ ( Q R , S ) + K ) .
This bound is independent of K ∗ \mathbf{K}_* K ∗ , since the barrier parameter K \mathbf{K} K was chosen independently of the truncation threshold. The preceding estimate bounds the boundary term independently of K ∗ \mathbf{K}_* K ∗ . For fixed j j j , we may therefore choose the truncation threshold K ∗ \mathbf{K}_* K ∗ sufficiently large, possibly depending on j j j , such that
max { max ∂ p a r Q ∗ ∣ ∇ v j ∣ , C ∗ ε j m − k − 1 k } < K ∗ 2 k ( 2 M ) k − 1 k . \max\left\{\max_{\partial_{\mathrm{par}}Q_*}|\nabla v_j|,C_*\varepsilon_j\mathbf{m}^{-\frac{k-1}{k}}\right\}<\frac{\mathbf{K}_*}{2k(2\mathbf{M})^{\frac{k-1}{k}}}. max { ∂ par Q ∗ max ∣∇ v j ∣ , C ∗ ε j m − k k − 1 } < 2 k ( 2 M ) k k − 1 K ∗ .
With this choice of K ∗ \mathbf{K}_* K ∗ , the chain rule, the bound u j , K ∗ ≤ 2 M u_{j,K_*}\le2\mathbf{M} u j , K ∗ ≤ 2 M , and (57) give
max Q ∗ ‾ ∣ ∇ u j , K ∗ ∣ ≤ k ( 2 M ) k − 1 k max Q ∗ ‾ ∣ ∇ v j ∣ ≤ 1 2 K ∗ . \max_{\overline{Q_*}}|\nabla u_{j,K_*}|\le k(2\mathbf{M})^{\frac{k-1}{k}}\max_{\overline{Q_*}}|\nabla v_j|\le\frac{1}{2}\mathbf{K}_*. Q ∗ max ∣∇ u j , K ∗ ∣ ≤ k ( 2 M ) k k − 1 Q ∗ max ∣∇ v j ∣ ≤ 2 1 K ∗ .
This contradicts the choice of t ∗ t_* t ∗ , since
max B R ‾ ∣ ∇ u j , K ∗ ( ⋅ , t ∗ ) ∣ = K ∗ . \max_{\overline{B_R}}|\nabla u_{j,K_*}(\cdot,t_*)|=\mathbf{K}_*. B R max ∣∇ u j , K ∗ ( ⋅ , t ∗ ) ∣ = K ∗ .
Consequently,
∣ ∇ u j , K ∗ ∣ < K ∗ throughout Q R , S . |\nabla u_{j,K_*}|<\mathbf{K}_*\qquad\text{throughout }Q_{R,S}. ∣∇ u j , K ∗ ∣ < K ∗ throughout Q R , S .
Thus, the gradient cutoff is identically one along the solution, and u j , K ∗ u_{j,K_*} u j , K ∗ solves the untruncated regularized equation (55) on the whole cylinder Q R , S Q_{R,S} Q R , S . For each j j j , we fix such a truncation threshold and set u j : = u j , K ∗ u_j:=u_{j,K_*} u j := u j , K ∗ . This gives the desired smooth solutions to (55) .
Step 2: Locally uniform convergence. Let Q ⋐ Q R , S Q\Subset Q_{R,S} Q ⋐ Q R , S be a cylinder. Choose another cylinder Q ^ \widehat{Q} Q such that Q ⋐ Q ^ ⋐ Q R , S Q\Subset\widehat{Q}\Subset Q_{R,S} Q ⋐ Q ⋐ Q R , S . Recall that 1 2 m ≤ u j ≤ 2 M \frac{1}{2}\mathbf{m}\le u_j\le2\mathbf{M} 2 1 m ≤ u j ≤ 2 M in Q R , S Q_{R,S} Q R , S . We cover Q ^ \widehat{Q} Q by finitely many cylinders of the form B r ( x ) × ( s − h , s ] B_r(x)\times(s-h,s] B r ( x ) × ( s − h , s ] , whose corresponding larger cylinders B 2 r ( x ) × ( s − 2 h , s ] B_{2r}(x)\times(s-2h,s] B 2 r ( x ) × ( s − 2 h , s ] are compactly contained in Q R , S Q_{R,S} Q R , S . Applying Proposition 4.2 on these cylinders gives
∥ ∇ ( u j 1 k ) ∥ L ∞ ( Q ^ ) ≤ C ( M 1 k + ε j m k − 1 k ) . \|\nabla(u_j^{\frac{1}{k}})\|_{L^\infty(\widehat{Q})}\le C\left(\mathbf{M}^{\frac{1}{k}}+\frac{\varepsilon_j}{\mathbf{m}^{\frac{k-1}{k}}}\right). ∥∇ ( u j k 1 ) ∥ L ∞ ( Q ) ≤ C ( M k 1 + m k k − 1 ε j ) .
Here C C C depends only on N N N , p p p , k k k and the fixed cylinders Q ^ \widehat{Q} Q and Q R , S Q_{R,S} Q R , S ; the factors arising from the bounds 1 2 m ≤ u j ≤ 2 M \frac{1}{2}\mathbf{m}\le u_j\le2\mathbf{M} 2 1 m ≤ u j ≤ 2 M have been absorbed into C C C . By the chain rule, we have ∇ u j = k u j k − 1 k ∇ ( u j 1 k ) \nabla u_j=ku_j^{\frac{k-1}{k}}\nabla(u_j^{\frac{1}{k}}) ∇ u j = k u j k k − 1 ∇ ( u j k 1 ) . Using u j ≤ 2 M u_j\le2\mathbf{M} u j ≤ 2 M and 0 < ε j < 1 0<\varepsilon_j<1 0 < ε j < 1 , we therefore obtain
L : = sup j ∈ N ∥ ∇ u j ∥ L ∞ ( Q ^ ) ≤ C k ( 2 M ) k − 1 k ( M 1 k + 1 m k − 1 k ) < ∞ , \mathbf{L}:=\sup_{j\in\mathbb{N}}\|\nabla u_j\|_{L^\infty(\widehat{Q})}\le Ck(2\mathbf{M})^{\frac{k-1}{k}}\left(\mathbf{M}^{\frac{1}{k}}+\frac{1}{\mathbf{m}^{\frac{k-1}{k}}}\right)<\infty, L := j ∈ N sup ∥∇ u j ∥ L ∞ ( Q ) ≤ C k ( 2 M ) k k − 1 ( M k 1 + m k k − 1 1 ) < ∞ ,
where we used that m > 0 \mathbf{m}>0 m > 0 and M \mathbf{M} M are fixed.
To prove that the sequence ( u j ) (u_j) ( u j ) is equicontinuous on Q Q Q , it remains to establish a modulus of continuity in time that is independent of j j j . Let η ∈ C ∞ ( B 1 ) \eta\in C^\infty(B_1) η ∈ C ∞ ( B 1 ) be a standard mollifying kernel satisfying η ≥ 0 \eta\ge0 η ≥ 0 and ∫ B 1 η ( y ) d y = 1 \int_{B_1}\eta(y)\,dy=1 ∫ B 1 η ( y ) d y = 1 , and set η δ ( y ) : = δ − N η ( y / δ ) \eta_\delta(y):=\delta^{-N}\eta(y/\delta) η δ ( y ) := δ − N η ( y / δ ) for δ > 0 \delta>0 δ > 0 . For 0 < δ < dist ( Q , ∂ Q ^ ) 0<\delta<\operatorname{dist}(Q,\partial\widehat{Q}) 0 < δ < dist ( Q , ∂ Q ) , we define the spatial mollification by
( u j p − 1 ) δ ( x , t ) : = ∫ B δ ( x ) u j ( y , t ) p − 1 η δ ( x − y ) d y , ( x , t ) ∈ Q . (58) (u_j^{p-1})_\delta(x,t):=\int_{B_\delta(x)}u_j(y,t)^{p-1}\eta_\delta(x-y)\,dy,\qquad(x,t)\in Q. \tag*{(58)} ( u j p − 1 ) δ ( x , t ) := ∫ B δ ( x ) u j ( y , t ) p − 1 η δ ( x − y ) d y , ( x , t ) ∈ Q . (58)
The restriction on δ \delta δ ensures that the integral involves only values of u j u_j u j in Q ^ \widehat{Q} Q . The spatial Lipschitz bounds for u j u_j u j on Q ^ \widehat{Q} Q and the normalization of η δ \eta_\delta η δ give
∣ ( u j p − 1 ) δ ( x , t ) − u j p − 1 ( x , t ) ∣ ≤ δ ∥ ∇ ( u j p − 1 ) ∥ L ∞ ( Q ^ ) ≤ C ( p , m , M ) L δ (59) \left|(u_j^{p-1})_\delta(x,t)-u_j^{p-1}(x,t)\right|\leq\delta\left\|\nabla(u_j^{p-1})\right\|_{L^\infty(\widehat{Q})}\leq C(p,\mathbf{m},\mathbf{M})L\delta \tag*{(59)} ( u j p − 1 ) δ ( x , t ) − u j p − 1 ( x , t ) ≤ δ ∇ ( u j p − 1 ) L ∞ ( Q ) ≤ C ( p , m , M ) L δ (59)
for every ( x , t ) ∈ Q (x,t)\in Q ( x , t ) ∈ Q . Similarly,
∣ u j p − 1 ( x 1 , t ) − u j p − 1 ( x 2 , t ) ∣ ≤ ∥ ∇ ( u j p − 1 ) ∥ L ∞ ( Q ) ∣ x 1 − x 2 ∣ ≤ C ( p , m , M ) L ∣ x 1 − x 2 ∣ (60) \left|u_j^{p-1}(x_1,t)-u_j^{p-1}(x_2,t)\right|
\leq\left\|\nabla(u_j^{p-1})\right\|_{L^\infty(Q)}|x_1-x_2|
\leq C(p,\mathbf{m},\mathbf{M})L|x_1-x_2| \tag*{(60)} u j p − 1 ( x 1 , t ) − u j p − 1 ( x 2 , t ) ≤ ∇ ( u j p − 1 ) L ∞ ( Q ) ∣ x 1 − x 2 ∣ ≤ C ( p , m , M ) L ∣ x 1 − x 2 ∣ (60)
whenever ( x 1 , t ) , ( x 2 , t ) ∈ Q (x_1,t),(x_2,t)\in Q ( x 1 , t ) , ( x 2 , t ) ∈ Q . In both estimates, we use the chain rule
∇ ( u j p − 1 ) = ( p − 1 ) u j p − 2 ∇ u j . \nabla(u_j^{p-1})=(p-1)u_j^{p-2}\nabla u_j. ∇ ( u j p − 1 ) = ( p − 1 ) u j p − 2 ∇ u j .
Together with the bounds 1 2 m ≤ u j ≤ 2 M \frac{1}{2}m\leq u_j\leq2M 2 1 m ≤ u j ≤ 2 M , this allows us to choose
C ( p , m , M ) : = ( p − 1 ) max { ( 1 2 m ) p − 2 , ( 2 M ) p − 2 } . C(p,\mathbf{m},\mathbf{M}):=(p-1)\max\left\{\left(\frac{1}{2}m\right)^{p-2},(2M)^{p-2}\right\}. C ( p , m , M ) := ( p − 1 ) max { ( 2 1 m ) p − 2 , ( 2 M ) p − 2 } .
Fix x x x and times t 1 < t 2 t_1<t_2 t 1 < t 2 such that ( x , t 1 ) , ( x , t 2 ) ∈ Q (x,t_1),(x,t_2)\in Q ( x , t 1 ) , ( x , t 2 ) ∈ Q . In the weak formulation of (55) , we use test functions of the form φ ( y , τ ) : = η δ ( x − y ) ζ ( τ ) \varphi(y,\tau):=\eta_\delta(x-y)\zeta(\tau) φ ( y , τ ) := η δ ( x − y ) ζ ( τ ) , where ζ \zeta ζ is a smooth temporal cutoff approximating the characteristic function of [ t 1 , t 2 ] [t_1,t_2] [ t 1 , t 2 ] . Passing to the limit in this approximation gives
∣ ( u j p − 1 ) δ ( x , t 2 ) − ( u j p − 1 ) δ ( x , t 1 ) ∣ = ∣ ∫ t 1 t 2 ∫ B δ ( x ) a ε j ( ∇ u j ( y , τ ) ) ⋅ ∇ y η δ ( x − y ) d y d τ ∣ ≤ C ( L + ε j ) p − 1 t 2 − t 1 δ . (61) \begin{aligned}
\left|(u_j^{p-1})_\delta(x,t_2)-(u_j^{p-1})_\delta(x,t_1)\right|
=\left|\int_{t_1}^{t_2}\int_{B_\delta(x)}\mathbf{a}_{\varepsilon_j}(\nabla u_j(y,\tau))\cdot\nabla_y\eta_\delta(x-y)\,dy\,d\tau\right| \\
&\leq C(L+\varepsilon_j)^{p-1}\frac{t_2-t_1}{\delta}. \tag*{(61)}
\end{aligned} ( u j p − 1 ) δ ( x , t 2 ) − ( u j p − 1 ) δ ( x , t 1 ) = ∫ t 1 t 2 ∫ B δ ( x ) a ε j ( ∇ u j ( y , τ )) ⋅ ∇ y η δ ( x − y ) d y d τ ≤ C ( L + ε j ) p − 1 δ t 2 − t 1 . (61)
Here we used
∣ a ε j ( ∇ u j ) ∣ ≤ ( ε j + ∣ ∇ u j ∣ ) p − 1 ≤ ( ε j + L ) p − 1 in Q ^ , \left|\mathbf{a}_{\varepsilon_j}(\nabla u_j)\right|\leq(\varepsilon_j+|\nabla u_j|)^{p-1}\leq(\varepsilon_j+L)^{p-1}\quad\text{in }\widehat{Q}, a ε j ( ∇ u j ) ≤ ( ε j + ∣∇ u j ∣ ) p − 1 ≤ ( ε j + L ) p − 1 in Q ,
together with
∫ B δ ( x ) ∣ ∇ y η δ ( x − y ) ∣ d y = 1 δ ∫ B 1 ∣ ∇ η ( y ) ∣ d y . \int_{B_\delta(x)}|\nabla_y\eta_\delta(x-y)|\,dy=\frac{1}{\delta}\int_{B_1}|\nabla\eta(y)|\,dy. ∫ B δ ( x ) ∣ ∇ y η δ ( x − y ) ∣ d y = δ 1 ∫ B 1 ∣∇ η ( y ) ∣ d y .
Thus, we may take C = ∥ ∇ η ∥ L 1 ( B 1 ) C=\|\nabla\eta\|_{L^1(B_1)} C = ∥∇ η ∥ L 1 ( B 1 ) , which is independent of j , δ , t 1 , t 2 j,\delta,t_1,t_2 j , δ , t 1 , t 2 .
Let ( x 1 , t 1 ) , ( x 2 , t 2 ) ∈ Q (x_1,t_1),(x_2,t_2)\in Q ( x 1 , t 1 ) , ( x 2 , t 2 ) ∈ Q . Since Q Q Q is a cylinder, we also have ( x 1 , t 2 ) ∈ Q (x_1,t_2)\in Q ( x 1 , t 2 ) ∈ Q . Combining (59) , (60) , and (61) , and using 0 < ε j < 1 0<\varepsilon_j<1 0 < ε j < 1 , we obtain
∣ u j p − 1 ( x 1 , t 1 ) − u j p − 1 ( x 2 , t 2 ) ∣ ≤ ∣ u j p − 1 ( x 1 , t 1 ) − ( u j p − 1 ) δ ( x 1 , t 1 ) ∣ + ∣ ( u j p − 1 ) δ ( x 1 , t 1 ) − ( u j p − 1 ) δ ( x 1 , t 2 ) ∣ + ∣ ( u j p − 1 ) δ ( x 1 , t 2 ) − u j p − 1 ( x 1 , t 2 ) ∣ + ∣ u j p − 1 ( x 1 , t 2 ) − u j p − 1 ( x 2 , t 2 ) ∣ ≤ C [ δ + ∣ t 2 − t 1 ∣ δ + ∣ x 1 − x 2 ∣ ] . \begin{aligned}
\left|u_j^{p-1}(x_1,t_1)-u_j^{p-1}(x_2,t_2)\right|
&\leq\left|u_j^{p-1}(x_1,t_1)-(u_j^{p-1})_\delta(x_1,t_1)\right|\\
&\quad+\left|(u_j^{p-1})_\delta(x_1,t_1)-(u_j^{p-1})_\delta(x_1,t_2)\right|\\
&\quad+\left|(u_j^{p-1})_\delta(x_1,t_2)-u_j^{p-1}(x_1,t_2)\right|\\
&\quad+\left|u_j^{p-1}(x_1,t_2)-u_j^{p-1}(x_2,t_2)\right|\\
&\leq C\left[\delta+\frac{|t_2-t_1|}{\delta}+|x_1-x_2|\right].
\end{aligned} u j p − 1 ( x 1 , t 1 ) − u j p − 1 ( x 2 , t 2 ) ≤ u j p − 1 ( x 1 , t 1 ) − ( u j p − 1 ) δ ( x 1 , t 1 ) + ( u j p − 1 ) δ ( x 1 , t 1 ) − ( u j p − 1 ) δ ( x 1 , t 2 ) + ( u j p − 1 ) δ ( x 1 , t 2 ) − u j p − 1 ( x 1 , t 2 ) + u j p − 1 ( x 1 , t 2 ) − u j p − 1 ( x 2 , t 2 ) ≤ C [ δ + δ ∣ t 2 − t 1 ∣ + ∣ x 1 − x 2 ∣ ] .
for every 0 < δ < dist ( Q , ∂ Q ^ ) 0<\delta<\operatorname{dist}(Q,\partial\widehat{Q}) 0 < δ < dist ( Q , ∂ Q ) . Here C C C depends only on p , m , M , L p,\mathbf{m},\mathbf{M},L p , m , M , L and the fixed mollifying kernel η \eta η , and is independent of j , δ j,\delta j , δ , and the chosen points. Choosing first δ \delta δ and then ∣ t 2 − t 1 ∣ |t_2-t_1| ∣ t 2 − t 1 ∣ and ∣ x 1 − x 2 ∣ |x_1-x_2| ∣ x 1 − x 2 ∣ sufficiently small, we conclude that ( u j p − 1 ) j ∈ N (u_j^{p-1})_{j\in\mathbb{N}} ( u j p − 1 ) j ∈ N is equicontinuous on Q Q Q . Since σ ↦ σ 1 p − 1 \sigma\mapsto\sigma^{\frac{1}{p-1}} σ ↦ σ p − 1 1 is Lipschitz on the fixed interval [ ( 1 2 m ) p − 1 , ( 2 M ) p − 1 ] \left[\left(\frac{1}{2}m\right)^{p-1},(2M)^{p-1}\right] [ ( 2 1 m ) p − 1 , ( 2 M ) p − 1 ] , the sequence ( u j ) j ∈ N (u_j)_{j\in\mathbb{N}} ( u j ) j ∈ N is equicontinuous on Q Q Q as well. The sequence is also uniformly bounded. Applying the Arzelà–Ascoli theorem on compact cylinders and using a diagonal argument along an exhaustion of Q R , S Q_{R,S} Q R , S , we therefore obtain a subsequence, not relabeled, such that
u j → w locally uniformly in Q R , S as j → ∞ , u_j\to w\qquad\text{locally uniformly in }Q_{R,S}\text{ as }j\to\infty, u j → w locally uniformly in Q R , S as j → ∞ ,
where w : Q R , S → [ 1 2 m , 2 M ] w:Q_{R,S}\to\left[\frac{1}{2}m,2M\right] w : Q R , S → [ 2 1 m , 2 M ] is continuous.
Step 3: Initial and boundary values of the limit map. To show that the limit w w w attains the prescribed boundary values, we fix a lateral boundary point
( x b , t b ) ∈ ∂ B R × [ − S , 0 ] . (x_b,t_b)\in\partial B_R\times[-S,0]. ( x b , t b ) ∈ ∂ B R × [ − S , 0 ] .
We choose an exterior ball B ϱ ( y ) ⊂ Ω ∖ B R ‾ B_\varrho(y)\subset\Omega\setminus\overline{B_R} B ϱ ( y ) ⊂ Ω ∖ B R that is externally tangent to B R B_R B R at x b x_b x b . For a parameter γ > 0 \gamma>0 γ > 0 satisfying
( γ + 1 ) min { 1 , p − 1 } > N − 1 , (\gamma+1)\min\{1,p-1\}>N-1, ( γ + 1 ) min { 1 , p − 1 } > N − 1 ,
we define
f ( x ) : = 1 − ( ϱ ∣ x − y ∣ ) γ , x ∈ B R . f(x):=1-\left(\frac{\varrho}{|x-y|}\right)^\gamma,\qquad x\in B_R. f ( x ) := 1 − ( ∣ x − y ∣ ϱ ) γ , x ∈ B R .
We note that f ( x b ) = 0 f(x_b)=0 f ( x b ) = 0 and f > 0 f>0 f > 0 in B R B_R B R . Let K > 0 \mathbf{K}>0 K > 0 be a constant to be chosen later. For each x ∈ B R x\in B_R x ∈ B R , we choose an orthonormal basis of R N \mathbb{R}^N R N consisting of the radial unit vector x − y ∣ x − y ∣ \frac{x-y}{|x-y|} ∣ x − y ∣ x − y and N − 1 N-1 N − 1 vectors orthogonal to x − y x-y x − y . These remaining vectors are tangent to the sphere ∂ B ∣ x − y ∣ ( y ) \partial B_{|x-y|}(y) ∂ B ∣ x − y ∣ ( y ) at x x x . The vectors of this basis are eigenvectors of both matrices, as follows from the explicit formulas
∇ 2 f ( x ) = γ ϱ γ ∣ x − y ∣ γ + 2 [ I − ( γ + 2 ) x − y ∣ x − y ∣ ⊗ x − y ∣ x − y ∣ ] \nabla^2f(x)=\frac{\gamma\varrho^\gamma}{|x-y|^{\gamma+2}}\left[I-(\gamma+2)\frac{x-y}{|x-y|}\otimes\frac{x-y}{|x-y|}\right] ∇ 2 f ( x ) = ∣ x − y ∣ γ + 2 γ ϱ γ [ I − ( γ + 2 ) ∣ x − y ∣ x − y ⊗ ∣ x − y ∣ x − y ]
and
D ξ a ε j ( K ∇ f ( x ) ) = ( ε j 2 + K 2 ∣ ∇ f ( x ) ∣ 2 ) p − 2 2 [ I + ( p − 2 ) K 2 ∣ ∇ f ( x ) ∣ 2 ε j 2 + K 2 ∣ ∇ f ( x ) ∣ 2 x − y ∣ x − y ∣ ⊗ x − y ∣ x − y ∣ ] . D_{\xi}a_{\varepsilon_j}(\mathbf{K}\nabla f(x))
=\left(\varepsilon_j^2+\mathbf{K}^2|\nabla f(x)|^2\right)^{\frac{p-2}{2}}
\left[I+(p-2)\frac{\mathbf{K}^2|\nabla f(x)|^2}{\varepsilon_j^2+\mathbf{K}^2|\nabla f(x)|^2}\frac{x-y}{|x-y|}\otimes\frac{x-y}{|x-y|}\right]. D ξ a ε j ( K ∇ f ( x )) = ( ε j 2 + K 2 ∣∇ f ( x ) ∣ 2 ) 2 p − 2 [ I + ( p − 2 ) ε j 2 + K 2 ∣∇ f ( x ) ∣ 2 K 2 ∣∇ f ( x ) ∣ 2 ∣ x − y ∣ x − y ⊗ ∣ x − y ∣ x − y ] .
Indeed, the tensor product appearing in both formulas acts as the identity in the radial direction and vanishes in every tangential direction. Computing the Frobenius inner product in this common orthonormal eigenbasis, we obtain
− D ξ a ε j ( K ∇ f ( x ) ) ⋅ ( K ∇ 2 f ( x ) ) = ∣ a ε j ( K ∇ f ( x ) ) ∣ ∣ x − y ∣ [ ( γ + 1 ) ( 1 + ( p − 2 ) K 2 ∣ ∇ f ( x ) ∣ 2 ε j 2 + K 2 ∣ ∇ f ( x ) ∣ 2 ) − ( N − 1 ) ] ≥ ∣ a ε j ( K ∇ f ( x ) ) ∣ ∣ x − y ∣ [ ( γ + 1 ) min { 1 , p − 1 } − ( N − 1 ) ] > 0. (62) \begin{aligned}
-D_{\xi}a_{\varepsilon_j}(\mathbf{K}\nabla f(x))\cdot(\mathbf{K}\nabla^2f(x))
=\frac{|\mathbf{a}_{\varepsilon_j}(\mathbf{K}\nabla f(x))|}{|x-y|}
\left[(\gamma+1)\left(1+(p-2)\frac{\mathbf{K}^2|\nabla f(x)|^2}{\varepsilon_j^2+\mathbf{K}^2|\nabla f(x)|^2}\right)-(N-1)\right] \\
&\geq\frac{|\mathbf{a}_{\varepsilon_j}(\mathbf{K}\nabla f(x))|}{|x-y|}
\left[(\gamma+1)\min\{1,p-1\}-(N-1)\right]>0. \tag*{(62)}
\end{aligned} − D ξ a ε j ( K ∇ f ( x )) ⋅ ( K ∇ 2 f ( x )) = ∣ x − y ∣ ∣ a ε j ( K ∇ f ( x )) ∣ [ ( γ + 1 ) ( 1 + ( p − 2 ) ε j 2 + K 2 ∣∇ f ( x ) ∣ 2 K 2 ∣∇ f ( x ) ∣ 2 ) − ( N − 1 ) ] ≥ ∣ x − y ∣ ∣ a ε j ( K ∇ f ( x )) ∣ [ ( γ + 1 ) min { 1 , p − 1 } − ( N − 1 ) ] > 0. (62)
The strict positivity follows from the choice of γ \gamma γ , together with K > 0 \mathbf{K}>0 K > 0 and ∣ ∇ f ( x ) ∣ > 0 |\nabla f(x)|>0 ∣∇ f ( x ) ∣ > 0 . To make the dependence on K \mathbf{K} K explicit, we use the definition of the flux to write
∣ a ε j ( K ∇ f ( x ) ) ∣ = K p − 1 ∣ ∇ f ( x ) ∣ p − 1 ( 1 + ε j 2 K 2 ∣ ∇ f ( x ) ∣ 2 ) p − 2 2 . |\mathbf{a}_{\varepsilon_j}(\mathbf{K}\nabla f(x))|
=\mathbf{K}^{p-1}|\nabla f(x)|^{p-1}
\left(1+\frac{\varepsilon_j^2}{\mathbf{K}^2|\nabla f(x)|^2}\right)^{\frac{p-2}{2}}. ∣ a ε j ( K ∇ f ( x )) ∣ = K p − 1 ∣∇ f ( x ) ∣ p − 1 ( 1 + K 2 ∣∇ f ( x ) ∣ 2 ε j 2 ) 2 p − 2 .
Since ∣ ∇ f ∣ |\nabla f| ∣∇ f ∣ has a strictly positive minimum on B R ‾ \overline{B_R} B R and 0 < ε j < 1 0<\varepsilon_j<1 0 < ε j < 1 , the last factor is bounded below by a positive constant, uniformly in x x x , j j j , and K ≥ 1 \mathbf{K}\geq1 K ≥ 1 . Together with the positive minimum of 1 ∣ x − y ∣ \frac{1}{|x-y|} ∣ x − y ∣ 1 and the choice of γ \gamma γ , this shows that the right-hand side of (5.7) is bounded below by a positive constant times K p − 1 \mathbf{K}^{p-1} K p − 1 . Since p > 1 p>1 p > 1 , it can therefore be made arbitrarily large by choosing K \mathbf{K} K sufficiently large, uniformly in x ∈ B R ‾ x\in\overline{B_R} x ∈ B R and j ∈ N j\in\mathbb{N} j ∈ N .
For parameters δ ∈ ( 0 , 1 ) \delta\in(0,1) δ ∈ ( 0 , 1 ) and Λ > 0 \Lambda>0 Λ > 0 , we define the barrier functions
w ± ( x , t ) : = u ( x b , t b ) ± [ 2 δ + K f ( x ) + Λ ( t − t b ) 2 ] , ( x , t ) ∈ Q R , S . w_\pm(x,t):=u(x_b,t_b)\pm\left[2\delta+\mathbf{K}f(x)+\Lambda(t-t_b)^2\right],\qquad(x,t)\in Q_{R,S}. w ± ( x , t ) := u ( x b , t b ) ± [ 2 δ + K f ( x ) + Λ ( t − t b ) 2 ] , ( x , t ) ∈ Q R , S .
For fixed δ ∈ ( 0 , 1 ) \delta\in(0,1) δ ∈ ( 0 , 1 ) , the continuity of u u u and the uniform convergence g j → u g_j \to u g j → u on ∂ p a r Q R , S \partial_{\mathrm{par}}Q_{R,S} ∂ par Q R , S yield a neighborhood U U U of ( x b , t b ) (x_b,t_b) ( x b , t b ) and an index j δ ∈ N j_\delta\in\mathbb{N} j δ ∈ N such that
∣ g j ( x , t ) − u ( x b , t b ) ∣ ≤ ∣ g j ( x , t ) − u ( x , t ) ∣ + ∣ u ( x , t ) − u ( x b , t b ) ∣ < 2 δ |g_j(x,t)-u(x_b,t_b)| \le|g_j(x,t)-u(x,t)|+|u(x,t)-u(x_b,t_b)|<2\delta ∣ g j ( x , t ) − u ( x b , t b ) ∣ ≤ ∣ g j ( x , t ) − u ( x , t ) ∣ + ∣ u ( x , t ) − u ( x b , t b ) ∣ < 2 δ
for every ( x , t ) ∈ U ∩ ∂ p a r Q R , S (x,t)\in U\cap\partial_{\mathrm{par}}Q_{R,S} ( x , t ) ∈ U ∩ ∂ par Q R , S and every j ≥ j δ j\ge j_\delta j ≥ j δ . Since f ≥ 0 f\ge0 f ≥ 0 on B ‾ R \overline{B}_R B R and Λ ( t − t b ) 2 ≥ 0 \Lambda(t-t_b)^2\ge0 Λ ( t − t b ) 2 ≥ 0 , it follows that
w − ≤ g j ≤ w + on U ∩ ∂ p a r Q R , S , j ≥ j δ . w_- \le g_j \le w_+ \qquad\text{on } U\cap\partial_{\mathrm{par}}Q_{R,S},\qquad j\ge j_\delta. w − ≤ g j ≤ w + on U ∩ ∂ par Q R , S , j ≥ j δ .
The neighborhood U U U and the index j δ j_\delta j δ can be chosen independently of K K K and Λ \Lambda Λ . Outside U U U , the function f ( x ) + ( t − t b ) 2 f(x)+(t-t_b)^2 f ( x ) + ( t − t b ) 2 has a strictly positive minimum on the closed parabolic boundary. We can therefore choose Λ \Lambda Λ and K K K sufficiently large in terms of M M M and the positive minimum above, independently of j j j , such that
w − ≤ g j = u j ≤ w + on ∂ p a r Q R , S , j ≥ j δ . w_- \le g_j=u_j\le w_+ \qquad\text{on } \partial_{\mathrm{par}}Q_{R,S},\qquad j\ge j_\delta. w − ≤ g j = u j ≤ w + on ∂ par Q R , S , j ≥ j δ .
Our next goal is to verify that w + w_+ w + is a supersolution to (56) . Since t , t b ∈ [ − S , 0 ] t,t_b\in[-S,0] t , t b ∈ [ − S , 0 ] , we have
∂ t w + = 2 Λ ( t − t b ) ≥ − 2 Λ S . \partial_t w_+=2\Lambda(t-t_b)\ge-2\Lambda S. ∂ t w + = 2Λ ( t − t b ) ≥ − 2Λ S .
Moreover, ∇ w + = K ∇ f \nabla w_+=K\nabla f ∇ w + = K ∇ f and ∇ 2 w + = K ∇ 2 f \nabla^2w_+=K\nabla^2f ∇ 2 w + = K ∇ 2 f . Keeping Λ \Lambda Λ fixed, it therefore suffices to choose K K K sufficiently large that
− w + 2 − p p − 1 D ξ a ε j ( K ∇ f ( x ) ) ⋅ ( K ∇ 2 f ( x ) ) ≥ 2 Λ S in Q R , S . -\frac{w_+^{2-p}}{p-1}D_\xi a_{\varepsilon_j}\bigl(K\nabla f(x)\bigr)\cdot\bigl(K\nabla^2f(x)\bigr)\ge2\Lambda S \qquad\text{in }Q_{R,S}. − p − 1 w + 2 − p D ξ a ε j ( K ∇ f ( x ) ) ⋅ ( K ∇ 2 f ( x ) ) ≥ 2Λ S in Q R , S .
This ensures that
∂ t w + − w + 2 − p p − 1 D ξ a ε j ( ∇ w + ) ⋅ ∇ 2 w + ≥ 0. \partial_t w_+-\frac{w_+^{2-p}}{p-1}D_\xi a_{\varepsilon_j}(\nabla w_+)\cdot\nabla^2w_+\ge0. ∂ t w + − p − 1 w + 2 − p D ξ a ε j ( ∇ w + ) ⋅ ∇ 2 w + ≥ 0.
If 1 < p ≤ 2 1<p\le2 1 < p ≤ 2 , then w + ≥ u ( x b , t b ) ≥ m w_+\ge u(x_b,t_b)\ge m w + ≥ u ( x b , t b ) ≥ m , and hence
w + 2 − p p − 1 ≥ m 2 − p p − 1 > 0. \frac{w_+^{2-p}}{p-1}\ge\frac{m^{2-p}}{p-1}>0. p − 1 w + 2 − p ≥ p − 1 m 2 − p > 0.
Together with (62) and the uniform growth in K K K established above, this allows us, after fixing Λ \Lambda Λ , to choose K K K sufficiently large so that the spatial contribution is at least 2 Λ S 2\Lambda S 2Λ S throughout Q R , S Q_{R,S} Q R , S , uniformly in j j j . Thus, in this case, w + w_+ w + is a supersolution to (56) throughout Q R , S Q_{R,S} Q R , S for every j ∈ N j\in\mathbb{N} j ∈ N . If p > 2 p>2 p > 2 , then at every point where w + ≤ 2 M w_+\le2M w + ≤ 2 M we have
w + 2 − p p − 1 ≥ ( 2 M ) 2 − p p − 1 > 0. \frac{w_+^{2-p}}{p-1}\ge\frac{(2M)^{2-p}}{p-1}>0. p − 1 w + 2 − p ≥ p − 1 ( 2 M ) 2 − p > 0.
With Λ \Lambda Λ fixed, the same argument therefore shows that, for sufficiently large K K K , the supersolution inequality holds at all such points, uniformly in j j j . Consequently, truncating from above at the constant solution 2 M 2M 2 M , we obtain that min { w + , 2 M } \min\{w_+,2M\} min { w + , 2 M } is a weak supersolution to the regularized divergence equation (55) throughout Q R , S Q_{R,S} Q R , S , for every j ∈ N j\in\mathbb{N} j ∈ N . Similarly, after increasing K K K if necessary, w − w_- w − satisfies the subsolution inequality for (56) wherever w − > 0 w_->0 w − > 0 if p ≥ 2 p\ge2 p ≥ 2 , and wherever w − > 1 2 m w_->\frac{1}{2}m w − > 2 1 m if 1 < p < 2 1<p<2 1 < p < 2 . For p > 2 p>2 p > 2 , the positivity restriction is needed because the coefficient contains the negative power w − 2 − p w_-^{2-p} w − 2 − p . For p = 2 p=2 p = 2 , this coefficient equals one, so the restriction can be omitted. Consequently, for every p > 1 p>1 p > 1 , the truncated function max { w − , 1 2 m } \max\{w_-,\frac{1}{2}m\} max { w − , 2 1 m } is a weak subsolution to (55) throughout Q R , S Q_{R,S} Q R , S , for every j ∈ N j\in\mathbb{N} j ∈ N . We apply the comparison principle for (55) with max { w − , 1 2 m } \max\{w_-,\frac{1}{2}m\} max { w − , 2 1 m } as the subsolution. As the supersolution, we use w + w_+ w + if 1 < p ≤ 2 1<p\le2 1 < p ≤ 2 , and min { w + , 2 M } \min\{w_+,2M\} min { w + , 2 M } if p > 2 p>2 p > 2 . Both comparison functions are bounded below by 1 2 m > 0 \frac{1}{2}m>0 2 1 m > 0 . The previously established boundary ordering, together with 1 2 m ≤ u j ≤ 2 M \frac{1}{2}m\le u_j\le2M 2 1 m ≤ u j ≤ 2 M , ensures that these comparison functions bound u j u_j u j on the parabolic boundary for every j ≥ j δ j\ge j_\delta j ≥ j δ . The comparison principle therefore yields
w − ≤ u j ≤ w + in Q R , S , j ≥ j δ . (63) w_-\le u_j\le w_+ \qquad\text{in }Q_{R,S},\qquad j\ge j_\delta. \tag*{(63)} w − ≤ u j ≤ w + in Q R , S , j ≥ j δ . (63)
Keeping δ \delta δ , Λ \Lambda Λ , and K \mathbf{K} K fixed, we now pass to the locally uniform limit as j → ∞ j \to\infty j → ∞ . Since the barriers are independent of j j j , this gives
u ( x b , t b ) − [ 2 δ + K f ( x ) + Λ ( t − t b ) ] 2 ≤ w ( x , t ) ≤ u ( x b , t b ) + [ 2 δ + K f ( x ) + Λ ( t − t b ) ] 2 u(x_b,t_b)-\left[2\delta+\mathbf{K}f(x)+\Lambda(t-t_b)\right]^2
\le w(x,t)\le u(x_b,t_b)+\left[2\delta+\mathbf{K}f(x)+\Lambda(t-t_b)\right]^2 u ( x b , t b ) − [ 2 δ + K f ( x ) + Λ ( t − t b ) ] 2 ≤ w ( x , t ) ≤ u ( x b , t b ) + [ 2 δ + K f ( x ) + Λ ( t − t b ) ] 2
for every ( x , t ) ∈ Q R , S (x,t)\in Q_{R,S} ( x , t ) ∈ Q R , S . Since f ( x b ) = 0 f(x_b)=0 f ( x b ) = 0 , letting first ( x , t ) → ( x b , t b ) (x,t)\to(x_b,t_b) ( x , t ) → ( x b , t b ) and then δ ↓ 0 \delta\downarrow0 δ ↓ 0 gives
w ( x , t ) → u ( x b , t b ) as Q R , S ∋ ( x , t ) → ( x b , t b ) . w(x,t)\to u(x_b,t_b)\qquad\text{as }Q_{R,S}\ni(x,t)\to(x_b,t_b). w ( x , t ) → u ( x b , t b ) as Q R , S ∋ ( x , t ) → ( x b , t b ) .
Thus, w w w attains the boundary values u u u on ∂ B R × [ − S , 0 ] \partial B_R\times[-S,0] ∂ B R × [ − S , 0 ] .
Next, fix an initial boundary point ( x ∗ , − S ) (x_*, -S) ( x ∗ , − S ) with x ∗ ∈ B R x_*\in B_R x ∗ ∈ B R . For parameters δ ∈ ( 0 , 1 ) \delta\in(0,1) δ ∈ ( 0 , 1 ) and K , Λ > 0 \mathbf{K},\Lambda>0 K , Λ > 0 , we define the barriers
w ± ( x , t ) : = u ( x ∗ , − S ) ± [ 2 δ + K ∣ x − x ∗ ∣ q + Λ ( t + S ) ] , ( x , t ) ∈ Q R , S , w_\pm(x,t):=u(x_*,-S)\pm\left[2\delta+\mathbf{K}|x-x_*|^q+\Lambda(t+S)\right],\qquad(x,t)\in Q_{R,S}, w ± ( x , t ) := u ( x ∗ , − S ) ± [ 2 δ + K ∣ x − x ∗ ∣ q + Λ ( t + S ) ] , ( x , t ) ∈ Q R , S ,
where q : = max { 2 , p ′ } q:=\max\{2,p'\} q := max { 2 , p ′ } and p ′ = p p − 1 p'=\frac{p}{p-1} p ′ = p − 1 p . We next verify that D ξ a ε j ( ∇ w ± ) ⋅ ∇ 2 w ± D_\xi a_{\varepsilon_j}(\nabla w_\pm)\cdot\nabla^2w_\pm D ξ a ε j ( ∇ w ± ) ⋅ ∇ 2 w ± is bounded in Q R , S Q_{R,S} Q R , S , uniformly in j j j . If p ≥ 2 p\ge2 p ≥ 2 , then q = 2 q=2 q = 2 , and hence
∇ w ± = ± 2 K ( x − x ∗ ) , ∇ 2 w ± = ± 2 K I . \nabla w_\pm=\pm2\mathbf{K}(x-x_*),\qquad\nabla^2w_\pm=\pm2\mathbf{K}I. ∇ w ± = ± 2 K ( x − x ∗ ) , ∇ 2 w ± = ± 2 K I .
Since ∣ x − x ∗ ∣ ≤ 2 R |x-x_*|\le2R ∣ x − x ∗ ∣ ≤ 2 R and 0 < ε j < 1 0<\varepsilon_j<1 0 < ε j < 1 , the formula for D ξ a ε j D_\xi a_{\varepsilon_j} D ξ a ε j gives a bound depending only on N N N , p p p , R R R , K \mathbf{K} K . If 1 < p < 2 1<p<2 1 < p < 2 , then q = p ′ > 2 q=p'>2 q = p ′ > 2 . For x ≠ x ∗ x\ne x_* x = x ∗ , the same formula and p − 2 < 0 p-2<0 p − 2 < 0 give
∣ D ξ a ε j ( ∇ w ± ) ⋅ ∇ 2 w ± ∣ ≤ C ∣ ∇ w ± ∣ p − 2 ∣ ∇ 2 w ± ∣ ≤ C K p − 1 ∣ x − x ∗ ∣ ( q − 1 ) ( p − 2 ) + q − 2 = C K p − 1 . \left|D_\xi a_{\varepsilon_j}(\nabla w_\pm)\cdot\nabla^2w_\pm\right|
\le C|\nabla w_\pm|^{p-2}|\nabla^2w_\pm|
\le C\mathbf{K}^{p-1}|x-x_*|^{(q-1)(p-2)+q-2}
=C\mathbf{K}^{p-1}. D ξ a ε j ( ∇ w ± ) ⋅ ∇ 2 w ± ≤ C ∣∇ w ± ∣ p − 2 ∣ ∇ 2 w ± ∣ ≤ C K p − 1 ∣ x − x ∗ ∣ ( q − 1 ) ( p − 2 ) + q − 2 = C K p − 1 .
where C C C depends only on N N N , p p p . Here we used
( q − 1 ) ( p − 2 ) + q − 2 = ( q − 1 ) ( p − 1 ) − 1 = 0. (q-1)(p-2)+q-2=(q-1)(p-1)-1=0. ( q − 1 ) ( p − 2 ) + q − 2 = ( q − 1 ) ( p − 1 ) − 1 = 0.
The estimate extends to x = x ∗ x=x_* x = x ∗ by continuity, since ε j > 0 \varepsilon_j>0 ε j > 0 and w ± w_\pm w ± is twice continuously differentiable in space. Thus, in both cases, the bound depends only on N N N , p p p , R R R , K \mathbf{K} K and is independent of j j j and Λ \Lambda Λ .
Fix K \mathbf{K} K as above, and let C = C ( N , p , R , K ) C=C(N,p,R,\mathbf{K}) C = C ( N , p , R , K ) be the uniform bound for the spatial contraction obtained above. If p ≥ 2 p\ge2 p ≥ 2 , then w + ≥ m w_+\ge\mathbf{m} w + ≥ m gives
w + 2 − p p − 1 ≤ m 2 − p p − 1 . \frac{w_+^{2-p}}{p-1}\le\frac{\mathbf{m}^{2-p}}{p-1}. p − 1 w + 2 − p ≤ p − 1 m 2 − p .
Choosing Λ > C m 2 − p / ( p − 1 ) \Lambda>C\mathbf{m}^{2-p}/(p-1) Λ > C m 2 − p / ( p − 1 ) therefore makes w + w_+ w + a supersolution to (56) in Q R , S Q_{R,S} Q R , S for every j ∈ N j\in\mathbb{N} j ∈ N . If 1 < p < 2 1<p<2 1 < p < 2 , we have
w + ( x , t ) ≤ M + 2 + K ( 2 R ) q + Λ S . w_+(x,t)\le\mathbf{M}+2+\mathbf{K}(2R)^q+\Lambda S. w + ( x , t ) ≤ M + 2 + K ( 2 R ) q + Λ S .
Since 0 < 2 − p < 1 0<2-p<1 0 < 2 − p < 1 , it follows that
lim Λ → ∞ C ( M + 2 + K ( 2 R ) q + Λ S ) 2 − p ( p − 1 ) Λ = 0. \lim_{\Lambda\to\infty}\frac{C\left(\mathbf{M}+2+\mathbf{K}(2R)^q+\Lambda S\right)^{2-p}}{(p-1)\Lambda}=0. Λ → ∞ lim ( p − 1 ) Λ C ( M + 2 + K ( 2 R ) q + Λ S ) 2 − p = 0.
Choosing Λ \Lambda Λ sufficiently large therefore makes w + w_+ w + a supersolution to (56) in Q R , S Q_{R,S} Q R , S for every j ∈ N j\in\mathbb{N} j ∈ N .
Since ∂ t w − = − Λ \partial_t w_-=-\Lambda ∂ t w − = − Λ , the bound for the spatial term yields the subsolution inequality for sufficiently large Λ \Lambda Λ wherever w − 2 − p w_-^{2-p} w − 2 − p is uniformly bounded. For p > 2 p>2 p > 2 , this factor may diverge as w − ↓ 0 w_-\downarrow0 w − ↓ 0 . We therefore work on { w − > 1 2 m } \{w_->\frac12\mathbf{m}\} { w − > 2 1 m } , where 1 2 m < w − ≤ M \frac12\mathbf{m}<w_-\le\mathbf{M} 2 1 m < w − ≤ M provides a uniform bound for every p > 1 p>1 p > 1 . Truncating from below at the constant solution 1 2 m \frac12\mathbf{m} 2 1 m then gives the weak subsolution max { w − , 1 2 m } \max\{w_-,\frac12\mathbf{m}\} max { w − , 2 1 m } .
To apply the comparison principle with these barriers, we check that the previously established boundary ordering remains valid. Since t + S ≥ 0 t+S\ge0 t + S ≥ 0 , increasing Λ \Lambda Λ decreases w − w_- w − and increases w + w_+ w + . Moreover, the lower truncation cannot exceed the boundary values, because g j = u j ≥ 1 2 m g_j=u_j\ge\frac{1}{2}\mathbf{m} g j = u j ≥ 2 1 m there. Consequently,
max { w − , 1 2 m } ≤ g j = u j ≤ w + on ∂ p a r Q R , S , j ≥ j δ . \max\left\{w_-,\frac{1}{2}\mathbf{m}\right\}\le g_j=u_j\le w_+\quad\text{on }\partial_{\mathrm{par}}Q_{R,S},\qquad j\ge j_\delta. max { w − , 2 1 m } ≤ g j = u j ≤ w + on ∂ par Q R , S , j ≥ j δ .
The comparison principle for (55) gives w − ≤ u j ≤ w + w_-\le u_j\le w_+ w − ≤ u j ≤ w + in Q R , S Q_{R,S} Q R , S for every j ≥ j δ j\ge j_\delta j ≥ j δ . Keeping δ \delta δ , K \mathbf{K} K , and Λ \Lambda Λ fixed and passing to the locally uniform limit as j → ∞ j\to\infty j → ∞ , we obtain
∣ w ( x , t ) − u ( x ∗ , − S ) ∣ ≤ 2 δ + K ∣ x − x ∗ ∣ q + Λ ( t + S ) for every ( x , t ) ∈ Q R , S . |w(x,t)-u(x_*, -S)|\le2\delta+\mathbf{K}|x-x_*|^q+\Lambda(t+S)\qquad\text{for every }(x,t)\in Q_{R,S}. ∣ w ( x , t ) − u ( x ∗ , − S ) ∣ ≤ 2 δ + K ∣ x − x ∗ ∣ q + Λ ( t + S ) for every ( x , t ) ∈ Q R , S .
Letting first ( x , t ) → ( x ∗ , − S ) (x,t)\to(x_*,-S) ( x , t ) → ( x ∗ , − S ) within Q R , S Q_{R,S} Q R , S and then δ ↓ 0 \delta\downarrow0 δ ↓ 0 , we conclude that
w ( x , t ) → u ( x ∗ , − S ) as Q R , S ∋ ( x , t ) → ( x ∗ , − S ) . w(x,t)\to u(x_*,-S)\qquad\text{as }Q_{R,S}\ni(x,t)\to(x_*,-S). w ( x , t ) → u ( x ∗ , − S ) as Q R , S ∋ ( x , t ) → ( x ∗ , − S ) .
Thus, w w w attains the prescribed initial values on B R × { − S } B_R\times\{-S\} B R × { − S } . Together with the lateral boundary limits, this shows that w w w extends continuously to ∂ p a r Q R , S \partial_{\mathrm{par}}Q_{R,S} ∂ par Q R , S with boundary values u u u , completing the identification of the initial and boundary values of the limit map.
Step 4: Identification of the limit map. It remains to show that w w w is a weak solution of (55) in Q R , S Q_{R,S} Q R , S and to conclude w = u w=u w = u . Fix a cylinder Q r , s ⋐ Q R , S Q_{r,s}\Subset Q_{R,S} Q r , s ⋐ Q R , S . Since this cylinder is arbitrary, it suffices to verify the weak formulation of (55) in Q r , s Q_{r,s} Q r , s . By the gradient bound obtained in Step 2 from Proposition 4.2 , we have
sup j ∈ N ∥ ∇ u j ∥ L ∞ ( Q r , s , R N ) < ∞ . \sup_{j\in\mathbb{N}}\|\nabla u_j\|_{L^\infty(Q_{r,s},\mathbb{R}^N)}<\infty. j ∈ N sup ∥∇ u j ∥ L ∞ ( Q r , s , R N ) < ∞.
Since 0 < ε j < 1 0<\varepsilon_j<1 0 < ε j < 1 and
∣ a ε j ( ∇ u j ) ∣ ≤ ( 1 + ∣ ∇ u j ∣ ) p − 1 , |a_{\varepsilon_j}(\nabla u_j)|\le(1+|\nabla u_j|)^{p-1}, ∣ a ε j ( ∇ u j ) ∣ ≤ ( 1 + ∣∇ u j ∣ ) p − 1 ,
the preceding gradient bound also gives
sup j ∈ N ∥ a ε j ( ∇ u j ) ∥ L ∞ ( Q r , s , R N ) < ∞ . \sup_{j\in\mathbb{N}}\|a_{\varepsilon_j}(\nabla u_j)\|_{L^\infty(Q_{r,s},\mathbb{R}^N)}<\infty. j ∈ N sup ∥ a ε j ( ∇ u j ) ∥ L ∞ ( Q r , s , R N ) < ∞.
By weak compactness on Q r , s Q_{r,s} Q r , s , we can pass to a subsequence, still indexed by j j j , such that
∇ u j ⇀ ∇ w in L p ( Q r , s , R N ) , a ε j ( ∇ u j ) ⇀ a o in L p ′ ( Q r , s , R N ) . (64) \begin{aligned}
\nabla u_j&\rightharpoonup\nabla w &&\text{in }L^p(Q_{r,s},\mathbb{R}^N),\\
a_{\varepsilon_j}(\nabla u_j)&\rightharpoonup a_o &&\text{in }L^{p'}(Q_{r,s},\mathbb{R}^N). \tag*{(64)}
\end{aligned} ∇ u j a ε j ( ∇ u j ) ⇀ ∇ w ⇀ a o in L p ( Q r , s , R N ) , in L p ′ ( Q r , s , R N ) . (64)
as j → ∞ j\to\infty j → ∞ , for some vector field a o ∈ L p ′ ( Q r , s , R N ) a_o\in L^{p'}(Q_{r,s},\mathbb{R}^N) a o ∈ L p ′ ( Q r , s , R N ) . The first weak limit is ∇ w \nabla w ∇ w because the locally uniform convergence u j → w u_j\to w u j → w implies that the spatial gradients converge to ∇ w \nabla w ∇ w in the sense of distributions. Since u j p − 1 → w p − 1 u_j^{p-1}\to w^{p-1} u j p − 1 → w p − 1 locally uniformly in Q r , s Q_{r,s} Q r , s and a ε j ( ∇ u j ) ⇀ a o a_{\varepsilon_j}(\nabla u_j)\rightharpoonup a_o a ε j ( ∇ u j ) ⇀ a o in L p ′ ( Q r , s , R N ) L^{p'}(Q_{r,s},\mathbb{R}^N) L p ′ ( Q r , s , R N ) by (5.8), we may pass to the limit in the weak formulation of (55) to obtain
∂ t ( w p − 1 ) = div a o weakly in Q r , s . (65) \partial_t(w^{p-1})=\operatorname{div}a_o\qquad\text{weakly in }Q_{r,s}. \tag*{(65)} ∂ t ( w p − 1 ) = div a o weakly in Q r , s . (65)
To conclude that w w w is a weak solution of (1) in Q r , s Q_{r,s} Q r , s , it remains to show that
a o = ∣ ∇ w ∣ p − 2 ∇ w almost everywhere in Q r , s . a_o=|\nabla w|^{p-2}\nabla w\qquad\text{almost everywhere in }Q_{r,s}. a o = ∣∇ w ∣ p − 2 ∇ w almost everywhere in Q r , s .
Let φ ∈ C 0 ∞ ( Q r , s ) \varphi\in C^\infty_0(Q_{r,s}) φ ∈ C 0 ∞ ( Q r , s ) be nonnegative, and let t − t_- t − denote the lower endpoint of the time interval of Q r , s Q_{r,s} Q r , s . For h > 0 h>0 h > 0 , we define the exponential time mollification of w w w by
⟦ w ⟧ h ( x , t ) : = 1 h ∫ t − t e τ − t h w ( x , τ ) d τ , ( x , t ) ∈ Q r , s . \llbracket w\rrbracket_h(x,t):=\frac{1}{h}\int_{t_-}^{t}e^{\frac{\tau-t}{h}}w(x,\tau)\,d\tau,\qquad(x,t)\in Q_{r,s}. [ [ w ] ] h ( x , t ) := h 1 ∫ t − t e h τ − t w ( x , τ ) d τ , ( x , t ) ∈ Q r , s .
Differentiating with respect to t t t , we obtain
∂ t ⟦ w ⟧ h = w − ⟦ w ⟧ h h . (66) \partial_t\llbracket w\rrbracket_h=\frac{w-\llbracket w\rrbracket_h}{h}. \tag*{(66)} ∂ t [ [ w ] ] h = h w − [ [ w ] ] h . (66)
For fixed h > 0 h>0 h > 0 , this time derivative is bounded because w w w is bounded. Moreover, ∇ ⟦ w ⟧ h = ⟦ ∇ w ⟧ h \nabla\llbracket w\rrbracket_h=\llbracket\nabla w\rrbracket_h ∇ [ [ w ] ] h = [ [ ∇ w ] ] h belongs to L p ( Q r , s , R N ) L^p(Q_{r,s},\mathbb{R}^N) L p ( Q r , s , R N ) . Since φ \varphi φ has compact support in Q r , s Q_{r,s} Q r , s , we can test (65) with φ [ w ] h \varphi[w]_h φ [ w ] h . Expanding the time derivative and the spatial gradient of this product gives
∬ Q r , s w p − 1 ∂ t [ w ] h φ d x d t + ∬ Q r , s w p − 1 [ w ] h ∂ t φ d x d t = ∬ Q r , s φ a o ⋅ ∇ [ w ] h d x d t + ∬ Q r , s [ w ] h a o ⋅ ∇ φ d x d t . \begin{aligned}
\iint_{Q_{r,s}} w^{p-1}\partial_t[w]_h\,\varphi\,\mathrm{d}x\,\mathrm{d}t+\iint_{Q_{r,s}}w^{p-1}[w]_h\,\partial_t\varphi\,\mathrm{d}x\,\mathrm{d}t\\
=\iint_{Q_{r,s}}\varphi a_o\cdot\nabla[w]_h\,\mathrm{d}x\,\mathrm{d}t+\iint_{Q_{r,s}}[w]_h a_o\cdot\nabla\varphi\,\mathrm{d}x\,\mathrm{d}t.
\end{aligned} ∬ Q r , s w p − 1 ∂ t [ w ] h φ d x d t + ∬ Q r , s w p − 1 [ w ] h ∂ t φ d x d t = ∬ Q r , s φ a o ⋅ ∇ [ w ] h d x d t + ∬ Q r , s [ w ] h a o ⋅ ∇ φ d x d t .
In the first integral on the left-hand side, we add and subtract [ w ] h p − 1 [w]_h^{p-1} [ w ] h p − 1 and use (66) . This gives
∬ Q r , s w p − 1 ∂ t [ w ] h φ d x d t = ∬ Q r , s [ w ] h p − 1 ∂ t [ w ] h φ d x d t + 1 h ∬ Q r , s ( w p − 1 − [ w ] h p − 1 ) ( w − [ w ] h ) φ d x d t ≥ ∬ Q r , s [ w ] h p − 1 ∂ t [ w ] h φ d x d t = − 1 p ∬ Q r , s [ w ] h p ∂ t φ d x d t . \begin{aligned}
\iint_{Q_{r,s}}w^{p-1}\partial_t[w]_h\,\varphi\,\mathrm{d}x\,\mathrm{d}t
={}&\iint_{Q_{r,s}}[w]_h^{p-1}\partial_t[w]_h\,\varphi\,\mathrm{d}x\,\mathrm{d}t\\
&+\frac{1}{h}\iint_{Q_{r,s}}\left(w^{p-1}-[w]_h^{p-1}\right)(w-[w]_h)\varphi\,\mathrm{d}x\,\mathrm{d}t\\
\ge{}&\iint_{Q_{r,s}}[w]_h^{p-1}\partial_t[w]_h\,\varphi\,\mathrm{d}x\,\mathrm{d}t
=-\frac{1}{p}\iint_{Q_{r,s}}[w]_h^p\partial_t\varphi\,\mathrm{d}x\,\mathrm{d}t.
\end{aligned} ∬ Q r , s w p − 1 ∂ t [ w ] h φ d x d t = ≥ ∬ Q r , s [ w ] h p − 1 ∂ t [ w ] h φ d x d t + h 1 ∬ Q r , s ( w p − 1 − [ w ] h p − 1 ) ( w − [ w ] h ) φ d x d t ∬ Q r , s [ w ] h p − 1 ∂ t [ w ] h φ d x d t = − p 1 ∬ Q r , s [ w ] h p ∂ t φ d x d t .
The indicated product is nonnegative because s ↦ s p − 1 s\mapsto s^{p-1} s ↦ s p − 1 is increasing on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) . Together with φ ≥ 0 \varphi\ge0 φ ≥ 0 , this gives the inequality. In the last equality, we integrated by parts in time. Substituting this lower bound into the preceding identity and moving the term containing ∇ φ \nabla\varphi ∇ φ to the right-hand side, we obtain
∬ Q r , s φ a o ⋅ ∇ [ w ] h d x d t ≥ ∬ Q r , s ( w p − 1 [ w ] h − 1 p [ w ] h p ) ∂ t φ d x d t − ∬ Q r , s [ w ] h a o ⋅ ∇ φ d x d t . \iint_{Q_{r,s}}\varphi a_o\cdot\nabla[w]_h\,\mathrm{d}x\,\mathrm{d}t
\ge\iint_{Q_{r,s}}\left(w^{p-1}[w]_h-\frac{1}{p}[w]_h^p\right)\partial_t\varphi\,\mathrm{d}x\,\mathrm{d}t-\iint_{Q_{r,s}}[w]_h a_o\cdot\nabla\varphi\,\mathrm{d}x\,\mathrm{d}t. ∬ Q r , s φ a o ⋅ ∇ [ w ] h d x d t ≥ ∬ Q r , s ( w p − 1 [ w ] h − p 1 [ w ] h p ) ∂ t φ d x d t − ∬ Q r , s [ w ] h a o ⋅ ∇ φ d x d t .
As h ↓ 0 h\downarrow0 h ↓ 0 , we have
[ w ] h → w in L p ( Q r , s ) , ∇ [ w ] h → ∇ w in L p ( Q r , s , R N ) , [w]_h\to w\quad\text{in }L^p(Q_{r,s}),\qquad\nabla[w]_h\to\nabla w\quad\text{in }L^p(Q_{r,s},\mathbb{R}^N), [ w ] h → w in L p ( Q r , s ) , ∇ [ w ] h → ∇ w in L p ( Q r , s , R N ) ,
with strong convergence in both spaces. In particular, the coefficient of ∂ t φ \partial_t\varphi ∂ t φ converges in L 1 ( Q r , s ) L^1(Q_{r,s}) L 1 ( Q r , s ) to p − 1 p w p \frac{p-1}{p}w^p p p − 1 w p . Since a o ∈ L p ′ ( Q r , s , R N ) a_o\in L^{p'}(Q_{r,s},\mathbb{R}^N) a o ∈ L p ′ ( Q r , s , R N ) , we may also pass to the limit in the two integrals involving a o a_o a o . We conclude that
∬ Q r , s φ a o ⋅ ∇ w d x d t ≥ p − 1 p ∬ Q r , s w p ∂ t φ d x d t − ∬ Q r , s w a o ⋅ ∇ φ d x d t . (67) \iint_{Q_{r,s}}\varphi a_o\cdot\nabla w\,\mathrm{d}x\,\mathrm{d}t\ge\frac{p-1}{p}\iint_{Q_{r,s}}w^p\partial_t\varphi\,\mathrm{d}x\,\mathrm{d}t-\iint_{Q_{r,s}}w a_o\cdot\nabla\varphi\,\mathrm{d}x\,\mathrm{d}t. \tag*{(67)} ∬ Q r , s φ a o ⋅ ∇ w d x d t ≥ p p − 1 ∬ Q r , s w p ∂ t φ d x d t − ∬ Q r , s w a o ⋅ ∇ φ d x d t . (67)
for every nonnegative φ ∈ C 0 ∞ ( Q r , s ) \varphi\in C^\infty_0(Q_{r,s}) φ ∈ C 0 ∞ ( Q r , s ) . Since u j u_j u j is smooth, testing (55) with φ u j \varphi u_j φ u j and integrating by parts gives
∬ Q r , s φ a ε j ( ∇ u j ) ⋅ ∇ u j d x d t = p − 1 p ∬ Q r , s u j p ∂ t φ d x d t − ∬ Q r , s u j a ε j ( ∇ u j ) ⋅ ∇ φ d x d t . (68) \begin{aligned}
\iint_{Q_{r,s}}\varphi a_{\varepsilon_j}(\nabla u_j)\cdot\nabla u_j\,\mathrm{d}x\,\mathrm{d}t \\
&=\frac{p-1}{p}\iint_{Q_{r,s}}u_j^p\partial_t\varphi\,\mathrm{d}x\,\mathrm{d}t-\iint_{Q_{r,s}}u_j a_{\varepsilon_j}(\nabla u_j)\cdot\nabla\varphi\,\mathrm{d}x\,\mathrm{d}t. \tag*{(68)}
\end{aligned} ∬ Q r , s φ a ε j ( ∇ u j ) ⋅ ∇ u j d x d t = p p − 1 ∬ Q r , s u j p ∂ t φ d x d t − ∬ Q r , s u j a ε j ( ∇ u j ) ⋅ ∇ φ d x d t . (68)
Since Q r , s ⋐ Q R , S Q_{r,s}\Subset Q_{R,S} Q r , s ⋐ Q R , S , the locally uniform convergence u j → w u_j\to w u j → w implies that u j p → w p u_j^p\to w^p u j p → w p uniformly on Q r , s Q_{r,s} Q r , s . Moreover, u j ∇ φ → w ∇ φ u_j\nabla\varphi\to w\nabla\varphi u j ∇ φ → w ∇ φ strongly in L p ( Q r , s , R N ) L^{p}(Q_{r,s},\mathbb{R}^N) L p ( Q r , s , R N ) . Together with the weak convergence of a ε j ( ∇ u j ) a_{\varepsilon_j}(\nabla u_j) a ε j ( ∇ u j ) in (5.8), this allows us to pass to the limit on the right-hand side of (68) . We obtain
lim j → ∞ ∬ Q r , s φ a ε j ( ∇ u j ) ⋅ ∇ u j d x d t = p − 1 p ∬ Q r , s w p ∂ t φ d x d t − ∬ Q r , s w a o ⋅ ∇ φ d x d t . ≤ ∬ Q r , s φ a o ⋅ ∇ w d x d t , \begin{aligned}
\lim_{j\to\infty}\iint_{Q_{r,s}}\varphi a_{\varepsilon_j}(\nabla u_j)\cdot\nabla u_j\,\mathrm{d}x\,\mathrm{d}t
=\frac{p-1}{p}\iint_{Q_{r,s}}w^p\partial_t\varphi\,\mathrm{d}x\,\mathrm{d}t-\iint_{Q_{r,s}}w a_o\cdot\nabla\varphi\,\mathrm{d}x\,\mathrm{d}t. \\
&\le\iint_{Q_{r,s}} \varphi\mathbf{a}_o \cdot\nabla w\,dx\,dt,
\end{aligned} j → ∞ lim ∬ Q r , s φ a ε j ( ∇ u j ) ⋅ ∇ u j d x d t = p p − 1 ∬ Q r , s w p ∂ t φ d x d t − ∬ Q r , s w a o ⋅ ∇ φ d x d t . ≤ ∬ Q r , s φ a o ⋅ ∇ w d x d t ,
where the last inequality follows from (67) since φ ≥ 0 \varphi\ge0 φ ≥ 0 . Fix ξ ∈ L p ( Q r , s , R N ) \xi\in L^p(Q_{r,s},\mathbb{R}^N) ξ ∈ L p ( Q r , s , R N ) and a nonnegative φ ∈ C 0 ∞ ( Q r , s ) \varphi\in C^\infty_0(Q_{r,s}) φ ∈ C 0 ∞ ( Q r , s ) . Since a ε j ( ξ ) → ∣ ξ ∣ p − 2 ξ \mathbf{a}_{\varepsilon_j}(\xi) \to|\xi|^{p-2}\xi a ε j ( ξ ) → ∣ ξ ∣ p − 2 ξ almost everywhere and ∣ a ε j ( ξ ) ∣ ≤ ( 1 + ∣ ξ ∣ ) p − 1 |\mathbf{a}_{\varepsilon_j}(\xi)| \le(1+|\xi|)^{p-1} ∣ a ε j ( ξ ) ∣ ≤ ( 1 + ∣ ξ ∣ ) p − 1 , dominated convergence gives a ε j ( ξ ) → ∣ ξ ∣ p − 2 ξ \mathbf{a}_{\varepsilon_j}(\xi) \to|\xi|^{p-2}\xi a ε j ( ξ ) → ∣ ξ ∣ p − 2 ξ strongly in L p ′ ( Q r , s , R N ) L^{p'}(Q_{r,s},\mathbb{R}^N) L p ′ ( Q r , s , R N ) . Combining this convergence with (5.8) and the preceding estimate for the term containing a ε j ( ∇ u j ) ⋅ ∇ u j \mathbf{a}_{\varepsilon_j}(\nabla u_j)\cdot\nabla u_j a ε j ( ∇ u j ) ⋅ ∇ u j , we obtain
∬ Q r , s φ ( a o − ∣ ξ ∣ p − 2 ξ ) ⋅ ( ∇ w − ξ ) d x d t ≥ lim j → ∞ ∬ Q r , s φ ( a ε j ( ∇ u j ) − a ε j ( ξ ) ) ⋅ ( ∇ u j − ξ ) d x d t ≥ 0. \iint_{Q_{r,s}} \varphi\left(\mathbf{a}_o-|\xi|^{p-2}\xi\right)\cdot(\nabla w-\xi)\,dx\,dt
\ge\lim_{j\to\infty}\iint_{Q_{r,s}}\varphi\left(\mathbf{a}_{\varepsilon_j}(\nabla u_j)-\mathbf{a}_{\varepsilon_j}(\xi)\right)\cdot(\nabla u_j-\xi)\,dx\,dt\ge0. ∬ Q r , s φ ( a o − ∣ ξ ∣ p − 2 ξ ) ⋅ ( ∇ w − ξ ) d x d t ≥ j → ∞ lim ∬ Q r , s φ ( a ε j ( ∇ u j ) − a ε j ( ξ ) ) ⋅ ( ∇ u j − ξ ) d x d t ≥ 0.
The last inequality follows from the monotonicity of a ε j \mathbf{a}_{\varepsilon_j} a ε j and the nonnegativity of φ \varphi φ . Let e ℓ e_\ell e ℓ denote the ℓ \ell ℓ -th standard basis vector of R N \mathbb{R}^N R N . For λ > 0 \lambda>0 λ > 0 , we apply the preceding inequality first with ξ = ∇ w − λ e ℓ \xi=\nabla w-\lambda e_\ell ξ = ∇ w − λ e ℓ and then with ξ = ∇ w + λ e ℓ \xi=\nabla w+\lambda e_\ell ξ = ∇ w + λ e ℓ . Dividing by λ \lambda λ gives
∬ Q r , s φ ∣ ∇ w − λ e ℓ ∣ p − 2 ( ∇ w − λ e ℓ ) ⋅ e ℓ d x d t ≤ ∬ Q r , s φ a o ⋅ e ℓ d x d t ≤ ∬ Q r , s φ ∣ ∇ w + λ e ℓ ∣ p − 2 ( ∇ w + λ e ℓ ) ⋅ e ℓ d x d t . \iint_{Q_{r,s}}\varphi|\nabla w-\lambda e_\ell|^{p-2}(\nabla w-\lambda e_\ell)\cdot e_\ell\,dx\,dt
\le\iint_{Q_{r,s}}\varphi\mathbf{a}_o\cdot e_\ell\,dx\,dt
\le\iint_{Q_{r,s}}\varphi|\nabla w+\lambda e_\ell|^{p-2}(\nabla w+\lambda e_\ell)\cdot e_\ell\,dx\,dt. ∬ Q r , s φ ∣∇ w − λ e ℓ ∣ p − 2 ( ∇ w − λ e ℓ ) ⋅ e ℓ d x d t ≤ ∬ Q r , s φ a o ⋅ e ℓ d x d t ≤ ∬ Q r , s φ ∣∇ w + λ e ℓ ∣ p − 2 ( ∇ w + λ e ℓ ) ⋅ e ℓ d x d t .
For 0 < λ ≤ 1 0<\lambda\le1 0 < λ ≤ 1 , the first and last integrands are bounded in absolute value by φ ( 1 + ∣ ∇ w ∣ ) p − 1 \varphi(1+|\nabla w|)^{p-1} φ ( 1 + ∣∇ w ∣ ) p − 1 , which is integrable on Q r , s Q_{r,s} Q r , s . By continuity of the map ζ ↦ ∣ ζ ∣ p − 2 ζ \zeta\mapsto|\zeta|^{p-2}\zeta ζ ↦ ∣ ζ ∣ p − 2 ζ , we may let λ ↓ 0 \lambda\downarrow0 λ ↓ 0 and apply dominated convergence to obtain
∬ Q r , s φ ( a o − ∣ ∇ w ∣ p − 2 ∇ w ) ⋅ e ℓ d x d t = 0. \iint_{Q_{r,s}}\varphi\left(\mathbf{a}_o-|\nabla w|^{p-2}\nabla w\right)\cdot e_\ell\,dx\,dt=0. ∬ Q r , s φ ( a o − ∣∇ w ∣ p − 2 ∇ w ) ⋅ e ℓ d x d t = 0.
Since this holds for every nonnegative φ ∈ C 0 ∞ ( Q r , s ) \varphi\in C^\infty_0(Q_{r,s}) φ ∈ C 0 ∞ ( Q r , s ) and every ℓ = 1 , … , N \ell=1,\ldots,N ℓ = 1 , … , N , we conclude that
a o = ∣ ∇ w ∣ p − 2 ∇ w \mathbf{a}_o=|\nabla w|^{p-2}\nabla w a o = ∣∇ w ∣ p − 2 ∇ w
almost everywhere in Q r , s Q_{r,s} Q r , s .
Substituting this identity into (65) shows that w w w is a weak solution of (1) in Q r , s Q_{r,s} Q r , s . Since Q r , s ⋐ Q R , S Q_{r,s}\Subset Q_{R,S} Q r , s ⋐ Q R , S was arbitrary, w w w is a weak solution in Q R , S Q_{R,S} Q R , S .
It remains to show that w = u w=u w = u . Since the energy bounds for w w w are local, we apply the comparison principle on the cylinders Q n : = B R − 1 n ( 0 ) × ( − S + 1 n , 1 n ) Q_n:=B_{R-\frac{1}{n}}(0)\times\left(-S+\frac{1}{n},\frac{1}{n}\right) Q n := B R − n 1 ( 0 ) × ( − S + n 1 , n 1 ) , where n ∈ N n\in\mathbb{N} n ∈ N is sufficiently large that 1 n < R \frac{1}{n}<R n 1 < R and 2 n < S \frac{2}{n}<S n 2 < S . Then Q n ⋐ Q R , S Q_n\Subset Q_{R,S} Q n ⋐ Q R , S and Q n ↑ Q R , S Q_n\uparrow Q_{R,S} Q n ↑ Q R , S . By Step 3, w − u w-u w − u extends continuously to ∂ p a r Q R , S \partial_{\mathrm{par}}Q_{R,S} ∂ par Q R , S and vanishes there. Since this boundary is compact and every point of ∂ p a r Q n \partial_{\mathrm{par}}Q_n ∂ par Q n lies within distance 1 / n 1/n 1/ n of it, we obtain
δ n : = sup ∂ p a r Q n ∣ w − u ∣ ⟶ 0 as n → ∞ . \delta_n:=\sup_{\partial_{\mathrm{par}}Q_n}|w-u|\longrightarrow0\qquad\text{as }n\to\infty. δ n := ∂ par Q n sup ∣ w − u ∣ ⟶ 0 as n → ∞.
For n n n sufficiently large, we have δ n < m \delta_n<m δ n < m . Since u ≥ m u\ge m u ≥ m , the definition of δ n \delta_n δ n gives
( 1 − δ n m ) u ≤ w ≤ ( 1 + δ n m ) u on ∂ p a r Q n . \left(1-\frac{\delta_n}{m}\right)u\le w\le\left(1+\frac{\delta_n}{m}\right)u
\qquad\text{on }\partial_{\mathrm{par}}Q_n. ( 1 − m δ n ) u ≤ w ≤ ( 1 + m δ n ) u on ∂ par Q n .
Both factors are positive. By the homogeneity of (1) , the two multiples of u u u are therefore weak solutions of (1) . Since Q n ⋐ Q R , S Q_n\Subset Q_{R,S} Q n ⋐ Q R , S , the comparison principle from [9 ] applies on Q n Q_n Q n and yields
( 1 − δ n m ) u ≤ w ≤ ( 1 + δ n m ) u in Q n . \left(1-\frac{\delta_n}{m}\right)u\le w\le\left(1+\frac{\delta_n}{m}\right)u
\qquad\text{in }Q_n. ( 1 − m δ n ) u ≤ w ≤ ( 1 + m δ n ) u in Q n .
Fix ( x , t ) ∈ Q R , S (x,t) \in Q_{R,S} ( x , t ) ∈ Q R , S . Then ( x , t ) ∈ Q n (x,t) \in Q_n ( x , t ) ∈ Q n for all sufficiently large n n n . Letting n → ∞ n \to\infty n → ∞ in the preceding inequalities gives w ( x , t ) = u ( x , t ) w(x,t)=u(x,t) w ( x , t ) = u ( x , t ) . Thus w = u w=u w = u in Q R , S Q_{R,S} Q R , S , which completes the proof. □ \square □
Proposition 5.2 (Positive cylinders). Let r , h > 0 r,h>0 r , h > 0 , and let u u u be a continuous weak solution to (1) in a neighborhood of B 2 r ( x o ) × [ s − 2 h , s ] B_{2r}(x_o) \times[s-2h,s] B 2 r ( x o ) × [ s − 2 h , s ] . Suppose that
0 < m ≤ u ≤ M in B 2 r ( x o ) × ( s − 2 h , s ] . 0<m\leq u\leq M \qquad\text{in } B_{2r}(x_o)\times(s-2h,s]. 0 < m ≤ u ≤ M in B 2 r ( x o ) × ( s − 2 h , s ] . Then, for every k ≥ k o k\geq k_o k ≥ k o , with k o k_o k o defined in (9) , we have
∥ ∇ ( u 1 k ) ∥ L ∞ ( B r ( x o ) × ( s − h , s ) ) ≤ C M 1 k ( 1 r + 1 h 1 p ) . (69) \left\lVert\nabla\left(u^{\frac{1}{k}}\right)\right\rVert_{L^\infty(B_r(x_o)\times(s-h,s))}\leq CM^{\frac{1}{k}}\left(\frac{1}{r}+\frac{1}{h^{\frac{1}{p}}}\right). \tag*{(69)} ∇ ( u k 1 ) L ∞ ( B r ( x o ) × ( s − h , s )) ≤ C M k 1 ( r 1 + h p 1 1 ) . (69) where C = C ( N , p , k ) C=C(N,p,k) C = C ( N , p , k ) is independent of m m m , M M M , r r r , h h h .
Proof. By Lemma 5.1, there exist ε j ↓ 0 \varepsilon_j\downarrow0 ε j ↓ 0 and smooth solutions u j u_j u j to (55) in B 2 r ( x o ) × ( s − 2 h , s ] B_{2r}(x_o)\times(s-2h,s] B 2 r ( x o ) × ( s − 2 h , s ] such that 1 2 m ≤ u j ≤ 2 M \frac{1}{2}m\leq u_j\leq2M 2 1 m ≤ u j ≤ 2 M and u j → u u_j\to u u j → u locally uniformly in this cylinder. Applying (54) and absorbing the fixed factors of 2 into the constant, we obtain
∥ ∇ ( u j 1 k ) ∥ L ∞ ( B r ( x o ) × ( s − h , s ) ) ≤ C M 1 k ( 1 r + 1 h 1 p ) + C ε j m − k − 1 k . \left\lVert\nabla\left(u_j^{\frac{1}{k}}\right)\right\rVert_{L^\infty(B_r(x_o)\times(s-h,s))}\leq CM^{\frac{1}{k}}\left(\frac{1}{r}+\frac{1}{h^{\frac{1}{p}}}\right)+C\varepsilon_jm^{-\frac{k-1}{k}}. ∇ ( u j k 1 ) L ∞ ( B r ( x o ) × ( s − h , s )) ≤ C M k 1 ( r 1 + h p 1 1 ) + C ε j m − k k − 1 .
Since m > 0 m>0 m > 0 is fixed, the last term tends to zero as j → ∞ j\to\infty j → ∞ . Moreover, u j 1 k → u 1 k u_j^{\frac{1}{k}}\to u^{\frac{1}{k}} u j k 1 → u k 1 locally uniformly in B 2 r ( x o ) × ( s − 2 h , s ] B_{2r}(x_o)\times(s-2h,s] B 2 r ( x o ) × ( s − 2 h , s ] . Thus, passing to the limit in the spatial Lipschitz estimates gives
∣ u ( x , t ) 1 k − u ( y , t ) 1 k ∣ ≤ C M 1 k ( 1 r + 1 h 1 p ) ∣ x − y ∣ \left|u(x,t)^{\frac{1}{k}}-u(y,t)^{\frac{1}{k}}\right|\leq CM^{\frac{1}{k}}\left(\frac{1}{r}+\frac{1}{h^{\frac{1}{p}}}\right)|x-y| u ( x , t ) k 1 − u ( y , t ) k 1 ≤ C M k 1 ( r 1 + h p 1 1 ) ∣ x − y ∣
for every x , y ∈ B r ( x o ) x,y\in B_r(x_o) x , y ∈ B r ( x o ) and t ∈ ( s − h , s ) t\in(s-h,s) t ∈ ( s − h , s ) . This implies (69) . □ \square □
Proof of Theorem 1.1
In this section, we first estimate how quickly a nonnegative solution can become positive inside a ball where it initially vanishes. Comparison with an exponential barrier shows that, for short times, the solution remains exponentially small in a smaller concentric ball. The precise result is stated in the following lemma.
We then prove local spatial Lipschitz continuity by combining this barrier estimate with the a priori gradient estimate on positive cylinders from Proposition 5.2. While the gradient estimate applies directly away from the zero set, the exponential smallness provided by the barrier estimate ensures that the gradient remains uniformly bounded as the solution becomes positive. In fact, we obtain a slightly stronger estimate showing that the gradient vanishes as the solution approaches the zero set; see Theorem 6.3 .
Barrier estimates close to the zero time.
Lemma 6.1. There exist constants c o , τ o > 0 c_o,\tau_o>0 c o , τ o > 0 , depending only on N , p N,p N , p , with the following property. Suppose that u u u is a nonnegative weak solution to (1) in B ϱ ( x o ) × ( t ∗ , t ∗ + τ o ϱ p ] B_\varrho(x_o)\times(t_*,t_*+\tau_o\varrho^p] B ϱ ( x o ) × ( t ∗ , t ∗ + τ o ϱ p ] , satisfies u ≤ M u\le M u ≤ M , and is continuous up to the initial time, with
u ( x , t ∗ ) = 0 for every x ∈ B ϱ ( x o ) . u(x,t_*)=0 \qquad\text{for every }x\in B_\varrho(x_o). u ( x , t ∗ ) = 0 for every x ∈ B ϱ ( x o ) . Then
u ( x , t ) ≤ M exp [ − c o ( ϱ p t − t ∗ ) 1 p − 1 ] (70) u(x,t)\le M\exp\left[-c_o\left(\frac{\varrho^p}{t-t_*}\right)^{\frac{1}{p-1}}\right] \tag*{(70)} u ( x , t ) ≤ M exp [ − c o ( t − t ∗ ϱ p ) p − 1 1 ] (70) for any
( x , t ) ∈ B ϱ / 2 ( x o ) × ( t ∗ , t ∗ + 1 2 τ o ϱ p ] . (x,t)\in B_{\varrho/2}(x_o)\times\left(t_*,t_*+\frac{1}{2}\tau_o\varrho^p\right]. ( x , t ) ∈ B ϱ /2 ( x o ) × ( t ∗ , t ∗ + 2 1 τ o ϱ p ] . If the solution is defined only up to an earlier terminal time T T T , the same estimate holds for t < T t<T t < T in the indicated time interval. It also holds at t = T t=T t = T whenever u u u extends continuously to that time.
Proof. If M = 0 M=0 M = 0 , the assertion is immediate. Therefore, we assume M > 0 M>0 M > 0 . The change of variables
y = x − x o ϱ , τ = t − t ∗ ϱ p y=\frac{x-x_o}{\varrho},\qquad\tau=\frac{t-t_*}{\varrho^p} y = ϱ x − x o , τ = ϱ p t − t ∗
maps the cylinder B ϱ ( x o ) × ( t ∗ , t ∗ + τ o ϱ p ] B_\varrho(x_o)\times(t_*,t_*+\tau_o\varrho^p] B ϱ ( x o ) × ( t ∗ , t ∗ + τ o ϱ p ] onto Q ~ : = B 1 × ( 0 , τ o ] \widetilde Q:=B_1\times(0,\tau_o] Q := B 1 × ( 0 , τ o ] . By Lemma 2.2 the rescaled function
v ( y , τ ) : = u ( x o + ϱ y , t ∗ + ϱ p τ ) M v(y,\tau):=\frac{u(x_o+\varrho y,t_*+\varrho^p\tau)}{M} v ( y , τ ) := M u ( x o + ϱ y , t ∗ + ϱ p τ )
is a nonnegative weak solution to
∂ τ ( v p − 1 ) − div ( ∣ ∇ v ∣ p − 2 ∇ v ) = 0 in Q ~ . \partial_\tau(v^{p-1})-\operatorname{div}(|\nabla v|^{p-2}\nabla v)=0 \qquad\text{in }\widetilde Q. ∂ τ ( v p − 1 ) − div ( ∣∇ v ∣ p − 2 ∇ v ) = 0 in Q .
Moreover, 0 ≤ v ≤ 1 0\le v\le1 0 ≤ v ≤ 1 in Q ~ \widetilde Q Q , and v v v is continuous up to the initial time, with v ( ⋅ , 0 ) = 0 v(\cdot,0)=0 v ( ⋅ , 0 ) = 0 . Thus, proving (70) reduces to proving
v ( y , τ ) ≤ exp [ − c o τ − 1 p − 1 ] in B 1 / 2 × ( 0 , 1 2 τ o ] . (71) v(y,\tau)\le\exp\left[-c_o\tau^{-\frac{1}{p-1}}\right]\qquad\text{in }B_{1/2}\times\left(0,\frac{1}{2}\tau_o\right]. \tag*{(71)} v ( y , τ ) ≤ exp [ − c o τ − p − 1 1 ] in B 1/2 × ( 0 , 2 1 τ o ] . (71)
In the following, all spatial derivatives are taken with respect to y y y . Let p ′ = p p − 1 > 1 p'=\frac{p}{p-1}>1 p ′ = p − 1 p > 1 be the Hölder conjugate of p p p . We define
ψ ( y ) : = ( 1 − ∣ y ∣ p ′ ) p for y ∈ B ‾ 1 . \psi(y):=(1-|y|^{p'})^p \qquad\text{for }y\in\overline{B}_1. ψ ( y ) := ( 1 − ∣ y ∣ p ′ ) p for y ∈ B 1 .
For 0 < ∣ y ∣ < 1 0<|y|<1 0 < ∣ y ∣ < 1 , we have
∇ ψ ( y ) = − ( p ′ ) 2 ( 1 − ∣ y ∣ p ′ ) p − 1 ∣ y ∣ p ′ − 2 y . \nabla\psi(y)=-(p')^2(1-|y|^{p'})^{p-1}|y|^{p'-2}y. ∇ ψ ( y ) = − ( p ′ ) 2 ( 1 − ∣ y ∣ p ′ ) p − 1 ∣ y ∣ p ′ − 2 y .
Since ( p ′ − 1 ) p = p ′ (p'-1)p=p' ( p ′ − 1 ) p = p ′ , we obtain
∣ ∇ ψ ∣ p = ( p ′ ) 2 p ( 1 − ∣ y ∣ p ′ ) ( p ′ − 1 ) p ∣ y ∣ ( p ′ − 1 ) p = ( p ′ ) 2 p ∣ y ∣ p ′ ψ ≤ ( p ′ ) 2 p ψ . (72) \begin{aligned}
|\nabla\psi|^p&=(p')^{2p}(1-|y|^{p'})^{(p'-1)p}|y|^{(p'-1)p}\\
&=(p')^{2p}|y|^{p'}\psi\le(p')^{2p}\psi. \tag*{(72)}
\end{aligned} ∣∇ ψ ∣ p = ( p ′ ) 2 p ( 1 − ∣ y ∣ p ′ ) ( p ′ − 1 ) p ∣ y ∣ ( p ′ − 1 ) p = ( p ′ ) 2 p ∣ y ∣ p ′ ψ ≤ ( p ′ ) 2 p ψ . (72)
Moreover, using ( p ′ − 1 ) ( p − 1 ) = 1 (p'-1)(p-1)=1 ( p ′ − 1 ) ( p − 1 ) = 1 , we find
∣ ∇ ψ ∣ p − 2 ∇ ψ = − ( p ′ ) 2 ( p − 1 ) ( 1 − ∣ y ∣ p ′ ) ( p ′ − 1 ) ( p − 1 ) ∣ y ∣ ( p ′ − 1 ) ( p − 1 ) − 1 y = − ( p ′ ) 2 ( p − 1 ) ( 1 − ∣ y ∣ p ′ ) y . (73) \begin{aligned}
|\nabla\psi|^{p-2}\nabla\psi
&=-(p')^{2(p-1)}(1-|y|^{p'})^{(p'-1)(p-1)}|y|^{(p'-1)(p-1)-1}y\\
&=-(p')^{2(p-1)}(1-|y|^{p'})y. \tag*{(73)}
\end{aligned} ∣∇ ψ ∣ p − 2 ∇ ψ = − ( p ′ ) 2 ( p − 1 ) ( 1 − ∣ y ∣ p ′ ) ( p ′ − 1 ) ( p − 1 ) ∣ y ∣ ( p ′ − 1 ) ( p − 1 ) − 1 y = − ( p ′ ) 2 ( p − 1 ) ( 1 − ∣ y ∣ p ′ ) y . (73)
Since p ′ > 1 p'>1 p ′ > 1 , the last expression extends to a C 1 C^1 C 1 vector field on B 1 B_1 B 1 . Taking its divergence yields
Δ p ψ = − ( p ′ ) 2 ( p − 1 ) div ( ( 1 − ∣ y ∣ p ′ ) y ) = − ( p ′ ) 2 ( p − 1 ) ( N ( 1 − ∣ y ∣ p ′ ) − p ′ ∣ y ∣ p ′ ) = ( p ′ ) 2 ( p − 1 ) ( ( N + p ′ ) ∣ y ∣ p ′ − N ) in B 1 . \begin{aligned}
\Delta_p\psi
&=-(p')^{2(p-1)}\operatorname{div}((1-|y|^{p'})y)\\
&=-(p')^{2(p-1)}\left(N(1-|y|^{p'})-p'|y|^{p'}\right)\\
&=(p')^{2(p-1)}\left((N+p')|y|^{p'}-N\right)\qquad\text{in }B_1.
\end{aligned} Δ p ψ = − ( p ′ ) 2 ( p − 1 ) div (( 1 − ∣ y ∣ p ′ ) y ) = − ( p ′ ) 2 ( p − 1 ) ( N ( 1 − ∣ y ∣ p ′ ) − p ′ ∣ y ∣ p ′ ) = ( p ′ ) 2 ( p − 1 ) ( ( N + p ′ ) ∣ y ∣ p ′ − N ) in B 1 .
To obtain a lower bound for Δ p ψ \Delta_p\psi Δ p ψ in terms of ψ \psi ψ , we distinguish two cases. If
∣ y ∣ p ′ ≥ N N + p ′ , |y|^{p'} \ge\frac{N}{N+p'}, ∣ y ∣ p ′ ≥ N + p ′ N ,
then Δ p ψ ≥ 0 \Delta_p\psi\ge0 Δ p ψ ≥ 0 . In the complementary region,
1 − ∣ y ∣ p ′ > p ′ N + p ′ , 1-|y|^{p'}>\frac{p'}{N+p'}, 1 − ∣ y ∣ p ′ > N + p ′ p ′ ,
and hence
ψ ( y ) > ( p ′ N + p ′ ) p ′ . \psi(y)>\left(\frac{p'}{N+p'}\right)^{p'}. ψ ( y ) > ( N + p ′ p ′ ) p ′ .
Since Δ p ψ ≥ − N ( p ′ ) 2 ( p − 1 ) \Delta_p\psi\ge-N(p')^{2(p-1)} Δ p ψ ≥ − N ( p ′ ) 2 ( p − 1 ) , we obtain
Δ p ψ ≥ − N ( p ′ ) 2 ( p − 1 ) ≥ − N ( p ′ ) 2 ( p − 1 ) ( N + p ′ p ′ ) p ′ ψ . \Delta_p\psi\ge-N(p')^{2(p-1)}\ge-N(p')^{2(p-1)}\left(\frac{N+p'}{p'}\right)^{p'}\psi. Δ p ψ ≥ − N ( p ′ ) 2 ( p − 1 ) ≥ − N ( p ′ ) 2 ( p − 1 ) ( p ′ N + p ′ ) p ′ ψ .
Combining the two regions yields
Δ p ψ ≥ − C o ψ in B 1 , C o : = N ( p ′ ) 2 ( p − 1 ) ( N + p ′ p ′ ) p ′ . (74) \Delta_p\psi\ge-C_o\psi\qquad\text{in } B_1,\qquad C_o:=N(p')^{2(p-1)}\left(\frac{N+p'}{p'}\right)^{p'}. \tag*{(74)} Δ p ψ ≥ − C o ψ in B 1 , C o := N ( p ′ ) 2 ( p − 1 ) ( p ′ N + p ′ ) p ′ . (74)
The constant C o C_o C o depends only on N N N , p p p .
Using the preceding estimates for ψ \psi ψ , we now construct a barrier. For λ > 0 \lambda>0 λ > 0 and δ ∈ ( 0 , 1 2 τ o ] \delta\in\left(0,\frac{1}{2}\tau_o\right] δ ∈ ( 0 , 2 1 τ o ] , define
f ( y , τ ) : = − λ ψ ( y ) ( τ + δ ) 1 p − 1 , V δ ( y , τ ) : = exp ( f ( y , τ ) ) . f(y,\tau):=-\frac{\lambda\psi(y)}{(\tau+\delta)^{\frac{1}{p-1}}},\qquad V_\delta(y,\tau):=\exp(f(y,\tau)). f ( y , τ ) := − ( τ + δ ) p − 1 1 λ ψ ( y ) , V δ ( y , τ ) := exp ( f ( y , τ )) .
The parameter λ \lambda λ will be chosen below, independently of δ \delta δ . Since ψ ∈ C 1 ( B 1 ‾ ) \psi\in C^1(\overline{B_1}) ψ ∈ C 1 ( B 1 ) , both f f f and V δ V_\delta V δ are continuously differentiable in space. Moreover,
∣ ∇ f ∣ p − 2 ∇ f = − λ p − 1 ∣ ∇ ψ ∣ p − 2 ∇ ψ τ + δ . |\nabla f|^{p-2}\nabla f=-\frac{\lambda^{p-1}|\nabla\psi|^{p-2}\nabla\psi}{\tau+\delta}. ∣∇ f ∣ p − 2 ∇ f = − τ + δ λ p − 1 ∣∇ ψ ∣ p − 2 ∇ ψ .
By (73) , this flux is continuously differentiable in space. Thus, the divergences below are well defined pointwise, including at the origin y = 0 y=0 y = 0 . The chain rule gives
∂ τ ( V δ p − 1 ) = ( p − 1 ) ∂ τ f e ( p − 1 ) f . \partial_\tau\left(V_\delta^{p-1}\right)=(p-1)\partial_\tau f\,e^{(p-1)f}. ∂ τ ( V δ p − 1 ) = ( p − 1 ) ∂ τ f e ( p − 1 ) f .
Similarly, since ∇ V δ = e f ∇ f \nabla V_\delta=e^f\nabla f ∇ V δ = e f ∇ f , we have
∣ ∇ V δ ∣ p − 2 ∇ V δ = e ( p − 1 ) f ∣ ∇ f ∣ p − 2 ∇ f . |\nabla V_\delta|^{p-2}\nabla V_\delta=e^{(p-1)f}|\nabla f|^{p-2}\nabla f. ∣∇ V δ ∣ p − 2 ∇ V δ = e ( p − 1 ) f ∣∇ f ∣ p − 2 ∇ f .
Taking the divergence and applying the product rule, we obtain
Δ p V δ = ∇ ( e ( p − 1 ) f ) ⋅ ( ∣ ∇ f ∣ p − 2 ∇ f ) + e ( p − 1 ) f Δ p f = e ( p − 1 ) f [ ( p − 1 ) ∣ ∇ f ∣ p + Δ p f ] . \begin{aligned}
\Delta_pV_\delta=\nabla\left(e^{(p-1)f}\right)\cdot\left(|\nabla f|^{p-2}\nabla f\right)+e^{(p-1)f}\Delta_pf \\
&=e^{(p-1)f}\left[(p-1)|\nabla f|^p+\Delta_pf\right].
\end{aligned} Δ p V δ = ∇ ( e ( p − 1 ) f ) ⋅ ( ∣∇ f ∣ p − 2 ∇ f ) + e ( p − 1 ) f Δ p f = e ( p − 1 ) f [ ( p − 1 ) ∣∇ f ∣ p + Δ p f ] .
Consequently,
∂ τ ( V δ p − 1 ) − Δ p V δ = V δ p − 1 R , (75) \partial_\tau\left(V_\delta^{p-1}\right)-\Delta_pV_\delta=V_\delta^{p-1}\mathbf{R}, \tag*{(75)} ∂ τ ( V δ p − 1 ) − Δ p V δ = V δ p − 1 R , (75)
where
R : = ( p − 1 ) ∂ τ f − Δ p f − ( p − 1 ) ∣ ∇ f ∣ p . \mathbf{R}:=(p-1)\partial_\tau f-\Delta_pf-(p-1)|\nabla f|^p. R := ( p − 1 ) ∂ τ f − Δ p f − ( p − 1 ) ∣∇ f ∣ p .
By (75) , V δ V_\delta V δ is a supersolution whenever R ≥ 0 \mathbf{R}\ge0 R ≥ 0 . We compute the three terms in R \mathbf{R} R separately. Since 1 + 1 p − 1 = p ′ 1+\frac{1}{p-1}=p' 1 + p − 1 1 = p ′ , differentiation with respect to time gives
( p − 1 ) ∂ τ f = λ ψ ( τ + δ ) 1 p − 1 + 1 = λ ψ ( τ + δ ) p ′ . (p-1)\partial_\tau f=\frac{\lambda\psi}{(\tau+\delta)^{\frac{1}{p-1}+1}}=\frac{\lambda\psi}{(\tau+\delta)^{p'}}. ( p − 1 ) ∂ τ f = ( τ + δ ) p − 1 1 + 1 λ ψ = ( τ + δ ) p ′ λ ψ .
For the gradient term, we obtain
( p − 1 ) ∣ ∇ f ∣ p = ( p − 1 ) λ p ∣ ∇ ψ ∣ p ( τ + δ ) p p − 1 = ( p − 1 ) λ p ∣ ∇ ψ ∣ p ( τ + δ ) p ′ . (p-1)|\nabla f|^p=\frac{(p-1)\lambda^p|\nabla\psi|^p}{(\tau+\delta)^{\frac{p}{p-1}}}=\frac{(p-1)\lambda^p|\nabla\psi|^p}{(\tau+\delta)^{p'}}. ( p − 1 ) ∣∇ f ∣ p = ( τ + δ ) p − 1 p ( p − 1 ) λ p ∣∇ ψ ∣ p = ( τ + δ ) p ′ ( p − 1 ) λ p ∣∇ ψ ∣ p .
Finally, taking the spatial divergence of
∣ ∇ f ∣ p − 2 ∇ f = − λ p − 1 τ + δ ∣ ∇ ψ ∣ p − 2 ∇ ψ |\nabla f|^{p-2}\nabla f=-\frac{\lambda^{p-1}}{\tau+\delta}|\nabla\psi|^{p-2}\nabla\psi ∣∇ f ∣ p − 2 ∇ f = − τ + δ λ p − 1 ∣∇ ψ ∣ p − 2 ∇ ψ
yields
Δ p f = − λ p − 1 τ + δ Δ p ψ . \Delta_p f=-\frac{\lambda^{p-1}}{\tau+\delta}\Delta_p\psi. Δ p f = − τ + δ λ p − 1 Δ p ψ .
Substituting these expressions into the definition of R \mathbf{R} R , we find
R = λ ψ ( τ + δ ) p ′ + λ p − 1 τ + δ Δ p ψ − ( p − 1 ) λ p ( τ + δ ) p ′ ∣ ∇ ψ ∣ p . \mathbf{R}=\frac{\lambda\psi}{(\tau+\delta)^{p'}}+\frac{\lambda^{p-1}}{\tau+\delta}\Delta_p\psi-\frac{(p-1)\lambda^p}{(\tau+\delta)^{p'}}|\nabla\psi|^p. R = ( τ + δ ) p ′ λ ψ + τ + δ λ p − 1 Δ p ψ − ( τ + δ ) p ′ ( p − 1 ) λ p ∣∇ ψ ∣ p .
Using (72) and (74) , we obtain
R ≥ ψ [ 1 − ( p − 1 ) ( p ′ ) 2 p λ p − 1 ] λ ( τ + δ ) p ′ − C 0 λ p − 1 τ + δ . (76) \mathbf{R}\geq\psi\left[1-(p-1)(p')^{2p}\lambda^{p-1}\right]\frac{\lambda}{(\tau+\delta)^{p'}}-\frac{C_0\lambda^{p-1}}{\tau+\delta}. \tag*{(76)} R ≥ ψ [ 1 − ( p − 1 ) ( p ′ ) 2 p λ p − 1 ] ( τ + δ ) p ′ λ − τ + δ C 0 λ p − 1 . (76)
We now choose λ \lambda λ and τ 0 \tau_0 τ 0 so that the right-hand side of (76) is nonnegative. First, we choose λ \lambda λ small enough to make the first term positive. We then restrict τ + δ \tau+\delta τ + δ by choosing τ 0 \tau_0 τ 0 sufficiently small so that this positive term dominates the negative one. Both choices will depend only on N , p N,p N , p . The precise choices are as follows. Specifically, let λ > 0 \lambda>0 λ > 0 satisfy
λ p − 1 = 1 2 ( p − 1 ) ( p ′ ) 2 p . \lambda^{p-1}=\frac{1}{2(p-1)(p')^{2p}}. λ p − 1 = 2 ( p − 1 ) ( p ′ ) 2 p 1 .
With this choice of λ \lambda λ , the expression in the inner square brackets in (76) equals 1 2 \frac{1}{2} 2 1 . Using
1 τ + δ = ( τ + δ ) 1 p − 1 ( τ + δ ) p ′ \frac{1}{\tau+\delta}=\frac{(\tau+\delta)^{\frac{1}{p-1}}}{(\tau+\delta)^{p'}} τ + δ 1 = ( τ + δ ) p ′ ( τ + δ ) p − 1 1
we can therefore rewrite (76) as
R ≥ ψ ( τ + δ ) p ′ [ 1 2 λ − C 0 λ p − 1 ( τ + δ ) 1 p − 1 ] . \mathbf{R}\geq\frac{\psi}{(\tau+\delta)^{p'}}\left[\frac{1}{2}\lambda-C_0\lambda^{p-1}(\tau+\delta)^{\frac{1}{p-1}}\right]. R ≥ ( τ + δ ) p ′ ψ [ 2 1 λ − C 0 λ p − 1 ( τ + δ ) p − 1 1 ] .
To make the expression in brackets nonnegative, we choose τ 0 > 0 \tau_0>0 τ 0 > 0 such that
τ 0 1 p − 1 = λ 2 − p 2 C 0 . \tau_0^{\frac{1}{p-1}}=\frac{\lambda^{2-p}}{2C_0}. τ 0 p − 1 1 = 2 C 0 λ 2 − p .
Indeed, whenever 0 ≤ τ 0\leq\tau 0 ≤ τ and τ + δ ≤ τ 0 \tau+\delta\leq\tau_0 τ + δ ≤ τ 0 , we have
C 0 λ p − 1 ( τ + δ ) 1 p − 1 ≤ C 0 λ p − 1 τ 0 1 p − 1 = 1 2 λ . C_0\lambda^{p-1}(\tau+\delta)^{\frac{1}{p-1}}\leq C_0\lambda^{p-1}\tau_0^{\frac{1}{p-1}}=\frac{1}{2}\lambda. C 0 λ p − 1 ( τ + δ ) p − 1 1 ≤ C 0 λ p − 1 τ 0 p − 1 1 = 2 1 λ .
Thus, R ≥ 0 \mathbf{R}\geq0 R ≥ 0 for y ∈ B 1 y\in B_1 y ∈ B 1 and 0 ≤ τ ≤ 1 2 τ 0 0\leq\tau\leq\frac{1}{2}\tau_0 0 ≤ τ ≤ 2 1 τ 0 , provided that 0 < δ ≤ 1 2 τ 0 0<\delta\leq\frac{1}{2}\tau_0 0 < δ ≤ 2 1 τ 0 . Both λ \lambda λ and τ 0 \tau_0 τ 0 depend only on N , p N,p N , p . We conclude that
∂ τ ( V δ p ′ ) − Δ p V δ ≥ 0 in B 1 × ( 0 , 1 2 τ 0 ] . \partial_\tau(V_\delta^{p'})-\Delta_pV_\delta\geq0\qquad\text{in }B_1\times\left(0,\frac{1}{2}\tau_0\right]. ∂ τ ( V δ p ′ ) − Δ p V δ ≥ 0 in B 1 × ( 0 , 2 1 τ 0 ] .
Moreover, since ψ ≥ 0 \psi\geq0 ψ ≥ 0 on B ‾ 1 \overline{B}_1 B 1 and ψ = 0 \psi=0 ψ = 0 on ∂ B 1 \partial B_1 ∂ B 1 , we have
0 < V δ ≤ 1 on B ‾ 1 × [ 0 , τ 0 ) , V δ = 1 on ∂ B 1 × [ 0 , τ 0 ) . 0<V_\delta\leq1\text{ on }\overline{B}_1\times[0,\tau_0),\qquad V_\delta=1\text{ on }\partial B_1\times[0,\tau_0). 0 < V δ ≤ 1 on B 1 × [ 0 , τ 0 ) , V δ = 1 on ∂ B 1 × [ 0 , τ 0 ) .
We now use the comparison principle for (1) (see [9 Theorem 1.2] ) to compare v v v with the barrier. Since no continuity of v v v up to ∂ B 1 \partial B_1 ∂ B 1 is assumed, we first work in B 1 − h B_{1-h} B 1 − h , where 0 < h < 1 0<h<1 0 < h < 1 , and subsequently let h ↓ 0 h\downarrow0 h ↓ 0 . On the boundary of the smaller ball B 1 − h B_{1-h} B 1 − h , the function V δ V_\delta V δ is strictly less than one, so it need not dominate v v v there. Therefore, we multiply V δ V_\delta V δ by a suitable positive constant so that the resulting function is at least one on this boundary. The homogeneity of the equation ensures that this multiplication preserves the supersolution property. The value of ψ \psi ψ on ∂ B 1 − h \partial B_{1-h} ∂ B 1 − h is
ψ h : = ( 1 − ( 1 − h ) p ′ ) p ′ . \psi_h:=\left(1-(1-h)^{p'}\right)^{p'}. ψ h := ( 1 − ( 1 − h ) p ′ ) p ′ .
For h ∈ ( 0 , 1 ) h \in(0,1) h ∈ ( 0 , 1 ) and δ > 0 \delta> 0 δ > 0 define
W ( y , τ ) : = exp ( λ ψ h δ 1 p − 1 ) V δ ( y , τ ) . W(y,\tau):=\exp\left(\frac{\lambda\psi_h}{\delta^{\frac{1}{p-1}}}\right)V_\delta(y,\tau). W ( y , τ ) := exp ( δ p − 1 1 λ ψ h ) V δ ( y , τ ) .
We have
∂ τ ( W p − 1 ) − Δ p W = exp ( ( p − 1 ) λ ψ h δ 1 p − 1 ) [ ∂ τ ( V δ p − 1 ) − Δ p V δ ] ≥ 0. \partial_\tau(W^{p-1})-\Delta_pW=\exp\left(\frac{(p-1)\lambda\psi_h}{\delta^{\frac{1}{p-1}}}\right)\left[\partial_\tau(V_\delta^{p-1})-\Delta_pV_\delta\right]\geq0. ∂ τ ( W p − 1 ) − Δ p W = exp ( δ p − 1 1 ( p − 1 ) λ ψ h ) [ ∂ τ ( V δ p − 1 ) − Δ p V δ ] ≥ 0.
in B 1 × ( 0 , 1 2 τ o ] B_1\times(0,\frac{1}{2}\tau_o] B 1 × ( 0 , 2 1 τ o ] . Fix 0 < δ < 1 2 τ o 0<\delta<\frac{1}{2}\tau_o 0 < δ < 2 1 τ o and 0 < h < 1 0<h<1 0 < h < 1 . We compare v v v and W W W on B 1 − h × ( 0 , 1 2 τ o ] B_{1-h}\times(0,\frac{1}{2}\tau_o] B 1 − h × ( 0 , 2 1 τ o ] . The closure of this cylinder lies inside the region where v v v is continuous and W W W is a supersolution. On the lateral boundary, where ∣ y ∣ = 1 − h |y|=1-h ∣ y ∣ = 1 − h , we have
W ( y , τ ) = exp [ λ ψ h ( 1 δ 1 p − 1 − 1 ( τ + δ ) 1 p − 1 ) ] ≥ 1 ≥ v ( y , τ ) . W(y,\tau)=\exp\left[\lambda\psi_h\left(\frac{1}{\delta^{\frac{1}{p-1}}}-\frac{1}{(\tau+\delta)^{\frac{1}{p-1}}}\right)\right]\geq1\geq v(y,\tau). W ( y , τ ) = exp [ λ ψ h ( δ p − 1 1 1 − ( τ + δ ) p − 1 1 1 ) ] ≥ 1 ≥ v ( y , τ ) .
To obtain the ordering at the initial time, observe that
v ( ⋅ , 0 ) = 0 < W ( ⋅ , 0 ) in B 1 − h . v(\cdot,0)=0<W(\cdot,0)\qquad\text{in }B_{1-h}. v ( ⋅ , 0 ) = 0 < W ( ⋅ , 0 ) in B 1 − h .
Together with the ordering on the lateral boundary, this shows that v ≤ W v\leq W v ≤ W on the parabolic boundary of B 1 − h × ( 0 , 1 2 τ o ] B_{1-h}\times(0,\frac{1}{2}\tau_o] B 1 − h × ( 0 , 2 1 τ o ] . Moreover, W W W satisfies
W ≥ exp ( − λ δ 1 p − 1 ) > 0 W\geq\exp\left(-\frac{\lambda}{\delta^{\frac{1}{p-1}}}\right)>0 W ≥ exp ( − δ p − 1 1 λ ) > 0
throughout the comparison cylinder. Since 0 ≤ v ≤ 1 0\leq v\leq1 0 ≤ v ≤ 1 , all assumptions of the comparison principle [9 Theorem 1.2] are satisfied. Consequently,
v ≤ W in B 1 − h × ( 0 , 1 2 τ o ] . v\leq W\qquad\text{in }B_{1-h}\times(0,\tfrac{1}{2}\tau_o]. v ≤ W in B 1 − h × ( 0 , 2 1 τ o ] .
whence
v ( y , τ ) ≤ exp ( λ ψ h δ 1 p − 1 ) V δ ( y , τ ) in B 1 − h × ( 0 , 1 2 τ o ] . v(y,\tau)\leq\exp\left(\frac{\lambda\psi_h}{\delta^{\frac{1}{p-1}}}\right)V_\delta(y,\tau)\qquad\text{in }B_{1-h}\times(0,\tfrac{1}{2}\tau_o]. v ( y , τ ) ≤ exp ( δ p − 1 1 λ ψ h ) V δ ( y , τ ) in B 1 − h × ( 0 , 2 1 τ o ] .
Fix ( y , τ ) ∈ B 1 × ( 0 , 1 2 τ o ] (y,\tau)\in B_1\times(0,\frac{1}{2}\tau_o] ( y , τ ) ∈ B 1 × ( 0 , 2 1 τ o ] . For 0 < h < 1 − ∣ y ∣ 0<h<1-|y| 0 < h < 1 − ∣ y ∣ , we have y ∈ B 1 − h y\in B_{1-h} y ∈ B 1 − h , so the preceding estimate gives
v ( y , τ ) ≤ exp ( λ ψ h δ 1 p − 1 ) V δ ( y , τ ) . v(y,\tau)\leq\exp\left(\frac{\lambda\psi_h}{\delta^{\frac{1}{p-1}}}\right)V_\delta(y,\tau). v ( y , τ ) ≤ exp ( δ p − 1 1 λ ψ h ) V δ ( y , τ ) .
Since ψ h → 0 \psi_h\to0 ψ h → 0 as h ↓ 0 h\downarrow0 h ↓ 0 and δ \delta δ remains fixed, the exponential factor converges to one. Thus, v ( y , τ ) ≤ V δ ( y , τ ) v(y,\tau)\leq V_\delta(y,\tau) v ( y , τ ) ≤ V δ ( y , τ ) . As ( y , τ ) (y,\tau) ( y , τ ) was arbitrary, this holds throughout B 1 × ( 0 , 1 2 τ o ] B_1\times(0,\frac{1}{2}\tau_o] B 1 × ( 0 , 2 1 τ o ] . We now let δ ↓ 0 \delta\downarrow0 δ ↓ 0 and conclude that
v ( y , τ ) ≤ exp [ − λ ψ ( y ) τ 1 p − 1 ] . v(y,\tau)\leq\exp\left[-\frac{\lambda\psi(y)}{\tau^{\frac{1}{p-1}}}\right]. v ( y , τ ) ≤ exp [ − τ p − 1 1 λ ψ ( y ) ] .
For y ∈ B 1 / 2 ‾ y\in\overline{B_{1/2}} y ∈ B 1/2 , we have
ψ ( y ) ≥ ( 1 − 1 2 p ′ ) p ′ . \psi(y)\geq\left(1-\frac{1}{2^{p'}}\right)^{p'}. ψ ( y ) ≥ ( 1 − 2 p ′ 1 ) p ′ .
Thus, setting
c o : = λ ( 1 − 1 2 p ′ ) p ′ > 0 , c_o:=\lambda\left(1-\frac{1}{2^{p'}}\right)^{p'}>0, c o := λ ( 1 − 2 p ′ 1 ) p ′ > 0 ,
we obtain (71) . Since λ \lambda λ depends only on N N N and p p p , the same is true for c o c_o c o . Returning to the original variables, we obtain
u ( x , t ) ≤ M exp [ − c o ( ϱ p t − t ∗ ) 1 p − 1 ] u(x,t)\leq M\exp\left[-c_o\left(\frac{\varrho^p}{t-t_*}\right)^{\frac{1}{p-1}}\right] u ( x , t ) ≤ M exp [ − c o ( t − t ∗ ϱ p ) p − 1 1 ]
for every ( x , t ) ∈ B ϱ / 2 ( x o ) × ( t ∗ , t ∗ + 1 2 τ o ϱ p ] (x,t)\in B_{\varrho/2}(x_o)\times(t_*,t_*+\frac{1}{2}\tau_o\varrho^p] ( x , t ) ∈ B ϱ /2 ( x o ) × ( t ∗ , t ∗ + 2 1 τ o ϱ p ] .
!Spatial cross-section of the comparison cylinders, showing the cylinders B 1 × ( 0 , τ o − δ ] B_1 \times(0,\tau_o-\delta] B 1 × ( 0 , τ o − δ ] and B 1 − h × ( 0 , 1 2 τ o ) B_{1-h} \times(0,\frac{1}{2}\tau_o) B 1 − h × ( 0 , 2 1 τ o ) , the comparison principle, and the lifted barrier bbox=0.23,0.13,0.78,0.53
Figure 3. A spatial cross-section of the comparison cylinders. The solid orange boundary is the parabolic boundary on which v ≤ W v \le W v ≤ W is established. The dashed upper time face is not part of the parabolic boundary. The diagram is schematic, with 0 < 1 2 τ o < τ o − δ 0 < \frac{1}{2}\tau_o < \tau_o-\delta 0 < 2 1 τ o < τ o − δ .
Finally, recall that τ o \tau_o τ o was chosen solely in the construction of the barrier and depends only on N N N and p p p . Suppose now that u u u is defined only on B ϱ ( x o ) × ( t ∗ , T ) B_{\varrho}(x_o) \times(t_*,T) B ϱ ( x o ) × ( t ∗ , T ) , where t ∗ < T < t ∗ + τ o ϱ p t_* < T < t_*+\tau_o\varrho^p t ∗ < T < t ∗ + τ o ϱ p . For any
0 < σ < min { T − t ∗ ϱ p , τ o 2 } , 0 < \sigma< \min\left\{\frac{T-t_*}{\varrho^p},\frac{\tau_o}{2}\right\}, 0 < σ < min { ϱ p T − t ∗ , 2 τ o } ,
we can carry out the same comparison on B 1 − h × ( 0 , σ ] B_{1-h} \times(0,\sigma] B 1 − h × ( 0 , σ ] . The barrier and all constants remain unchanged. Thus, the estimate holds whenever t ∗ < t < T t_* < t < T t ∗ < t < T and t ≤ t ∗ + 1 2 τ o ϱ p t \le t_*+\frac{1}{2}\tau_o\varrho^p t ≤ t ∗ + 2 1 τ o ϱ p , with equality at the latter time obtained by continuity. If T ≤ t ∗ + 1 2 τ o ϱ p T \le t_*+\frac{1}{2}\tau_o\varrho^p T ≤ t ∗ + 2 1 τ o ϱ p and u u u extends continuously to t = T t=T t = T , passing to the limit as t ↑ T t \uparrow T t ↑ T also gives the estimate at T T T .
Remark 6.2. Exponential-type comparison function have already been used before; see, for example, [12 ] in the context of singular porous medium equations, or [27 ] for doubly nonlinear parabolic equations.
Spatial Lipschitz continuity
We now combine the barrier estimate of Lemma 6.1 with the gradient estimate on positive cylinders from Proposition 5.2 to establish the following result, which is stronger than Theorem 1.1 and implies it as a consequence. In fact, it quantifies the decay of the gradient as the solution approaches the zero set.
Theorem 6.3. Let u u u be a locally bounded, nonnegative local weak solution of (1.1) in Ω T \Omega_T Ω T . Let k o k_o k o be defined in (3.1). For every k ≥ k o k \ge k_o k ≥ k o , there exists a constant C = C ( N , p , k ) > 0 C = C(N,p,k) > 0 C = C ( N , p , k ) > 0 with the following property. If Q 2 R ( z o ) ⋐ Ω T Q_{2R}(z_o) \Subset\Omega_T Q 2 R ( z o ) ⋐ Ω T and
0 ≤ u ≤ M in Q 2 R ( z o ) , 0 \le u \le M \quad\text{in } Q_{2R}(z_o), 0 ≤ u ≤ M in Q 2 R ( z o ) ,
then
∥ ∇ ( u 1 k ) ∥ L ∞ ( Q R ( z o ) ) ≤ C M 1 k R . \left\lVert\nabla\left(u^{\frac{1}{k}}\right)\right\rVert_{L^\infty(Q_R(z_o))} \leq\frac{C M^{\frac{1}{k}}}{R}. ∇ ( u k 1 ) L ∞ ( Q R ( z o )) ≤ R C M k 1 .
Moreover,
∣ ∇ u ∣ ≤ C M 1 k R u 1 − 1 k a.e. in Q R ( z o ) . |\nabla u| \leq\frac{C M^{\frac{1}{k}}}{R}u^{1-\frac{1}{k}}
\qquad\text{a.e. in } Q_R(z_o). ∣∇ u ∣ ≤ R C M k 1 u 1 − k 1 a.e. in Q R ( z o ) .
Proof. If M = 0 M=0 M = 0 , the estimate is immediate. Otherwise, by Lemma 2.2 , we consider the rescaled function
v ( y , τ ) : = 1 M u ( x o + R y , t o + R p τ ) , v(y,\tau):=\frac{1}{M}u(x_o+Ry,t_o+R^p\tau), v ( y , τ ) := M 1 u ( x o + R y , t o + R p τ ) ,
which is a solution to (1) on Q 2 : = B 2 ( 0 ) × ( − 2 p , 0 ] Q_2:=B_2(0)\times(-2^p,0] Q 2 := B 2 ( 0 ) × ( − 2 p , 0 ] and satisfies 0 ≤ v ≤ 1 0\leq v\leq1 0 ≤ v ≤ 1 . Fix k = k o k=k_o k = k o , with k o k_o k o defined in (9) , so that k k k depends only on N , p N,p N , p . Let t ∗ ∈ [ − 2 p , 0 ] t_*\in[-2^p,0] t ∗ ∈ [ − 2 p , 0 ] be the transition time given by Proposition 2.4 applied on Q 2 Q_2 Q 2 , i.e. v v v vanishes throughout B 2 B_2 B 2 before t ∗ t_* t ∗ and is strictly positive after t ∗ t_* t ∗ . If t ∗ = 0 t_*=0 t ∗ = 0 , then v ≡ 0 v\equiv0 v ≡ 0 in Q 2 Q_2 Q 2 , and the estimate is immediate. Otherwise, fix ( x ˉ , t ˉ ) ∈ Q 1 (\bar{x},\bar{t})\in Q_1 ( x ˉ , t ˉ ) ∈ Q 1 . Let τ o = τ o ( N , p ) > 0 \tau_o=\tau_o(N,p)>0 τ o = τ o ( N , p ) > 0 be the parameter from Lemma 6.1 , and choose
0 < τ 1 < min { 2 p − 1 , 1 2 2 − p τ o } . 0<\tau_1<\min\left\{2^p-1,\frac{1}{2}2^{-p}\tau_o\right\}. 0 < τ 1 < min { 2 p − 1 , 2 1 2 − p τ o } .
We distinguish three cases. Suppose first that t ˉ ≥ t ∗ + τ 1 \bar{t}\geq t_*+\tau_1 t ˉ ≥ t ∗ + τ 1 , so that t ˉ \bar{t} t ˉ is separated from the transition time by at least τ 1 \tau_1 τ 1 . The closed cylinder
B 1 2 ( x ˉ ) ‾ × [ t ˉ − 1 2 τ 1 , t ˉ ] ⊂ B 2 ( 0 ) × ( t ∗ , 0 ] \overline{B_{\frac{1}{2}}(\bar{x})}\times\left[\bar{t}-\frac{1}{2}\tau_1,\bar{t}\right]\subset B_2(0)\times(t_*,0] B 2 1 ( x ˉ ) × [ t ˉ − 2 1 τ 1 , t ˉ ] ⊂ B 2 ( 0 ) × ( t ∗ , 0 ]
lies entirely in the region where u u u is strictly positive. By continuity and compactness, there exists m > 0 m>0 m > 0 such that
m ≤ v ≤ 1 on B 1 2 ( x ˉ ) ‾ × [ t ˉ − 1 2 τ 1 , t ˉ ] . m\leq v\leq1\qquad\text{on }\overline{B_{\frac{1}{2}}(\bar{x})}\times\left[\bar{t}-\frac{1}{2}\tau_1,\bar{t}\right]. m ≤ v ≤ 1 on B 2 1 ( x ˉ ) × [ t ˉ − 2 1 τ 1 , t ˉ ] .
Therefore, we can apply Proposition 5.2 with r = 1 4 r=\frac{1}{4} r = 4 1 and h = 1 4 τ 1 h=\frac{1}{4}\tau_1 h = 4 1 τ 1 . Since v 1 k ≤ 1 v^{\frac{1}{k}}\leq1 v k 1 ≤ 1 , this yields
∥ ∇ ( v 1 k ) ∥ L ∞ ( B 1 4 ( x ˉ ) × ( t ˉ − 1 4 τ 1 , t ˉ ) ) ≤ C [ 4 + ( 4 τ 1 ) 1 p ] ≤ C ( N , p , k ) . \left\lVert\nabla\left(v^{\frac{1}{k}}\right)\right\rVert_{L^\infty\left(B_{\frac{1}{4}}(\bar{x})\times\left(\bar{t}-\frac{1}{4}\tau_1,\bar{t}\right)\right)}
\leq C\left[4+\left(\frac{4}{\tau_1}\right)^{\frac{1}{p}}\right]\leq C(N,p,k). ∇ ( v k 1 ) L ∞ ( B 4 1 ( x ˉ ) × ( t ˉ − 4 1 τ 1 , t ˉ ) ) ≤ C [ 4 + ( τ 1 4 ) p 1 ] ≤ C ( N , p , k ) .
Here τ 1 \tau_1 τ 1 is fixed in terms of N , p N,p N , p , and the bound is independent of m m m .
Next, suppose that t ∗ < t ˉ < t ∗ + τ 1 t_*<\bar{t}<t_*+\tau_1 t ∗ < t ˉ < t ∗ + τ 1 . Thus, ( x ˉ , t ˉ ) (\bar{x},\bar{t}) ( x ˉ , t ˉ ) lies in the positivity set, but its time coordinate is close to the transition time. Since t ˉ > − 1 \bar{t}>-1 t ˉ > − 1 and τ 1 < 2 p − 1 \tau_1<2^p-1 τ 1 < 2 p − 1 , we have
t ∗ > t ˉ − τ 1 > − 1 − ( 2 p − 1 ) = − 2 p . t_*>\bar{t}-\tau_1>-1-(2^p-1)=-2^p. t ∗ > t ˉ − τ 1 > − 1 − ( 2 p − 1 ) = − 2 p .
In particular, the initial time t ∗ t_* t ∗ lies inside the time interval of Q 2 Q_2 Q 2 , and v ( ⋅ , t ∗ ) = 0 v(\cdot,t_*)=0 v ( ⋅ , t ∗ ) = 0 in B 2 B_2 B 2 . We apply Lemma 6.1 with ϱ = 1 2 \varrho=\frac{1}{2} ϱ = 2 1 , center x ˉ \bar{x} x ˉ , and upper bound M = 1 M=1 M = 1 , using its version for shorter time intervals if necessary. This gives
v ( x , t ) ≤ exp [ − c o ( 2 − p t − t ∗ ) 1 p − 1 ] v(x,t)\leq\exp\left[-c_o\left(\frac{2^{-p}}{t-t_*}\right)^{\frac{1}{p-1}}\right] v ( x , t ) ≤ exp [ − c o ( t − t ∗ 2 − p ) p − 1 1 ]
for every ( x , t ) (x,t) ( x , t ) satisfying
x ∈ B 1 4 ( x ˉ ) , t ∗ < t ≤ min { 0 , t ∗ + 1 2 2 − p τ o } . x\in B_{\frac{1}{4}}(\bar{x}),\qquad t_*<t\leq\min\left\{0,t_*+\frac{1}{2}2^{-p}\tau_o\right\}. x ∈ B 4 1 ( x ˉ ) , t ∗ < t ≤ min { 0 , t ∗ + 2 1 2 − p τ o } .
Since t ˉ − t ∗ < τ 1 < τ o / 2 p + 1 \bar{t}-t_*<\tau_1<\tau_o/2^{p+1} t ˉ − t ∗ < τ 1 < τ o / 2 p + 1 , this estimate applies, in particular, for all t ∈ ( t ∗ , t ˉ ] t\in(t_*,\bar{t}] t ∈ ( t ∗ , t ˉ ] . Here c o c_o c o depends only on N N N and p p p . By the choice of τ 1 \tau_1 τ 1 , we have
t ∗ < 1 2 ( t ∗ + t ˉ ) < t ˉ < t ∗ + τ 1 < t ∗ + 1 2 2 − p τ o . t_*<\frac{1}{2}(t_*+\bar{t})<\bar{t}<t_*+\tau_1<t_*+\frac{1}{2}2^{-p}\tau_o. t ∗ < 2 1 ( t ∗ + t ˉ ) < t ˉ < t ∗ + τ 1 < t ∗ + 2 1 2 − p τ o .
Since the right-hand side of the preceding estimate is increasing in t t t , it follows that
sup B 1 4 ( x ˉ ) × ( 1 2 ( t ∗ + t ˉ ) , t ˉ ) v ≤ exp [ − c 1 ( t ˉ − t ∗ ) 1 p − 1 ] = : M ∗ , c 1 : = c o 2 − p p − 1 > 0. \sup_{B_{\frac{1}{4}}(\bar{x})\times\left(\frac{1}{2}(t_*+\bar{t}),\bar{t}\right)}v
\leq\exp\left[-\frac{c_1}{(\bar{t}-t_*)^{\frac{1}{p-1}}}\right]=:M_*,
\qquad c_1:=c_o2^{-\frac{p}{p-1}}>0. B 4 1 ( x ˉ ) × ( 2 1 ( t ∗ + t ˉ ) , t ˉ ) sup v ≤ exp [ − ( t ˉ − t ∗ ) p − 1 1 c 1 ] =: M ∗ , c 1 := c o 2 − p − 1 p > 0.
By positivity and continuity, u u u also has a positive lower bound on B 1 4 ( x ˉ ) × [ 1 2 ( t ∗ + t ˉ ) , t ˉ ] B_{\frac{1}{4}}(\bar{x}) \times[\frac{1}{2}(t_*+\bar{t}),\bar{t}] B 4 1 ( x ˉ ) × [ 2 1 ( t ∗ + t ˉ ) , t ˉ ] . Therefore, we can apply Proposition 5.2 with ( r , h , s , M ) (r,h,s,\mathbf{M}) ( r , h , s , M ) replaced by ( 1 8 , 1 4 ( t ˉ − t ∗ ) , t ˉ , M ∗ ) (\frac{1}{8},\frac{1}{4}(\bar{t}-t_*),\bar{t},\mathbf{M}_*) ( 8 1 , 4 1 ( t ˉ − t ∗ ) , t ˉ , M ∗ ) . This yields
∥ ∇ ( v 1 k ) ∥ L ∞ ( B 1 8 ( x ˉ ) × ( t ˉ − 1 4 ( t ˉ − t ∗ ) , t ˉ ) ) ≤ C exp [ − c 1 k ( t ˉ − t ∗ ) 1 p − 1 ] [ δ + ( 4 t ˉ − t ∗ ) 1 p ] ≤ C ( N , p , k ) . \left\|\nabla\left(v^{\frac{1}{k}}\right)\right\|_{L^\infty\left(B_{\frac{1}{8}}(\bar{x})\times(\bar{t}-\frac{1}{4}(\bar{t}-t_*),\bar{t})\right)}
\leq C\exp\left[-\frac{c_1}{k(\bar{t}-t_*)^{\frac{1}{p-1}}}\right]\left[\delta+\left(\frac{4}{\bar{t}-t_*}\right)^{\frac{1}{p}}\right]\leq C(N,p,k). ∇ ( v k 1 ) L ∞ ( B 8 1 ( x ˉ ) × ( t ˉ − 4 1 ( t ˉ − t ∗ ) , t ˉ ) ) ≤ C exp [ − k ( t ˉ − t ∗ ) p − 1 1 c 1 ] [ δ + ( t ˉ − t ∗ 4 ) p 1 ] ≤ C ( N , p , k ) .
The last inequality follows because the function
\sigma\longmapsto\exp\left[-\frac{c_1}{k\sigma^{\frac{1}{p-1}}}\right]\left[1+\frac{1}{\sigma^{\frac{1}{p}}\right]}
is bounded above by a constant C = C ( N , p , k ) C=C(N,p,k) C = C ( N , p , k ) for all σ > 0 \sigma>0 σ > 0 . In particular, the resulting constant is independent of t ∗ t_* t ∗ and t ˉ \bar{t} t ˉ .
Finally, if t ˉ ≤ t ∗ \bar{t}\leq t_* t ˉ ≤ t ∗ , then v ( ⋅ , t ˉ ) ≡ 0 v(\cdot,\bar{t})\equiv0 v ( ⋅ , t ˉ ) ≡ 0 in B 2 ( 0 ) B_2(0) B 2 ( 0 ) . Consequently, u 1 k ( ⋅ , t ˉ ) u^{\frac{1}{k}}(\cdot,\bar{t}) u k 1 ( ⋅ , t ˉ ) is constant, and its spatial gradient vanishes.
The first two cases yield a uniform bound C ( N , p , k ) C(N,p,k) C ( N , p , k ) for the spatial gradient of v 1 k v^{\frac{1}{k}} v k 1 in the positive region Q 1 ∩ { t > t ∗ } Q_1\cap\{t>t_*\} Q 1 ∩ { t > t ∗ } . Indeed, this region can be covered by cylinders of the type considered above, and the bounds are independent of their location. For t ≤ t ∗ t\leq t_* t ≤ t ∗ , the function v 1 k ( ⋅ , t ) v^{\frac{1}{k}}(\cdot,t) v k 1 ( ⋅ , t ) is identically zero, so its spatial gradient vanishes. Combining these estimates, we obtain
∥ ∇ ( v 1 k ) ∥ L ∞ ( Q 1 ) ≤ C ( N , p , k ) . \left\|\nabla\left(v^{\frac{1}{k}}\right)\right\|_{L^\infty(Q_1)}\leq C(N,p,k). ∇ ( v k 1 ) L ∞ ( Q 1 ) ≤ C ( N , p , k ) .
Rescaling back to the original solution u u u on Q 2 R ( z o ) Q_{2R}(z_o) Q 2 R ( z o ) , we have shown that
∥ ∇ ( u 1 k ) ∥ L ∞ ( Q R ( z o ) ) ≤ C ( N , p , k ) M 1 k R , \left\|\nabla\left(u^{\frac{1}{k}}\right)\right\|_{L^\infty(Q_R(z_o))}\leq\frac{C(N,p,k)M^{\frac{1}{k}}}{R}, ∇ ( u k 1 ) L ∞ ( Q R ( z o )) ≤ R C ( N , p , k ) M k 1 ,
which is the first assertion of the theorem. In particular, this shows ∇ ( u 1 k ) ≤ C M 1 k / R \nabla(u^{\frac{1}{k}})\leq CM^{\frac{1}{k}}/R ∇ ( u k 1 ) ≤ C M k 1 / R a.e. in Q R ( z o ) Q_R(z_o) Q R ( z o ) . By the spatial Sobolev chain rule, we have ∇ u = k u 1 − 1 k ∇ ( u 1 k ) \nabla u=ku^{1-\frac{1}{k}}\nabla(u^{\frac{1}{k}}) ∇ u = k u 1 − k 1 ∇ ( u k 1 ) almost everywhere in Q R ( z o ) Q_R(z_o) Q R ( z o ) , which in combination with the preceding inequality yields the second claim. □
At this point, Theorem 1.1 is an immediate consequence of the preceding Theorem 6.3 .
Proof of Theorem 1.1. Set M : = ∥ u ∥ L ∞ ( Q 2 R ( z o ) ) M:=\|u\|_{L^\infty(Q_{2R}(z_o))} M := ∥ u ∥ L ∞ ( Q 2 R ( z o )) and choose k = k o k=k_o k = k o , where k o k_o k o is the parameter defined in (3.1). From the second inequality of Theorem 6.3 , we have
∥ ∇ u ∥ L ∞ ( Q R ( z o ) ) ≤ C M 1 k R ∥ u 1 − 1 k ∥ L ∞ ( Q R ( z o ) ) ≤ C R ∥ u ∥ L ∞ ( Q 2 R ( z o ) ) , \|\nabla u\|_{L^\infty(Q_R(z_o))}\leq\frac{CM^{\frac{1}{k}}}{R}\|u^{1-\frac{1}{k}}\|_{L^\infty(Q_R(z_o))}\leq\frac{C}{R}\|u\|_{L^\infty(Q_{2R}(z_o))}, ∥∇ u ∥ L ∞ ( Q R ( z o )) ≤ R C M k 1 ∥ u 1 − k 1 ∥ L ∞ ( Q R ( z o )) ≤ R C ∥ u ∥ L ∞ ( Q 2 R ( z o )) ,
with a constant C = C ( N , p ) C=C(N,p) C = C ( N , p ) . This completes the proof of Theorem 1.1 . □
C1alpha-estimates
For solutions to (1.1) that are strictly positive and bounded, regularity results for equations of parabolic p p p -Laplace type yield Hölder estimates for the gradient. We recall the following result from [6 Theorem 6.8] .
Lemma 7.1. Suppose U U U is a weak solution to (1.1) in Q 4 Q_4 Q 4 , satisfying
1 2 ≤ U ≤ 2 a.e. in Q 4 . \frac{1}{2}\leq U\leq2\quad\text{a.e. in }Q_4. 2 1 ≤ U ≤ 2 a.e. in Q 4 .
Then U U U has a continuous spatial gradient on Q 1 Q_1 Q 1 . Moreover, for some β ∈ ( 0 , 1 ) \beta\in(0,1) β ∈ ( 0 , 1 ) and C > 1 C>1 C > 1 depending only on N N N and p p p , we have
sup Q 1 ∣ ∇ U ∣ ≤ C , ∣ ∇ U ( z ) − ∇ U ( ζ ) ∣ ≤ C dist p a r ( z , ζ ) β \sup_{Q_1}|\nabla U|\leq C,\qquad|\nabla U(z)-\nabla U(\zeta)|\leq C\operatorname{dist}_{\mathrm{par}}(z,\zeta)^\beta Q 1 sup ∣∇ U ∣ ≤ C , ∣∇ U ( z ) − ∇ U ( ζ ) ∣ ≤ C dist par ( z , ζ ) β
for any z , ζ ∈ Q 1 z,\zeta\in Q_1 z , ζ ∈ Q 1 . We are now in a position to give the
Proof of Theorem 1.2 . The proof is divided into three steps. In the first two steps, we establish the qualitative property of continuity of the gradient, first on the positivity set and then on the zero set. In the final step, we then prove the claimed Hölder continuity of the gradient.
Step 1: Continuous differentiability on the positivity set. In this section, we aim to show that ∇ u \nabla u ∇ u is continuous throughout { u > 0 } ∩ Q 1 \{u > 0\} \cap Q_1 { u > 0 } ∩ Q 1 . We work henceforth with 0 ≤ u ≤ 1 0 \le u \le1 0 ≤ u ≤ 1 in Q 2 = B 2 × ( − 2 p , 0 ] Q_2 = B_2 \times(-2^p, 0] Q 2 = B 2 × ( − 2 p , 0 ] . The gradient bound from Theorem 6.3 and Proposition 2.3 give constants L L L , H H H and α , η ∈ ( 0 , 1 ) \alpha, \eta\in(0,1) α , η ∈ ( 0 , 1 ) , depending only on N N N and p p p , such that on Q 7 4 Q_{\frac{7}{4}} Q 4 7 we have that
∣ u α ( x , t ) − u α ( y , t ) ∣ ≤ L ∣ x − y ∣ , ∣ ∇ u ∣ ≤ L u 1 − α a.e. , (77) |u^\alpha(x,t) - u^\alpha(y,t)| \le L|x-y|,\qquad|\nabla u| \le Lu^{1-\alpha}\quad\text{a.e.}, \tag*{(77)} ∣ u α ( x , t ) − u α ( y , t ) ∣ ≤ L ∣ x − y ∣ , ∣∇ u ∣ ≤ L u 1 − α a.e. , (77)
∣ u ( z ) − u ( ζ ) ∣ ≤ H dist p a r ( z , ζ ) η . (78) |u(z) - u(\zeta)| \le H\operatorname{dist}_{\mathrm{par}}(z,\zeta)^\eta. \tag*{(78)} ∣ u ( z ) − u ( ζ ) ∣ ≤ H dist par ( z , ζ ) η . (78)
We used a.e. for the gradient since we do not know yet if it is continuous. Pick ζ = ( y , s ) ∈ Q 1 \zeta= (y,s) \in Q_1 ζ = ( y , s ) ∈ Q 1 such that u ( ζ ) > 0 u(\zeta) > 0 u ( ζ ) > 0 ; it suffices to show that for any ε > 0 \varepsilon> 0 ε > 0 , there is δ ( ε ) > 0 \delta(\varepsilon) > 0 δ ( ε ) > 0 , such that
∣ ∇ u ( z ) − ∇ u ( ζ ) ∣ ≤ ε |\nabla u(z) - \nabla u(\zeta)| \le\varepsilon ∣∇ u ( z ) − ∇ u ( ζ ) ∣ ≤ ε
for any z ∈ Q δ ( ζ ) z \in Q_{\delta}(\zeta) z ∈ Q δ ( ζ ) . To this end, we first fix c ∈ ( 0 , 1 ) c \in(0,1) c ∈ ( 0 , 1 ) , depending only on N N N and p p p , so small that
4 c < 3 / 4 , ( 4 c ) p < ( 7 / 4 ) p − 1 , H ( 8 c ) η ≤ 1 / 2. (79) 4c < 3/4,\qquad(4c)^p < (7/4)^p - 1,\qquad H(8c)^\eta\le1/2. \tag*{(79)} 4 c < 3/4 , ( 4 c ) p < ( 7/4 ) p − 1 , H ( 8 c ) η ≤ 1/2. (79)
For another point z = ( x , t ) ∈ Q 1 z = (x,t) \in Q_1 z = ( x , t ) ∈ Q 1 , we set
δ = dist p a r ( z , ζ ) , m = max { u ( z ) , u ( ζ ) } ∈ ( 0 , 1 ] , ϱ = c m 1 η . \delta= \operatorname{dist}_{\mathrm{par}}(z,\zeta),\qquad m = \max\{u(z),u(\zeta)\} \in(0,1],\qquad\varrho= c m^{\frac{1}{\eta}}. δ = dist par ( z , ζ ) , m = max { u ( z ) , u ( ζ )} ∈ ( 0 , 1 ] , ϱ = c m η 1 .
We consider the case δ ≤ ϱ / 2 \delta\le\varrho/2 δ ≤ ϱ /2 . Denote by z m = ( x m , t m ) z_m = (x_m,t_m) z m = ( x m , t m ) either z z z or ζ \zeta ζ where the maximum m m m is attained, and set t + = max { t , s } t_+ = \max\{t,s\} t + = max { t , s } to the later time. Use the cylinder
C ϱ = B 4 ϱ ( x m ) × ( t + − ( 4 ϱ ) p , t + ] . C_\varrho= B_{4\varrho}(x_m) \times(t_+ - (4\varrho)^p,t_+]. C ϱ = B 4 ϱ ( x m ) × ( t + − ( 4 ϱ ) p , t + ] .
The first two conditions in (79) and the fact m ≤ 1 m \le1 m ≤ 1 place this cylinder in Q 7 4 Q_{\frac{7}{4}} Q 4 7 . Observe that every w ∈ C ϱ w \in C_\varrho w ∈ C ϱ satisfies dist p a r ( w , z m ) ≤ 8 ϱ \operatorname{dist}_{\mathrm{par}}(w,z_m) \le8\varrho dist par ( w , z m ) ≤ 8 ϱ : the spatial separation is at most 4 ϱ 4\varrho 4 ϱ , and the time separation is at most ( 4 ϱ ) p (4\varrho)^p ( 4 ϱ ) p , because 0 ≤ t + − t m ≤ δ p ≤ ( ϱ / 2 ) p 0 \le t_+ - t_m \le\delta^p \le(\varrho/2)^p 0 ≤ t + − t m ≤ δ p ≤ ( ϱ /2 ) p . Hence, by (78) and {(7.3)}3 _3 3 we get for all w ∈ C ϱ w \in C_\varrho w ∈ C ϱ that
∣ u ( w ) − u ( z m ) ∣ ≤ H dist p a r ( w , z m ) η ≤ H ( 8 ϱ ) η = H ( 8 c ) η m ≤ 1 2 m |u(w) - u(z_m)| \le H\operatorname{dist}_{\mathrm{par}}(w,z_m)^\eta\le H(8\varrho)^\eta= H(8c)^\eta m \le\tfrac12 m ∣ u ( w ) − u ( z m ) ∣ ≤ H dist par ( w , z m ) η ≤ H ( 8 ϱ ) η = H ( 8 c ) η m ≤ 2 1 m
which in turn yields that
1 2 m ≤ u ( w ) ≤ 3 2 m , for all w ∈ C ϱ . (80) \tfrac12m \le u(w) \le\tfrac32m,\qquad\text{for all }w \in C_\varrho. \tag*{(80)} 2 1 m ≤ u ( w ) ≤ 2 3 m , for all w ∈ C ϱ . (80)
Both points z z z and ζ \zeta ζ lie in B ϱ ‾ ( x m ) × [ t + − ϱ p , t + ] \overline{B_\varrho}(x_m) \times[t_+ - \varrho^p,t_+] B ϱ ( x m ) × [ t + − ϱ p , t + ] since we have assumed their distance δ \delta δ to be less than ϱ / 2 \varrho/2 ϱ /2 .
Define on Q 4 Q_4 Q 4
U ( Y , S ) : = u ( x m + ϱ Y , t + + ϱ p S ) m , U(Y,S) := \frac{u(x_m+\varrho Y,t_+ + \varrho^p S)}{m}, U ( Y , S ) := m u ( x m + ϱ Y , t + + ϱ p S ) ,
which again solves (1) . By (80) , we have 1 / 2 ≤ U ≤ 3 / 2 1/2 \le U \le3/2 1/2 ≤ U ≤ 3/2 . First apply Lemma 7.1 to U U U and then scale back to get that
∣ ∇ u ( z ) − ∇ u ( ζ ) ∣ ≤ C m ϱ ( δ ϱ ) β = C δ β m 1 − 1 + β η , (81) |\nabla u(z) - \nabla u(\zeta)| \le C\frac{m}{\varrho}\left(\frac{\delta}{\varrho}\right)^\beta= C\delta^\beta m^{1-\frac{1+\beta}{\eta}}, \tag*{(81)} ∣∇ u ( z ) − ∇ u ( ζ ) ∣ ≤ C ϱ m ( ϱ δ ) β = C δ β m 1 − η 1 + β , (81)
for any z ∈ Q 1 z \in Q_1 z ∈ Q 1 , as long as the distance verifies that δ = dist p a r ( z , ζ ) ≤ 1 2 ϱ = 1 2 c m 1 η \delta= \operatorname{dist}_{\mathrm{par}}(z,\zeta) \le\tfrac12\varrho= \tfrac12cm^{\frac1\eta} δ = dist par ( z , ζ ) ≤ 2 1 ϱ = 2 1 c m η 1 .
Note that the factor c − 1 − β c^{-1-\beta} c − 1 − β has been absorbed in the constant C C C of the above estimate. The assumption δ ≤ 1 2 ϱ \delta\le\tfrac12\varrho δ ≤ 2 1 ϱ is in particular satisfied whenever δ ≤ 1 2 c u ( ζ ) 1 η \delta\le\tfrac12cu(\zeta)^{\frac1\eta} δ ≤ 2 1 c u ( ζ ) η 1 . Therefore, estimate (81) implies that ∇ u \nabla u ∇ u is continuous in any point ζ ∈ Q 1 \zeta\in Q_1 ζ ∈ Q 1 with u ( ζ ) > 0 u(\zeta) > 0 u ( ζ ) > 0 .
Step 2: Continuous differentiability on the zero set. Let z o = ( x o , t o ) ∈ Q 1 z_o=(x_o,t_o)\in Q_1 z o = ( x o , t o ) ∈ Q 1 with u ( z o ) = 0 u(z_o)=0 u ( z o ) = 0 . For a small step-size h ∈ R N ∖ { 0 } h\in\mathbb{R}^N\setminus\{0\} h ∈ R N ∖ { 0 } , estimate (77) 1 _1 1 at the fixed time t o t_o t o gives
0 ≤ u ( x o + h , t o ) ≤ L 1 α ∣ h ∣ 1 α . (82) 0\leq u(x_o+h,t_o)\leq L^{\frac{1}{\alpha}}|h|^{\frac{1}{\alpha}}. \tag*{(82)} 0 ≤ u ( x o + h , t o ) ≤ L α 1 ∣ h ∣ α 1 . (82)
Since α < 1 \alpha<1 α < 1 , we divide (82) by ∣ h ∣ |h| ∣ h ∣ and let h → 0 h\to0 h → 0 to get Fréchet differentiability in the spatial variables at z o z_o z o , with ∇ u ( z o ) = 0 \nabla u(z_o)=0 ∇ u ( z o ) = 0 .
On { u > 0 } ∩ Q 1 \{u>0\}\cap Q_1 { u > 0 } ∩ Q 1 both u u u and ∇ u \nabla u ∇ u are continuous. The almost everywhere weighted bound in (77) consequently holds pointwise there. With the derivative just defined at zeros it becomes
∣ ∇ u ( z ) ∣ ≤ L u ( z ) 1 − α for every z ∈ Q 1 . (83) |\nabla u(z)|\leq Lu(z)^{1-\alpha}\qquad\text{for every }z\in Q_1. \tag*{(83)} ∣∇ u ( z ) ∣ ≤ Lu ( z ) 1 − α for every z ∈ Q 1 . (83)
Consider a sequence Q 1 ∋ z j → z o Q_1\ni z_j\to z_o Q 1 ∋ z j → z o ; use (83) with z j z_j z j , send j → ∞ j\to\infty j → ∞ , recall the continuity of u u u on Q 1 Q_1 Q 1 , and use u ( z o ) = 0 u(z_o)=0 u ( z o ) = 0 to conclude that the right-hand side tends to zero. Thus ∇ u \nabla u ∇ u is jointly continuous across the zero set. We remark that an unweighted Lipschitz estimate would not by itself suffice for this step.
Step 3: Hölder continuity of the gradient. It remains to obtain a uniform Hölder modulus for the gradient. We continue in the normalized setting 0 ≤ u ≤ 1 0\leq u\leq1 0 ≤ u ≤ 1 in Q 2 Q_2 Q 2 and write
A = 1 + β η − 1 > 0 , s ∗ = β A + 1 − α = β η 1 + β − α η , (84) A=\frac{1+\beta}{\eta}-1>0,\qquad s_*=\frac{\beta}{A+1-\alpha}=\frac{\beta\eta}{1+\beta-\alpha\eta}, \tag*{(84)} A = η 1 + β − 1 > 0 , s ∗ = A + 1 − α β = 1 + β − α η β η , (84)
with the parameter β ∈ ( 0 , 1 ) \beta\in(0,1) β ∈ ( 0 , 1 ) from Lemma 7.1 and the parameters α , η ∈ ( 0 , 1 ) \alpha,\eta\in(0,1) α , η ∈ ( 0 , 1 ) from (77) and (78) . All these parameters are now fixed by N N N and p p p . Because α η < 1 \alpha\eta<1 α η < 1 , we have 0 < s ∗ < η 0<s_*<\eta 0 < s ∗ < η . Fix c c c from (79) and choose δ o = δ o ( N , p ) ∈ ( 0 , 1 ) \delta_o=\delta_o(N,p)\in(0,1) δ o = δ o ( N , p ) ∈ ( 0 , 1 ) such that
δ o 1 − s ∗ η ≤ 1 2 c . \delta_o^{1-\frac{s_*}{\eta}}\leq\frac{1}{2}c. δ o 1 − η s ∗ ≤ 2 1 c .
For two points z , ζ ∈ Q 1 z,\zeta\in Q_1 z , ζ ∈ Q 1 , recall the notation δ = dist p a r ( z , ζ ) \delta=\operatorname{dist}_{\mathrm{par}}(z,\zeta) δ = dist par ( z , ζ ) and m = max { u ( z ) , u ( ζ ) } m=\max\{u(z),u(\zeta)\} m = max { u ( z ) , u ( ζ )} . The case δ = 0 \delta=0 δ = 0 is immediate. If δ ≥ δ o \delta\geq\delta_o δ ≥ δ o , (83) yields the desired estimate (87) for a sufficiently large constant C C C . Suppose 0 < δ < δ o 0<\delta<\delta_o 0 < δ < δ o .
If m ≤ δ s ∗ m\leq\delta^{s_*} m ≤ δ s ∗ , the pointwise weighted bound gives
∣ ∇ u ( z ) − ∇ u ( ζ ) ∣ ≤ 2 L m 1 − α ≤ 2 L δ s ∗ ( 1 − α ) . (85) |\nabla u(z)-\nabla u(\zeta)|\leq2Lm^{1-\alpha}\leq2L\delta^{s_*(1-\alpha)}. \tag*{(85)} ∣∇ u ( z ) − ∇ u ( ζ ) ∣ ≤ 2 L m 1 − α ≤ 2 L δ s ∗ ( 1 − α ) . (85)
This case includes pairs for which both values vanish.
If m > δ s ∗ m>\delta^{s_*} m > δ s ∗ , put ϱ = c m 1 η \varrho=cm^{\frac{1}{\eta}} ϱ = c m η 1 . Then
δ ϱ < c − 1 δ 1 − s ∗ η ≤ 1 2 , \frac{\delta}{\varrho}<c^{-1}\delta^{1-\frac{s_*}{\eta}}\leq\frac{1}{2}, ϱ δ < c − 1 δ 1 − η s ∗ ≤ 2 1 ,
so construction from Step 1 and (81) apply. In particular, that construction also proves positivity at the other point; it is not assumed in advance. From (81) and A > 0 A>0 A > 0 we deduce
∣ ∇ u ( z ) − ∇ u ( ζ ) ∣ ≤ C δ β m − A ≤ C δ β − s ∗ A . (86) |\nabla u(z)-\nabla u(\zeta)|\leq C\delta^\beta m^{-A}\leq C\delta^{\beta-s_*A}. \tag*{(86)} ∣∇ u ( z ) − ∇ u ( ζ ) ∣ ≤ C δ β m − A ≤ C δ β − s ∗ A . (86)
The choice of s ∗ s_* s ∗ in (84) balances the exponents:
s ∗ ( 1 − α ) = β − s ∗ A = β η ( 1 − α ) 1 + β − α η = : γ > 0. s_*(1-\alpha)=\beta-s_*A=\frac{\beta\eta(1-\alpha)}{1+\beta-\alpha\eta}=:\gamma>0. s ∗ ( 1 − α ) = β − s ∗ A = 1 + β − α η β η ( 1 − α ) =: γ > 0.
With the exponent γ \gamma γ fixed above, equations (85) and (86) prove
∣ ∇ u ( z ) − ∇ u ( ζ ) ∣ ≤ C dist p a r ( z , ζ ) γ on Q 1 . (87) |\nabla u(z)-\nabla u(\zeta)|\leq C\operatorname{dist}_{\mathrm{par}}(z,\zeta)^\gamma\qquad\text{on }Q_1. \tag*{(87)} ∣∇ u ( z ) − ∇ u ( ζ ) ∣ ≤ C dist par ( z , ζ ) γ on Q 1 . (87)
In the general case u : Q 2 R ( z o ) → [ 0 , M ] u:Q_{2R}(z_o)\to[0,M] u : Q 2 R ( z o ) → [ 0 , M ] , we consider the rescaled function
U ( Y , S ) = M − 1 u ( x o + R Y , t o + R p S ) for ( Y , S ) ∈ Q 2 . U(Y,S)=M^{-1}u(x_o+RY,t_o+R^pS)\qquad\text{for }(Y,S)\in Q_2. U ( Y , S ) = M − 1 u ( x o + R Y , t o + R p S ) for ( Y , S ) ∈ Q 2 .
This is a solution to (1) with 0 ≤ U ≤ 1 0\leq U\leq1 0 ≤ U ≤ 1 , so that (87) is applicable with U U U in place of u u u . Since the spatial gradient scales by ∇ u = ( M / R ) ∇ U \nabla u=(M/R)\nabla U ∇ u = ( M / R ) ∇ U and the metric by dist p a r ( Y , S ) = dist p a r ( z , ζ ) / R \operatorname{dist}_{\mathrm{par}}(Y,S)=\operatorname{dist}_{\mathrm{par}}(z,\zeta)/R dist par ( Y , S ) = dist par ( z , ζ ) / R , we obtain the asserted bound (2) . This completes the proof of the theorem.
AI disclosure. The development of this work and the preparation of the manuscript were partly assisted by AI tools. Verena Bögelein , Fachbereich Mathematik, Universität Salzburg, Hellbrunner Str. 34, 5020 Salzburg, Austria Email address: verena.boegelein@plus.ac.at
Frank Duzaar , Fachbereich Mathematik, Universität Salzburg, Hellbrunner Str. 34, 5020 Salzburg, Austria Email address: frankjohannes.duzaar@plus.ac.at
Ugo Gianazza , Dipartimento di Matematica “F. Casorati”, Università di Pavia, Via Ferrata 5, 27100 Pavia, Italy Email address: ugogia04@unipv.it
Naian Liao , Fachbereich Mathematik, Universität Salzburg, Hellbrunner Str. 34, 5020 Salzburg, Austria Email address: naian.liao@plus.ac.at
Christoph Scheven , Fakultät für Mathematik, Universität Duisburg-Essen, Thea-Leymann-Str. 9, 45127 Essen, Germany Email address: christoph.scheven@uni-due.de