Introduction
The aim of this paper is to show, in arguably the simplest meaningful situation, how second order shape derivatives can be computed for shape optimization problems involving operators of fractional order. To this end we compute the second shape derivative of the fractional torsion energy for smooth domains and the exponent s = 1 / 2 s = 1/2 s = 1/2 . While second order shape derivatives are well understood in the local setting, see e.g. [6 , 20 ] , and first order derivatives for fractional problems have received considerable attention, see e.g. [7 , 9 , 10 , 23 ] , second order derivatives in the fractional setting have been less studied. We expect the approach developed here to be useful for related problems, such as fractional capacities and first Dirichlet eigenvalues, and for other exponents s ∈ ( 0 , 1 ) s \in(0, 1) s ∈ ( 0 , 1 ) .
Our initial motivation was quantitative stability of the ball. For problems in which the ball is a volume-constrained optimizer (a minimizer or a maximizer, depending on the problem), computing the second variation and establishing its coercivity are essential steps in carrying out the program of [6 ] , namely a proof of sharp quantitative stability through the study of nearly spherical sets and a selection principle. We were indeed on this path when we found, in [15 ] , a much shorter and less technical proof of quantitative stability, adapting the strategy of [19 ] . We nevertheless decided to write down a rigorous derivation of the second variation, since it may be useful for other purposes. With this in mind we also prove its coercivity at the ball under volume-preserving variations.
Following the full strategy of [6 ] would require further results. First, one would need weaker regularity assumptions on the domains, or a regularity bootstrap for the nearly spherical sets arising in the contradiction argument; see [4 , 22 ] . Second, one would need continuity estimates for the second variation in the relevant Sobolev norm, here H 1 / 2 H^{1/2} H 1/2 . This might require the framework of [16 , 17 , 18 ] , instead of the regularity results of [3 ] used here. Finally, the selection principle would require an adaptation of the regularity theory for fractional Bernoulli free boundary problems developed in [8 ] .
Let Ω ⊂ R n \Omega\subset\mathbb{R}^n Ω ⊂ R n , n ≥ 2 n \ge2 n ≥ 2 , be a bounded open set with C ∞ C^\infty C ∞ boundary. The fractional torsion function u Ω u_\Omega u Ω is the solution of
( − Δ ) 1 / 2 u Ω = 1 in Ω , u Ω = 0 in R n ∖ Ω , (1) (-\Delta)^{1/2}u_\Omega=1 \text{ in }\Omega,\qquad u_\Omega=0 \text{ in }\mathbb{R}^n\setminus\Omega, \tag*{(1)} ( − Δ ) 1/2 u Ω = 1 in Ω , u Ω = 0 in R n ∖ Ω , (1)
where ( − Δ ) 1 / 2 (-\Delta)^{1/2} ( − Δ ) 1/2 is the fractional Laplacian with Fourier symbol ∣ ξ ∣ |\xi| ∣ ξ ∣ ,
( − Δ ) 1 / 2 v ( x ) = c n P.V. ∫ R n v ( x ) − v ( y ) ∣ x − y ∣ n + 1 d y , c n = Γ ( ( n + 1 ) / 2 ) π ( n + 1 ) / 2 , (-\Delta)^{1/2}v(x)=c_n\operatorname{P.V.}\int_{\mathbb{R}^n}\frac{v(x)-v(y)}{|x-y|^{n+1}}\,\mathrm{d}y,\qquad c_n=\frac{\Gamma((n+1)/2)}{\pi^{(n+1)/2}}, ( − Δ ) 1/2 v ( x ) = c n P.V. ∫ R n ∣ x − y ∣ n + 1 v ( x ) − v ( y ) d y , c n = π ( n + 1 ) /2 Γ (( n + 1 ) /2 ) ,
and the fractional torsional rigidity , or torsion energy, of Ω \Omega Ω is
T ( Ω ) = ∫ Ω u Ω . T(\Omega)=\int_\Omega u_\Omega. T ( Ω ) = ∫ Ω u Ω .
The fractional Saint-Venant inequality states that balls are the (unique) maximizers of the fractional torsion under volume constraint, see e.g. [5 , 15 ] . It is then a natural question to compute the corresponding first and second order shape derivatives. Let us first recall what is known for the first order derivatives from the works [7 , 9 , 10 , 11 , 23 ] . For this we need to introduce some notation.
Problem (1) is understood in the energy space
V = { v ∈ H 1 / 2 ( R n ) : v = 0 a.e. in R n ∖ Ω } , V=\{v\in H^{1/2}(\mathbb{R}^n):v=0\ \text{a.e. in }\mathbb{R}^n\setminus\Omega\}, V = { v ∈ H 1/2 ( R n ) : v = 0 a.e. in R n ∖ Ω } ,
a ( v 1 , v 2 ) = c n 2 ∬ R 2 n ( v 1 ( y ) − v 1 ( z ) ) ( v 2 ( y ) − v 2 ( z ) ) ∣ y − z ∣ n + 1 d y d z , (2) a(v_1,v_2)=\frac{c_n}{2}\iint_{\mathbb{R}^{2n}}\frac{(v_1(y)-v_1(z))(v_2(y)-v_2(z))}{|y-z|^{n+1}}\,\mathrm{d}y\,\mathrm{d}z, \tag*{(2)} a ( v 1 , v 2 ) = 2 c n ∬ R 2 n ∣ y − z ∣ n + 1 ( v 1 ( y ) − v 1 ( z )) ( v 2 ( y ) − v 2 ( z )) d y d z , (2)
so that a ( v 1 , v 2 ) = ∫ R n v 1 ( − Δ ) 1 / 2 v 2 a(v_1,v_2)=\int_{\mathbb{R}^n}v_1(-\Delta)^{1/2}v_2 a ( v 1 , v 2 ) = ∫ R n v 1 ( − Δ ) 1/2 v 2 for smooth functions. Then u Ω u_\Omega u Ω is the unique element of V V V with a ( u Ω , v ) = ∫ Ω v a(u_\Omega,v)=\int_\Omega v a ( u Ω , v ) = ∫ Ω v for every v ∈ V v\in V v ∈ V , and
T ( Ω ) = max v ∈ V { 2 ∫ Ω v − a ( v , v ) } , (3) T(\Omega)=\max_{v\in V}\left\{2\int_\Omega v-a(v,v)\right\}, \tag*{(3)} T ( Ω ) = v ∈ V max { 2 ∫ Ω v − a ( v , v ) } , (3)
the maximum being attained at v = u Ω v=u_\Omega v = u Ω .
We write ν \nu ν for the outer unit normal to ∂ Ω \partial\Omega ∂ Ω , H = div ∂ Ω ν H=\operatorname{div}_{\partial\Omega}\nu H = div ∂ Ω ν for the mean curvature and d d d for the distance to ∂ Ω \partial\Omega ∂ Ω . We fix r 0 > 0 r_0>0 r 0 > 0 such that ( x , r ) ↦ x − r ν ( x ) (x,r)\mapsto x-r\nu(x) ( x , r ) ↦ x − r ν ( x ) is a diffeomorphism from ∂ Ω × ( − 2 r 0 , 2 r 0 ) \partial\Omega\times(-2r_0,2r_0) ∂ Ω × ( − 2 r 0 , 2 r 0 ) onto a neighbourhood of ∂ Ω \partial\Omega ∂ Ω , and a function δ ∈ C ∞ ( Ω ‾ ) \delta\in C^\infty(\overline{\Omega}) δ ∈ C ∞ ( Ω ) , positive in Ω \Omega Ω , with δ = d \delta=d δ = d on { y ∈ Ω : d ( y ) < r 0 } \{y\in\Omega:d(y)<r_0\} { y ∈ Ω : d ( y ) < r 0 } . Thus d ( x − r ν ( x ) ) = δ ( x − r ν ( x ) ) = r d(x-r\nu(x))=\delta(x-r\nu(x))=r d ( x − r ν ( x )) = δ ( x − r ν ( x )) = r for x ∈ ∂ Ω x\in\partial\Omega x ∈ ∂ Ω and 0 < r < r 0 0<r<r_0 0 < r < r 0 .
The torsion function behaves like d \sqrt{d} d at the boundary [24 ] : by Lemma 2.1 below, u Ω = δ U u_\Omega=\sqrt{\delta}\,U u Ω = δ U with U ∈ C ∞ ( Ω ‾ ) U\in C^\infty(\overline{\Omega}) U ∈ C ∞ ( Ω ) . The function U U U depends on the choice of δ \delta δ away from ∂ Ω \partial\Omega ∂ Ω , but its behaviour near ∂ Ω \partial\Omega ∂ Ω does not, since there U ( x − r ν ( x ) ) = r − 1 / 2 u Ω ( x − r ν ( x ) ) U(x-r\nu(x))=r^{-1/2}u_\Omega(x-r\nu(x)) U ( x − r ν ( x )) = r − 1/2 u Ω ( x − r ν ( x )) . The torsion function enters our formulas only through two boundary quantities: the weighted trace ψ \psi ψ of u Ω u_\Omega u Ω and its first normal coefficient κ \kappa κ , defined for x ∈ ∂ Ω x\in\partial\Omega x ∈ ∂ Ω by
ψ ( x ) = lim r ↓ 0 u Ω ( x − r ν ( x ) ) r = U ( x ) , κ ( x ) = ∂ r U ( x − r ν ( x ) ) ∣ r = 0 , \psi(x)=\lim_{r\downarrow0}\frac{u_\Omega(x-r\nu(x))}{\sqrt{r}}=U(x),\qquad\kappa(x)=\left.\partial_rU(x-r\nu(x))\right|_{r=0}, ψ ( x ) = r ↓ 0 lim r u Ω ( x − r ν ( x )) = U ( x ) , κ ( x ) = ∂ r U ( x − r ν ( x )) ∣ r = 0 ,
so that uniformly in x ∈ ∂ Ω x\in\partial\Omega x ∈ ∂ Ω
u Ω ( x − r ν ( x ) ) = ψ ( x ) r + κ ( x ) r 3 / 2 + O ( r 5 / 2 ) . (4) u_\Omega(x-r\nu(x))=\psi(x)\sqrt{r}+\kappa(x)r^{3/2}+O(r^{5/2}). \tag*{(4)} u Ω ( x − r ν ( x )) = ψ ( x ) r + κ ( x ) r 3/2 + O ( r 5/2 ) . (4)
In terms of these coefficients, the first variation of T T T is as follows. Let Φ t \Phi_t Φ t be any smooth family of diffeomorphisms of R n \mathbb{R}^n R n with Φ 0 = Id \Phi_0=\operatorname{Id} Φ 0 = Id , let X = ∂ t Φ t ∣ t = 0 X=\left.\partial_t\Phi_t\right|_{t=0} X = ∂ t Φ t ∣ t = 0 and φ = X ⋅ ν \varphi=X\cdot\nu φ = X ⋅ ν on ∂ Ω \partial\Omega ∂ Ω . Let Ω t = Φ t ( Ω ) \Omega_t=\Phi_t(\Omega) Ω t = Φ t ( Ω ) . Then
d d t T ( Ω t ) ∣ t = 0 = c ∗ ∫ ∂ Ω ψ 2 φ d H n − 1 , c ∗ : = Γ ( 3 / 2 ) 2 = π 4 . (5) \left.\frac{d}{dt}T(\Omega_t)\right|_{t=0}=c_*\int_{\partial\Omega}\psi^2\varphi\,d\mathcal{H}^{n-1},\qquad c_*:=\Gamma(3/2)^2=\frac{\pi}{4}. \tag*{(5)} d t d T ( Ω t ) t = 0 = c ∗ ∫ ∂ Ω ψ 2 φ d H n − 1 , c ∗ := Γ ( 3/2 ) 2 = 4 π . (5)
As in [10 ] where the fractional Dirichlet eigenvalue is considered, we recover the first variation formula (5) through a short application of the generalized fractional Pohozaev identity of [10 ] ; see Remark 3.6.
We now turn to the second order derivative. The difficulty comes from understanding the PDE which is solved by the Eulerian derivative q = ∂ t u Ω t ∣ t = 0 q=\left.\partial_tu_{\Omega_t}\right|_{t=0} q = ∂ t u Ω t ∣ t = 0 . Indeed, by differentiating the PDE solved by u Ω t u_{\Omega_t} u Ω t we can guess that it is 1 / 2 1/2 1/2 -harmonic in Ω \Omega Ω and equal to zero outside Ω \Omega Ω . What makes it both non-trivial and non-standard is that it blows up like the inverse square root of the distance to the boundary. In particular, it lies outside the energy space in which (1) is posed. These are ”large solutions” in the language of [1 , 26 ] . One of the main points of the analysis is thus to understand a certain Dirichlet-to-Neumann-type operator D Ω \mathcal{D}_{\Omega} D Ω , defined in (7) below, which acts on the boundary data of these large 1 / 2 1/2 1/2 -harmonic functions. Here we will only need an elementary construction of these functions and of D Ω \mathcal{D}_{\Omega} D Ω (Proposition 2.4) which we now describe. For every f ∈ C ∞ ( ∂ Ω ) f\in C^\infty(\partial\Omega) f ∈ C ∞ ( ∂ Ω ) we prove that there is a unique function Q f Q_f Q f , vanishing outside Ω \Omega Ω and with δ Q f ∈ C ∞ ( Ω ‾ ) \sqrt{\delta}Q_f\in C^\infty(\overline{\Omega}) δ Q f ∈ C ∞ ( Ω ) , such that
( − Δ ) 1 / 2 Q f = 0 in Ω , ( δ Q f ) ∣ ∂ Ω = f . (6) (-\Delta)^{1/2}Q_f=0\ \text{in }\Omega,\qquad\left.(\sqrt{\delta}Q_f)\right|_{\partial\Omega}=f. \tag*{(6)} ( − Δ ) 1/2 Q f = 0 in Ω , ( δ Q f ) ∂ Ω = f . (6)
Here the equation holds in the sense of distributions: Q f Q_f Q f is integrable, but it blows up like d − 1 / 2 d^{-1/2} d − 1/2 at ∂ Ω \partial\Omega ∂ Ω and is not in L 2 L^2 L 2 unless f = 0 f=0 f = 0 . We set, for x ∈ ∂ Ω x\in\partial\Omega x ∈ ∂ Ω ,
D Ω f ( x ) = ∂ r ( r Q f ( x − r ν ( x ) ) ) ∣ r = 0 , (7) \mathcal{D}_{\Omega}f(x)=\left.\partial_r\left(\sqrt{r}Q_f(x-r\nu(x))\right)\right|_{r=0}, \tag*{(7)} D Ω f ( x ) = ∂ r ( r Q f ( x − r ν ( x )) ) r = 0 , (7)
that is, Q f ( x − r ν ( x ) ) = f ( x ) r − 1 / 2 + D Ω f ( x ) r + O ( r 3 / 2 ) Q_f(x-r\nu(x))=f(x)r^{-1/2}+\mathcal{D}_{\Omega}f(x)\sqrt{r}+O(r^{3/2}) Q f ( x − r ν ( x )) = f ( x ) r − 1/2 + D Ω f ( x ) r + O ( r 3/2 ) . The construction of D Ω \mathcal{D}_{\Omega} D Ω in Proposition 2.4 only uses the regularity result Lemma 2.1 .
We now give the main statement. We will limit ourselves to the computation in the case of normal graphs. This is very common in the study of shape derivatives, see [20 ] . Fix φ ∈ C ∞ ( ∂ Ω ) \varphi\in C^\infty(\partial\Omega) φ ∈ C ∞ ( ∂ Ω ) . For x ∈ ∂ Ω x\in\partial\Omega x ∈ ∂ Ω and ∣ r ∣ ≤ r 0 / 2 |r|\le r_0/2 ∣ r ∣ ≤ r 0 /2 we set
X ( x − r ν ( x ) ) = φ ( x ) ν ( x ) (8) X(x-r\nu(x))=\varphi(x)\nu(x) \tag*{(8)} X ( x − r ν ( x )) = φ ( x ) ν ( x ) (8)
which we extend to a smooth vector field X X X compactly supported in { d < r 0 } \{d<r_0\} { d < r 0 } . We then set
Φ t ( x ) = x + t X ( x ) , \Phi_t(x)=x+tX(x), Φ t ( x ) = x + tX ( x ) ,
which is a diffeomorphism for ∣ t ∣ |t| ∣ t ∣ small enough. We write
Ω t = Φ t ( Ω ) , ∂ Ω t = { x + t φ ( x ) ν ( x ) : x ∈ ∂ Ω } . \Omega_t=\Phi_t(\Omega),\qquad\partial\Omega_t=\{x+t\varphi(x)\nu(x):x\in\partial\Omega\}. Ω t = Φ t ( Ω ) , ∂ Ω t = { x + tφ ( x ) ν ( x ) : x ∈ ∂ Ω } .
Theorem 1.1. Let Ω ⊂ R n \Omega\subset\mathbb{R}^n Ω ⊂ R n be a bounded open set with C ∞ C^\infty C ∞ boundary, and let φ ∈ C ∞ ( ∂ Ω ) \varphi\in C^\infty(\partial\Omega) φ ∈ C ∞ ( ∂ Ω ) . The function t ↦ T ( Ω t ) t\mapsto T(\Omega_t) t ↦ T ( Ω t ) is of class C 2 C^2 C 2 near 0 0 0 , and
d 2 d t 2 T ( Ω t ) ∣ t = 0 = H Ω ( φ ) : = c ∗ ∫ ∂ Ω { ψ φ D Ω ( ψ φ ) + ( H ψ 2 − 3 ψ κ ) φ 2 } d H n − 1 . (9) \left.\frac{d^2}{dt^2}T(\Omega_t)\right|_{t=0}=\mathcal{H}_{\Omega}(\varphi):=c_*\int_{\partial\Omega}\left\{\psi\varphi\mathcal{D}_{\Omega}(\psi\varphi)+(H\psi^2-3\psi\kappa)\varphi^2\right\}\,d\mathcal{H}^{n-1}. \tag*{(9)} d t 2 d 2 T ( Ω t ) t = 0 = H Ω ( φ ) := c ∗ ∫ ∂ Ω { ψ φ D Ω ( ψ φ ) + ( H ψ 2 − 3 ψ κ ) φ 2 } d H n − 1 . (9)
The natural variable in (9) is the weighted displacement ψ φ \psi\varphi ψ φ . The term H ψ 2 H\psi^2 H ψ 2 comes from the variation of the area element, and the term − 3 ψ κ -3\psi\kappa − 3 ψ κ from following the moving boundary point in the expansion (4) (see (17) below).
In the last part of the paper we compute the second variation of the unit ball for volume constrained (to first order) variations. We show that this Hessian is negative and coercive in H 1 / 2 ( S n − 1 ) H^{1/2}(\mathbb{S}^{n-1}) H 1/2 ( S n − 1 ) modulo translations. This result might appear surprising at first sight since this coercivity is in the same norm as for the local case, see [6 ] . This comes from the fact that as in the local case, the operator involved (here D Ω ( ψ φ ) D_{\Omega}(\psi\varphi) D Ω ( ψ φ ) ) is a Dirichlet-to-Neumann map which is thus of order one.
Let P \mathcal{P} P denote harmonic extension to B = B 1 B=B_1 B = B 1 , Λ = ∂ ν P \Lambda=\partial_\nu\mathcal{P} Λ = ∂ ν P the classical Dirichlet-to-Neumann operator, and φ ℓ \varphi_\ell φ ℓ the projection onto spherical harmonics of degree ℓ \ell ℓ . With ∥ ⋅ ∥ 2 \|\cdot\|_2 ∥ ⋅ ∥ 2 denoting the L 2 ( S n − 1 ) L^2(\mathbb{S}^{n-1}) L 2 ( S n − 1 ) norm, set
∥ φ ∥ H ∗ 1 / 2 2 : = ∑ ℓ ≥ 0 ( 1 + ℓ ) ∥ φ ℓ ∥ 2 2 = ∥ φ ∥ 2 2 + ∫ B ∣ ∇ P φ ∣ 2 d y . \|\varphi\|_{H_*^{1/2}}^2 := \sum_{\ell\ge0}(1+\ell)\|\varphi_\ell\|_2^2 = \|\varphi\|_2^2+\int_B|\nabla\mathcal{P}\varphi|^2\,dy. ∥ φ ∥ H ∗ 1/2 2 := ℓ ≥ 0 ∑ ( 1 + ℓ ) ∥ φ ℓ ∥ 2 2 = ∥ φ ∥ 2 2 + ∫ B ∣∇ P φ ∣ 2 d y .
which is equivalent to the usual H 1 / 2 ( S n − 1 ) H^{1/2}(\mathbb{S}^{n-1}) H 1/2 ( S n − 1 ) norm. Let ψ \psi ψ be the weighted trace of u B u_B u B and set λ ∗ = c ∗ ψ 2 \lambda_* = c_*\psi^2 λ ∗ = c ∗ ψ 2 . For any φ ∈ C ∞ ( S n − 1 ) \varphi\in C^\infty(\mathbb{S}^{n-1}) φ ∈ C ∞ ( S n − 1 ) with
∫ S n − 1 φ d H n − 1 = 0 , (10) \int_{\mathbb{S}^{n-1}}\varphi\,d\mathcal{H}^{n-1}=0, \tag*{(10)} ∫ S n − 1 φ d H n − 1 = 0 , (10)
define the quadratic form
Q ( φ ) : = λ ∗ ∫ S n − 1 ( φ Λ φ − φ 2 ) d H n − 1 = λ ∗ ∑ ℓ ≥ 0 ( ℓ − 1 ) ∥ φ ℓ ∥ 2 2 . (11) Q(\varphi):=\lambda_*\int_{\mathbb{S}^{n-1}}(\varphi\Lambda\varphi-\varphi^2)\,d\mathcal{H}^{n-1}
=\lambda_*\sum_{\ell\ge0}(\ell-1)\|\varphi_\ell\|_2^2. \tag*{(11)} Q ( φ ) := λ ∗ ∫ S n − 1 ( φ Λ φ − φ 2 ) d H n − 1 = λ ∗ ℓ ≥ 0 ∑ ( ℓ − 1 ) ∥ φ ℓ ∥ 2 2 . (11)
For h ∈ C ∞ ( S n − 1 ) h\in C^\infty(\mathbb{S}^{n-1}) h ∈ C ∞ ( S n − 1 ) , we set Ω h = { ρ θ : ρ ∈ [ 0 , 1 + h ( θ ) ] , θ ∈ S n − 1 } \Omega_h=\{\rho\theta:\rho\in[0,1+h(\theta)],\ \theta\in\mathbb{S}^{n-1}\} Ω h = { ρθ : ρ ∈ [ 0 , 1 + h ( θ )] , θ ∈ S n − 1 } .
Theorem 1.2. For any smooth volume-preserving path Ω h ( t ) \Omega_{h(t)} Ω h ( t ) with h ( 0 ) = 0 h(0)=0 h ( 0 ) = 0 , setting φ = h ˙ ( 0 ) \varphi=\dot h(0) φ = h ˙ ( 0 ) (which satisfies (1.10) by the volume preserving property), we have
− d 2 d t 2 T ( Ω h ( t ) ) ∣ t = 0 = Q ( φ ) . (12) -\left.\frac{d^2}{dt^2}T(\Omega_{h(t)})\right|_{t=0}=Q(\varphi). \tag*{(12)} − d t 2 d 2 T ( Ω h ( t ) ) t = 0 = Q ( φ ) . (12) If moreover φ \varphi φ satisfies
∫ S n − 1 φ ν d H n − 1 = 0 , (13) \int_{\mathbb{S}^{n-1}}\varphi\nu\,d\mathcal{H}^{n-1}=0, \tag*{(13)} ∫ S n − 1 φ ν d H n − 1 = 0 , (13) then
λ ∗ 3 ∥ φ ∥ H ∗ 1 / 2 2 ≤ Q ( φ ) ≤ λ ∗ ∥ φ ∥ H ∗ 1 / 2 2 . (14) \frac{\lambda_*}{3}\|\varphi\|_{H_*^{1/2}}^2\le Q(\varphi)\le\lambda_*\|\varphi\|_{H_*^{1/2}}^2. \tag*{(14)} 3 λ ∗ ∥ φ ∥ H ∗ 1/2 2 ≤ Q ( φ ) ≤ λ ∗ ∥ φ ∥ H ∗ 1/2 2 . (14) The constant 1 / 3 1/3 1/3 is attained exactly on spherical harmonics of degree two.
Conditions (1.10) and (1.13) correspond to (first order) volume preserving and translation modes.
A formal derivation of the second variation
The following computation explains the structure of (1.9). It assumes that all the objects involved are smooth in t t t and that all expansions may be differentiated; it is not used in the proof.
For x ∈ ∂ Ω x\in\partial\Omega x ∈ ∂ Ω set x t = Φ t ( x ) = x + t φ ( x ) ν ( x ) x_t=\Phi_t(x)=x+t\varphi(x)\nu(x) x t = Φ t ( x ) = x + tφ ( x ) ν ( x ) , write ν t \nu_t ν t for the outer normal to Ω t \Omega_t Ω t , and u t = u Ω t u_t=u_{\Omega_t} u t = u Ω t , ψ t \psi_t ψ t , κ t \kappa_t κ t for the torsion function of Ω t \Omega_t Ω t and its boundary coefficients. The normal velocity of ∂ Ω t \partial\Omega_t ∂ Ω t at x t x_t x t is φ t ( x t ) = φ ( x ) ν ( x ) ⋅ ν t ( x t ) \varphi_t(x_t)=\varphi(x)\nu(x)\cdot\nu_t(x_t) φ t ( x t ) = φ ( x ) ν ( x ) ⋅ ν t ( x t ) .
First variation at time t t t . By (1.5) applied at Ω t \Omega_t Ω t , we have
d d t T ( Ω t ) = c ∗ ∫ ∂ Ω t ψ t 2 φ t d H n − 1 . \frac{d}{dt}T(\Omega_t)=c_*\int_{\partial\Omega_t}\psi_t^2\varphi_t\,d\mathcal{H}^{n-1}. d t d T ( Ω t ) = c ∗ ∫ ∂ Ω t ψ t 2 φ t d H n − 1 .
Differentiation of the boundary integral. Differentiating once more using the standard formula for boundary integrals, see [20 ] , and noting that ∂ t φ t ( x t ) ∣ t = 0 = 0 \left.\partial_t\varphi_t(x_t)\right|_{t=0}=0 ∂ t φ t ( x t ) ∣ t = 0 = 0 , we obtain
d 2 d t 2 T ( Ω t ) ∣ t = 0 = c ∗ ∫ ∂ Ω { 2 ψ ψ ˙ φ + H ψ 2 φ 2 } d H n − 1 , ψ ˙ ( x ) : = d d t ψ t ( x t ) ∣ t = 0 , (15) \left.\frac{d^2}{dt^2}T(\Omega_t)\right|_{t=0}
=c_*\int_{\partial\Omega}\left\{2\psi\dot\psi\varphi+H\psi^2\varphi^2\right\}\,d\mathcal{H}^{n-1},
\qquad
\dot\psi(x):=\left.\frac{d}{dt}\psi_t(x_t)\right|_{t=0}, \tag*{(15)} d t 2 d 2 T ( Ω t ) t = 0 = c ∗ ∫ ∂ Ω { 2 ψ ψ ˙ φ + H ψ 2 φ 2 } d H n − 1 , ψ ˙ ( x ) := d t d ψ t ( x t ) t = 0 , (15)
and everything reduces to the computation of ψ ˙ \dot\psi ψ ˙ .
Differentiation of the weighted trace. Recall that q = ∂ t u t ∣ t = 0 q=\left.\partial_tu_t\right|_{t=0} q = ∂ t u t ∣ t = 0 . We have
ψ t ( x t ) = lim r → 0 r − 1 / 2 u t ( x t − r ν t ( x t ) ) , ∂ t ( x t − r ν t ( x t ) ) ∣ t = 0 = φ ν + r ∇ ∂ Ω φ . \psi_t(x_t)=\lim_{r\to0}r^{-1/2}u_t\bigl(x_t-r\nu_t(x_t)\bigr),\qquad\left.\partial_t\bigl(x_t-r\nu_t(x_t)\bigr)\right|_{t=0}=\varphi\nu+r\nabla_{\partial\Omega}\varphi. ψ t ( x t ) = r → 0 lim r − 1/2 u t ( x t − r ν t ( x t ) ) , ∂ t ( x t − r ν t ( x t ) ) t = 0 = φ ν + r ∇ ∂ Ω φ .
Differentiating at fixed r r r gives
d d t [ u t ( x t − r ν t ( x t ) ) r ] ∣ t = 0 = 1 r { q ( x − r ν ) + φ ν ⋅ ∇ u Ω ( x − r ν ) + r ∇ ∂ Ω φ ⋅ ∇ u Ω ( x − r ν ) } . \left.\frac{d}{dt}\left[\frac{u_t(x_t-r\nu_t(x_t))}{\sqrt{r}}\right]\right|_{t=0}
=\frac{1}{\sqrt{r}}\left\{q(x-r\nu)+\varphi\nu\cdot\nabla u_\Omega(x-r\nu)+r\nabla_{\partial\Omega}\varphi\cdot\nabla u_\Omega(x-r\nu)\right\}. d t d [ r u t ( x t − r ν t ( x t )) ] t = 0 = r 1 { q ( x − r ν ) + φ ν ⋅ ∇ u Ω ( x − r ν ) + r ∇ ∂ Ω φ ⋅ ∇ u Ω ( x − r ν ) } .
By (4) , ν ⋅ ∇ u Ω ( x − r ν ) = − ψ 2 r − 3 2 κ r + O ( r 3 / 2 ) \nu\cdot\nabla u_\Omega(x-r\nu)=-\frac{\psi}{2\sqrt{r}}-\frac{3}{2}\kappa\sqrt{r}+O(r^{3/2}) ν ⋅ ∇ u Ω ( x − r ν ) = − 2 r ψ − 2 3 κ r + O ( r 3/2 ) , while the tangential part of ∇ u Ω ( x − r ν ) \nabla u_\Omega(x-r\nu) ∇ u Ω ( x − r ν ) is O ( r ) O(\sqrt{r}) O ( r ) . We shall see below that
q ( x − r ν ( x ) ) = ψ φ 2 r + B φ r + O ( r 3 / 2 ) (16) q(x-r\nu(x))=\frac{\psi\varphi}{2\sqrt{r}}+B_\varphi\sqrt{r}+O(r^{3/2}) \tag*{(16)} q ( x − r ν ( x )) = 2 r ψ φ + B φ r + O ( r 3/2 ) (16)
for some function B φ B_\varphi B φ . The two terms of order r − 1 / 2 r^{-1/2} r − 1/2 cancel, and letting r → 0 r\to0 r → 0 gives
ψ ˙ = B φ − 3 2 κ φ . (17) \dot{\psi}=B_\varphi-\frac{3}{2}\kappa\varphi. \tag*{(17)} ψ ˙ = B φ − 2 3 κ φ . (17)
**Identification of B φ . ∗ ∗ F o r B_\varphi.** For B φ . ∗ ∗ F or \chi\in C^\infty(\overline{\Omega})$ and ∣ t ∣ |t| ∣ t ∣ small, ∫ u t ( − Δ ) 1 / 2 χ = ∫ χ \int u_t(-\Delta)^{1/2}\chi=\int\chi ∫ u t ( − Δ ) 1/2 χ = ∫ χ ; differentiating in t t t shows that q q q is 1 2 \frac{1}{2} 2 1 -harmonic in Ω \Omega Ω . It also vanishes outside Ω ‾ \overline{\Omega} Ω . Let w = ∂ t ( u t ∘ Φ t ) ∣ t = 0 w=\left.\partial_t(u_t\circ\Phi_t)\right|_{t=0} w = ∂ t ( u t ∘ Φ t ) ∣ t = 0 be the material derivative. By the chain rule we have
q = w − X ⋅ ∇ u Ω . q=w-X\cdot\nabla u_\Omega. q = w − X ⋅ ∇ u Ω .
The function w w w vanishes like d \sqrt{d} d on ∂ Ω \partial\Omega ∂ Ω , whereas X ⋅ ∇ u Ω = − ψ φ 2 d − 1 / 2 + O ( d ) X\cdot\nabla u_\Omega=-\frac{\psi\varphi}{2}d^{-1/2}+O(\sqrt{d}) X ⋅ ∇ u Ω = − 2 ψ φ d − 1/2 + O ( d ) by (4) . Thus q q q is a large 1 2 \frac{1}{2} 2 1 -harmonic function with weighted trace ψ φ / 2 \psi\varphi/2 ψ φ /2 . By uniqueness in (6) , q = Q ψ φ / 2 q=Q_{\psi\varphi/2} q = Q ψ φ /2 , which gives (1.16) with
B φ = 1 2 D Ω ( ψ φ ) . (18) B_\varphi=\frac{1}{2}D_\Omega(\psi\varphi). \tag*{(18)} B φ = 2 1 D Ω ( ψ φ ) . (18)
Inserting (1.17) and (1.18) into (1.15) gives (1.9)`.
Two steps of this computation are formal. The identification q = Q ψ φ / 2 q=Q_{\psi\varphi/2} q = Q ψ φ /2 requires some care, because q q q is not in L 2 L^2 L 2 (as soon as φ ≢ 0 \varphi\not\equiv0 φ ≡ 0 ) and therefore lies outside the variational framework in which u t u_t u t is defined; this is done rigorously on the fixed domain Ω \Omega Ω in Proposition 3.3. The genuinely formal step is the derivation of (17) , which interchanges ∂ t \partial_t ∂ t and lim r → 0 \lim_{r\to0} lim r → 0 . Justifying it would require expansions of u t u_t u t and of the large solutions on Ω t \Omega_t Ω t , together with their t t t -derivatives, that are uniform in t t t . The proof below avoids this interchange.
Strategy of the proof
The rigorous proof computes d 2 d t 2 T ( Ω t ) ∣ t = 0 \left.\frac{d^2}{dt^2}T(\Omega_t)\right|_{t=0} d t 2 d 2 T ( Ω t ) t = 0 without ever differentiating ψ t \psi_t ψ t . All steps but one take place on the fixed domain Ω \Omega Ω .
We pull the problem back to Ω \Omega Ω by the diffeomorphisms Φ t = Id + t X φ \Phi_t=\operatorname{Id}+tX_\varphi Φ t = Id + t X φ of (8) , whose velocity is constant along the normals near ∂ Ω \partial\Omega ∂ Ω . The pulled-back states v t = u t ∘ Φ t v_t=u_t\circ\Phi_t v t = u t ∘ Φ t solve a t ( v t , v ) = ℓ t ( v ) a_t(v_t,v)=\ell_t(v) a t ( v t , v ) = ℓ t ( v ) in the fixed space V V V , where a t a_t a t and ℓ t \ell_t ℓ t are obtained from (recall (2) ) a a a and ∫ v \int v ∫ v by a change of variables. From (3) we get the exact identity T ( Ω t ) = G ( t ) + a t ( v t − u Ω , v t − u Ω ) T(\Omega_t)=G(t)+a_t(v_t-u_\Omega,v_t-u_\Omega) T ( Ω t ) = G ( t ) + a t ( v t − u Ω , v t − u Ω ) with G ( t ) = 2 ℓ t ( u Ω ) − a t ( u Ω , u Ω ) G(t)=2\ell_t(u_\Omega)-a_t(u_\Omega,u_\Omega) G ( t ) = 2 ℓ t ( u Ω ) − a t ( u Ω , u Ω ) , and therefore
d 2 d t 2 T ( Ω t ) ∣ t = 0 = G ′ ′ ( 0 ) + 2 a ( w , w ) , \left.\frac{d^2}{dt^2}T(\Omega_t)\right|_{t=0}=G''(0)+2a(w,w), d t 2 d 2 T ( Ω t ) t = 0 = G ′′ ( 0 ) + 2 a ( w , w ) ,
where we recall that w = ∂ t v t ∣ t = 0 w=\left.\partial_tv_t\right|_{t=0} w = ∂ t v t ∣ t = 0 is the material derivative, see Proposition 3.2. The aim is then to rewrite the right-hand side as a boundary integral. We start with the term a ( w , w ) a(w,w) a ( w , w ) . For this we show in Proposition 3.3 that q = w − X ⋅ ∇ u Ω = Q ψ φ / 2 q=w-X\cdot\nabla u_\Omega=Q_{\psi\varphi/2} q = w − X ⋅ ∇ u Ω = Q ψ φ /2 . To prove this equality, we first show that the d − 1 / 2 d^{-1/2} d − 1/2 singularities of q q q and of Q ψ φ / 2 Q_{\psi\varphi/2} Q ψ φ /2 cancel, so that their difference lies in V V V and can be treated by a more standard energy argument. This shows that w = δ W w=\sqrt{\delta}\,W w = δ W with W W W smooth and gives the boundary value of W W W (Lemma 3.4 ). A bilinear Pohozaev identity from [10 ] then turns a ( w , w ) a(w,w) a ( w , w ) into a volume integral plus a boundary integral (Corollary 3.7 ).
It thus remains to compute G ′ ′ ( 0 ) G''(0) G ′′ ( 0 ) . The same Pohozaev identity, applied on Ω t \Omega_t Ω t to the transported function u Ω ∘ Φ t − 1 u_{\Omega}\circ\Phi_t^{-1} u Ω ∘ Φ t − 1 and pulled back to Ω \Omega Ω , gives
G ′ ( t ) = 2 ∫ Ω ( L t u Ω − 1 ) X ^ t ⋅ ∇ u Ω J t + B ( t ) , G'(t)=2\int_{\Omega}(L_tu_{\Omega}-1)\widehat{X}_t\cdot\nabla u_{\Omega}J_t+\mathcal{B}(t), G ′ ( t ) = 2 ∫ Ω ( L t u Ω − 1 ) X t ⋅ ∇ u Ω J t + B ( t ) ,
where L t L_t L t is the pull-back of ( − Δ ) 1 / 2 (-\Delta)^{1/2} ( − Δ ) 1/2 by Φ t \Phi_t Φ t , J t J_t J t the Jacobian, X ^ t \widehat{X}_t X t a transported velocity, and B ( t ) \mathcal{B}(t) B ( t ) an explicit boundary integral (Lemma 3.8). Only the kernel of the operator depends on t t t here. Since L 0 u Ω = 1 L_0u_{\Omega}=1 L 0 u Ω = 1 , computing G ′ ′ ( 0 ) G''(0) G ′′ ( 0 ) amounts to passing to the limit in ( L t u Ω − 1 ) / t (L_tu_{\Omega}-1)/t ( L t u Ω − 1 ) / t , tested against X ⋅ ∇ u Ω X\cdot\nabla u_{\Omega} X ⋅ ∇ u Ω . The distributional limit is easy to identify, but X ⋅ ∇ u Ω X\cdot\nabla u_{\Omega} X ⋅ ∇ u Ω behaves like d − 1 / 2 d^{-1/2} d − 1/2 , so we also need the uniform bound, see Lemma 3.11 ,
∥ L t u Ω − L 0 u Ω ∥ L ∞ ( Ω ) ≤ C ∣ t ∣ , \lVert L_tu_{\Omega}-L_0u_{\Omega}\rVert_{L^\infty(\Omega)}\le C|t|, ∥ L t u Ω − L 0 u Ω ∥ L ∞ ( Ω ) ≤ C ∣ t ∣ ,
which is the main technical estimate of the paper and which may be seen as a simplified version of related estimates from [3 ] . A naive Taylor expansion only gives a bound of order ∣ t ∣ d − 1 / 2 |t|d^{-1/2} ∣ t ∣ d − 1/2 . The proof uses instead that, after the kernel of L t L_t L t is frozen at the evaluation point, the difference L t − L 0 L_t-L_0 L t − L 0 becomes an operator P b \mathcal{P}_b P b with an even angular density b b b , and that such operators annihilate + l + \sqrt{\vphantom{+}l_+} + l + for every affine function l l l (Lemma 2.5). The volume integral in G ′ ′ ( 0 ) G''(0) G ′′ ( 0 ) then cancels exactly the one produced by the Pohozaev identity, and only boundary terms remain.
Organisation of the paper
Organisation of the paper. Section 2 2 2 collects the results on nonlocal equations that we use: boundary regularity, large solutions and the operator D Ω D_{\Omega} D Ω , and the estimate for frozen angular operators. Section 3 3 3 contains the proof of Theorem 1.1 , and Section 4 4 4 the proof of Theorem 1.2 .
Preliminaries on nonlocal equations
Throughout this section Ω ⊂ R n \Omega\subset\mathbb{R}^n Ω ⊂ R n is a bounded open set with C ∞ C^\infty C ∞ boundary, and r 0 r_0 r 0 , δ \delta δ , V V V , a a a are as in the introduction. Functions defined in Ω \Omega Ω are extended by zero to R n \mathbb{R}^n R n without further mention. For a function g ∈ L 1 ( R n ) g\in L^1(\mathbb{R}^n) g ∈ L 1 ( R n ) with compact support, ( − Δ ) 1 / 2 g (-\Delta)^{1/2}g ( − Δ ) 1/2 g is the distribution χ ↦ ∫ g ( − Δ ) 1 / 2 χ \chi\mapsto\int g(-\Delta)^{1/2}\chi χ ↦ ∫ g ( − Δ ) 1/2 χ , which is well defined since ( − Δ ) 1 / 2 χ (-\Delta)^{1/2}\chi ( − Δ ) 1/2 χ is smooth and decays like ∣ x ∣ − n − 1 |x|^{-n-1} ∣ x ∣ − n − 1 for χ ∈ C c ∞ ( R n ) \chi\in C^\infty_c(\mathbb{R}^n) χ ∈ C c ∞ ( R n ) .
The energy space
2.1. The energy space. Since ∂ Ω \partial\Omega ∂ Ω is smooth, C c ∞ ( Ω ) C^\infty_c(\Omega) C c ∞ ( Ω ) is dense in V V V ; this holds for every bounded open set with continuous boundary, see [13 ] . Fix a bounded set A ⊂ R n ∖ Ω A\subset\mathbb{R}^n\setminus\Omega A ⊂ R n ∖ Ω of positive measure at positive distance from Ω \Omega Ω . Since every v ∈ V v\in V v ∈ V vanishes on A A A , keeping only the blocks Ω × A \Omega\times A Ω × A and A × Ω A\times\Omega A × Ω in the definition of a a a gives
a ( v , v ) ≥ c n ∫ Ω ∣ v ( y ) ∣ 2 ( ∫ A d z ∣ y − z ∣ n + 1 ) d y ≥ c − 1 ∥ v ∥ L 2 ( Ω ) 2 . (19) a(v,v)\ge c_n\int_{\Omega}|v(y)|^2\left(\int_A\frac{dz}{|y-z|^{n+1}}\right)dy\ge c^{-1}\lVert v\rVert_{L^2(\Omega)}^2. \tag*{(19)} a ( v , v ) ≥ c n ∫ Ω ∣ v ( y ) ∣ 2 ( ∫ A ∣ y − z ∣ n + 1 d z ) d y ≥ c − 1 ∥ v ∥ L 2 ( Ω ) 2 . (19)
with c c c depending on Ω \Omega Ω and A A A . Hence ∥ v ∥ V : = a ( v , v ) 1 / 2 \lVert v\rVert_V:=a(v,v)^{1/2} ∥ v ∥ V := a ( v , v ) 1/2 is a norm on V V V , equivalent to the H 1 / 2 ( R n ) H^{1/2}(\mathbb{R}^n) H 1/2 ( R n ) norm, and a a a is coercive on V V V .
We shall use the comparison principle in the following form: if v ∈ V v\in V v ∈ V and a ( v , ϕ ) ≥ 0 a(v,\phi)\ge0 a ( v , ϕ ) ≥ 0 for every ϕ ∈ V \phi\in V ϕ ∈ V with ϕ ≥ 0 \phi\ge0 ϕ ≥ 0 , then v ≥ 0 v\ge0 v ≥ 0 . Indeed, testing with ϕ = v − \phi=v^- ϕ = v − and using a ( v + , v − ) ≤ 0 a(v^+,v^-)\le0 a ( v + , v − ) ≤ 0 , which holds because ( v + ( y ) − v + ( z ) ) ( v − ( y ) − v − ( z ) ) ≤ 0 (v^+(y)-v^+(z))(v^-(y)-v^-(z))\le0 ( v + ( y ) − v + ( z )) ( v − ( y ) − v − ( z )) ≤ 0 pointwise, we get $0\le a(v,v^-)= a ( v + , v − ) − a ( v − , v − ) ≤ − a ( v − , v − ) a(v^{+},v^{-})-a(v^{-},v^{-})\le-a(v^{-},v^{-}) a ( v + , v − ) − a ( v − , v − ) ≤ − a ( v − , v − ) , so v − = 0 v^{-}=0 v − = 0 . In particular u ≥ 0 u\ge0 u ≥ 0 , and comparison with the torsion function of a ball containing Ω \Omega Ω shows that u u u is bounded.
Boundary regularity
Lemma 2.1 below is the boundary regularity input of the paper. It follows from [3 Theorem 1.4] and [3 Theorem 2.2 and Corollary 2.3] .
Lemma 2.1. For every v ∈ V v\in V v ∈ V , f ∈ C ∞ ( Ω ‾ ) f\in C^{\infty}(\overline{\Omega}) f ∈ C ∞ ( Ω ) and A ∈ C ∞ ( Ω ‾ ) A\in C^{\infty}(\overline{\Omega}) A ∈ C ∞ ( Ω ) ,
( − Δ ) 1 / 2 v = f in Ω ⟹ v / δ ∈ C ∞ ( Ω ‾ ) , (-\Delta)^{1/2}v=f\ \text{ in }\Omega\quad\Longrightarrow\quad v/\sqrt{\delta}\in C^{\infty}(\overline{\Omega}), ( − Δ ) 1/2 v = f in Ω ⟹ v / δ ∈ C ∞ ( Ω ) , ( − Δ ) 1 / 2 ( δ A ) ∈ C ∞ ( Ω ‾ ) . (20) (-\Delta)^{1/2}(\sqrt{\delta}A)\in C^{\infty}(\overline{\Omega}). \tag*{(20)} ( − Δ ) 1/2 ( δ A ) ∈ C ∞ ( Ω ) . (20)
We shall also need the following energy membership property which essentially follows from the well-known fact that the function x ↦ ( x n ) + 1 / 2 x\mapsto(x_n)_{+}^{1/2} x ↦ ( x n ) + 1/2 belongs to H l o c 1 / 2 ( R n ) H_{\mathrm{loc}}^{1/2}(\mathbb{R}^{n}) H loc 1/2 ( R n ) , see e.g. [21 Lemmas 3.1 and 3.2] , together with a localization argument.
Lemma 2.2. If A ∈ C 1 ( Ω ‾ ) A\in C^{1}(\overline{\Omega}) A ∈ C 1 ( Ω ) , then δ A ∈ V \sqrt{\delta}A\in V δ A ∈ V .
Singular lifting
To construct large 1 / 2 1/2 1/2 -harmonic functions, we first write down an explicit function with the prescribed singularity f d − 1 / 2 f\,d^{-1/2} f d − 1/2 and check that its image under ( − Δ ) 1 / 2 (-\Delta)^{1/2} ( − Δ ) 1/2 is smooth up to the boundary. The remaining correction is then found in the energy space V V V .
Lemma 2.3. Let f ∈ C ∞ ( ∂ Ω ) f\in C^{\infty}(\partial\Omega) f ∈ C ∞ ( ∂ Ω ) and let F ∈ C ∞ ( Ω ‾ ) F\in C^{\infty}(\overline{\Omega}) F ∈ C ∞ ( Ω ) be supported in { d < r 0 } \{d<r_{0}\} { d < r 0 } , with F ( x − r ν ( x ) ) = f ( x ) F(x-r\nu(x))=f(x) F ( x − r ν ( x )) = f ( x ) for x ∈ ∂ Ω x\in\partial\Omega x ∈ ∂ Ω and 0 ≤ r < r 0 / 2 0\le r<r_{0}/2 0 ≤ r < r 0 /2 . Let ζ f = δ − 1 / 2 F \zeta_f=\delta^{-1/2}F ζ f = δ − 1/2 F in Ω \Omega Ω and ζ f = 0 \zeta_f=0 ζ f = 0 outside Ω \Omega Ω . Then ζ f ∈ L 1 ( R n ) \zeta_f\in L^{1}(\mathbb{R}^{n}) ζ f ∈ L 1 ( R n ) and ( − Δ ) 1 / 2 ζ f ∈ C ∞ ( Ω ‾ ) (-\Delta)^{1/2}\zeta_f\in C^{\infty}(\overline{\Omega}) ( − Δ ) 1/2 ζ f ∈ C ∞ ( Ω ) .
Proof. The function ζ f \zeta_f ζ f is integrable because r − 1 / 2 r^{-1/2} r − 1/2 is integrable near 0. Let A = 2 F ∇ δ A=2F\nabla\delta A = 2 F ∇ δ , then
ζ f = div ( δ A ) − δ div A \zeta_f=\operatorname{div}(\sqrt{\delta}A)-\sqrt{\delta}\operatorname{div}A ζ f = div ( δ A ) − δ div A
in the sense of distributions on R n \mathbb{R}^{n} R n . The divergence produces no measure on ∂ Ω \partial\Omega ∂ Ω , because δ A \sqrt{\delta}A δ A belongs to W 1 , 1 ( R n ) W^{1,1}(\mathbb{R}^{n}) W 1 , 1 ( R n ) (its gradient is O ( d − 1 / 2 ) O(d^{-1/2}) O ( d − 1/2 ) ) and vanishes on ∂ Ω \partial\Omega ∂ Ω . Since ( − Δ ) 1 / 2 (-\Delta)^{1/2} ( − Δ ) 1/2 commutes with derivatives,
( − Δ ) 1 / 2 ζ f = ∑ i ∂ i ( − Δ ) 1 / 2 ( δ A i ) − ( − Δ ) 1 / 2 ( δ div A ) , (-\Delta)^{1/2}\zeta_f=\sum_i\partial_i(-\Delta)^{1/2}(\sqrt{\delta}A_i)-(-\Delta)^{1/2}(\sqrt{\delta}\operatorname{div}A), ( − Δ ) 1/2 ζ f = i ∑ ∂ i ( − Δ ) 1/2 ( δ A i ) − ( − Δ ) 1/2 ( δ div A ) ,
and the second line of (20) shows that the restriction of the right-hand side to Ω \Omega Ω belongs to C ∞ ( Ω ‾ ) C^{\infty}(\overline{\Omega}) C ∞ ( Ω ) . □
The Dirichlet-to-Neumann operator D Ω D_{\Omega} D Ω
Proposition 2.4. For every f ∈ C ∞ ( ∂ Ω ) f\in C^{\infty}(\partial\Omega) f ∈ C ∞ ( ∂ Ω ) there is a unique function Q f Q_f Q f , vanishing outside Ω \Omega Ω , with δ Q f ∈ C ∞ ( Ω ‾ ) \sqrt{\delta}Q_f\in C^{\infty}(\overline{\Omega}) δ Q f ∈ C ∞ ( Ω ) and satisfying (6) . With ζ f \zeta_f ζ f as in Lemma 2.3, Q f = ζ f + ω f Q_f=\zeta_f+\omega_f Q f = ζ f + ω f , where ω f ∈ V \omega_f\in V ω f ∈ V solves ( − Δ ) 1 / 2 ω f = − ( − Δ ) 1 / 2 ζ f (-\Delta)^{1/2}\omega_f=-(-\Delta)^{1/2}\zeta_f ( − Δ ) 1/2 ω f = − ( − Δ ) 1/2 ζ f in Ω \Omega Ω , and
D Ω f = ( ω f / δ ) ∣ ∂ Ω . (21) D_{\Omega}f=\left(\omega_f/\sqrt{\delta}\right)\big|_{\partial\Omega}. \tag*{(21)} D Ω f = ( ω f / δ ) ∂ Ω . (21) Moreover Q f ( x − r ν ( x ) ) = f ( x ) r − 1 / 2 + D Ω f ( x ) r + O ( r 3 / 2 ) Q_f(x-r\nu(x))=f(x)r^{-1/2}+D_{\Omega}f(x)\sqrt{r}+O(r^{3/2}) Q f ( x − r ν ( x )) = f ( x ) r − 1/2 + D Ω f ( x ) r + O ( r 3/2 ) uniformly in x ∈ ∂ Ω x\in\partial\Omega x ∈ ∂ Ω . Uniqueness holds in the larger class δ Q ∈ C 2 ( Ω ‾ ) \sqrt{\delta}Q\in C^{2}(\overline{\Omega}) δ Q ∈ C 2 ( Ω ) ; in particular Q f Q_f Q f does not depend on the choice of F F F .
Proof. Existence. Let g f = ( − Δ ) 1 / 2 ζ f g_f=(-\Delta)^{1/2}\zeta_f g f = ( − Δ ) 1/2 ζ f , which belongs to C ∞ ( Ω ‾ ) C^{\infty}(\overline{\Omega}) C ∞ ( Ω ) by Lemma 2.3. By (19) , v ↦ ∫ Ω g f v v\mapsto\int_{\Omega}g_fv v ↦ ∫ Ω g f v is a bounded linear form on V V V , and the Lax–Milgram theorem gives a unique ω f ∈ V \omega_f\in V ω f ∈ V with a ( ω f , v ) = − ∫ Ω g f v a(\omega_f,v)=-\int_{\Omega}g_fv a ( ω f , v ) = − ∫ Ω g f v for every v ∈ V v\in V v ∈ V . Then Q f = ζ f + ω f Q_f=\zeta_f+\omega_f Q f = ζ f + ω f vanishes outside Ω \Omega Ω and ( − Δ ) 1 / 2 Q f = 0 (-\Delta)^{1/2}Q_f=0 ( − Δ ) 1/2 Q f = 0 in $\Omega.
Boundary behaviour. The first line of (20) gives δ h f \sqrt{\delta}\,h_f δ h f with h f ∈ C ∞ ( Ω ‾ ) h_f \in C^\infty(\overline{\Omega}) h f ∈ C ∞ ( Ω ) , hence
δ Q f = F + δ h f . \sqrt{\delta}\,Q_f = F+\delta h_f. δ Q f = F + δ h f .
This shows that δ Q f ∈ C ∞ ( Ω ‾ ) \sqrt{\delta}\,Q_f \in C^\infty(\overline{\Omega}) δ Q f ∈ C ∞ ( Ω ) and that its boundary value is f f f . For x ∈ ∂ Ω x \in\partial\Omega x ∈ ∂ Ω and small r > 0 r>0 r > 0 we have δ ( x − r ν ( x ) ) = r \delta(x-r\nu(x))=r δ ( x − r ν ( x )) = r and F ( x − r ν ( x ) ) = f ( x ) F(x-r\nu(x))=f(x) F ( x − r ν ( x )) = f ( x ) , so
r Q f ( x − r ν ( x ) ) = f ( x ) + r h f ( x − r ν ( x ) ) . \sqrt{r}\,Q_f(x-r\nu(x))=f(x)+r h_f(x-r\nu(x)). r Q f ( x − r ν ( x )) = f ( x ) + r h f ( x − r ν ( x )) .
Differentiating at r = 0 r=0 r = 0 gives D Ω f = h f ∣ ∂ Ω D_\Omega f=h_f|_{\partial\Omega} D Ω f = h f ∣ ∂ Ω , which is (21) ; here it matters that F F F is constant along normals, so that no derivative of F F F appears. The expansion of Q f Q_f Q f follows from the same identity and the smoothness of h f h_f h f .
Uniqueness. Let Q 1 , Q 2 Q_1,Q_2 Q 1 , Q 2 be two solutions of (6) with the same datum f f f and with δ Q i ∈ C 2 ( Ω ‾ ) \sqrt{\delta}\,Q_i\in C^2(\overline{\Omega}) δ Q i ∈ C 2 ( Ω ) . Set R = Q 1 − Q 2 R=Q_1-Q_2 R = Q 1 − Q 2 and P = δ R ∈ C 2 ( Ω ‾ ) P=\sqrt{\delta}\,R\in C^2(\overline{\Omega}) P = δ R ∈ C 2 ( Ω ) , so that P = 0 P=0 P = 0 on ∂ Ω \partial\Omega ∂ Ω . Near ∂ Ω \partial\Omega ∂ Ω , where r = δ r=\delta r = δ is a coordinate, P ( z , r ) = r A ( z , r ) P(z,r)=rA(z,r) P ( z , r ) = r A ( z , r ) with A ( z , r ) = ∫ 0 1 ∂ r P ( z , θ r ) d θ A(z,r)=\int_0^1\partial_rP(z,\theta r)\,d\theta A ( z , r ) = ∫ 0 1 ∂ r P ( z , θ r ) d θ , and A ∈ C 1 A\in C^1 A ∈ C 1 because ∂ r P ∈ C 1 \partial_rP\in C^1 ∂ r P ∈ C 1 . Away from the boundary δ \delta δ is bounded below. Hence P = δ A P=\delta A P = δ A with A ∈ C 1 ( Ω ‾ ) A\in C^1(\overline{\Omega}) A ∈ C 1 ( Ω ) , and R = δ A R=\sqrt{\delta}\,A R = δ A belongs to V V V by Lemma 2.2. Since R R R is 1 2 \frac12 2 1 -harmonic in Ω \Omega Ω , a ( R , ϕ ) = 0 a(R,\phi)=0 a ( R , ϕ ) = 0 for ϕ ∈ C ∞ ( Ω ‾ ) \phi\in C^\infty(\overline{\Omega}) ϕ ∈ C ∞ ( Ω ) , hence for every ϕ ∈ V \phi\in V ϕ ∈ V by density. Taking ϕ = R \phi=R ϕ = R gives R = 0 R=0 R = 0 . Two different choices of F F F give two solutions of (6) , which therefore coincide. □
In particular D Ω f D_\Omega f D Ω f is computed by solving one ordinary Dirichlet problem, for ω f \omega_f ω f , and then using (21) .
Angular operators
For an even function b ∈ L ∞ ( S n − 1 ) b\in L^\infty(\mathbb{S}^{n-1}) b ∈ L ∞ ( S n − 1 ) , that is b ( − θ ) = b ( θ ) b(-\theta)=b(\theta) b ( − θ ) = b ( θ ) , we consider the operator
P b v ( y ) : = c n P.V. ∫ R n ( v ( y ) − v ( z ) ) b ( ( z − y ) / ∣ z − y ∣ ) ∣ z − y ∣ n + 1 d z . \mathcal{P}_b v(y):=c_n\,\operatorname{P.V.}\int_{\mathbb{R}^n}\frac{(v(y)-v(z))b((z-y)/|z-y|)}{|z-y|^{n+1}}\,dz. P b v ( y ) := c n P.V. ∫ R n ∣ z − y ∣ n + 1 ( v ( y ) − v ( z )) b (( z − y ) /∣ z − y ∣ ) d z .
For b ≡ 1 b\equiv1 b ≡ 1 this is ( − Δ ) 1 / 2 (-\Delta)^{1/2} ( − Δ ) 1/2 . Passing to polar coordinates around y y y and pairing the directions θ \theta θ and − θ -\theta − θ , which is possible since b b b is even, we obtain
P b v ( y ) = c n 2 ∫ S n − 1 b ( θ ) ∫ 0 ∞ 2 v ( y ) − v ( y + ρ θ ) − v ( y − ρ θ ) ρ 2 d ρ d θ . (22) \mathcal{P}_b v(y)=\frac{c_n}{2}\int_{\mathbb{S}^{n-1}}b(\theta)\int_0^\infty\frac{2v(y)-v(y+\rho\theta)-v(y-\rho\theta)}{\rho^2}\,d\rho\,d\theta. \tag*{(22)} P b v ( y ) = 2 c n ∫ S n − 1 b ( θ ) ∫ 0 ∞ ρ 2 2 v ( y ) − v ( y + ρθ ) − v ( y − ρθ ) d ρ d θ . (22)
Notice that as opposed to the original definition of P b \mathcal{P}_b P b this integral will always be absolutely convergent in the cases we are interested in.
The following lemma is the key estimate for the transported forcing in § 2.12 . It may be seen as a simplified version of [3 Corollary 2.3] . We include a proof because it is elementary, does not require b b b to be positive and uses only its boundedness.
Lemma 2.5. Let U ∈ C ∞ ( Ω ‾ ) U\in C^\infty(\overline{\Omega}) U ∈ C ∞ ( Ω ) and u = δ U u=\sqrt{\delta}\,U u = δ U . For every bounded even function b : S n − 1 → R b:\mathbb{S}^{n-1}\to\mathbb{R} b : S n − 1 → R ,
∥ P b u ∥ L ∞ ( Ω ) ≤ C ∥ b ∥ L ∞ ( S n − 1 ) , \|\mathcal{P}_b u\|_{L^\infty(\Omega)}\le C\|b\|_{L^\infty(\mathbb{S}^{n-1})}, ∥ P b u ∥ L ∞ ( Ω ) ≤ C ∥ b ∥ L ∞ ( S n − 1 ) , where C C C depends only on Ω \Omega Ω , δ \delta δ and ∥ U ∥ C 2 ( Ω ‾ ) \|U\|_{C^2(\overline{\Omega})} ∥ U ∥ C 2 ( Ω ) .
Proof. Let d d d be the signed distance to ∂ Ω \partial\Omega ∂ Ω , positive in Ω \Omega Ω , which is smooth on { ∣ d ∣ < 2 r 0 } \{|d|<2r_0\} { ∣ d ∣ < 2 r 0 } , and extend U U U smoothly to a neighbourhood of Ω ‾ \overline{\Omega} Ω . Let C Ω C_\Omega C Ω be a bound for D 2 d D^2d D 2 d on { ∣ d ∣ < 2 r 0 } \{|d|<2r_0\} { ∣ d ∣ < 2 r 0 } , and fix R 0 ∈ ( 0 , r 0 / 5 ] R_0\in(0,r_0/5] R 0 ∈ ( 0 , r 0 /5 ] with 16 C Ω R 0 ≤ 1 16C_\Omega R_0\le1 16 C Ω R 0 ≤ 1 . Fix y ∈ Ω y\in\Omega y ∈ Ω . We will treat separately the cases d ( y ) ≥ R 0 d(y)\ge R_0 d ( y ) ≥ R 0 and d ( y ) < R 0 d(y)<R_0 d ( y ) < R 0 .
We start with the easier case d ( y ) ≥ R 0 d(y)\ge R_0 d ( y ) ≥ R 0 . Since d d d is 1-Lipschitz, B R 0 / 2 ( y ) B_{R_0/2}(y) B R 0 /2 ( y ) is contained in the compact set K = { z ∈ Ω ‾ : d ( z ) ≥ R 0 / 2 } K=\{z\in\overline{\Omega}:d(z)\ge R_0/2\} K = { z ∈ Ω : d ( z ) ≥ R 0 /2 } , on which δ \delta δ is bounded below by a positive constant. Hence M : = sup K ∣ D 2 u ∣ < ∞ M:=\sup_K|D^2u|<\infty M := sup K ∣ D 2 u ∣ < ∞ . For 0 < ρ < R 0 / 2 0<\rho<R_0/2 0 < ρ < R 0 /2 , Taylor’s formula applied to @page 8/21 starts_mid=0 ends_mid=1 Boundary behaviour. The first line of (20) gives δ h f \sqrt{\delta}\,h_f δ h f with h f ∈ C ∞ ( Ω ‾ ) h_f \in C^\infty(\overline{\Omega}) h f ∈ C ∞ ( Ω ) , hence
δ Q f = F + δ h f . \sqrt{\delta}\,Q_f = F+\delta h_f. δ Q f = F + δ h f .
This shows that δ Q f ∈ C ∞ ( Ω ‾ ) \sqrt{\delta}\,Q_f \in C^\infty(\overline{\Omega}) δ Q f ∈ C ∞ ( Ω ) and that its boundary value is f f f . For x ∈ ∂ Ω x \in\partial\Omega x ∈ ∂ Ω and small r > 0 r>0 r > 0 we have δ ( x − r ν ( x ) ) = r \delta(x-r\nu(x))=r δ ( x − r ν ( x )) = r and F ( x − r ν ( x ) ) = f ( x ) F(x-r\nu(x))=f(x) F ( x − r ν ( x )) = f ( x ) , so
r Q f ( x − r ν ( x ) ) = f ( x ) + r h f ( x − r ν ( x ) ) . \sqrt{r}\,Q_f(x-r\nu(x))=f(x)+r h_f(x-r\nu(x)). r Q f ( x − r ν ( x )) = f ( x ) + r h f ( x − r ν ( x )) .
Differentiating at r = 0 r=0 r = 0 gives D Ω f = h f ∣ ∂ Ω D_\Omega f=h_f|_{\partial\Omega} D Ω f = h f ∣ ∂ Ω , which is (21) ; here it matters that F F F is constant along normals, so that no derivative of F F F appears. The expansion of Q f Q_f Q f follows from the same identity and the smoothness of h f h_f h f .
Uniqueness. Let Q 1 , Q 2 Q_1,Q_2 Q 1 , Q 2 be two solutions of (6) with the same datum f f f and with δ Q i ∈ C 2 ( Ω ‾ ) \sqrt{\delta}\,Q_i\in C^2(\overline{\Omega}) δ Q i ∈ C 2 ( Ω ) . Set R = Q 1 − Q 2 R=Q_1-Q_2 R = Q 1 − Q 2 and P = δ R ∈ C 2 ( Ω ‾ ) P=\sqrt{\delta}\,R\in C^2(\overline{\Omega}) P = δ R ∈ C 2 ( Ω ) , so that P = 0 P=0 P = 0 on ∂ Ω \partial\Omega ∂ Ω . Near ∂ Ω \partial\Omega ∂ Ω , where r = δ r=\delta r = δ is a coordinate, P ( z , r ) = r A ( z , r ) P(z,r)=rA(z,r) P ( z , r ) = r A ( z , r ) with A ( z , r ) = ∫ 0 1 ∂ r P ( z , θ r ) d θ A(z,r)=\int_0^1\partial_rP(z,\theta r)\,d\theta A ( z , r ) = ∫ 0 1 ∂ r P ( z , θ r ) d θ , and A ∈ C 1 A\in C^1 A ∈ C 1 because ∂ r P ∈ C 1 \partial_rP\in C^1 ∂ r P ∈ C 1 . Away from the boundary δ \delta δ is bounded below. Hence P = δ A P=\delta A P = δ A with A ∈ C 1 ( Ω ‾ ) A\in C^1(\overline{\Omega}) A ∈ C 1 ( Ω ) , and R = δ A R=\sqrt{\delta}\,A R = δ A belongs to V V V by Lemma 2.2. Since R R R is 1 2 \frac12 2 1 -harmonic in Ω \Omega Ω , a ( R , ϕ ) = 0 a(R,\phi)=0 a ( R , ϕ ) = 0 for ϕ ∈ C ∞ ( Ω ‾ ) \phi\in C^\infty(\overline{\Omega}) ϕ ∈ C ∞ ( Ω ) , hence for every ϕ ∈ V \phi\in V ϕ ∈ V by density. Taking ϕ = R \phi=R ϕ = R gives R = 0 R=0 R = 0 . Two different choices of F F F give two solutions of (6) , which therefore coincide. □
In particular D Ω f D_\Omega f D Ω f is computed by solving one ordinary Dirichlet problem, for ω f \omega_f ω f , and then using (21) .
Angular operators
For an even function b ∈ L ∞ ( S n − 1 ) b\in L^\infty(\mathbb{S}^{n-1}) b ∈ L ∞ ( S n − 1 ) , that is b ( − θ ) = b ( θ ) b(-\theta)=b(\theta) b ( − θ ) = b ( θ ) , we consider the operator
P b v ( y ) : = c n P.V. ∫ R n ( v ( y ) − v ( z ) ) b ( ( z − y ) / ∣ z − y ∣ ) ∣ z − y ∣ n + 1 d z . \mathcal{P}_b v(y):=c_n\,\operatorname{P.V.}\int_{\mathbb{R}^n}\frac{(v(y)-v(z))b((z-y)/|z-y|)}{|z-y|^{n+1}}\,dz. P b v ( y ) := c n P.V. ∫ R n ∣ z − y ∣ n + 1 ( v ( y ) − v ( z )) b (( z − y ) /∣ z − y ∣ ) d z .
For b ≡ 1 b\equiv1 b ≡ 1 this is ( − Δ ) 1 / 2 (-\Delta)^{1/2} ( − Δ ) 1/2 . Passing to polar coordinates around y y y and pairing the directions θ \theta θ and − θ -\theta − θ , which is possible since b b b is even, we obtain
P b v ( y ) = c n 2 ∫ S n − 1 b ( θ ) ∫ 0 ∞ 2 v ( y ) − v ( y + ρ θ ) − v ( y − ρ θ ) ρ 2 d ρ d θ . (23) \mathcal{P}_b v(y)=\frac{c_n}{2}\int_{\mathbb{S}^{n-1}}b(\theta)\int_0^\infty\frac{2v(y)-v(y+\rho\theta)-v(y-\rho\theta)}{\rho^2}\,d\rho\,d\theta. \tag*{(23)} P b v ( y ) = 2 c n ∫ S n − 1 b ( θ ) ∫ 0 ∞ ρ 2 2 v ( y ) − v ( y + ρθ ) − v ( y − ρθ ) d ρ d θ . (23)
Notice that as opposed to the original definition of P b \mathcal{P}_b P b this integral will always be absolutely convergent in the cases we are interested in.
The following lemma is the key estimate for the transported forcing in § 2.12 . It may be seen as a simplified version of [3 Corollary 2.3] . We include a proof because it is elementary, does not require b b b to be positive and uses only its boundedness.
Lemma 2.5. Let U ∈ C ∞ ( Ω ‾ ) U\in C^\infty(\overline{\Omega}) U ∈ C ∞ ( Ω ) and u = δ U u=\sqrt{\delta}\,U u = δ U . For every bounded even function b : S n − 1 → R b:\mathbb{S}^{n-1}\to\mathbb{R} b : S n − 1 → R ,
∥ P b u ∥ L ∞ ( Ω ) ≤ C ∥ b ∥ L ∞ ( S n − 1 ) , \|\mathcal{P}_b u\|_{L^\infty(\Omega)}\le C\|b\|_{L^\infty(\mathbb{S}^{n-1})}, ∥ P b u ∥ L ∞ ( Ω ) ≤ C ∥ b ∥ L ∞ ( S n − 1 ) , where C C C depends only on Ω \Omega Ω , δ \delta δ and ∥ U ∥ C 2 ( Ω ‾ ) \|U\|_{C^2(\overline{\Omega})} ∥ U ∥ C 2 ( Ω ) .
Proof. Let d \mathbf{d} d be the signed distance to ∂ Ω \partial\Omega ∂ Ω , positive in Ω \Omega Ω , which is smooth on { ∣ d ∣ < 2 r 0 } \{|d|<2r_0\} { ∣ d ∣ < 2 r 0 } , and extend U U U smoothly to a neighbourhood of Ω ‾ \overline{\Omega} Ω . Let C Ω C_\Omega C Ω be a bound for D 2 d D^2d D 2 d on { ∣ d ∣ < 2 r 0 } \{|d|<2r_0\} { ∣ d ∣ < 2 r 0 } , and fix R 0 ∈ ( 0 , r 0 / 5 ] R_0\in(0,r_0/5] R 0 ∈ ( 0 , r 0 /5 ] with 16 C Ω R 0 ≤ 1 16C_\Omega R_0\le1 16 C Ω R 0 ≤ 1 . Fix y ∈ Ω y\in\Omega y ∈ Ω . We will treat separately the cases d ( y ) ≥ R 0 d(y)\ge R_0 d ( y ) ≥ R 0 and d ( y ) < R 0 d(y)<R_0 d ( y ) < R 0 .
We start with the easier case d ( y ) ≥ R 0 d(y)\ge R_0 d ( y ) ≥ R 0 . Since d d d is 1-Lipschitz, B R 0 / 2 ( y ) B_{R_0/2}(y) B R 0 /2 ( y ) is contained in the compact set K = { z ∈ Ω ‾ : d ( z ) ≥ R 0 / 2 } K=\{z\in\overline{\Omega}:d(z)\ge R_0/2\} K = { z ∈ Ω : d ( z ) ≥ R 0 /2 } , on which δ \delta δ is bounded below by a positive constant. Hence M : = sup K ∣ D 2 u ∣ < ∞ M:=\sup_K|D^2u|<\infty M := sup K ∣ D 2 u ∣ < ∞ . For 0 < ρ < R 0 / 2 0<\rho<R_0/2 0 < ρ < R 0 /2 , Taylor’s formula applied to ρ ↦ u ( y + ρ θ ) + u ( y − ρ θ ) \rho\mapsto u(y+\rho\theta)+u(y-\rho\theta) ρ ↦ u ( y + ρθ ) + u ( y − ρθ ) , whose derivative vanishes at ρ = 0 \rho=0 ρ = 0 , gives ∣ 2 u ( y ) − u ( y + ρ θ ) − u ( y − ρ θ ) ∣ ≤ M ρ 2 |2u(y)-u(y+\rho\theta)-u(y-\rho\theta)|\le M\rho^2 ∣2 u ( y ) − u ( y + ρθ ) − u ( y − ρθ ) ∣ ≤ M ρ 2 . For ρ ≥ R 0 / 2 \rho\ge R_0/2 ρ ≥ R 0 /2 the numerator is bounded by 4 ∥ u ∥ ∞ 4\|u\|_\infty 4∥ u ∥ ∞ . Hence
∣ P b u ( y ) ∣ ≤ c n 2 ∣ S n − 1 ∣ ∥ b ∥ ∞ ( M R 0 2 + 8 ∥ u ∥ ∞ R 0 ) ≤ C ∥ b ∥ ∞ . |\mathcal{P}_b u(y)|\le\frac{c_n}{2}|S^{n-1}|\|b\|_\infty\left(\frac{MR_0}{2}+\frac{8\|u\|_\infty}{R_0}\right)\le C\|b\|_\infty. ∣ P b u ( y ) ∣ ≤ 2 c n ∣ S n − 1 ∣∥ b ∥ ∞ ( 2 M R 0 + R 0 8∥ u ∥ ∞ ) ≤ C ∥ b ∥ ∞ .
If instead d ( y ) < R 0 d(y)<R_0 d ( y ) < R 0 , then y = x − d ( y ) ν ( x ) y=x-d(y)\nu(x) y = x − d ( y ) ν ( x ) for a unique x ∈ ∂ Ω x\in\partial\Omega x ∈ ∂ Ω ; d d d and U U U are smooth on B 5 R 0 ( x ) B_{5R_0}(x) B 5 R 0 ( x ) , and u = U d + u=U\sqrt{d_+} u = U d + there. We first prove that
P b ( ( l ) + 1 / 2 ) ( y ) = 0 whenever l is affine and l ( y ) > 0 , (24) \mathcal{P}_b\bigl((l)_+^{1/2}\bigr)(y)=0\qquad\text{whenever } l \text{ is affine and } l(y)>0, \tag*{(24)} P b ( ( l ) + 1/2 ) ( y ) = 0 whenever l is affine and l ( y ) > 0 , (24)
Let thus l l l be an affine function with l ( y ) = r > 0 l(y)=r>0 l ( y ) = r > 0 , and let v = ( l ) + 1 / 2 v=(l)_+^{1/2} v = ( l ) + 1/2 . Fix θ ∈ S n − 1 \theta\in S^{n-1} θ ∈ S n − 1 and put α = ∇ l ⋅ θ \alpha=\nabla l\cdot\theta α = ∇ l ⋅ θ . Since l ( y ± ρ θ ) = r ± α ρ l(y\pm\rho\theta)=r\pm\alpha\rho l ( y ± ρθ ) = r ± α ρ , the inner integral in (2.4) is
I ( r , α ) : = ∫ 0 ∞ 2 r − ( r + α ρ ) + − ( r − α ρ ) + ρ 2 d ρ . I(r,\alpha):=\int_0^\infty\frac{2\sqrt r-\sqrt{(r+\alpha\rho)_+}-\sqrt{(r-\alpha\rho)_+}}{\rho^2}\,d\rho. I ( r , α ) := ∫ 0 ∞ ρ 2 2 r − ( r + α ρ ) + − ( r − α ρ ) + d ρ .
We claim that it converges absolutely and that
I ( r , α ) = 0 for all r > 0 , α ∈ R . (25) I(r,\alpha)=0\qquad\text{for all } r>0,\ \alpha\in\mathbb{R}. \tag*{(25)} I ( r , α ) = 0 for all r > 0 , α ∈ R . (25)
After integration this would prove the claimed (24) . The integrand is even in α \alpha α and vanishes for α = 0 \alpha=0 α = 0 . For α ≠ 0 \alpha\ne0 α = 0 , the substitution ρ = r s / ∣ α ∣ \rho=rs/|\alpha| ρ = r s /∣ α ∣ gives
I ( r , α ) = ∣ α ∣ r I 0 , I 0 : = ∫ 0 ∞ g ( s ) s 2 d s , g ( s ) : = 2 − 1 + s − ( 1 − s ) + . I(r,\alpha)=\frac{|\alpha|}{\sqrt r}I_0,\qquad I_0:=\int_0^\infty\frac{g(s)}{s^2}\,ds,\qquad g(s):=2-\sqrt{1+s}-\sqrt{(1-s)_+}. I ( r , α ) = r ∣ α ∣ I 0 , I 0 := ∫ 0 ∞ s 2 g ( s ) d s , g ( s ) := 2 − 1 + s − ( 1 − s ) + .
Clearly ∣ I 0 ∣ < ∞ |I_0|<\infty ∣ I 0 ∣ < ∞ and we only need to prove that it is zero. Let
F − ( s ) = log 1 − 1 − s 1 + 1 − s ( 0 < s < 1 ) , F + ( s ) = log 1 + s − 1 1 + s + 1 ( s > 0 ) . F_-(s)=\log\frac{1-\sqrt{1-s}}{1+\sqrt{1-s}}\quad(0<s<1),\qquad F_+(s)=\log\frac{\sqrt{1+s}-1}{\sqrt{1+s}+1}\quad(s>0). F − ( s ) = log 1 + 1 − s 1 − 1 − s ( 0 < s < 1 ) , F + ( s ) = log 1 + s + 1 1 + s − 1 ( s > 0 ) .
Integrating by parts it may be checked that
2 I 0 = ∫ 0 1 ( F − ′ − F + ′ ) − ∫ 1 ∞ F + ′ . 2I_0=\int_0^1(F_-' - F_+')-\int_1^\infty F_+'. 2 I 0 = ∫ 0 1 ( F − ′ − F + ′ ) − ∫ 1 ∞ F + ′ .
Since ( F − − F + ) ( 0 ) = 0 (F_--F_+)(0)=0 ( F − − F + ) ( 0 ) = 0 , F − ( 1 ) = 0 F_-(1)=0 F − ( 1 ) = 0 and F + → 0 F_+\to0 F + → 0 at infinity, this concludes the proof of I 0 = 0 I_0=0 I 0 = 0 and thus of (25) .
Recall now that y = x − r ν ( x ) y=x-r\nu(x) y = x − r ν ( x ) with 0 < r < R 0 0<r<R_0 0 < r < R 0 , and set
l y ( z ) = − ( z − x ) ⋅ ν ( x ) , h y ( z ) = U ( y ) ( l y ( z ) ) + , e y = u − h y . l_y(z)=-(z-x)\cdot\nu(x),\qquad h_y(z)=U(y)\sqrt{(l_y(z))_+},\qquad e_y=u-h_y. l y ( z ) = − ( z − x ) ⋅ ν ( x ) , h y ( z ) = U ( y ) ( l y ( z ) ) + , e y = u − h y .
Then l y l_y l y is affine, l y ( y ) = r l_y(y)=r l y ( y ) = r , and u ( y ) = U ( y ) r = h y ( y ) u(y)=U(y)\sqrt r=h_y(y) u ( y ) = U ( y ) r = h y ( y ) , so e y ( y ) = 0 e_y(y)=0 e y ( y ) = 0 . By (24) , P b h y ( y ) = 0 \mathcal{P}_b h_y(y)=0 P b h y ( y ) = 0 , hence by linearity, P b u ( y ) = P b e y ( y ) \mathcal{P}_b u(y)=\mathcal{P}_b e_y(y) P b u ( y ) = P b e y ( y ) so that it is enough to bound P b e y ( y ) \mathcal{P}_b e_y(y) P b e y ( y ) . We bound separately the contributions to (2.4) of ∣ z − y ∣ < r / 2 |z-y|<r/2 ∣ z − y ∣ < r /2 , of r / 2 ≤ ∣ z − y ∣ < R 0 r/2\le|z-y|<R_0 r /2 ≤ ∣ z − y ∣ < R 0 , and of ∣ z − y ∣ ≥ R 0 |z-y|\ge R_0 ∣ z − y ∣ ≥ R 0 .
Before doing so we record that since d ( x ) = 0 d(x)=0 d ( x ) = 0 and ∇ d ( x ) = − ν ( x ) = ∇ l y \nabla d(x)=-\nu(x)=\nabla l_y ∇ d ( x ) = − ν ( x ) = ∇ l y , Taylor’s formula gives, for z ∈ B 5 R 0 ( x ) z\in B_{5R_0}(x) z ∈ B 5 R 0 ( x ) ,
∣ d ( z ) − l y ( z ) ∣ ≤ C Ω ∣ z − x ∣ 2 , ∣ ∇ d ( z ) − ∇ l y ∣ ≤ C Ω ∣ z − x ∣ . (26) |d(z)-l_y(z)|\le C_\Omega|z-x|^2,\qquad|\nabla d(z)-\nabla l_y|\le C_\Omega|z-x|. \tag*{(26)} ∣ d ( z ) − l y ( z ) ∣ ≤ C Ω ∣ z − x ∣ 2 , ∣∇ d ( z ) − ∇ l y ∣ ≤ C Ω ∣ z − x ∣. (26)
We start with z ∈ B r / 2 ( y ) z\in B_{r/2}(y) z ∈ B r /2 ( y ) where both d d d and l y l_y l y are at least r / 2 r/2 r /2 , and ∣ z − x ∣ ≤ 2 r |z-x|\le2r ∣ z − x ∣ ≤ 2 r . Writing e y = ( U − U ( y ) ) d + U ( y ) ( d − l y ) e_y=(U-U(y))\sqrt d+U(y)(\sqrt d-\sqrt{l_y}) e y = ( U − U ( y )) d + U ( y ) ( d − l y ) and differentiating twice, we get
∥ D 2 e y ∥ L ∞ ( B r / 2 ( y ) ) ≤ C r − 1 / 2 . \|D^2e_y\|_{L^\infty(B_{r/2}(y))}\le Cr^{-1/2}. ∥ D 2 e y ∥ L ∞ ( B r /2 ( y )) ≤ C r − 1/2 .
The second difference thus satisfies ∣ 2 e y ( y ) − e y ( y + ρ θ ) − e y ( y − ρ θ ) ∣ ≤ C r − 1 / 2 ρ 2 |2e_y(y)-e_y(y+\rho\theta)-e_y(y-\rho\theta)|\le Cr^{-1/2}\rho^2 ∣2 e y ( y ) − e y ( y + ρθ ) − e y ( y − ρθ ) ∣ ≤ C r − 1/2 ρ 2 for ρ < r / 2 \rho<r/2 ρ < r /2 , and the near field contributes at most of order ∥ b ∥ ∞ r \|b\|_\infty\sqrt r ∥ b ∥ ∞ r to P b e y ( y ) \mathcal{P}_b e_y(y) P b e y ( y ) . Let now r / 2 ≤ ρ ≤ R 0 r/2 \le\rho\le R_0 r /2 ≤ ρ ≤ R 0 and consider z z z with ρ ≤ ∣ z − y ∣ < 2 ρ \rho\le|z-y| < 2\rho ρ ≤ ∣ z − y ∣ < 2 ρ , so that ∣ z − x ∣ ≤ 4 ρ |z-x| \le4\rho ∣ z − x ∣ ≤ 4 ρ . Since ∣ U ( z ) − U ( y ) ∣ ≤ C ρ |U(z)-U(y)| \le C\rho ∣ U ( z ) − U ( y ) ∣ ≤ C ρ and d ( z ) + ≤ C ρ \sqrt{d(z)_+} \le C\sqrt{\rho} d ( z ) + ≤ C ρ ,
∣ e y ( z ) ∣ ≤ C ρ 3 / 2 + C ∣ d ( z ) + − l y ( z ) + ∣ . |e_y(z)| \le C\rho^{3/2}+C\left|\sqrt{d(z)_+}-\sqrt{l_y(z)_+}\right|. ∣ e y ( z ) ∣ ≤ C ρ 3/2 + C d ( z ) + − l y ( z ) + .
The bound ∣ d + − ( l y ) + ∣ ≤ ∣ d − l y ∣ 1 / 2 ≤ C ρ \left|\sqrt{d_+}-\sqrt{(l_y)_+}\right|\le|d-l_y|^{1/2}\le C\rho d + − ( l y ) + ≤ ∣ d − l y ∣ 1/2 ≤ C ρ is not sufficient: it would give a contribution of order one from each dyadic annulus, hence a logarithmic divergence. We use instead that d d d and l y l_y l y differ by O ( ρ 2 ) O(\rho^2) O ( ρ 2 ) in the normal direction. Set a = l y ( z ) a=l_y(z) a = l y ( z ) and ε ρ = 16 C Ω ρ 2 \varepsilon_\rho=16C\Omega\rho^2 ε ρ = 16 C Ω ρ 2 , so that ∣ d ( z ) − a ∣ ≤ ε ρ |d(z)-a|\le\varepsilon_\rho ∣ d ( z ) − a ∣ ≤ ε ρ by (26) . Since s ↦ s + s\mapsto\sqrt{s_+} s ↦ s + is nondecreasing, both d ( z ) + \sqrt{d(z)_+} d ( z ) + and a + \sqrt{a_+} a + lie between ( a − ε ρ ) + \sqrt{(a-\varepsilon_\rho)_+} ( a − ε ρ ) + and ( a + ε ρ ) + \sqrt{(a+\varepsilon_\rho)_+} ( a + ε ρ ) + , so
∣ d ( z ) + − a + ∣ ≤ ( a + ε ρ ) + − ( a − ε ρ ) + . \left|\sqrt{d(z)_+}-\sqrt{a_+}\right|\le\sqrt{(a+\varepsilon_\rho)_+}-\sqrt{(a-\varepsilon_\rho)_+}. d ( z ) + − a + ≤ ( a + ε ρ ) + − ( a − ε ρ ) + .
By the choice of R 0 R_0 R 0 , ε ρ ≤ ρ \varepsilon_\rho\le\rho ε ρ ≤ ρ . Integrating in the normal coordinate a a a ,
∫ − 4 ρ 4 ρ ( ( a + ε ρ ) + − ( a − ε ρ ) + ) d a = 2 3 ( ( 4 ρ + ε ρ ) 3 / 2 − ( 4 ρ − ε ρ ) 3 / 2 ) ≤ C ε ρ ρ = C ρ 5 / 2 , \int_{-4\rho}^{4\rho}\left(\sqrt{(a+\varepsilon_\rho)_+}-\sqrt{(a-\varepsilon_\rho)_+}\right)\,da
=\frac{2}{3}\left((4\rho+\varepsilon_\rho)^{3/2}-(4\rho-\varepsilon_\rho)^{3/2}\right)
\le C\varepsilon_\rho\sqrt{\rho}=C\rho^{5/2}, ∫ − 4 ρ 4 ρ ( ( a + ε ρ ) + − ( a − ε ρ ) + ) d a = 3 2 ( ( 4 ρ + ε ρ ) 3/2 − ( 4 ρ − ε ρ ) 3/2 ) ≤ C ε ρ ρ = C ρ 5/2 ,
and, since the tangential cross-sections have measure at most O ( ρ n − 1 ) O(\rho^{n-1}) O ( ρ n − 1 ) ,
∫ ρ ≤ ∣ z − y ∣ < 2 ρ ∣ e y ( z ) ∣ d z ≤ C ρ n + 3 / 2 . \int_{\rho\le|z-y|<2\rho}|e_y(z)|\,dz\le C\rho^{n+3/2}. ∫ ρ ≤ ∣ z − y ∣ < 2 ρ ∣ e y ( z ) ∣ d z ≤ C ρ n + 3/2 .
As e y ( y ) = 0 e_y(y)=0 e y ( y ) = 0 and ∣ z − y ∣ − n − 1 ≤ ρ − n − 1 |z-y|^{-n-1}\le\rho^{-n-1} ∣ z − y ∣ − n − 1 ≤ ρ − n − 1 on the annulus, each annulus contributes at most O ( ∥ b ∥ ∞ ρ 1 / 2 ) O(\|b\|_\infty\rho^{1/2}) O ( ∥ b ∥ ∞ ρ 1/2 ) to P b e y ( y ) \mathcal{P}_b e_y(y) P b e y ( y ) . Summing over ρ = 2 j r / 2 < R 0 \rho=2^j r/2<R_0 ρ = 2 j r /2 < R 0 , j ≥ 0 j\ge0 j ≥ 0 , gives a bound O ( ∥ b ∥ ∞ ) O(\|b\|_\infty) O ( ∥ b ∥ ∞ ) independent of r r r .
For ∣ z − y ∣ ≥ R 0 |z-y|\ge R_0 ∣ z − y ∣ ≥ R 0 , ∣ e y ( z ) ∣ ≤ ∥ u ∥ ∞ + ∣ U ( y ) ∣ r + ∣ z − y ∣ ≤ C ( 1 + ∣ z − y ∣ 1 / 2 ) |e_y(z)|\le\|u\|_\infty+|U(y)|\sqrt{r+|z-y|}\le C(1+|z-y|^{1/2}) ∣ e y ( z ) ∣ ≤ ∥ u ∥ ∞ + ∣ U ( y ) ∣ r + ∣ z − y ∣ ≤ C ( 1 + ∣ z − y ∣ 1/2 ) , which is integrable against ∣ z − y ∣ − n − 1 |z-y|^{-n-1} ∣ z − y ∣ − n − 1 on { ∣ z − y ∣ ≥ R 0 } \{|z-y|\ge R_0\} { ∣ z − y ∣ ≥ R 0 } . The tail contributes at most of the order of ∥ b ∥ ∞ \|b\|_\infty ∥ b ∥ ∞ .
Summing the three contributions we conclude that as claimed
∣ P b u ( y ) ∣ = ∣ P b e y ( y ) ∣ ≤ C ∥ b ∥ ∞ . |\mathcal{P}_b u(y)|=|\mathcal{P}_b e_y(y)|\le C\|b\|_\infty. ∣ P b u ( y ) ∣ = ∣ P b e y ( y ) ∣ ≤ C ∥ b ∥ ∞ .
Proof of Theorem 1.1
Deformations
In this section the normal deformation direction φ ∈ C ∞ ( ∂ Ω ) \varphi\in C^\infty(\partial\Omega) φ ∈ C ∞ ( ∂ Ω ) is fixed. For the corresponding family of deformations Φ t \Phi_t Φ t we set Ω t = Φ t ( Ω ) \Omega_t=\Phi_t(\Omega) Ω t = Φ t ( Ω ) , u t = u Ω t u_t=u_{\Omega_t} u t = u Ω t and J t = det D Φ t J_t=\det D\Phi_t J t = det D Φ t . For ∣ t ∣ |t| ∣ t ∣ small, Ω t \Omega_t Ω t has C ∞ C^\infty C ∞ boundary, and we fix δ t ∈ C ∞ ( Ω t ‾ ) \delta_t\in C^\infty(\overline{\Omega_t}) δ t ∈ C ∞ ( Ω t ) , positive in Ω t \Omega_t Ω t and equal to the distance to ∂ Ω t \partial\Omega_t ∂ Ω t near ∂ Ω t \partial\Omega_t ∂ Ω t , with δ 0 = δ \delta_0=\delta δ 0 = δ . The function δ ∘ Φ t − 1 \delta\circ\Phi_t^{-1} δ ∘ Φ t − 1 is smooth on Ω t ‾ \overline{\Omega_t} Ω t , positive in Ω t \Omega_t Ω t , and vanishes on ∂ Ω t \partial\Omega_t ∂ Ω t with non-zero gradient. Writing it in normal coordinates of ∂ Ω t \partial\Omega_t ∂ Ω t and using Taylor’s formula with integral remainder in the normal variable, we obtain
δ ∘ Φ t − 1 = γ t δ t , γ t ∈ C ∞ ( Ω t ‾ ) , γ t > 0. (27) \delta\circ\Phi_t^{-1}=\gamma_t\delta_t,\qquad\gamma_t\in C^\infty(\overline{\Omega_t}),\qquad\gamma_t>0. \tag*{(27)} δ ∘ Φ t − 1 = γ t δ t , γ t ∈ C ∞ ( Ω t ) , γ t > 0. (27)
On the boundary, γ t \gamma_t γ t is the normal derivative of δ ∘ Φ t − 1 \delta\circ\Phi_t^{-1} δ ∘ Φ t − 1 : since ∇ δ = − ν \nabla\delta=-\nu ∇ δ = − ν on ∂ Ω \partial\Omega ∂ Ω ,
γ t ( Φ t ( x ) ) = β t ( x ) : = ∣ D Φ t ( x ) − T ν ( x ) ∣ , x ∈ ∂ Ω . (28) \gamma_t(\Phi_t(x))=\beta_t(x):=|D\Phi_t(x)^{-T}\nu(x)|,\qquad x\in\partial\Omega. \tag*{(28)} γ t ( Φ t ( x )) = β t ( x ) := ∣ D Φ t ( x ) − T ν ( x ) ∣ , x ∈ ∂ Ω. (28)
Since D Φ t = Id + t D X D\Phi_t=\operatorname{Id}+tDX D Φ t = Id + t D X we have
β 0 = 1 , ∂ t β t ∣ t = 0 = − ν ⋅ D X ν = 0 on ∂ Ω . (29) \beta_0=1,\qquad\left.\partial_t\beta_t\right|_{t=0}=-\nu\cdot DX\nu=0\quad\text{on }\partial\Omega. \tag*{(29)} β 0 = 1 , ∂ t β t ∣ t = 0 = − ν ⋅ D X ν = 0 on ∂ Ω. (29)
Definition 3.1. The material derivative is w = ∂ t v t ∣ t = 0 w=\left.\partial_t v_t\right|_{t=0} w = ∂ t v t ∣ t = 0 , where v t = u t ∘ Φ t v_t=u_t\circ\Phi_t v t = u t ∘ Φ t . The Eulerian derivative is q = ∂ t u t ∣ t = 0 q=\left.\partial_tu_t\right|_{t=0} q = ∂ t u t ∣ t = 0 in D ′ ( R n ) \mathcal{D}'(\mathbb{R}^n) D ′ ( R n ) .
By Proposition 3.2 below, w w w exists in V V V , and by Proposition 3.3 , q q q exists and q = w − X ⋅ ∇ u q = w - X \cdot\nabla u q = w − X ⋅ ∇ u .
The function w w w vanishes like d \sqrt{d} d on ∂ Ω \partial\Omega ∂ Ω (Lemma 3.4 ), whereas X ⋅ ∇ u X \cdot\nabla u X ⋅ ∇ u behaves like − ψ d 2 d − 1 / 2 -\frac{\psi_d}{2}d^{-1/2} − 2 ψ d d − 1/2 . Hence, as soon as φ ≢ 0 \varphi\not\equiv0 φ ≡ 0 , q q q does not belong to L 2 L^2 L 2 , and a fortiori not to V V V : the Eulerian derivative lies outside the variational framework in which u t u_t u t is defined. This is the main difficulty in the computation of the second variation.
The bulk second variation
We first pull the problem back to the fixed domain Ω \Omega Ω and write the second derivative of T ( Ω t ) T(\Omega_t) T ( Ω t ) in terms of the material derivative. Changing variables by Φ t \Phi_t Φ t , we define on V V V
a t ( v 1 , v 2 ) = c n 2 ∬ ( v 1 ( y ) − v 1 ( z ) ) ( v 2 ( y ) − v 2 ( z ) ) J t ( y ) J t ( z ) ∣ Φ t ( y ) − Φ t ( z ) ∣ n + 1 d y d z , a_t(v_1,v_2)=\frac{c_n}{2}\iint\frac{(v_1(y)-v_1(z))(v_2(y)-v_2(z))J_t(y)J_t(z)}{|\Phi_t(y)-\Phi_t(z)|^{n+1}}\,dy\,dz, a t ( v 1 , v 2 ) = 2 c n ∬ ∣ Φ t ( y ) − Φ t ( z ) ∣ n + 1 ( v 1 ( y ) − v 1 ( z )) ( v 2 ( y ) − v 2 ( z )) J t ( y ) J t ( z ) d y d z ,
ℓ t ( v ) = ∫ Ω v J t . \ell_t(v)=\int_{\Omega} vJ_t. ℓ t ( v ) = ∫ Ω v J t .
Thus a 0 = a a_0=a a 0 = a , ℓ 0 ( v ) = ∫ Ω v \ell_0(v)=\int_{\Omega}v ℓ 0 ( v ) = ∫ Ω v , and a t ( v 1 , v 2 ) = a ( v 1 ∘ Φ t − 1 , v 2 ∘ Φ t − 1 ) a_t(v_1,v_2)=a(v_1\circ\Phi_t^{-1},v_2\circ\Phi_t^{-1}) a t ( v 1 , v 2 ) = a ( v 1 ∘ Φ t − 1 , v 2 ∘ Φ t − 1 ) . We write a 0 ′ a'_0 a 0 ′ , a 0 ′ ′ a''_0 a 0 ′′ , ℓ 0 ′ \ell'_0 ℓ 0 ′ , ℓ 0 ′ ′ \ell''_0 ℓ 0 ′′ for the derivatives at t = 0 t=0 t = 0 , which exist in L ( V , V ∗ ) \mathcal{L}(V,V^*) L ( V , V ∗ ) and V ∗ V^* V ∗ by the proof below.
Proposition 3.2. The map t ↦ v t = u t ∘ Φ t t\mapsto v_t=u_t\circ\Phi_t t ↦ v t = u t ∘ Φ t is of class C 2 C^2 C 2 with values in V V V , the material derivative w w w is the unique solution in V V V of
a ( w , v ) = ℓ 0 ′ ( v ) − a 0 ′ ( u , v ) ( v ∈ V ) , (30) a(w,v)=\ell'_0(v)-a'_0(u,v) \qquad(v\in V), \tag*{(30)} a ( w , v ) = ℓ 0 ′ ( v ) − a 0 ′ ( u , v ) ( v ∈ V ) , (30) and
d 2 d t 2 T ( Ω t ) ∣ t = 0 = 2 ℓ 0 ′ ′ ( u ) − a 0 ′ ′ ( u , u ) + 2 a ( w , w ) . (31) \left.\frac{d^2}{dt^2}T(\Omega_t)\right|_{t=0}=2\ell''_0(u)-a''_0(u,u)+2a(w,w). \tag*{(31)} d t 2 d 2 T ( Ω t ) t = 0 = 2 ℓ 0 ′′ ( u ) − a 0 ′′ ( u , u ) + 2 a ( w , w ) . (31)
Proof. The equation on the fixed domain. Composition with the uniformly bi-Lipschitz maps Φ t \Phi_t Φ t preserves H 1 / 2 ( R n ) H^{1/2}(\mathbb{R}^n) H 1/2 ( R n ) , with uniform bounds, and u t ∘ Φ t u_t\circ\Phi_t u t ∘ Φ t vanishes outside Ω \Omega Ω ; hence v t ∈ V v_t\in V v t ∈ V for every t t t . Changing variables in the weak formulation of {(1.1)} on Ω t \Omega_t Ω t and testing with v ∘ Φ t − 1 v\circ\Phi_t^{-1} v ∘ Φ t − 1 , v ∈ V v\in V v ∈ V , we obtain
a t ( v t , v ) = ℓ t ( v ) ( v ∈ V ) . (32) a_t(v_t,v)=\ell_t(v) \qquad(v\in V). \tag*{(32)} a t ( v t , v ) = ℓ t ( v ) ( v ∈ V ) . (32)
Regularity of a t a_t a t and ℓ t \ell_t ℓ t . We isolate the dependence on t t t in the multiplier
ρ t ( y , z ) = J t ( y ) J t ( z ) ( ∣ y − z ∣ ∣ Φ t ( y ) − Φ t ( z ) ∣ ) n + 1 , \rho_t(y,z)=J_t(y)J_t(z)\left(\frac{|y-z|}{|\Phi_t(y)-\Phi_t(z)|}\right)^{n+1}, ρ t ( y , z ) = J t ( y ) J t ( z ) ( ∣ Φ t ( y ) − Φ t ( z ) ∣ ∣ y − z ∣ ) n + 1 ,
a t ( v 1 , v 2 ) = c n 2 ∬ ( v 1 ( y ) − v 1 ( z ) ) ( v 2 ( y ) − v 2 ( z ) ) ∣ y − z ∣ n + 1 ρ t ( y , z ) d y d z . (33) a_t(v_1,v_2)=\frac{c_n}{2}\iint\frac{(v_1(y)-v_1(z))(v_2(y)-v_2(z))}{|y-z|^{n+1}}\rho_t(y,z)\,dy\,dz. \tag*{(33)} a t ( v 1 , v 2 ) = 2 c n ∬ ∣ y − z ∣ n + 1 ( v 1 ( y ) − v 1 ( z )) ( v 2 ( y ) − v 2 ( z )) ρ t ( y , z ) d y d z . (33)
Let Ξ t ( y , z ) = ( Φ t ( y ) − Φ t ( z ) ) / ∣ y − z ∣ \Xi_t(y,z)=(\Phi_t(y)-\Phi_t(z))/|y-z| Ξ t ( y , z ) = ( Φ t ( y ) − Φ t ( z )) /∣ y − z ∣ for y ≠ z y\ne z y = z . The map g ↦ ( g ( y ) − g ( z ) ) / ∣ y − z ∣ g\mapsto(g(y)-g(z))/|y-z| g ↦ ( g ( y ) − g ( z )) /∣ y − z ∣ is linear and bounded from W 1 , ∞ ( R n ; R n ) W^{1,\infty}(\mathbb{R}^n;\mathbb{R}^n) W 1 , ∞ ( R n ; R n ) to L ∞ ( R 2 n ; R n ) L^\infty(\mathbb{R}^{2n};\mathbb{R}^n) L ∞ ( R 2 n ; R n ) . Since t ↦ Φ t − Id t\mapsto\Phi_t-\operatorname{Id} t ↦ Φ t − Id is C 2 C^2 C 2 with values in W 1 , ∞ W^{1,\infty} W 1 , ∞ , t ↦ Ξ t t\mapsto\Xi_t t ↦ Ξ t is C 2 C^2 C 2 with values in L ∞ ( R 2 n ) L^\infty(\mathbb{R}^{2n}) L ∞ ( R 2 n ) . The uniform bi-Lipschitz bounds show that Ξ t \Xi_t Ξ t takes values in a fixed compact subset of R n ∖ { 0 } \mathbb{R}^n\setminus\{0\} R n ∖ { 0 } , on which ξ ↦ ∣ ξ ∣ − n − 1 \xi\mapsto|\xi|^{-n-1} ξ ↦ ∣ ξ ∣ − n − 1 is smooth, so t ↦ ∣ Ξ t ∣ − n − 1 t\mapsto|\Xi_t|^{-n-1} t ↦ ∣ Ξ t ∣ − n − 1 is also C 2 C^2 C 2 into L ∞ L^\infty L ∞ . Since J t J_t J t is a polynomial in the entries of D Φ t D\Phi_t D Φ t , we conclude that
ρ t ∈ C 2 ( ( − t 0 , t 0 ) ; L ∞ ( R 2 n ) ) . (34) \rho_t\in C^2((-t_0,t_0);L^\infty(\mathbb{R}^{2n})). \tag*{(34)} ρ t ∈ C 2 (( − t 0 , t 0 ) ; L ∞ ( R 2 n )) . (34)
Note that this is differentiability with respect to the L ∞ L^\infty L ∞ norm, not merely a pointwise bound on ∂ t k ρ t \partial_t^k\rho_t ∂ t k ρ t . For m ∈ L ∞ ( R 2 n ) m\in L^\infty(\mathbb{R}^{2n}) m ∈ L ∞ ( R 2 n ) let B m B_m B m be the bilinear form defined as in (33) with ρ t \rho_t ρ t replaced by m m m . The Cauchy–Schwarz inequality gives
∣ B m ( v 1 , v 2 ) ∣ ≤ ∥ m ∥ L ∞ ∥ v 1 ∥ V ∥ v 2 ∥ V . (35) \lvert B_m(v_1,v_2)\rvert\leq\lVert m\rVert_{L^\infty}\lVert v_1\rVert_V\lVert v_2\rVert_V. \tag*{(35)} ∣ B m ( v 1 , v 2 )∣ ≤ ∥ m ∥ L ∞ ∥ v 1 ∥ V ∥ v 2 ∥ V . (35)
so m ↦ B m m\mapsto B_m m ↦ B m is a bounded linear map from L ∞ ( R 2 n ) L^\infty(\mathbb{R}^{2n}) L ∞ ( R 2 n ) into L ( V , V ∗ ) \mathcal{L}(V,V^*) L ( V , V ∗ ) . Since a t = B ρ t a_t=B_{\rho_t} a t = B ρ t , (34) and (35) give a t ∈ C 2 ( ( − t 0 , t 0 ) ; L ( V , V ∗ ) ) a_t\in C^2((-t_0,t_0);\mathcal{L}(V,V^*)) a t ∈ C 2 (( − t 0 , t 0 ) ; L ( V , V ∗ )) , with ∂ t k a t = B ∂ t k ρ t \partial_t^k a_t=B_{\partial_t^k\rho_t} ∂ t k a t = B ∂ t k ρ t . The uniform bounds 0 < c ≤ ρ t ≤ C 0<c\leq\rho_t\leq C 0 < c ≤ ρ t ≤ C give for v ∈ V v\in V v ∈ V ,
c a ( v , v ) ≤ a t ( v , v ) ≤ C a ( v , v ) . (36) c\,a(v,v)\leq a_t(v,v)\leq C\,a(v,v). \tag*{(36)} c a ( v , v ) ≤ a t ( v , v ) ≤ C a ( v , v ) . (36)
By (19) , ∣ ∫ Ω v J ∣ ≤ C ∥ J ∥ L ∞ ∥ v ∥ V \lvert\int_\Omega vJ\rvert\leq C\lVert J\rVert_{L^\infty}\lVert v\rVert_V ∣ ∫ Ω v J ∣ ≤ C ∥ J ∥ L ∞ ∥ v ∥ V ; since J t ∈ C t 2 L x ∞ J_t\in C^2_tL^\infty_x J t ∈ C t 2 L x ∞ , we get ℓ t ∈ C 2 ( ( − t 0 , t 0 ) ; V ∗ ) \ell_t\in C^2((-t_0,t_0);V^*) ℓ t ∈ C 2 (( − t 0 , t 0 ) ; V ∗ ) .
Differentiability of v t v_t v t . Let A t ∈ L ( V , V ∗ ) A_t\in\mathcal{L}(V,V^*) A t ∈ L ( V , V ∗ ) be defined by ⟨ A t v 1 , v 2 ⟩ = a t ( v 1 , v 2 ) \langle A_tv_1,v_2\rangle=a_t(v_1,v_2) ⟨ A t v 1 , v 2 ⟩ = a t ( v 1 , v 2 ) , so that (32) reads A t v t = ℓ t A_tv_t=\ell_t A t v t = ℓ t . By (36) and the Lax–Milgram theorem, each A t A_t A t is invertible with ∥ A t − 1 ∥ ≤ c − 1 \lVert A_t^{-1}\rVert\leq c^{-1} ∥ A t − 1 ∥ ≤ c − 1 . Inversion is smooth on the open set of invertible operators, so t ↦ A t − 1 t\mapsto A_t^{-1} t ↦ A t − 1 is C 2 C^2 C 2 into L ( V ∗ , V ) \mathcal{L}(V^*,V) L ( V ∗ , V ) , and v t = A t − 1 ℓ t ∈ C 2 ( ( − t 0 , t 0 ) ; V ) v_t=A_t^{-1}\ell_t\in C^2((-t_0,t_0);V) v t = A t − 1 ℓ t ∈ C 2 (( − t 0 , t 0 ) ; V ) . Differentiating (32) at t = 0 t=0 t = 0 gives a 0 ′ ( u , v ) + a ( w , v ) = ℓ 0 ′ ( v ) a'_0(u,v)+a(w,v)=\ell'_0(v) a 0 ′ ( u , v ) + a ( w , v ) = ℓ 0 ′ ( v ) , that is (30) , which determines w w w by coercivity.
The second variation. Let F t ( v ) = 2 ℓ t ( v ) − a t ( v , v ) F_t(v)=2\ell_t(v)-a_t(v,v) F t ( v ) = 2 ℓ t ( v ) − a t ( v , v ) . By (32) , F t ( v t ) = ℓ t ( v t ) = ∫ Ω t u t = T ( Ω t ) F_t(v_t)=\ell_t(v_t)=\int_{\Omega_t}u_t=T(\Omega_t) F t ( v t ) = ℓ t ( v t ) = ∫ Ω t u t = T ( Ω t ) . Set G ( t ) = F t ( u ) = 2 ℓ t ( u ) − a t ( u , u ) G(t)=F_t(u)=2\ell_t(u)-a_t(u,u) G ( t ) = F t ( u ) = 2 ℓ t ( u ) − a t ( u , u ) and h = v t − u h=v_t-u h = v t − u . Using (32) ,
F t ( v t ) − F t ( u ) = 2 ℓ t ( h ) − 2 a t ( u , h ) − a t ( h , h ) = 2 a t ( v t , h ) − 2 a t ( u , h ) − a t ( h , h ) = a t ( h , h ) , F_t(v_t)-F_t(u)=2\ell_t(h)-2a_t(u,h)-a_t(h,h)=2a_t(v_t,h)-2a_t(u,h)-a_t(h,h)=a_t(h,h), F t ( v t ) − F t ( u ) = 2 ℓ t ( h ) − 2 a t ( u , h ) − a t ( h , h ) = 2 a t ( v t , h ) − 2 a t ( u , h ) − a t ( h , h ) = a t ( h , h ) ,
which gives the exact identity
T ( Ω t ) = G ( t ) + a t ( v t − u , v t − u ) . (37) T(\Omega_t)=G(t)+a_t(v_t-u,v_t-u). \tag*{(37)} T ( Ω t ) = G ( t ) + a t ( v t − u , v t − u ) . (37)
The last term is C 2 C^2 C 2 in t t t , and since v t − u = t w + o ( t ) v_t-u=tw+o(t) v t − u = tw + o ( t ) in V V V and a t → a a_t\to a a t → a in operator norm, it equals t 2 a ( w , w ) + o ( t 2 ) t^2a(w,w)+o(t^2) t 2 a ( w , w ) + o ( t 2 ) ; its second derivative at t = 0 t=0 t = 0 is 2 a ( w , w ) 2a(w,w) 2 a ( w , w ) . Since u u u is fixed, G ′ ′ ( 0 ) = 2 ℓ 0 ′ ′ ( u ) − a 0 ′ ′ ( u , u ) G''(0)=2\ell''_0(u)-a''_0(u,u) G ′′ ( 0 ) = 2 ℓ 0 ′′ ( u ) − a 0 ′′ ( u , u ) , and (31) follows. □
The Eulerian derivative
We now identify the Eulerian derivative with the large solution of Proposition 2.4 with datum ψ φ / 2 \psi\varphi/2 ψ φ /2 . This is the step where the singular behaviour of q q q is handled.
Proposition 3.3. The Eulerian derivative is the integrable function
q = w − X ⋅ ∇ u = Q ψ φ / 2 , (38) q=w-X\cdot\nabla u=Q_{\psi\varphi/2}, \tag*{(38)} q = w − X ⋅ ∇ u = Q ψ φ /2 , (38) so that, for x ∈ ∂ Ω x\in\partial\Omega x ∈ ∂ Ω ,
q ( x − r ν ( x ) ) = ψ φ 2 r + B φ r + O ( r 3 / 2 ) , B φ = 1 2 D Ω ( ψ φ ) . (39) q(x-r\nu(x))=\frac{\psi\varphi}{2\sqrt{r}}+B_\varphi\sqrt{r}+O(r^{3/2}),\qquad B_\varphi=\frac{1}{2}D_\Omega(\psi\varphi). \tag*{(39)} q ( x − r ν ( x )) = 2 r ψ φ + B φ r + O ( r 3/2 ) , B φ = 2 1 D Ω ( ψ φ ) . (39)
Proof. Since ∣ ∇ u ∣ ≤ C δ − 1 / 2 \lvert\nabla u\rvert\leq C\delta^{-1/2} ∣ ∇ u ∣ ≤ C δ − 1/2 near ∂ Ω \partial\Omega ∂ Ω , the gradient is integrable. Moreover, u u u vanishes on ∂ Ω \partial\Omega ∂ Ω , so its zero extension creates no jump and u = δ U ∈ W 1 , 1 ( R n ) u=\sqrt{\delta}\,U\in W^{1,1}(\mathbb{R}^n) u = δ U ∈ W 1 , 1 ( R n ) . Let χ ∈ C c ∞ ( R n ) \chi\in C^\infty_c(\mathbb{R}^n) χ ∈ C c ∞ ( R n ) . Changing variables,
∫ R n u t χ = ∫ Ω v t ( χ ∘ Φ t ) J t , \int_{\mathbb{R}^n}u_t\chi=\int_\Omega v_t(\chi\circ\Phi_t)J_t, ∫ R n u t χ = ∫ Ω v t ( χ ∘ Φ t ) J t ,
and since v t ∈ C t 1 V v_t\in C^1_tV v t ∈ C t 1 V , this is differentiable at t = 0 t=0 t = 0 with derivative
⟨ q , χ ⟩ = ∫ Ω w χ + ∫ Ω u X ⋅ ∇ χ + ∫ Ω u χ div X . \langle q,\chi\rangle=\int_\Omega w\chi+\int_\Omega uX\cdot\nabla\chi+\int_\Omega u\chi\operatorname{div}X. ⟨ q , χ ⟩ = ∫ Ω w χ + ∫ Ω u X ⋅ ∇ χ + ∫ Ω u χ div X .
Since u X χ ∈ W 1 , 1 ( R n ) uX\chi\in W^{1,1}(\mathbb{R}^n) u X χ ∈ W 1 , 1 ( R n ) has compact support, ∫ R n div ( u X χ ) = 0 \int_{\mathbb{R}^n}\operatorname{div}(uX\chi)=0 ∫ R n div ( u X χ ) = 0 , and the product rule turns the last two terms into − ∫ χ X ⋅ ∇ u -\int\chi X\cdot\nabla u − ∫ χ X ⋅ ∇ u . Hence
q = w − X ⋅ ∇ u ∈ L 1 ( R n ) . q=w-X\cdot\nabla u\in L^1(\mathbb{R}^n). q = w − X ⋅ ∇ u ∈ L 1 ( R n ) .
Let now χ ∈ C c ∞ ( Ω ) \chi\in C^\infty_c(\Omega) χ ∈ C c ∞ ( Ω ) . For ∣ t ∣ |t| ∣ t ∣ small the support of χ \chi χ is contained in Ω t \Omega_t Ω t , so ∫ u t ( − Δ ) 1 / 2 χ = ∫ χ \int u_t(-\Delta)^{1/2}\chi=\int\chi ∫ u t ( − Δ ) 1/2 χ = ∫ χ , which does not depend on t t t . All the functions u t u_t u t , and hence q q q , are supported in a fixed compact set K K K . Let ζ ∈ C c ∞ ( R n ) \zeta\in C^\infty_c(\mathbb{R}^n) ζ ∈ C c ∞ ( R n ) with ζ = 1 \zeta=1 ζ = 1 near K K K . Then ∫ u t ζ ( − Δ ) 1 / 2 χ = ∫ χ \int u_t\zeta(-\Delta)^{1/2}\chi=\int\chi ∫ u t ζ ( − Δ ) 1/2 χ = ∫ χ for all small ∣ t ∣ |t| ∣ t ∣ , and the formula for ⟨ q , ⋅ ⟩ \langle q,\cdot\rangle ⟨ q , ⋅ ⟩ obtained above gives ⟨ q , ( − Δ ) 1 / 2 χ ⟩ = ⟨ q , ζ ( − Δ ) 1 / 2 χ ⟩ = 0 \langle q,(-\Delta)^{1/2}\chi\rangle=\langle q,\zeta(-\Delta)^{1/2}\chi\rangle=0 ⟨ q , ( − Δ ) 1/2 χ ⟩ = ⟨ q , ζ ( − Δ ) 1/2 χ ⟩ = 0 , that is, q q q is 1 2 \frac12 2 1 -harmonic in Ω \Omega Ω in the sense of distributions.
Let f = ψ φ / 2 f=\psi\varphi/2 f = ψ φ /2 , and let F F F and ζ f = δ − 1 / 2 F \zeta_f=\delta^{-1/2}F ζ f = δ − 1/2 F be as in Lemma 2.3 . From u = δ U u=\sqrt{\delta}\,U u = δ U ,
X ⋅ ∇ u + ζ f = δ − 1 / 2 G + δ X ⋅ ∇ U , G : = 1 2 U X ⋅ ∇ δ + F . X\cdot\nabla u+\zeta_f=\delta^{-1/2}G+\sqrt{\delta}\,X\cdot\nabla U,\qquad G:=\frac12 U X\cdot\nabla\delta+F. X ⋅ ∇ u + ζ f = δ − 1/2 G + δ X ⋅ ∇ U , G := 2 1 U X ⋅ ∇ δ + F .
The function G G G is smooth up to ∂ Ω \partial\Omega ∂ Ω . On ∂ Ω \partial\Omega ∂ Ω we have U = ψ U=\psi U = ψ , ∇ δ = − ν \nabla\delta=-\nu ∇ δ = − ν , hence X ⋅ ∇ δ = − φ X\cdot\nabla\delta=-\varphi X ⋅ ∇ δ = − φ , and F = ψ φ / 2 F=\psi\varphi/2 F = ψ φ /2 , so that
G ∣ ∂ Ω = − 1 2 ψ φ + 1 2 ψ φ = 0. G|_{\partial\Omega}=-\frac12\psi\varphi+\frac12\psi\varphi=0. G ∣ ∂ Ω = − 2 1 ψ φ + 2 1 ψ φ = 0.
This is the point of the choice f = ψ φ / 2 f=\psi\varphi/2 f = ψ φ /2 : neither X ⋅ ∇ u X\cdot\nabla u X ⋅ ∇ u nor ζ f \zeta_f ζ f belongs to V V V , but their δ − 1 / 2 \delta^{-1/2} δ − 1/2 singularities cancel. As in the proof of uniqueness in Proposition 2.4, G = δ A G=\delta A G = δ A with A ∈ C ∞ ( Ω ‾ ) A\in C^\infty(\overline{\Omega}) A ∈ C ∞ ( Ω ) , and
X ⋅ ∇ u + ζ f = δ ( A + X ⋅ ∇ U ) ∈ V (40) X\cdot\nabla u+\zeta_f=\sqrt{\delta}\,(A+X\cdot\nabla U)\in V \tag*{(40)} X ⋅ ∇ u + ζ f = δ ( A + X ⋅ ∇ U ) ∈ V (40)
by Lemma 2.2 .
By Proposition 2.4, Q f = ζ f + ω f Q_f=\zeta_f+\omega_f Q f = ζ f + ω f with ω f ∈ V \omega_f\in V ω f ∈ V . Hence
q − Q f = w − ( X ⋅ ∇ u + ζ f ) − ω f ∈ V , q-Q_f=w-(X\cdot\nabla u+\zeta_f)-\omega_f\in V, q − Q f = w − ( X ⋅ ∇ u + ζ f ) − ω f ∈ V ,
each of the three terms being in V V V , by Proposition 3.2 , (3.14) and Proposition 2.4. Both q q q and Q f Q_f Q f are 1 2 \frac12 2 1 -harmonic in Ω \Omega Ω , by the argument above and (6) , so a ( q − Q f , ϕ ) = 0 a(q-Q_f,\phi)=0 a ( q − Q f , ϕ ) = 0 for ϕ ∈ C c ∞ ( Ω ) \phi\in C^\infty_c(\Omega) ϕ ∈ C c ∞ ( Ω ) , and by density for every ϕ ∈ V \phi\in V ϕ ∈ V . Taking ϕ = q − Q f \phi=q-Q_f ϕ = q − Q f gives q = Q f q=Q_f q = Q f . The expansion (39) follows from Proposition 2.4 and the linearity of D Ω D_\Omega D Ω . □
The material derivative
Lemma 3.4. We have w = δ W w=\sqrt{\delta}\,W w = δ W with W ∈ C ∞ ( Ω ‾ ) W\in C^\infty(\overline{\Omega}) W ∈ C ∞ ( Ω ) , and f w : = ( − Δ ) 1 / 2 w ∈ C ∞ ( Ω ‾ ) f_w:=(-\Delta)^{1/2}w\in C^\infty(\overline{\Omega}) f w := ( − Δ ) 1/2 w ∈ C ∞ ( Ω ) . Moreover ψ w : = W ∣ ∂ Ω \psi_w:=W|_{\partial\Omega} ψ w := W ∣ ∂ Ω is given by
ψ w = B φ − 3 2 κ φ . (41) \psi_w=B\varphi-\frac32\kappa\varphi. \tag*{(41)} ψ w = B φ − 2 3 κ φ . (41)
Proposition 3.2 only gives w ∈ V w\in V w ∈ V , which does not imply that w / δ w/\sqrt{\delta} w / δ has a boundary trace. The lemma provides the structure of w w w needed to apply the Pohozaev identity of Lemma 3.5 to the pair ( u , w ) (u,w) ( u , w ) .
Proof. Let f = ψ φ / 2 f=\psi\varphi/2 f = ψ φ /2 so that by Proposition 2.4, ω f = δ h f \omega_f=\sqrt{\delta}\,h_f ω f = δ h f with h f ∈ C ∞ ( Ω ‾ ) h_f\in C^\infty(\overline{\Omega}) h f ∈ C ∞ ( Ω ) . By (38) ,
w = q + X ⋅ ∇ u = ( X ⋅ ∇ u + ζ f ) + ω f = δ ( A + X ⋅ ∇ U + h f ) , (42) w=q+X\cdot\nabla u=(X\cdot\nabla u+\zeta_f)+\omega_f=\sqrt{\delta}\,(A+X\cdot\nabla U+h_f), \tag*{(42)} w = q + X ⋅ ∇ u = ( X ⋅ ∇ u + ζ f ) + ω f = δ ( A + X ⋅ ∇ U + h f ) , (42)
so w = δ W w=\sqrt{\delta}\,W w = δ W with W ∈ C ∞ ( Ω ‾ ) W\in C^\infty(\overline{\Omega}) W ∈ C ∞ ( Ω ) . The second line of (20) then gives f w ∈ C ∞ ( Ω ‾ ) f_w\in C^\infty(\overline{\Omega}) f w ∈ C ∞ ( Ω ) .
To compute ψ w \psi_w ψ w , we expand w = q + X ⋅ ∇ u w=q+X\cdot\nabla u w = q + X ⋅ ∇ u along the normal ray x − r ν ( x ) x-r\nu(x) x − r ν ( x ) , x ∈ ∂ Ω x\in\partial\Omega x ∈ ∂ Ω . There ∇ δ = − ν \nabla\delta=-\nu ∇ δ = − ν and U ( x − r ν ) = ψ + κ r + O ( r 2 ) U(x-r\nu)=\psi+\kappa r+O(r^2) U ( x − r ν ) = ψ + κ r + O ( r 2 ) , so
X ⋅ ∇ u = − U 2 r X ⋅ ν + r X ⋅ ∇ U . X\cdot\nabla u=-\frac{U}{2\sqrt{r}}X\cdot\nu+\sqrt{r}\,X\cdot\nabla U. X ⋅ ∇ u = − 2 r U X ⋅ ν + r X ⋅ ∇ U .
Recalling that ν ⋅ D X ν = 0 \nu\cdot DX\nu= 0 ν ⋅ D X ν = 0 and thus X ( x − r ν ) ⋅ ν = φ + O ( r 2 ) X(x-r\nu)\cdot\nu=\varphi+O(r^2) X ( x − r ν ) ⋅ ν = φ + O ( r 2 ) , the first term equals − ψ φ 2 r − 1 2 r κ φ + O ( r 3 / 2 ) -\frac{\psi\varphi}{2\sqrt{r}}-\frac{1}{2}\sqrt{r}\kappa\varphi+O(r^{3/2}) − 2 r ψ φ − 2 1 r κ φ + O ( r 3/2 ) . On ∂ Ω \partial\Omega ∂ Ω , ∇ U = ∇ ∂ Ω ψ − κ ν \nabla U=\nabla_{\partial\Omega}\psi-\kappa\nu ∇ U = ∇ ∂ Ω ψ − κ ν , hence X ⋅ ∇ U = − κ φ X\cdot\nabla U=-\kappa\varphi X ⋅ ∇ U = − κ φ , and the second term equals − r κ φ + O ( r 3 / 2 ) -\sqrt{r}\kappa\varphi+O(r^{3/2}) − r κ φ + O ( r 3/2 ) . Adding (39) , the terms of order r − 1 / 2 r^{-1/2} r − 1/2 cancel and
w ( x − r ν ( x ) ) = r [ B φ − 3 2 κ φ ] + O ( r 3 / 2 ) , w(x-r\nu(x))=\sqrt{r}\left[B\varphi-\frac{3}{2}\kappa\varphi\right]+O(r^{3/2}), w ( x − r ν ( x )) = r [ B φ − 2 3 κ φ ] + O ( r 3/2 ) ,
which gives (41) .
The bilinear Pohozaev identity
By Proposition 3.2, d 2 d t 2 T ( Ω t ) ∣ t = 0 = G ′ ′ ( 0 ) + 2 a ( w , w ) \left.\frac{d^2}{dt^2}T(\Omega_t)\right|_{t=0}=G^{\prime\prime}(0)+2a(w,w) d t 2 d 2 T ( Ω t ) t = 0 = G ′′ ( 0 ) + 2 a ( w , w ) . We now use a Pohozaev identity to write a ( w , w ) a(w,w) a ( w , w ) as a volume integral plus a boundary integral. We recall the following fractional Pohozaev identity from [10 Theorem 1.3, (1.10)] .
Lemma 3.5. For i = 1 , 2 i=1,2 i = 1 , 2 let v i = δ V i v_i=\sqrt{\delta}V_i v i = δ V i with V i ∈ C ∞ ( Ω ‾ ) V_i\in C^\infty(\overline{\Omega}) V i ∈ C ∞ ( Ω ) , and set f i = ( − Δ ) 1 / 2 v i f_i=(-\Delta)^{1/2}v_i f i = ( − Δ ) 1/2 v i and ψ i = V i ∣ ∂ Ω \psi_i=V_i|_{\partial\Omega} ψ i = V i ∣ ∂ Ω . For every Lipschitz vector field X X X on R n \mathbb{R}^n R n , with φ = X ⋅ ν \varphi=X\cdot\nu φ = X ⋅ ν ,
a 0 ′ [ X ] ( v 1 , v 2 ) = − ∫ Ω { f 1 X ⋅ ∇ v 2 + f 2 X ⋅ ∇ v 1 } d y − c ∗ ∫ ∂ Ω ψ 1 ψ 2 φ d H n − 1 . (43) a'_0[X](v_1,v_2)=-\int_{\Omega}\{f_1X\cdot\nabla v_2+f_2X\cdot\nabla v_1\}\,dy-c_*\int_{\partial\Omega}\psi_1\psi_2\varphi\,d\mathcal{H}^{n-1}. \tag*{(43)} a 0 ′ [ X ] ( v 1 , v 2 ) = − ∫ Ω { f 1 X ⋅ ∇ v 2 + f 2 X ⋅ ∇ v 1 } d y − c ∗ ∫ ∂ Ω ψ 1 ψ 2 φ d H n − 1 . (43) For X ( y ) = y X(y)=y X ( y ) = y and v 1 = v 2 v_1=v_2 v 1 = v 2 , (43) reduces to the fractional Pohozaev identity of Ros-Oton and Serra [25 ] , since then a 0 ′ [ X ] = ( n − 1 ) a a'_0[X]=(n-1)a a 0 ′ [ X ] = ( n − 1 ) a .
Remark 3.6. The case v 1 = v 2 = u v_1=v_2=u v 1 = v 2 = u of (43) gives a proof of the first variation formula (5) . Indeed, (37) shows that d d t T ( Ω t ) ∣ t = 0 = G ′ ( 0 ) = 2 ℓ 0 ′ ( u ) − a 0 ′ ( u , u ) \left.\frac{d}{dt}T(\Omega_t)\right|_{t=0}=G'(0)=2\ell'_0(u)-a'_0(u,u) d t d T ( Ω t ) t = 0 = G ′ ( 0 ) = 2 ℓ 0 ′ ( u ) − a 0 ′ ( u , u ) . Since u ∈ W 1 , 1 ( R n ) u\in W^{1,1}(\mathbb{R}^n) u ∈ W 1 , 1 ( R n ) vanishes on ∂ Ω \partial\Omega ∂ Ω , ℓ 0 ′ ( u ) = ∫ Ω u div X = − ∫ Ω X ⋅ ∇ u \ell'_0(u)=\int_{\Omega}u\operatorname{div}X=-\int_{\Omega}X\cdot\nabla u ℓ 0 ′ ( u ) = ∫ Ω u div X = − ∫ Ω X ⋅ ∇ u , while (43) with f 1 = f 2 = 1 f_1=f_2=1 f 1 = f 2 = 1 gives a 0 ′ ( u , u ) = − 2 ∫ Ω X ⋅ ∇ u − c ∗ ∫ ∂ Ω ψ 2 φ a'_0(u,u)=-2\int_{\Omega}X\cdot\nabla u-c_*\int_{\partial\Omega}\psi^2\varphi a 0 ′ ( u , u ) = − 2 ∫ Ω X ⋅ ∇ u − c ∗ ∫ ∂ Ω ψ 2 φ . The volume integrals cancel, and (5) follows.
Corollary 3.7. With the notation previously introduced we have
a ( w , w ) = ∫ Ω f w X ⋅ ∇ u d y + c ∗ ∫ ∂ Ω ψ ψ w φ d H n − 1 . (44) a(w,w)=\int_{\Omega}f_wX\cdot\nabla u\,dy+c_*\int_{\partial\Omega}\psi\psi_w\varphi\,d\mathcal{H}^{n-1}. \tag*{(44)} a ( w , w ) = ∫ Ω f w X ⋅ ∇ u d y + c ∗ ∫ ∂ Ω ψ ψ w φ d H n − 1 . (44)
Proof. By Lemma 3.4, w = δ W w=\sqrt{\delta}W w = δ W with W W W smooth and f w f_w f w smooth, so Lemma 3.5 applies to v 1 = u v_1=u v 1 = u , v 2 = w v_2=w v 2 = w and gives
a 0 ′ ( u , w ) = − ∫ Ω X ⋅ ∇ w − ∫ Ω f w X ⋅ ∇ u − c ∗ ∫ ∂ Ω ψ ψ w φ . a'_0(u,w)=-\int_{\Omega}X\cdot\nabla w-\int_{\Omega}f_wX\cdot\nabla u-c_*\int_{\partial\Omega}\psi\psi_w\varphi. a 0 ′ ( u , w ) = − ∫ Ω X ⋅ ∇ w − ∫ Ω f w X ⋅ ∇ u − c ∗ ∫ ∂ Ω ψ ψ w φ .
The zero extension of w w w belongs to W 1 , 1 ( R n ) W^{1,1}(\mathbb{R}^n) W 1 , 1 ( R n ) and vanishes on ∂ Ω \partial\Omega ∂ Ω , so ℓ 0 ′ ( w ) = ∫ Ω w div X = − ∫ Ω X ⋅ ∇ w \ell'_0(w)=\int_{\Omega}w\operatorname{div}X=-\int_{\Omega}X\cdot\nabla w ℓ 0 ′ ( w ) = ∫ Ω w div X = − ∫ Ω X ⋅ ∇ w . Inserting both identities in (30) with v = w v=w v = w , that is a ( w , w ) = ℓ 0 ′ ( w ) − a 0 ′ ( u , w ) a(w,w)=\ell'_0(w)-a'_0(u,w) a ( w , w ) = ℓ 0 ′ ( w ) − a 0 ′ ( u , w ) , the two integrals of X ⋅ ∇ w X\cdot\nabla w X ⋅ ∇ w cancel and (44) follows.
From (44) we see that 2 a ( w , w ) 2a(w,w) 2 a ( w , w ) contains the volume integral 2 ∫ Ω f w X ⋅ ∇ u 2\int_{\Omega}f_wX\cdot\nabla u 2 ∫ Ω f w X ⋅ ∇ u . We show in the next subsection that G ′ ′ ( 0 ) G^{\prime\prime}(0) G ′′ ( 0 ) contains the same integral with the opposite sign.
The transported forcing
Let
u ~ t = u ∘ Φ t − 1 , f ~ t = ( − Δ ) 1 / 2 u ~ t . \widetilde{u}_t=u\circ\Phi_t^{-1},\qquad\widetilde{f}_t=(-\Delta)^{1/2}\widetilde{u}_t. u t = u ∘ Φ t − 1 , f t = ( − Δ ) 1/2 u t .
The function u ~ t \widetilde{u}_t u t is the torsion function of Ω \Omega Ω transported to Ω t \Omega_t Ω t by Φ t \Phi_t Φ t ; it is not the torsion function of Ω t \Omega_t Ω t . By (27) ,
u ~ t = δ ∘ Φ t − 1 ( U ∘ Φ t − 1 ) = δ t A t , A t : = γ t ( U ∘ Φ t − 1 ) ∈ C ∞ ( Ω t ‾ ) . (45) \widetilde{u}_t=\sqrt{\delta\circ\Phi_t^{-1}}\left(U\circ\Phi_t^{-1}\right)=\sqrt{\delta_t}\,A_t,\qquad A_t:=\sqrt{\gamma_t}\left(U\circ\Phi_t^{-1}\right)\in C^\infty(\overline{\Omega_t}). \tag*{(45)} u t = δ ∘ Φ t − 1 ( U ∘ Φ t − 1 ) = δ t A t , A t := γ t ( U ∘ Φ t − 1 ) ∈ C ∞ ( Ω t ) . (45)
and the second line of (20) , applied on the domain Ω t \Omega_t Ω t , gives f ~ t ∈ C ∞ ( Ω t ‾ ) \widetilde f_t \in C^\infty(\overline{\Omega_t}) f t ∈ C ∞ ( Ω t ) for each t t t . Pulling back by Φ t \Phi_t Φ t , we get f ~ t ∘ Φ t = L t u \widetilde f_t \circ\Phi_t=L_tu f t ∘ Φ t = L t u in Ω \Omega Ω , where
L t v ( y ) = c n P.V. ∫ R n v ( y ) − v ( z ) ∣ y − z ∣ n + 1 ϱ t ( y , z ) d z , ϱ t ( y , z ) = J t ( z ) ( ∣ y − z ∣ ∣ Φ t y − Φ t z ∣ ) n + 1 . (46) L_tv(y)=c_n\operatorname{P.V.}\int_{\mathbb{R}^n}\frac{v(y)-v(z)}{|y-z|^{n+1}}\varrho_t(y,z)\,dz,\qquad\varrho_t(y,z)=J_t(z)\left(\frac{|y-z|}{|\Phi_ty-\Phi_tz|}\right)^{n+1}. \tag*{(46)} L t v ( y ) = c n P.V. ∫ R n ∣ y − z ∣ n + 1 v ( y ) − v ( z ) ϱ t ( y , z ) d z , ϱ t ( y , z ) = J t ( z ) ( ∣ Φ t y − Φ t z ∣ ∣ y − z ∣ ) n + 1 . (46)
Here the principal value is taken over the balls B ε ( y ) B_\varepsilon(y) B ε ( y ) , whereas the change of variables produces the excised sets S ε = { z : ∣ Φ t z − Φ t y ∣ < ε } S_\varepsilon=\{z:|\Phi_tz-\Phi_ty|<\varepsilon\} S ε = { z : ∣ Φ t z − Φ t y ∣ < ε } . The two limits coincide for v v v smooth near the fixed point y ∈ Ω y\in\Omega y ∈ Ω , as is the case for u u u . Indeed, writing h = z − y h=z-y h = z − y , the leading singular term of the integrand is − J t ( y ) ∇ v ( y ) ⋅ h / ∣ D Φ t ( y ) h ∣ n + 1 -J_t(y)\nabla v(y)\cdot h/|D\Phi_t(y)h|^{n+1} − J t ( y ) ∇ v ( y ) ⋅ h /∣ D Φ t ( y ) h ∣ n + 1 , which is odd; the remainder is O ( ∣ h ∣ 1 − n ) O(|h|^{1-n}) O ( ∣ h ∣ 1 − n ) and is locally integrable. The odd term integrates to zero between a centred ball and the centred ellipsoid { ∣ D Φ t ( y ) h ∣ < ε } \{|D\Phi_t(y)h|<\varepsilon\} { ∣ D Φ t ( y ) h ∣ < ε } . By Taylor’s formula, replacing this ellipsoid by S ε − y S_\varepsilon-y S ε − y changes its radial boundary by O ( ε 2 ) O(\varepsilon^2) O ( ε 2 ) , producing an error O ( ε ) O(\varepsilon) O ( ε ) .
In (46) , the domain and the function are fixed, and only the kernel depends on t t t : ( L t ) (L_t) ( L t ) is a smooth family of operators of order one. Moreover L 0 = ( − Δ ) 1 / 2 L_0=(-\Delta)^{1/2} L 0 = ( − Δ ) 1/2 , so that L 0 u = 1 L_0u=1 L 0 u = 1 in Ω \Omega Ω , and ρ t ( y , z ) = J t ( y ) ϱ t ( y , z ) \rho_t(y,z)=J_t(y)\varrho_t(y,z) ρ t ( y , z ) = J t ( y ) ϱ t ( y , z ) , where ρ t \rho_t ρ t is the multiplier of (33) .
Lemma 3.8. Let X t = Φ ˙ t ∘ Φ t − 1 X_t=\dot{\Phi}_t\circ\Phi_t^{-1} X t = Φ ˙ t ∘ Φ t − 1 , X ^ t = ( D Φ t ) − 1 Φ ˙ t \widehat X_t=(D\Phi_t)^{-1}\dot{\Phi}_t X t = ( D Φ t ) − 1 Φ ˙ t , let ν t \nu_t ν t be the outer normal to Ω t \Omega_t Ω t , and let
B ( t ) = c ∗ ∫ ∂ Ω t ψ ~ t 2 ( X t ⋅ ν t ) d H n − 1 , ψ ~ t = u ~ t δ t ∣ ∂ Ω t . B(t)=c_*\int_{\partial\Omega_t}\widetilde{\psi}_t^2(X_t\cdot\nu_t)\,d\mathcal{H}^{n-1},\qquad\widetilde{\psi}_t=\left.\frac{\widetilde u_t}{\sqrt{\delta_t}}\right|_{\partial\Omega_t}. B ( t ) = c ∗ ∫ ∂ Ω t ψ t 2 ( X t ⋅ ν t ) d H n − 1 , ψ t = δ t u t ∂ Ω t . Then, for all small t t t ,
G ′ ( t ) = 2 ∫ Ω ( L t u − 1 ) X ^ t ⋅ ∇ u J t d y + B ( t ) . (47) G'(t)=2\int_{\Omega}(L_tu-1)\widehat X_t\cdot\nabla u\,J_t\,dy+B(t). \tag*{(47)} G ′ ( t ) = 2 ∫ Ω ( L t u − 1 ) X t ⋅ ∇ u J t d y + B ( t ) . (47)
Proof. We have G ( t ) = 2 ∫ Ω t u ~ t − a ( u ~ t , u ~ t ) G(t)=2\int_{\Omega_t}\widetilde u_t-a(\widetilde u_t,\widetilde u_t) G ( t ) = 2 ∫ Ω t u t − a ( u t , u t ) . Since ∂ t u ~ t = − X t ⋅ ∇ u ~ t \partial_t\widetilde u_t=-X_t\cdot\nabla\widetilde u_t ∂ t u t = − X t ⋅ ∇ u t and u ~ t \widetilde u_t u t vanishes on ∂ Ω t \partial\Omega_t ∂ Ω t ,
d d t ∫ Ω t u ~ t = − ∫ Ω t X t ⋅ ∇ u ~ t . \frac{d}{dt}\int_{\Omega_t}\widetilde u_t=-\int_{\Omega_t}X_t\cdot\nabla\widetilde u_t. d t d ∫ Ω t u t = − ∫ Ω t X t ⋅ ∇ u t .
For the second term, let Ψ h = Φ t + h ∘ Φ t − 1 \Psi_h=\Phi_{t+h}\circ\Phi_t^{-1} Ψ h = Φ t + h ∘ Φ t − 1 , so that Ψ 0 = Id \Psi_0=\operatorname{Id} Ψ 0 = Id , ∂ h Ψ h ∣ h = 0 = X t \left.\partial_h\Psi_h\right|_{h=0}=X_t ∂ h Ψ h ∣ h = 0 = X t , and a ( u ~ t + h , u ~ t + h ) = a ( u ~ t ∘ Ψ h − 1 , u ~ t ∘ Ψ h − 1 ) a(\widetilde u_{t+h},\widetilde u_{t+h})=a(\widetilde u_t\circ\Psi_h^{-1},\widetilde u_t\circ\Psi_h^{-1}) a ( u t + h , u t + h ) = a ( u t ∘ Ψ h − 1 , u t ∘ Ψ h − 1 ) . Hence d d t a ( u ~ t , u ~ t ) \frac{d}{dt}a(\widetilde u_t,\widetilde u_t) d t d a ( u t , u t ) is the first-order form a 0 ′ [ X t ] a_0'[X_t] a 0 ′ [ X t ] , computed on the domain Ω t \Omega_t Ω t , applied to ( u ~ t , u ~ t ) (\widetilde u_t,\widetilde u_t) ( u t , u t ) . By Lemma 3.5 on Ω t \Omega_t Ω t , with v 1 = v 2 = u ~ t = δ t A t v_1=v_2=\widetilde u_t=\sqrt{\delta_t}A_t v 1 = v 2 = u t = δ t A t (see (45) ) and regularised distance δ t \delta_t δ t ,
d d t a ( u ~ t , u ~ t ) = − 2 ∫ Ω t f ~ t X t ⋅ ∇ u ~ t − B ( t ) . \frac{d}{dt}a(\widetilde u_t,\widetilde u_t)=-2\int_{\Omega_t}\widetilde f_tX_t\cdot\nabla\widetilde u_t-B(t). d t d a ( u t , u t ) = − 2 ∫ Ω t f t X t ⋅ ∇ u t − B ( t ) .
Combining the two derivatives,
G ′ ( t ) = 2 ∫ Ω t ( f ~ t − 1 ) X t ⋅ ∇ u ~ t + B ( t ) . G'(t)=2\int_{\Omega_t}(\widetilde f_t-1)X_t\cdot\nabla\widetilde u_t+B(t). G ′ ( t ) = 2 ∫ Ω t ( f t − 1 ) X t ⋅ ∇ u t + B ( t ) .
Changing variables by Φ t \Phi_t Φ t and using f ~ t ∘ Φ t = L t u \widetilde f_t\circ\Phi_t=L_tu f t ∘ Φ t = L t u and ( X t ⋅ ∇ u ~ t ) ∘ Φ t = X ^ t ⋅ ∇ u (X_t\cdot\nabla\widetilde u_t)\circ\Phi_t=\widehat X_t\cdot\nabla u ( X t ⋅ ∇ u t ) ∘ Φ t = X t ⋅ ∇ u , we obtain (47) . □
Since L 0 u − 1 = 0 L_0u-1=0 L 0 u − 1 = 0 , the derivatives of X ^ t \widehat X_t X t and J t J_t J t do not contribute to G ′ ′ ( 0 ) G''(0) G ′′ ( 0 ) : we only need the limit of ( L t u − 1 ) / t (L_tu-1)/t ( L t u − 1 ) / t , tested against X ⋅ ∇ u X\cdot\nabla u X ⋅ ∇ u . The boundary term is explicit.
Lemma 3.9. We have ψ ~ t ∘ Φ t = β t ψ \widetilde{\psi}_t\circ\Phi_t=\sqrt{\beta_t}\,\psi ψ t ∘ Φ t = β t ψ on ∂ Ω \partial\Omega ∂ Ω , and
B ′ ( 0 ) = c ∗ ∫ ∂ Ω H ψ 2 φ 2 d H n − 1 . (48) B'(0)=c_*\int_{\partial\Omega}H\psi^2\varphi^2\,d\mathcal{H}^{n-1}. \tag*{(48)} B ′ ( 0 ) = c ∗ ∫ ∂ Ω H ψ 2 φ 2 d H n − 1 . (48)
Proof. By (45) and (28) , ψ ~ t ( x t ) = A t ( x t ) = γ t ( x t ) U ( x ) = β t ( x ) ψ ( x ) \widetilde{\psi}_t(x_t)=A_t(x_t)=\sqrt{\gamma_t(x_t)}U(x)=\sqrt{\beta_t(x)}\psi(x) ψ t ( x t ) = A t ( x t ) = γ t ( x t ) U ( x ) = β t ( x ) ψ ( x ) . On ∂ Ω \partial\Omega ∂ Ω , Φ t ( x ) = x t \Phi_t(x)=x_t Φ t ( x ) = x t and X t ( x t ) = Φ ˙ t ( x ) = φ ( x ) ν ( x ) X_t(x_t)=\dot{\Phi}_t(x)=\varphi(x)\nu(x) X t ( x t ) = Φ ˙ t ( x ) = φ ( x ) ν ( x ) . Pulling back by x ↦ x t x\mapsto x_t x ↦ x t ,
B ( t ) = c ∗ ∫ ∂ Ω ψ 2 β t φ ν ⋅ ν t ( x t ) J t ∂ d H n − 1 , B(t)=c_*\int_{\partial\Omega}\psi^2\beta_t\varphi\nu\cdot\nu_t(x_t)J_t^{\partial}\,d\mathcal{H}^{n-1}, B ( t ) = c ∗ ∫ ∂ Ω ψ 2 β t φ ν ⋅ ν t ( x t ) J t ∂ d H n − 1 ,
where J t ∂ J_t^{\partial} J t ∂ is the surface Jacobian of x ↦ x t x\mapsto x_t x ↦ x t . By (29) , β 0 = 1 \beta_0=1 β 0 = 1 and ∂ t β t ∣ t = 0 = 0 \left.\partial_t\beta_t\right|_{t=0}=0 ∂ t β t ∣ t = 0 = 0 . Since ν t ( x t ) \nu_t(x_t) ν t ( x t ) is a unit vector, ν ⋅ ∂ t ν t ( x t ) ∣ t = 0 = 0 \left.\nu\cdot\partial_t\nu_t(x_t)\right|_{t=0}=0 ν ⋅ ∂ t ν t ( x t ) ∣ t = 0 = 0 , and ∂ t J t ∂ ∣ t = 0 = div ∂ Ω ( φ ν ) = H φ \left.\partial_tJ_t^{\partial}\right|_{t=0}=\operatorname{div}_{\partial\Omega}(\varphi\nu)=H\varphi ∂ t J t ∂ t = 0 = div ∂ Ω ( φ ν ) = H φ . This gives (48) .
We now identify the limit of ( L t u − 1 ) / t (L_tu-1)/t ( L t u − 1 ) / t . Recall from Lemma 3.4 that f w = ( ( − Δ ) 1 / 2 w ) ∣ Ω ∈ C ∞ ( Ω ‾ ) f_w=((-\Delta)^{1/2}w)|_{\Omega}\in C^\infty(\overline{\Omega}) f w = (( − Δ ) 1/2 w ) ∣ Ω ∈ C ∞ ( Ω ) .
Lemma 3.10. ( f ~ t − 1 ) / t → − f w (\widetilde{f}_t-1)/t\to-f_w ( f t − 1 ) / t → − f w in D ′ ( Ω ) D'(\Omega) D ′ ( Ω ) as t → 0 t\to0 t → 0 .
Proof. Since u ∈ W 1 , 1 ( R n ) u\in W^{1,1}(\mathbb{R}^n) u ∈ W 1 , 1 ( R n ) and Φ t = Id + t X \Phi_t=\operatorname{Id}+tX Φ t = Id + tX with X X X smooth and compactly supported, ( u ~ t − u ) / t → − X ⋅ ∇ u (\widetilde{u}_t-u)/t\to-X\cdot\nabla u ( u t − u ) / t → − X ⋅ ∇ u in L 1 ( R n ) L^1(\mathbb{R}^n) L 1 ( R n ) . Let χ ∈ C c ∞ ( Ω ) \chi\in C^\infty_c(\Omega) χ ∈ C c ∞ ( Ω ) . For ∣ t ∣ |t| ∣ t ∣ small the support of χ \chi χ is contained in Ω t \Omega_t Ω t , so that ∫ u ~ t ( − Δ ) 1 / 2 χ = ∫ f ~ t χ \int\widetilde{u}_t(-\Delta)^{1/2}\chi=\int\widetilde{f}_t\chi ∫ u t ( − Δ ) 1/2 χ = ∫ f t χ and ∫ u ( − Δ ) 1 / 2 χ = ∫ f χ \int u(-\Delta)^{1/2}\chi=\int f\chi ∫ u ( − Δ ) 1/2 χ = ∫ f χ . Since ( − Δ ) 1 / 2 χ (-\Delta)^{1/2}\chi ( − Δ ) 1/2 χ is bounded,
∫ f ~ t − 1 t χ ⟶ − ∫ X ⋅ ∇ u ( − Δ ) 1 / 2 χ . \int\frac{\widetilde{f}_t-1}{t}\chi\longrightarrow-\int X\cdot\nabla u(-\Delta)^{1/2}\chi. ∫ t f t − 1 χ ⟶ − ∫ X ⋅ ∇ u ( − Δ ) 1/2 χ .
By (38) , − X ⋅ ∇ u = q − w -X\cdot\nabla u=q-w − X ⋅ ∇ u = q − w . Since q q q is $1/2$-harmonic in Ω \Omega Ω (Proposition 3.3) and ( − Δ ) 1 / 2 w = f w (-\Delta)^{1/2}w=f_w ( − Δ ) 1/2 w = f w in $\Omega, the right-hand side equals − ∫ f w χ -\int f_w\chi − ∫ f w χ .
This convergence is not enough to pass to the limit in (47) , because X ⋅ ∇ u X\cdot\nabla u X ⋅ ∇ u behaves like d − 1 / 2 d^{-1/2} d − 1/2 : it is integrable, but it is not a test function. We need the following estimate.
Lemma 3.11. Let U ∈ C ∞ ( Ω ‾ ) U\in C^\infty(\overline{\Omega}) U ∈ C ∞ ( Ω ) , u = δ U u=\sqrt{\delta}U u = δ U , and let L t L_t L t be as in (46) . There are t 0 > 0 t_0>0 t 0 > 0 and C C C , depending only on Ω \Omega Ω , δ \delta δ , ∥ U ∥ C 2 ( Ω ‾ ) \|U\|_{C^2(\overline{\Omega})} ∥ U ∥ C 2 ( Ω ) and sup ∣ t ∣ ≤ t 0 ∣ t ∣ − 1 ∥ Φ t − Id ∥ C 2 \sup_{|t|\leq t_0}|t|^{-1}\|\Phi_t-\operatorname{Id}\|_{C^2} sup ∣ t ∣ ≤ t 0 ∣ t ∣ − 1 ∥ Φ t − Id ∥ C 2 , such that
∥ L t u − L 0 u ∥ L ∞ ( Ω ) ≤ C ∣ t ∣ , ∣ t ∣ ≤ t 0 . (49) \|L_tu-L_0u\|_{L^\infty(\Omega)}\leq C|t|,\qquad|t|\leq t_0. \tag*{(49)} ∥ L t u − L 0 u ∥ L ∞ ( Ω ) ≤ C ∣ t ∣ , ∣ t ∣ ≤ t 0 . (49)
The difficulty is the boundary behaviour of u u u : ∣ ∇ u ∣ ≲ d − 1 / 2 |\nabla u|\lesssim d^{-1/2} ∣∇ u ∣ ≲ d − 1/2 and ∣ D 2 u ∣ ≲ d − 3 / 2 |D^2u|\lesssim d^{-3/2} ∣ D 2 u ∣ ≲ d − 3/2 . If one splits the principal value in (46) at ∣ y − z ∣ = d ( y ) / 2 |y-z|=d(y)/2 ∣ y − z ∣ = d ( y ) /2 , freezes the kernel at y y y , and expands u ( y ) − u ( z ) u(y)-u(z) u ( y ) − u ( z ) to second order, the linear term cancels by parity of the frozen kernel, but the quadratic term only gives ∣ L t u − L 0 u ∣ ≲ ∣ t ∣ d − 1 / 2 |L_tu-L_0u|\lesssim|t|d^{-1/2} ∣ L t u − L 0 u ∣ ≲ ∣ t ∣ d − 1/2 . The proof uses instead the exact cancellation of Lemma 2.5.
Proof. We freeze the moving kernel at the evaluation point. Let
K t ( y , z ) = J t ( z ) ∣ Φ t z − Φ t y ∣ n + 1 , K t f r ( y , z ) = J t ( y ) ∣ D Φ t ( y ) ( z − y ) ∣ n + 1 , E t = K t − K t f r , K_t(y,z)=\frac{J_t(z)}{|\Phi_tz-\Phi_ty|^{n+1}},\qquad K_t^{\mathrm{fr}}(y,z)=\frac{J_t(y)}{|D\Phi_t(y)(z-y)|^{n+1}},\qquad E_t=K_t-K_t^{\mathrm{fr}}, K t ( y , z ) = ∣ Φ t z − Φ t y ∣ n + 1 J t ( z ) , K t fr ( y , z ) = ∣ D Φ t ( y ) ( z − y ) ∣ n + 1 J t ( y ) , E t = K t − K t fr ,
so that
( L t − L 0 ) u ( y ) = c n P.V. ∫ R n ( u ( y ) − u ( z ) ) b t , y ( ( z − y ) / ∣ z − y ∣ ) ∣ z − y ∣ n + 1 d z + c n P.V. ∫ R n ( u ( y ) − u ( z ) ) E t ( y , z ) d z , (L_t-L_0)u(y)=c_n\operatorname{P.V.}\int_{\mathbb{R}^n}(u(y)-u(z))\frac{b_{t,y}((z-y)/|z-y|)}{|z-y|^{n+1}}\,dz
+c_n\operatorname{P.V.}\int_{\mathbb{R}^n}(u(y)-u(z))E_t(y,z)\,dz, ( L t − L 0 ) u ( y ) = c n P.V. ∫ R n ( u ( y ) − u ( z )) ∣ z − y ∣ n + 1 b t , y (( z − y ) /∣ z − y ∣ ) d z + c n P.V. ∫ R n ( u ( y ) − u ( z )) E t ( y , z ) d z ,
where
b t , y ( θ ) = J t ( y ) ∣ D Φ t ( y ) θ ∣ n + 1 − 1 , θ ∈ S n − 1 . b_{t,y}(\theta)=\frac{J_t(y)}{|D\Phi_t(y)\theta|^{n+1}}-1,\qquad\theta\in S^{n-1}. b t , y ( θ ) = ∣ D Φ t ( y ) θ ∣ n + 1 J t ( y ) − 1 , θ ∈ S n − 1 .
which is even and satisfies sup y ∥ b t , y ∥ ∞ ≤ C ∣ t ∣ \sup_y \|b_{t,y}\|_\infty\le C|t| sup y ∥ b t , y ∥ ∞ ≤ C ∣ t ∣ . We estimate each of the terms separately. By Lemma 2.5, applied at each fixed y y y with b = b t , y b=b_{t,y} b = b t , y ,
∣ P b t , y u ( y ) ∣ ≤ C ∥ b t , y ∥ ∞ ≤ C ∣ t ∣ . |\mathcal{P}_{b_{t,y}}u(y)| \le C\|b_{t,y}\|_\infty\le C|t|. ∣ P b t , y u ( y ) ∣ ≤ C ∥ b t , y ∥ ∞ ≤ C ∣ t ∣.
We now turn to the term involving E t E_t E t . Since Φ 0 = Id \Phi_0=\operatorname{Id} Φ 0 = Id , we have
∥ D Φ t − Id ∥ ∞ + ∥ D 2 Φ t ∥ ∞ + ∥ J t − 1 ∥ ∞ + ∥ ∇ J t ∥ ∞ ≤ C ∣ t ∣ . \|D\Phi_t-\operatorname{Id}\|_\infty+\|D^2\Phi_t\|_\infty+\|J_t-1\|_\infty+\|\nabla J_t\|_\infty\le C|t|. ∥ D Φ t − Id ∥ ∞ + ∥ D 2 Φ t ∥ ∞ + ∥ J t − 1 ∥ ∞ + ∥∇ J t ∥ ∞ ≤ C ∣ t ∣.
Using ∣ Φ t z − Φ t y − D Φ t ( y ) ( z − y ) ∣ ≤ C ∣ t ∣ ∣ z − y ∣ 2 |\Phi_tz-\Phi_ty-D\Phi_t(y)(z-y)|\le C|t||z-y|^2 ∣ Φ t z − Φ t y − D Φ t ( y ) ( z − y ) ∣ ≤ C ∣ t ∣∣ z − y ∣ 2 , ∣ J t ( z ) − J t ( y ) ∣ ≤ C ∣ t ∣ ∣ z − y ∣ |J_t(z)-J_t(y)|\le C|t||z-y| ∣ J t ( z ) − J t ( y ) ∣ ≤ C ∣ t ∣∣ z − y ∣ , and the fact that both K t K_t K t and K t f r K_t^{\mathrm{fr}} K t fr differ from ∣ z − y ∣ − n − 1 |z-y|^{-n-1} ∣ z − y ∣ − n − 1 by at most ∣ t ∣ ∣ z − y ∣ − n − 1 |t||z-y|^{-n-1} ∣ t ∣∣ z − y ∣ − n − 1 , we get
∣ E t ( y , z ) ∣ ≤ C ∣ t ∣ { ∣ z − y ∣ − n , 0 < ∣ z − y ∣ ≤ 1 , ∣ z − y ∣ − n − 1 , ∣ z − y ∣ > 1. |E_t(y,z)|\le C|t|
\begin{cases}
|z-y|^{-n}, & 0<|z-y|\le1,\\
|z-y|^{-n-1}, & |z-y|>1.
\end{cases} ∣ E t ( y , z ) ∣ ≤ C ∣ t ∣ { ∣ z − y ∣ − n , ∣ z − y ∣ − n − 1 , 0 < ∣ z − y ∣ ≤ 1 , ∣ z − y ∣ > 1.
The error kernel is one order less singular than ∣ z − y ∣ − n − 1 |z-y|^{-n-1} ∣ z − y ∣ − n − 1 , and the global Hölder continuity of u u u is enough to control it. Indeed, the zero extension of δ \delta δ is Lipschitz, so the zero extension of u u u is C 1 / 2 C^{1/2} C 1/2 on R n \mathbb{R}^n R n , and
sup y ∈ Ω ∣ ∫ R n ( u ( y ) − u ( z ) ) E t ( y , z ) d z ∣ ≤ C ∣ t ∣ ( [ u ] C 1 / 2 ( R n ) ∫ 0 1 ρ − 1 / 2 d ρ + ∥ u ∥ ∞ ∫ 1 ∞ ρ − 2 d ρ ) ≤ C ∣ t ∣ . \sup_{y\in\Omega}\left|\int_{\mathbb{R}^n}(u(y)-u(z))E_t(y,z)\,dz\right|
\le C|t|\left([u]_{C^{1/2}(\mathbb{R}^n)}\int_0^1\rho^{-1/2}\,d\rho+\|u\|_\infty\int_1^\infty\rho^{-2}\,d\rho\right)\le C|t|. y ∈ Ω sup ∫ R n ( u ( y ) − u ( z )) E t ( y , z ) d z ≤ C ∣ t ∣ ( [ u ] C 1/2 ( R n ) ∫ 0 1 ρ − 1/2 d ρ + ∥ u ∥ ∞ ∫ 1 ∞ ρ − 2 d ρ ) ≤ C ∣ t ∣.
This proves (49) . □
Proposition 3.12. We have
G ′ ′ ( 0 ) = − 2 ∫ Ω f w X ⋅ ∇ u d y + c ∗ ∫ ∂ Ω H ψ 2 φ 2 d H n − 1 . (50) \mathcal{G}''(0)=-2\int_\Omega f_w X\cdot\nabla u\,dy+c_*\int_{\partial\Omega}H\psi^2\varphi^2\,d\mathcal{H}^{n-1}. \tag*{(50)} G ′′ ( 0 ) = − 2 ∫ Ω f w X ⋅ ∇ u d y + c ∗ ∫ ∂ Ω H ψ 2 φ 2 d H n − 1 . (50)
Proof. Let g t = ( L t u − 1 ) / t g_t=(L_tu-1)/t g t = ( L t u − 1 ) / t . By Lemma 3.11 and L 0 u = 1 L_0u=1 L 0 u = 1 , ∥ g t ∥ L ∞ ( Ω ) ≤ C \|g_t\|_{L^\infty(\Omega)}\le C ∥ g t ∥ L ∞ ( Ω ) ≤ C . Let χ ∈ C c ∞ ( Ω ) \chi\in C^\infty_c(\Omega) χ ∈ C c ∞ ( Ω ) and χ t = ( χ ∘ Φ t − 1 ) ∣ det D Φ t − 1 ∣ \chi_t=(\chi\circ\Phi_t^{-1})|\det D\Phi_t^{-1}| χ t = ( χ ∘ Φ t − 1 ) ∣ det D Φ t − 1 ∣ . Since g t = ( ( f ~ t − 1 ) / t ) ∘ Φ t g_t=((\widetilde f_t-1)/t)\circ\Phi_t g t = (( f t − 1 ) / t ) ∘ Φ t and ∥ ( f ~ t − 1 ) / t ∥ L ∞ ( Ω t ) = ∥ g t ∥ L ∞ ( Ω ) \|(\widetilde f_t-1)/t\|_{L^\infty(\Omega_t)}=\|g_t\|_{L^\infty(\Omega)} ∥ ( f t − 1 ) / t ∥ L ∞ ( Ω t ) = ∥ g t ∥ L ∞ ( Ω ) ,
∫ Ω g t χ = ∫ f ~ t − 1 t χ t = ∫ f ~ t − 1 t χ + O ( ∥ χ t − χ ∥ L 1 ) , \int_\Omega g_t\chi=\int\frac{\widetilde f_t-1}{t}\chi_t=\int\frac{\widetilde f_t-1}{t}\chi+O(\|\chi_t-\chi\|_{L^1}), ∫ Ω g t χ = ∫ t f t − 1 χ t = ∫ t f t − 1 χ + O ( ∥ χ t − χ ∥ L 1 ) ,
and Lemma 3.10 gives ∫ g t χ → − ∫ f w χ \int g_t\chi\to-\int f_w\chi ∫ g t χ → − ∫ f w χ . Since ( g t ) (g_t) ( g t ) is bounded in L ∞ ( Ω ) L^\infty(\Omega) L ∞ ( Ω ) and C c ∞ ( Ω ) C^\infty_c(\Omega) C c ∞ ( Ω ) is dense in L 1 ( Ω ) L^1(\Omega) L 1 ( Ω ) ,
g t ⇀ ∗ − f w in L ∞ ( Ω ) . g_t \overset{*}{\rightharpoonup}-f_w\qquad\text{in }L^\infty(\Omega). g t ⇀ ∗ − f w in L ∞ ( Ω ) .
On the other hand X ^ t → X \widehat X_t\to X X t → X and J t → 1 J_t\to1 J t → 1 uniformly, and ∣ ∇ u ∣ ≤ C δ − 1 / 2 ∈ L 1 ( Ω ) |\nabla u|\le C\delta^{-1/2}\in L^1(\Omega) ∣∇ u ∣ ≤ C δ − 1/2 ∈ L 1 ( Ω ) , so X ^ t ⋅ ∇ u J t → X ⋅ ∇ u \widehat X_t\cdot\nabla u\,J_t\to X\cdot\nabla u X t ⋅ ∇ u J t → X ⋅ ∇ u strongly in L 1 ( Ω ) L^1(\Omega) L 1 ( Ω ) . Hence
∫ Ω g t X ^ t ⋅ ∇ u J t ⟶ − ∫ Ω f w X ⋅ ∇ u . (51) \int_\Omega g_t\widehat X_t\cdot\nabla u\,J_t\longrightarrow-\int_\Omega f_wX\cdot\nabla u. \tag*{(51)} ∫ Ω g t X t ⋅ ∇ u J t ⟶ − ∫ Ω f w X ⋅ ∇ u . (51)
By Proposition 3.2 , G \mathcal{G} G is of class C 2 C^2 C 2 , and by Lemma 3.8, G ′ ( 0 ) = B ( 0 ) \mathcal{G}'(0)=B(0) G ′ ( 0 ) = B ( 0 ) . Therefore
G ′ ′ ( 0 ) = lim t → 0 G ′ ( t ) − G ′ ( 0 ) t = 2 lim t → 0 ∫ Ω g t X ^ t ⋅ ∇ u J t + B ′ ( 0 ) , \mathcal{G}''(0)=\lim_{t\to0}\frac{\mathcal{G}'(t)-\mathcal{G}'(0)}{t}
=2\lim_{t\to0}\int_\Omega g_t\widehat X_t\cdot\nabla u\,J_t+B'(0), G ′′ ( 0 ) = t → 0 lim t G ′ ( t ) − G ′ ( 0 ) = 2 t → 0 lim ∫ Ω g t X t ⋅ ∇ u J t + B ′ ( 0 ) ,
so that (51) and (48) give (50) . □
Conclusion of the proof
Proof of Theorem 1.1 . By Proposition 3.2 , t ↦ T ( Ω t ) t \mapsto T(\Omega_t) t ↦ T ( Ω t ) is of class C 2 C^2 C 2 . Adding (50) and twice (44) , the volume integrals cancel, and
d 2 d t 2 T ( Ω t ) ∣ t = 0 = c ∗ ∫ ∂ Ω { H ψ 2 φ 2 + 2 ψ φ ψ w } d H n − 1 . \left.\frac{d^2}{dt^2}T(\Omega_t)\right|_{t=0}=c_*\int_{\partial\Omega}\left\{H\psi^2\varphi^2+2\psi\varphi\psi_w\right\}\,d\mathcal{H}^{n-1}. d t 2 d 2 T ( Ω t ) t = 0 = c ∗ ∫ ∂ Ω { H ψ 2 φ 2 + 2 ψ φ ψ w } d H n − 1 .
For the normal extension, ψ w = 1 2 D Ω ( ψ φ ) − 3 2 κ φ \psi_w=\frac{1}{2}\mathcal{D}_{\Omega}(\psi\varphi)-\frac{3}{2}\kappa\varphi ψ w = 2 1 D Ω ( ψ φ ) − 2 3 κ φ by Lemma 3.4 and (39) , which gives (9) . □
The case of the ball
In this section we prove Theorem 1.2 . We first compute explicitly the first and second shape derivatives at the ball. Let B = B 1 B=B_1 B = B 1 , so that ν ( x ) = x \nu(x)=x ν ( x ) = x and H = n − 1 H=n-1 H = n − 1 . We use the notation P , Λ , φ ℓ \mathcal{P},\Lambda,\varphi_\ell P , Λ , φ ℓ introduced in Section 1. The torsion function of B B B is given by [12 , 14 ]
u ( y ) = A n 1 − ∣ y ∣ 2 , with A n = Γ ( n / 2 ) π Γ ( ( n + 1 ) / 2 ) , ψ = 2 A n , κ = − ψ 4 . (52) u(y)=A_n\sqrt{1-|y|^2},\qquad\text{with }A_n=\frac{\Gamma(n/2)}{\sqrt{\pi}\,\Gamma((n+1)/2)},\qquad\psi=\sqrt{2}A_n,\qquad\kappa=-\frac{\psi}{4}. \tag*{(52)} u ( y ) = A n 1 − ∣ y ∣ 2 , with A n = π Γ (( n + 1 ) /2 ) Γ ( n /2 ) , ψ = 2 A n , κ = − 4 ψ . (52)
The boundary coefficients follow by expanding 1 − ∣ x − r ν ( x ) ∣ 2 = 2 r ( 1 − r / 4 + O ( r 2 ) ) \sqrt{1-|x-r\nu(x)|^2}=\sqrt{2r}(1-r/4+O(r^2)) 1 − ∣ x − r ν ( x ) ∣ 2 = 2 r ( 1 − r /4 + O ( r 2 )) . In particular λ ∗ = c ∗ ψ 2 = π 4 A n 2 \lambda_*=c_*\psi^2=\frac{\pi}{4}A_n^2 λ ∗ = c ∗ ψ 2 = 4 π A n 2 .
Lemma 4.1. For f ∈ C ∞ ( S n − 1 ) f\in C^\infty(\mathbb{S}^{n-1}) f ∈ C ∞ ( S n − 1 ) ,
Q f ( y ) = 2 1 − ∣ y ∣ 2 P f ( y ) ( y ∈ B ) , D B f = f 4 − Λ f . Q_f(y)=\frac{\sqrt{2}}{\sqrt{1-|y|^2}}\mathcal{P}f(y)\quad(y\in B),\qquad\mathcal{D}_Bf=\frac{f}{4}-\Lambda f. Q f ( y ) = 1 − ∣ y ∣ 2 2 P f ( y ) ( y ∈ B ) , D B f = 4 f − Λ f . Proof. The Martin kernel of B B B for ( − Δ ) 1 / 2 (-\Delta)^{1/2} ( − Δ ) 1/2 is proportional to ( 1 − ∣ y ∣ 2 ) 1 / 2 ∣ y − z ∣ − n (1-|y|^2)^{1/2}|y-z|^{-n} ( 1 − ∣ y ∣ 2 ) 1/2 ∣ y − z ∣ − n , z ∈ S n − 1 z\in\mathbb{S}^{n-1} z ∈ S n − 1 [2 (1.8)] ; see also [1 ] . Comparing it with the classical Poisson kernel shows that the displayed function, extended by zero, is 1 2 \frac{1}{2} 2 1 -harmonic in B B B . It has smooth weighted trace f f f , so Proposition 2.4 identifies it with Q f Q_f Q f . Moreover,
r Q f ( x − r ν ) = ( 1 − r 2 ) − 1 / 2 P f ( x − r ν ) = f + r ( f 4 − Λ f ) + O ( r 2 ) , \sqrt{r}\,Q_f(x-r\nu)=\left(1-\frac{r}{2}\right)^{-1/2}\mathcal{P}f(x-r\nu)=f+r\left(\frac{f}{4}-\Lambda f\right)+O(r^2), r Q f ( x − r ν ) = ( 1 − 2 r ) − 1/2 P f ( x − r ν ) = f + r ( 4 f − Λ f ) + O ( r 2 ) , which gives the formula for D B \mathcal{D}_B D B . □
Recall the notation Ω h \Omega_h Ω h from the introduction and set
J ( h ) = T ( Ω h ) , V ( h ) = ∣ Ω h ∣ . \mathcal{J}(h)=T(\Omega_h),\qquad\mathcal{V}(h)=|\Omega_h|. J ( h ) = T ( Ω h ) , V ( h ) = ∣ Ω h ∣.
For φ ∈ C ∞ ( S n − 1 ) \varphi\in C^\infty(\mathbb{S}^{n-1}) φ ∈ C ∞ ( S n − 1 ) we write T φ ( t ) = J ( t φ ) T_\varphi(t)=\mathcal{J}(t\varphi) T φ ( t ) = J ( tφ ) and V φ ( t ) = V ( t φ ) V_\varphi(t)=\mathcal{V}(t\varphi) V φ ( t ) = V ( tφ ) . Combining (5) , Theorem 1.1 together with (52) and Lemma 4.1 yields
T φ ′ ( 0 ) = λ ∗ ∫ S n − 1 φ d H n − 1 , T φ ′ ′ ( 0 ) = λ ∗ ∫ S n − 1 ( n φ 2 − φ Λ φ ) d H n − 1 . (53) T_\varphi'(0)=\lambda_*\int_{\mathbb{S}^{n-1}}\varphi\,d\mathcal{H}^{n-1},\qquad T_\varphi''(0)=\lambda_*\int_{\mathbb{S}^{n-1}}\left(n\varphi^2-\varphi\Lambda\varphi\right)\,d\mathcal{H}^{n-1}. \tag*{(53)} T φ ′ ( 0 ) = λ ∗ ∫ S n − 1 φ d H n − 1 , T φ ′′ ( 0 ) = λ ∗ ∫ S n − 1 ( n φ 2 − φ Λ φ ) d H n − 1 . (53)
On the other hand, polar coordinates give
V ( h ) = 1 n ∫ S n − 1 ( 1 + h ) n d H n − 1 , (54) \mathcal{V}(h)=\frac{1}{n}\int_{\mathbb{S}^{n-1}}(1+h)^n\,d\mathcal{H}^{n-1}, \tag*{(54)} V ( h ) = n 1 ∫ S n − 1 ( 1 + h ) n d H n − 1 , (54)
and hence
V φ ′ ( 0 ) = ∫ S n − 1 φ d H n − 1 , V φ ′ ′ ( 0 ) = ( n − 1 ) ∫ S n − 1 φ 2 d H n − 1 . (55) V_\varphi'(0)=\int_{\mathbb{S}^{n-1}}\varphi\,d\mathcal{H}^{n-1},\qquad V_\varphi''(0)=(n-1)\int_{\mathbb{S}^{n-1}}\varphi^2\,d\mathcal{H}^{n-1}. \tag*{(55)} V φ ′ ( 0 ) = ∫ S n − 1 φ d H n − 1 , V φ ′′ ( 0 ) = ( n − 1 ) ∫ S n − 1 φ 2 d H n − 1 . (55)
Lemma 4.2. Let t ↦ h ( t ) t \mapsto h(t) t ↦ h ( t ) be C 2 C^2 C 2 with values in C m ( S n − 1 ) C^m(\mathbb{S}^{n-1}) C m ( S n − 1 ) for every m m m , with h ( 0 ) = 0 h(0)=0 h ( 0 ) = 0 and ∣ Ω h ( t ) ∣ = ∣ B ∣ |\Omega_{h(t)}|=|B| ∣ Ω h ( t ) ∣ = ∣ B ∣ . Then φ = h ˙ ( 0 ) \varphi=\dot{h}(0) φ = h ˙ ( 0 ) has zero mean and
− d 2 d t 2 T ( Ω h ( t ) ) ∣ t = 0 = λ ∗ ∫ S n − 1 ( φ Λ φ − φ 2 ) d H n − 1 . (56) -\left.\frac{d^2}{dt^2}T(\Omega_{h(t)})\right|_{t=0}=\lambda_*\int_{\mathbb{S}^{n-1}}(\varphi\Lambda\varphi-\varphi^2)\,d\mathcal{H}^{n-1}. \tag*{(56)} − d t 2 d 2 T ( Ω h ( t ) ) t = 0 = λ ∗ ∫ S n − 1 ( φ Λ φ − φ 2 ) d H n − 1 . (56) Proof. Differentiating V ( h ( t ) ) = ∣ B ∣ \mathcal{V}(h(t))=|B| V ( h ( t )) = ∣ B ∣ at t = 0 t=0 t = 0 with (54) gives ∫ S n − 1 φ = 0 \int_{\mathbb{S}^{n-1}}\varphi=0 ∫ S n − 1 φ = 0 . We compare − T -T − T with the Lagrangian
L ( h ) = − J ( h ) + λ ∗ ( V ( h ) − ∣ B ∣ ) . \mathcal{L}(h)=-\mathcal{J}(h)+\lambda_*(\mathcal{V}(h)-|B|). L ( h ) = − J ( h ) + λ ∗ ( V ( h ) − ∣ B ∣ ) . Along the path, L ( h ( t ) ) = − T ( Ω h ( t ) ) \mathcal{L}(h(t))=-T(\Omega_{h(t)}) L ( h ( t )) = − T ( Ω h ( t ) ) , since the volume is fixed. We first show that L \mathcal{L} L is of class C 2 C^2 C 2 near 0 0 0 in C 2 ( S n − 1 ) C^2(\mathbb{S}^{n-1}) C 2 ( S n − 1 ) . By (54) , V \mathcal{V} V is a polynomial in h h h . For J \mathcal{J} J , choose a smooth cutoff χ \chi χ , supported in ( − r 0 , r 0 ) (-r_0,r_0) ( − r 0 , r 0 ) and equal to one near zero, and set X h ( ρ θ ) = χ ( 1 − ρ ) h ( θ ) θ X_h(\rho\theta)=\chi(1-\rho)h(\theta)\theta X h ( ρθ ) = χ ( 1 − ρ ) h ( θ ) θ . The map h ↦ X h h\mapsto X_h h ↦ X h is bounded and linear from C 2 C^2 C 2 to W 1 , ∞ W^{1,\infty} W 1 , ∞ , and Id + X h \operatorname{Id}+X_h Id + X h is a bi-Lipschitz map of B B B onto Ω h \Omega_h Ω h for small h h h . The proof of Proposition 3.2 applies with the Banach parameter h h h in place of t t t and Id + X h \operatorname{Id}+X_h Id + X h in place of Φ t \Phi_t Φ t : it only uses that the deformation depends in a C 2 C^2 C 2 way on the parameter with values in W 1 , ∞ W^{1,\infty} W 1 , ∞ , with uniform bi-Lipschitz bounds. Hence J \mathcal{J} J is of class C 2 C^2 C 2 .
We now compute the derivatives of L \mathcal{L} L at 0 0 0 in the direction φ \varphi φ . Set
L ( t ) = L ( t φ ) = − T φ ( t ) + λ ∗ ( V φ ( t ) − ∣ B ∣ ) . L(t)=\mathcal{L}(t\varphi)=-T_\varphi(t)+\lambda_*(\mathcal{V}_\varphi(t)-|B|). L ( t ) = L ( tφ ) = − T φ ( t ) + λ ∗ ( V φ ( t ) − ∣ B ∣ ) . By (53) and (55) ,
L ′ ( 0 ) = 0 , L ′ ′ ( 0 ) = − T φ ′ ′ ( 0 ) + λ ∗ V φ ′ ′ ( 0 ) = λ ∗ ∫ S n − 1 ( φ Λ φ − φ 2 ) d H n − 1 . L'(0)=0,\qquad L''(0)=-T_\varphi''(0)+\lambda_*\mathcal{V}_\varphi''(0)=\lambda_*\int_{\mathbb{S}^{n-1}}(\varphi\Lambda\varphi-\varphi^2)\,d\mathcal{H}^{n-1}. L ′ ( 0 ) = 0 , L ′′ ( 0 ) = − T φ ′′ ( 0 ) + λ ∗ V φ ′′ ( 0 ) = λ ∗ ∫ S n − 1 ( φ Λ φ − φ 2 ) d H n − 1 . Since L \mathcal{L} L is C 2 C^2 C 2 , D L ( 0 ) [ φ ] = L ′ ( 0 ) D\mathcal{L}(0)[\varphi]=L'(0) D L ( 0 ) [ φ ] = L ′ ( 0 ) and D 2 L ( 0 ) [ φ , φ ] = L ′ ′ ( 0 ) D^2\mathcal{L}(0)[\varphi,\varphi]=L''(0) D 2 L ( 0 ) [ φ , φ ] = L ′′ ( 0 ) . Finally, t ↦ h ( t ) t\mapsto h(t) t ↦ h ( t ) is C 2 C^2 C 2 with values in C 2 ( S n − 1 ) C^2(\mathbb{S}^{n-1}) C 2 ( S n − 1 ) , so the chain rule gives
d 2 d t 2 L ( h ( t ) ) ∣ t = 0 = D 2 L ( 0 ) [ φ , φ ] + D L ( 0 ) [ h ¨ ( 0 ) ] = λ ∗ ∫ S n − 1 ( φ Λ φ − φ 2 ) d H n − 1 , \left.\frac{d^2}{dt^2}\mathcal{L}(h(t))\right|_{t=0}=D^2\mathcal{L}(0)[\varphi,\varphi]+D\mathcal{L}(0)[\ddot{h}(0)]=\lambda_*\int_{\mathbb{S}^{n-1}}(\varphi\Lambda\varphi-\varphi^2)\,d\mathcal{H}^{n-1}, d t 2 d 2 L ( h ( t )) t = 0 = D 2 L ( 0 ) [ φ , φ ] + D L ( 0 ) [ h ¨ ( 0 )] = λ ∗ ∫ S n − 1 ( φ Λ φ − φ 2 ) d H n − 1 , because h ¨ ( 0 ) \ddot{h}(0) h ¨ ( 0 ) is smooth. Since L ( h ( t ) ) = − T ( Ω h ( t ) ) \mathcal{L}(h(t))=-T(\Omega_{h(t)}) L ( h ( t )) = − T ( Ω h ( t ) ) , this is (56) . □
Proof of Theorem 1.2 . Let h ( t ) h(t) h ( t ) be a smooth volume-preserving path, in the sense of Lemma 4.2 . By that lemma, φ = h ˙ ( 0 ) \varphi=\dot{h}(0) φ = h ˙ ( 0 ) satisfies (10) , and (12) holds with Q Q Q given by the first expression in (11) . Since P φ ℓ ( ρ θ ) = ρ ℓ φ ℓ ( θ ) P_{\varphi_\ell}(\rho\theta)=\rho^\ell\varphi_\ell(\theta) P φ ℓ ( ρθ ) = ρ ℓ φ ℓ ( θ ) , we have Λ φ ℓ = ℓ φ ℓ \Lambda\varphi_\ell=\ell\varphi_\ell Λ φ ℓ = ℓ φ ℓ , and orthogonality gives the second expression in (11) . Conditions (10) and (13) are precisely φ 0 = φ 1 = 0 \varphi_0=\varphi_1=0 φ 0 = φ 1 = 0 , and
1 3 ( 1 + ℓ ) ≤ ℓ − 1 ≤ 1 + ℓ ( ℓ ≥ 2 ) . \frac{1}{3}(1+\ell)\leq\ell-1\leq1+\ell\qquad(\ell\geq2). 3 1 ( 1 + ℓ ) ≤ ℓ − 1 ≤ 1 + ℓ ( ℓ ≥ 2 ) .
This proves (14) , with equality in the lower bound exactly on harmonics of degree two. The upper constant is also optimal, since ( ℓ − 1 ) / ( ℓ + 1 ) → 1 (\ell-1)/(\ell+1)\to1 ( ℓ − 1 ) / ( ℓ + 1 ) → 1 . □
Acknowledgments
The authors acknowledge financial support from the ANR grant STOIQUES. MN is member of INdAM/GNAMPA. BR is a member of INdAM–GNAMPA and has been partially supported by the “INdAM - GNAMPA Project”, code CUP E53C25002010001.
Use of AI. The proofs in this paper were found mostly with the help of a large language model (LLM), which was also used to help write up the paper. The work involved substantial interaction between the authors and the model. The authors have checked the results and the text, and take full responsibility for the content of the paper.
References
LAMA, Univ Paris Est Creteil, Univ Gustave Eiffel, UPEM, CNRS, F-94010 Créteil, France
Email address: michael.goldman@cnrs.fr
Dipartimento di Matematica, Università di Pisa, Pisa, Italy
Email address: matteo.novaga@unipi.it
Dipartimento di Matematica, Università di Bologna, Bologna, Italy
Email address: berardo.ruffini@unibo.it