Introduction
Let n ≥ 2 n \ge2 n ≥ 2 , s ∈ ( 0 , 1 ) s \in(0,1) s ∈ ( 0 , 1 ) , and let Ω ⊂ R n \Omega\subset\mathbb{R}^{n} Ω ⊂ R n be an open set with 0 < ∣ Ω ∣ < ∞ 0 < |\Omega| < \infty 0 < ∣Ω∣ < ∞ . We denote by λ s ( Ω ) \lambda_s(\Omega) λ s ( Ω ) the first eigenvalue of the fractional Laplacian ( − Δ ) s (-\Delta)^s ( − Δ ) s in Ω \Omega Ω with homogeneous exterior Dirichlet condition; the precise normalization is fixed in § 2 . If B Ω B_{\Omega} B Ω is a ball with ∣ B Ω ∣ = ∣ Ω ∣ |B_{\Omega}| = |\Omega| ∣ B Ω ∣ = ∣Ω∣ , the fractional Faber–Krahn inequality states that λ s ( Ω ) ≥ λ s ( B Ω ) \lambda_s(\Omega) \ge\lambda_s(B_{\Omega}) λ s ( Ω ) ≥ λ s ( B Ω ) ; it follows from the fractional Pólya–Szegő inequality, see [1 , 3 ] . A relatively natural question in this context is to quantify this inequality and prove that sets with λ s ( Ω ) \lambda_s(\Omega) λ s ( Ω ) close to λ s ( B Ω ) \lambda_s(B_{\Omega}) λ s ( B Ω ) are close to B Ω B_{\Omega} B Ω , up to translation. See e.g. [3 , 4 , 5 , 12 , 13 , 17 ] . We measure the distance of Ω \Omega Ω from balls by its Fraenkel asymmetry
A ( Ω ) = inf a ∈ R n ∣ Ω △ B Ω ( a ) ∣ ∣ Ω ∣ , \mathcal{A}(\Omega) = \inf_{a \in\mathbb{R}^{n}} \frac{|\Omega\mathbin{\triangle} B_{\Omega}(a)|}{|\Omega|}, A ( Ω ) = a ∈ R n inf ∣Ω∣ ∣Ω △ B Ω ( a ) ∣ ,
where B Ω ( a ) B_{\Omega}(a) B Ω ( a ) is the ball of center a a a with ∣ B Ω ( a ) ∣ = ∣ Ω ∣ |B_{\Omega}(a)| = |\Omega| ∣ B Ω ( a ) ∣ = ∣Ω∣ .
Brasco, Cinti and Vita [3, Theorem 1.3] proved that for every open set of finite measure
λ s ( Ω ) λ s ( B Ω ) − 1 ≥ c A ( Ω ) 3 / s \frac{\lambda_s(\Omega)}{\lambda_s(B_{\Omega})} - 1 \ge c\mathcal{A}(\Omega)^{3/s} λ s ( B Ω ) λ s ( Ω ) − 1 ≥ c A ( Ω ) 3/ s
and conjectured [3 Remark 1.6] that the exponent 3 / s 3/s 3/ s can be replaced by 2 2 2 . For the torsional rigidity, the exponent was improved to 2 + 1 / s 2+1/s 2 + 1/ s in [19 Theorem 1.8] . The exponent 2 2 2 is optimal also in the fractional setting, as shown by the volume-preserving ellipsoidal perturbations in Lemma 5.1 . Our main result confirms this conjecture.
Theorem 1.1. For every n ≥ 2 n \ge2 n ≥ 2 and s ∈ ( 0 , 1 ) s \in(0,1) s ∈ ( 0 , 1 ) there is c = c ( n , s ) > 0 c = c(n,s) > 0 c = c ( n , s ) > 0 such that every open set Ω ⊂ R n \Omega\subset\mathbb{R}^{n} Ω ⊂ R n with 0 < ∣ Ω ∣ < ∞ 0 < |\Omega| < \infty 0 < ∣Ω∣ < ∞ satisfies
λ s ( Ω ) λ s ( B Ω ) − 1 ≥ c A ( Ω ) 2 . \frac{\lambda_s(\Omega)}{\lambda_s(B_{\Omega})} - 1 \ge c\mathcal{A}(\Omega)^2. λ s ( B Ω ) λ s ( Ω ) − 1 ≥ c A ( Ω ) 2 .
As in [5 , 17 ] , we actually prove the corresponding quantitative inequality for the fractional torsional rigidity. Let T s ( Ω ) T_s(\Omega) T s ( Ω ) denote the fractional torsional rigidity of Ω \Omega Ω (see § 2 ), and let B = B 1 B = B_1 B = B 1 be the unit ball.
Theorem 1.2. For every n ≥ 2 n \ge2 n ≥ 2 and s ∈ ( 0 , 1 ) s \in(0,1) s ∈ ( 0 , 1 ) there is C = C ( n , s ) > 0 C = C(n,s) > 0 C = C ( n , s ) > 0 such that every open set Ω ⊂ R n \Omega\subset\mathbb{R}^{n} Ω ⊂ R n with ∣ Ω ∣ = ∣ B ∣ |\Omega| = |B| ∣Ω∣ = ∣ B ∣ satisfies
A ( Ω ) 2 ≤ C ( T s ( B ) − T s ( Ω ) ) . \mathcal{A}(\Omega)^2 \le C\bigl(T_s(B) - T_s(\Omega)\bigr). A ( Ω ) 2 ≤ C ( T s ( B ) − T s ( Ω ) ) .
2020 Mathematics Subject Classification. 35R11, 49Q10, 35P15.
Key words and phrases. Fractional Laplacian, Faber–Krahn inequality, Saint-Venant inequality, torsional rigidity, quantitative stability, Fraenkel asymmetry, Caffarelli–Silvestre extension.
Theorem 1.2 is proved in § 4 . Theorem 1.1 follows from it and from the fractional Köhler–Jobin inequality from [2 ] . Since the argument is identical to that in the local case [5 , 17 ] , we omit the proof of this last derivation.
As in [17 ] , the constants in Theorems 1.1 and 1.2 are in principle computable. We do not track their numerical values. We expect the proof to apply mutatis mutandis to the case of the fractional capacity, which would lead to a similar sharp extension of the results obtained in [7 ] .
Strategy of the proof
Our proof of Theorem 1.2 follows the strategy of [17 ] . We refine the level-set/symmetrization approach of [3 ] in a similar spirit as [17 ] refines the proof from [14 ] . For Ω \Omega Ω a given set, we write u Ω u_{\Omega} u Ω for its torsion function and u Ω ∗ u_{\Omega}^{*} u Ω ∗ for the corresponding Schwarz rearrangement. Although not always explicitly presented in this form, the starting point in [3 , 14 ] is to split the torsion deficit as the sum of a deficit between the energy of u Ω u_{\Omega} u Ω and u Ω ∗ u_{\Omega}^{*} u Ω ∗ plus the energy of u Ω ∗ − u B u_{\Omega}^{*}-u_{B} u Ω ∗ − u B , see Lemma 2.2 here. By coarea, this is then transferred to a control on the isoperimetric deficit of the level sets of u Ω u_{\Omega} u Ω . It is then natural to try to use the quantitative isoperimetric inequality from [13 ] to control the Fraenkel asymmetry of these level sets. The difficulty is then to relate these asymmetries to the asymmetry of Ω \Omega Ω . In [3 , 14 ] , this is done by a mean-value-type argument which leads to a nonsharp exponent. The central idea of [17 ] is to replace this Chebyshev-type argument by the fundamental theorem of calculus which then requires estimating the derivative with respect to the height, or somewhat equivalently with respect to volume, of this Fraenkel asymmetry. Interestingly, bounding this derivative involves the strong form of the quantitative isoperimetric inequality, due to Fusco and Julin [12 ] (see also [18 ] ) rather than the one involving the Fraenkel asymmetry from [13 ] .
As in [3 ] , we work with the Caffarelli–Silvestre [6 ] extension U U U of the torsion function u Ω u_{\Omega} u Ω and with the horizontal level sets E t , z = { U ( ⋅ , z ) > t } E_{t,z}=\{U(\cdot,z)>t\} E t , z = { U ( ⋅ , z ) > t } at each height z > 0 z>0 z > 0 . Rearranging U U U in the horizontal variables, we write the torsion deficit as the weighted integral of four nonnegative defects (Lemma 3.1 ). The first two defects D H D_{H} D H and D I D_{I} D I are the same as in [17 ] : D H D_{H} D H comes from the Cauchy–Schwarz inequality on ∂ E t , z \partial E_{t,z} ∂ E t , z and measures the oscillation of ∣ ∇ x U ∣ \lvert\nabla_{x}U\rvert ∣ ∇ x U ∣ , while D I D_{I} D I is the isoperimetric defect of E t , z E_{t,z} E t , z . The last two defects are new compared with the local case. The vertical defect D V D_{V} D V measures the oscillation of ∂ z U \partial_{z}U ∂ z U on ∂ E t , z \partial E_{t,z} ∂ E t , z , and D B D_{B} D B measures the weighted distance between the horizontally rearranged extension and the extension of the torsion function of the ball.
We normalize ∣ Ω ∣ = ω n \lvert\Omega\rvert=\omega_{n} ∣ Ω ∣ = ω n . Following [17 ] , at every height z z z we parametrize the level sets by the radius of a ball of the same volume rather than by t t t . Thus we choose t ( ρ , z ) t(\rho,z) t ( ρ , z ) so that ∣ E t ( ρ , z ) , z ∣ = ω n ρ n \lvert E_{t(\rho,z),z}\rvert=\omega_{n}\rho^{n} ∣ E t ( ρ , z ) , z ∣ = ω n ρ n , set E ρ , z = E t ( ρ , z ) , z E_{\rho,z}=E_{t(\rho,z),z} E ρ , z = E t ( ρ , z ) , z , and consider the unnormalized Fraenkel asymmetry, or moving ball distance in the language of [17 ] , q ( ρ , z ) = inf a ∣ E ρ , z △ B ρ ( a ) ∣ q(\rho,z)=\inf_{a}\lvert E_{\rho,z}\mathbin{\triangle}B_{\rho}(a)\rvert q ( ρ , z ) = inf a ∣ E ρ , z △ B ρ ( a )∣ . To follow the proof of [17 ] closely, we would ideally like to control the derivative of q q q along a curve connecting the point ( 1 , 0 ) (1,0) ( 1 , 0 ) to a point ( ρ , z ) (\rho,z) ( ρ , z ) with ρ , z ≈ 1 / 2 \rho,z\approx1/2 ρ , z ≈ 1/2 . The corner value q ( 1 , 0 ) = ω n A ( Ω ) q(1,0)=\omega_{n}\mathcal{A}(\Omega) q ( 1 , 0 ) = ω n A ( Ω ) , understood as a joint limit, is thus the quantity we want to control. However, there are two difficulties. The first, which might be only technical, concerns regularity in z z z . Following [17 ] , we can prove that q q q is Lipschitz, hence absolutely continuous, in the horizontal variable ρ \rho ρ , see (31) , but we are not able to obtain a similar result in the z z z variable. The key insight, see Proposition 3.6 and Remark 3.7 , is that while q q q is not necessarily absolutely continuous in z z z , the weighted quantity w / w B w/w_{B} w / w B is better behaved. Here w w w is the radial speed of the level-sets of the extension of u Ω u_{\Omega} u Ω (and similarly for w B w_{B} w B ). The second difficulty is that the weights are degenerating as ( ρ , z ) (\rho,z) ( ρ , z ) approaches ( 1 , 0 ) (1,0) ( 1 , 0 ) . To circumvent this issue, we replace pointwise bounds by averaged ones on horizontal windows of variable size in the proof of Theorem 3.10 .
As in [5 , 17 ] , we actually first prove Theorem 1.2 for sets which have a diameter bounded by a universal constant. To reduce the general case to this one we use an iterated truncation/De Giorgi type argument in Proposition 4.2 in a similar way as in [5 Lemma 5.3] and [17 Lemma 4.1] .
Organization of the paper
§ 2 fixes the notation and proves the splitting of the torsion deficit into a horizontal rearrangement loss and a distance to the ball. § 3 proves the sharp quantitative Saint-Venant inequality for sets contained in a fixed ball. § 4 removes the confinement and § 5 proves the sharpness of the exponent. Appendix A contains the estimates for the extension of the torsion function of the ball and the boundary control of the radial profile.
AI disclosure and development of this project
We started thinking about the question of the quantitative stability for the fractional capacity after completing the work [15 ] . After realizing that the approach there would most likely not yield the desired result, we started exploring the adaptation to the non-local setting of the selection principle strategy from [5 ] . We faced, however, several technical difficulties, the first of which is the lack of computations of second-order shape derivatives for non-local problems, see [8 , 9 , 20 ] for (tedious) first-order computations. In spring 2026, with the help of LLMs, we made substantial progress on that research program and we are reasonably optimistic that it could be carried out at the price of extensive technical work. As an illustration of what this route might look like, we provide in the companion paper [16 ] the computation of the second-order shape derivative of the torsion in the case of smooth sets, and show linear quantitative stability of the ball. After the paper [17 ] appeared on arXiv it was very tempting to check if their much more direct strategy could be adapted to the non-local setting. This was done principally by the LLMs with some interactions and simplifications from our end. The first draft of the paper was also generated before thorough human inspection and editing. We take full responsibility for the content of this paper.
Preliminaries and notation
Fix n ≥ 2 n \ge2 n ≥ 2 and s ∈ ( 0 , 1 ) s \in(0,1) s ∈ ( 0 , 1 ) . Constants denoted by C , c > 0 C,c>0 C , c > 0 may depend on n , s n,s n , s ; dependence on any further parameter is displayed. We write A ≲ B A \lesssim B A ≲ B for A ≤ C B A \le CB A ≤ C B and A ≍ B A \asymp B A ≍ B if A ≲ B ≲ A A \lesssim B \lesssim A A ≲ B ≲ A . For an open set Ω ⊂ R n \Omega\subset\mathbb{R}^{n} Ω ⊂ R n of finite measure, set
H 0 s ( Ω ) = C c ∞ ( Ω ) ‾ H s ( R n ) . H_{0}^{s}(\Omega)=\overline{C_{c}^{\infty}(\Omega)}^{H^{s}(\mathbb{R}^{n})}. H 0 s ( Ω ) = C c ∞ ( Ω ) H s ( R n ) .
and
E s ( u , v ) = c n , s 2 ∬ R n × R n ( u ( x ) − u ( y ) ) ( v ( x ) − v ( y ) ) ∣ x − y ∣ n + 2 s d x d y , c n , s = 4 s Γ ( n / 2 + s ) π n / 2 ∣ Γ ( − s ) ∣ . \mathcal{E}_{s}(u,v)=\frac{c_{n,s}}{2}\iint_{\mathbb{R}^{n}\times\mathbb{R}^{n}}\frac{(u(x)-u(y))(v(x)-v(y))}{\lvert x-y\rvert^{n+2s}}\,dx\,dy,\qquad c_{n,s}=\frac{4^{s}\Gamma(n/2+s)}{\pi^{n/2}\lvert\Gamma(-s)\rvert}. E s ( u , v ) = 2 c n , s ∬ R n × R n ∣ x − y ∣ n + 2 s ( u ( x ) − u ( y )) ( v ( x ) − v ( y )) d x d y , c n , s = π n /2 ∣ Γ ( − s )∣ 4 s Γ ( n /2 + s ) .
We write E s ( u ) = E s ( u , u ) \mathcal{E}_{s}(u)=\mathcal{E}_{s}(u,u) E s ( u ) = E s ( u , u ) . With the Fourier convention f ^ ( ξ ) = ∫ R n e − i x ⋅ ξ f ( x ) d x \widehat{f}(\xi)=\int_{\mathbb{R}^{n}}e^{-ix\cdot\xi}f(x)\,dx f ( ξ ) = ∫ R n e − i x ⋅ ξ f ( x ) d x , this is the quadratic form associated with the multiplier ∣ ξ ∣ 2 s \lvert\xi\rvert^{2s} ∣ ξ ∣ 2 s . Thus
( − Δ ) s f ( x ) = c n , s PV ∫ R n f ( x ) − f ( y ) ∣ x − y ∣ n + 2 s d y (-\Delta)^{s}f(x)=c_{n,s}\operatorname{PV}\int_{\mathbb{R}^{n}}\frac{f(x)-f(y)}{\lvert x-y\rvert^{n+2s}}\,dy ( − Δ ) s f ( x ) = c n , s PV ∫ R n ∣ x − y ∣ n + 2 s f ( x ) − f ( y ) d y
whenever the principal value is defined. The first eigenvalue is
λ s ( Ω ) = inf 0 ≠ v ∈ H 0 s ( Ω ) E s ( v ) ∥ v ∥ 2 2 . \lambda_{s}(\Omega)=\inf_{0\ne v\in H_{0}^{s}(\Omega)}\frac{\mathcal{E}_{s}(v)}{\lVert v\rVert_{2}^{2}}. λ s ( Ω ) = 0 = v ∈ H 0 s ( Ω ) inf ∥ v ∥ 2 2 E s ( v ) .
The torsion function u Ω u_{\Omega} u Ω is the unique minimizer of
u Ω = arg min v ∈ H 0 s ( Ω ) { 1 2 E s ( v ) − ∫ Ω v d x } . u_{\Omega}=\operatorname*{arg\,min}_{v\in H_{0}^{s}(\Omega)}\left\{\frac{1}{2}\mathcal{E}_{s}(v)-\int_{\Omega}v\,dx\right\}. u Ω = v ∈ H 0 s ( Ω ) arg min { 2 1 E s ( v ) − ∫ Ω v d x } .
Equivalently, for every φ ∈ H 0 s ( Ω ) \varphi\in H_{0}^{s}(\Omega) φ ∈ H 0 s ( Ω ) ,
E s ( u Ω , φ ) = ∫ Ω φ d x , (1) \mathcal{E}_{s}(u_{\Omega},\varphi)=\int_{\Omega}\varphi\,dx, \tag*{(1)} E s ( u Ω , φ ) = ∫ Ω φ d x , (1)
which is the weak formulation of
{ ( − Δ ) s u Ω = 1 in Ω , u Ω = 0 in R n ∖ Ω . \begin{cases}
(-\Delta)^{s}u_{\Omega}=1 & \text{in }\Omega,\\
u_{\Omega}=0 & \text{in }\mathbb{R}^{n}\setminus\Omega.
\end{cases} { ( − Δ ) s u Ω = 1 u Ω = 0 in Ω , in R n ∖ Ω.
The torsional rigidity is
T s ( Ω ) = ∫ Ω u Ω d x = E s ( u Ω ) = max v ∈ H 0 s ( Ω ) { 2 ∫ Ω v d x − E s ( v ) } . (2) T_{s}(\Omega)=\int_{\Omega}u_{\Omega}\,dx=\mathcal{E}_{s}(u_{\Omega})=\max_{v\in H_{0}^{s}(\Omega)}\left\{2\int_{\Omega}v\,dx-\mathcal{E}_{s}(v)\right\}. \tag*{(2)} T s ( Ω ) = ∫ Ω u Ω d x = E s ( u Ω ) = v ∈ H 0 s ( Ω ) max { 2 ∫ Ω v d x − E s ( v ) } . (2)
Since u Ω ( x ) = r 2 s u Ω ( x / r ) u_{\Omega}(x)=r^{2s}u_{\Omega}(x/r) u Ω ( x ) = r 2 s u Ω ( x / r ) for r > 0 r>0 r > 0 , we have
T s ( r Ω ) = r n + 2 s T s ( Ω ) , λ s ( r Ω ) = r − 2 s λ s ( Ω ) . (3) T_s(r\Omega)=r^{n+2s}T_s(\Omega), \qquad\lambda_s(r\Omega)=r^{-2s}\lambda_s(\Omega). \tag*{(3)} T s ( r Ω ) = r n + 2 s T s ( Ω ) , λ s ( r Ω ) = r − 2 s λ s ( Ω ) . (3)
We work under the normalization ∣ Ω ∣ = ω n = ∣ B ∣ |\Omega|=\omega_n=|B| ∣Ω∣ = ω n = ∣ B ∣ , where B = B 1 B=B_1 B = B 1 , and set
δ T ( Ω ) = T s ( B ) − T s ( Ω ) . \delta_T(\Omega)=T_s(B)-T_s(\Omega). δ T ( Ω ) = T s ( B ) − T s ( Ω ) .
The Caffarelli–Silvestre extension of f ∈ H s ( R n ) f\in H^s(\mathbb{R}^n) f ∈ H s ( R n ) is
Ext s f ( x , z ) = P z ( s ) ∗ f ( x ) , P z ( s ) ( x ) = Γ ( n / 2 + s ) π n / 2 Γ ( s ) z 2 s ( ∣ x ∣ 2 + z 2 ) ( n + 2 s ) / 2 . (4) \operatorname{Ext}_s f(x,z)=P_z^{(s)}*f(x), \qquad P_z^{(s)}(x)=\frac{\Gamma(n/2+s)}{\pi^{n/2}\Gamma(s)}\frac{z^{2s}}{(|x|^2+z^2)^{(n+2s)/2}}. \tag*{(4)} Ext s f ( x , z ) = P z ( s ) ∗ f ( x ) , P z ( s ) ( x ) = π n /2 Γ ( s ) Γ ( n /2 + s ) ( ∣ x ∣ 2 + z 2 ) ( n + 2 s ) /2 z 2 s . (4)
It solves div ( z 1 − 2 s ∇ Ext s f ) = 0 \operatorname{div}(z^{1-2s}\nabla\operatorname{Ext}_s f)=0 div ( z 1 − 2 s ∇ Ext s f ) = 0 in R + n + 1 = R n × ( 0 , ∞ ) \mathbb{R}^{n+1}_+=\mathbb{R}^n\times(0,\infty) R + n + 1 = R n × ( 0 , ∞ ) and [6 ]
d s = 2 2 s − 1 Γ ( s ) Γ ( 1 − s ) , E s ( f ) = d s ∫ R + n + 1 z 1 − 2 s ∣ ∇ Ext s f ∣ 2 d x d z . (5) d_s=2^{2s-1}\frac{\Gamma(s)}{\Gamma(1-s)}, \qquad\mathcal{E}_s(f)=d_s\int_{\mathbb{R}^{n+1}_+}z^{1-2s}|\nabla\operatorname{Ext}_s f|^2\,dx\,dz. \tag*{(5)} d s = 2 2 s − 1 Γ ( 1 − s ) Γ ( s ) , E s ( f ) = d s ∫ R + n + 1 z 1 − 2 s ∣∇ Ext s f ∣ 2 d x d z . (5)
Moreover,
E s ( f ) = d s min F ( x , 0 ) = f ∫ R + n + 1 z 1 − 2 s ∣ ∇ F ∣ 2 d x d z (6) \mathcal{E}_s(f)=d_s\min_{F(x,0)=f}\int_{\mathbb{R}^{n+1}_+}z^{1-2s}|\nabla F|^2\,dx\,dz \tag*{(6)} E s ( f ) = d s F ( x , 0 ) = f min ∫ R + n + 1 z 1 − 2 s ∣∇ F ∣ 2 d x d z (6)
and
( − Δ ) s f = − d s lim z ↓ 0 z 1 − 2 s ∂ z Ext s f . (-\Delta)^s f=-d_s\lim_{z\downarrow0}z^{1-2s}\partial_z\operatorname{Ext}_s f. ( − Δ ) s f = − d s z ↓ 0 lim z 1 − 2 s ∂ z Ext s f .
in the distributional trace sense. Hence U = Ext s u Ω U=\operatorname{Ext}_s u_{\Omega} U = Ext s u Ω satisfies
{ div ( z 1 − 2 s ∇ U ) = 0 in R + n + 1 , U = 0 on ( R n ∖ Ω ) × { 0 } , − d s lim z ↓ 0 z 1 − 2 s U z = 1 on Ω × { 0 } . \begin{cases}
\operatorname{div}(z^{1-2s}\nabla U)=0 & \text{in }\mathbb{R}^{n+1}_+,\\
U=0 & \text{on }(\mathbb{R}^n\setminus\Omega)\times\{0\},\\
-d_s\lim_{z\downarrow0}z^{1-2s}U_z=1 & \text{on }\Omega\times\{0\}.
\end{cases} ⎩ ⎨ ⎧ div ( z 1 − 2 s ∇ U ) = 0 U = 0 − d s lim z ↓ 0 z 1 − 2 s U z = 1 in R + n + 1 , on ( R n ∖ Ω ) × { 0 } , on Ω × { 0 } .
By this we mean that for every finite-energy Φ \Phi Φ whose trace belongs to H 0 s ( Ω ) H^s_0(\Omega) H 0 s ( Ω ) ,
d s ∫ R + n + 1 z 1 − 2 s ∇ U ⋅ ∇ Φ d x d z = ∫ Ω Φ ( x , 0 ) d x . d_s\int_{\mathbb{R}^{n+1}_+}z^{1-2s}\nabla U\cdot\nabla\Phi\,dx\,dz=\int_{\Omega}\Phi(x,0)\,dx. d s ∫ R + n + 1 z 1 − 2 s ∇ U ⋅ ∇Φ d x d z = ∫ Ω Φ ( x , 0 ) d x .
For a nonnegative measurable function f f f vanishing at infinity we denote by f ∗ f^* f ∗ its Schwarz symmetric decreasing rearrangement, and we use the fractional Pólya–Szegő inequality E s ( f ∗ ) ≤ E s ( f ) \mathcal{E}_s(f^*)\leq\mathcal{E}_s(f) E s ( f ∗ ) ≤ E s ( f ) , see [1 ] . Let U ∗ ( ⋅ , z ) U^*(\cdot,z) U ∗ ( ⋅ , z ) be the Schwarz rearrangement of U ( ⋅ , z ) U(\cdot,z) U ( ⋅ , z ) in the horizontal variables, let U ^ = Ext s ( u Ω ∗ ) \widehat{U}=\operatorname{Ext}_s(u_\Omega^*) U = Ext s ( u Ω ∗ ) , and let u B u_B u B and U B U_B U B be the torsion function of B B B and its extension.
For z > 0 z>0 z > 0 , set M ( z ) = max x U ( x , z ) M(z)=\max_x U(x,z) M ( z ) = max x U ( x , z ) and
D = { ( t , z ) : z > 0 , 0 < t < M ( z ) } . D=\{(t,z):z>0,\ 0<t<M(z)\}. D = {( t , z ) : z > 0 , 0 < t < M ( z )} .
For ( t , z ) ∈ D (t,z)\in D ( t , z ) ∈ D define
E t , z = { U ( ⋅ , z ) > t } , m ( t , z ) = ∣ E t , z ∣ , ρ ( t , z ) = ( m ( t , z ) ω n ) 1 / n , m ( t ( ρ , z ) , z ) = ω n ρ n . E_{t,z}=\{U(\cdot,z)>t\}, \qquad m(t,z)=|E_{t,z}|, \qquad\rho(t,z)=\left(\frac{m(t,z)}{\omega_n}\right)^{1/n}, \qquad m(t(\rho,z),z)=\omega_n\rho^n. E t , z = { U ( ⋅ , z ) > t } , m ( t , z ) = ∣ E t , z ∣ , ρ ( t , z ) = ( ω n m ( t , z ) ) 1/ n , m ( t ( ρ , z ) , z ) = ω n ρ n .
By Lemma 3.1 below, m ( ⋅ , z ) m(\cdot,z) m ( ⋅ , z ) is continuous and strictly decreasing, so that t ( ρ , z ) t(\rho,z) t ( ρ , z ) is well defined for every ρ > 0 \rho>0 ρ > 0 , and t ( ρ , z ) = U ∗ ( ρ e 1 , z ) t(\rho,z)=U^*(\rho e_1,z) t ( ρ , z ) = U ∗ ( ρ e 1 , z ) . We also set
E ρ , z = E t ( ρ , z ) , z , w ( ρ , z ) = − ∂ ρ t ( ρ , z ) , P ρ = n ω n ρ n − 1 . E_{\rho,z}=E_{t(\rho,z),z}, \qquad w(\rho,z)=-\partial_\rho t(\rho,z), \qquad P_\rho=n\omega_n\rho^{n-1}. E ρ , z = E t ( ρ , z ) , z , w ( ρ , z ) = − ∂ ρ t ( ρ , z ) , P ρ = n ω n ρ n − 1 .
For the ball profile, put
t B ( ρ , z ) = U B ( ρ e 1 , z ) , w B ( ρ , z ) = − ∂ ρ t B ( ρ , z ) , t_B(\rho,z)=U_B(\rho e_1,z), \qquad w_B(\rho,z)=-\partial_\rho t_B(\rho,z), t B ( ρ , z ) = U B ( ρ e 1 , z ) , w B ( ρ , z ) = − ∂ ρ t B ( ρ , z ) ,
v ( ρ , z ) = t ( ρ , z ) − t B ( ρ , z ) . (7) v(\rho,z)=t(\rho,z)-t_B(\rho,z). \tag*{(7)} v ( ρ , z ) = t ( ρ , z ) − t B ( ρ , z ) . (7)
Notice that
w = w B − v ρ . w=w_B-v_\rho. w = w B − v ρ .
For 0 < ℓ < 1 / 4 0<\ell<1/4 0 < ℓ < 1/4 , write
R = ( 1 − ℓ , 1 ) × ( 0 , ℓ ) . (8) \mathcal{R}=(1-\ell,1)\times(0,\ell). \tag*{(8)} R = ( 1 − ℓ , 1 ) × ( 0 , ℓ ) . (8)
We say that ( t , z ) ∈ D (t,z) \in\mathcal{D} ( t , z ) ∈ D , or the corresponding level, is regular if ∇ x U ( ⋅ , z ) \nabla_x U(\cdot,z) ∇ x U ( ⋅ , z ) does not vanish on { U ( ⋅ , z ) = t } \{U(\cdot,z)=t\} { U ( ⋅ , z ) = t } ; a pair ( ρ , z ) (\rho,z) ( ρ , z ) is regular if ( t ( ρ , z ) , z ) (t(\rho,z),z) ( t ( ρ , z ) , z ) is. With g = ∣ ∇ x U ∣ g=|\nabla_x U| g = ∣ ∇ x U ∣ and h = U z h=U_z h = U z , set
α = ∫ ∂ E t , z 1 g d H n − 1 , β = ∫ ∂ E t , z g d H n − 1 , \alpha=\int_{\partial E_{t,z}}\frac{1}{g}\,d\mathcal{H}^{n-1},\qquad\beta=\int_{\partial E_{t,z}}g\,d\mathcal{H}^{n-1}, α = ∫ ∂ E t , z g 1 d H n − 1 , β = ∫ ∂ E t , z g d H n − 1 ,
γ = ∫ ∂ E t , z h g d H n − 1 , η = ∫ ∂ E t , z h 2 g d H n − 1 . (9) \gamma=\int_{\partial E_{t,z}}\frac{h}{g}\,d\mathcal{H}^{n-1},\qquad\eta=\int_{\partial E_{t,z}}\frac{h^2}{g}\,d\mathcal{H}^{n-1}. \tag*{(9)} γ = ∫ ∂ E t , z g h d H n − 1 , η = ∫ ∂ E t , z g h 2 d H n − 1 . (9)
For the local torsion problem, Gauss–Green and − Δ u = 1 -\Delta u=1 − Δ u = 1 give β ( t ) = m ( t ) \beta(t)=m(t) β ( t ) = m ( t ) which is used in [17 Proposition 2.5] to simplify some computations. Here we cannot rely on such a simplification.
Let P = Per ( E t , z ) P=\operatorname{Per}(E_{t,z}) P = Per ( E t , z ) and P ∗ = P ρ ( t , z ) P_* = P_{\rho(t,z)} P ∗ = P ρ ( t , z ) . Define
D H = α β − P 2 , D I = P 2 − P ∗ 2 , D_H=\alpha\beta-P^2,\qquad D_I=P^2-P_*^2, D H = α β − P 2 , D I = P 2 − P ∗ 2 ,
D V = α η − γ 2 , D B = ∣ ( − P ∗ γ ) − α ∇ t B ( ρ ( t , z ) , z ) ∣ 2 . D_V=\alpha\eta-\gamma^2,\qquad D_B=\left|\begin{pmatrix}-P_*\\ \gamma\end{pmatrix}-\alpha\nabla_t B(\rho(t,z),z)\right|^2. D V = α η − γ 2 , D B = ( − P ∗ γ ) − α ∇ t B ( ρ ( t , z ) , z ) 2 .
The quotients involving these quantities are set equal to zero on nonregular levels. In radius coordinates they are evaluated at ( t ( ρ , z ) , z ) (t(\rho,z),z) ( t ( ρ , z ) , z ) .
Finally, for a cap A > 0 A>0 A > 0 , define
Q ( t , z ) = inf a ∈ R n ∣ E t , z △ B ρ ( t , z ) ( a ) ∣ , q ( ρ , z ) = Q ( t ( ρ , z ) , z ) , Q(t,z)=\inf_{a\in\mathbb{R}^n}|E_{t,z}\mathbin{\triangle}B_{\rho(t,z)}(a)|,\qquad q(\rho,z)=Q(t(\rho,z),z), Q ( t , z ) = a ∈ R n inf ∣ E t , z △ B ρ ( t , z ) ( a ) ∣ , q ( ρ , z ) = Q ( t ( ρ , z ) , z ) ,
Q A = min { Q , A } , q A = min { q , A } . (10) Q_A=\min\{Q,A\},\qquad q_A=\min\{q,A\}. \tag*{(10)} Q A = min { Q , A } , q A = min { q , A } . (10)
The cap plays a role only in the proof of (47) , where A = ω n A ( Ω ) A=\omega_n A(\Omega) A = ω n A ( Ω ) : it makes the error terms there proportional to A A A , see the discussion before (48) . At regular levels, the fixed-volume direction is
Z = ∂ z + γ α ∂ t , Z m = 0. (11) \mathcal{Z}=\partial_z+\frac{\gamma}{\alpha}\partial_t,\qquad\mathcal{Z}m=0. \tag*{(11)} Z = ∂ z + α γ ∂ t , Z m = 0. (11)
Horizontal rearrangement of the extension
Lemma 2.1. Let Ω ⊂ R n \Omega\subset\mathbb{R}^n Ω ⊂ R n be bounded and open, with ∣ Ω ∣ = ω n |\Omega|=\omega_n ∣Ω∣ = ω n , and let U = Ext s u Ω U=\operatorname{Ext}_s u_\Omega U = Ext s u Ω . For every z > 0 z>0 z > 0 , the function U ( ⋅ , z ) U(\cdot,z) U ( ⋅ , z ) is nonnegative, belongs to H 1 ( R n ) H^1(\mathbb{R}^n) H 1 ( R n ) , and vanishes at infinity. It is positive and real analytic, its positive superlevel sets are bounded, and its level sets and horizontal critical set are Lebesgue null. The horizontal rearrangement U ∗ U^* U ∗ has weak derivatives in R + n + 1 \mathbb{R}^{n+1}_+ R + n + 1 and, for almost every z > 0 z>0 z > 0 ,
∥ ∇ x U ∗ ( ⋅ , z ) ∥ 2 ≤ ∥ ∇ x U ( ⋅ , z ) ∥ 2 , ∥ ∂ z U ∗ ( ⋅ , z ) ∥ 2 ≤ ∥ U z ( ⋅ , z ) ∥ 2 . (12) \|\nabla_x U^*(\cdot,z)\|_2\leq\|\nabla_x U(\cdot,z)\|_2,\qquad\|\partial_z U^*(\cdot,z)\|_2\leq\|U_z(\cdot,z)\|_2. \tag*{(12)} ∥ ∇ x U ∗ ( ⋅ , z ) ∥ 2 ≤ ∥ ∇ x U ( ⋅ , z ) ∥ 2 , ∥ ∂ z U ∗ ( ⋅ , z ) ∥ 2 ≤ ∥ U z ( ⋅ , z ) ∥ 2 . (12) In particular U ∗ U^* U ∗ has finite weighted Dirichlet energy. Moreover, as z ↓ 0 z\downarrow0 z ↓ 0 ,
∥ U ∗ ( ⋅ , z ) − u Ω ∗ ∥ 2 ≤ ∥ U ( ⋅ , z ) − u Ω ∥ 2 ⟶ 0 , (13) \|U^*(\cdot,z)-u_\Omega^*\|_2\leq\|U(\cdot,z)-u_\Omega\|_2\longrightarrow0, \tag*{(13)} ∥ U ∗ ( ⋅ , z ) − u Ω ∗ ∥ 2 ≤ ∥ U ( ⋅ , z ) − u Ω ∥ 2 ⟶ 0 , (13) so that the bottom trace of U ∗ U^* U ∗ is u Ω ∗ u_\Omega^* u Ω ∗ .
Proof. Write u = u Ω u=u_\Omega u = u Ω . It is nonnegative, belongs to L 1 ∩ L 2 L^1\cap L^2 L 1 ∩ L 2 , and has bounded support. For z > 0 z>0 z > 0 the kernel P z ( s ) P_z^{(s)} P z ( s ) and its gradient are integrable, so U ( ⋅ , z ) = P z ( s ) ∗ u U(\cdot,z)=P_z^{(s)}*u U ( ⋅ , z ) = P z ( s ) ∗ u and ∇ x U ( ⋅ , z ) = ( ∇ P z ( s ) ) ∗ u \nabla_xU(\cdot,z)=(\nabla P_z^{(s)})*u ∇ x U ( ⋅ , z ) = ( ∇ P z ( s ) ) ∗ u belong to L 2 ( R n ) L^2(\mathbb{R}^n) L 2 ( R n ) , and U ( ⋅ , z ) U(\cdot,z) U ( ⋅ , z ) tends to zero at infinity. Real analyticity follows from the formula for P z ( s ) P_z^{(s)} P z ( s ) . Its positive superlevel sets are bounded, uniformly when z z z ranges in a compact subinterval of ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) and t t t stays bounded away from zero. Since at least one coordinate derivative is a nonzero analytic function,
∣ { x : ∇ x U ( x , z ) = 0 } ∣ = 0. |\{x:\nabla_xU(x,z)=0\}|=0. ∣ { x : ∇ x U ( x , z ) = 0 } ∣ = 0.
Likewise, every level set has zero n n n -dimensional measure.
Horizontal Pólya–Szegő gives the first inequality in (12) . The Hardy–Littlewood–Pólya inequality gives
∥ U ∗ ( ⋅ , z + ε ) − U ∗ ( ⋅ , z ) ∥ 2 ≤ ∥ U ( ⋅ , z + ε ) − U ( ⋅ , z ) ∥ 2 . \|U^*(\cdot,z+\varepsilon)-U^*(\cdot,z)\|_2\leq\|U(\cdot,z+\varepsilon)-U(\cdot,z)\|_2. ∥ U ∗ ( ⋅ , z + ε ) − U ∗ ( ⋅ , z ) ∥ 2 ≤ ∥ U ( ⋅ , z + ε ) − U ( ⋅ , z ) ∥ 2 .
On every compact height interval, z ↦ U ∗ ( ⋅ , z ) z\mapsto U^*(\cdot,z) z ↦ U ∗ ( ⋅ , z ) is therefore Lipschitz with values in L 2 ( R n ) L^2(\mathbb{R}^n) L 2 ( R n ) , and difference quotients give a finite vertical derivative satisfying the second inequality in (12) for almost every z z z . Multiplying these slice inequalities by z 1 − 2 s z^{1-2s} z 1 − 2 s and integrating shows that U ∗ U^* U ∗ has finite weighted Dirichlet energy. On compact positive height intervals the weight is bounded above and below. By (12) , U ∗ ∈ L l o c 2 ( ( 0 , ∞ ) ; H 1 ( R n ) ) U^* \in L^2_{\mathrm{loc}}((0,\infty);H^1(\mathbb{R}^n)) U ∗ ∈ L loc 2 (( 0 , ∞ ) ; H 1 ( R n )) and ∂ z U ∗ ∈ L l o c 2 ( ( 0 , ∞ ) ; L 2 ( R n ) ) \partial_z U^* \in L^2_{\mathrm{loc}}((0,\infty);L^2(\mathbb{R}^n)) ∂ z U ∗ ∈ L loc 2 (( 0 , ∞ ) ; L 2 ( R n )) , so by Fubini the slice-wise gradient ∇ x U ∗ \nabla_x U^* ∇ x U ∗ and the weak vertical derivative are the joint weak derivatives, and U ∗ ∈ H l o c 1 ( R n × ( 0 , ∞ ) ) U^* \in H^1_{\mathrm{loc}}(\mathbb{R}^n \times(0,\infty)) U ∗ ∈ H loc 1 ( R n × ( 0 , ∞ )) . Finally, the kernel P z ( s ) P_z^{(s)} P z ( s ) is an approximate identity, so U ( ⋅ , z ) → u U(\cdot,z) \to u U ( ⋅ , z ) → u in L 2 ( R n ) L^2(\mathbb{R}^n) L 2 ( R n ) , and the Hardy–Littlewood–Pólya inequality gives (13) . □
Exact torsion and extension splittings
Lemma 2.2. Let Ω ⊂ R n \Omega\subset\mathbb{R}^n Ω ⊂ R n be bounded and open, with ∣ Ω ∣ = ω n |\Omega|=\omega_n ∣Ω∣ = ω n , let u = u Ω u=u_\Omega u = u Ω and U = Ext s u U=\operatorname{Ext}_s u U = Ext s u , and let U ∗ U^* U ∗ and U B U_B U B be as above. Then
δ T ( Ω ) = E s ( u ) − E s ( u ∗ ) + E s ( u ∗ − u B ) . (14) \delta_T(\Omega)=\mathcal{E}_s(u)-\mathcal{E}_s(u^*)+\mathcal{E}_s(u^*-u_B). \tag*{(14)} δ T ( Ω ) = E s ( u ) − E s ( u ∗ ) + E s ( u ∗ − u B ) . (14) The difference E s ( u ) − E s ( u ∗ ) \mathcal{E}_s(u)-\mathcal{E}_s(u^*) E s ( u ) − E s ( u ∗ ) and the last term are nonnegative. In extension variables, with d s d_s d s as in (5) , one also has
δ T ( Ω ) = d s ∫ R + n + 1 z 1 − 2 s ( ∣ ∇ U ∣ 2 − ∣ ∇ U ∗ ∣ 2 + ∣ ∇ ( U ∗ − U B ) ∣ 2 ) d x d z . (15) \delta_T(\Omega)=d_s\int_{\mathbb{R}^{n+1}_+}z^{1-2s}\left(|\nabla U|^2-|\nabla U^*|^2+|\nabla(U^*-U_B)|^2\right)\,dx\,dz. \tag*{(15)} δ T ( Ω ) = d s ∫ R + n + 1 z 1 − 2 s ( ∣∇ U ∣ 2 − ∣∇ U ∗ ∣ 2 + ∣∇ ( U ∗ − U B ) ∣ 2 ) d x d z . (15) In particular, by (12) ,
∫ R + n + 1 z 1 − 2 s ∣ ∇ ( U ∗ − U B ) ∣ 2 d x d z ≤ C δ T ( Ω ) . (16) \int_{\mathbb{R}^{n+1}_+}z^{1-2s}|\nabla(U^*-U_B)|^2\,dx\,dz\leq C\delta_T(\Omega). \tag*{(16)} ∫ R + n + 1 z 1 − 2 s ∣∇ ( U ∗ − U B ) ∣ 2 d x d z ≤ C δ T ( Ω ) . (16)
Proof. Fractional Pólya–Szegő gives E s ( u ∗ ) ≤ E s ( u ) \mathcal{E}_s(u^*)\leq\mathcal{E}_s(u) E s ( u ∗ ) ≤ E s ( u ) . The rearranged function belongs to H 0 s ( B ) H^s_0(B) H 0 s ( B ) : it is in H s ( R n ) H^s(\mathbb{R}^n) H s ( R n ) and is supported in B ‾ \overline{B} B . For 0 < r < 1 0<r<1 0 < r < 1 , the inward dilation u ∗ ( x / r ) u^*(x/r) u ∗ ( x / r ) is supported in B r B_r B r and converges to u ∗ u^* u ∗ in H s ( R n ) H^s(\mathbb{R}^n) H s ( R n ) as r ↑ 1 r\uparrow1 r ↑ 1 . Mollification at a scale smaller than 1 − r 1-r 1 − r then gives an approximation by C c ∞ ( B ) C^\infty_c(B) C c ∞ ( B ) functions. Thus it is an admissible test in the ball torsion equation under the fixed closure convention. We then have
E s ( u B , u ∗ ) = ∫ B u ∗ = T s ( Ω ) , E s ( u B ) = T s ( B ) , E s ( u ) = T s ( Ω ) . \mathcal{E}_s(u_B,u^*)=\int_B u^*=T_s(\Omega),\qquad\mathcal{E}_s(u_B)=T_s(B),\qquad\mathcal{E}_s(u)=T_s(\Omega). E s ( u B , u ∗ ) = ∫ B u ∗ = T s ( Ω ) , E s ( u B ) = T s ( B ) , E s ( u ) = T s ( Ω ) .
Consequently, after expanding the squares,
E s ( u ) − E s ( u ∗ ) + E s ( u ∗ − u B ) = T s ( Ω ) + T s ( B ) − 2 T s ( Ω ) = δ T ( Ω ) , \mathcal{E}_s(u)-\mathcal{E}_s(u^*)+\mathcal{E}_s(u^*-u_B)=T_s(\Omega)+T_s(B)-2T_s(\Omega)=\delta_T(\Omega), E s ( u ) − E s ( u ∗ ) + E s ( u ∗ − u B ) = T s ( Ω ) + T s ( B ) − 2 T s ( Ω ) = δ T ( Ω ) ,
which proves (14) and the nonnegativity of its two grouped terms.
In this proof write
⟨ F , G ⟩ = d s ∫ R + n + 1 z 1 − 2 s ∇ F ⋅ ∇ G d x d z , ∥ F ∥ 2 = ⟨ F , F ⟩ . \langle F,G\rangle=d_s\int_{\mathbb{R}^{n+1}_+}z^{1-2s}\nabla F\cdot\nabla G\,dx\,dz,\qquad\lVert F\rVert^2=\langle F,F\rangle. ⟨ F , G ⟩ = d s ∫ R + n + 1 z 1 − 2 s ∇ F ⋅ ∇ G d x d z , ∥ F ∥ 2 = ⟨ F , F ⟩ .
In particular, ∥ Ext s f ∥ 2 = E s ( f ) \lVert\operatorname{Ext}_s f\rVert^2=\mathcal{E}_s(f) ∥ Ext s f ∥ 2 = E s ( f ) . Recall that U ^ = Ext s ( u ∗ ) \widehat{U}=\operatorname{Ext}_s(u^*) U = Ext s ( u ∗ ) . Lemma 2.1 shows that U ∗ U^* U ∗ has finite weighted energy and bottom trace u ∗ u^* u ∗ . Thus U ∗ − U ^ U^*-\widehat{U} U ∗ − U has zero weighted trace. For any finite-energy weighted-harmonic extension G G G , the variation G + ε ( U ∗ − U ^ ) G+\varepsilon(U^*-\widehat{U}) G + ε ( U ∗ − U ) has the same trace as G G G . Using the variational characterization (6) of E s a n d d i f f e r e n t i a t i o n a t \mathcal{E}_s and differentiation at E s an dd i f f er e n t ia t i o na t \varepsilon=0$ give
⟨ G , U ∗ − U ^ ⟩ = 0. (17) \langle G,U^*-\widehat{U}\rangle=0. \tag*{(17)} ⟨ G , U ∗ − U ⟩ = 0. (17)
Apply (17) to G = U ^ G=\widehat{U} G = U and to G = U ^ − U B = Ext s ( u ∗ − u B ) G=\widehat{U}-U_B=\operatorname{Ext}_s(u^*-u_B) G = U − U B = Ext s ( u ∗ − u B ) . We obtain
∥ U ∗ ∥ 2 = ∥ U ^ ∥ 2 + ∥ U ∗ − U ^ ∥ 2 , \lVert U^*\rVert^2=\lVert\widehat{U}\rVert^2+\lVert U^*-\widehat{U}\rVert^2, ∥ U ∗ ∥ 2 = ∥ U ∥ 2 + ∥ U ∗ − U ∥ 2 ,
∥ U ∗ − U B ∥ 2 = ∥ U ∗ − U ^ ∥ 2 + ∥ U ^ − U B ∥ 2 . \lVert U^*-U_B\rVert^2=\lVert U^*-\widehat{U}\rVert^2+\lVert\widehat{U}-U_B\rVert^2. ∥ U ∗ − U B ∥ 2 = ∥ U ∗ − U ∥ 2 + ∥ U − U B ∥ 2 .
Substitution into (14) cancels the common ∥ U ∗ − U ^ ∥ 2 \lVert U^*-\widehat{U}\rVert^2 ∥ U ∗ − U ∥ 2 term and gives
δ T ( Ω ) = ∥ U ∥ 2 − ∥ U ∗ ∥ 2 + ∥ U ∗ − U B ∥ 2 , \delta_T(\Omega)=\lVert U\rVert^2-\lVert U^*\rVert^2+\lVert U^*-U_B\rVert^2, δ T ( Ω ) = ∥ U ∥ 2 − ∥ U ∗ ∥ 2 + ∥ U ∗ − U B ∥ 2 ,
which is (15) . □
Ball speed and boundary profile
Two further estimates are used in § 3 ; since their proofs are independent of the rest of the argument, they are given in Appendix A. Lemma A.1 shows that there is ℓ ∈ ( 0 , 1 / 16 ) \ell\in(0,1/16) ℓ ∈ ( 0 , 1/16 ) , depending only on n n n and s s s , such that the radial speed w B w_B w B of the ball extension satisfies, on the rectangle R \mathcal{R} R of (8) ,
w B ( ρ , z ) ≳ ( 1 − ρ + z ) s − 1 , ∣ ∇ w B ( ρ , z ) ∣ ≲ w B ( ρ , z ) 1 − ρ + z . w_B(\rho,z) \gtrsim(1-\rho+z)^{s-1}, \qquad\lvert\nabla w_B(\rho,z)\rvert\lesssim\frac{w_B(\rho,z)}{1-\rho+z}. w B ( ρ , z ) ≳ ( 1 − ρ + z ) s − 1 , ∣ ∇ w B ( ρ , z )∣ ≲ 1 − ρ + z w B ( ρ , z ) .
Lemma A.2 shows that the profile error v = t − t B v=t-t_B v = t − t B satisfies
∫ R z 1 − 2 s ∣ ∇ v ∣ 2 d ρ d z + ∫ 1 − ℓ 1 ∣ v ( ρ , 0 ) ∣ 2 ( 1 − ρ ) 2 s d ρ ≤ C ( ℓ ) δ T ( Ω ) , \int_{\mathcal{R}} z^{1-2s}\lvert\nabla v\rvert^2\,d\rho\,dz+\int_{1-\ell}^{1}\frac{\lvert v(\rho,0)\rvert^2}{(1-\rho)^{2s}}\,d\rho\leq C(\ell)\delta_T(\Omega), ∫ R z 1 − 2 s ∣ ∇ v ∣ 2 d ρ d z + ∫ 1 − ℓ 1 ( 1 − ρ ) 2 s ∣ v ( ρ , 0 ) ∣ 2 d ρ ≤ C ( ℓ ) δ T ( Ω ) ,
where v ( ⋅ , 0 ) v(\cdot,0) v ( ⋅ , 0 ) is the bottom trace of v v v .
Quantitative torsion inequality for bounded sets
Four-defect identity
The next lemma is the counterpart of the level-set identity [17 Theorem 2.4] for the local torsion function and of its expression in radius coordinates [17 Proposition 2.5] . The defects D H D_H D H and D I D_I D I are those of [17 ] , computed on the horizontal slices of the extension; the vertical defect D V D_V D V and the profile defect D B D_B D B have no local counterpart.
Using α w = P ρ \alpha w=P_\rho α w = P ρ , proved below, the two horizontal defects can still be combined as
w P ρ 2 ( D H + D I ) = β P ρ − w ≥ 0. \frac{w}{P_\rho^2}(D_H+D_I)=\frac{\beta}{P_\rho}-w\geq0. P ρ 2 w ( D H + D I ) = P ρ β − w ≥ 0.
In the local case, β = m = ω n ρ n \beta=m=\omega_n\rho^n β = m = ω n ρ n reduces the right-hand side to ρ / n − w \rho/n-w ρ / n − w , the nonnegative difference from the ball speed. For the extension, β / P ρ \beta/P_\rho β / P ρ is not prescribed by the level-set volume, as explained after (9) . This identity therefore does not give a pointwise comparison between w w w and w B w_B w B . The profile defect D B D_B D B provides a nonnegative energy control of their difference instead.
Lemma 3.1. Let Ω ⊂ R n \Omega\subset\mathbb{R}^n Ω ⊂ R n be bounded and open, with ∣ Ω ∣ = ω n \lvert\Omega\rvert=\omega_n ∣ Ω ∣ = ω n , let u = u Ω u=u_\Omega u = u Ω and U = Ext s u U=\operatorname{Ext}_s u U = Ext s u . The function m ( ⋅ , z ) m(\cdot,z) m ( ⋅ , z ) is continuous, strictly decreasing and locally absolutely continuous on ( 0 , M ( z ) ) (0,M(z)) ( 0 , M ( z )) , with m ( t , z ) → + ∞ m(t,z)\to+\infty m ( t , z ) → + ∞ as t ↓ 0 t\downarrow0 t ↓ 0 and m ( t , z ) → 0 m(t,z)\to0 m ( t , z ) → 0 as t ↑ M ( z ) t\uparrow M(z) t ↑ M ( z ) . The nonregular pairs form a null set, both in the ( t , z ) (t,z) ( t , z ) and in the ( ρ , z ) (\rho,z) ( ρ , z ) variables, and at regular pairs
m t = − α , m z = γ , ∂ ρ t = − P ρ α , ∂ z t = γ α . (18) m_t=-\alpha,\qquad m_z=\gamma,\qquad\partial_\rho t=-\frac{P_\rho}{\alpha},\qquad\partial_z t=\frac{\gamma}{\alpha}. \tag*{(18)} m t = − α , m z = γ , ∂ ρ t = − α P ρ , ∂ z t = α γ . (18) All four defects are nonnegative and
δ T ( Ω ) = T s ( B ) − T s ( Ω ) = d s ∫ 0 ∞ z 1 − 2 s ∫ 0 M ( z ) D H + D I + D B + D V α d t d z . (19) \delta_T(\Omega)=T_s(B)-T_s(\Omega)=d_s\int_0^\infty z^{1-2s}\int_0^{M(z)}\frac{D_H+D_I+D_B+D_V}{\alpha}\,dt\,dz. \tag*{(19)} δ T ( Ω ) = T s ( B ) − T s ( Ω ) = d s ∫ 0 ∞ z 1 − 2 s ∫ 0 M ( z ) α D H + D I + D B + D V d t d z . (19) More precisely,
∫ R + n + 1 z 1 − 2 s ( ∣ ∇ x U ∣ 2 − ∣ ∇ x U ∗ ∣ 2 ) d x d z = ∫ 0 ∞ z 1 − 2 s ∫ 0 M ( z ) D H + D I α d t d z . (20) \int_{\mathbb{R}^{n+1}_+}z^{1-2s}\left(\lvert\nabla_xU\rvert^2-\lvert\nabla_xU^*\rvert^2\right)\,dx\,dz=\int_0^\infty z^{1-2s}\int_0^{M(z)}\frac{D_H+D_I}{\alpha}\,dt\,dz. \tag*{(20)} ∫ R + n + 1 z 1 − 2 s ( ∣ ∇ x U ∣ 2 − ∣ ∇ x U ∗ ∣ 2 ) d x d z = ∫ 0 ∞ z 1 − 2 s ∫ 0 M ( z ) α D H + D I d t d z . (20) ∫ R + n + 1 z 1 − 2 s ( ∣ U z ∣ 2 − ∣ ∂ z U ∗ ∣ 2 ) d x d z = ∫ 0 ∞ z 1 − 2 s ∫ 0 M ( z ) D V α d t d z . (21) \int_{\mathbb{R}^{n+1}_+}z^{1-2s}\left(\lvert U_z\rvert^2-\lvert\partial_zU^*\rvert^2\right)\,dx\,dz=\int_0^\infty z^{1-2s}\int_0^{M(z)}\frac{D_V}{\alpha}\,dt\,dz. \tag*{(21)} ∫ R + n + 1 z 1 − 2 s ( ∣ U z ∣ 2 − ∣ ∂ z U ∗ ∣ 2 ) d x d z = ∫ 0 ∞ z 1 − 2 s ∫ 0 M ( z ) α D V d t d z . (21) ∫ R + n + 1 z 1 − 2 s ∣ ∇ ( U ∗ − U B ) ∣ 2 d x d z = ∫ 0 ∞ z 1 − 2 s ∫ 0 M ( z ) D B α d t d z . (22) \int_{\mathbb{R}^{n+1}_+}z^{1-2s}\lvert\nabla(U^*-U_B)\rvert^2\,dx\,dz=\int_0^\infty z^{1-2s}\int_0^{M(z)}\frac{D_B}{\alpha}\,dt\,dz. \tag*{(22)} ∫ R + n + 1 z 1 − 2 s ∣ ∇ ( U ∗ − U B ) ∣ 2 d x d z = ∫ 0 ∞ z 1 − 2 s ∫ 0 M ( z ) α D B d t d z . (22) In radius coordinates this reads
δ T ( Ω ) = d s ∫ 0 ∞ z 1 − 2 s ∫ 0 ∞ w 2 P ρ ( D H + D I + D B + D V ) d ρ d z . (23) \delta_T(\Omega)=d_s\int_0^\infty z^{1-2s}\int_0^\infty\frac{w^2}{P_\rho}(D_H+D_I+D_B+D_V)\,d\rho\,dz. \tag*{(23)} δ T ( Ω ) = d s ∫ 0 ∞ z 1 − 2 s ∫ 0 ∞ P ρ w 2 ( D H + D I + D B + D V ) d ρ d z . (23) Consequently, on the fixed rectangle R \mathcal{R} R ,
∫ R z 1 − 2 s [ w 2 P ρ ( D H + D I + D V ) + P ρ ∣ ∇ v ∣ 2 ] d ρ d z ≤ C δ T ( Ω ) . (24) \int_{\mathcal{R}}z^{1-2s}\left[\frac{w^2}{P_\rho}(D_H+D_I+D_V)+P_\rho\lvert\nabla v\rvert^2\right]\,d\rho\,dz\leq C\delta_T(\Omega). \tag*{(24)} ∫ R z 1 − 2 s [ P ρ w 2 ( D H + D I + D V ) + P ρ ∣ ∇ v ∣ 2 ] d ρ d z ≤ C δ T ( Ω ) . (24)
Proof. Coarea gives
m ( a , z ) − m ( b , z ) = ∫ a b α ( t , z ) d t , 0 < a < b < M ( z ) . (25) m(a,z)-m(b,z)=\int_a^b \alpha(t,z)\,dt,\qquad0<a<b<M(z). \tag*{(25)} m ( a , z ) − m ( b , z ) = ∫ a b α ( t , z ) d t , 0 < a < b < M ( z ) . (25)
The function m ( ⋅ , z ) m(\cdot,z) m ( ⋅ , z ) is locally absolutely continuous and strictly decreasing: the preimage of every nonempty interval of values inside ( 0 , M ( z ) ) (0,M(z)) ( 0 , M ( z )) contains a nonempty open set. Its endpoint limits are + ∞ +\infty + ∞ at t = 0 t=0 t = 0 and zero at t = M ( z ) t=M(z) t = M ( z ) . Hence t ( ρ , z ) t(\rho,z) t ( ρ , z ) is uniquely defined for every ρ > 0 \rho>0 ρ > 0 .
At a regular pair ( t , z ) (t,z) ( t , z ) the outward normal to ∂ E t , z \partial E_{t,z} ∂ E t , z is − ∇ x U / g -\nabla_x U/g − ∇ x U / g . At fixed t t t , its outward normal velocity is h / g h/g h / g , obtained by differentiating U ( x , z ) = t U(x,z)=t U ( x , z ) = t . This gives the first two identities in (18) ; the last two follow by differentiating m ( t ( ρ , z ) , z ) = ω n ρ n m(t(\rho,z),z)=\omega_n\rho^n m ( t ( ρ , z ) , z ) = ω n ρ n .
For fixed z z z , coarea and the nullity of the horizontal critical set give
∫ R n ∣ ∇ x U ∣ 2 d x = ∫ 0 M ( z ) β d t , ∫ R n ∣ U z ∣ 2 d x = ∫ 0 M ( z ) η d t . \int_{\mathbb{R}^n}|\nabla_x U|^2\,dx=\int_0^{M(z)}\beta\,dt,\qquad
\int_{\mathbb{R}^n}|U_z|^2\,dx=\int_0^{M(z)}\eta\,dt. ∫ R n ∣ ∇ x U ∣ 2 d x = ∫ 0 M ( z ) β d t , ∫ R n ∣ U z ∣ 2 d x = ∫ 0 M ( z ) η d t .
The monotone map t ↦ ρ ( t , z ) t\mapsto\rho(t,z) t ↦ ρ ( t , z ) is locally absolutely continuous, with ∣ ∂ t ρ ∣ = α / P ∗ |\partial_t\rho|=\alpha/P_* ∣ ∂ t ρ ∣ = α / P ∗ . Polar integration and (18) therefore give, for a.e. z z z ,
∫ R n ∣ ∇ x U ∗ ∣ 2 d x = ∫ 0 M ( z ) P ∗ 2 α d t , ∫ R n ∣ ∂ z U ∗ ∣ 2 d x = ∫ 0 M ( z ) γ 2 α d t . \int_{\mathbb{R}^n}|\nabla_x U^*|^2\,dx=\int_0^{M(z)}\frac{P_*^2}{\alpha}\,dt,\qquad
\int_{\mathbb{R}^n}|\partial_z U^*|^2\,dx=\int_0^{M(z)}\frac{\gamma^2}{\alpha}\,dt. ∫ R n ∣ ∇ x U ∗ ∣ 2 d x = ∫ 0 M ( z ) α P ∗ 2 d t , ∫ R n ∣ ∂ z U ∗ ∣ 2 d x = ∫ 0 M ( z ) α γ 2 d t .
Multiplying by z 1 − 2 s z^{1-2s} z 1 − 2 s , integrating in height, and subtracting proves (20) and (21) , since α β − P ∗ 2 = D H + D I \alpha\beta-P_*^2=D_H+D_I α β − P ∗ 2 = D H + D I and α η − γ 2 = D V \alpha\eta-\gamma^2=D_V α η − γ 2 = D V .
For the profile error, both U ∗ U^* U ∗ and U B U_B U B are radial in x x x . Using the same change of variables, now with P ρ ∣ d ρ ∣ = α d t P_\rho|d\rho|=\alpha\,dt P ρ ∣ d ρ ∣ = α d t , gives
∫ R + n + 1 z 1 − 2 s ∣ ∇ ( U ∗ − U B ) ∣ 2 d x d z = ∫ 0 ∞ z 1 − 2 s ∫ 0 ∞ P ρ ∣ ∇ v ( ρ , z ) ∣ 2 d ρ d z = ∫ 0 ∞ z 1 − 2 s ∫ 0 M ( z ) α ∣ ∇ v ( ρ ( t , z ) , z ) ∣ 2 d t d z = ∫ 0 ∞ z 1 − 2 s ∫ 0 M ( z ) D B α d t d z . \begin{aligned}
\int_{\mathbb{R}^{n+1}_+}z^{1-2s}|\nabla(U^*-U_B)|^2\,dx\,dz
=\int_0^\infty z^{1-2s}\int_0^\infty P_\rho|\nabla v(\rho,z)|^2\,d\rho\,dz \\
&=\int_0^\infty z^{1-2s}\int_0^{M(z)}\alpha|\nabla v(\rho(t,z),z)|^2\,dt\,dz \\
&=\int_0^\infty z^{1-2s}\int_0^{M(z)}\frac{D_B}{\alpha}\,dt\,dz.
\end{aligned} ∫ R + n + 1 z 1 − 2 s ∣∇ ( U ∗ − U B ) ∣ 2 d x d z = ∫ 0 ∞ z 1 − 2 s ∫ 0 ∞ P ρ ∣∇ v ( ρ , z ) ∣ 2 d ρ d z = ∫ 0 ∞ z 1 − 2 s ∫ 0 M ( z ) α ∣∇ v ( ρ ( t , z ) , z ) ∣ 2 d t d z = ∫ 0 ∞ z 1 − 2 s ∫ 0 M ( z ) α D B d t d z .
The last equality follows from α ∇ t = ( − P ∗ , γ ) \alpha\nabla t=(-P_*,\gamma) α ∇ t = ( − P ∗ , γ ) at the corresponding radius. This proves (22) . All changes of variables can first be made on compact level intervals and then extended by monotone convergence of the nonnegative energy densities.
Combining (20) , (21) and (22) with the splitting (15) of Lemma 2.2 proves (19) . Cauchy–Schwarz gives D H , D V ≥ 0 D_H,D_V\ge0 D H , D V ≥ 0 , classical isoperimetry gives D I ≥ 0 D_I\ge0 D I ≥ 0 , and D B D_B D B is a squared norm. Finally, ∣ d t ∣ = w d ρ |dt|=w\,d\rho ∣ d t ∣ = w d ρ , α w = P ρ \alpha w=P_\rho α w = P ρ , and D B = α 2 ∣ ∇ v ∣ 2 D_B=\alpha^2|\nabla v|^2 D B = α 2 ∣∇ v ∣ 2 give (23) and, by restriction to R \mathcal{R} R , (24) . □
A nonsharp quantitative Saint-Venant inequality
For the reader’s convenience, we use Lemma 3.1 to give a short derivation of the nonsharp quantitative Saint-Venant inequality of [3 , 19 ] . It uses only the isoperimetric defect and the profile defect, and provides the initial smallness needed in Proposition 4.2.
Proposition 3.2. For every n ≥ 2 n\ge2 n ≥ 2 and s ∈ ( 0 , 1 ) s\in(0,1) s ∈ ( 0 , 1 ) there is c = c ( n , s ) > 0 c=c(n,s)>0 c = c ( n , s ) > 0 such that every open set Ω ⊂ R n \Omega\subset\mathbb{R}^n Ω ⊂ R n with ∣ Ω ∣ = ω n |\Omega|=\omega_n ∣Ω∣ = ω n satisfies
T s ( B ) − T s ( Ω ) ≥ c A ( Ω ) 2 + 1 / s . T_s(B)-T_s(\Omega)\ge c\mathcal{A}(\Omega)^{2+1/s}. T s ( B ) − T s ( Ω ) ≥ c A ( Ω ) 2 + 1/ s .
Proof. By truncation we may assume that Ω \Omega Ω is bounded (but without a universal bound on the diameter). We write δ = δ T ( Ω ) \delta=\delta_T(\Omega) δ = δ T ( Ω ) and A = A ( Ω ) \mathcal{A}=\mathcal{A}(\Omega) A = A ( Ω ) . We may assume that A > 0 \mathcal{A}>0 A > 0 since otherwise there is nothing to prove. Fix ℓ \ell ℓ of Lemma A.1, and put
τ = ε A , I τ = ( 1 − 2 τ , 1 − τ ) , t τ = b τ s , z τ = τ 1 + 1 / ( 2 s ) . \tau=\varepsilon\mathcal{A},\qquad I_\tau=(1-2\tau,1-\tau),\qquad t_\tau=b\tau^s,\qquad z_\tau=\tau^{1+1/(2s)}. τ = ε A , I τ = ( 1 − 2 τ , 1 − τ ) , t τ = b τ s , z τ = τ 1 + 1/ ( 2 s ) .
Here b = b ( n , s ) > 0 b=b(n,s)>0 b = b ( n , s ) > 0 is chosen small enough below, and then ε = ε ( n , s ) > 0 \varepsilon=\varepsilon(n,s)>0 ε = ε ( n , s ) > 0 is chosen so that 2 τ < ℓ 2\tau<\ell 2 τ < ℓ and the lower bound (27) below holds. Since A ≤ 2 \mathcal{A}\le2 A ≤ 2 , we may also assume τ < 1 \tau<1 τ < 1 ; hence z τ ≤ τ z_\tau\le\tau z τ ≤ τ .
Let G = { u > 2 t τ } G=\{u>2t_{\tau}\} G = { u > 2 t τ } . If ∣ G ∣ < ω n ( 1 − τ ) n |G|<\omega_n(1-\tau)^n ∣ G ∣ < ω n ( 1 − τ ) n , then u ∗ ( ρ e 1 ) ≤ 2 t τ u^*(\rho e_1)\le2t_{\tau} u ∗ ( ρ e 1 ) ≤ 2 t τ for almost every ρ ≥ 1 − τ \rho\ge1-\tau ρ ≥ 1 − τ . The ball formula (A.2) gives u B ( ρ e 1 ) ≥ c τ s u_B(\rho e_1)\ge c\tau^s u B ( ρ e 1 ) ≥ c τ s on ( 1 − τ , 1 − τ / 2 ) (1-\tau,1-\tau/2) ( 1 − τ , 1 − τ /2 ) . Choose b b b so that 2 t τ 2t_{\tau} 2 t τ is at most half this lower bound. Lemma A.2 then gives
δ ≥ c ∫ 1 − τ 1 − τ / 2 ∣ u ∗ ( ρ e 1 ) − u B ( ρ e 1 ) ∣ 2 ( 1 − ρ ) 2 s d ρ ≥ c τ ≥ c A 2 + 1 / s . \delta\ge c\int_{1-\tau}^{1-\tau/2}\frac{|u^*(\rho e_1)-u_B(\rho e_1)|^2}{(1-\rho)^{2s}}\,d\rho\ge c\tau\ge cA^{2+1/s}. δ ≥ c ∫ 1 − τ 1 − τ /2 ( 1 − ρ ) 2 s ∣ u ∗ ( ρ e 1 ) − u B ( ρ e 1 ) ∣ 2 d ρ ≥ c τ ≥ c A 2 + 1/ s .
We may therefore assume that
∣ Ω ∖ G ∣ ≤ ω n ( 1 − ( 1 − τ ) n ) ≲ τ . (26) |\Omega\setminus G|\le\omega_n\bigl(1-(1-\tau)^n\bigr)\lesssim\tau. \tag*{(26)} ∣Ω ∖ G ∣ ≤ ω n ( 1 − ( 1 − τ ) n ) ≲ τ . (26)
The weighted Cauchy–Schwarz inequality along vertical lines and the strong L 2 L^2 L 2 trace give for fixed z > 0 z>0 z > 0 ,
∥ U ( ⋅ , z ) − u ∥ 2 2 ≤ C z 2 s ∫ 0 z ζ 1 − 2 s ∥ U z ( ⋅ , ζ ) ∥ 2 2 d ζ \|U(\cdot,z)-u\|_2^2\le Cz^{2s}\int_0^z\zeta^{1-2s}\|U_z(\cdot,\zeta)\|_2^2\,d\zeta ∥ U ( ⋅ , z ) − u ∥ 2 2 ≤ C z 2 s ∫ 0 z ζ 1 − 2 s ∥ U z ( ⋅ , ζ ) ∥ 2 2 d ζ
≲ z 2 s E s ( u ) ≤ z 2 s T s ( B ) . \lesssim z^{2s}\mathcal{E}_s(u)\le z^{2s}T_s(B). ≲ z 2 s E s ( u ) ≤ z 2 s T s ( B ) .
For ρ ∈ I τ \rho\in I_\tau ρ ∈ I τ and 0 < z < z τ 0<z<z_\tau 0 < z < z τ , let E = E ρ , z E=E_{\rho,z} E = E ρ , z and t = t ( ρ , z ) t=t(\rho,z) t = t ( ρ , z ) . If t ≤ t τ t\le t_\tau t ≤ t τ , then ∣ G ∖ E ∣ ≤ t τ − 2 ∥ U ( ⋅ , z ) − u ∥ 2 2 |G\setminus E|\le t_\tau^{-2}\|U(\cdot,z)-u\|_2^2 ∣ G ∖ E ∣ ≤ t τ − 2 ∥ U ( ⋅ , z ) − u ∥ 2 2 . Since ∣ E ∣ = ω n ρ n ≤ ∣ Ω ∣ |E|=\omega_n\rho^n\le|\Omega| ∣ E ∣ = ω n ρ n ≤ ∣Ω∣ , we have ∣ E ∖ Ω ∣ ≤ ∣ Ω ∖ E ∣ |E\setminus\Omega|\le|\Omega\setminus E| ∣ E ∖ Ω∣ ≤ ∣Ω ∖ E ∣ ; using (26) , we obtain
∣ Ω △ E ∣ ≲ τ + t τ − 2 ∥ U ( ⋅ , z ) − u ∥ 2 2 . |\Omega\triangle E|\lesssim\tau+t_\tau^{-2}\|U(\cdot,z)-u\|_2^2. ∣Ω△ E ∣ ≲ τ + t τ − 2 ∥ U ( ⋅ , z ) − u ∥ 2 2 .
If t > t τ t>t_\tau t > t τ , since u = 0 u=0 u = 0 and U ( ⋅ , z ) > t > t τ U(\cdot,z)>t>t_\tau U ( ⋅ , z ) > t > t τ on E ∖ Ω E\setminus\Omega E ∖ Ω , Chebyshev yields ∣ E ∖ Ω ∣ ≤ t τ − 2 ∥ U ( ⋅ , z ) − u ∥ 2 2 |E\setminus\Omega|\le t_\tau^{-2}\|U(\cdot,z)-u\|_2^2 ∣ E ∖ Ω∣ ≤ t τ − 2 ∥ U ( ⋅ , z ) − u ∥ 2 2 . Now ∣ Ω △ E ∣ = ∣ Ω ∣ − ∣ E ∣ + 2 ∣ E ∖ Ω ∣ |\Omega\triangle E|=|\Omega|-|E|+2|E\setminus\Omega| ∣Ω△ E ∣ = ∣Ω∣ − ∣ E ∣ + 2∣ E ∖ Ω∣ and ∣ Ω ∣ − ∣ E ∣ = ω n ( 1 − ρ n ) ≲ τ |\Omega|-|E|=\omega_n(1-\rho^n)\lesssim\tau ∣Ω∣ − ∣ E ∣ = ω n ( 1 − ρ n ) ≲ τ on I τ I_\tau I τ give the same bound. Thus, in both cases,
∣ Ω △ E ρ , z ∣ ≲ τ + z 2 s τ 2 ≲ τ . |\Omega\triangle E_{\rho,z}|\lesssim\tau+\frac{z^{2s}}{\tau^2}\lesssim\tau. ∣Ω△ E ρ , z ∣ ≲ τ + τ 2 z 2 s ≲ τ .
For every center a ∈ R n a\in\mathbb{R}^n a ∈ R n , the triangle inequality yields
∣ E ρ , z △ B ρ ( a ) ∣ ≥ ∣ Ω △ B 1 ( a ) ∣ − ∣ Ω △ E ρ , z ∣ − ω n ( 1 − ρ n ) . |E_{\rho,z}\triangle B_\rho(a)|\ge|\Omega\triangle B_1(a)|-|\Omega\triangle E_{\rho,z}|-\omega_n(1-\rho^n). ∣ E ρ , z △ B ρ ( a ) ∣ ≥ ∣Ω△ B 1 ( a ) ∣ − ∣Ω△ E ρ , z ∣ − ω n ( 1 − ρ n ) .
Taking the infimum over a a a and then choosing ε \varepsilon ε small enough gives
q ( ρ , z ) ≥ ω n A − C τ ≥ 1 2 ω n A on I τ × ( 0 , z τ ) . (27) q(\rho,z)\ge\omega_n A-C\tau\ge\tfrac12\omega_n A\qquad\text{on }I_\tau\times(0,z_\tau). \tag*{(27)} q ( ρ , z ) ≥ ω n A − C τ ≥ 2 1 ω n A on I τ × ( 0 , z τ ) . (27)
The quantitative isoperimetric inequality gives D I ≳ ρ 2 n − 2 A ( E ρ , z ) 2 D_I\gtrsim\rho^{2n-2}A(E_{\rho,z})^2 D I ≳ ρ 2 n − 2 A ( E ρ , z ) 2 . Since q = ω n ρ n A ( E ρ , z ) q=\omega_n\rho^nA(E_{\rho,z}) q = ω n ρ n A ( E ρ , z ) , it follows that q 2 ≲ ρ 2 D I q^2\lesssim\rho^2D_I q 2 ≲ ρ 2 D I . Hence (27) gives D I ≳ A 2 D_I\gtrsim A^2 D I ≳ A 2 on the rectangle above. Retaining D I D_I D I and the radial part of D B D_B D B in (24) , and using P ρ ≍ 1 P_\rho\asymp1 P ρ ≍ 1 there, we find
δ ≳ ∫ 0 z τ z 1 − 2 s ∫ I τ ( A 2 w 2 + ( w B − w ) 2 ) d ρ d z . \delta\gtrsim\int_0^{z_\tau}z^{1-2s}\int_{I_\tau}\bigl(A^2w^2+(w_B-w)^2\bigr)\,d\rho\,dz. δ ≳ ∫ 0 z τ z 1 − 2 s ∫ I τ ( A 2 w 2 + ( w B − w ) 2 ) d ρ d z .
For 0 < A ≤ 2 0<A\le2 0 < A ≤ 2 , A 2 w 2 + ( w B − w ) 2 ≳ A 2 w B 2 A^2w^2+(w_B-w)^2\gtrsim A^2w_B^2 A 2 w 2 + ( w B − w ) 2 ≳ A 2 w B 2 . Lemma A.1 gives w B ≳ τ s − 1 w_B\gtrsim\tau^{s-1} w B ≳ τ s − 1 on this rectangle. Consequently,
δ ≳ A 2 τ 2 s − 1 z τ 2 − 2 s = A 2 τ 1 / s ≳ A 2 + 1 / s . \delta\gtrsim A^2\tau^{2s-1}z_\tau^{2-2s}=A^2\tau^{1/s}\gtrsim A^{2+1/s}. δ ≳ A 2 τ 2 s − 1 z τ 2 − 2 s = A 2 τ 1/ s ≳ A 2 + 1/ s .
Horizontal estimates
Throughout this subsection Ω ⊂ R n \Omega\subset\mathbb{R}^n Ω ⊂ R n is bounded and open, with ∣ Ω ∣ = ω n |\Omega|=\omega_n ∣Ω∣ = ω n , u = u Ω u=u_\Omega u = u Ω , U = Ext s u U=\operatorname{Ext}_s u U = Ext s u , and E ρ , z E_{\rho,z} E ρ , z , q ( ρ , z ) q(\rho,z) q ( ρ , z ) are as in Section 2. In [17 ] the level sets of the torsion function are contained in Ω \Omega Ω , hence in B R B_R B R when Ω ⊂ B R \Omega\subset B_R Ω ⊂ B R ; this confinement enters the strong isoperimetric estimate [17 Proposition 3.5] . At positive height U ( ⋅ , z ) U(\cdot,z) U ( ⋅ , z ) is positive on all of R n \mathbb{R}^n R n , and the confinement of the level sets of volume at most ω n \omega_n ω n has to be proved.
Lemma 3.3. Let Ω ⊂ B R \Omega\subset B_R Ω ⊂ B R be open with ∣ Ω ∣ = ω n |\Omega|=\omega_n ∣Ω∣ = ω n , and let z 0 > 0 z_0>0 z 0 > 0 . Set L = R + 2 4 R 2 + z 0 2 L=R+2\sqrt{4R^2+z_0^2} L = R + 2 4 R 2 + z 0 2 . Then
E ρ , z ⊂ B L , 0 < ρ ≤ 1 , 0 < z ≤ z 0 . (28) E_{\rho,z}\subset B_L,\qquad0<\rho\le1,\quad0<z\le z_0. \tag*{(28)} E ρ , z ⊂ B L , 0 < ρ ≤ 1 , 0 < z ≤ z 0 . (28)
Proof. Since ∣ Ω ∣ = ω n |\Omega|=\omega_n ∣Ω∣ = ω n and Ω ⊂ B R \Omega\subset B_R Ω ⊂ B R , necessarily R ≥ 1 R\ge1 R ≥ 1 . The torsion function is nonnegative, supported in Ω \Omega Ω , and T s ( Ω ) = ∫ u > 0 T_s(\Omega)=\int u>0 T s ( Ω ) = ∫ u > 0 . By (4) , for x , y ∈ B R x,y\in B_R x , y ∈ B R and 0 < z ≤ z 0 0<z\le z_0 0 < z ≤ z 0 ,
P z ( s ) ( x − y ) ≥ Γ ( n / 2 + s ) π n / 2 Γ ( s ) z 2 s ( 4 R 2 + z 0 2 ) ( n + 2 s ) / 2 . P_z^{(s)}(x-y)\ge\frac{\Gamma(n/2+s)}{\pi^{n/2}\Gamma(s)}\frac{z^{2s}}{(4R^2+z_0^2)^{(n+2s)/2}}. P z ( s ) ( x − y ) ≥ π n /2 Γ ( s ) Γ ( n /2 + s ) ( 4 R 2 + z 0 2 ) ( n + 2 s ) /2 z 2 s .
Integrating against u ( y ) u(y) u ( y ) therefore gives, throughout B R B_R B R ,
U ( x , z ) ≥ Γ ( n / 2 + s ) π n / 2 Γ ( s ) z 2 s T s ( Ω ) ( 4 R 2 + z 0 2 ) ( n + 2 s ) / 2 = t z . U(x,z) \ge\frac{\Gamma(n/2+s)}{\pi^{n/2}\Gamma(s)}\frac{z^{2s}T_s(\Omega)}{(4R^2+z_0^2)^{(n+2s)/2}}=t_z. U ( x , z ) ≥ π n /2 Γ ( s ) Γ ( n /2 + s ) ( 4 R 2 + z 0 2 ) ( n + 2 s ) /2 z 2 s T s ( Ω ) = t z .
Since B R ⊂ E t z , z B_R\subset E_{t_z,z} B R ⊂ E t z , z , we have
∣ E t z , z ∣ ≥ ∣ B R ∣ ≥ ω n = ∣ E t ( 1 , z ) , z ∣ |E_{t_z,z}|\ge|B_R|\ge\omega_n=|E_{t(1,z),z}| ∣ E t z , z ∣ ≥ ∣ B R ∣ ≥ ω n = ∣ E t ( 1 , z ) , z ∣
and thus
t ( 1 , z ) ≥ t z . (29) t(1,z)\ge t_z. \tag*{(29)} t ( 1 , z ) ≥ t z . (29)
For ∣ x ∣ ≥ L > R |x|\ge L>R ∣ x ∣ ≥ L > R and y ∈ Ω y\in\Omega y ∈ Ω , we have ∣ x − y ∣ ≥ L − R |x-y|\ge L-R ∣ x − y ∣ ≥ L − R . Dropping the nonnegative z 2 z^2 z 2 from the kernel denominator yields
U ( x , z ) ≤ Γ ( n / 2 + s ) π n / 2 Γ ( s ) z 2 s T s ( Ω ) ( L − R ) n + 2 s . (30) U(x,z)\le\frac{\Gamma(n/2+s)}{\pi^{n/2}\Gamma(s)}\frac{z^{2s}T_s(\Omega)}{(L-R)^{n+2s}}. \tag*{(30)} U ( x , z ) ≤ π n /2 Γ ( s ) Γ ( n /2 + s ) ( L − R ) n + 2 s z 2 s T s ( Ω ) . (30)
For L = R + 2 4 R 2 + z 0 2 L=R+2\sqrt{4R^2+z_0^2} L = R + 2 4 R 2 + z 0 2 , the right-hand side of (30) is then 2 − ( n + 2 s ) 2^{-(n+2s)} 2 − ( n + 2 s ) times the right-hand side of (29) . Since 2 − ( n + 2 s ) < 1 2^{-(n+2s)}<1 2 − ( n + 2 s ) < 1 , it follows that U ( x , z ) < t ( 1 , z ) U(x,z)<t(1,z) U ( x , z ) < t ( 1 , z ) whenever ∣ x ∣ ≥ L |x|\ge L ∣ x ∣ ≥ L . This confines E 1 , z E_{1,z} E 1 , z to B L B_L B L for every 0 < z ≤ z 0 0<z\le z_0 0 < z ≤ z 0 . Finally, t ( ρ , z ) ≥ t ( 1 , z ) t(\rho,z)\ge t(1,z) t ( ρ , z ) ≥ t ( 1 , z ) for ρ ≤ 1 \rho\le1 ρ ≤ 1 , so E ρ , z ⊂ E 1 , z E_{\rho,z}\subset E_{1,z} E ρ , z ⊂ E 1 , z and (28) follows. □
We now record the horizontal estimates. At each fixed height, they follow from items (1)–(3) of [17 Proposition 3.6] together with the estimate on good radii in the proof of [17 Theorem 3.7] . It builds on [17 Lemmas 3.1 and 3.2] . The only additional input in our setting is the confinement from Lemma 3.3. Notice that smoothness of ∂ Ω \partial\Omega ∂ Ω is assumed in [17 Lemma 3.2] only to obtain a uniform first-order expansion of the torsion function near the level sets. Here U ( ⋅ , z ) U(\cdot,z) U ( ⋅ , z ) is smooth on R n \mathbb{R}^n R n , so the argument carries over. We refer the reader to [17 ] for the proofs.
Proposition 3.4. Let Ω ⊂ B R \Omega\subset B_R Ω ⊂ B R with R ≥ 1 R\ge1 R ≥ 1 . All defects below are evaluated at ( t ( ρ , z ) , z ) (t(\rho,z),z) ( t ( ρ , z ) , z ) . For every z > 0 z>0 z > 0 the infimum defining q ( ρ , z ) q(\rho,z) q ( ρ , z ) is attained for every ρ > 0 \rho>0 ρ > 0 , and for 0 < ρ , ρ ′ ≤ 1 0<\rho,\rho'\le1 0 < ρ , ρ ′ ≤ 1 ,
∣ q ( ρ , z ) − q ( ρ ′ , z ) ∣ ≤ 2 ω n ∣ ρ n − ( ρ ′ ) n ∣ ≤ 2 n ω n ∣ ρ − ρ ′ ∣ , 0 ≤ q ( ρ , z ) ≤ 2 ω n . (31) |q(\rho,z)-q(\rho',z)|\le2\omega_n|\rho^n-(\rho')^n|\le2n\omega_n|\rho-\rho'|,\qquad0\le q(\rho,z)\le2\omega_n. \tag*{(31)} ∣ q ( ρ , z ) − q ( ρ ′ , z ) ∣ ≤ 2 ω n ∣ ρ n − ( ρ ′ ) n ∣ ≤ 2 n ω n ∣ ρ − ρ ′ ∣ , 0 ≤ q ( ρ , z ) ≤ 2 ω n . (31) In particular, q ( ⋅ , z ) q(\cdot,z) q ( ⋅ , z ) is locally Lipschitz on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) .
Fix Z > 0 Z>0 Z > 0 , and ρ ∗ ∈ ( 0 , 1 ) \rho_*\in(0,1) ρ ∗ ∈ ( 0 , 1 ) . There is C = C ( R , Z , ρ ∗ ) > 0 C=C(R,Z,\rho_*)>0 C = C ( R , Z , ρ ∗ ) > 0 such that
q ( ρ , z ) 2 + ( ∂ ρ q ( ρ , z ) ) 2 ≤ C ( D H + D I ) (32) q(\rho,z)^2+(\partial_\rho q(\rho,z))^2\le C(D_H+D_I) \tag*{(32)} q ( ρ , z ) 2 + ( ∂ ρ q ( ρ , z ) ) 2 ≤ C ( D H + D I ) (32) for almost every ( ρ , z ) ∈ [ ρ ∗ , 1 ] × ( 0 , Z ) (\rho,z)\in[\rho_*,1]\times(0,Z) ( ρ , z ) ∈ [ ρ ∗ , 1 ] × ( 0 , Z ) . Consequently,
∫ 0 Z z 1 − 2 s ∫ ρ ∗ 1 w 2 ( q 2 + ( ∂ ρ q ) 2 ) d ρ d z ≤ C δ T ( Ω ) . \int_0^Z z^{1-2s}\int_{\rho_*}^1 w^2\left(q^2+(\partial_\rho q)^2\right)\,d\rho\,dz\le C\delta_T(\Omega). ∫ 0 Z z 1 − 2 s ∫ ρ ∗ 1 w 2 ( q 2 + ( ∂ ρ q ) 2 ) d ρ d z ≤ C δ T ( Ω ) .
Scalar differentiation in level coordinates
We now turn to the treatment of the vertical derivatives, which is new compared with [17 ] . For this, we first revert to the original ( t , z ) (t,z) ( t , z ) variables.
Lemma 3.5. Let Ω ⊂ R n \Omega\subset\mathbb{R}^n Ω ⊂ R n be bounded and open, with ∣ Ω ∣ = ω n |\Omega|=\omega_n ∣Ω∣ = ω n , let u = u Ω u=u_\Omega u = u Ω and U = Ext s u U=\operatorname{Ext}_s u U = Ext s u , and let A > 0 A>0 A > 0 . The functions m m m , Q Q Q , Q A Q_A Q A of (10) are continuous on D D D , and they belong to W l o c 1 ( D ) W_{\mathrm{loc}}^1(D) W loc 1 ( D ) .
Writing
ϑ ( t , z ) = ∫ ∂ E t , z ∣ h ∣ g d H n − 1 \vartheta(t,z)=\int_{\partial E_{t,z}}\frac{|h|}{g}\,d\mathcal{H}^{n-1} ϑ ( t , z ) = ∫ ∂ E t , z g ∣ h ∣ d H n − 1 we have
∣ ( Q A ) t ∣ ≤ 2 α , ∣ ( Q A ) z ∣ ≤ 2 ϑ . (33) |(Q_A)_t|\le2\alpha,\qquad|(Q_A)_z|\le2\vartheta. \tag*{(33)} ∣ ( Q A ) t ∣ ≤ 2 α , ∣ ( Q A ) z ∣ ≤ 2 ϑ . (33) Furthermore, recalling the definition (11) of Z \mathcal{Z} Z ,
∣ Z Q A ∣ 2 ≤ D V a.e. in D . (34) |\mathcal{Z}Q_A|^2\le D_V\qquad\text{a.e. in }D. \tag*{(34)} ∣ Z Q A ∣ 2 ≤ D V a.e. in D . (34)
Proof. Step 1: the asymmetry is 2-Lipschitz for the symmetric difference. If E , F E,F E , F have finite measure and B E , B F B_E,B_F B E , B F denote balls with ∣ B E ∣ = ∣ E ∣ |B_E|=|E| ∣ B E ∣ = ∣ E ∣ , ∣ B F ∣ = ∣ F ∣ |B_F|=|F| ∣ B F ∣ = ∣ F ∣ , comparison with a common center gives
∣ inf a ∣ E △ B E ( a ) ∣ − inf a ∣ F △ B F ( a ) ∣ ∣ ≤ ∣ E △ F ∣ + ∣ ∣ E ∣ − ∣ F ∣ ∣ ≤ 2 ∣ E △ F ∣ . (35) \left|\inf_a|E\mathbin{\triangle}B_E(a)|-\inf_a|F\mathbin{\triangle}B_F(a)|\right|\le|E\mathbin{\triangle}F|+\bigl||E|-|F|\bigr|\le2|E\mathbin{\triangle}F|. \tag*{(35)} a inf ∣ E △ B E ( a ) ∣ − a inf ∣ F △ B F ( a ) ∣ ≤ ∣ E △ F ∣ + ∣ E ∣ − ∣ F ∣ ≤ 2∣ E △ F ∣. (35)
Taking a cap preserves this Lipschitz bound, since r ↦ min { r , A } r \mapsto\min\{r,A\} r ↦ min { r , A } is 1-Lipschitz and the cap is independent of the pair. Everything in the first two assertions follows by combining (35) with estimates for ∣ E t 2 , z 2 △ E t 1 , z 1 ∣ |E_{t_2,z_2}\mathbin{\triangle}E_{t_1,z_1}| ∣ E t 2 , z 2 △ E t 1 , z 1 ∣ .
Step 2: continuity. On the rectangles under consideration all the sets E t , z E_{t,z} E t , z lie in one fixed bounded spatial set, U U U is continuous, and its positive level sets have zero Lebesgue measure. Dominated convergence therefore makes ( t , z ) ↦ 1 E t , z (t,z)\mapsto\mathbf{1}_{E_{t,z}} ( t , z ) ↦ 1 E t , z continuous from D D D to L 1 ( R n ) L^1(\mathbb{R}^n) L 1 ( R n ) . Hence m m m is continuous, and by (35) so are Q Q Q and Q A Q_A Q A .
Step 3: horizontal absolute continuity. At fixed z z z the sets E t , z E_{t,z} E t , z are nested, so for t 1 < t 2 t_1<t_2 t 1 < t 2
∣ E t 1 , z △ E t 2 , z ∣ = m ( t 1 , z ) − m ( t 2 , z ) = ∫ t 1 t 2 α ( t , z ) d t |E_{t_1,z}\mathbin{\triangle}E_{t_2,z}|=m(t_1,z)-m(t_2,z)=\int_{t_1}^{t_2}\alpha(t,z)\,dt ∣ E t 1 , z △ E t 2 , z ∣ = m ( t 1 , z ) − m ( t 2 , z ) = ∫ t 1 t 2 α ( t , z ) d t
by (25) . With (35) this gives absolute continuity in t t t and ∣ ( Q A ) t ∣ ≤ 2 α |(Q_A)_t|\le2\alpha ∣ ( Q A ) t ∣ ≤ 2 α .
Step 4: vertical slicing bound. We claim that for almost every t t t and every z 1 < z 2 z_1<z_2 z 1 < z 2 in the rectangle,
∣ E t , z 2 △ E t , z 1 ∣ ≤ ∫ z 1 z 2 ϑ ( t , z ) d z . (36) |E_{t,z_2}\mathbin{\triangle}E_{t,z_1}|\le\int_{z_1}^{z_2}\vartheta(t,z)\,dz. \tag*{(36)} ∣ E t , z 2 △ E t , z 1 ∣ ≤ ∫ z 1 z 2 ϑ ( t , z ) d z . (36)
For almost every t t t , let μ t \mu_t μ t be the positive measure
μ t ( I ) = ∫ R n ∣ D z 1 { U ( x , ⋅ ) > t } ∣ ( I ) d x . \mu_t(I)=\int_{\mathbb{R}^n}|D_z\mathbf{1}_{\{U(x,\cdot)>t\}}|(I)\,dx. μ t ( I ) = ∫ R n ∣ D z 1 { U ( x , ⋅ ) > t } ∣ ( I ) d x .
It is locally finite on the parameter rectangles under consideration. For a nonnegative continuous compactly supported function f ( t , z ) f(t,z) f ( t , z ) , one-dimensional coarea followed by horizontal coarea gives
∫ d t ∫ f ( t , z ) d μ t ( z ) = ∫ R n ∫ ∣ U z ( x , z ) ∣ f ( U ( x , z ) , z ) d x d z = ∫ ∫ f ( t , z ) ϑ ( t , z ) d t d z . \begin{aligned}
\int dt\int f(t,z)\,d\mu_t(z)=\int_{\mathbb{R}^n}\int|U_z(x,z)|f(U(x,z),z)\,dx\,dz \\
&=\int\int f(t,z)\vartheta(t,z)\,dt\,dz.
\end{aligned} ∫ d t ∫ f ( t , z ) d μ t ( z ) = ∫ R n ∫ ∣ U z ( x , z ) ∣ f ( U ( x , z ) , z ) d x d z = ∫∫ f ( t , z ) ϑ ( t , z ) d t d z .
Disintegration of these positive measures gives d μ t ( z ) = ϑ ( t , z ) d z d\mu_t(z)=\vartheta(t,z)\,dz d μ t ( z ) = ϑ ( t , z ) d z for almost every t t t , which gives (36) . Initially this holds at almost every pair of endpoints. It extends to all endpoints on the interval by the L x 1 L^1_x L x 1 continuity of Step 2. The Lipschitz comparison (35) now gives absolute continuity of Q A Q_A Q A in z z z and its second bound in (33) .
Step 5: the fixed-volume direction. We finally prove (34) . Since g g g is bounded below near every regular pair, the bounds (33) show that Q A Q_A Q A is Lipschitz in a neighborhood of every regular pair; in particular, its weak and pointwise derivatives agree almost everywhere. Fix a regular pair ( t , z ) (t,z) ( t , z ) where Q A Q_A Q A is differentiable, set c ∗ = γ ( t , z ) / α ( t , z ) c_*=\gamma(t,z)/\alpha(t,z) c ∗ = γ ( t , z ) / α ( t , z ) and ( t ε , z ε ) = ( t + ε c ∗ , z + ε ) (t_\varepsilon,z_\varepsilon)=(t+\varepsilon c_*,z+\varepsilon) ( t ε , z ε ) = ( t + ε c ∗ , z + ε ) , so that
Z Q A ( t , z ) = lim ε → 0 Q A ( t ε , z ε ) − Q A ( t , z ) ε . \mathcal{Z}Q_A(t,z)=\lim_{\varepsilon\to0}\frac{Q_A(t_\varepsilon,z_\varepsilon)-Q_A(t,z)}{\varepsilon}. Z Q A ( t , z ) = ε → 0 lim ε Q A ( t ε , z ε ) − Q A ( t , z ) .
By (35) ,
∣ Q A ( t ε , z ε ) − Q A ( t , z ) ∣ ≤ ∣ E t ε , z ε △ E t , z ∣ + ∣ m ( t ε , z ε ) − m ( t , z ) ∣ . |Q_A(t_\varepsilon,z_\varepsilon)-Q_A(t,z)|\le|E_{t_\varepsilon,z_\varepsilon}\mathbin{\triangle}E_{t,z}|+|m(t_\varepsilon,z_\varepsilon)-m(t,z)|. ∣ Q A ( t ε , z ε ) − Q A ( t , z ) ∣ ≤ ∣ E t ε , z ε △ E t , z ∣ + ∣ m ( t ε , z ε ) − m ( t , z ) ∣.
Since Z m = 0 \mathcal{Z}m=0 Z m = 0 by (11) , m ( t ε , z ε ) = m ( t , z ) + o ( ∣ ε ∣ ) m(t_\varepsilon,z_\varepsilon)=m(t,z)+o(|\varepsilon|) m ( t ε , z ε ) = m ( t , z ) + o ( ∣ ε ∣ ) , so the second term does not contribute in the limit.
We compute the first term by a normal graph. The surface Σ = ∂ E t , z \Sigma=\partial E_{t,z} Σ = ∂ E t , z is compact and of class C 1 C^1 C 1 , with g > 0 g>0 g > 0 on it, and its outer normal is ν = − ∇ x U / g \nu=-\nabla_xU/g ν = − ∇ x U / g . Parametrize a tubular neighborhood by x = y + σ ν ( y ) x=y+\sigma\nu(y) x = y + σ ν ( y ) , y ∈ Σ y\in\Sigma y ∈ Σ . Since
∂ σ U ( y + σ ν ( y ) , z ) ∣ σ = 0 = ∇ x U ⋅ ν = − g < 0 , \left.\partial_\sigma U(y+\sigma\nu(y),z)\right|_{\sigma=0}=\nabla_xU\cdot\nu=-g<0, ∂ σ U ( y + σ ν ( y ) , z ) ∣ σ = 0 = ∇ x U ⋅ ν = − g < 0 ,
the implicit function theorem provides, for each y ∈ Σ y\in\Sigma y ∈ Σ and all small ε \varepsilon ε , a unique small r ε ( y ) r_\varepsilon(y) r ε ( y ) with U ( y + r ε ( y ) ν ( y ) , z ε ) = t ε U(y+r_\varepsilon(y)\nu(y),z_\varepsilon)=t_\varepsilon U ( y + r ε ( y ) ν ( y ) , z ε ) = t ε . Differentiating this identity at ε = 0 \varepsilon=0 ε = 0 gives − g ∂ ε r ε ∣ ε = 0 + h = c ∗ -g\,\partial_\varepsilon r_\varepsilon|_{\varepsilon=0}+h=c_* − g ∂ ε r ε ∣ ε = 0 + h = c ∗ , that is,
r ε ( y ) = ε h ( y ) − c ∗ g ( y ) + o ( ∣ ε ∣ ) uniformly on Σ . r_\varepsilon(y)=\varepsilon\frac{h(y)-c_*}{g(y)}+o(|\varepsilon|)\qquad\text{uniformly on }\Sigma. r ε ( y ) = ε g ( y ) h ( y ) − c ∗ + o ( ∣ ε ∣ ) uniformly on Σ.
Thus E t ε , z ε E_{t_\varepsilon,z_\varepsilon} E t ε , z ε is, near Σ \Sigma Σ , the normal graph of r ε r_\varepsilon r ε over E t , z E_{t,z} E t , z , and the area formula in normal coordinates gives
lim ε → 0 ∣ E t ε , z ε △ E t , z ∣ ∣ ε ∣ = ∫ ∂ E t , z ∣ h − c ∗ ∣ g d H n − 1 . \lim_{\varepsilon\to0}\frac{\lvert E_{t_\varepsilon,z_\varepsilon}\mathbin{\triangle} E_{t,z}\rvert}{\lvert\varepsilon\rvert}
=\int_{\partial E_{t,z}}\frac{\lvert h-c_*\rvert}{g}\,\mathrm{d}\mathcal{H}^{n-1}. ε → 0 lim ∣ ε ∣ ∣ E t ε , z ε △ E t , z ∣ = ∫ ∂ E t , z g ∣ h − c ∗ ∣ d H n − 1 .
Recalling that c ∗ = γ / α c_*=\gamma/\alpha c ∗ = γ / α , we obtain
∣ Z Q A ( t , z ) ∣ ≤ ∫ ∂ E t , z ∣ h − γ / α ∣ g d H n − 1 . \lvert ZQ_A(t,z)\rvert\leq\int_{\partial E_{t,z}}\frac{\lvert h-\gamma/\alpha\rvert}{g}\,\mathrm{d}\mathcal{H}^{n-1}. ∣ Z Q A ( t , z )∣ ≤ ∫ ∂ E t , z g ∣ h − γ / α ∣ d H n − 1 .
Cauchy–Schwarz concludes the proof of (34) . □
Weighted vertical flux and weak change of variables
Proposition 3.6. Let ℓ \ell ℓ be as in Lemma A.1, Ω ⊂ R n \Omega\subset\mathbb{R}^n Ω ⊂ R n be bounded and open with ∣ Ω ∣ = ω n \lvert\Omega\rvert=\omega_n ∣ Ω ∣ = ω n , and let R R R be the rectangle (8) . For each A ∈ ( 0 , 2 ω n ] A\in(0,2\omega_n] A ∈ ( 0 , 2 ω n ] , define almost everywhere on R R R
J A ( ρ , z ) = w ( ρ , z ) ( Z Q A ) ( t ( ρ , z ) , z ) . J_A(\rho,z)=w(\rho,z)(ZQ_A)(t(\rho,z),z). J A ( ρ , z ) = w ( ρ , z ) ( Z Q A ) ( t ( ρ , z ) , z ) . Then
∣ J A ∣ 2 ≤ w 2 D V , ∫ R z 1 − 2 s ∣ J A ∣ 2 ρ ρ d ρ d z ≤ C δ T ( Ω ) . (37) \lvert J_A\rvert^2\leq w^2D_V,\qquad\int_R z^{1-2s}\frac{\lvert J_A\rvert^2}{\rho_\rho}\,\mathrm{d}\rho\,\mathrm{d}z\leq C\delta T(\Omega). \tag*{(37)} ∣ J A ∣ 2 ≤ w 2 D V , ∫ R z 1 − 2 s ρ ρ ∣ J A ∣ 2 d ρ d z ≤ C δ T ( Ω ) . (37) For every φ ∈ C c ∞ ( R ) \varphi\in C^\infty_c(R) φ ∈ C c ∞ ( R ) one has
∫ R J A φ = − ∫ R w q A φ z − ∫ R t z ∂ ρ ( q A φ ) . (38) \int_R J_A\varphi=-\int_R wq_A\varphi_z-\int_R t_z\partial_\rho(q_A\varphi). \tag*{(38)} ∫ R J A φ = − ∫ R w q A φ z − ∫ R t z ∂ ρ ( q A φ ) . (38) In particular, the following identity holds in distributions:
∂ z ( w q A w B ) = J A w B − ∂ ρ ( v z q A w B ) + v z ∂ ρ ( q A w B ) + v ρ q A ∂ z w B w B 2 . (39) \partial_z\left(\frac{wq_A}{w_B}\right)=\frac{J_A}{w_B}-\partial_\rho\left(\frac{v_zq_A}{w_B}\right)+v_z\partial_\rho\left(\frac{q_A}{w_B}\right)+\frac{v_\rho q_A\partial_z w_B}{w_B^2}. \tag*{(39)} ∂ z ( w B w q A ) = w B J A − ∂ ρ ( w B v z q A ) + v z ∂ ρ ( w B q A ) + w B 2 v ρ q A ∂ z w B . (39)
Proof. The pointwise estimate in (37) follows from (34) . Using the definition of D V D_V D V in (23) and integrating gives the second bound.
A null set in ( t , z ) (t,z) ( t , z ) pulls back to a null set in ( ρ , z ) (\rho,z) ( ρ , z ) : for almost every fixed height, t ↦ ρ ( t , z ) t\mapsto\rho(t,z) t ↦ ρ ( t , z ) is locally absolutely continuous, so it maps the corresponding null section to a null set, and Fubini applies. Thus the definition of J A J_A J A is independent of the almost-everywhere representatives of the weak derivatives of Q A Q_A Q A . The weighted energy estimate also shows that J A ∈ L l o c 2 ( R ) J_A\in L^2_{\mathrm{loc}}(R) J A ∈ L loc 2 ( R ) , because the height weight is bounded above and below on compact subsets of R R R .
We now prove (38) . The chain rule for Sobolev functions and the fact that m m m stays positive on compact subsets of D D D give
ρ ∈ W l o c 1 , 1 ( D ) , ρ t = − α P ρ , ρ z = γ P ρ . \rho\in W^{1,1}_{\mathrm{loc}}(D),\qquad\rho_t=-\frac{\alpha}{P_\rho},\qquad\rho_z=\frac{\gamma}{P_\rho}. ρ ∈ W loc 1 , 1 ( D ) , ρ t = − P ρ α , ρ z = P ρ γ .
For a fixed test function φ \varphi φ , put ψ ( t , z ) = φ ( ρ ( t , z ) , z ) \psi(t,z)=\varphi(\rho(t,z),z) ψ ( t , z ) = φ ( ρ ( t , z ) , z ) . It is bounded and belongs to W l o c 1 , 1 ( D ) W^{1,1}_{\mathrm{loc}}(D) W loc 1 , 1 ( D ) . Its weak derivatives are
ψ t = − α P ρ φ ρ , ψ z = φ z + γ P ρ φ ρ , \psi_t=-\frac{\alpha}{P_\rho}\varphi_\rho,\qquad\psi_z=\varphi_z+\frac{\gamma}{P_\rho}\varphi_\rho, ψ t = − P ρ α φ ρ , ψ z = φ z + P ρ γ φ ρ ,
where the derivatives of φ \varphi φ are evaluated at ( ρ ( t , z ) , z ) (\rho(t,z),z) ( ρ ( t , z ) , z ) . Since Q A Q_A Q A and ψ \psi ψ are bounded Sobolev functions on a common compact subset of D D D , their product belongs to W 1 , 1 W^{1,1} W 1 , 1 there, and the weak Leibniz rule in the z z z variable gives
∫ D ( Z Q A ) ψ = ∫ D γ α ( Q A ) t ψ − ∫ D Q A ( φ z + γ P ρ φ ρ ) , \int_D (ZQ_A)\psi=\int_D\frac{\gamma}{\alpha}(Q_A)_t\psi-\int_D Q_A\left(\varphi_z+\frac{\gamma}{P_\rho}\varphi_\rho\right), ∫ D ( Z Q A ) ψ = ∫ D α γ ( Q A ) t ψ − ∫ D Q A ( φ z + P ρ γ φ ρ ) ,
where the functions involving φ \varphi φ are evaluated at ( ρ ( t , z ) , z ) (\rho(t,z),z) ( ρ ( t , z ) , z ) . Every term is integrable on this compact support. In particular, ∣ γ ( Q A ) t / α ∣ ≤ 2 ∣ γ ∣ \lvert\gamma(Q_A)_t/\alpha\rvert\leq2\lvert\gamma\rvert ∣ γ ( Q A ) t / α ∣ ≤ 2 ∣ γ ∣ by (33) .
For each fixed z z z , make the change of variables ρ = ρ ( t , z ) \rho=\rho(t,z) ρ = ρ ( t , z ) . By (7) , ∣ d t ∣ = w d ρ \lvert\mathrm{d}t\rvert=w\,\mathrm{d}\rho ∣ d t ∣ = w d ρ , and q A ( ρ , z ) = Q A ( t ( ρ , z ) , z ) q_A(\rho,z)=Q_A(t(\rho,z),z) q A ( ρ , z ) = Q A ( t ( ρ , z ) , z ) . Hence
( q A ) ρ = − w ( Q A ) t ( t ( ρ , z ) , z ) a.e. (q_A)_\rho=-w(Q_A)_t(t(\rho,z),z)\qquad\text{a.e.} ( q A ) ρ = − w ( Q A ) t ( t ( ρ , z ) , z ) a.e.
The preceding identity consequently becomes
∫ R J A φ = − ∫ R γ α ( q A ) ρ φ − ∫ R w q A ( φ z + γ P ρ φ ρ ) . \int_{\mathcal{R}} J_A\varphi=-\int_{\mathcal{R}}\frac{\gamma}{\alpha}(q_A)_\rho\varphi-\int_{\mathcal{R}}wq_A\left(\varphi_z+\frac{\gamma}{P_\rho}\varphi_\rho\right). ∫ R J A φ = − ∫ R α γ ( q A ) ρ φ − ∫ R w q A ( φ z + P ρ γ φ ρ ) .
Using γ w / P ρ = γ / α = t z \gamma w/P_\rho=\gamma/\alpha=t_z γ w / P ρ = γ / α = t z gives precisely (38) .
In distributions, (38) reads
∂ z ( w q A ) = J A − ∂ ρ ( t z q A ) + t z ( q A ) ρ . \partial_z(wq_A)=J_A-\partial_\rho(t_zq_A)+t_z(q_A)_\rho. ∂ z ( w q A ) = J A − ∂ ρ ( t z q A ) + t z ( q A ) ρ .
Insert t z = ( t B ) z + v z t_z=(t_B)_z+v_z t z = ( t B ) z + v z and ∂ ρ ( t B ) z = − ∂ z w B \partial_\rho(t_B)_z=-\partial_z w_B ∂ ρ ( t B ) z = − ∂ z w B , and multiply by the smooth positive function 1 / w B 1/w_B 1/ w B on compact subsets of R \mathcal{R} R . The product rule and w B − w = v ρ w_B-w=v_\rho w B − w = v ρ give (39) . □
Corollary 3.8 (Horizontal and vertical flux estimate). Let R ≥ 1 R\ge1 R ≥ 1 , let Ω ⊂ B R \Omega\subset B_R Ω ⊂ B R be open with ∣ Ω ∣ = ω n |\Omega|=\omega_n ∣Ω∣ = ω n , and let ℓ \ell ℓ be as in Lemma A.1. Then, for every A ∈ ( 0 , 2 ω n ] A\in(0,2\omega_n] A ∈ ( 0 , 2 ω n ] , almost everywhere on R \mathcal{R} R ,
q 2 + ( q ρ ) + 2 + ∣ ( Z Q A ) ( t ( ρ , z ) , z ) ∣ 2 ≤ C ( R , ℓ ) ( D H + D I ) + D V . (41) q^2+(q_\rho)^2_++|(ZQ_A)(t(\rho,z),z)|^2\le C(R,\ell)(D_H+D_I)+D_V. \tag*{(41)} q 2 + ( q ρ ) + 2 + ∣ ( Z Q A ) ( t ( ρ , z ) , z ) ∣ 2 ≤ C ( R , ℓ ) ( D H + D I ) + D V . (41) Consequently,
∫ R z 1 − 2 s [ w 2 ( q 2 + ( q ρ ) + 2 ) + ∣ J A ∣ 2 + ∣ ∇ v ∣ 2 ] d ρ d z + ∫ 1 − ℓ 1 ∣ v ( ρ , 0 ) ∣ 2 ( 1 − ρ ) 2 s d ρ ≤ C ( R ) δ T ( Ω ) . (42) \int_{\mathcal{R}}z^{1-2s}\left[w^2\left(q^2+(q_\rho)^2_+\right)+|J_A|^2+|\nabla v|^2\right]\,d\rho\,dz+\int_{1-\ell}^1\frac{|v(\rho,0)|^2}{(1-\rho)^{2s}}\,d\rho\le C(R)\delta_T(\Omega). \tag*{(42)} ∫ R z 1 − 2 s [ w 2 ( q 2 + ( q ρ ) + 2 ) + ∣ J A ∣ 2 + ∣∇ v ∣ 2 ] d ρ d z + ∫ 1 − ℓ 1 ( 1 − ρ ) 2 s ∣ v ( ρ , 0 ) ∣ 2 d ρ ≤ C ( R ) δ T ( Ω ) . (42) Proof. Estimate (41) follows from (32) in Proposition 3.4, with ρ ∗ = 1 − ℓ \rho_* =1-\ell ρ ∗ = 1 − ℓ and Z = ℓ Z=\ell Z = ℓ , and from (34) .
Multiply (41) by z 1 − 2 s w 2 / P ρ z^{1-2s}w^2/P_\rho z 1 − 2 s w 2 / P ρ and integrate over R \mathcal{R} R . Since J A = w ( Z Q A ) ( t ( ρ , z ) , z ) J_A=w(ZQ_A)(t(\rho,z),z) J A = w ( Z Q A ) ( t ( ρ , z ) , z ) , the radial four-defect identity (23) gives
∫ R z 1 − 2 s 1 P ρ [ w 2 ( q 2 + ( q ρ ) + 2 ) + ∣ J A ∣ 2 ] d ρ d z ≤ C ( R , ℓ ) δ T ( Ω ) . \int_{\mathcal{R}}z^{1-2s}\frac{1}{P_\rho}\left[w^2\left(q^2+(q_\rho)^2_+\right)+|J_A|^2\right]\,d\rho\,dz\le C(R,\ell)\delta_T(\Omega). ∫ R z 1 − 2 s P ρ 1 [ w 2 ( q 2 + ( q ρ ) + 2 ) + ∣ J A ∣ 2 ] d ρ d z ≤ C ( R , ℓ ) δ T ( Ω ) . On 1 − ℓ < ρ < 1 1-\ell<\rho<1 1 − ℓ < ρ < 1 , the radial factor P ρ = n ω n ρ n − 1 P_\rho=n\omega_n\rho^{n-1} P ρ = n ω n ρ n − 1 is bounded above and below by positive constants depending only on n n n and ℓ \ell ℓ . Thus the preceding estimate controls the first two bulk terms of (42) without the factor P ρ − 1 P_\rho^{-1} P ρ − 1 . Finally, Lemma A.2 gives
∫ R z 1 − 2 s ∣ ∇ v ∣ 2 d ρ d z + ∫ 1 − ℓ 1 ∣ v ( ρ , 0 ) ∣ 2 ( 1 − ρ ) 2 s d ρ ≤ C ( ℓ ) δ T ( Ω ) . \int_{\mathcal{R}}z^{1-2s}|\nabla v|^2\,d\rho\,dz+\int_{1-\ell}^1\frac{|v(\rho,0)|^2}{(1-\rho)^{2s}}\,d\rho\le C(\ell)\delta_T(\Omega). ∫ R z 1 − 2 s ∣∇ v ∣ 2 d ρ d z + ∫ 1 − ℓ 1 ( 1 − ρ ) 2 s ∣ v ( ρ , 0 ) ∣ 2 d ρ ≤ C ( ℓ ) δ T ( Ω ) . This proves (42) . Since the choice of ℓ \ell ℓ in Lemma A.1 depends only on n n n and s s s , all ℓ \ell ℓ -dependence is absorbed into C ( R ) C(R) C ( R ) . □
The corner value of the moving-ball distance
The asymmetry of Ω \Omega Ω is recovered from q q q only in the limit ρ → 1 \rho\to1 ρ → 1 , z → 0 z\to0 z → 0 ; this replaces the identity q ( 1 ) = ω n A ( Ω ) q(1)=\omega_n A(\Omega) q ( 1 ) = ω n A ( Ω ) in the proof of [17 Theorem 3.7] .
Lemma 3.9. Let Ω ⊂ R n \Omega\subset\mathbb{R}^n Ω ⊂ R n be bounded and open with ∣ Ω ∣ = ω n |\Omega|=\omega_n ∣Ω∣ = ω n , and let E z = E 1 , z E_z=E_{1,z} E z = E 1 , z . Then ∣ E z △ Ω ∣ → 0 |E_z\mathbin{\triangle}\Omega|\to0 ∣ E z △ Ω∣ → 0 as z ↓ 0 z\downarrow0 z ↓ 0 , and
∣ q ( ρ , z ) − ω n A ( Ω ) ∣ ≤ 2 n ω n ( 1 − ρ ) + ∣ E z △ Ω ∣ ( 0 < ρ ≤ 1 , z > 0 ) . (43) |q(\rho,z)-\omega_nA(\Omega)|\le2n\omega_n(1-\rho)+|E_z\mathbin{\triangle}\Omega|\qquad(0<\rho\le1,\ z>0). \tag*{(43)} ∣ q ( ρ , z ) − ω n A ( Ω ) ∣ ≤ 2 n ω n ( 1 − ρ ) + ∣ E z △ Ω∣ ( 0 < ρ ≤ 1 , z > 0 ) . (43) In particular,
lim ρ → 1 − z → 0 + q ( ρ , z ) = ω n A ( Ω ) . (44) \lim_{\substack{\rho\to1^-\\ z\to0^+}}q(\rho,z)=\omega_nA(\Omega). \tag*{(44)} ρ → 1 − z → 0 + lim q ( ρ , z ) = ω n A ( Ω ) . (44)
Proof. By Lemma 3.1 , ∣ E z ∣ = ω n = ∣ Ω ∣ |E_z|=\omega_n=|\Omega| ∣ E z ∣ = ω n = ∣Ω∣ for every z > 0 z>0 z > 0 , so ∣ E z △ Ω ∣ = 2 ∣ Ω ∖ E z ∣ |E_z\mathbin{\triangle}\Omega|=2|\Omega\setminus E_z| ∣ E z △ Ω∣ = 2∣Ω ∖ E z ∣ . Let t z = t ( 1 , z ) t_z=t(1,z) t z = t ( 1 , z ) . Since U ( ⋅ , z ) ≥ t z U(\cdot,z)\ge t_z U ( ⋅ , z ) ≥ t z on E z E_z E z and U ( ⋅ , z ) ≤ t z U(\cdot,z)\le t_z U ( ⋅ , z ) ≤ t z outside it, and ∣ Ω ∣ = ∣ E z ∣ |\Omega|=|E_z| ∣Ω∣ = ∣ E z ∣ , we have
∫ E z U ( ⋅ , z ) ≥ ∫ Ω U ( ⋅ , z ) , \int_{E_z} U(\cdot,z)\ge\int_{\Omega}U(\cdot,z), ∫ E z U ( ⋅ , z ) ≥ ∫ Ω U ( ⋅ , z ) ,
that is, ∫ Ω ∖ E z U ( ⋅ , z ) ≤ ∫ E z ∖ Ω U ( ⋅ , z ) \int_{\Omega\setminus E_z}U(\cdot,z)\le\int_{E_z\setminus\Omega}U(\cdot,z) ∫ Ω ∖ E z U ( ⋅ , z ) ≤ ∫ E z ∖ Ω U ( ⋅ , z ) . Using u = 0 u=0 u = 0 outside Ω \Omega Ω , we obtain
∫ Ω ∖ E z u ≤ ∫ Ω ∖ E z ∣ U ( ⋅ , z ) − u ∣ + ∫ E z ∖ Ω ∣ U ( ⋅ , z ) − u ∣ ≤ ( 2 ω n ) 1 / 2 ∥ U ( ⋅ , z ) − u ∥ 2 = : ε z , \int_{\Omega\setminus E_z}u\le\int_{\Omega\setminus E_z}|U(\cdot,z)-u|+\int_{E_z\setminus\Omega}|U(\cdot,z)-u|\le(2\omega_n)^{1/2}\|U(\cdot,z)-u\|_2=:\varepsilon_z, ∫ Ω ∖ E z u ≤ ∫ Ω ∖ E z ∣ U ( ⋅ , z ) − u ∣ + ∫ E z ∖ Ω ∣ U ( ⋅ , z ) − u ∣ ≤ ( 2 ω n ) 1/2 ∥ U ( ⋅ , z ) − u ∥ 2 =: ε z ,
and ε z → 0 \varepsilon_z\to0 ε z → 0 by Lemma 2.1 .
We also need that u > 0 u>0 u > 0 almost everywhere in Ω \Omega Ω . Let B ′ ⋐ Ω B'\Subset\Omega B ′ ⋐ Ω be a ball. The function f + = ( u B ′ − u ) + f_+=(u_{B'}-u)_+ f + = ( u B ′ − u ) + lies in H s ( R n ) H^s(\mathbb{R}^n) H s ( R n ) and vanishes outside B ′ B' B ′ , so f + ∈ H 0 s ( B ′ ) ⊂ H 0 s ( Ω ) f_+\in H_0^s(B')\subset H^s_0(\Omega) f + ∈ H 0 s ( B ′ ) ⊂ H 0 s ( Ω ) . Testing the two torsion equations with f + f_+ f + gives E s ( f , f + ) = ∫ B ′ f + − ∫ Ω f + ≤ 0 \mathcal{E}_s(f,f_+)=\int_{B'}f_+-\int_{\Omega}f_+\le0 E s ( f , f + ) = ∫ B ′ f + − ∫ Ω f + ≤ 0 , while E s ( f , f + ) ≥ E s ( f + ) \mathcal{E}_s(f,f_+)\ge\mathcal{E}_s(f_+) E s ( f , f + ) ≥ E s ( f + ) . Hence f + = 0 f_+=0 f + = 0 , that is, u ≥ u B ′ > 0 u\ge u_{B'}>0 u ≥ u B ′ > 0 in B ′ B' B ′ . Covering Ω \Omega Ω by countably many such balls proves the claim.
For σ > 0 \sigma>0 σ > 0 we have Ω ∖ E z ⊂ ( Ω ∩ { u ≤ σ } ) ∪ ( { u > σ } ∖ E z ) \Omega\setminus E_z\subset(\Omega\cap\{u\le\sigma\})\cup(\{u>\sigma\}\setminus E_z) Ω ∖ E z ⊂ ( Ω ∩ { u ≤ σ }) ∪ ({ u > σ } ∖ E z ) , so that
∣ Ω ∖ E z ∣ ≤ ∣ Ω ∩ { 0 < u ≤ σ } ∣ + σ − 1 ε z . |\Omega\setminus E_z|\le|\Omega\cap\{0<u\le\sigma\}|+\sigma^{-1}\varepsilon_z. ∣Ω ∖ E z ∣ ≤ ∣Ω ∩ { 0 < u ≤ σ } ∣ + σ − 1 ε z .
Letting first z ↓ 0 z\downarrow0 z ↓ 0 and then σ ↓ 0 \sigma\downarrow0 σ ↓ 0 (using ∣ Ω ∣ < ∞ |\Omega|<\infty ∣Ω∣ < ∞ ) gives ∣ Ω ∖ E z ∣ → 0 |\Omega\setminus E_z|\to0 ∣Ω ∖ E z ∣ → 0 .
For every center a a a , ∣ E z △ B 1 ( a ) ∣ − ∣ Ω △ B 1 ( a ) ∣ ≤ ∣ E z △ Ω ∣ |E_z\mathbin{\triangle}B_1(a)|-|\Omega\mathbin{\triangle}B_1(a)|\le|E_z\mathbin{\triangle}\Omega| ∣ E z △ B 1 ( a ) ∣ − ∣Ω △ B 1 ( a ) ∣ ≤ ∣ E z △ Ω∣ . Taking the infimum over a a a and recalling ∣ Ω ∣ = ω n |\Omega|=\omega_n ∣Ω∣ = ω n gives ∣ q ( 1 , z ) − ω n A ( Ω ) ∣ ≤ ∣ E z △ Ω ∣ |q(1,z)-\omega_n\mathcal{A}(\Omega)|\le|E_z\mathbin{\triangle}\Omega| ∣ q ( 1 , z ) − ω n A ( Ω ) ∣ ≤ ∣ E z △ Ω∣ . Together with (31) , this proves (43) and hence (44) . □
Quantitative torsion inequality for confined sets
We now combine the weighted action estimate (42) with the weak flux identity (39) to prove the quantitative estimate for confined sets. The following theorem is the counterpart of [17 Theorem 3.7] . The proof is substantially more technical than its local counterpart.
Theorem 3.10. For every n ≥ 2 n\ge2 n ≥ 2 , s ∈ ( 0 , 1 ) s\in(0,1) s ∈ ( 0 , 1 ) , and R ≥ 1 R\ge1 R ≥ 1 there is C ( R ) > 0 C(R)>0 C ( R ) > 0 such that every open set Ω ⊂ B R \Omega\subset B_R Ω ⊂ B R with ∣ Ω ∣ = ω n |\Omega|=\omega_n ∣Ω∣ = ω n satisfies
A ( Ω ) 2 ≤ C ( R ) δ T ( Ω ) . (45) \mathcal{A}(\Omega)^2\le C(R)\delta_T(\Omega). \tag*{(45)} A ( Ω ) 2 ≤ C ( R ) δ T ( Ω ) . (45)
Proof. Let ℓ \ell ℓ , c 0 c_0 c 0 and C 0 C_0 C 0 be as in Lemma A.1 ; they depend only on n n n and s s s . Proposition 3.4 yields that q q q has horizontal Lipschitz constant Λ = 2 n ω n \Lambda=2n\omega_n Λ = 2 n ω n and 0 ≤ q ≤ Λ 0\le q\le\Lambda 0 ≤ q ≤ Λ . Set A = ω n A ( Ω ) A=\omega_n\mathcal{A}(\Omega) A = ω n A ( Ω ) ; by Lemma 3.9 , q ( 1 , z ) → A q(1,z)\to A q ( 1 , z ) → A as z ↓ 0 z\downarrow0 z ↓ 0 . If A = 0 A=0 A = 0 , the conclusion follows from δ T ( Ω ) ≥ 0 \delta_T(\Omega)\ge0 δ T ( Ω ) ≥ 0 ; assume henceforth that A > 0 A>0 A > 0 . Define
J = ∫ R ∫ 0 1 z 1 − 2 s [ w 2 ( q 2 + ( q ρ ) 2 ) + ∣ J A ∣ 2 + ∣ ∇ v ∣ 2 ] d ρ d z + ∫ 1 − ℓ 1 ∣ v ( ρ , 0 ) ∣ 2 ( 1 − ρ ) 2 s d ρ . (46) \mathcal{J}=\int_{\mathbb{R}}\int_0^1 z^{1-2s}\left[w^2\left(q^2+(q_\rho)^2\right)+|J_A|^2+|\nabla v|^2\right]\,d\rho\,dz+\int_{1-\ell}^1\frac{|v(\rho,0)|^2}{(1-\rho)^{2s}}\,d\rho. \tag*{(46)} J = ∫ R ∫ 0 1 z 1 − 2 s [ w 2 ( q 2 + ( q ρ ) 2 ) + ∣ J A ∣ 2 + ∣∇ v ∣ 2 ] d ρ d z + ∫ 1 − ℓ 1 ( 1 − ρ ) 2 s ∣ v ( ρ , 0 ) ∣ 2 d ρ . (46)
We claim that
A 2 ≤ C J , (47) A^2\le C\mathcal{J}, \tag*{(47)} A 2 ≤ C J , (47)
where C C C depends only on s , ℓ , c 0 , C 0 s,\ell,c_0,C_0 s , ℓ , c 0 , C 0 . By Corollary 3.8, this would conclude the proof of (45) .
We thus prove (47) . Choose a nonnegative χ ∈ C ∞ ( ( 1 , 2 ) ) \chi\in C^\infty((1,2)) χ ∈ C ∞ (( 1 , 2 )) with ∫ χ = 1 \int\chi=1 ∫ χ = 1 , put r 0 = ℓ / 4 r_0=\ell/4 r 0 = ℓ /4 , and fix K ≥ max { 1 , 2 Λ / r 0 } K\ge\max\{1,2\Lambda/r_0\} K ≥ max { 1 , 2Λ/ r 0 } . Write τ = A / K \tau=A/K τ = A / K , so 0 < τ ≤ r 0 / 2 0<\tau\le r_0/2 0 < τ ≤ r 0 /2 .
Let us first give an idea of the proof. Formally, one would like to compare A A A with an interior value by integrating a derivative in z z z . However, q A q_A q A has no useful Sobolev regularity in z z z , and bulk L 2 L^2 L 2 estimates do not control a prescribed horizontal slice. The weak flux identity (39) allows us instead to differentiate horizontal averages of w q A / w B wq_A/w_B w q A / w B at positive heights. We average on windows of width λ ( z ) = z + τ \lambda(z)=z+\tau λ ( z ) = z + τ near ρ = 1 \rho=1 ρ = 1 . On such a window 1 − ρ + z ≳ λ ( z ) 1-\rho+z\gtrsim\lambda(z) 1 − ρ + z ≳ λ ( z ) , which matches the ball-speed estimate; the scale regularizes the corner at size τ \tau τ and makes comparisons compatible to z z z for z ≫ τ z\gg\tau z ≫ τ . The corner limit and the Hardy boundary term bound the liminf of the average at the bottom from below, a one-sided derivative estimate propagates this bound to a fixed height interval, and there the term w 2 q 2 w^2q^2 w 2 q 2 in J \mathcal{J} J controls the average. The cap in q A q_A q A is what makes this work at the sharp scale. At positive heights q q q need not be of order A A A , while in the error terms it appears undifferentiated against derivatives of size λ − 1 \lambda^{-1} λ − 1 . The bound q A ≤ A q_A\le A q A ≤ A turns these factors into A / λ ≤ A / τ = K A/\lambda\le A/\tau=K A / λ ≤ A / τ = K . With only q ≤ 2 ω n q\le2\omega_n q ≤ 2 ω n they would be of size τ − 1 \tau^{-1} τ − 1 , and the argument would give A 3 ≲ J A^3\lesssim\mathcal{J} A 3 ≲ J instead of A 2 ≲ J A^2\lesssim\mathcal{J} A 2 ≲ J .
We now start the proof of (47) . Set
λ ( z ) = z + τ , I λ ( z ) = ( 1 − 2 λ ( z ) , 1 − λ ( z ) ) , \lambda(z)=z+\tau,\qquad I_{\lambda(z)}=(1-2\lambda(z),1-\lambda(z)), λ ( z ) = z + τ , I λ ( z ) = ( 1 − 2 λ ( z ) , 1 − λ ( z )) ,
φ ( ρ , z ) = 1 λ ( z ) χ ( 1 − ρ λ ( z ) ) , H ( z ) = ∫ ρ ∞ w q A w B φ d ρ . (48) \varphi(\rho,z)=\frac{1}{\lambda(z)}\chi\left(\frac{1-\rho}{\lambda(z)}\right),\qquad H(z)=\int_{\rho}^{\infty}\frac{wq_A}{w_B}\varphi\,\mathrm{d}\rho. \tag*{(48)} φ ( ρ , z ) = λ ( z ) 1 χ ( λ ( z ) 1 − ρ ) , H ( z ) = ∫ ρ ∞ w B w q A φ d ρ . (48)
For 0 ≤ z ≤ r 0 0\le z\le r_0 0 ≤ z ≤ r 0 , we have τ ≤ λ ( z ) ≤ 3 r 0 / 2 \tau\le\lambda(z)\le3r_0/2 τ ≤ λ ( z ) ≤ 3 r 0 /2 , and hence 2 λ ( z ) ≤ 3 ℓ / 4 < ℓ 2\lambda(z)\le3\ell/4<\ell 2 λ ( z ) ≤ 3 ℓ /4 < ℓ . Thus the horizontal support remains in a compact subset of ( 1 − ℓ , 1 ) (1-\ell,1) ( 1 − ℓ , 1 ) for each fixed A > 0 A>0 A > 0 . Write λ = λ ( z ) \lambda=\lambda(z) λ = λ ( z ) below. On I λ I_\lambda I λ , λ < 1 − ρ < 2 λ \lambda<1-\rho<2\lambda λ < 1 − ρ < 2 λ and λ < 1 − ρ + z < 3 λ \lambda<1-\rho+z<3\lambda λ < 1 − ρ + z < 3 λ . By (A.1),
w B ≍ λ s − 1 , ∣ ∇ w B ∣ w B ≤ C λ − 1 , ∣ φ ∣ ≤ C λ − 1 , ∣ φ ρ ∣ + ∣ φ z ∣ ≤ C λ − 2 . (49) w_B\asymp\lambda^{s-1},\qquad\frac{|\nabla w_B|}{w_B}\le C\lambda^{-1},\qquad|\varphi|\le C\lambda^{-1},\qquad|\varphi_\rho|+|\varphi_z|\le C\lambda^{-2}. \tag*{(49)} w B ≍ λ s − 1 , w B ∣∇ w B ∣ ≤ C λ − 1 , ∣ φ ∣ ≤ C λ − 1 , ∣ φ ρ ∣ + ∣ φ z ∣ ≤ C λ − 2 . (49)
Notice that, by definition, ∫ φ ( ρ , z ) d ρ = 1 \int\varphi(\rho,z)\,\mathrm{d}\rho=1 ∫ φ ( ρ , z ) d ρ = 1 . Since λ ′ ( z ) = 1 \lambda'(z)=1 λ ′ ( z ) = 1 , direct differentiation gives
φ z = − 1 λ 2 χ ( 1 − ρ λ ) − 1 − ρ λ 3 χ ′ ( 1 − ρ λ ) = ∂ ρ ( 1 − ρ λ φ ) . (50) \varphi_z=-\frac{1}{\lambda^2}\chi\left(\frac{1-\rho}{\lambda}\right)-\frac{1-\rho}{\lambda^3}\chi'\left(\frac{1-\rho}{\lambda}\right)=\partial_\rho\left(\frac{1-\rho}{\lambda}\varphi\right). \tag*{(50)} φ z = − λ 2 1 χ ( λ 1 − ρ ) − λ 3 1 − ρ χ ′ ( λ 1 − ρ ) = ∂ ρ ( λ 1 − ρ φ ) . (50)
We first explain how to differentiate H H H away from the bottom. As noted in Remark 3.7, we may write (39) as ∂ z ( w q A / w B ) = − ∂ ρ B A + C A \partial_z(wq_A/w_B)=-\partial_\rho B_A+C_A ∂ z ( w q A / w B ) = − ∂ ρ B A + C A , where B A , C A B_A,C_A B A , C A are given by (40) . Testing against φ ( ρ , z ) θ ( z ) \varphi(\rho,z)\theta(z) φ ( ρ , z ) θ ( z ) , with θ ∈ C c ∞ ( ( 0 , r 0 ) ) \theta\in C^\infty_c((0,r_0)) θ ∈ C c ∞ (( 0 , r 0 )) , yields in the sense of distributions
H ′ ( z ) = ∫ ρ ∞ ( B A φ ρ + C A φ + w q A w B φ z ) d ρ . (51) H'(z)=\int_{\rho}^{\infty}\left(B_A\varphi_\rho+C_A\varphi+\frac{wq_A}{w_B}\varphi_z\right)\,\mathrm{d}\rho. \tag*{(51)} H ′ ( z ) = ∫ ρ ∞ ( B A φ ρ + C A φ + w B w q A φ z ) d ρ . (51)
On every compact height interval in ( 0 , r 0 ) (0,r_0) ( 0 , r 0 ) the right-hand side is integrable: q A , ( q A ) ρ q_A,(q_A)_\rho q A , ( q A ) ρ are bounded, v ∈ H l o c 1 ( R ) v\in H^1_{\mathrm{loc}}(\mathbb{R}) v ∈ H loc 1 ( R ) , and the coefficients are smooth on the horizontal support. Thus H ∈ W l o c 1 ( ( 0 , r 0 ) ) H\in W^1_{\mathrm{loc}}((0,r_0)) H ∈ W loc 1 (( 0 , r 0 )) .
Using w = w B − v ρ w=w_B-v_\rho w = w B − v ρ and (50) , horizontal integration by parts gives
∫ w q A w B φ z d ρ = − ∫ 1 − ρ λ ( q A ) ρ φ d ρ − ∫ v ρ q A w B ∂ ρ ( 1 − ρ λ φ ) d ρ = − ∫ 1 − ρ λ w w B ( q A ) ρ φ d ρ − ∫ v ρ w B ∂ ρ ( 1 − ρ λ q A φ ) d ρ , \begin{aligned}
\int\frac{wq_A}{w_B}\varphi_z\,\mathrm{d}\rho
=-\int\frac{1-\rho}{\lambda}(q_A)_\rho\varphi\,\mathrm{d}\rho-\int\frac{v_\rho q_A}{w_B}\partial_\rho\left(\frac{1-\rho}{\lambda}\varphi\right)\,\mathrm{d}\rho\\
&=-\int\frac{1-\rho}{\lambda}\frac{w}{w_B}(q_A)_\rho\varphi\,\mathrm{d}\rho-\int\frac{v_\rho}{w_B}\partial_\rho\left(\frac{1-\rho}{\lambda}q_A\varphi\right)\,\mathrm{d}\rho,
\end{aligned} ∫ w B w q A φ z d ρ = − ∫ λ 1 − ρ ( q A ) ρ φ d ρ − ∫ w B v ρ q A ∂ ρ ( λ 1 − ρ φ ) d ρ = − ∫ λ 1 − ρ w B w ( q A ) ρ φ d ρ − ∫ w B v ρ ∂ ρ ( λ 1 − ρ q A φ ) d ρ ,
where the second equality uses 1 = w / w B + v ρ / w B 1=w/w_B+v_\rho/w_B 1 = w / w B + v ρ / w B . Substituting in (51) gives the exact identity
H ′ ( z ) = ∫ I λ [ J A w B − 1 − ρ λ w w B ( q A ) ρ ] φ d ρ + ∫ I λ v z ∂ ρ ( q A φ w B ) d ρ + ∫ I λ v ρ q A ( w B ) z w B 2 φ d ρ − ∫ I λ v ρ w B ∂ ρ ( 1 − ρ λ q A φ ) d ρ . (52) \begin{aligned}
H'(z)={}&\int_{I_\lambda}\left[\frac{J_A}{w_B}-\frac{1-\rho}{\lambda}\frac{w}{w_B}(q_A)_\rho\right]\varphi\,\mathrm{d}\rho\\
&+\int_{I_\lambda}v_z\partial_\rho\left(\frac{q_A\varphi}{w_B}\right)\,\mathrm{d}\rho+\int_{I_\lambda}\frac{v_\rho q_A(w_B)_z}{w_B^2}\varphi\,\mathrm{d}\rho\\
&-\int_{I_\lambda}\frac{v_\rho}{w_B}\partial_\rho\left(\frac{1-\rho}{\lambda}q_A\varphi\right)\,\mathrm{d}\rho. \tag*{(52)}
\end{aligned} H ′ ( z ) = ∫ I λ [ w B J A − λ 1 − ρ w B w ( q A ) ρ ] φ d ρ + ∫ I λ v z ∂ ρ ( w B q A φ ) d ρ + ∫ I λ w B 2 v ρ q A ( w B ) z φ d ρ − ∫ I λ w B v ρ ∂ ρ ( λ 1 − ρ q A φ ) d ρ . (52)
We estimate this derivative from below. Since w ≥ 0 w\ge0 w ≥ 0 , ( 1 − ρ ) / λ > 0 (1-\rho)/\lambda>0 ( 1 − ρ ) / λ > 0 , φ ≥ 0 \varphi\ge0 φ ≥ 0 , and ( ( q A ) ρ ) + ≤ ( q ρ ) + ((q_A)_\rho)_+\le(q_\rho)_+ (( q A ) ρ ) + ≤ ( q ρ ) + , the first line of (52) is bounded below by
− C λ 1 / 2 − s [ ∥ J A ( ⋅ , z ) ∥ L 2 ( I λ ) + ∥ w ( q ρ ) + ( ⋅ , z ) ∥ L 2 ( I λ ) ] . -C\lambda^{1/2-s}\left[\|J_A(\cdot,z)\|_{L^2(I_\lambda)}+\|w(q_\rho)_+(\cdot,z)\|_{L^2(I_\lambda)}\right]. − C λ 1/2 − s [ ∥ J A ( ⋅ , z ) ∥ L 2 ( I λ ) + ∥ w ( q ρ ) + ( ⋅ , z ) ∥ L 2 ( I λ ) ] .
Indeed, φ / w B ≤ C λ − s \varphi/w_B\le C\lambda^{-s} φ / w B ≤ C λ − s , ( 1 − ρ ) / λ ≤ 2 (1-\rho)/\lambda\le2 ( 1 − ρ ) / λ ≤ 2 , and ∣ I λ ∣ = λ |I_\lambda|=\lambda ∣ I λ ∣ = λ . For the remaining terms, the product rule, 0 ≤ q A ≤ A 0\le q_A\le A 0 ≤ q A ≤ A , ∣ ( q A ) ρ ∣ ≤ Λ |(q_A)_\rho|\le\Lambda ∣ ( q A ) ρ ∣ ≤ Λ , and (49) give
∣ ∂ ρ ( q A φ w B ) ∣ ≤ C ( Λ λ − s + A λ − 1 − s ) , \left|\partial_\rho\left(\frac{q_A\varphi}{w_B}\right)\right|\le C(\Lambda\lambda^{-s}+A\lambda^{-1-s}), ∂ ρ ( w B q A φ ) ≤ C ( Λ λ − s + A λ − 1 − s ) ,
q A ∣ ( w B ) z ∣ w B 2 ∣ φ ∣ ≤ C A λ − 1 − s , \frac{q_A|(w_B)_z|}{w_B^2}|\varphi|\le CA\lambda^{-1-s}, w B 2 q A ∣ ( w B ) z ∣ ∣ φ ∣ ≤ C A λ − 1 − s ,
1 w B ∣ ∂ ρ ( 1 − ρ λ q A φ ) ∣ ≤ C ( Λ λ − s + A λ − 1 − s ) . (53) \frac{1}{w_B}\left|\partial_\rho\left(\frac{1-\rho}{\lambda}q_A\varphi\right)\right|\le C(\Lambda\lambda^{-s}+A\lambda^{-1-s}). \tag*{(53)} w B 1 ∂ ρ ( λ 1 − ρ q A φ ) ≤ C ( Λ λ − s + A λ − 1 − s ) . (53)
For example, in the first line the derivative of q A q_A q A contributes C Λ λ − s C\Lambda\lambda^{-s} C Λ λ − s , while the derivatives of φ \varphi φ and 1 / w B 1/w_B 1/ w B each contribute C A λ − 1 − s CA\lambda^{-1-s} C A λ − 1 − s . In the last line, ∣ ∂ ρ ( ( 1 − ρ ) / λ ) ∣ = λ − 1 |\partial_\rho((1-\rho)/\lambda)|=\lambda^{-1} ∣ ∂ ρ (( 1 − ρ ) / λ ) ∣ = λ − 1 gives the same power. Cauchy–Schwarz on I λ I_\lambda I λ therefore yields
H ′ ( z ) ≥ − C λ 1 / 2 − s [ ∥ J A ( ⋅ , z ) ∥ L 2 ( I λ ) + ∥ w ( q ρ ) + ( ⋅ , z ) ∥ L 2 ( I λ ) + ( Λ + A λ ) ∥ ∇ v ( ⋅ , z ) ∥ L 2 ( I λ ) ] . H'(z)\ge-C\lambda^{1/2-s}\left[\|J_A(\cdot,z)\|_{L^2(I_\lambda)}+\|w(q_\rho)_+(\cdot,z)\|_{L^2(I_\lambda)}+\left(\Lambda+\frac{A}{\lambda}\right)\|\nabla v(\cdot,z)\|_{L^2(I_\lambda)}\right]. H ′ ( z ) ≥ − C λ 1/2 − s [ ∥ J A ( ⋅ , z ) ∥ L 2 ( I λ ) + ∥ w ( q ρ ) + ( ⋅ , z ) ∥ L 2 ( I λ ) + ( Λ + λ A ) ∥∇ v ( ⋅ , z ) ∥ L 2 ( I λ ) ] .
Since λ ≥ τ = A / K \lambda\ge\tau=A/K λ ≥ τ = A / K , we have A / λ ≤ K A/\lambda\le K A / λ ≤ K , and obtain for almost every z ∈ ( 0 , r 0 ) z\in(0,r_0) z ∈ ( 0 , r 0 )
H ′ ( z ) ≥ − C λ ( z ) 1 / 2 − s [ ∥ J A ( ⋅ , z ) ∥ L 2 ( I λ ( z ) ) + ∥ w ( q ρ ) + ( ⋅ , z ) ∥ L 2 ( I λ ( z ) ) + ∥ ∇ v ( ⋅ , z ) ∥ L 2 ( I λ ( z ) ) ] . (54) H'(z)\ge-C\lambda(z)^{1/2-s}\left[\|J_A(\cdot,z)\|_{L^2(I_{\lambda(z)})}+\|w(q_\rho)_+(\cdot,z)\|_{L^2(I_{\lambda(z)})}+\|\nabla v(\cdot,z)\|_{L^2(I_{\lambda(z)})}\right]. \tag*{(54)} H ′ ( z ) ≥ − C λ ( z ) 1/2 − s [ ∥ J A ( ⋅ , z ) ∥ L 2 ( I λ ( z ) ) + ∥ w ( q ρ ) + ( ⋅ , z ) ∥ L 2 ( I λ ( z ) ) + ∥∇ v ( ⋅ , z ) ∥ L 2 ( I λ ( z ) ) ] . (54)
We next claim that
lim inf z ↓ 0 H ( z ) ≥ A − 2 Λ τ − C τ J . (55) \liminf_{z\downarrow0}H(z)\ge A-2\Lambda\tau-C\sqrt{\tau\mathcal{J}}. \tag*{(55)} z ↓ 0 lim inf H ( z ) ≥ A − 2Λ τ − C τ J . (55)
Since w = w B − v ρ w=w_B-v_\rho w = w B − v ρ , horizontal integration by parts gives
H ( z ) = ∫ I λ q A φ d ρ + ∫ I λ v ∂ ρ ( q A φ w B ) d ρ . (56) H(z)=\int_{I_\lambda}q_A\varphi\,d\rho+\int_{I_\lambda}v\,\partial_\rho\left(\frac{q_A\varphi}{w_B}\right)\,d\rho. \tag*{(56)} H ( z ) = ∫ I λ q A φ d ρ + ∫ I λ v ∂ ρ ( w B q A φ ) d ρ . (56)
By horizontal Lipschitz continuity, q A ( ρ , z ) ≥ min { q ( 1 , z ) − 2 Λ λ ( z ) , A } q_A(\rho,z)\ge\min\{q(1,z)-2\Lambda\lambda(z),A\} q A ( ρ , z ) ≥ min { q ( 1 , z ) − 2Λ λ ( z ) , A } on I λ I_\lambda I λ . Since q ( 1 , z ) → A q(1,z)\to A q ( 1 , z ) → A , λ ( z ) → τ \lambda(z)\to\tau λ ( z ) → τ , and ∫ φ = 1 \int\varphi=1 ∫ φ = 1 ,
lim inf z ↓ 0 ∫ I λ q A φ d ρ ≥ A − 2 Λ τ . \liminf_{z\downarrow0}\int_{I_\lambda}q_A\varphi\,d\rho\ge A-2\Lambda\tau. z ↓ 0 lim inf ∫ I λ q A φ d ρ ≥ A − 2Λ τ .
For z < τ z<\tau z < τ , (53) and A / τ = K A/\tau=K A / τ = K bound the coefficient in the second term of (56) by C τ − s C\tau^{-s} C τ − s . Write v ( ⋅ , z ) = v 0 + [ v ( ⋅ , z ) − v 0 ] v(\cdot,z)=v_0+[v(\cdot,z)-v_0] v ( ⋅ , z ) = v 0 + [ v ( ⋅ , z ) − v 0 ] , where v 0 v_0 v 0 is the bottom trace. The strong L 2 ( ( 1 − ℓ , 1 ) ) L^2((1-\ell,1)) L 2 (( 1 − ℓ , 1 )) convergence of the slices of v v v given by Lemma A.2 gives
∣ ∫ I λ ( v − v 0 ) ∂ ρ ( q A φ w B ) d ρ ∣ ≤ C τ − s λ ∥ v ( ⋅ , z ) − v 0 ∥ L 2 ( ( 1 − ℓ , 1 ) ) ⟶ 0. \left|\int_{I_\lambda}(v-v_0)\partial_\rho\left(\frac{q_A\varphi}{w_B}\right)\,d\rho\right|\le C\tau^{-s}\sqrt{\lambda}\|v(\cdot,z)-v_0\|_{L^2((1-\ell,1))}\longrightarrow0. ∫ I λ ( v − v 0 ) ∂ ρ ( w B q A φ ) d ρ ≤ C τ − s λ ∥ v ( ⋅ , z ) − v 0 ∥ L 2 (( 1 − ℓ , 1 )) ⟶ 0.
Also I λ ( z ) → I τ = ( 1 − 2 τ , 1 − τ ) I_{\lambda(z)}\to I_\tau=(1-2\tau,1-\tau) I λ ( z ) → I τ = ( 1 − 2 τ , 1 − τ ) in measure, so integrability of v 0 v_0 v 0 implies ∫ I λ ( z ) ∣ v 0 ∣ d ρ → ∫ I τ ∣ v 0 ∣ d ρ \int_{I_{\lambda(z)}}|v_0|\,d\rho\to\int_{I_\tau}|v_0|\,d\rho ∫ I λ ( z ) ∣ v 0 ∣ d ρ → ∫ I τ ∣ v 0 ∣ d ρ . The Hardy term in (46) gives
∫ I τ ∣ v 0 ∣ d ρ ≤ ( ∫ I τ ∣ v 0 ∣ 2 ( 1 − ρ ) 2 s d ρ ) 1 / 2 ( ∫ I τ ( 1 − ρ ) 2 s d ρ ) 1 / 2 ≤ C τ s + 1 / 2 J . \int_{I_\tau}|v_0|\,d\rho\le\left(\int_{I_\tau}\frac{|v_0|^2}{(1-\rho)^{2s}}\,d\rho\right)^{1/2}\left(\int_{I_\tau}(1-\rho)^{2s}\,d\rho\right)^{1/2}\le C\tau^{s+1/2}\sqrt{\mathcal{J}}. ∫ I τ ∣ v 0 ∣ d ρ ≤ ( ∫ I τ ( 1 − ρ ) 2 s ∣ v 0 ∣ 2 d ρ ) 1/2 ( ∫ I τ ( 1 − ρ ) 2 s d ρ ) 1/2 ≤ C τ s + 1/2 J .
Consequently the limsup of the absolute value of the second term in (56) is at most
C τ − s τ s + 1 / 2 J = C τ J . C\tau^{-s}\tau^{s+1/2}\sqrt{\mathcal{J}}=C\sqrt{\tau\mathcal{J}}. C τ − s τ s + 1/2 J = C τ J .
This proves (55) .
We now propagate this information directly to a fixed height. Local absolute continuity and (54) give, for ε < z \varepsilon<z ε < z ,
H ( ε ) ≤ H ( z ) + C ∫ ε z λ ( z ′ ) 1 / 2 − s [ ∥ J A ( ⋅ , z ′ ) ∥ L 2 ( I λ ( z ′ ) ) + ∥ w ( q ρ ) + ( ⋅ , z ′ ) ∥ L 2 ( I λ ( z ′ ) ) + ∥ ∇ v ( ⋅ , z ′ ) ∥ L 2 ( I λ ( z ′ ) ) ] d z ′ ≤ H ( z ) + C J , H(\varepsilon)\le H(z)+C\int_\varepsilon^z\lambda(z')^{1/2-s}\left[\|J_A(\cdot,z')\|_{L^2(I_{\lambda(z')})}+\|w(q_\rho)_+(\cdot,z')\|_{L^2(I_{\lambda(z')})}+\|\nabla v(\cdot,z')\|_{L^2(I_{\lambda(z')})}\right]\,dz'\le H(z)+C\sqrt{\mathcal{J}}, H ( ε ) ≤ H ( z ) + C ∫ ε z λ ( z ′ ) 1/2 − s [ ∥ J A ( ⋅ , z ′ ) ∥ L 2 ( I λ ( z ′ ) ) + ∥ w ( q ρ ) + ( ⋅ , z ′ ) ∥ L 2 ( I λ ( z ′ ) ) + ∥∇ v ( ⋅ , z ′ ) ∥ L 2 ( I λ ( z ′ ) ) ] d z ′ ≤ H ( z ) + C J ,
where weighted Cauchy–Schwarz and (46) apply because
∫ 0 r 0 λ ( z ) 1 − 2 s z 2 s − 1 d z = ∫ 0 r 0 ( z + τ z ) 1 − 2 s d z ≤ C r 0 , \int_0^{r_0}\lambda(z)^{1-2s}z^{2s-1}\,dz=\int_0^{r_0}\left(\frac{z+\tau}{z}\right)^{1-2s}\,dz\le Cr_0, ∫ 0 r 0 λ ( z ) 1 − 2 s z 2 s − 1 d z = ∫ 0 r 0 ( z z + τ ) 1 − 2 s d z ≤ C r 0 ,
uniformly in 0 < τ ≤ r 0 / 2 0<\tau\le r_0/2 0 < τ ≤ r 0 /2 . Combining this with (55) yields,
A − 2 Λ τ − C τ J ≤ H ( z ) + C J . (57) A-2\Lambda\tau-C\sqrt{\tau\mathcal{J}}\le H(z)+C\sqrt{\mathcal{J}}. \tag*{(57)} A − 2Λ τ − C τ J ≤ H ( z ) + C J . (57)
Finally, for z ∈ [ r 0 / 2 , r 0 ] z \in[r_0/2,r_0] z ∈ [ r 0 /2 , r 0 ] , we have λ ( z ) ≍ r 0 \lambda(z) \asymp r_0 λ ( z ) ≍ r 0 , φ ≤ C \varphi\le C φ ≤ C , w B ≍ 1 w_B \asymp1 w B ≍ 1 , and z 1 − 2 s ≍ 1 z^{1-2s} \asymp1 z 1 − 2 s ≍ 1 . Since 0 ≤ q A ≤ q 0 \le q_A \le q 0 ≤ q A ≤ q and w ≥ 0 w \ge0 w ≥ 0 ,
∫ r 0 / 2 r 0 H ( z ) d z ≤ C ∫ r 0 / 2 r 0 ∫ I λ ( z ) w q d ρ d z ≤ C ( ∫ r 0 / 2 r 0 ∫ I λ ( z ) z 1 − 2 s w 2 q 2 d ρ d z ) 1 / 2 ≤ C F . \int_{r_0/2}^{r_0} H(z)\,dz \le C\int_{r_0/2}^{r_0}\int_{I_{\lambda(z)}} wq\,d\rho\,dz
\le C\left(\int_{r_0/2}^{r_0}\int_{I_{\lambda(z)}} z^{1-2s}w^2q^2\,d\rho\,dz\right)^{1/2}
\le C\sqrt{\mathcal{F}}. ∫ r 0 /2 r 0 H ( z ) d z ≤ C ∫ r 0 /2 r 0 ∫ I λ ( z ) w q d ρ d z ≤ C ( ∫ r 0 /2 r 0 ∫ I λ ( z ) z 1 − 2 s w 2 q 2 d ρ d z ) 1/2 ≤ C F .
Integrating (57) in ( r 0 / 2 , r 0 ) (r_0/2,r_0) ( r 0 /2 , r 0 ) gives
( 1 − 2 Λ K ) A = A − 2 Λ τ ≤ C F + C τ F ≤ C F . \left(1-\frac{2\Lambda}{K}\right)A=A-2\Lambda\tau\le C\sqrt{\mathcal{F}}+C\sqrt{\tau\mathcal{F}}\le C\sqrt{\mathcal{F}}. ( 1 − K 2Λ ) A = A − 2Λ τ ≤ C F + C τ F ≤ C F .
Since K ≥ 8 Λ K \ge8\Lambda K ≥ 8Λ , this proves (3.30) and hence the theorem.
Reduction to bounded sets
In this section we remove the assumption Ω ⊂ B R \Omega\subset B_R Ω ⊂ B R of Theorem 3.10 by a truncation argument. As in [17 Lemma 4.1] , which follows [5 Lemma 5.3] , a nonsharp stability estimate provides the initial smallness of the mass of Ω \Omega Ω outside a ball, and an iteration on the tail volumes then reduces it to the order of the deficit. The nonsharp input is Proposition 3.2, see also [3 , 19 ] . The local L ∞ L^\infty L ∞ tail estimate for the torsion function used in [5 , 17 ] is replaced by an L 1 L^1 L 1 tail estimate.
The exact cutoff error
Lemma 4.1. Let u ∈ H s ( R n ) u \in H^s(\mathbb{R}^n) u ∈ H s ( R n ) and let χ \chi χ be bounded and Lipschitz. Then
E s ( χ u ) = E s ( u , χ 2 u ) + R χ (58) \mathcal{E}_s(\chi u)=\mathcal{E}_s(u,\chi^2u)+\mathcal{R}_\chi \tag*{(58)} E s ( χ u ) = E s ( u , χ 2 u ) + R χ (58)
where
R χ = c n , s 2 ∬ u ( x ) u ( y ) ( χ ( x ) − χ ( y ) ) 2 ∣ x − y ∣ n + 2 s d x d y . \mathcal{R}_\chi=\frac{c_{n,s}}{2}\iint\frac{u(x)u(y)(\chi(x)-\chi(y))^2}{|x-y|^{n+2s}}\,dx\,dy. R χ = 2 c n , s ∬ ∣ x − y ∣ n + 2 s u ( x ) u ( y ) ( χ ( x ) − χ ( y ) ) 2 d x d y .
If u = u Ω u=u_\Omega u = u Ω , 0 ≤ χ ≤ 1 0\le\chi\le1 0 ≤ χ ≤ 1 , and χ u ∈ H 0 s ( D ) \chi u\in H^s_0(D) χ u ∈ H 0 s ( D ) for an open D ⊂ Ω D\subset\Omega D ⊂ Ω , then
T s ( Ω ) − T s ( D ) ≤ ∫ Ω ( 1 − χ ) 2 u d x + R χ . (59) T_s(\Omega)-T_s(D)\le\int_\Omega(1-\chi)^2u\,dx+\mathcal{R}_\chi. \tag*{(59)} T s ( Ω ) − T s ( D ) ≤ ∫ Ω ( 1 − χ ) 2 u d x + R χ . (59)
Proof. For real numbers a , b , c , d a,b,c,d a , b , c , d , ( c a − d b ) 2 − ( a − b ) ( c 2 a − d 2 b ) = a b ( c − d ) 2 (ca-db)^2-(a-b)(c^2a-d^2b)=ab(c-d)^2 ( c a − d b ) 2 − ( a − b ) ( c 2 a − d 2 b ) = ab ( c − d ) 2 . Apply this pointwise to the two values of u u u and χ \chi χ and integrate. This gives (58) . Since χ \chi χ is Lipschitz, the error term R χ \mathcal{R}_\chi R χ is absolutely integrable.
If u = u Ω u=u_\Omega u = u Ω , testing (1) with χ 2 u \chi^2u χ 2 u gives E s ( u , χ 2 u ) = ∫ Ω χ 2 u \mathcal{E}_s(u,\chi^2u)=\int_\Omega\chi^2u E s ( u , χ 2 u ) = ∫ Ω χ 2 u . Insert χ u \chi u χ u into the variational formula for T s ( D ) T_s(D) T s ( D ) to obtain (59) .
Truncation
Proposition 4.2. There exist R ∗ , C , δ ∗ > 0 R_*,C,\delta_*>0 R ∗ , C , δ ∗ > 0 , depending only on n n n and s s s , such that every open set Ω ⊂ R n \Omega\subset\mathbb{R}^n Ω ⊂ R n with ∣ Ω ∣ = ω n |\Omega|=\omega_n ∣Ω∣ = ω n and δ T ( Ω ) ≤ δ ∗ \delta_T(\Omega)\le\delta_* δ T ( Ω ) ≤ δ ∗ admits an open set Ω ~ ⊂ B R ∗ \widetilde{\Omega}\subset B_{R_*} Ω ⊂ B R ∗ satisfying
∣ Ω ~ ∣ = ω n , δ T ( Ω ~ ) ≤ C δ T ( Ω ) , |\widetilde{\Omega}|=\omega_n,\qquad\delta_T(\widetilde{\Omega})\le C\delta_T(\Omega), ∣ Ω ∣ = ω n , δ T ( Ω ) ≤ C δ T ( Ω ) ,
A ( Ω ) ≤ A ( Ω ~ ) + C δ T ( Ω ) . (60) \mathcal{A}(\Omega)\le\mathcal{A}(\widetilde{\Omega})+C\delta_T(\Omega). \tag*{(60)} A ( Ω ) ≤ A ( Ω ) + C δ T ( Ω ) . (60)
Proof. Write δ = δ T ( Ω ) \delta=\delta_T(\Omega) δ = δ T ( Ω ) and u = u Ω u=u_\Omega u = u Ω . Since n ≥ 2 n\ge2 n ≥ 2 and s < 1 s<1 s < 1 , we have n > 2 s n>2s n > 2 s .
First,
∥ u ∥ L ∞ ( R n ) ≤ C . (61) \|u\|_{L^\infty(\mathbb{R}^n)}\le C. \tag*{(61)} ∥ u ∥ L ∞ ( R n ) ≤ C . (61)
To prove this, set 2 s ∗ = 2 n / ( n − 2 s ) 2_s^*=2n/(n-2s) 2 s ∗ = 2 n / ( n − 2 s ) and μ ( t ) = ∣ { u > t } ∣ \mu(t)=|\{u>t\}| μ ( t ) = ∣ { u > t } ∣ . Let f t = ( u − t ) + f_t=(u-t)_+ f t = ( u − t ) + . Since for all real a , b a,b a , b ,
( a − b ) ( ( a − t ) + − ( b − t ) + ) ≥ ( ( a − t ) + − ( b − t ) + ) 2 , (a-b)\big((a-t)_+-(b-t)_+\big)\ge\big((a-t)_+-(b-t)_+\big)^2, ( a − b ) ( ( a − t ) + − ( b − t ) + ) ≥ ( ( a − t ) + − ( b − t ) + ) 2 ,
testing (1) with f t f_t f t gives E s ( f t ) ≤ E s ( u , f t ) = ∫ Ω f t \mathcal{E}_s(f_t)\le\mathcal{E}_s(u,f_t)=\int_\Omega f_t E s ( f t ) ≤ E s ( u , f t ) = ∫ Ω f t . Combining this with fractional Sobolev and Hölder’s inequality (recall that f t f_t f t is supported in { u > t } \{u>t\} { u > t } ), we get
∥ f t ∥ 2 s ∗ 2 ≤ C E s ( f t ) ≤ C ∫ Ω f t ≤ C μ ( t ) 1 − 1 / 2 s ∗ ∥ f t ∥ 2 s ∗ . \|f_t\|_{2_s^*}^2\le C\mathcal{E}_s(f_t)\le C\int_\Omega f_t\le C\mu(t)^{1-1/2_s^*}\|f_t\|_{2_s^*}. ∥ f t ∥ 2 s ∗ 2 ≤ C E s ( f t ) ≤ C ∫ Ω f t ≤ C μ ( t ) 1 − 1/ 2 s ∗ ∥ f t ∥ 2 s ∗ .
Consequently, for t ′ > t ≥ 0 t' > t \ge0 t ′ > t ≥ 0 ,
μ ( t ′ ) ≤ ( C t ′ − t ) 2 s ∗ / s μ ( t ) 2 s ∗ / s − 1 . \mu(t') \le\left(\frac{C}{t'-t}\right)^{2s^*/s}\mu(t)^{2s^*/s-1}. μ ( t ′ ) ≤ ( t ′ − t C ) 2 s ∗ / s μ ( t ) 2 s ∗ / s − 1 .
With t j = Θ ( 1 − 2 − j ) t_j=\Theta(1-2^{-j}) t j = Θ ( 1 − 2 − j ) and Θ = K μ ( 0 ) 2 s / n \Theta=K\mu(0)^{2s/n} Θ = K μ ( 0 ) 2 s / n , induction yields μ ( t j ) ≤ 2 − n j / ( 2 s ) μ ( 0 ) \mu(t_j)\le2^{-nj/(2s)}\mu(0) μ ( t j ) ≤ 2 − nj / ( 2 s ) μ ( 0 ) if K K K is sufficiently large.
Suppose 0 ≤ χ ≤ 1 0\le\chi\le1 0 ≤ χ ≤ 1 , ∣ ∇ χ ∣ ≤ C 1 |\nabla\chi|\le C_1 ∣∇ χ ∣ ≤ C 1 , and χ \chi χ is constant on B k B_k B k , where C 1 C_1 C 1 is fixed independently of k k k . If χ ( x ) ≠ χ ( y ) \chi(x)\ne\chi(y) χ ( x ) = χ ( y ) , at least one of x , y x,y x , y lies outside B k B_k B k . Since u ≥ 0 u\ge0 u ≥ 0 , symmetry and (61) give
0 ≤ R χ ≤ C ∫ Ω ∖ B k u ( x ) ∫ R n u ( y ) min { 1 , C 1 2 ∣ x − y ∣ 2 } ∣ x − y ∣ n + 2 s d y d x ≤ C ∫ Ω ∖ B k u . (62) \begin{aligned}
0\le R_\chi\le C\int_{\Omega\setminus B_k}u(x)\int_{\mathbb{R}^n}u(y)\frac{\min\{1,C_1^2|x-y|^2\}}{|x-y|^{n+2s}}\,dy\,dx \\
&\le C\int_{\Omega\setminus B_k}u. \tag*{(62)}
\end{aligned} 0 ≤ R χ ≤ C ∫ Ω ∖ B k u ( x ) ∫ R n u ( y ) ∣ x − y ∣ n + 2 s min { 1 , C 1 2 ∣ x − y ∣ 2 } d y d x ≤ C ∫ Ω ∖ B k u . (62)
For k ≥ 1 k\ge1 k ≥ 1 , put S k = Ω ∖ B k S_k=\Omega\setminus B_k S k = Ω ∖ B k and choose a smooth cutoff χ k \chi_k χ k such that
0 ≤ χ k ≤ 1 , χ k = 0 on B k , χ k = 1 outside B k + 1 , ∣ ∇ χ k ∣ ≤ C . 0\le\chi_k\le1,\qquad\chi_k=0\text{ on }B_k,\qquad\chi_k=1\text{ outside }B_{k+1},\qquad|\nabla\chi_k|\le C. 0 ≤ χ k ≤ 1 , χ k = 0 on B k , χ k = 1 outside B k + 1 , ∣∇ χ k ∣ ≤ C .
By (58) and (62) ,
E s ( χ k u ) = ∫ Ω χ k 2 u + R χ k ≤ C ∫ S k u ≤ C ∣ S k ∣ . \mathcal{E}_s(\chi_k u)=\int_\Omega\chi_k^2u+R_{\chi_k}\le C\int_{S_k}u\le C|S_k|. E s ( χ k u ) = ∫ Ω χ k 2 u + R χ k ≤ C ∫ S k u ≤ C ∣ S k ∣.
Since χ k u = u \chi_k u=u χ k u = u on S k + 1 S_{k+1} S k + 1 , Sobolev and Hölder imply
∫ S k + 1 u ≤ ∣ S k + 1 ∣ 1 − 1 / 2 s ∗ ∥ χ k u ∥ 2 s ∗ ≤ C ∣ S k ∣ 1 / 2 + s / n E s ( χ k u ) 1 / 2 ≤ C ∣ S k ∣ 1 + s / n . (63) \begin{aligned}
\int_{S_{k+1}}u\le|S_{k+1}|^{1-1/2_s^*}\|\chi_k u\|_{2_s^*} \\
&\le C|S_k|^{1/2+s/n}\mathcal{E}_s(\chi_k u)^{1/2}\le C|S_k|^{1+s/n}. \tag*{(63)}
\end{aligned} ∫ S k + 1 u ≤ ∣ S k + 1 ∣ 1 − 1/ 2 s ∗ ∥ χ k u ∥ 2 s ∗ ≤ C ∣ S k ∣ 1/2 + s / n E s ( χ k u ) 1/2 ≤ C ∣ S k ∣ 1 + s / n . (63)
Next choose a smooth cutoff ζ k \zeta_k ζ k satisfying
0 ≤ ζ k ≤ 1 , ζ k = 1 on B k + 1 , supp ζ k ⋐ B k + 2 , ∣ ∇ ζ k ∣ ≤ C . 0\le\zeta_k\le1,\qquad\zeta_k=1\text{ on }B_{k+1},\qquad\operatorname{supp}\zeta_k\Subset B_{k+2},\qquad|\nabla\zeta_k|\le C. 0 ≤ ζ k ≤ 1 , ζ k = 1 on B k + 1 , supp ζ k ⋐ B k + 2 , ∣∇ ζ k ∣ ≤ C .
For D k = Ω ∩ B k + 2 D_k=\Omega\cap B_{k+2} D k = Ω ∩ B k + 2 , applying (59) , (62) with inner radius k + 1 k+1 k + 1 , and (63) , we obtain
0 ≤ T s ( Ω ) − T s ( D k ) ≤ ∫ Ω ( 1 − ζ k ) 2 u + R ζ k ≤ C ∫ S k + 1 u ≤ C ∣ S k ∣ 1 + s / n . (64) \begin{aligned}
0\le T_s(\Omega)-T_s(D_k)\le\int_\Omega(1-\zeta_k)^2u+R_{\zeta_k} \\
&\le C\int_{S_{k+1}}u\le C|S_k|^{1+s/n}. \tag*{(64)}
\end{aligned} 0 ≤ T s ( Ω ) − T s ( D k ) ≤ ∫ Ω ( 1 − ζ k ) 2 u + R ζ k ≤ C ∫ S k + 1 u ≤ C ∣ S k ∣ 1 + s / n . (64)
Proposition 3.2 gives A ( Ω ) ≤ C δ s / ( 2 s + 1 ) \mathcal{A}(\Omega)\le C\delta^{s/(2s+1)} A ( Ω ) ≤ C δ s / ( 2 s + 1 ) . If δ = 0 \delta=0 δ = 0 , then A ( Ω ) = 0 \mathcal{A}(\Omega)=0 A ( Ω ) = 0 and we take Ω ~ = B \widetilde{\Omega}=B Ω = B . Otherwise translate Ω \Omega Ω so that B 1 B_1 B 1 realizes its Fraenkel asymmetry. Set
a k = ∣ Ω ∖ B k ∣ ω n , σ = 1 + s n . a_k=\frac{|\Omega\setminus B_k|}{\omega_n},\qquad\sigma=1+\frac{s}{n}. a k = ω n ∣Ω ∖ B k ∣ , σ = 1 + n s .
Then
a 1 = 1 2 A ( Ω ) ≤ C δ s / ( 2 s + 1 ) . a_1=\frac{1}{2}\mathcal{A}(\Omega)\le C\delta^{s/(2s+1)}. a 1 = 2 1 A ( Ω ) ≤ C δ s / ( 2 s + 1 ) .
The fractional Saint-Venant inequality T s ( D k ) ≤ T s ( B r k ) T_s(D_k)\le T_s(B_{r_k}) T s ( D k ) ≤ T s ( B r k ) , where ∣ B r k ∣ = ∣ D k ∣ |B_{r_k}|=|D_k| ∣ B r k ∣ = ∣ D k ∣ , which follows from Lemma 2.2 and (3) , together with (64) , give
T s ( B ) − δ − C a k σ ≤ T s ( D k ) ≤ T s ( B ) ( 1 − a k + 2 ) 1 + 2 s / n . T_s(B)-\delta-Ca_k^\sigma\le T_s(D_k)\le T_s(B)(1-a_{k+2})^{1+2s/n}. T s ( B ) − δ − C a k σ ≤ T s ( D k ) ≤ T s ( B ) ( 1 − a k + 2 ) 1 + 2 s / n .
Since 1 − ( 1 − a ) 1 + 2 s / n ≥ a 1-(1-a)^{1+2s/n}\ge a 1 − ( 1 − a ) 1 + 2 s / n ≥ a for a ∈ [ 0 , 1 ] a\in[0,1] a ∈ [ 0 , 1 ] , it follows that
a k + 2 ≤ C ( δ + a k σ ) . a_{k+2}\le C(\delta+a_k^\sigma). a k + 2 ≤ C ( δ + a k σ ) .
For 0 < δ ≤ 1 0<\delta\le1 0 < δ ≤ 1 , induction gives
a 2 j + 1 ≤ C j δ min { s σ j / ( 2 s + 1 ) , 1 } . a_{2j+1}\le C_j\delta^{\min\{s\sigma^j/(2s+1),1\}}. a 2 j + 1 ≤ C j δ m i n { s σ j / ( 2 s + 1 ) , 1 } .
Choose N N N with s σ N / ( 2 s + 1 ) ≥ 1 s\sigma^N/(2s+1)\ge1 s σ N / ( 2 s + 1 ) ≥ 1 and put k ∗ = 2 N + 1 k_*=2N+1 k ∗ = 2 N + 1 . Then
a k ∗ ≤ C N δ . (65) a_{k_*}\le C_N\delta. \tag*{(65)} a k ∗ ≤ C N δ . (65)
Let
D = Ω ∩ B k ∗ + 2 , a = ∣ Ω ∖ D ∣ ω n , r = ( 1 − a ) 1 / n . D=\Omega\cap B_{k_*+2},\qquad a=\frac{|\Omega\setminus D|}{\omega_n},\qquad r=(1-a)^{1/n}. D = Ω ∩ B k ∗ + 2 , a = ω n ∣Ω ∖ D ∣ , r = ( 1 − a ) 1/ n .
Choose δ ∗ ≤ min { 1 , ( 2 C N ) − 1 } \delta_* \le\min\{1,(2C_N)^{-1}\} δ ∗ ≤ min { 1 , ( 2 C N ) − 1 } . Then a ≤ a k ∗ ≤ 1 / 2 a \le a_{k_*} \le1/2 a ≤ a k ∗ ≤ 1/2 , and Ω ~ = r − 1 D \widetilde{\Omega}=r^{-1}D Ω = r − 1 D satisfies
∣ Ω ~ ∣ = ω n , Ω ~ ⊂ B R ∗ , R ∗ = 2 1 / n ( k ∗ + 2 ) . |\widetilde{\Omega}|=\omega_n,\qquad\widetilde{\Omega}\subset B_{R_*},\qquad R_*=2^{1/n}(k_*+2). ∣ Ω ∣ = ω n , Ω ⊂ B R ∗ , R ∗ = 2 1/ n ( k ∗ + 2 ) .
By (64) and (65) ,
T s ( Ω ~ ) − T s ( D ) ≤ C a k ∗ σ ≤ C δ . T_s(\widetilde{\Omega})-T_s(D)\le Ca_{k_*}^{\sigma}\le C\delta. T s ( Ω ) − T s ( D ) ≤ C a k ∗ σ ≤ C δ .
Since T s ( Ω ~ ) = r − ( n + 2 s ) T s ( D ) ≥ T s ( D ) T_s(\widetilde{\Omega})=r^{-(n+2s)}T_s(D)\ge T_s(D) T s ( Ω ) = r − ( n + 2 s ) T s ( D ) ≥ T s ( D ) , this yields
δ T ( Ω ~ ) ≤ δ + T s ( Ω ~ ) − T s ( D ) ≤ C δ . \delta_T(\widetilde{\Omega})\le\delta+T_s(\widetilde{\Omega})-T_s(D)\le C\delta. δ T ( Ω ) ≤ δ + T s ( Ω ) − T s ( D ) ≤ C δ .
If B 1 ( b ) B_1(b) B 1 ( b ) realizes the asymmetry of Ω ~ \widetilde{\Omega} Ω , comparison with B 1 ( r b ) B_1(rb) B 1 ( r b ) gives
ω n A ( Ω ) ≤ ∣ Ω ∖ D ∣ + ∣ D △ B r ( r b ) ∣ + ∣ B r ( r b ) △ B 1 ( r b ) ∣ = 2 ω n a + r n ω n A ( Ω ~ ) . \begin{aligned}
\omega_n\mathcal{A}(\Omega)\le|\Omega\setminus D|+|D\mathbin{\triangle}B_r(rb)|+|B_r(rb)\mathbin{\triangle}B_1(rb)| \\
&=2\omega_n a+r^n\omega_n\mathcal{A}(\widetilde{\Omega}).
\end{aligned} ω n A ( Ω ) ≤ ∣Ω ∖ D ∣ + ∣ D △ B r ( r b ) ∣ + ∣ B r ( r b ) △ B 1 ( r b ) ∣ = 2 ω n a + r n ω n A ( Ω ) .
The last assertion of (60) follows from a ≤ C N δ a\le C_N\delta a ≤ C N δ and r ≤ 1 r\le1 r ≤ 1 .
Proof of Theorem 1.2. If δ T ( Ω ) ≤ δ ∗ \delta_T(\Omega)\le\delta_* δ T ( Ω ) ≤ δ ∗ , apply Proposition 4.2 and then Theorem 3.10 to Ω ~ \widetilde{\Omega} Ω to obtain
A ( Ω ) 2 ≤ C ( A ( Ω ~ ) 2 + δ T ( Ω ) 2 ) ≤ C δ T ( Ω ) . \mathcal{A}(\Omega)^2\le C\bigl(\mathcal{A}(\widetilde{\Omega})^2+\delta_T(\Omega)^2\bigr)\le C\delta_T(\Omega). A ( Ω ) 2 ≤ C ( A ( Ω ) 2 + δ T ( Ω ) 2 ) ≤ C δ T ( Ω ) .
If δ T ( Ω ) > δ ∗ \delta_T(\Omega)>\delta_* δ T ( Ω ) > δ ∗ , the conclusion follows from A ( Ω ) ≤ 2 \mathcal{A}(\Omega)\le2 A ( Ω ) ≤ 2 .
Optimality of the exponent
We now show that the exponent 2 2 2 in Theorems 1.1 and 1.2 is indeed optimal. As in [5 ] it can be shown by estimating the deficits on ellipsoids.
Lemma 5.1. Let
L ε = diag ( e ε , e − ε , 1 , … , 1 ) , Ω ε = L ε B . L_\varepsilon=\operatorname{diag}(e^\varepsilon,e^{-\varepsilon},1,\ldots,1),\qquad\Omega_\varepsilon=L_\varepsilon B. L ε = diag ( e ε , e − ε , 1 , … , 1 ) , Ω ε = L ε B . For every fixed n ≥ 2 n\ge2 n ≥ 2 and s ∈ ( 0 , 1 ) s\in(0,1) s ∈ ( 0 , 1 ) , these sets satisfy ∣ Ω ε ∣ = ω n |\Omega_\varepsilon|=\omega_n ∣ Ω ε ∣ = ω n and, for all sufficiently small ∣ ε ∣ |\varepsilon| ∣ ε ∣ ,
c ∣ ε ∣ ≤ A ( Ω ε ) ≤ C ∣ ε ∣ , c|\varepsilon|\le\mathcal{A}(\Omega_\varepsilon)\le C|\varepsilon|, c ∣ ε ∣ ≤ A ( Ω ε ) ≤ C ∣ ε ∣ , 0 ≤ λ s ( Ω ε ) λ s ( B ) − 1 ≤ M s ( ε ) − 1 , 0 ≤ T s ( B ) − T s ( Ω ε ) T s ( B ) ≤ 1 − 1 M s ( ε ) , 0\le\frac{\lambda_s(\Omega_\varepsilon)}{\lambda_s(B)}-1\le M_s(\varepsilon)-1,\qquad0\le\frac{T_s(B)-T_s(\Omega_\varepsilon)}{T_s(B)}\le1-\frac{1}{M_s(\varepsilon)}, 0 ≤ λ s ( B ) λ s ( Ω ε ) − 1 ≤ M s ( ε ) − 1 , 0 ≤ T s ( B ) T s ( B ) − T s ( Ω ε ) ≤ 1 − M s ( ε ) 1 , where
M s ( ε ) = 1 ∣ S n − 1 ∣ ∫ S n − 1 ∣ L ε − T θ ∣ 2 s d H n − 1 ( θ ) M_s(\varepsilon)=\frac{1}{|\mathbb{S}^{n-1}|}\int_{\mathbb{S}^{n-1}}|L_\varepsilon^{-T}\theta|^{2s}\,d\mathcal{H}^{n-1}(\theta) M s ( ε ) = ∣ S n − 1 ∣ 1 ∫ S n − 1 ∣ L ε − T θ ∣ 2 s d H n − 1 ( θ ) = 1 + 4 s ( n + 2 s ) n ( n + 2 ) ε 2 + O ( ε 4 ) . (66) =1+\frac{4s(n+2s)}{n(n+2)}\varepsilon^2+O(\varepsilon^4). \tag*{(66)} = 1 + n ( n + 2 ) 4 s ( n + 2 s ) ε 2 + O ( ε 4 ) . (66) The constants and the remainder bound depend only on n , s n,s n , s .
Proof. The volume assertion follows from det L ε = 1 \det L_\varepsilon=1 det L ε = 1 . For any radial f ∈ H 0 s ( B ) f\in H^s_0(B) f ∈ H 0 s ( B ) , put f ε ( x ) = f ( L ε − 1 x ) f_\varepsilon(x)=f(L_\varepsilon^{-1}x) f ε ( x ) = f ( L ε − 1 x ) . Linear changes of variables in smooth compactly supported approximations show that f ε ∈ H 0 s ( Ω ε ) f_\varepsilon\in H^s_0(\Omega_\varepsilon) f ε ∈ H 0 s ( Ω ε ) . Its integral and L 2 L^2 L 2 norm agree with those of f f f . Moreover, f ε ^ ( ξ ) = f ^ ( L ε T ξ ) \widehat{f_\varepsilon}(\xi)=\widehat f(L_\varepsilon^T\xi) f ε ( ξ ) = f ( L ε T ξ ) , so the Fourier representation of the quadratic form and polar integration give
E s ( f ε ) = 1 ( 2 π ) n ∫ R n ∣ L ε − T ξ ∣ 2 s ∣ f ^ ( ξ ) ∣ 2 d ξ = M s ( ε ) E s ( f ) . (67) \begin{aligned}
\mathcal{E}_s(f_\varepsilon)=\frac{1}{(2\pi)^n}\int_{\mathbb{R}^n}|L_\varepsilon^{-T}\xi|^{2s}|\widehat f(\xi)|^2\,d\xi\\
&=M_s(\varepsilon)\mathcal{E}_s(f). \tag*{(67)}
\end{aligned} E s ( f ε ) = ( 2 π ) n 1 ∫ R n ∣ L ε − T ξ ∣ 2 s ∣ f ( ξ ) ∣ 2 d ξ = M s ( ε ) E s ( f ) . (67)
Here f ^ \widehat f f is radial. A nonnegative radial first eigenfunction ϕ \phi ϕ of B B B can be chosen by Schwarz rearrangement of a minimizer of the Rayleigh quotient. Applying (67) to ϕ \phi ϕ gives λ s ( Ω ε ) ≤ M s ( ε ) λ s ( B ) \lambda_s(\Omega_\varepsilon)\le M_s(\varepsilon)\lambda_s(B) λ s ( Ω ε ) ≤ M s ( ε ) λ s ( B ) . For torsion, take f = u B f=u_B f = u B and use the trial functions a f ε af_\varepsilon a f ε , a ∈ R a\in\mathbb{R} a ∈ R , in (2) . Since
∫ f ε = E s ( u B ) = T s ( B ) , \int f_\varepsilon=\mathcal{E}_s(u_B)=T_s(B), ∫ f ε = E s ( u B ) = T s ( B ) ,
T s ( Ω ε ) ≥ max a ∈ R ( 2 a − a 2 M s ( ε ) ) T s ( B ) = T s ( B ) M s ( ε ) . T_s(\Omega_\varepsilon)\ge\max_{a\in\mathbb{R}}\left(2a-a^2M_s(\varepsilon)\right)T_s(B)=\frac{T_s(B)}{M_s(\varepsilon)}. T s ( Ω ε ) ≥ a ∈ R max ( 2 a − a 2 M s ( ε ) ) T s ( B ) = M s ( ε ) T s ( B ) .
The qualitative Faber–Krahn and Saint-Venant inequalities give the nonnegativity of both deficits, and also M s ( ε ) ≥ 1 M_s(\varepsilon)\ge1 M s ( ε ) ≥ 1 . To compute the angular factor, write a ( θ ) = θ 1 2 − θ 2 2 a(\theta)=\theta_1^2-\theta_2^2 a ( θ ) = θ 1 2 − θ 2 2 and b ( θ ) = θ 1 2 + θ 2 2 b(\theta)=\theta_1^2+\theta_2^2 b ( θ ) = θ 1 2 + θ 2 2 . Uniformly on the sphere,
∣ L ε − T θ ∣ 2 s = 1 − 2 s ε a ( θ ) + ε 2 ( 2 s b ( θ ) + 2 s ( s − 1 ) a ( θ ) 2 ) + O ( ∣ ε ∣ 3 ) . \lvert L_\varepsilon^{-T}\theta\rvert^{2s}=1-2s\varepsilon a(\theta)+\varepsilon^2\left(2s b(\theta)+2s(s-1)a(\theta)^2\right)+O(\lvert\varepsilon\rvert^3). ∣ L ε − T θ ∣ 2 s = 1 − 2 s ε a ( θ ) + ε 2 ( 2 s b ( θ ) + 2 s ( s − 1 ) a ( θ ) 2 ) + O (∣ ε ∣ 3 ) .
The spherical averages of a a a , b b b , and a 2 a^2 a 2 are respectively 0 0 0 , 2 / n 2/n 2/ n , and 4 / [ n ( n + 2 ) ] 4/[n(n+2)] 4/ [ n ( n + 2 )] . This gives the quadratic coefficient in (66) . Interchanging the first two coordinates shows that M s M_s M s is even; its smoothness near zero and Taylor’s theorem improve the remainder to O ( ε 4 ) O(\varepsilon^4) O ( ε 4 ) .
It remains to estimate the asymmetry, including the infimum over translations. The radial function of Ω ε \Omega_\varepsilon Ω ε is
r ε ( θ ) = ∣ L ε − 1 θ ∣ − 1 = 1 + ε ( θ 1 2 − θ 2 2 ) + O ( ε 2 ) , r_\varepsilon(\theta)=\lvert L_\varepsilon^{-1}\theta\rvert^{-1}=1+\varepsilon(\theta_1^2-\theta_2^2)+O(\varepsilon^2), r ε ( θ ) = ∣ L ε − 1 θ ∣ − 1 = 1 + ε ( θ 1 2 − θ 2 2 ) + O ( ε 2 ) ,
with uniform remainder. Consequently,
∣ Ω ε △ B ∣ = 1 n ∫ S n − 1 ∣ r ε ( θ ) n − 1 ∣ d H n − 1 ( θ ) = ∣ ε ∣ ∫ S n − 1 ∣ θ 1 2 − θ 2 2 ∣ d H n − 1 ( θ ) + O ( ε 2 ) . \begin{aligned}
\lvert\Omega_\varepsilon\mathbin{\triangle} B\rvert=\frac{1}{n}\int_{\mathbb{S}^{n-1}}\lvert r_\varepsilon(\theta)^n-1\rvert\,d\mathcal{H}^{n-1}(\theta) \\
&=\lvert\varepsilon\rvert\int_{\mathbb{S}^{n-1}}\lvert\theta_1^2-\theta_2^2\rvert\,d\mathcal{H}^{n-1}(\theta)+O(\varepsilon^2).
\end{aligned} ∣ Ω ε △ B ∣ = n 1 ∫ S n − 1 ∣ r ε ( θ ) n − 1 ∣ d H n − 1 ( θ ) = ∣ ε ∣ ∫ S n − 1 ∣ θ 1 2 − θ 2 2 ∣ d H n − 1 ( θ ) + O ( ε 2 ) .
The coefficient is positive because n ≥ 2 n\ge2 n ≥ 2 . Since Ω ε = − Ω ε \Omega_\varepsilon=-\Omega_\varepsilon Ω ε = − Ω ε , for every center a ∈ R n a\in\mathbb{R}^n a ∈ R n ,
∣ B 1 ( a ) △ B 1 ( − a ) ∣ ≤ 2 ∣ Ω ε △ B 1 ( a ) ∣ . \lvert B_1(a)\mathbin{\triangle}B_1(-a)\rvert\le2\lvert\Omega_\varepsilon\mathbin{\triangle}B_1(a)\rvert. ∣ B 1 ( a ) △ B 1 ( − a )∣ ≤ 2 ∣ Ω ε △ B 1 ( a )∣ .
The symmetric difference of two unit balls is nondecreasing in the distance between their centers, as follows by slicing parallel to that displacement. Thus
∣ Ω ε △ B ∣ ≤ ∣ Ω ε △ B 1 ( a ) ∣ + ∣ B 1 ( a ) △ B ∣ \lvert\Omega_\varepsilon\mathbin{\triangle}B\rvert\le\lvert\Omega_\varepsilon\mathbin{\triangle}B_1(a)\rvert+\lvert B_1(a)\mathbin{\triangle}B\rvert ∣ Ω ε △ B ∣ ≤ ∣ Ω ε △ B 1 ( a )∣ + ∣ B 1 ( a ) △ B ∣
≤ 3 ∣ Ω ε △ B 1 ( a ) ∣ . \le3\lvert\Omega_\varepsilon\mathbin{\triangle}B_1(a)\rvert. ≤ 3 ∣ Ω ε △ B 1 ( a )∣ .
Taking the infimum over a a a proves
∣ Ω ε △ B ∣ 3 ω n ≤ A ( Ω ε ) ≤ ∣ Ω ε △ B ∣ ω n . \frac{\lvert\Omega_\varepsilon\mathbin{\triangle}B\rvert}{3\omega_n}\le\mathcal{A}(\Omega_\varepsilon)\le\frac{\lvert\Omega_\varepsilon\mathbin{\triangle}B\rvert}{\omega_n}. 3 ω n ∣ Ω ε △ B ∣ ≤ A ( Ω ε ) ≤ ω n ∣ Ω ε △ B ∣ .
which completes the proof. □
Appendix A. Ball speed and boundary control of the radial profile
Estimates for the ball speed
Lemma A.1. Let U B = Ext s u B U_B=\operatorname{Ext}_s u_B U B = Ext s u B , with w B = − ∂ ρ t B w_B=-\partial_\rho t_B w B = − ∂ ρ t B . There are ℓ ∈ ( 0 , 1 / 16 ) \ell\in(0,1/16) ℓ ∈ ( 0 , 1/16 ) and constants c 0 , C 0 > 0 c_0,C_0>0 c 0 , C 0 > 0 , depending only on n , s n,s n , s , such that the ball speed is positive on R \mathbb{R} R and
c 0 ( 1 − ρ + z ) s − 1 ≤ w B ( ρ , z ) ≤ C 0 ( 1 − ρ + z ) s − 1 , c_0(1-\rho+z)^{s-1}\le w_B(\rho,z)\le C_0(1-\rho+z)^{s-1}, c 0 ( 1 − ρ + z ) s − 1 ≤ w B ( ρ , z ) ≤ C 0 ( 1 − ρ + z ) s − 1 , ∣ ∇ w B ( ρ , z ) ∣ ≤ C 0 w B ( ρ , z ) 1 − ρ + z . (68) \lvert\nabla w_B(\rho,z)\rvert\le C_0\frac{w_B(\rho,z)}{1-\rho+z}. \tag*{(68)} ∣ ∇ w B ( ρ , z )∣ ≤ C 0 1 − ρ + z w B ( ρ , z ) . (68) Moreover, w B w_B w B is smooth up to every compact subinterval of the bottom edge and ( w B ) z ( ρ , 0 ) = 0 (w_B)_z(\rho,0)=0 ( w B ) z ( ρ , 0 ) = 0 for 0 < ρ < 1 0<\rho<1 0 < ρ < 1 .
Proof. The explicit ball formula [10 Theorem 1] , with exponent s s s and fractional order 2 s 2s 2 s , gives in our normalization
u B ( x ) = κ n , s ( 1 − ∣ x ∣ 2 ) + s , κ n , s = Γ ( n / 2 ) 2 2 s Γ ( 1 + s ) Γ ( n / 2 + s ) . (69) u_B(x)=\kappa_{n,s}(1-\lvert x\rvert^2)_+^s,\qquad\kappa_{n,s}=\frac{\Gamma(n/2)}{2^{2s}\Gamma(1+s)\Gamma(n/2+s)}. \tag*{(69)} u B ( x ) = κ n , s ( 1 − ∣ x ∣ 2 ) + s , κ n , s = 2 2 s Γ ( 1 + s ) Γ ( n /2 + s ) Γ ( n /2 ) . (69)
The operator in [10 ] has the opposite sign; reversing it gives the positive multiplier ∣ ξ ∣ 2 s \lvert\xi\rvert^{2s} ∣ ξ ∣ 2 s . In particular,
w B ( ρ , 0 ) = 2 s κ n , s ρ ( 1 − ρ 2 ) s − 1 ≍ ( 1 − ρ ) s − 1 ( 1 / 2 < ρ < 1 ) . (70) w_B(\rho,0)=2s\kappa_{n,s}\rho(1-\rho^2)^{s-1}\asymp(1-\rho)^{s-1}\qquad(1/2<\rho<1). \tag*{(70)} w B ( ρ , 0 ) = 2 s κ n , s ρ ( 1 − ρ 2 ) s − 1 ≍ ( 1 − ρ ) s − 1 ( 1/2 < ρ < 1 ) . (70)
We justify below that this radial derivative of the trace is also the bottom trace of w B w_B w B .
Weighted local estimates and even reflection. We use the local boundedness, Hölder continuity, and Harnack estimates for weak solutions of
div ( ∣ z ∣ 1 − 2 s ∇ f ) = 0 \operatorname{div}(\lvert z\rvert^{1-2s}\nabla f)=0 div (∣ z ∣ 1 − 2 s ∇ f ) = 0
in [11 Section 2.3] . The weight ∣ z ∣ 1 − 2 s \lvert z\rvert^{1-2s} ∣ z ∣ 1 − 2 s belongs to A 2 A_2 A 2 precisely because − 1 < 1 − 2 s < 1 -1<1-2s<1 − 1 < 1 − 2 s < 1 . The constants depend on the dimension and s s s , and the estimates are invariant under simultaneous dilation of the spatial variables and height. Harnack applies to nonnegative solutions, including balls that meet z = 0 z=0 z = 0 .
We also need tangential derivative estimates, which follow from these inputs as follows. The coefficient is independent of the horizontal variables, so horizontal difference quotients solve the same equation. The weighted Caccioppoli inequality, iterated on nested balls, bounds the weighted L 2 L^2 L 2 norms of all horizontal derivatives. Local boundedness applied to those derivatives then gives, on a smaller ball of radius comparable to r r r ,
sup ∣ D y m f ∣ ≤ C m r − m sup ∣ f ∣ ( m ≥ 1 ) , (71) \sup|D_y^m f| \leq C_m r^{-m}\sup|f| \qquad(m \geq1), \tag*{(71)} sup ∣ D y m f ∣ ≤ C m r − m sup ∣ f ∣ ( m ≥ 1 ) , (71)
where the supremum on the right is over a fixed larger ball and y y y denotes the horizontal variables in the dimension being considered. For signed derivatives one applies local boundedness to their positive and negative parts. Hölder continuity gives continuous representatives of all these horizontal derivatives across z = 0 z=0 z = 0 .
If f f f is even in z z z and is a weak solution across z = 0 z=0 z = 0 , its weighted normal flux has zero trace. Indeed, ∂ z ( ∣ z ∣ 1 − 2 s f z ) = − ∣ z ∣ 1 − 2 s Δ Y f \partial_z(|z|^{1-2s}f_z)=-|z|^{1-2s}\Delta_Yf ∂ z ( ∣ z ∣ 1 − 2 s f z ) = − ∣ z ∣ 1 − 2 s Δ Y f has an integrable right-hand side locally, by the preceding tangential regularity. The flux is continuous across zero and odd, so its trace there is zero. Integration in z > 0 z>0 z > 0 therefore gives the exact identity
f z ( y , z ) = − z 2 s − 1 ∫ 0 z t 1 − 2 s Δ Y f ( y , t ) d t = − z 2 s − 1 ∫ 0 1 t 1 − 2 s Δ Y f ( y , t z ) d t . (72) f_z(y,z)=-z^{2s-1}\int_0^z t^{1-2s}\Delta_Yf(y,t)\,\mathrm{d}t=-z^{2s-1}\int_0^1 t^{1-2s}\Delta_Yf(y,tz)\,\mathrm{d}t. \tag*{(72)} f z ( y , z ) = − z 2 s − 1 ∫ 0 z t 1 − 2 s Δ Y f ( y , t ) d t = − z 2 s − 1 ∫ 0 1 t 1 − 2 s Δ Y f ( y , t z ) d t . (72)
In particular f z ( y , 0 ) = 0 f_z(y,0)=0 f z ( y , 0 ) = 0 , and
∣ f z ( y , z ) ∣ ≤ z 2 − 2 s sup 0 ≤ t ≤ z ∣ Δ Y f ( y , t ) ∣ . (73) |f_z(y,z)|\leq\frac{z}{2-2s}\sup_{0\leq t\leq z}|\Delta_Yf(y,t)|. \tag*{(73)} ∣ f z ( y , z ) ∣ ≤ 2 − 2 s z 0 ≤ t ≤ z sup ∣ Δ Y f ( y , t ) ∣. (73)
The last expression in (A.5), first for all horizontal derivatives and then differentiated repeatedly in z z z , proves smoothness across the bottom. More precisely, continuity of all horizontal derivatives starts the induction on the number of normal derivatives; at each step differentiation under the integral uses an integrable power t 1 − 2 s + j t^{1-2s+j} t 1 − 2 s + j , j ≥ 0 j\geq0 j ≥ 0 . The even extension satisfies the same formula for negative z z z .
The ball equation and its zero-flux derivative. On each bottom patch compactly contained in B B B , the weak torsion condition is
− d s lim z ↓ 0 z 1 − 2 s ∂ z U B = 1. -d_s\lim_{z\downarrow0}z^{1-2s}\partial_zU_B=1. − d s z ↓ 0 lim z 1 − 2 s ∂ z U B = 1.
Since ∂ z ( z 2 s / ( 2 s d s ) ) = z 2 s − 1 / d s \partial_z(z^{2s}/(2s d_s))=z^{2s-1}/d_s ∂ z ( z 2 s / ( 2 s d s )) = z 2 s − 1 / d s , the function U B + z 2 s / ( 2 s d s ) U_B+z^{2s}/(2s d_s) U B + z 2 s / ( 2 s d s ) has zero weak weighted normal flux on this patch. Its even reflection is a weak solution of div ( ∣ z ∣ 1 − 2 s ∇ f ) = 0 \operatorname{div}(|z|^{1-2s}\nabla f)=0 div ( ∣ z ∣ 1 − 2 s ∇ f ) = 0 across the patch, by [6 Lemma 4.1] , or directly by testing on the two half-balls. The added term has finite local weighted energy because ∫ 0 r z 2 s − 1 d z < ∞ \int_0^r z^{2s-1}\,\mathrm{d}z<\infty ∫ 0 r z 2 s − 1 d z < ∞ . The preceding local argument shows that this corrected reflection is smooth. Its radial derivative is − w B -w_B − w B , since the correction is independent of the horizontal variables. Thus w B w_B w B has a smooth even reflection, its bottom value is (A.3), and ( w B ) z ( ρ , 0 ) = 0 (w_B)_z(\rho,0)=0 ( w B ) z ( ρ , 0 ) = 0 .
The convolution of two nonnegative radial nondecreasing functions is radial nondecreasing, so w B ≥ 0 w_B\geq0 w B ≥ 0 at positive height. Differentiating the weighted radial equation for t B t_B t B gives
( w B ) ρ ρ + n − 1 ρ ( w B ) ρ + ( w B ) z z + 1 − 2 s z ( w B ) z − n − 1 ρ 2 w B = 0. (w_B)_{\rho\rho}+\frac{n-1}{\rho}(w_B)_\rho+(w_B)_{zz}+\frac{1-2s}{z}(w_B)_z-\frac{n-1}{\rho^2}w_B=0. ( w B ) ρρ + ρ n − 1 ( w B ) ρ + ( w B ) z z + z 1 − 2 s ( w B ) z − ρ 2 n − 1 w B = 0.
Equivalently, k ( ρ , z ) = w B ( ρ , z ) / ρ k(\rho,z)=w_B(\rho,z)/\rho k ( ρ , z ) = w B ( ρ , z ) / ρ satisfies
k ρ ρ + n + 1 ρ k ρ + k z z + 1 − 2 s z k z = 0. (74) k_{\rho\rho}+\frac{n+1}{\rho}k_\rho+k_{zz}+\frac{1-2s}{z}k_z=0. \tag*{(74)} k ρρ + ρ n + 1 k ρ + k z z + z 1 − 2 s k z = 0. (74)
Thus k ( ∣ y ∣ , z ) k(|y|,z) k ( ∣ y ∣ , z ) , for y ∈ R n + 2 y\in\mathbb{R}^{n+2} y ∈ R n + 2 , solves div ( z 1 − 2 s ∇ k ) = 0 \operatorname{div}(z^{1-2s}\nabla k)=0 div ( z 1 − 2 s ∇ k ) = 0 away from ∣ y ∣ = 0 |y|=0 ∣ y ∣ = 0 . It is nonnegative and not identically zero, hence positive at positive height by the strong maximum principle. On bottom annuli with 0 < ∣ y ∣ < 1 0<|y|<1 0 < ∣ y ∣ < 1 , its smooth even reflection solves the corresponding equation with weight ∣ z ∣ 1 − 2 s |z|^{1-2s} ∣ z ∣ 1 − 2 s ; its normal derivative is O ( z ) O(z) O ( z ) , so the weighted flux is O ( z 2 − 2 s ) → 0 O(z^{2-2s})\to0 O ( z 2 − 2 s ) → 0 and there is no interface source. All subsequent estimates stay away from ∣ y ∣ = 0 |y|=0 ∣ y ∣ = 0 .
The regime 0 < z ≤ 1 − ρ 0<z\leq1-\rho 0 < z ≤ 1 − ρ . Write d = 1 − ρ d=1-\rho d = 1 − ρ and y 0 = ρ e 1 ∈ R n + 2 y_0=\rho e_1\in\mathbb{R}^{n+2} y 0 = ρ e 1 ∈ R n + 2 . For d d d small, the cylinder ∣ y − y 0 ∣ < d / 2 |y-y_0|<d/2 ∣ y − y 0 ∣ < d /2 , ∣ z ∣ < 2 d |z|<2d ∣ z ∣ < 2 d lies in the reflected region and has ∣ y ∣ > 1 / 2 |y|>1/2 ∣ y ∣ > 1/2 . A bounded Harnack chain of balls with radius comparable to d d d joins ( y 0 , 0 ) (y_0,0) ( y 0 , 0 ) to ( y 0 , z ) (y_0,z) ( y 0 , z ) and covers a smaller neighborhood of the segment between them. The weighted Harnack inequality and (A.3) imply
k ( ρ , z ) ≍ k ( ρ , 0 ) ≍ d s − 1 ( 0 < z ≤ d ) . (75) k(\rho,z) \asymp k(\rho,0) \asymp d^{s-1} \qquad(0<z\le d). \tag*{(75)} k ( ρ , z ) ≍ k ( ρ , 0 ) ≍ d s − 1 ( 0 < z ≤ d ) . (75)
The number of balls and their relative radii are independent of d , z d,z d , z . Applying (A.4) for m = 1 , 2 m=1,2 m = 1 , 2 on these balls and using Harnack to compare their suprema with k ( ρ , z ) k(\rho,z) k ( ρ , z ) gives
∣ k ρ ( ρ , z ) ∣ ≤ C d − 1 k ( ρ , z ) , sup 0 ≤ t ≤ z ∣ Δ Y k ( y 0 , t ) ∣ ≤ C d − 2 k ( ρ , z ) . |k_\rho(\rho,z)|\le Cd^{-1}k(\rho,z),\qquad\sup_{0\le t\le z}|\Delta_Y k(y_0,t)|\le Cd^{-2}k(\rho,z). ∣ k ρ ( ρ , z ) ∣ ≤ C d − 1 k ( ρ , z ) , 0 ≤ t ≤ z sup ∣ Δ Y k ( y 0 , t ) ∣ ≤ C d − 2 k ( ρ , z ) .
Consequently (A.6) gives
∣ k z ( ρ , z ) ∣ ≤ C z d − 2 k ( ρ , z ) ≤ C d − 1 k ( ρ , z ) . |k_z(\rho,z)|\le Czd^{-2}k(\rho,z)\le Cd^{-1}k(\rho,z). ∣ k z ( ρ , z ) ∣ ≤ C z d − 2 k ( ρ , z ) ≤ C d − 1 k ( ρ , z ) .
Since w B = ρ k w_B=\rho k w B = ρ k , ρ ≍ 1 \rho\asymp1 ρ ≍ 1 , and d + z ≍ d d+z\asymp d d + z ≍ d , this proves (A.1) in this regime, after choosing ℓ \ell ℓ small.
The regime z > 1 − ρ z>1-\rho z > 1 − ρ . Join ( ρ e 1 , z ) (\rho e_1,z) ( ρ e 1 , z ) to ( ( 1 − z ) e 1 , z ) ((1-z)e_1,z) (( 1 − z ) e 1 , z ) by a bounded Harnack chain of balls of radius comparable to z z z . These balls remain in the upper half-space and have radial coordinates greater than 1 / 2 1/2 1/2 . At the second endpoint (A.8) applies, with bottom distance z z z , so k ( ρ , z ) ≍ z s − 1 k(\rho,z)\asymp z^{s-1} k ( ρ , z ) ≍ z s − 1 . On an interior ball of radius z / 8 z/8 z /8 about the first endpoint, scaling by z z z makes the coefficient in (A.7) smooth and uniformly elliptic, with bounds depending only on n , s n,s n , s . The interior gradient estimate and Harnack give ∣ ∇ k ( ρ , z ) ∣ ≤ C k ( ρ , z ) / z |\nabla k(\rho,z)|\le Ck(\rho,z)/z ∣∇ k ( ρ , z ) ∣ ≤ C k ( ρ , z ) / z there. Multiplying by ρ ≍ 1 \rho\asymp1 ρ ≍ 1 and using 1 − ρ + z ≍ z 1-\rho+z\asymp z 1 − ρ + z ≍ z proves the remaining estimates in (A.1). □
Boundary control of the radial profile
Lemma A.2. For every f ∈ H s ( R n ) f\in H^s(\mathbb{R}^n) f ∈ H s ( R n ) vanishing a.e. outside B B B ,
∫ B ∣ f ( x ) ∣ 2 ( 1 − ∣ x ∣ ) 2 s d x ≤ C E s ( f ) . (76) \int_B\frac{|f(x)|^2}{(1-|x|)^{2s}}\,dx\le C\mathcal{E}_s(f). \tag*{(76)} ∫ B ( 1 − ∣ x ∣ ) 2 s ∣ f ( x ) ∣ 2 d x ≤ C E s ( f ) . (76)
Let now Ω ⊂ R n \Omega\subset\mathbb{R}^n Ω ⊂ R n be bounded and open, with ∣ Ω ∣ = ω n |\Omega|=\omega_n ∣Ω∣ = ω n , let u = u Ω u=u_\Omega u = u Ω , let 0 < ℓ < 1 / 4 0<\ell<1/4 0 < ℓ < 1/4 , and consider the profile error v = t − t B v=t-t_B v = t − t B of (7) , with v ( ⋅ , 0 ) v(\cdot,0) v ( ⋅ , 0 ) defined as the radial profile of u ∗ − u B u^*-u_B u ∗ − u B . Then
∫ R z 1 − 2 s ∣ ∇ v ∣ 2 d ρ d z + ∫ 1 − ℓ 1 ∣ v ( ρ , 0 ) ∣ 2 ( 1 − ρ ) 2 s d ρ ≤ C ( ℓ ) δ T ( Ω ) . (77) \int_{\mathbb{R}}z^{1-2s}|\nabla v|^2\,d\rho\,dz+\int_{1-\ell}^{1}\frac{|v(\rho,0)|^2}{(1-\rho)^{2s}}\,d\rho\le C(\ell)\delta_T(\Omega). \tag*{(77)} ∫ R z 1 − 2 s ∣∇ v ∣ 2 d ρ d z + ∫ 1 − ℓ 1 ( 1 − ρ ) 2 s ∣ v ( ρ , 0 ) ∣ 2 d ρ ≤ C ( ℓ ) δ T ( Ω ) . (77)
Moreover v ( ⋅ , z ) → v ( ⋅ , 0 ) v(\cdot,z)\to v(\cdot,0) v ( ⋅ , z ) → v ( ⋅ , 0 ) in L 2 ( ( 1 − ℓ , 1 ) ) L^2((1-\ell,1)) L 2 (( 1 − ℓ , 1 )) as z ↓ 0 z\downarrow0 z ↓ 0 .
Proof. The exterior Hardy inequality. Since f f f vanishes a.e. outside B B B , the integrand of E s ( f ) \mathcal{E}_s(f) E s ( f ) equals ∣ f ( x ) ∣ 2 ∣ x − y ∣ − n − 2 s |f(x)|^2|x-y|^{-n-2s} ∣ f ( x ) ∣ 2 ∣ x − y ∣ − n − 2 s on B × B c B\times B^c B × B c and ∣ f ( y ) ∣ 2 ∣ x − y ∣ − n − 2 s |f(y)|^2|x-y|^{-n-2s} ∣ f ( y ) ∣ 2 ∣ x − y ∣ − n − 2 s on B c × B B^c\times B B c × B . Discarding the remaining nonnegative contribution of B × B B\times B B × B and using the symmetry of these two regions,
E s ( f ) ≥ c n , s 2 ⋅ 2 ∫ B ∣ f ( x ) ∣ 2 ∫ B c ∣ x − y ∣ − n − 2 s d y d x = c n , s ∫ B ∣ f ( x ) ∣ 2 ∫ B c ∣ x − y ∣ − n − 2 s d y d x . \mathcal{E}_s(f)\ge\frac{c_{n,s}}{2}\cdot2\int_B|f(x)|^2\int_{B^c}|x-y|^{-n-2s}\,dy\,dx
=c_{n,s}\int_B|f(x)|^2\int_{B^c}|x-y|^{-n-2s}\,dy\,dx. E s ( f ) ≥ 2 c n , s ⋅ 2 ∫ B ∣ f ( x ) ∣ 2 ∫ B c ∣ x − y ∣ − n − 2 s d y d x = c n , s ∫ B ∣ f ( x ) ∣ 2 ∫ B c ∣ x − y ∣ − n − 2 s d y d x .
Put d = 1 − ∣ x ∣ ∈ ( 0 , 1 ] d=1-|x|\in(0,1] d = 1 − ∣ x ∣ ∈ ( 0 , 1 ] . If x ≠ 0 x\ne0 x = 0 , let e = x / ∣ x ∣ e=x/|x| e = x /∣ x ∣ ; if x = 0 x=0 x = 0 , take any unit vector. Every point of the ball B d / 2 ( ( 1 + d ) e ) B_{d/2}((1+d)e) B d /2 (( 1 + d ) e ) is at distance at least 1 + d − d / 2 > 1 1+d-d/2>1 1 + d − d /2 > 1 from the origin, so that ball lies outside B B B ; since ∣ x − ( 1 + d ) e ∣ = 2 d |x-(1+d)e|=2d ∣ x − ( 1 + d ) e ∣ = 2 d , its points are at distance at most 5 d / 2 5d/2 5 d /2 from x x x . Hence the inner integral is at least
ω n ( d 2 ) n ( 5 d 2 ) − n − 2 s = ω n 2 2 s 5 − n − 2 s d − 2 s , \omega_n\left(\frac d2\right)^n\left(\frac{5d}{2}\right)^{-n-2s}
=\omega_n2^{2s}5^{-n-2s}d^{-2s}, ω n ( 2 d ) n ( 2 5 d ) − n − 2 s = ω n 2 2 s 5 − n − 2 s d − 2 s ,
which proves (A.9) with C = ( c n , s ω n 2 2 s 5 − n − 2 s ) − 1 C=(c_{n,s}\omega_n2^{2s}5^{-n-2s})^{-1} C = ( c n , s ω n 2 2 s 5 − n − 2 s ) − 1 .
The boundary term. Apply (A.9) to f = u ∗ − u B f=u^*-u_B f = u ∗ − u B . This is admissible: u B ∈ H 0 s ( B ) u_B\in H^s_0(B) u B ∈ H 0 s ( B ) , and Lemma 2.2 shows u ∗ ∈ H 0 s ( B ) u^*\in H^s_0(B) u ∗ ∈ H 0 s ( B ) ; in particular both belong to H s ( R n ) H^s(\mathbb{R}^n) H s ( R n ) and vanish a.e. outside B B B , the latter because ∣ { u ∗ > 0 } ∣ = ∣ Ω ∣ = ∣ B ∣ \lvert\{u^*>0\}\rvert=|\Omega|=|B| ∣{ u ∗ > 0 }∣ = ∣Ω∣ = ∣ B ∣ and u ∗ u^* u ∗ is radial nonincreasing. Both functions being radial, f ( ρ e 1 ) = v ( ρ , 0 ) f(\rho e_1)=v(\rho,0) f ( ρ e 1 ) = v ( ρ , 0 ) , and polar integration gives
∫ 0 1 ∣ v ( ρ , 0 ) ∣ 2 ( 1 − ρ ) 2 s P ρ d ρ = ∫ B ∣ f ( x ) ∣ 2 ( 1 − ∣ x ∣ ) 2 s d x ≤ C E s ( u ∗ − u B ) . \int_0^1\frac{|v(\rho,0)|^2}{(1-\rho)^{2s}}P_\rho\,d\rho
=\int_B\frac{|f(x)|^2}{(1-|x|)^{2s}}\,dx
\le C\mathcal{E}_s(u^*-u_B). ∫ 0 1 ( 1 − ρ ) 2 s ∣ v ( ρ , 0 ) ∣ 2 P ρ d ρ = ∫ B ( 1 − ∣ x ∣ ) 2 s ∣ f ( x ) ∣ 2 d x ≤ C E s ( u ∗ − u B ) .
By (14) and the nonnegativity of E s ( u ) − E s ( u ∗ ) E_s(u)-E_s(u^*) E s ( u ) − E s ( u ∗ ) we have E s ( u ∗ − u B ) ≤ δ T ( Ω ) E_s(u^*-u_B)\leq\delta_T(\Omega) E s ( u ∗ − u B ) ≤ δ T ( Ω ) . On ρ ∈ [ 1 − ℓ , 1 ] \rho\in[1-\ell,1] ρ ∈ [ 1 − ℓ , 1 ] , with ℓ < 1 / 4 \ell<1/4 ℓ < 1/4 , the radial factor P ρ = n ω n ρ n − 1 P_\rho=n\omega_n\rho^{n-1} P ρ = n ω n ρ n − 1 is bounded above and below by positive dimensional constants. This proves the boundary part of (A.10).
The interior term. Both U ∗ U^* U ∗ and U B U_B U B are radial in x x x , so U ∗ − U B U^*-U_B U ∗ − U B has radial profile v v v and, as in the proof of Lemma 3.1 ,
∫ R + n z 1 − 2 s ∣ ∇ ( U ∗ − U B ) ∣ 2 d x d z = ∫ 0 ∞ z 1 − 2 s ∫ 0 ∞ P ρ ∣ ∇ v ∣ 2 d ρ d z . \int_{\mathbb{R}^n_+}z^{1-2s}|\nabla(U^*-U_B)|^2\,dx\,dz=\int_0^\infty z^{1-2s}\int_0^\infty P_\rho|\nabla v|^2\,d\rho\,dz. ∫ R + n z 1 − 2 s ∣∇ ( U ∗ − U B ) ∣ 2 d x d z = ∫ 0 ∞ z 1 − 2 s ∫ 0 ∞ P ρ ∣∇ v ∣ 2 d ρ d z .
Restricting the nonnegative integrand to R \mathcal{R} R , bounding P ρ P_\rho P ρ from below there, and using (16) gives
∫ R z 1 − 2 s ∣ ∇ v ∣ 2 d ρ d z ≤ C δ T ( Ω ) . \int_{\mathcal{R}}z^{1-2s}|\nabla v|^2\,d\rho\,dz\leq C\delta_T(\Omega). ∫ R z 1 − 2 s ∣∇ v ∣ 2 d ρ d z ≤ C δ T ( Ω ) .
Identification of the bottom trace. The kernel P z ( s ) P_z^{(s)} P z ( s ) is an approximate identity, so U B ( ⋅ , z ) → u B U_B(\cdot,z)\to u_B U B ( ⋅ , z ) → u B and U ( ⋅ , z ) → u U(\cdot,z)\to u U ( ⋅ , z ) → u in L 2 ( R n ) L^2(\mathbb{R}^n) L 2 ( R n ) , and (13) gives ∥ U ∗ ( ⋅ , z ) − u ∗ ∥ 2 ≤ ∥ U ( ⋅ , z ) − u ∥ 2 → 0 \|U^*(\cdot,z)-u^*\|_2\leq\|U(\cdot,z)-u\|_2\to0 ∥ U ∗ ( ⋅ , z ) − u ∗ ∥ 2 ≤ ∥ U ( ⋅ , z ) − u ∥ 2 → 0 . Hence
∫ 0 ∞ P ρ ∣ v ( ρ , z ) − v ( ρ , 0 ) ∣ 2 d ρ → 0 , \int_0^\infty P_\rho|v(\rho,z)-v(\rho,0)|^2\,d\rho\to0, ∫ 0 ∞ P ρ ∣ v ( ρ , z ) − v ( ρ , 0 ) ∣ 2 d ρ → 0 ,
and the two-sided bound for P ρ P_\rho P ρ gives the asserted L 2 ( ( 1 − ℓ , 1 ) ) L^2((1-\ell,1)) L 2 (( 1 − ℓ , 1 )) convergence. □
Acknowledgments
The authors acknowledge financial support from the ANR grant STOIQUES. MN and BR are members of INdAM/GNAMPA. BR is a member of INdAM–GNAMPA and has been partially supported by the “INdAM - GNAMPA Project”, code CUP E53C25002010001.
LAMA, CNRS, Université Paris-Est Créteil, 61 avenue du Général de Gaulle, 94010 Créteil, France
Email address: michael.goldman@cnrs.fr
Department of Mathematics, University of Pisa, 56127 Pisa, Italy
Email address: matteo.novaga@unipi.it
Department of Mathematics, University of Bologna, 40126 Bologna, Italy
Email address: berardo.ruffini@unibo.it