Introduction
The scalar-flat boundary Yamabe problem asks whether a compact Riemannian manifold with boundary admits a conformal metric with zero scalar curvature and constant boundary mean curvature. Let ( M n , g ) (M^n,g) ( M n , g ) be a smooth compact Riemannian manifold with nonempty boundary, n ≥ 3 n\geq3 n ≥ 3 , and let q = 2 ( n − 1 ) / ( n − 2 ) q=2(n-1)/(n-2) q = 2 ( n − 1 ) / ( n − 2 ) be the critical trace exponent. Denote the scalar curvature by Scal g \operatorname{Scal}_g Scal g , the averaged boundary mean curvature by H g H_g H g , and the volume and boundary measures by d V g dV_g d V g and d σ g d\sigma_g d σ g . We use the convention for which the Euclidean unit ball has H g = 1 H_g=1 H g = 1 . For u ∈ H 1 ( M ) u\in H^1(M) u ∈ H 1 ( M ) , the conformal energy is
N g ∘ ( u ) = ∫ M ( ∣ ∇ u ∣ g 2 + n − 2 4 ( n − 1 ) Scal g u 2 ) d V g + n − 2 2 ∫ ∂ M H g u 2 d σ g . (1) \mathcal{N}^{\circ}_g(u)=\int_M\left(|\nabla u|_g^2+\frac{n-2}{4(n-1)}\operatorname{Scal}_g u^2\right)dV_g+\frac{n-2}{2}\int_{\partial M}H_g u^2\,d\sigma_g. \tag*{(1)} N g ∘ ( u ) = ∫ M ( ∣∇ u ∣ g 2 + 4 ( n − 1 ) n − 2 Scal g u 2 ) d V g + 2 n − 2 ∫ ∂ M H g u 2 d σ g . (1)
Let Tr u \operatorname{Tr}u Tr u denote the boundary trace. The Escobar quotient and its infimum are
J g [ u ] = N g ∘ ( u ) ∥ u ∥ L q ( ∂ M ) 2 , C E s c ∗ ( M , g ) = inf u ∈ H 1 ( M ) Tr u ≠ 0 J g [ u ] . (2) \mathcal{J}_g[u]=\frac{\mathcal{N}^{\circ}_g(u)}{\|u\|_{L^q(\partial M)}^2},\qquad C^*_{\mathrm{Esc}}(M,g)=\inf_{\substack{u\in H^1(M)\\ \operatorname{Tr}u\ne0}}\mathcal{J}_g[u]. \tag*{(2)} J g [ u ] = ∥ u ∥ L q ( ∂ M ) 2 N g ∘ ( u ) , C Esc ∗ ( M , g ) = u ∈ H 1 ( M ) Tr u = 0 inf J g [ u ] . (2)
Write S ∗ S_* S ∗ for this infimum on the round hemisphere, equivalently the sharp half-space trace constant. Rescaled half-space optimizers, called boundary bubbles, have quotient tending to S ∗ S_* S ∗ . Their concentration accounts for the loss of compactness at this value. We study how the norm of the constrained gradient measures proximity to one concentrating bubble, and which geometric terms prevent such a bubble from being stationary.
Throughout the analytic arguments we assume strict coercivity, denoted ( Y ∂ + ) (Y_{\partial}^{+}) ( Y ∂ + ) :
N g ∘ ( u ) ≥ c Y ∥ u ∥ H 1 ( M ) 2 ( u ∈ H 1 ( M ) ) , c Y > 0. \mathcal{N}^{\circ}_g(u)\geq c_Y\|u\|_{H^1(M)}^2\qquad(u\in H^1(M)),\qquad c_Y>0. N g ∘ ( u ) ≥ c Y ∥ u ∥ H 1 ( M ) 2 ( u ∈ H 1 ( M )) , c Y > 0.
Polarization defines the energy inner product. The regularity and geometric bounds are stated with each result. The argument has two parts. A profile decomposition without a sign assumption determines the possible loss of compactness below the first mixed level. At the remaining single-bubble level S ∗ S_* S ∗ , we solve the equation in the complementary directions and study the resulting energy as a function of the center and scale. A Palais–Smale sequence may concentrate at any prescribed boundary point; a concentrating sequence of exact critical points must also make this reduced energy stationary. The quadratic and fourth-order expansions give successive geometric obstructions to this stationarity, under the hypotheses stated below.
Escobar developed the variational approach [19 ] and its scalar-flat formulation [18 , 20 , 21 ] ; the half-space optimizers are given in [17 Eq. (2)] . Write I I ˚ g \mathring{\mathrm{II}}_g II ˚ g for the trace-free second fundamental form and W g W_g W g for the Weyl tensor. A point is umbilic if I I ˚ g = 0 \mathring{\mathrm{II}}_g = 0 II ˚ g = 0 there, and the boundary is totally umbilic if this holds everywhere. The strict comparison at a non-umbilic point was proved by Escobar for n ≥ 6 n \ge6 n ≥ 6 and Marques for n = 4 , 5 n = 4, 5 n = 4 , 5 ; see [37 pp. 383–384] and [9 pp. 1–2] . Almaraz used fourth-order test functions for umbilic boundaries with W g ≠ 0 W_g \ne0 W g = 0 [9 Section 3] . Under a curvature-jet vanishing hypothesis, Mayer and Ndiaye proved existence for positive conformal invariant and totally umbilic boundary when n ≥ 6 n \ge6 n ≥ 6 [39 Theorem 1.1] .
Deficit estimates controlling the squared distance from extremals were proved by Bianchi and Egnell for the Euclidean Sobolev inequality [3 ] and by Ho for the half-space trace inequality [29 ] ; see also [2 Section 2] . Fractional Sobolev analogues were proved by Chen, Frank, and Weth [6 ] . Zhang, Zhou, and Zou proved stability and profile decomposition for the flat trace problem [ZZZ25, Theorems 1.2, 1.6, and 5.1]. On compact manifolds with nonnegative scalar curvature and conformal invariant below S ∗ S_* S ∗ , Borquez, Caju, and Van Den Bosch obtained estimates near the positive minimizing set [2 Theorem 1.1] . Quantitative stability and threshold compactness for the homogeneous Dirichlet Gagliardo–Nirenberg quotient on manifolds with boundary are studied in [40 ] .
For the Euclidean Sobolev equation, sharp quantitative stability estimates were proved by Figalli and Glaudo [22 ] and Deng, Sun, and Wei [DSW25]. Chen and Kim proved sharp estimates on closed manifolds with rates depending on the geometry and the concentrating functions [8 ] . Quantitative stability estimates for the flat trace problem appear in [ZZZ25]; Yu and Zhou treat one bubble for the p p p -Laplacian [46 ] . We determine how the norm of the constrained gradient controls the distance to the standard and corrected bubble families on a compact manifold with boundary.
Positive-solution compactness was proved under various hypotheses by Almaraz [10 ] , Almaraz, de Queiroz, and Wang [13 ] , Kim, Musso, and Wei [30 ] , and Ghimenti and Micheletti [25 , 26 ] . Gong, Kim, Musso, and Wei establish compactness through dimension 14 for general boundary and 21 for umbilic boundary, excluding the conformal hemisphere [24 Theorem 1] . Our stability estimates concern approximate critical points and distinguish the standard bubble family from the corrected family.
Let S = { u ∈ H 1 ( M ) : ∥ u ∥ L q ( ∂ M ) = 1 } \mathcal{S} = \{u \in H^1(M) : \|u\|_{L^q(\partial M)} = 1\} S = { u ∈ H 1 ( M ) : ∥ u ∥ L q ( ∂ M ) = 1 } . Polarization of N g ∘ \mathcal{N}_g^\circ N g ∘ gives the energy norm ∥ ⋅ ∥ E , g \|\cdot\|_{E,g} ∥ ⋅ ∥ E , g , and we put
r g ( u ) = ∥ grad S J g ( u ) ∥ E , g . r_g(u) = \|\operatorname{grad}_{\mathcal{S}}\mathcal{J}_g(u)\|_{E,g}. r g ( u ) = ∥ grad S J g ( u ) ∥ E , g .
A constrained Palais–Smale sequence at c c c satisfies J g [ u j ] → c \mathcal{J}_g[u_j] \to c J g [ u j ] → c and r g ( u j ) → 0 r_g(u_j) \to0 r g ( u j ) → 0 . Throughout, the smooth compact metric satisfies ( Y ∂ + ) (Y^+_\partial) ( Y ∂ + ) , and constants may depend on ( M , g ) (M,g) ( M , g ) . The finite-regularity results are stated in Theorem 7.6 and proved in Appendix C.
For the bubble expansions, choose a conformal representative with H g = 0 H_g = 0 H g = 0 . Let U x , ε ∈ S \mathcal{U}_{x,\varepsilon} \in\mathcal{S} U x , ε ∈ S be the normalized boundary bubble centered at x x x and of scale ε \varepsilon ε , with the cutoff specified in § 2 . Write Z 0 = ε ∂ ε U x , ε Z_0 = \varepsilon\partial_\varepsilon\mathcal{U}_{x,\varepsilon} Z 0 = ε ∂ ε U x , ε and Z a = ε ∂ x a U x , ε Z_a = \varepsilon\partial_{x^a}\mathcal{U}_{x,\varepsilon} Z a = ε ∂ x a U x , ε in local boundary coordinates, and set
X x , ε = { h ∈ H 1 ( M ) : ∫ ∂ M U x , ε q − 1 h d σ g = 0 , ⟨ h , Z j ⟩ E , g = 0 ( 0 ≤ j ≤ n − 1 ) } . X_{x,\varepsilon} = \left\{ h \in H^1(M) : \int_{\partial M} \mathcal{U}_{x,\varepsilon}^{q-1} h\,d\sigma_g = 0,\ \langle h,Z_j\rangle_{E,g} = 0 \quad(0 \le j \le n-1) \right\}. X x , ε = { h ∈ H 1 ( M ) : ∫ ∂ M U x , ε q − 1 h d σ g = 0 , ⟨ h , Z j ⟩ E , g = 0 ( 0 ≤ j ≤ n − 1 ) } .
The correction solves the equation transverse to changes of center and scale. For h ∈ X x , ε h \in X_{x,\varepsilon} h ∈ X x , ε put ret U ( h ) = ( U + h ) / ∥ U + h ∥ q \operatorname{ret}_{U}(h)=(U+h)/\lVert U+h\rVert_q ret U ( h ) = ( U + h ) / ∥ U + h ∥ q , the normalization map (a local retraction onto S \mathcal{S} S ). When n ≥ 5 n\geq5 n ≥ 5 , there is a unique small correction w x , ε ∈ X x , ε w_{x,\varepsilon}\in X_{x,\varepsilon} w x , ε ∈ X x , ε for which the differential of J g ∘ ret U x , ε \mathcal{J}_g\circ\operatorname{ret}_{U_{x,\varepsilon}} J g ∘ ret U x , ε vanishes on X x , ε X_{x,\varepsilon} X x , ε .
Define
Φ ( x , ε ) = ret U x , ε ( w x , ε ) , J 1 ( x , ε ) = J g [ Φ ( x , ε ) ] . \Phi(x,\varepsilon)=\operatorname{ret}_{U_{x,\varepsilon}}(w_{x,\varepsilon}),\qquad\mathcal{J}_1(x,\varepsilon)=\mathcal{J}_g[\Phi(x,\varepsilon)]. Φ ( x , ε ) = ret U x , ε ( w x , ε ) , J 1 ( x , ε ) = J g [ Φ ( x , ε )] .
Let B 0 \mathcal{B}_0 B 0 and B L S \mathcal{B}_{\mathrm{LS}} B LS be the standard and Lyapunov–Schmidt (LS) corrected families at sufficiently small scales, including both signs. All distances below are measured in the energy norm.
Theorem A (Quantitative stability and sharpness). Assume n ≥ 5 n\geq5 n ≥ 5 and H g = 0 H_g=0 H g = 0 . Let K ⊂ ∂ M K\subset\partial M K ⊂ ∂ M be compact with inf K ∣ I I ˚ g ∣ > 0 \inf_K\lvert\mathring{\mathrm{II}}_g\rvert>0 inf K ∣ II ˚ g ∣ > 0 . There are ρ , ε 0 , C > 0 \rho,\varepsilon_0,C>0 ρ , ε 0 , C > 0 such that every u ∈ S u\in\mathcal{S} u ∈ S whose normalized modulation gives
u = σ ret U x , ε ( v ) , v ∈ X x , ε , ∥ v ∥ E , g < ρ , x ∈ K , 0 < ε < ε 0 , σ ∈ { ± 1 } , u=\sigma\operatorname{ret}_{U_{x,\varepsilon}}(v),\qquad v\in X_{x,\varepsilon},\qquad\lVert v\rVert_{E,g}<\rho,\qquad x\in K,\qquad0<\varepsilon<\varepsilon_0,\qquad\sigma\in\{\pm1\}, u = σ ret U x , ε ( v ) , v ∈ X x , ε , ∥ v ∥ E , g < ρ , x ∈ K , 0 < ε < ε 0 , σ ∈ { ± 1 } , satisfies
∥ u − σ Φ ( x , ε ) ∥ E , g ≤ C r g ( u ) , ε 2 ≤ C r g ( u ) , \lVert u-\sigma\Phi(x,\varepsilon)\rVert_{E,g}\leq C r_g(u),\qquad\varepsilon^2\leq C r_g(u), ∥ u − σ Φ ( x , ε ) ∥ E , g ≤ C r g ( u ) , ε 2 ≤ C r g ( u ) , dist E ( u , B 0 ) ≤ C r g ( u ) 1 / 2 , dist E ( u , B L S ) ≤ C r g ( u ) . (3) \operatorname{dist}_E(u,\mathcal{B}_0)\leq C r_g(u)^{1/2},\qquad\operatorname{dist}_E(u,\mathcal{B}_{\mathrm{LS}})\leq C r_g(u). \tag*{(3)} dist E ( u , B 0 ) ≤ C r g ( u ) 1/2 , dist E ( u , B LS ) ≤ C r g ( u ) . (3) If J g [ u ] ≥ S ∗ \mathcal{J}_g[u]\geq S_* J g [ u ] ≥ S ∗ , then
dist E ( u , B 0 ) ≤ C r g ( u ) . \operatorname{dist}_E(u,\mathcal{B}_0)\leq C r_g(u). dist E ( u , B 0 ) ≤ C r g ( u ) . Whenever I I ˚ g \mathring{\mathrm{II}}_g II ˚ g is not identically zero on a connected boundary component, there are concentrating sequences showing that the powers 1 / 2 1/2 1/2 in the standard-family estimate and 1 1 1 in its improvement above S ∗ S_* S ∗ cannot be increased. Furthermore, these statements hold independently of any sign assumptions on the underlying functions.
Theorem A is proved in Theorem 6.6 and Proposition 6.7. The complementary correction has order ε \varepsilon ε , whereas the logarithmic scale derivative of the reduced functional has order ε 2 \varepsilon^2 ε 2 away from the umbilic locus. The negative energy term in the next theorem accounts both for the square-root estimate and for its linear improvement above S ∗ S_* S ∗ .
Theorem B (The exact quadratic coefficient). Assume n ≥ 5 n\geq5 n ≥ 5 and H g = 0 H_g=0 H g = 0 . Uniformly in x ∈ ∂ M x\in\partial M x ∈ ∂ M ,
J 1 ( x , ε ) / S ∗ = 1 + κ 3 r e d ( n ) ∣ I I ˚ g ( x ) ∣ 2 ε 2 + o ( ε 2 ) , \mathcal{J}_1(x,\varepsilon)/S_*=1+\kappa_3^{\mathrm{red}}(n)\lvert\mathring{\mathrm{II}}_g(x)\rvert^2\varepsilon^2+o(\varepsilon^2), J 1 ( x , ε ) / S ∗ = 1 + κ 3 red ( n ) ∣ II ˚ g ( x ) ∣ 2 ε 2 + o ( ε 2 ) , S ∗ − 1 ∂ ε J 1 ( x , ε ) = 2 κ 3 r e d ( n ) ∣ I I ˚ g ( x ) ∣ 2 ε + o ( ε ) . S_*^{-1}\partial_\varepsilon\mathcal{J}_1(x,\varepsilon)=2\kappa_3^{\mathrm{red}}(n)\lvert\mathring{\mathrm{II}}_g(x)\rvert^2\varepsilon+o(\varepsilon). S ∗ − 1 ∂ ε J 1 ( x , ε ) = 2 κ 3 red ( n ) ∣ II ˚ g ( x ) ∣ 2 ε + o ( ε ) . The reduced coefficient is strictly negative. Its exact value and the complementary correction are
δ κ 3 L S ( n ) = ( 2 n − 1 ) ( n + 2 ) 2 ( n − 1 ) 2 ( n + 1 ) ( n − 3 ) ( n − 4 ) , \delta\kappa_3^{\mathrm{LS}}(n)=\frac{(2n-1)(n+2)}{2(n-1)^2(n+1)(n-3)(n-4)}, δ κ 3 LS ( n ) = 2 ( n − 1 ) 2 ( n + 1 ) ( n − 3 ) ( n − 4 ) ( 2 n − 1 ) ( n + 2 ) , κ 3 r e d ( n ) = − ( n − 2 ) ( n 2 − 2 n − 2 ) 2 ( n − 1 ) 2 ( n + 1 ) ( n − 3 ) ( n − 4 ) < 0. (4) \kappa_3^{\mathrm{red}}(n)=-\frac{(n-2)(n^2-2n-2)}{2(n-1)^2(n+1)(n-3)(n-4)}<0. \tag*{(4)} κ 3 red ( n ) = − 2 ( n − 1 ) 2 ( n + 1 ) ( n − 3 ) ( n − 4 ) ( n − 2 ) ( n 2 − 2 n − 2 ) < 0. (4) The standard-bubble coefficient is κ 3 b a r e ( n ) = ( 6 − n ) / ( 2 ( n − 1 ) ( n − 3 ) ( n − 4 ) ) \kappa_3^{\mathrm{bare}}(n)=(6-n)/(2(n-1)(n-3)(n-4)) κ 3 bare ( n ) = ( 6 − n ) / ( 2 ( n − 1 ) ( n − 3 ) ( n − 4 )) , and κ 3 r e d = κ 3 b a r e − δ κ 3 L S \kappa_3^{\mathrm{red}}=\kappa_3^{\mathrm{bare}}-\delta\kappa_3^{\mathrm{LS}} κ 3 red = κ 3 bare − δ κ 3 LS . In particular, δ κ 3 L S ( 5 ) = 21 / 128 \delta\kappa_3^{\mathrm{LS}}(5)=21/128 δ κ 3 LS ( 5 ) = 21/128 and κ 3 r e d ( 5 ) = − 13 / 128 \kappa_3^{\mathrm{red}}(5)=-13/128 κ 3 red ( 5 ) = − 13/128 .
The quadratic form involving the inverse linearized operator is evaluated by an explicit particular solution and a harmonic Robin correction, using spherical harmonics as in [24 Proposition 3.3] . Its evaluation and differentiated expansion are proved in Proposition 5.15 and Proposition 6.4.
The compactness statements use a profile decomposition without a sign assumption, with a remainder converging strongly to zero in H 1 H^1 H 1 , based on [11 Lemma 2.4] . Put
Y = C E s c ∗ ( M , g ) > 0 , c min = ( S ∗ n − 1 + Y n − 1 ) 1 / ( n − 1 ) , Λ 2 = 2 1 / ( n − 1 ) S ∗ . Y=C_{\mathrm{Esc}}^*(M,g)>0,\qquad c_{\min}=(S_*^{n-1}+Y^{n-1})^{1/(n-1)},\qquad\Lambda_2=2^{1/(n-1)}S_*. Y = C Esc ∗ ( M , g ) > 0 , c m i n = ( S ∗ n − 1 + Y n − 1 ) 1/ ( n − 1 ) , Λ 2 = 2 1/ ( n − 1 ) S ∗ .
Theorem C (Palais–Smale levels and compactness of critical points).
(i) Assume n ≥ 3 n \ge3 n ≥ 3 . Every constrained Palais–Smale sequence with level c < S ∗ c < S_* c < S ∗ or S ∗ < c < c m i x S_* < c < c_{\mathrm{mix}} S ∗ < c < c mix is precompact in H 1 ( M ) H^1(M) H 1 ( M ) . The endpoint c m i x c_{\mathrm{mix}} c mix admits a noncompact Palais–Smale sequence. For S ∗ < c < Λ 2 S_* < c < \Lambda_2 S ∗ < c < Λ 2 , a noncompact constrained Palais–Smale sequence at c c c exists if and only if
d = ( c n − 1 − S ∗ n − 1 ) 1 / ( n − 1 ) d = \left(c^{n-1} - S_*^{n-1}\right)^{1/(n-1)} d = ( c n − 1 − S ∗ n − 1 ) 1/ ( n − 1 )
is a critical value of J g ∣ S \mathcal{J}_g|_{\mathcal{S}} J g ∣ S .
(ii) Assume n ≥ 5 n \ge5 n ≥ 5 . In a representative with H g = 0 H_g = 0 H g = 0 , let ( u j ) ⊂ S (u_j) \subset\mathcal{S} ( u j ) ⊂ S be constrained critical points with J g [ u j ] → S ∗ \mathcal{J}_g[u_j] \to S_* J g [ u j ] → S ∗ that are not precompact. After extraction there are σ ∈ { ± 1 } \sigma\in\{\pm1\} σ ∈ { ± 1 } , ε j ↓ 0 \varepsilon_j \downarrow0 ε j ↓ 0 , and x j → p ∈ ∂ M x_j \to p \in\partial M x j → p ∈ ∂ M such that
u j = σ Φ ( x j , ε j ) , I I ˚ g ( p ) = 0. u_j = \sigma\Phi(x_j,\varepsilon_j), \qquad\mathring{\mathrm{II}}_g(p) = 0. u j = σ Φ ( x j , ε j ) , II ˚ g ( p ) = 0.
The umbilicity conclusion is conformally invariant. In any conformal representative, if the boundary is nowhere umbilic, then for every c 0 < c m i x c_0 < c_{\mathrm{mix}} c 0 < c mix the set
{ u ∈ S : grad S J g ( u ) = 0 , J g [ u ] ≤ c 0 } \{u \in\mathcal{S} : \operatorname{grad}_{\mathcal{S}}\mathcal{J}_g(u) = 0,\ \mathcal{J}_g[u] \le c_0\} { u ∈ S : grad S J g ( u ) = 0 , J g [ u ] ≤ c 0 }
is compact in H 1 ( M ) H^1(M) H 1 ( M ) .
The Palais–Smale assertions are Theorems 4.3 and 4.7 . A stronger mass bound for sign-changing profiles was proved by Almaraz and Wang [14 Theorem 1.4] . The critical-point conclusions follow from Theorem 7.2 and Corollary 7.5. For n ≥ 5 n \ge5 n ≥ 5 , the nowhere-umbilic condition is open and dense among coercive C m C^m C m metrics, m ≥ 3 m \ge3 m ≥ 3 ; see Theorem 7.6 .
The preceding theorem leaves umbilic points as possible blow-up points of sequences of critical points. At an umbilic center, the order-ε \varepsilon ε coefficient of the complementary first variation vanishes. The next coefficient can occur at order ε 2 \varepsilon^2 ε 2 , and the correction obtained by completing the square contributes at order ε 4 \varepsilon^4 ε 4 . We call these leading covectors the geometric sources; their precise definitions are given in Remark 5.3. The resulting quartic coefficient has an explicit curvature formula under local total umbilicity.
For the fourth-order expansion assume n ≥ 7 n \ge7 n ≥ 7 and take R β ( ε ) = ε − β R_\beta(\varepsilon) = \varepsilon^{-\beta} R β ( ε ) = ε − β with 4 / ( n − 2 ) < β < 1 4/(n-2) < \beta< 1 4/ ( n − 2 ) < β < 1 , and let J 1 , β \mathcal{J}_{1,\beta} J 1 , β be the corresponding corrected one-bubble energy. Write E g = W g ( ν , ⋅ , ν , ⋅ ) E_g = W_g(\nu,\cdot,\nu,\cdot) E g = W g ( ν , ⋅ , ν , ⋅ ) on T ∂ M T\partial M T ∂ M , where ν \nu ν is a unit normal, and let W ‾ g \overline{W}_g W g be the intrinsic Weyl tensor of the boundary metric.
Theorem D (The fourth-order coefficient). Assume n ≥ 7 n \ge7 n ≥ 7 and H g = 0 H_g = 0 H g = 0 .
(i) At every umbilic point p p p there is a coefficient Ξ g r e d ( p ) \Xi^{\mathrm{red}}_g(p) Ξ g red ( p ) , independent of β \beta β , such that
J 1 , β ( p , ε ) / S ∗ = 1 + Ξ g r e d ( p ) ε 4 + o ( ε 4 ) , \mathcal{J}_{1,\beta}(p,\varepsilon)/S_* = 1 + \Xi^{\mathrm{red}}_g(p)\varepsilon^4 + o(\varepsilon^4), J 1 , β ( p , ε ) / S ∗ = 1 + Ξ g red ( p ) ε 4 + o ( ε 4 ) , S ∗ − 1 ∂ ε J 1 , β ( p , ε ) = 4 Ξ g r e d ( p ) ε 3 + o ( ε 3 ) . S_*^{-1}\partial_\varepsilon\mathcal{J}_{1,\beta}(p,\varepsilon) = 4\Xi^{\mathrm{red}}_g(p)\varepsilon^3 + o(\varepsilon^3). S ∗ − 1 ∂ ε J 1 , β ( p , ε ) = 4 Ξ g red ( p ) ε 3 + o ( ε 3 ) . The remainders are uniform on compact subsets of the umbilic locus.
(ii) If the boundary is totally umbilic near p p p , then
Ξ g r e d ( p ) = − e n ( n − 2 ) 2 ( n − 3 ) 2 ∣ E g ( p ) ∣ g 2 − b n ∣ W ‾ g ( p ) ∣ g 2 , (5) \Xi^{\mathrm{red}}_g(p) = -e_n\frac{(n-2)^2}{(n-3)^2}|E_g(p)|_g^2 - b_n|\overline{W}_g(p)|_g^2, \tag*{(5)} Ξ g red ( p ) = − e n ( n − 3 ) 2 ( n − 2 ) 2 ∣ E g ( p ) ∣ g 2 − b n ∣ W g ( p ) ∣ g 2 , (5) where
e n = 3 n 3 − 18 n 2 + 21 n + 14 3 ( n − 2 ) ( n − 1 ) 2 ( n + 1 ) ( n − 4 ) ( n − 5 ) ( n − 6 ) , b n = n − 2 24 ( n − 1 ) ( n − 3 ) ( n − 5 ) ( n − 6 ) . e_n = \frac{3n^3 - 18n^2 + 21n + 14}{3(n-2)(n-1)^2(n+1)(n-4)(n-5)(n-6)}, \qquad b_n = \frac{n-2}{24(n-1)(n-3)(n-5)(n-6)}. e n = 3 ( n − 2 ) ( n − 1 ) 2 ( n + 1 ) ( n − 4 ) ( n − 5 ) ( n − 6 ) 3 n 3 − 18 n 2 + 21 n + 14 , b n = 24 ( n − 1 ) ( n − 3 ) ( n − 5 ) ( n − 6 ) n − 2 . Both constants are positive. The optimal algebraic comparison is
Ξ g r e d ( p ) ≤ − c ( n ) ∣ W ‾ g ( p ) ∣ g 2 , c ( n ) = n − 2 4 ( n − 3 ) e n , \Xi^{\mathrm{red}}_g(p) \le-c(n)|\overline{W}_g(p)|_g^2, \qquad c(n) = \frac{n-2}{4(n-3)}e_n, Ξ g red ( p ) ≤ − c ( n ) ∣ W g ( p ) ∣ g 2 , c ( n ) = 4 ( n − 3 ) n − 2 e n , with equality precisely when W ‾ g ( p ) = 0 \overline{W}_g(p) = 0 W g ( p ) = 0 . In particular, Ξ g r e d ( p ) = 0 \Xi^{\mathrm{red}}_g(p) = 0 Ξ g red ( p ) = 0 if and only if W ‾ g ( p ) = 0 \overline{W}_g(p) = 0 W g ( p ) = 0 . For another representative g ~ = ψ 4 / ( n − 2 ) g \widetilde{g} = \psi^{4/(n-2)}g g = ψ 4/ ( n − 2 ) g with H g ~ = 0 H_{\widetilde{g}} = 0 H g = 0 ,
Ξ g ~ r e d ( p ) = ψ ( p ) − 8 / ( n − 2 ) Ξ g r e d ( p ) . \Xi^{\mathrm{red}}_{\widetilde{g}}(p) = \psi(p)^{-8/(n-2)}\Xi^{\mathrm{red}}_g(p). Ξ g red ( p ) = ψ ( p ) − 8/ ( n − 2 ) Ξ g red ( p ) .
(iii) Suppose that the whole boundary is totally umbilic. If a normalized sequence of constrained critical points at values tending to S ∗ S_* S ∗ is not precompact, then every limiting blow-up center p p p satisfies W g ( p ) = 0 W_g(p)=0 W g ( p ) = 0 . If inf ∂ M ∣ W g ∣ > 0 \inf_{\partial M}|W_g|>0 inf ∂ M ∣ W g ∣ > 0 , then for every c 0 < c m i x c_0<c_{\mathrm{mix}} c 0 < c mix the normalized constrained critical points with values at most c 0 c_0 c 0 form a compact subset of H 1 ( M ) H^1(M) H 1 ( M ) .
The expansion is Proposition 8.5 ; the curvature formula, conformal weight, and compactness consequence are Theorem 8.7 and Corollaries 8.8 and 8.9.
The two coefficient calculations use the same flat complementary Hessian. If S \mathscr{S} S and L \mathscr{L} L are the complementary differential and Hessian of the functional, respectively, solving the complementary equation changes the energy by
− 1 2 ⟨ L − 1 S , S ⟩ -\frac{1}{2}\langle\mathscr{L}^{-1}\mathscr{S},\mathscr{S}\rangle − 2 1 ⟨ L − 1 S , S ⟩
to leading order. The source is O ( ε ) O(\varepsilon) O ( ε ) in the quadratic calculation and O ( ε 2 ) O(\varepsilon^2) O ( ε 2 ) at umbilic centers with the cutoff used in Theorem D . We control the remainder and its weighted scale derivative. The explicit curvature calculation in Theorem D uses the same harmonic Robin pairing as Theorem B.
We use ∇ ∂ \nabla_{\partial} ∇ ∂ for the gradient of the induced boundary metric. The dimensional restrictions reflect the corresponding integrability: the profile decomposition holds for n ≥ 3 n\geq3 n ≥ 3 , the first geometric source has finite norm in the dual of the energy space for n ≥ 5 n\geq5 n ≥ 5 , and the fourth-order analysis requires n ≥ 7 n\geq7 n ≥ 7 . The standard-bubble calculations used below are stated in Proposition 2.5, with their sources in the companion paper [41 ] .
After proving coercivity and modulation, we establish the profile decomposition without a sign assumption. We then construct the corrected family and compute its quadratic coefficient. The differentiated expansion gives the stability estimates and the compactness results for critical points. The fourth-order calculation continues this argument at umbilic centers and concludes the main text. Supplementary deficit estimates, hemisphere rigidity, the finite-regularity genericity proof, and further consequences are given in the appendices.
Analytic and geometric preliminaries
Throughout, M M M is a compact smooth manifold of dimension n ≥ 3 n\geq3 n ≥ 3 , with nonempty smooth boundary, and g g g is a smooth Riemannian metric. Let g ˉ = g ∣ T ∂ M \bar{g}=g|_{T\partial M} g ˉ = g ∣ T ∂ M be the induced boundary metric and ν \nu ν the inward unit normal. Our conventions are
II ( X , Y ) = − ⟨ ∇ X ν , Y ⟩ g , K g = tr g ˉ II , H g = K g n − 1 , II ˚ = II − H g g ˉ . \operatorname{II}(X,Y)=-\langle\nabla_X\nu,Y\rangle_g,\qquad K_g=\operatorname{tr}_{\bar{g}}\operatorname{II},\qquad H_g=\frac{K_g}{n-1},\qquad\mathring{\operatorname{II}}=\operatorname{II}-H_g\bar{g}. II ( X , Y ) = − ⟨ ∇ X ν , Y ⟩ g , K g = tr g ˉ II , H g = n − 1 K g , II ˚ = II − H g g ˉ .
Thus the Euclidean unit ball has H g = 1 H_g=1 H g = 1 ; see [41 Remark 3.12] .
Energy and geometric assumptions
Retain q q q from the introduction, also denoted 2 ∂ ∗ 2_{\partial}^{*} 2 ∂ ∗ , and put a n = 4 ( n − 1 ) / ( n − 2 ) a_n=4(n-1)/(n-2) a n = 4 ( n − 1 ) / ( n − 2 ) . The conformal operators are
L g ∘ = − Δ g + a n − 1 Scal g , B g ∘ = − ∂ ν + n − 2 2 H g . (6) L_g^{\circ}=-\Delta_g+a_n^{-1}\operatorname{Scal}_g,\qquad B_g^{\circ}=-\partial_\nu+\frac{n-2}{2}H_g. \tag*{(6)} L g ∘ = − Δ g + a n − 1 Scal g , B g ∘ = − ∂ ν + 2 n − 2 H g . (6)
The polarization of (1) is
N g ∘ ( u , v ) = ∫ M ( ⟨ ∇ u , ∇ v ⟩ g + a n − 1 Scal g u v ) d V g + n − 2 2 ∫ ∂ M H g u v d σ g . (7) \mathcal{N}_g^{\circ}(u,v)=\int_M\left(\langle\nabla u,\nabla v\rangle_g+a_n^{-1}\operatorname{Scal}_g uv\right)\,dV_g+\frac{n-2}{2}\int_{\partial M}H_guv\,d\sigma_g. \tag*{(7)} N g ∘ ( u , v ) = ∫ M ( ⟨ ∇ u , ∇ v ⟩ g + a n − 1 Scal g uv ) d V g + 2 n − 2 ∫ ∂ M H g uv d σ g . (7)
We suppress the subscript g g g when the metric is fixed. Green’s identity gives, for smooth u u u ,
∫ M u L g ∘ v d V g + ∫ ∂ M u B g ∘ v d σ g = N g ∘ ( u ) , Q g ( u ) = a n N g ∘ ( u ) , (8) \int_M uL_g^{\circ}v\,dV_g+\int_{\partial M}uB_g^{\circ}v\,d\sigma_g=\mathcal{N}_g^{\circ}(u),\qquad Q_g(u)=a_n\mathcal{N}_g^{\circ}(u), \tag*{(8)} ∫ M u L g ∘ v d V g + ∫ ∂ M u B g ∘ v d σ g = N g ∘ ( u ) , Q g ( u ) = a n N g ∘ ( u ) , (8)
where Q g ( u ) = a n ∫ M ∣ ∇ u ∣ 2 + ∫ M Scal g u 2 + 2 ∫ ∂ M K g u 2 Q_g(u)=a_n\int_M|\nabla u|^2+\int_M\operatorname{Scal}_g u^2+2\int_{\partial M}K_gu^2 Q g ( u ) = a n ∫ M ∣∇ u ∣ 2 + ∫ M Scal g u 2 + 2 ∫ ∂ M K g u 2 . For H 1 H^1 H 1 functions, the operator pairing in (8) is defined by the energy bilinear form.
We retain the strict coercivity hypothesis ( Y ∂ + ) (Y^{+}_{\partial}) ( Y ∂ + ) of (1.3). The bounded coefficients and the trace inequality also give
N g ∘ ( u ) ≤ C ∥ u ∥ H 1 ( M ) 2 . (9) \mathcal{N}^{\circ}_{g}(u) \le C\lVert u\rVert_{H^{1}(M)}^{2}. \tag*{(9)} N g ∘ ( u ) ≤ C ∥ u ∥ H 1 ( M ) 2 . (9)
Thus ⟨ u , v ⟩ E , g = N g ∘ ( u , v ) \langle u,v\rangle_{E,g}=\mathcal{N}^{\circ}_{g}(u,v) ⟨ u , v ⟩ E , g = N g ∘ ( u , v ) is an inner product, and ∥ u ∥ E , g = N g ∘ ( u ) \lVert u\rVert_{E,g}=\sqrt{\mathcal{N}^{\circ}_{g}(u)} ∥ u ∥ E , g = N g ∘ ( u ) is equivalent to the H 1 H^{1} H 1 norm.
The quotient and sharp constant are defined in (2) . If C t r C_{\mathrm{tr}} C tr is the norm of the trace map H 1 ( M ) → L q ( ∂ M ) H^{1}(M)\to L^{q}(\partial M) H 1 ( M ) → L q ( ∂ M ) , then
C E s c ∗ ( M , g ) ≥ c Y C t r − 2 > 0. (10) C^{*}_{\mathrm{Esc}}(M,g)\ge c_{Y}C_{\mathrm{tr}}^{-2}>0. \tag*{(10)} C Esc ∗ ( M , g ) ≥ c Y C tr − 2 > 0. (10)
We write DefEsc g ( u ) = J g [ u ] − C E s c ∗ ( M , g ) \operatorname{DefEsc}_{g}(u)=\mathcal{J}_{g}[u]-C^{*}_{\mathrm{Esc}}(M,g) DefEsc g ( u ) = J g [ u ] − C Esc ∗ ( M , g ) . The hypothesis ( Y ∂ + ) (Y^{+}_{\partial}) ( Y ∂ + ) is standard in the boundary Yamabe problem; see [18 , 19 , 28 ] .
Write Rm g \operatorname{Rm}_{g} Rm g and Ric g \operatorname{Ric}_{g} Ric g for the Riemann and Ricci curvature tensors, respectively. The bounded-geometry assumption ( B G k + ) (\mathrm{BG}^{k+}) ( BG k + ) means that there is a uniform radius r 0 > 0 r_{0}>0 r 0 > 0 for boundary normal charts, interior geodesic charts, and the inward Fermi collar, together with bounds
sup M ∣ ∇ j Rm g ∣ ≤ C j ( 0 ≤ j ≤ k − 2 ) , sup ∂ M ∣ ∇ j I I ∣ ≤ C ~ j ( 0 ≤ j ≤ k − 1 ) . \sup_{M}\lvert\nabla^{j}\operatorname{Rm}_{g}\rvert\le C_{j}\quad(0\le j\le k-2),\qquad\sup_{\partial M}\lvert\nabla^{j}\mathrm{II}\rvert\le\widetilde{C}_{j}\quad(0\le j\le k-1). M sup ∣ ∇ j Rm g ∣ ≤ C j ( 0 ≤ j ≤ k − 2 ) , ∂ M sup ∣ ∇ j II ∣ ≤ C j ( 0 ≤ j ≤ k − 1 ) .
More precisely, inj ( ∂ M , g ‾ ) ≥ r 0 \operatorname{inj}(\partial M,\overline{g})\ge r_{0} inj ( ∂ M , g ) ≥ r 0 , a center at distance at least r ≤ r 0 r\le r_{0} r ≤ r 0 from the boundary has an interior chart of radius r r r , and the normal exponential map is a diffeomorphism from ∂ M × [ 0 , r 0 ) \partial M\times[0,r_{0}) ∂ M × [ 0 , r 0 ) onto the collar. Here ∇ ⊤ \nabla^{\top} ∇ ⊤ is the boundary Levi-Civita connection. The trace and mixed coordinate estimates are given in [41 Lemmas 3.1–3.2] . For a fixed smooth compact manifold their constants may depend on ( M , g ) (M,g) ( M , g ) and the chart radii. For sharp manifold trace inequalities with a lower-order boundary term, see [35 ] .
Curvature conventions. We take
R ( X , Y ) Z = ∇ X ∇ Y Z − ∇ Y ∇ X Z − ∇ [ X , Y ] Z R(X,Y)Z=\nabla_{X}\nabla_{Y}Z-\nabla_{Y}\nabla_{X}Z-\nabla_{[X,Y]}Z R ( X , Y ) Z = ∇ X ∇ Y Z − ∇ Y ∇ X Z − ∇ [ X , Y ] Z
and R i j k l = ⟨ R ( ∂ k , ∂ l ) ∂ j , ∂ i ⟩ g R_{ijkl}=\langle R(\partial_{k},\partial_{l})\partial_{j},\partial_{i}\rangle_{g} R ij k l = ⟨ R ( ∂ k , ∂ l ) ∂ j , ∂ i ⟩ g . The round sphere therefore has R i j k l = g i k g j l − g i l g j k R_{ijkl}=g_{ik}g_{jl}-g_{il}g_{jk} R ij k l = g ik g j l − g i l g j k . Let Σ t \Sigma_{t} Σ t be the parallel hypersurface at inward distance t t t , and differentiate the Fermi components of its second fundamental form. At t = 0 t=0 t = 0 ,
∂ t ( I I Σ t ) a b = − ( I I 2 ) a b + R a n b n , ( ∇ ν I I ) a b = ( I I 2 ) a b + R a n b n . \partial_{t}(\mathrm{II}_{\Sigma_{t}})_{ab}=-(\mathrm{II}^{2})_{ab}+R_{anbn},\qquad(\nabla_{\nu}\mathrm{II})_{ab}=(\mathrm{II}^{2})_{ab}+R_{anbn}. ∂ t ( II Σ t ) ab = − ( II 2 ) ab + R anbn , ( ∇ ν II ) ab = ( II 2 ) ab + R anbn .
The second identity uses the covariant normal derivative of the same tensor family. The Gauss equation is
Scal g ‾ = Scal g ∣ ∂ M − 2 Ric g ( ν , ν ) + K g 2 − ∣ I I ∣ 2 . \operatorname{Scal}_{\overline{g}}=\left.\operatorname{Scal}_{g}\right|_{\partial M}-2\operatorname{Ric}_{g}(\nu,\nu)+K_{g}^{2}-\lvert\mathrm{II}\rvert^{2}. Scal g = Scal g ∣ ∂ M − 2 Ric g ( ν , ν ) + K g 2 − ∣ II ∣ 2 .
Tangential indices a , b a,b a , b range from 1 1 1 to n − 1 n-1 n − 1 , repeated indices are summed, and tangential indices are raised with g ‾ \overline{g} g .
Model profiles and boundary bubbles
We use the closed half-space R + n = R n − 1 × [ 0 , ∞ ) \mathbb{R}^{n}_{+}=\mathbb{R}^{n-1}\times[0,\infty) R + n = R n − 1 × [ 0 , ∞ ) and the round hemisphere S + n = { x ∈ S n : x n + 1 ≥ 0 } S^{n}_{+}=\{x\in S^{n}:x_{n+1}\ge0\} S + n = { x ∈ S n : x n + 1 ≥ 0 } as manifolds with boundary. Equations in either domain refer to its interior. Write y = ( y ′ , t ) y=(y',t) y = ( y ′ , t ) , also denoted Y Y Y in rescaled calculations, B R + = B R ∩ R + n B^{+}_{R}=B_{R}\cap\mathbb{R}^{n}_{+} B R + = B R ∩ R + n , and ∇ tan = ( ∂ y a ) a = 1 n − 1 \nabla_{\tan}=(\partial_{y_{a}})_{a=1}^{n-1} ∇ t a n = ( ∂ y a ) a = 1 n − 1 . Flat volume and boundary measures are denoted by d y dy d y and d σ d\sigma d σ .
The space D 1 , 2 ( R + n ) = H ˙ 1 ( R + n ) D^{1,2}(\mathbb{R}^{n}_{+})=\dot{H}^{1}(\mathbb{R}^{n}_{+}) D 1 , 2 ( R + n ) = H ˙ 1 ( R + n ) is the completion, in the norm ∥ ∇ w ∥ L 2 \lVert\nabla w\rVert_{L^{2}} ∥ ∇ w ∥ L 2 , of restrictions of C c ∞ ( R n ) C^{\infty}_{c}(\mathbb{R}^{n}) C c ∞ ( R n ) to R + n \mathbb{R}^{n}_{+} R + n . Its elements belong to L 2 n / ( n − 2 ) ( R + n ) L^{2n/(n-2)}(\mathbb{R}^{n}_{+}) L 2 n / ( n − 2 ) ( R + n ) and have traces in L q ( ∂ R + n ) L^{q}(\partial\mathbb{R}^{n}_{+}) L q ( ∂ R + n ) . We use the analogous definition on R n \mathbb{R}^{n} R n .
The hemisphere constant S ∗ S_{*} S ∗ also equals the sharp flat half-space constant [17 , 4 ] . With Γ \Gamma Γ denoting Euler’s gamma function, we fix the optimizer [17 Eq. (2)]
U + ( y ′ , t ) = A n ( ∣ y ′ ∣ 2 + ( 1 + t ) 2 ) − ( n − 2 ) / 2 , A n 2 = 2 n − 2 Γ ( n / 2 ) ( n − 2 ) π n / 2 . U_{+}(y',t)=A_{n}\left(\lvert y'\rvert^{2}+(1+t)^{2}\right)^{-(n-2)/2},\qquad A_{n}^{2}=\frac{2^{n-2}\Gamma(n/2)}{(n-2)\pi^{n/2}}. U + ( y ′ , t ) = A n ( ∣ y ′ ∣ 2 + ( 1 + t ) 2 ) − ( n − 2 ) /2 , A n 2 = ( n − 2 ) π n /2 2 n − 2 Γ ( n /2 ) .
This choice gives ∥ ∇ U + ∥ 2 = 1 \lVert\nabla U_+\rVert_2=1 ∥ ∇ U + ∥ 2 = 1 and ∥ U + ( ⋅ , 0 ) ∥ q q = S ∗ − 1 \lVert U_+(\cdot,0)\rVert_q^q=S_*^{-1} ∥ U + ( ⋅ , 0 ) ∥ q q = S ∗ − 1 . Put t ∗ = S ∗ 1 / 2 t_*=S_*^{1/2} t ∗ = S ∗ 1/2 and
U ∗ = t ∗ U + . (11) U_*=t_*U_+. \tag*{(11)} U ∗ = t ∗ U + . (11)
Then ∥ U ∗ ( ⋅ , 0 ) ∥ q = 1 \lVert U_*(\cdot,0)\rVert_q=1 ∥ U ∗ ( ⋅ , 0 ) ∥ q = 1 , ∥ ∇ U ∗ ∥ 2 2 = S ∗ \lVert\nabla U_*\rVert_2^2=S_* ∥ ∇ U ∗ ∥ 2 2 = S ∗ , and
− Δ U ∗ = 0 , − ∂ t U ∗ = S ∗ U ∗ q − 1 on ∂ R + n . -\Delta U_*=0,\qquad-\partial_tU_*=S_*U_*^{q-1}\quad\text{on }\partial\mathbb{R}_+^n. − Δ U ∗ = 0 , − ∂ t U ∗ = S ∗ U ∗ q − 1 on ∂ R + n .
For U + U_+ U + the boundary equation coefficient is instead λ f l a t = S ∗ q / 2 \lambda_{\mathrm{flat}}=S_*^{q/2} λ flat = S ∗ q /2 . All profile moments below use this fixed U + U_+ U + .
Convention 2.1 (Cutoff function). Fix η ∈ C c ∞ ( [ 0 , ∞ ) ; [ 0 , 1 ] ) \eta\in C^\infty_c([0,\infty);[0,1]) η ∈ C c ∞ ([ 0 , ∞ ) ; [ 0 , 1 ]) equal to 1 on [ 0 , 1 ] [0,1] [ 0 , 1 ] and 0 on [ 2 , ∞ ) [2,\infty) [ 2 , ∞ ) . Set χ R ( y ) = η ( ∣ y ∣ / R ) \chi_R(y)=\eta(|y|/R) χ R ( y ) = η ( ∣ y ∣/ R ) . Since q − 2 = 2 / ( n − 2 ) q-2=2/(n-2) q − 2 = 2/ ( n − 2 ) ,
χ R q − 2 = η R 2 , χ R q − 1 = η R q − 1 , χ R q = η R 2 ( n − 1 ) , η R ( y ) = η ( ∣ y ∣ / R ) . \chi_R^{q-2}=\eta_R^2,\qquad\chi_R^{q-1}=\eta_R^{q-1},\qquad\chi_R^q=\eta_R^{2(n-1)},\qquad\eta_R(y)=\eta(|y|/R). χ R q − 2 = η R 2 , χ R q − 1 = η R q − 1 , χ R q = η R 2 ( n − 1 ) , η R ( y ) = η ( ∣ y ∣/ R ) . Put γ = q − 2 \gamma=q-2 γ = q − 2 , α = ( n − 2 ) / 2 \alpha=(n-2)/2 α = ( n − 2 ) /2 , and R = ε − β R=\varepsilon^{-\beta} R = ε − β . We use logarithmic scale derivatives D = ε ∂ ε \mathbf{D}=\varepsilon\partial_\varepsilon D = ε ∂ ε . For ρ = ∣ Y ∣ / R \rho=|Y|/R ρ = ∣ Y ∣/ R , differentiation at fixed model coordinates gives
D ( χ R γ ) = β ρ ( η γ ) ′ ( ρ ) , \mathbf{D}(\chi_R^\gamma)=\beta\rho(\eta^{\gamma})'(\rho), D ( χ R γ ) = β ρ ( η γ ) ′ ( ρ ) , whereas at fixed coordinates on M M M the annular factor is ( β − 1 ) ρ ( η γ ) ′ ( ρ ) (\beta-1)\rho(\eta^{\gamma})'(\rho) ( β − 1 ) ρ ( η γ ) ′ ( ρ ) . Their first two logarithmic derivatives are bounded. After multiplication by U + U_+ U + , each annular coefficient has L n − 1 ( ∂ R + n ) L^{n-1}(\partial\mathbb{R}_+^n) L n − 1 ( ∂ R + n ) norm O ( R − 1 ) O(R^{-1}) O ( R − 1 ) . We call V γ V^\gamma V γ a boundary potential. The estimate
∣ ∫ ∂ M V h ϕ ∣ ≤ ∥ V ∥ n − 1 ∥ h ∥ q ∥ ϕ ∥ q ≤ c Y − 1 c t r 2 ∥ V ∥ n − 1 ∥ h ∥ E ∥ ϕ ∥ E (12) \left|\int_{\partial M}Vh\phi\right|\leq\lVert V\rVert_{n-1}\lVert h\rVert_q\lVert\phi\rVert_q\leq c_Y^{-1}c_{\mathrm{tr}}^2\lVert V\rVert_{n-1}\lVert h\rVert_E\lVert\phi\rVert_E \tag*{(12)} ∫ ∂ M V h ϕ ≤ ∥ V ∥ n − 1 ∥ h ∥ q ∥ ϕ ∥ q ≤ c Y − 1 c tr 2 ∥ V ∥ n − 1 ∥ h ∥ E ∥ ϕ ∥ E (12) explains its action on functions in the energy space. Ordinary ∂ ε \partial_\varepsilon ∂ ε derivatives introduce an additional factor ε − 1 \varepsilon^{-1} ε − 1 .
Let Φ x \Phi_x Φ x be a Fermi chart centered at x ∈ ∂ M x\in\partial M x ∈ ∂ M . For w w w supported in B 2 ε R + B^+_{2\varepsilon R} B 2 εR + and 2 ε R < r 0 2\varepsilon R<r_0 2 εR < r 0 , define its rescaling in Fermi coordinates by
( T x , ε w ) ( z ) = ε − ( n − 2 ) / 2 w ( Φ x − 1 ( z ) / ε ) in the chart , (T_{x,\varepsilon}w)(z)=\varepsilon^{-(n-2)/2}w\left(\Phi_x^{-1}(z)/\varepsilon\right)\quad\text{in the chart}, ( T x , ε w ) ( z ) = ε − ( n − 2 ) /2 w ( Φ x − 1 ( z ) / ε ) in the chart ,
and extend it by zero outside. For general w w w we first insert χ R \chi_R χ R .
Remark 2.2 (Cutoff radius). Unless another exponent is specified, we take R ( ε ) = ε − 3 / 4 R(\varepsilon)=\varepsilon^{-3/4} R ( ε ) = ε − 3/4 . The support radius on M M M is 2 ε R ( ε ) = 2 ε 1 / 4 2\varepsilon R(\varepsilon)=2\varepsilon^{1/4} 2 εR ( ε ) = 2 ε 1/4 which tends to zero as ε → 0 \varepsilon\to0 ε → 0 .
Definition 2.3 (Bubbles and their parameter families). The boundary bubble and its boundary normalization are
U + , x , ε = T x , ε ( χ R ( ε ) U + ) , U x , ε = U + , x , ε ∥ U + , x , ε ∥ L q ( ∂ M ) . U_{+,x,\varepsilon}=T_{x,\varepsilon}(\chi_{R(\varepsilon)}U_+),\qquad U_{x,\varepsilon}=\frac{U_{+,x,\varepsilon}}{\lVert U_{+,x,\varepsilon}\rVert_{L^q(\partial M)}}. U + , x , ε = T x , ε ( χ R ( ε ) U + ) , U x , ε = ∥ U + , x , ε ∥ L q ( ∂ M ) U + , x , ε . Allowing positive scalar multiples gives
C ∂ = { a U + , x , ε : a > 0 } . \mathcal{C}_{\partial}=\{aU_{+,x,\varepsilon}:a>0\}. C ∂ = { a U + , x , ε : a > 0 } . Only scales whose supports lie in the stated charts are included. A concentrating subfamily will be specified in each modulation statement. For α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) , U x , ε ( α ) U_{x,\varepsilon}^{(\alpha)} U x , ε ( α ) denotes the same normalized construction with R ( ε ) = ε − α R(\varepsilon)=\varepsilon^{-\alpha} R ( ε ) = ε − α .
Convention 2.4 (Coefficients and diagonal limits). The geometric coefficients are obtained by expanding at fixed R R R and then letting R → ∞ R\to\infty R → ∞ . Under the stated dimension assumptions, the required moments converge. We evaluate the expansions along R ( ε ) = ε − 3 / 4 R(\varepsilon)=\varepsilon^{-3/4} R ( ε ) = ε − 3/4 unless otherwise specified; the stated remainders include the resulting truncation errors. The same choice is admissible for finitely many scales k ε k\varepsilon k ε , k = 1 , 2 , 3 k=1,2,3 k = 1 , 2 , 3 . Fixed-radius statements display their dependence on R R R explicitly.
The modulation directions are ∂ ε U + , x , ε \partial_\varepsilon U_{+,x,\varepsilon} ∂ ε U + , x , ε and ∂ x a U + , x , ε \partial_{x_a}U_{+,x,\varepsilon} ∂ x a U + , x , ε . We impose energy orthogonality to them, together with the boundary constraint described below.
Higher-order coordinate remainders use the stated (BG) assumptions.
The boundary constraint
On S = { u ∈ H 1 ( M ) : ∥ u ∥ L q ( ∂ M ) = 1 } S = \{u \in H^1(M) : \|u\|_{L^q(\partial M)} = 1\} S = { u ∈ H 1 ( M ) : ∥ u ∥ L q ( ∂ M ) = 1 } , the tangent space and normalization map are
T u S = { h : ∫ ∂ M ∣ u ∣ q − 2 u h d σ g = 0 } , ret u ( h ) = u + h ∥ u + h ∥ L q ( ∂ M ) . T_uS = \left\{h : \int_{\partial M} |u|^{q-2}uh\,\mathrm{d}\sigma_g = 0\right\}, \qquad\operatorname{ret}_u(h) = \frac{u+h}{\|u+h\|_{L^q(\partial M)}}. T u S = { h : ∫ ∂ M ∣ u ∣ q − 2 u h d σ g = 0 } , ret u ( h ) = ∥ u + h ∥ L q ( ∂ M ) u + h .
The constrained gradient is the energy Riesz representative of D ( J g ∣ S ) [ u ] [ h ] = 2 N g ∘ ( u , h ) D(\mathcal{J}_g|_S)[u][h] = 2\mathcal{N}^{\circ}_g(u,h) D ( J g ∣ S ) [ u ] [ h ] = 2 N g ∘ ( u , h ) . Thus a Palais–Smale sequence at level c c c satisfies J g [ u k ] → c \mathcal{J}_g[u_k] \to c J g [ u k ] → c and ∥ grad S J g ( u k ) ∥ E , g → 0 \|\operatorname{grad}_S\mathcal{J}_g(u_k)\|_{E,g} \to0 ∥ grad S J g ( u k ) ∥ E , g → 0 . We use H E − 1 ( M ) H^{-1}_E(M) H E − 1 ( M ) for the energy-norm dual; boundary distributions act through the trace. For n ≥ 5 n \ge5 n ≥ 5 , put γ = q − 2 = 2 / ( n − 2 ) ∈ ( 0 , 1 ) \gamma= q - 2 = 2/(n - 2) \in(0,1) γ = q − 2 = 2/ ( n − 2 ) ∈ ( 0 , 1 ) .
Standard-bubble expansions
The standard-bubble expansion supplies the geometric terms in the reduction.
Proposition 2.5 (Standard-bubble expansions). Let C ⊂ ∂ M C \subset\partial M C ⊂ ∂ M be an open boundary region with a fixed Fermi collar, and suppose that H g = 0 H_g = 0 H g = 0 on C C C . Use the normalized bubbles of Definition 2.3.
Assume n ≥ 5 n \ge5 n ≥ 5 and (BG3+), and use the cutoff R ( ε ) = ε − 3 / 4 R(\varepsilon) = \varepsilon^{-3/4} R ( ε ) = ε − 3/4 . Put κ 3 b a r e ( n ) = ( 6 − n ) / [ 2 ( n − 1 ) ( n − 3 ) ( n − 4 ) ] \kappa^{\mathrm{bare}}_3(n) = (6-n)/[2(n-1)(n-3)(n-4)] κ 3 bare ( n ) = ( 6 − n ) / [ 2 ( n − 1 ) ( n − 3 ) ( n − 4 )] . Uniformly for x x x in compact subsets of C C C ,
J g [ U x , ε ] S ∗ = 1 + κ 3 b a r e ( n ) ∣ I I ˚ g ( x ) ∣ 2 ε 2 + e 2 ( x , ε ) , (13) \frac{\mathcal{J}_g[U_{x,\varepsilon}]}{S_*} = 1 + \kappa^{\mathrm{bare}}_3(n)|\mathring{\mathrm{II}}_g(x)|^2\varepsilon^2 + e_2(x,\varepsilon), \tag*{(13)} S ∗ J g [ U x , ε ] = 1 + κ 3 bare ( n ) ∣ II ˚ g ( x ) ∣ 2 ε 2 + e 2 ( x , ε ) , (13) ∣ e 2 ( x , ε ) ∣ + ∣ ε ∂ ε e 2 ( x , ε ) ∣ = o ( ε 2 ) . |e_2(x,\varepsilon)| + |\varepsilon\partial_\varepsilon e_2(x,\varepsilon)| = o(\varepsilon^2). ∣ e 2 ( x , ε ) ∣ + ∣ ε ∂ ε e 2 ( x , ε ) ∣ = o ( ε 2 ) . In particular,
1 S ∗ ∂ ε J g [ U x , ε ] = 2 κ 3 b a r e ( n ) ∣ I I ˚ g ( x ) ∣ 2 ε + o ( ε ) . (14) \frac{1}{S_*}\partial_\varepsilon\mathcal{J}_g[U_{x,\varepsilon}] = 2\kappa^{\mathrm{bare}}_3(n)|\mathring{\mathrm{II}}_g(x)|^2\varepsilon+ o(\varepsilon). \tag*{(14)} S ∗ 1 ∂ ε J g [ U x , ε ] = 2 κ 3 bare ( n ) ∣ II ˚ g ( x ) ∣ 2 ε + o ( ε ) . (14) Assume n ≥ 6 n \ge6 n ≥ 6 and (BG4+). At an umbilic point p ∈ C p \in C p ∈ C , both the quadratic coefficient and the geometric cubic coefficient Θ g ( p ) \Theta_g(p) Θ g ( p ) vanish. Here Θ g \Theta_g Θ g is the cubic coefficient of J g / S ∗ \mathcal{J}_g/S_* J g / S ∗ , obtained from the fixed-R R R expansion followed by R → ∞ R \to\infty R → ∞ . Consequently,
J g [ U p , ε ] = S ∗ + o ( ε 3 ) ( n ≥ 7 ) . \mathcal{J}_g[U_{p,\varepsilon}] = S_* + o(\varepsilon^3) \qquad(n \ge7). J g [ U p , ε ] = S ∗ + o ( ε 3 ) ( n ≥ 7 ) . For n = 6 n = 6 n = 6 , the same conclusion holds for U p , ε ( α ) U_{p,\varepsilon}^{(\alpha)} U p , ε ( α ) with each fixed 3 / 4 < α < 1 3/4 < \alpha< 1 3/4 < α < 1 .
Proof. The quadratic coefficient is [41 Proposition 4.7] . Put D = ε ∂ ε \mathbf{D} = \varepsilon\partial_\varepsilon D = ε ∂ ε . In the rescaled Fermi integrals the first-order term vanishes by H g = 0 H_g = 0 H g = 0 and angular integration. The flat tail, the second-moment tail, and the Taylor error, including one D \mathbf{D} D derivative, are bounded by
C ( R 2 − n + ε 2 R 4 − n + ε 3 L n ( R ) ) . C\left(R^{2-n} + \varepsilon^2R^{4-n} + \varepsilon^3L_n(R)\right). C ( R 2 − n + ε 2 R 4 − n + ε 3 L n ( R ) ) .
where L 5 ( R ) = log ( 2 R ) L_5(R) = \log(2R) L 5 ( R ) = log ( 2 R ) and L n ( R ) = 1 L_n(R) = 1 L n ( R ) = 1 for n ≥ 6 n \ge6 n ≥ 6 . The radial derivative of the metric remainder has the same order under (BG3+); a derivative of χ R \chi_R χ R is supported on the same annulus and has the same decay. Scalar and boundary-density errors obey these bounds as well. The boundary norm is bounded away from zero. Thus division preserves the estimates, which are o ( ε 2 ) o(\varepsilon^2) o ( ε 2 ) for R = ε − 3 / 4 R = \varepsilon^{-3/4} R = ε − 3/4 . This proves (13) and its ordinary scale-derivative consequence.
Every cubic contraction contains an undifferentiated I I ˚ \mathring{\mathrm{II}} II ˚ by [41 Equation (B.21), Corollary B.10] . The quadratic and cubic coefficients therefore vanish at p p p . The expansion of [41 Lemma 4.11 and Equations (4.25), (4.27)] leaves the bound
C ( R 2 − n + ε 2 R 4 − n + ε 3 R 5 − n + ε 4 K n ( R ) ) . C\left(R^{2-n} + \varepsilon^2R^{4-n} + \varepsilon^3R^{5-n} + \varepsilon^4K_n(R)\right). C ( R 2 − n + ε 2 R 4 − n + ε 3 R 5 − n + ε 4 K n ( R ) ) .
where K 6 ( R ) = log ( 2 R ) K_6(R) = \log(2R) K 6 ( R ) = log ( 2 R ) and K n ( R ) = 1 K_n(R) = 1 K n ( R ) = 1 for n ≥ 7 n \ge7 n ≥ 7 . This is o ( ε 3 ) o(\varepsilon^3) o ( ε 3 ) for the cutoff R ( ε ) = ε − 3 / 4 R(\varepsilon) = \varepsilon^{-3/4} R ( ε ) = ε − 3/4 when n ≥ 7 n \ge7 n ≥ 7 . For n = 6 n = 6 n = 6 , take R = ε − α R = \varepsilon^{-\alpha} R = ε − α with α > 3 / 4 \alpha> 3/4 α > 3/4 , so the flat tail ε 4 α \varepsilon^{4\alpha} ε 4 α is also o ( ε 3 ) o(\varepsilon^3) o ( ε 3 ) . □
Coercivity and modulation
Coercivity on the hemisphere and conformal invariance
The one-bubble reduction rests on a spectral gap transverse to the optimizer orbit and a uniform choice of bubble parameters. The gap is the linearized estimate underlying half-space trace stability: Ho’s theorem [29 ] is recalled in [2 Section 2] , and a refined spectral estimate appears in [ZZZ25, Proposition 7.1]. Supplementary deficit estimates are given in Appendix A.
Fix the Cayley transform C : S + n ∖ { p ∞ } → R + n C : S^{n}_{+} \setminus\{p_{\infty}\} \to\mathbb{R}^{n}_{+} C : S + n ∖ { p ∞ } → R + n , where p ∞ ∈ ∂ S + n p_{\infty} \in\partial S^{n}_{+} p ∞ ∈ ∂ S + n . Its conformal pullback C ∗ U + C^{*}U_{+} C ∗ U + extends smoothly across p ∞ p_{\infty} p ∞ . Let F ∗ F_{*} F ∗ be its normalization to unit boundary L q L^{q} L q norm, let M S \mathcal{M}_{S} M S be its positive normalized optimizer orbit under boundary-preserving conformal maps, and let E = T F ∗ M S E = T_{F_{*}}\mathcal{M}_{S} E = T F ∗ M S . On the hemisphere the constraint is
S = { u ∈ H 1 ( S + n ) : ∥ u ∥ L q ( ∂ S + n ) = 1 } . (15) \mathcal{S} = \{u \in H^{1}(S^{n}_{+}) : \|u\|_{L^{q}(\partial S^{n}_{+})} = 1\}. \tag*{(15)} S = { u ∈ H 1 ( S + n ) : ∥ u ∥ L q ( ∂ S + n ) = 1 } . (15)
Lemma 3.1 (Equivalence and conformal invariance of the energy norm). The energy norm satisfies
∥ u ∥ E , g 2 = N g ∘ ( u ) . (16) \|u\|_{\mathcal{E},g}^{2} = \mathcal{N}^{\circ}_{g}(u). \tag*{(16)} ∥ u ∥ E , g 2 = N g ∘ ( u ) . (16) It is equivalent to the H 1 H^{1} H 1 norm on the round hemisphere, and on ( M , g ) (M,g) ( M , g ) under ( B G 2 + ) (\mathrm{BG2}^{+}) ( BG2 + ) and ( Y ∂ + ) (\mathrm{Y}_{\partial}^{+}) ( Y ∂ + ) . If g ^ = ϕ 4 / ( n − 2 ) g \widehat{g} = \phi^{4/(n-2)}g g = ϕ 4/ ( n − 2 ) g , multiplication by ϕ \phi ϕ preserves the energy norm and the Escobar deficit. In particular, the Cayley pullback is an isometry from H ˙ 1 ( R + n ) \dot{H}^{1}(\mathbb{R}^{n}_{+}) H ˙ 1 ( R + n ) to the hemisphere energy space and preserves the deficit and energy distances.
Proof. Green’s identity gives
∫ M u L g ∘ v d V g + ∫ ∂ M u B g ∘ v d σ g = N g ∘ ( u , v ) . (17) \int_{M} uL_{g}^{\circ}v\,dV_{g}+\int_{\partial M}uB_{g}^{\circ}v\,d\sigma_{g}=\mathcal{N}_{g}^{\circ}(u,v). \tag*{(17)} ∫ M u L g ∘ v d V g + ∫ ∂ M u B g ∘ v d σ g = N g ∘ ( u , v ) . (17)
On the round hemisphere, Scal g > 0 \operatorname{Scal}_{g}>0 Scal g > 0 and H g = 0 H_{g}=0 H g = 0 , so the energy and H 1 H^{1} H 1 norms are equivalent. On M M M , the bounds follow from (1.3) and (9) . Conformal covariance gives
N g ^ ∘ ( v , w ) = N g ∘ ( ϕ v , ϕ w ) , ∫ ∂ M ∣ v ∣ q d σ g ^ = ∫ ∂ M ∣ ϕ v ∣ q d σ g ; \mathcal{N}_{\widehat{g}}^{\circ}(v,w)=\mathcal{N}_{g}^{\circ}(\phi v,\phi w),\qquad\int_{\partial M}|v|^{q}\,d\sigma_{\widehat{g}}=\int_{\partial M}|\phi v|^{q}\,d\sigma_{g}; N g ∘ ( v , w ) = N g ∘ ( ϕ v , ϕw ) , ∫ ∂ M ∣ v ∣ q d σ g = ∫ ∂ M ∣ ϕ v ∣ q d σ g ;
see [41 Lemma 3.13] . For the Cayley map these identities first hold for functions vanishing near the pole. Such functions are dense because a cutoff at radius ρ \rho ρ has squared Dirichlet norm O ( ρ n − 2 ) O(\rho^{n-2}) O ( ρ n − 2 ) and squared L 2 L^{2} L 2 norm O ( ρ n ) O(\rho^{n}) O ( ρ n ) , with n ≥ 3 n\geq3 n ≥ 3 . Density in the completed half-space Dirichlet space then proves the assertion. □
Lemma 3.2 (Constrained second variation at F ∗ F_{*} F ∗ ). For the normalized optimizer F ∗ F_{*} F ∗ , one has
L g r o u n d ∘ F ∗ = 0 in S + n , B g r o u n d ∘ F ∗ = λ ∗ F ∗ q − 1 on ∂ S + n , L_{\mathrm{ground}}^{\circ}F_{*}=0\text{ in }S^{n}_{+},\qquad B_{\mathrm{ground}}^{\circ}F_{*}=\lambda_{*}F_{*}^{q-1}\text{ on }\partial S^{n}_{+}, L ground ∘ F ∗ = 0 in S + n , B ground ∘ F ∗ = λ ∗ F ∗ q − 1 on ∂ S + n , with λ ∗ = N g r o u n d ∘ ( F ∗ ) = J g r o u n d [ F ∗ ] = S ∗ \lambda_{*}=\mathcal{N}_{\mathrm{ground}}^{\circ}(F_{*})=\mathcal{J}_{\mathrm{ground}}[F_{*}]=S_{*} λ ∗ = N ground ∘ ( F ∗ ) = J ground [ F ∗ ] = S ∗ . For ϕ ∈ T F ∗ S = { ϕ : ∫ ∂ S + n F ∗ q − 1 ϕ = 0 } \phi\in T_{F_{*}}\mathcal{S}=\{\phi:\int_{\partial S^{n}_{+}}F_{*}^{q-1}\phi=0\} ϕ ∈ T F ∗ S = { ϕ : ∫ ∂ S + n F ∗ q − 1 ϕ = 0 } , the constrained Jacobi form, equal to half the Hessian, is
Q [ ϕ ] = N g r o u n d ∘ ( ϕ , ϕ ) − ( q − 1 ) λ ∗ ∫ ∂ S + n F ∗ q − 2 ϕ 2 d σ . (18) Q[\phi]=\mathcal{N}_{\mathrm{ground}}^{\circ}(\phi,\phi)-(q-1)\lambda_{*}\int_{\partial S^{n}_{+}}F_{*}^{q-2}\phi^{2}\,d\sigma. \tag*{(18)} Q [ ϕ ] = N ground ∘ ( ϕ , ϕ ) − ( q − 1 ) λ ∗ ∫ ∂ S + n F ∗ q − 2 ϕ 2 d σ . (18) We denote by
Q [ ϕ , η ] : = N g r o u n d ∘ ( ϕ , η ) − ( q − 1 ) λ ∗ ∫ ∂ S + n F ∗ q − 2 ϕ η d σ Q[\phi,\eta]:=\mathcal{N}_{\mathrm{ground}}^{\circ}(\phi,\eta)-(q-1)\lambda_{*}\int_{\partial S^{n}_{+}}F_{*}^{q-2}\phi\eta\,d\sigma Q [ ϕ , η ] := N ground ∘ ( ϕ , η ) − ( q − 1 ) λ ∗ ∫ ∂ S + n F ∗ q − 2 ϕ η d σ the associated symmetric bilinear form, so that Q [ ϕ ] = Q [ ϕ , ϕ ] Q[\phi]=Q[\phi,\phi] Q [ ϕ ] = Q [ ϕ , ϕ ] . Since H g r o u n d = 0 H_{\mathrm{ground}}=0 H ground = 0 , the boundary mean-curvature term is absent. Moreover, ker Q = E \ker Q=E ker Q = E on T F ∗ S T_{F_{*}}\mathcal{S} T F ∗ S , and there exists c ∗ > 0 c_{*}>0 c ∗ > 0 such that Q [ ϕ ] ≥ c ∗ ∥ ϕ ∥ H 1 ( S + n ) 2 Q[\phi]\geq c_{*}\|\phi\|_{H^{1}(S^{n}_{+})}^{2} Q [ ϕ ] ≥ c ∗ ∥ ϕ ∥ H 1 ( S + n ) 2 for all ϕ ∈ T F ∗ S ∩ E ⊥ \phi\in T_{F_{*}}\mathcal{S}\cap E^{\perp} ϕ ∈ T F ∗ S ∩ E ⊥ .
The coercivity constant relative to the energy norm can be chosen as 2 / ( n + 2 ) 2/(n+2) 2/ ( n + 2 ) , uniformly along M S \mathcal{M}_{S} M S . In particular,
L g r o u n d ∘ ϕ = 0 , B g r o u n d ∘ ϕ = ( q − 1 ) λ ∗ F ∗ q − 2 ϕ L_{\mathrm{ground}}^{\circ}\phi=0,\qquad B_{\mathrm{ground}}^{\circ}\phi=(q-1)\lambda_{*}F_{*}^{q-2}\phi L ground ∘ ϕ = 0 , B ground ∘ ϕ = ( q − 1 ) λ ∗ F ∗ q − 2 ϕ
has no nonzero solution in T F ∗ S ∩ E ⊥ T_{F_*}S \cap E^\perp T F ∗ S ∩ E ⊥ .
Proof. Write B ( u , v ) = N g r o u n d ∘ ( u , v ) \mathcal{B}(u,v)=\mathcal{N}^{\circ}_{\mathrm{ground}}(u,v) B ( u , v ) = N ground ∘ ( u , v ) and G ( u ) = ∫ ∂ S + n ∣ u ∣ q d σ \mathcal{G}(u)=\int_{\partial S^n_+}|u|^q\,d\sigma G ( u ) = ∫ ∂ S + n ∣ u ∣ q d σ . On G = 1 \mathcal{G}=1 G = 1 , the quotient equals B ( u , u ) \mathcal{B}(u,u) B ( u , u ) . The Euler–Lagrange equation is
B ( F ∗ , η ) = λ ∗ ∫ ∂ S + n F ∗ q − 1 η d σ ∀ η ∈ H 1 ( S + n ) . (19) \mathcal{B}(F_*,\eta)=\lambda_*\int_{\partial S^n_+}F_*^{q-1}\eta\,d\sigma\qquad\forall\eta\in H^1(S^n_+). \tag*{(19)} B ( F ∗ , η ) = λ ∗ ∫ ∂ S + n F ∗ q − 1 η d σ ∀ η ∈ H 1 ( S + n ) . (19)
Testing with F ∗ F_* F ∗ gives λ ∗ = B ( F ∗ , F ∗ ) = S ∗ \lambda_*=\mathcal{B}(F_*,F_*)=S_* λ ∗ = B ( F ∗ , F ∗ ) = S ∗ .
Differentiating ∫ ∣ u t ∣ q = 1 \int|u_t|^q=1 ∫ ∣ u t ∣ q = 1 twice along a constrained curve with u ˙ 0 = ϕ \dot{u}_0=\phi u ˙ 0 = ϕ gives
∫ ∂ S + n F ∗ q − 1 u ¨ 0 = − ( q − 1 ) ∫ ∂ S + n F ∗ q − 2 ϕ 2 . \int_{\partial S^n_+}F_*^{q-1}\ddot{u}_0=-(q-1)\int_{\partial S^n_+}F_*^{q-2}\phi^2. ∫ ∂ S + n F ∗ q − 1 u ¨ 0 = − ( q − 1 ) ∫ ∂ S + n F ∗ q − 2 ϕ 2 .
Substitution in the second derivative of B ( u t , u t ) \mathcal{B}(u_t,u_t) B ( u t , u t ) proves (18) .
For Z ∈ E Z\in E Z ∈ E , choose a curve of normalized optimizers F s F_s F s with F 0 = F ∗ F_0=F_* F 0 = F ∗ and F ˙ 0 = Z \dot{F}_0=Z F ˙ 0 = Z . Each satisfies (19) with the same multiplier S ∗ S_* S ∗ . Differentiation gives Q [ Z , η ] = 0 Q[Z,\eta]=0 Q [ Z , η ] = 0 for every η ∈ H 1 ( S + n ) \eta\in H^1(S^n_+) η ∈ H 1 ( S + n ) , so E ⊂ ker Q E\subset\ker Q E ⊂ ker Q . For the reverse inclusion and the gap, use the conformal map from the half-space to the Euclidean unit ball B n B^n B n whose conformal factor is ( 2 / ( ∣ y ′ ∣ 2 + ( 1 + y n ) 2 ) ) ( n − 2 ) / 2 \left(2/(|y'|^2+(1+y_n)^2)\right)^{(n-2)/2} ( 2/ ( ∣ y ′ ∣ 2 + ( 1 + y n ) 2 ) ) ( n − 2 ) /2 . It carries U + U_+ U + to a constant, so the corresponding map from the hemisphere carries F ∗ F_* F ∗ to the boundary-normalized constant optimizer on B n B^n B n . Conformal covariance preserves the energy, constraint, and Jacobi form. Put α = ( n − 2 ) / 2 \alpha=(n-2)/2 α = ( n − 2 ) /2 . On the ball,
∥ ϕ ∥ E , B 2 = ∫ B n ∣ ∇ ϕ ∣ 2 + α ∫ ∂ B n ϕ 2 , Q B [ ϕ ] = ∫ B n ∣ ∇ ϕ ∣ 2 − ∫ ∂ B n ϕ 2 . \|\phi\|_{E,B}^2=\int_{B^n}|\nabla\phi|^2+\alpha\int_{\partial B^n}\phi^2,\qquad Q_B[\phi]=\int_{B^n}|\nabla\phi|^2-\int_{\partial B^n}\phi^2. ∥ ϕ ∥ E , B 2 = ∫ B n ∣∇ ϕ ∣ 2 + α ∫ ∂ B n ϕ 2 , Q B [ ϕ ] = ∫ B n ∣∇ ϕ ∣ 2 − ∫ ∂ B n ϕ 2 .
because the constant optimizer satisfies λ ∗ F ∗ q − 2 = α \lambda_*F_*^{q-2}=\alpha λ ∗ F ∗ q − 2 = α and α ( q − 2 ) = 1 \alpha(q-2)=1 α ( q − 2 ) = 1 .
Write ϕ = h + w \phi=h+w ϕ = h + w , where h h h is the harmonic extension of its trace and w ∈ H 0 1 ( B n ) w\in H^1_0(B^n) w ∈ H 0 1 ( B n ) . Let h ℓ h_\ell h ℓ denote the trace component in spherical harmonics of degree ℓ \ell ℓ . For H ℓ ( r , θ ) = r ℓ h ℓ ( θ ) H_\ell(r,\theta)=r^\ell h_\ell(\theta) H ℓ ( r , θ ) = r ℓ h ℓ ( θ ) , Green’s identity, with ν o u t \nu_{\mathrm{out}} ν out the outward unit normal to the ball, gives
∫ B n ∇ H ℓ ⋅ ∇ H m = ∫ ∂ B n h ℓ ∂ ν o u t H m = δ ℓ m ℓ ∥ h ℓ ∥ 2 2 , ∫ B n ∇ H ℓ ⋅ ∇ w = 0. \int_{B^n}\nabla H_\ell\cdot\nabla H_m=\int_{\partial B^n}h_\ell\partial_{\nu_{\mathrm{out}}}H_m=\delta_{\ell m}\ell\|h_\ell\|_2^2,\qquad\int_{B^n}\nabla H_\ell\cdot\nabla w=0. ∫ B n ∇ H ℓ ⋅ ∇ H m = ∫ ∂ B n h ℓ ∂ ν out H m = δ ℓ m ℓ ∥ h ℓ ∥ 2 2 , ∫ B n ∇ H ℓ ⋅ ∇ w = 0.
Together with ∥ Tr ϕ ∥ 2 2 = ∑ ℓ ∥ h ℓ ∥ 2 2 \|\operatorname{Tr}\phi\|_2^2=\sum_{\ell}\|h_\ell\|_2^2 ∥ Tr ϕ ∥ 2 2 = ∑ ℓ ∥ h ℓ ∥ 2 2 , this yields
∥ ϕ ∥ E , B 2 = ∥ ∇ w ∥ L 2 ( B n ) 2 + ∑ ℓ ≥ 0 ( ℓ + α ) ∥ h ℓ ∥ L 2 ( ∂ B n ) 2 , \|\phi\|_{E,B}^2=\|\nabla w\|_{L^2(B^n)}^2+\sum_{\ell\geq0}(\ell+\alpha)\|h_\ell\|_{L^2(\partial B^n)}^2, ∥ ϕ ∥ E , B 2 = ∥∇ w ∥ L 2 ( B n ) 2 + ℓ ≥ 0 ∑ ( ℓ + α ) ∥ h ℓ ∥ L 2 ( ∂ B n ) 2 ,
Q B [ ϕ ] = ∥ ∇ w ∥ L 2 ( B n ) 2 + ∑ ℓ ≥ 0 ( ℓ − 1 ) ∥ h ℓ ∥ L 2 ( ∂ B n ) 2 . Q_B[\phi]=\|\nabla w\|_{L^2(B^n)}^2+\sum_{\ell\geq0}(\ell-1)\|h_\ell\|_{L^2(\partial B^n)}^2. Q B [ ϕ ] = ∥∇ w ∥ L 2 ( B n ) 2 + ℓ ≥ 0 ∑ ( ℓ − 1 ) ∥ h ℓ ∥ L 2 ( ∂ B n ) 2 .
The tangency condition requires the trace to have zero mean, so h 0 = 0 h_0=0 h 0 = 0 . The kernel is exactly the degree-one harmonic space, generated by the coordinate functions; these are the derivatives of the boundary-preserving conformal motions of the constant optimizer. Under the conformal change of variables, this space gives E E E . Furthermore, ⟨ ϕ , x i ⟩ E , B = ( 1 + α ) ∫ ∂ B n ϕ x i \langle\phi,x_i\rangle_{E,B}=(1+\alpha)\int_{\partial B^n}\phi x_i ⟨ ϕ , x i ⟩ E , B = ( 1 + α ) ∫ ∂ B n ϕ x i , so orthogonality to this kernel in the energy inner product also removes h 1 h_1 h 1 . Since
ℓ − 1 ℓ + α ≥ 1 2 + α = 2 n + 2 ( ℓ ≥ 2 ) , \frac{\ell-1}{\ell+\alpha}\geq\frac{1}{2+\alpha}=\frac{2}{n+2}\qquad(\ell\geq2), ℓ + α ℓ − 1 ≥ 2 + α 1 = n + 2 2 ( ℓ ≥ 2 ) ,
one obtains Q B [ ϕ ] ≥ 2 n + 2 ∥ ϕ ∥ E , B 2 Q_B[\phi]\geq\frac{2}{n+2}\|\phi\|_{E,B}^2 Q B [ ϕ ] ≥ n + 2 2 ∥ ϕ ∥ E , B 2 on the constrained complement. Applying the inverse conformal change and using equivalence of the energy and H 1 H^1 H 1 norms proves the stated kernel identity and coercivity. Testing the Jacobi system with ϕ \phi ϕ gives uniqueness on the complement; conformal covariance gives uniformity along the optimizer orbit. □
Remark 3.3 (Modulation on a general manifold). We choose the center, scale, and amplitude by the boundary constraint and energy orthogonality. The corresponding statement on the hemisphere is Lemma A.1 .
Lemma 3.4 (Modulation near C ∂ \mathcal{C}_{\partial} C ∂ ). Assume n ≥ 5 n \ge5 n ≥ 5 , ( Y ∂ + ) (Y_{\partial}^{+}) ( Y ∂ + ) , and ( B G 2 + ) (\mathrm{BG}^{2+}) ( BG 2 + ) , and fix a Fermi collar. There are sufficiently small constants ε 1 , δ 0 , δ 1 > 0 \varepsilon_1,\delta_0,\delta_1>0 ε 1 , δ 0 , δ 1 > 0 , with δ 0 < δ 1 \delta_0<\delta_1 δ 0 < δ 1 , for the following assertion. Define the concentrating family
C ∂ = { a U + , x , ε : a > 0 , x ∈ ∂ M , 0 < ε ≤ ε 1 / 2 } . \mathcal{C}_{\partial}=\left\{aU_{+,x,\varepsilon}:a>0,\ x\in\partial M,\ 0<\varepsilon\le\varepsilon_1/2\right\}. C ∂ = { a U + , x , ε : a > 0 , x ∈ ∂ M , 0 < ε ≤ ε 1 /2 } . If u ∈ H 1 ( M ) ∖ { 0 } u\in H^1(M)\setminus\{0\} u ∈ H 1 ( M ) ∖ { 0 } satisfies dist H 1 ( u , C ∂ ) ≤ δ 0 ∥ u ∥ H 1 ( M ) \operatorname{dist}_{H^1}(u,\mathcal{C}_{\partial})\le\delta_0\|u\|_{H^1(M)} dist H 1 ( u , C ∂ ) ≤ δ 0 ∥ u ∥ H 1 ( M ) , there are parameters ( a ♯ , x ♯ , ε ♯ ) (a^{\sharp},x^{\sharp},\varepsilon^{\sharp}) ( a ♯ , x ♯ , ε ♯ ) with a ♯ > 0 a^{\sharp}>0 a ♯ > 0 and 0 < ε ♯ ≤ ε 1 0<\varepsilon^{\sharp}\le\varepsilon_1 0 < ε ♯ ≤ ε 1 such that, with v ♯ = a ♯ U + , x ♯ , ε ♯ v^{\sharp}=a^{\sharp}U_{+,x^{\sharp},\varepsilon^{\sharp}} v ♯ = a ♯ U + , x ♯ , ε ♯ and r = u − v ♯ r=u-v^{\sharp} r = u − v ♯ ,
⟨ r , ∂ ε v ♯ ⟩ E , g = 0 , ⟨ r , ∂ x α v ♯ ⟩ E , g = 0 ( α = 1 , … , n − 1 ) . \langle r,\partial_\varepsilon v^{\sharp}\rangle_{E,g}=0,\qquad\langle r,\partial_{x_\alpha}v^{\sharp}\rangle_{E,g}=0\quad(\alpha=1,\ldots,n-1). ⟨ r , ∂ ε v ♯ ⟩ E , g = 0 , ⟨ r , ∂ x α v ♯ ⟩ E , g = 0 ( α = 1 , … , n − 1 ) . The remaining orthogonality condition is ∫ ∂ M ( v ♯ ) q − 1 r d σ g = 0 \int_{\partial M}(v^{\sharp})^{q-1}r\,d\sigma_g=0 ∫ ∂ M ( v ♯ ) q − 1 r d σ g = 0 . This solution is unique among triples satisfying these orthogonality conditions, 0 < ε ≤ ε 1 0<\varepsilon\le\varepsilon_1 0 < ε ≤ ε 1 , and ∥ u − a U + , x , ε ∥ H 1 ≤ δ 1 ∥ u ∥ H 1 \|u-aU_{+,x,\varepsilon}\|_{H^1}\le\delta_1\|u\|_{H^1} ∥ u − a U + , x , ε ∥ H 1 ≤ δ 1 ∥ u ∥ H 1 . The parameter map is C 1 C^1 C 1 , and
∥ r ∥ H 1 ( M ) ≤ C dist H 1 ( u , C ∂ ) . \|r\|_{H^1(M)}\le C\operatorname{dist}_{H^1}(u,\mathcal{C}_{\partial}). ∥ r ∥ H 1 ( M ) ≤ C dist H 1 ( u , C ∂ ) . The modulation constants depend on the fixed metric and, in the finite-regularity case, on its C 3 C^3 C 3 bounds.
Proof. Write Ψ ( a , ξ , ε ) = a U + , x ( ξ ) , ε \Psi(a,\xi,\varepsilon)=aU_{+,x(\xi),\varepsilon} Ψ ( a , ξ , ε ) = a U + , x ( ξ ) , ε in a boundary chart. The equations are
∫ ∂ M Ψ q − 1 ( u − Ψ ) = 0 , ⟨ u − Ψ , ∂ ε Ψ ⟩ E = 0 , ⟨ u − Ψ , ∂ ξ α Ψ ⟩ E = 0. \int_{\partial M}\Psi^{q-1}(u-\Psi)=0,\qquad\langle u-\Psi,\partial_\varepsilon\Psi\rangle_E=0,\qquad\langle u-\Psi,\partial_{\xi_\alpha}\Psi\rangle_E=0. ∫ ∂ M Ψ q − 1 ( u − Ψ ) = 0 , ⟨ u − Ψ , ∂ ε Ψ ⟩ E = 0 , ⟨ u − Ψ , ∂ ξ α Ψ ⟩ E = 0.
Amplitude homogeneity permits ∥ u ∥ H 1 = 1 \|u\|_{H^1}=1 ∥ u ∥ H 1 = 1 ; the relevant amplitudes then lie in a fixed compact subset of ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) . Near ( a 0 , ξ 0 , ε 0 ) (a_0,\xi_0,\varepsilon_0) ( a 0 , ξ 0 , ε 0 ) use parameters ( a , ( ξ − ξ 0 ) / ε 0 , log ( ε / ε 0 ) ) (a,(\xi-\xi_0)/\varepsilon_0,\log(\varepsilon/\varepsilon_0)) ( a , ( ξ − ξ 0 ) / ε 0 , log ( ε / ε 0 )) and multiply the center and scale equations by ε 0 \varepsilon_0 ε 0 and ε \varepsilon ε , respectively. The profile has uniformly bounded first two derivatives in these parameters. The case of C 3 C^3 C 3 metrics follows from Lemma C.1. In the flat limit the Jacobian has a nonzero amplitude entry and Jacobi block − a 0 2 ( n − 2 ) I / 4 -a_0^2(n-2)I/4 − a 0 2 ( n − 2 ) I /4 ; the mixed constraint entries vanish because translation and dilation preserve the boundary mass. Fermi rescaling therefore makes the Jacobian uniformly invertible. Terms involving second profile derivatives are multiplied by u − Ψ u-\Psi u − Ψ . The quantitative inverse function theorem gives a unique C 1 C^1 C 1 solution and a uniform local distance estimate. Choosing the logarithmic scale radius smaller than log 2 \log2 log 2 ensures ε ♯ ≤ ε 1 \varepsilon^{\sharp}\le\varepsilon_1 ε ♯ ≤ ε 1 .
For uniqueness throughout the stated neighborhood, when the relative scale and center parameters leave every compact set, the flat energy inner product of the corresponding bubbles tends to zero. Fermi rescaling and cutoff tails preserve the resulting positive separation. On a compact set of relative parameters, continuity and injectivity give positive separation outside the local inversion neighborhood. Thus sufficiently small ε 1 \varepsilon_1 ε 1 and δ 1 \delta_1 δ 1 place any two candidate triples in one such neighborhood, where they coincide.
Apply the uniform local estimate at approximate nearest bubbles to obtain ∥ u − v ♯ ∥ H 1 ≤ C dist H 1 ( u , C ∂ ) \|u-v^{\sharp}\|_{H^1}\le C\operatorname{dist}_{H^1}(u,\mathcal{C}_{\partial}) ∥ u − v ♯ ∥ H 1 ≤ C dist H 1 ( u , C ∂ ) . Choose C δ 0 < δ 1 C\delta_0<\delta_1 C δ 0 < δ 1 and rescale the amplitude.
Coercivity near a boundary bubble on ( M , g ) (M,g) ( M , g )
Theorem 3.6 (Coercivity near a boundary bubble on ( M , g ) (M,g) ( M , g ) ). *Assume n ≥ 5 n \ge5 n ≥ 5 , ( Y ∂ + ) (Y_{\partial}^{+}) ( Y ∂ + ) , and ( B G 2 + ) (BG^{2+}) ( B G 2 + ) . Use the energy inner product of (7) . Then there exist ε 0 > 0 \varepsilon_0 > 0 ε 0 > 0 and c ∗ > 0 c_* > 0 c ∗ > 0 , depending only on n n n and the ( B G 2 + ) (BG^{2+}) ( B G 2 + ) and ( Y ∂ + ) (Y_{\partial}^{+}) ( Y ∂ + ) data of ( M , g ) (M,g) ( M , g ) , such that for every x ∈ ∂ M x \in\partial M x ∈ ∂ M , every ε ∈ ( 0 , ε 0 ] \varepsilon\in(0,\varepsilon_0] ε ∈ ( 0 , ε 0 ] , and every w ∈ H 1 ( M ) w \in H^1(M) w ∈ H 1 ( M ) satisfying
⟨ w , ∂ ε U + , x , ε ⟩ E , g = ⟨ w , ∇ x U + , x , ε ⟩ E , g = 0 , ∫ ∂ M ( U + , x , ε ) q − 1 w d σ g = 0 , q = 2 ( n − 1 ) n − 2 , (20) \langle w,\partial_{\varepsilon}U_{+,x,\varepsilon}\rangle_{E,g}=\langle w,\nabla_xU_{+,x,\varepsilon}\rangle_{E,g}=0,\qquad\int_{\partial M}(U_{+,x,\varepsilon})^{q-1}w\,d\sigma_g=0,\qquad q=\frac{2(n-1)}{n-2}, \tag*{(20)} ⟨ w , ∂ ε U + , x , ε ⟩ E , g = ⟨ w , ∇ x U + , x , ε ⟩ E , g = 0 , ∫ ∂ M ( U + , x , ε ) q − 1 w d σ g = 0 , q = n − 2 2 ( n − 1 ) , (20) one has
⟨ D 2 J g [ U + , x , ε ] w , w ⟩ ≥ c ∗ ∥ w ∥ E , g 2 ≍ ∥ w ∥ H 1 ( M ) 2 . (21) \langle D^2J_g[U_{+,x,\varepsilon}]w,w\rangle\ge c_*\|w\|_{E,g}^2\asymp\|w\|_{H^1(M)}^2. \tag*{(21)} ⟨ D 2 J g [ U + , x , ε ] w , w ⟩ ≥ c ∗ ∥ w ∥ E , g 2 ≍ ∥ w ∥ H 1 ( M ) 2 . (21)
Proof. Suppose the uniform gap fails. Then there are x k ∈ ∂ M x_k\in\partial M x k ∈ ∂ M , ε k ↓ 0 \varepsilon_k\downarrow0 ε k ↓ 0 , and w k w_k w k such that
∥ w k ∥ E , g = 1 , {xref ‘ ( 3.6 ) ‘ h o l d s a t ( x k , ε k ) , D 2 J g [ U + , x k , ε k ] [ w k , w k ] ≤ k − 1 . (22) \|w_k\|_{E,g}=1,\qquad\text{\{xref}`(3.6)` holds at (x_k,\varepsilon_k),\qquad D^2J_g[U_{+,x_k,\varepsilon_k}][w_k,w_k]\le k^{-1}. \tag*{(22)} ∥ w k ∥ E , g = 1 , {xref ‘ ( 3.6 ) ‘ h o l d s a t ( x k , ε k ) , D 2 J g [ U + , x k , ε k ] [ w k , w k ] ≤ k − 1 . (22)
Write w k ⇀ W w_k\rightharpoonup W w k ⇀ W in H 1 ( M ) H^1(M) H 1 ( M ) and v k = w k − W v_k=w_k-W v k = w k − W . Strong L 2 L^2 L 2 convergence on M M M and ∂ M \partial M ∂ M implies
1 = ∥ W ∥ E 2 + ∥ v k ∥ E 2 + o ( 1 ) , ∥ v k ∥ E 2 = ∫ M ∣ ∇ v k ∣ 2 + o ( 1 ) . 1=\|W\|_E^2+\|v_k\|_E^2+o(1),\qquad\|v_k\|_E^2=\int_M|\nabla v_k|^2+o(1). 1 = ∥ W ∥ E 2 + ∥ v k ∥ E 2 + o ( 1 ) , ∥ v k ∥ E 2 = ∫ M ∣∇ v k ∣ 2 + o ( 1 ) .
For the critical rescaling w ~ k ( y ) = ε k ( n − 2 ) / 2 w k ( Φ x k ( ε k y ) ) \widetilde w_k(y)=\varepsilon_k^{(n-2)/2}w_k(\Phi_{x_k}(\varepsilon_k y)) w k ( y ) = ε k ( n − 2 ) /2 w k ( Φ x k ( ε k y )) , the fixed function W W W has vanishing gradient energy on each shrinking ball in M M M . Thus the coordinate bounds give, for fixed R R R ,
∫ B R + ∣ ∇ w ~ k ∣ 2 ≤ ( 1 + O ( ε k R ) ) ∫ M ∣ ∇ v k ∣ 2 + o ( 1 ) ≤ 1 + o ( 1 ) . (23) \int_{B_R^+}|\nabla\widetilde w_k|^2\le(1+O(\varepsilon_kR))\int_M|\nabla v_k|^2+o(1)\le1+o(1). \tag*{(23)} ∫ B R + ∣∇ w k ∣ 2 ≤ ( 1 + O ( ε k R )) ∫ M ∣∇ v k ∣ 2 + o ( 1 ) ≤ 1 + o ( 1 ) . (23)
On the full expanding chart, uniform coordinate comparison gives
∫ B r 0 / ε k + ∣ ∇ w ~ k ∣ 2 ≤ C ∥ w k ∥ H 1 ( M ) 2 ≤ C . (24) \int_{B_{r_0/\varepsilon_k}^+}|\nabla\widetilde w_k|^2\le C\|w_k\|_{H^1(M)}^2\le C. \tag*{(24)} ∫ B r 0 / ε k + ∣∇ w k ∣ 2 ≤ C ∥ w k ∥ H 1 ( M ) 2 ≤ C . (24)
The global Sobolev inequality also bounds ∥ w ~ k ∥ 2 n / ( n − 2 ) \|\widetilde w_k\|_{2n/(n-2)} ∥ w k ∥ 2 n / ( n − 2 ) . The trace inequality on the expanding half-ball gives
∥ w ~ k ( ⋅ , 0 ) ∥ L q ( ∣ y ′ ∣ < r 0 / ε k ) ≤ C ( ∥ ∇ w ~ k ∥ 2 + ε k r 0 ∥ w ~ k ∥ 2 ) ≤ C . (25) \|\widetilde w_k(\cdot,0)\|_{L^q(\lvert y'\rvert<r_0/\varepsilon_k)}
\le C\left(\|\nabla\widetilde w_k\|_2+\frac{\varepsilon_k}{r_0}\|\widetilde w_k\|_2\right)\le C. \tag*{(25)} ∥ w k ( ⋅ , 0 ) ∥ L q (∣ y ′ ∣ < r 0 / ε k ) ≤ C ( ∥∇ w k ∥ 2 + r 0 ε k ∥ w k ∥ 2 ) ≤ C . (25)
because ε k ∥ w ~ k ∥ 2 ≤ C ∥ w k ∥ L 2 ( M ) \varepsilon_k\|\widetilde w_k\|_2\le C\|w_k\|_{L^2(M)} ε k ∥ w k ∥ 2 ≤ C ∥ w k ∥ L 2 ( M ) . A diagonal subsequence converges weakly in local H 1 H^1 H 1 and strongly in local boundary L 2 L^2 L 2 to w ~ \widetilde w w . By (3.9),
∫ R + n ∣ ∇ w ~ ∣ 2 ≤ 1. (26) \int_{\mathbb{R}_+^n}|\nabla\widetilde w|^2\le1. \tag*{(26)} ∫ R + n ∣∇ w ∣ 2 ≤ 1. (26)
The bound in L 2 n / ( n − 2 ) L^{2n/(n-2)} L 2 n / ( n − 2 ) gives w ~ ∈ L 2 n / ( n − 2 ) \widetilde w\in L^{2n/(n-2)} w ∈ L 2 n / ( n − 2 ) , and hence w ~ ∈ D 1 , 2 ( R + n ) \widetilde w\in D^{1,2}(\mathbb{R}_+^n) w ∈ D 1 , 2 ( R + n ) .
For tangent directions, differentiation of the quotient gives
D 2 J g [ U k ] [ w k , w k ] = 2 ∥ U k ∥ q − 2 ( 1 − ( q − 1 ) λ k ∫ ∂ M U k q − 2 w k 2 ) , λ k = N g o ( U k ) ∥ U k ∥ q q ⟶ λ f l a t , D^2J_g[U_k][w_k,w_k]=2\|U_k\|_q^{-2}\left(1-(q-1)\lambda_k\int_{\partial M}U_k^{q-2}w_k^2\right),\qquad\lambda_k=\frac{\mathcal{N}^o_g(U_k)}{\|U_k\|_q^q}\longrightarrow\lambda_{\mathrm{flat}}, D 2 J g [ U k ] [ w k , w k ] = 2∥ U k ∥ q − 2 ( 1 − ( q − 1 ) λ k ∫ ∂ M U k q − 2 w k 2 ) , λ k = ∥ U k ∥ q q N g o ( U k ) ⟶ λ flat ,
where U k = U + , x k , ε k U_k=U_{+,x_k,\varepsilon_k} U k = U + , x k , ε k . The weighted integrals are bounded by (3.11), so a further subsequence satisfies
( q − 1 ) λ f l a t ∫ ∂ M U k q − 2 w ~ k 2 ⟶ m ≥ 1. (27) (q-1)\lambda_{\mathrm{flat}}\int_{\partial M}U_k^{q-2}\widetilde w_k^2\longrightarrow m\ge1. \tag*{(27)} ( q − 1 ) λ flat ∫ ∂ M U k q − 2 w k 2 ⟶ m ≥ 1. (27)
Local trace convergence passes this integral to the limit. Indeed, U + ( ⋅ , 0 ) ∈ L n − 1 U_+(\cdot,0)\in L^{n-1} U + ( ⋅ , 0 ) ∈ L n − 1 and
∫ ∣ y ′ ∣ > R U + q − 2 w ~ k 2 ≤ C ∥ U + q − 2 ∥ L n − 1 ( ∣ y ′ ∣ > R ) ⟶ 0 \int_{\lvert y'\rvert>R}U_+^{q-2}\widetilde w_k^2\le C\|U_+^{q-2}\|_{L^{n-1}(\lvert y'\rvert>R)}\longrightarrow0 ∫ ∣ y ′ ∣ > R U + q − 2 w k 2 ≤ C ∥ U + q − 2 ∥ L n − 1 (∣ y ′ ∣ > R ) ⟶ 0
@page 12/57 starts_mid=0 ends_mid=1
Coercivity near a boundary bubble on ( M , g ) (M,g) ( M , g )
Theorem 3.6 (Coercivity near a boundary bubble on ( M , g ) (M,g) ( M , g ) ). *Assume n ≥ 5 n \ge5 n ≥ 5 , ( Y ∂ + ) (Y_{\partial}^{+}) ( Y ∂ + ) , and ( B G 2 + ) (BG^{2+}) ( B G 2 + ) . Use the energy inner product of (7) . Then there exist ε 0 > 0 \varepsilon_0 > 0 ε 0 > 0 and c ∗ > 0 c_* > 0 c ∗ > 0 , depending only on n n n and the ( B G 2 + ) (BG^{2+}) ( B G 2 + ) and ( Y ∂ + ) (Y_{\partial}^{+}) ( Y ∂ + ) data of ( M , g ) (M,g) ( M , g ) , such that for every x ∈ ∂ M x \in\partial M x ∈ ∂ M , every ε ∈ ( 0 , ε 0 ] \varepsilon\in(0,\varepsilon_0] ε ∈ ( 0 , ε 0 ] , and every w ∈ H 1 ( M ) w \in H^1(M) w ∈ H 1 ( M ) satisfying
⟨ w , ∂ ε U + , x , ε ⟩ E , g = ⟨ w , ∇ x U + , x , ε ⟩ E , g = 0 , ∫ ∂ M ( U + , x , ε ) q − 1 w d σ g = 0 , q = 2 ( n − 1 ) n − 2 , (28) \langle w,\partial_{\varepsilon}U_{+,x,\varepsilon}\rangle_{E,g}=\langle w,\nabla_xU_{+,x,\varepsilon}\rangle_{E,g}=0,\qquad\int_{\partial M}(U_{+,x,\varepsilon})^{q-1}w\,d\sigma_g=0,\qquad q=\frac{2(n-1)}{n-2}, \tag*{(28)} ⟨ w , ∂ ε U + , x , ε ⟩ E , g = ⟨ w , ∇ x U + , x , ε ⟩ E , g = 0 , ∫ ∂ M ( U + , x , ε ) q − 1 w d σ g = 0 , q = n − 2 2 ( n − 1 ) , (28) one has
⟨ D 2 J g [ U + , x , ε ] w , w ⟩ ≥ c ∗ ∥ w ∥ E , g 2 ≍ ∥ w ∥ H 1 ( M ) 2 . (29) \langle D^2J_g[U_{+,x,\varepsilon}]w,w\rangle\ge c_*\|w\|_{E,g}^2\asymp\|w\|_{H^1(M)}^2. \tag*{(29)} ⟨ D 2 J g [ U + , x , ε ] w , w ⟩ ≥ c ∗ ∥ w ∥ E , g 2 ≍ ∥ w ∥ H 1 ( M ) 2 . (29)
Proof. Suppose the uniform gap fails. Then there are x k ∈ ∂ M x_k\in\partial M x k ∈ ∂ M , ε k ↓ 0 \varepsilon_k\downarrow0 ε k ↓ 0 , and w k w_k w k such that
∥ w k ∥ E , g = 1 , {xref ‘ ( 3.6 ) ‘ h o l d s a t ( x k , ε k ) , D 2 J g [ U + , x k , ε k ] [ w k , w k ] ≤ k − 1 . (30) \|w_k\|_{E,g}=1,\qquad\text{\{xref}`(3.6)` holds at (x_k,\varepsilon_k),\qquad D^2J_g[U_{+,x_k,\varepsilon_k}][w_k,w_k]\le k^{-1}. \tag*{(30)} ∥ w k ∥ E , g = 1 , {xref ‘ ( 3.6 ) ‘ h o l d s a t ( x k , ε k ) , D 2 J g [ U + , x k , ε k ] [ w k , w k ] ≤ k − 1 . (30)
Write w k ⇀ W w_k\rightharpoonup W w k ⇀ W in H 1 ( M ) H^1(M) H 1 ( M ) and v k = w k − W v_k=w_k-W v k = w k − W . Strong L 2 L^2 L 2 convergence on M M M and ∂ M \partial M ∂ M implies
1 = ∥ W ∥ E 2 + ∥ v k ∥ E 2 + o ( 1 ) , ∥ v k ∥ E 2 = ∫ M ∣ ∇ v k ∣ 2 + o ( 1 ) . 1=\|W\|_E^2+\|v_k\|_E^2+o(1),\qquad\|v_k\|_E^2=\int_M|\nabla v_k|^2+o(1). 1 = ∥ W ∥ E 2 + ∥ v k ∥ E 2 + o ( 1 ) , ∥ v k ∥ E 2 = ∫ M ∣∇ v k ∣ 2 + o ( 1 ) .
For the critical rescaling w ~ k ( y ) = ε k ( n − 2 ) / 2 w k ( Φ x k ( ε k y ) ) \widetilde w_k(y)=\varepsilon_k^{(n-2)/2}w_k(\Phi_{x_k}(\varepsilon_k y)) w k ( y ) = ε k ( n − 2 ) /2 w k ( Φ x k ( ε k y )) , the fixed function W W W has vanishing gradient energy on each shrinking ball in M M M . Thus the coordinate bounds give, for fixed R R R ,
∫ B R + ∣ ∇ w ~ k ∣ 2 ≤ ( 1 + O ( ε k R ) ) ∫ M ∣ ∇ v k ∣ 2 + o ( 1 ) ≤ 1 + o ( 1 ) . (31) \int_{B_R^+}|\nabla\widetilde w_k|^2\le(1+O(\varepsilon_kR))\int_M|\nabla v_k|^2+o(1)\le1+o(1). \tag*{(31)} ∫ B R + ∣∇ w k ∣ 2 ≤ ( 1 + O ( ε k R )) ∫ M ∣∇ v k ∣ 2 + o ( 1 ) ≤ 1 + o ( 1 ) . (31)
On the full expanding chart, uniform coordinate comparison gives
∫ B r 0 / ε k + ∣ ∇ w ~ k ∣ 2 ≤ C ∥ w k ∥ H 1 ( M ) 2 ≤ C . (32) \int_{B_{r_0/\varepsilon_k}^+}|\nabla\widetilde w_k|^2\le C\|w_k\|_{H^1(M)}^2\le C. \tag*{(32)} ∫ B r 0 / ε k + ∣∇ w k ∣ 2 ≤ C ∥ w k ∥ H 1 ( M ) 2 ≤ C . (32)
The global Sobolev inequality also bounds ∥ w ~ k ∥ 2 n / ( n − 2 ) \|\widetilde w_k\|_{2n/(n-2)} ∥ w k ∥ 2 n / ( n − 2 ) . The trace inequality on the expanding half-ball gives
∥ w ~ k ( ⋅ , 0 ) ∥ L q ( ∣ y ′ ∣ < r 0 / ε k ) ≤ C ( ∥ ∇ w ~ k ∥ 2 + ε k r 0 ∥ w ~ k ∥ 2 ) ≤ C . (33) \|\widetilde w_k(\cdot,0)\|_{L^q(\lvert y'\rvert<r_0/\varepsilon_k)}
\le C\left(\|\nabla\widetilde w_k\|_2+\frac{\varepsilon_k}{r_0}\|\widetilde w_k\|_2\right)\le C. \tag*{(33)} ∥ w k ( ⋅ , 0 ) ∥ L q (∣ y ′ ∣ < r 0 / ε k ) ≤ C ( ∥∇ w k ∥ 2 + r 0 ε k ∥ w k ∥ 2 ) ≤ C . (33)
because ε k ∥ w ~ k ∥ 2 ≤ C ∥ w k ∥ L 2 ( M ) \varepsilon_k\|\widetilde w_k\|_2\le C\|w_k\|_{L^2(M)} ε k ∥ w k ∥ 2 ≤ C ∥ w k ∥ L 2 ( M ) . A diagonal subsequence converges weakly in local H 1 H^1 H 1 and strongly in local boundary L 2 L^2 L 2 to w ~ \widetilde w w . By (3.9),
∫ R + n ∣ ∇ w ~ ∣ 2 ≤ 1. (34) \int_{\mathbb{R}_+^n}|\nabla\widetilde w|^2\le1. \tag*{(34)} ∫ R + n ∣∇ w ∣ 2 ≤ 1. (34)
The bound in L 2 n / ( n − 2 ) L^{2n/(n-2)} L 2 n / ( n − 2 ) gives w ~ ∈ L 2 n / ( n − 2 ) \widetilde w\in L^{2n/(n-2)} w ∈ L 2 n / ( n − 2 ) , and hence w ~ ∈ D 1 , 2 ( R + n ) \widetilde w\in D^{1,2}(\mathbb{R}_+^n) w ∈ D 1 , 2 ( R + n ) .
For tangent directions, differentiation of the quotient gives
D 2 J g [ U k ] [ w k , w k ] = 2 ∥ U k ∥ q − 2 ( 1 − ( q − 1 ) λ k ∫ ∂ M U k q − 2 w k 2 ) , λ k = N g o ( U k ) ∥ U k ∥ q q ⟶ λ f l a t , D^2J_g[U_k][w_k,w_k]=2\|U_k\|_q^{-2}\left(1-(q-1)\lambda_k\int_{\partial M}U_k^{q-2}w_k^2\right),\qquad\lambda_k=\frac{\mathcal{N}^o_g(U_k)}{\|U_k\|_q^q}\longrightarrow\lambda_{\mathrm{flat}}, D 2 J g [ U k ] [ w k , w k ] = 2∥ U k ∥ q − 2 ( 1 − ( q − 1 ) λ k ∫ ∂ M U k q − 2 w k 2 ) , λ k = ∥ U k ∥ q q N g o ( U k ) ⟶ λ flat ,
where U k = U + , x k , ε k U_k=U_{+,x_k,\varepsilon_k} U k = U + , x k , ε k . The weighted integrals are bounded by (3.11), so a further subsequence satisfies
( q − 1 ) λ f l a t ∫ ∂ M U k q − 2 w ~ k 2 ⟶ m ≥ 1. (35) (q-1)\lambda_{\mathrm{flat}}\int_{\partial M}U_k^{q-2}\widetilde w_k^2\longrightarrow m\ge1. \tag*{(35)} ( q − 1 ) λ flat ∫ ∂ M U k q − 2 w k 2 ⟶ m ≥ 1. (35)
Local trace convergence passes this integral to the limit. Indeed, U + ( ⋅ , 0 ) ∈ L n − 1 U_+(\cdot,0)\in L^{n-1} U + ( ⋅ , 0 ) ∈ L n − 1 and
∫ ∣ y ′ ∣ > R U + q − 2 w ~ k 2 ≤ C ∥ U + q − 2 ∥ L n − 1 ( ∣ y ′ ∣ > R ) ⟶ 0 \int_{\lvert y'\rvert>R}U_+^{q-2}\widetilde w_k^2\le C\|U_+^{q-2}\|_{L^{n-1}(\lvert y'\rvert>R)}\longrightarrow0 ∫ ∣ y ′ ∣ > R U + q − 2 w k 2 ≤ C ∥ U + q − 2 ∥ L n − 1 (∣ y ′ ∣ > R ) ⟶ 0
uniformly in k k k . The same tail estimates hold after insertion of the cutoff and the boundary Jacobian, both of which converge to 1 1 1 on fixed balls. Define
Q f l a t [ v ] = ∫ R + n ∣ ∇ v ∣ 2 − ( q − 1 ) λ f l a t ∫ ∂ R + n U + q − 2 v 2 . Q_{\mathrm{flat}}[v]=\int_{\mathbb{R}^{n}_{+}}|\nabla v|^{2}-(q-1)\lambda_{\mathrm{flat}}\int_{\partial\mathbb{R}^{n}_{+}}U_{+}^{q-2}v^{2}. Q flat [ v ] = ∫ R + n ∣∇ v ∣ 2 − ( q − 1 ) λ flat ∫ ∂ R + n U + q − 2 v 2 .
Then
( q − 1 ) λ f l a t ∫ ∂ R + n U + q − 2 w ~ 2 = m , Q f l a t [ w ~ ] ≤ 1 − m ≤ 0. (q-1)\lambda_{\mathrm{flat}}\int_{\partial\mathbb{R}^{n}_{+}}U_{+}^{q-2}\widetilde{w}^{2}=m,\qquad Q_{\mathrm{flat}}[\widetilde{w}]\leq1-m\leq0. ( q − 1 ) λ flat ∫ ∂ R + n U + q − 2 w 2 = m , Q flat [ w ] ≤ 1 − m ≤ 0.
In particular w ~ ≠ 0 \widetilde{w}\ne0 w = 0 . The boundary constraint passes to the limit by the critical trace tail estimate. The normalized scale and center orthogonality conditions pass by local weak convergence, the Dirichlet tails of the flat Jacobi fields, and the vanishing lower-order terms. Thus w ~ \widetilde{w} w is constrained and orthogonal to every flat translation and dilation field. Conformal invariance and Lemma 3.2 give Q f l a t [ w ~ ] > 0 Q_{\mathrm{flat}}[\widetilde{w}]>0 Q flat [ w ] > 0 , a contradiction. □
Lemma 3.7 (Estimate for the first variation). Assume n ≥ 5 n\geq5 n ≥ 5 , ( B G 2 + ) (\mathrm{BG}^{2+}) ( BG 2 + ) , and ( Y ∂ + ) (Y_{\partial}^{+}) ( Y ∂ + ) . There exist ε 0 > 0 \varepsilon_{0}>0 ε 0 > 0 and C g e o > 0 C_{\mathrm{geo}}>0 C geo > 0 , depending only on n n n and the ( B G 2 + ) (\mathrm{BG}^{2+}) ( BG 2 + ) and ( Y ∂ + ) (Y_{\partial}^{+}) ( Y ∂ + ) data, such that the following holds.
For every x ∈ ∂ M x\in\partial M x ∈ ∂ M , every ε ∈ ( 0 , ε 0 ] \varepsilon\in(0,\varepsilon_{0}] ε ∈ ( 0 , ε 0 ] , and every a > 0 a>0 a > 0 , set
v x , ε : = a U + , x , ε , v_{x,\varepsilon}:=aU_{+,x,\varepsilon}, v x , ε := a U + , x , ε , where U + , x , ε = T x , ε ( χ R ( ε ) U + ) U_{+,x,\varepsilon}=T_{x,\varepsilon}(\chi_{R(\varepsilon)}U_{+}) U + , x , ε = T x , ε ( χ R ( ε ) U + ) is the boundary bubble of Definition 2.3 , with the cutoff chosen as in Remark 2.2. Then
∣ L [ v x , ε ] ( ψ ) ∣ : = ∣ D J g [ v x , ε ] ( ψ ) ∣ ≤ C g e o ε a − 1 ∥ ψ ∥ E , g \left|L[v_{x,\varepsilon}](\psi)\right|:=\left|DJ_{g}[v_{x,\varepsilon}](\psi)\right|\leq C_{\mathrm{geo}}\varepsilon a^{-1}\|\psi\|_{E,g} ∣ L [ v x , ε ] ( ψ ) ∣ := ∣ D J g [ v x , ε ] ( ψ ) ∣ ≤ C geo ε a − 1 ∥ ψ ∥ E , g for every ψ ∈ T v x , ε S v x , ε \psi\in T_{v_{x,\varepsilon}}S_{v_{x,\varepsilon}} ψ ∈ T v x , ε S v x , ε , where
S v x , ε : = { F ∈ H 1 ( M ) : ∥ F ∥ L q ( ∂ M ) = ∥ v x , ε ∥ L q ( ∂ M ) } , q = 2 ∂ ∗ . S_{v_{x,\varepsilon}}:=\left\{F\in H^{1}(M):\|F\|_{L^{q}(\partial M)}=\|v_{x,\varepsilon}\|_{L^{q}(\partial M)}\right\},\qquad q=2_{\partial}^{*}. S v x , ε := { F ∈ H 1 ( M ) : ∥ F ∥ L q ( ∂ M ) = ∥ v x , ε ∥ L q ( ∂ M ) } , q = 2 ∂ ∗ .
Proof. Write v = v x , ε v=v_{x,\varepsilon} v = v x , ε and q = 2 ∂ ∗ q=2_{\partial}^{*} q = 2 ∂ ∗ . For ψ ∈ T v S v \psi\in T_{v}S_{v} ψ ∈ T v S v , the first variation is
L [ v ] ( ψ ) = 2 ∥ v ∥ L q ( ∂ M ) 2 N g ∘ ( v , ψ ) , (36) L[v](\psi)=\frac{2}{\|v\|_{L^{q}(\partial M)}^{2}}\mathcal{N}_{g}^{\circ}(v,\psi), \tag*{(36)} L [ v ] ( ψ ) = ∥ v ∥ L q ( ∂ M ) 2 2 N g ∘ ( v , ψ ) , (36)
since the derivative of the denominator vanishes on T v S v T_{v}S_{v} T v S v . By the Green identity,
N g ∘ ( v , ψ ) = ∫ M ψ L g ∘ v d V g + ∫ ∂ M ψ B g ∘ v d σ g . \mathcal{N}_{g}^{\circ}(v,\psi)=\int_{M}\psi L_{g}^{\circ}v\,dV_{g}+\int_{\partial M}\psi B_{g}^{\circ}v\,d\sigma_{g}. N g ∘ ( v , ψ ) = ∫ M ψ L g ∘ v d V g + ∫ ∂ M ψ B g ∘ v d σ g .
It suffices by homogeneity to take a = 1 a=1 a = 1 : D J g [ a U ] ( ψ ) = a − 1 D J g [ U ] ( ψ ) DJ_{g}[aU](\psi)=a^{-1}DJ_{g}[U](\psi) D J g [ a U ] ( ψ ) = a − 1 D J g [ U ] ( ψ ) , and the tangent condition is unchanged by this amplitude. Write ψ ~ \widetilde{\psi} ψ for the critically rescaled test function. After cutoff in the fixed collar, its Dirichlet and critical trace norms are bounded by C ∥ ψ ∥ E , g C\|\psi\|_{E,g} C ∥ ψ ∥ E , g . Write P R = χ R U + P_{R}=\chi_{R}U_{+} P R = χ R U + . Testing the flat equation by χ R ψ ~ \chi_{R}\widetilde{\psi} χ R ψ gives the exact cancellation
∫ ∇ P R ⋅ ∇ ψ ~ − λ f l a t ∫ ∂ R + n P R q − 1 ψ ~ = ∫ U + ∇ χ R ⋅ ∇ ψ ~ − ∫ ψ ~ ∇ U + ⋅ ∇ χ R + λ f l a t ∫ ∂ R + n ( χ R − χ R q − 1 ) U + q − 1 ψ ~ . (37) \int\nabla P_{R}\cdot\nabla\widetilde{\psi}-\lambda_{\mathrm{flat}}\int_{\partial\mathbb{R}^{n}_{+}}P_{R}^{q-1}\widetilde{\psi}
=\int U_{+}\nabla\chi_{R}\cdot\nabla\widetilde{\psi}-\int\widetilde{\psi}\nabla U_{+}\cdot\nabla\chi_{R}+\lambda_{\mathrm{flat}}\int_{\partial\mathbb{R}^{n}_{+}}(\chi_{R}-\chi_{R}^{q-1})U_{+}^{q-1}\widetilde{\psi}. \tag*{(37)} ∫ ∇ P R ⋅ ∇ ψ − λ flat ∫ ∂ R + n P R q − 1 ψ = ∫ U + ∇ χ R ⋅ ∇ ψ − ∫ ψ ∇ U + ⋅ ∇ χ R + λ flat ∫ ∂ R + n ( χ R − χ R q − 1 ) U + q − 1 ψ . (37)
The integrals on the right are supported on R < ∣ Y ∣ < 2 R R<|Y|<2R R < ∣ Y ∣ < 2 R . Cauchy–Schwarz, Hardy’s inequality, and the trace inequality bound their sum by C R − α ∥ ψ ~ ∥ D 1 , 2 CR^{-\alpha}\|\widetilde{\psi}\|_{D^{1,2}} C R − α ∥ ψ ∥ D 1 , 2 . In the rescaled energy pairing, the principal coefficient error is O ( ε ∣ Y ∣ ) O(\varepsilon|Y|) O ( ε ∣ Y ∣ ) ; its pairing with the bubble is bounded by
C ε ∥ ∣ Y ∣ ∇ U + ∥ L 2 ∥ ψ ~ ∥ D 1 , 2 . C\varepsilon\||Y|\nabla U_{+}\|_{L^{2}}\|\widetilde{\psi}\|_{D^{1,2}}. C ε ∥∣ Y ∣∇ U + ∥ L 2 ∥ ψ ∥ D 1 , 2 .
The weighted norm is finite for n ≥ 5 n\geq5 n ≥ 5 . The boundary mean-curvature term is bounded by C ε ∥ U + ∥ L q ′ ∥ ψ ~ ∥ L q C\varepsilon\|U_{+}\|_{L^{q'}}\|\widetilde{\psi}\|_{L^{q}} C ε ∥ U + ∥ L q ′ ∥ ψ ∥ L q , q ′ = q / ( q − 1 ) q'=q/(q-1) q ′ = q / ( q − 1 ) ; U + ( ⋅ , 0 ) ∈ L q ′ U_{+}(\cdot,0)\in L^{q'} U + ( ⋅ , 0 ) ∈ L q ′ in the same dimension range. The scalar term is controlled by Hardy’s inequality, with bound
C ε 2 ∥ ∣ Y ∣ U + ∥ L 2 ( B 2 R + ) ∥ ψ ~ ∥ D 1 , 2 ≤ C ε 2 ℓ n ′ ( R ) ∥ ψ ∥ E , g , C\varepsilon^{2}\||Y|U_{+}\|_{L^{2}(B_{2R}^{+})}\|\widetilde{\psi}\|_{D^{1,2}}\leq C\varepsilon^{2}\ell_{n}'(R)\|\psi\|_{E,g}, C ε 2 ∥∣ Y ∣ U + ∥ L 2 ( B 2 R + ) ∥ ψ ∥ D 1 , 2 ≤ C ε 2 ℓ n ′ ( R ) ∥ ψ ∥ E , g ,
where ℓ 5 ′ ( R ) = R 1 / 2 \ell'_5(R)=R^{1/2} ℓ 5 ′ ( R ) = R 1/2 , ℓ 6 ′ ( R ) = ( 1 + log R ) 1 / 2 \ell'_6(R)=(1+\log R)^{1/2} ℓ 6 ′ ( R ) = ( 1 + log R ) 1/2 , and ℓ n ( R ) = 1 \ell_n(R)=1 ℓ n ( R ) = 1 for n ≥ 7 n\geq7 n ≥ 7 . The boundary density error is O ( ε 2 ) O(\varepsilon^2) O ( ε 2 ) , since ∣ Y ′ ∣ 2 U + q − 1 ∈ L q ′ ( R n − 1 ) |Y'|^2U_+^{q-1}\in L^{q'}(\mathbb{R}^{n-1}) ∣ Y ′ ∣ 2 U + q − 1 ∈ L q ′ ( R n − 1 ) for n ≥ 5 n\geq5 n ≥ 5 . The flat boundary term vanishes by tangency, up to the boundary-density error just estimated. The annular terms in (37) are bounded by
C ω ( R ) ∥ ψ ∥ E , g , ω ( R ) = O ( R − ( n − 2 ) / 2 ) . C\omega(R)\|\psi\|_{E,g},\qquad\omega(R)=O\left(R^{-(n-2)/2}\right). C ω ( R ) ∥ ψ ∥ E , g , ω ( R ) = O ( R − ( n − 2 ) /2 ) .
For R = ε − 3 / 4 R=\varepsilon^{-3/4} R = ε − 3/4 both ε 2 ℓ n ′ ( R ) \varepsilon^2\ell'_n(R) ε 2 ℓ n ′ ( R ) and ω ( R ) \omega(R) ω ( R ) are o ( ε ) o(\varepsilon) o ( ε ) . The quotient denominator is bounded above and below independently of the center and scale. Equation (3.14 now gives the claimed estimate, including the factor a − 1 a^{-1} a − 1 .
Compactness and Palais–Smale decompositions
Palais–Smale sequences are precompact below S ∗ S_* S ∗ and admit a finite profile decomposition at every level. The profile masses determine the first mixed level and characterize the noncompact levels between one and two bubbles. For the concentration–compactness and profile-decomposition background, see [32 , 33 , 23 ] .
Profile decomposition without a sign assumption
Proposition 4.1 (Boundary profile decomposition without a sign assumption). Assume ( Y ∂ + ) (Y^+_\partial) ( Y ∂ + ) and ( B G 2 + ) (\mathrm{BG}^+_2) ( BG 2 + ) . Let ( u k ) ⊂ S (u_k)\subset S ( u k ) ⊂ S be a constrained Palais–Smale sequence at level c > 0 c>0 c > 0 , with no sign assumption. Then c ≥ C E s c ∗ ( M , g ) > 0 c\geq C^*_{\mathrm{Esc}}(M,g)>0 c ≥ C Esc ∗ ( M , g ) > 0 . After passing to a subsequence, there exist a weak limit u ∞ ∈ H 1 ( M ) u_\infty\in H^1(M) u ∞ ∈ H 1 ( M ) , an integer J ≥ 0 J\geq0 J ≥ 0 , boundary points x k j ∈ ∂ M x_k^j\in\partial M x k j ∈ ∂ M , scales ε k j ↓ 0 \varepsilon_k^j\downarrow0 ε k j ↓ 0 , nonzero profiles V ( j ) ∈ H ˙ 1 ( R + n ) V^{(j)}\in\dot H^1(\mathbb{R}^n_+) V ( j ) ∈ H ˙ 1 ( R + n ) , 1 ≤ j ≤ J 1\leq j\leq J 1 ≤ j ≤ J , and cutoff radii R k j → ∞ R_k^j\to\infty R k j → ∞ with ε k j R k j → 0 \varepsilon_k^jR_k^j\to0 ε k j R k j → 0 , such that
u k = u ∞ + ∑ j = 1 J T x k j , ε k j ( χ R k j V ( j ) ) + r k , ∥ r k ∥ H 1 ( M ) → 0. (38) u_k=u_\infty+\sum_{j=1}^{J}\mathcal{T}_{x_k^j,\varepsilon_k^j}\left(\chi_{R_k^j}V^{(j)}\right)+r_k,\qquad\|r_k\|_{H^1(M)}\to0. \tag*{(38)} u k = u ∞ + j = 1 ∑ J T x k j , ε k j ( χ R k j V ( j ) ) + r k , ∥ r k ∥ H 1 ( M ) → 0. (38)
Write d ∂ d_{\partial} d ∂ for the distance of the induced boundary metric. The centers and scales satisfy the separation condition
ε k i ε k j + ε k j ε k i + d ∂ ( x k i , x k j ) 2 ε k i ε k j ⟶ ∞ ( i ≠ j ) . (39) \frac{\varepsilon_k^i}{\varepsilon_k^j}+\frac{\varepsilon_k^j}{\varepsilon_k^i}+\frac{d_{\partial}(x_k^i,x_k^j)^2}{\varepsilon_k^i\varepsilon_k^j}\longrightarrow\infty\qquad(i\neq j). \tag*{(39)} ε k j ε k i + ε k i ε k j + ε k i ε k j d ∂ ( x k i , x k j ) 2 ⟶ ∞ ( i = j ) . (39)
The weak limit solves
L g ∘ u ∞ = 0 in M , B g ∘ u ∞ = c ∣ u ∞ ∣ q − 2 u ∞ on ∂ M , L_g^\circ u_\infty=0\quad\text{in }M,\qquad B_g^\circ u_\infty=c|u_\infty|^{q-2}u_\infty\quad\text{on }\partial M, L g ∘ u ∞ = 0 in M , B g ∘ u ∞ = c ∣ u ∞ ∣ q − 2 u ∞ on ∂ M ,
and each profile solves the critical boundary problem
− Δ V ( j ) = 0 in R + n , − ∂ y n V ( j ) = c ∣ V ( j ) ∣ q − 2 V ( j ) on ∂ R + n . -\Delta V^{(j)}=0\quad\text{in }\mathbb{R}^n_+,\qquad-\partial_{y_n}V^{(j)}=c|V^{(j)}|^{q-2}V^{(j)}\quad\text{on }\partial\mathbb{R}^n_+. − Δ V ( j ) = 0 in R + n , − ∂ y n V ( j ) = c ∣ V ( j ) ∣ q − 2 V ( j ) on ∂ R + n .
The boundary mass and energy split:
1 = ∫ ∂ M ∣ u ∞ ∣ q d σ g + ∑ j = 1 J m j + o ( 1 ) , (40) 1=\int_{\partial M}|u_\infty|^q\,d\sigma_g+\sum_{j=1}^{J}m_j+o(1), \tag*{(40)} 1 = ∫ ∂ M ∣ u ∞ ∣ q d σ g + j = 1 ∑ J m j + o ( 1 ) , (40)
c = N g ∘ ( u ∞ ) + ∑ j = 1 J E j + N g ∘ ( r k ) + o ( 1 ) , (41) c=\mathcal{N}_g^\circ(u_\infty)+\sum_{j=1}^{J}E_j+\mathcal{N}_g^\circ(r_k)+o(1), \tag*{(41)} c = N g ∘ ( u ∞ ) + j = 1 ∑ J E j + N g ∘ ( r k ) + o ( 1 ) , (41)
where m j : = ∫ ∂ R + n ∣ V ( j ) ∣ q d y ′ m_j:=\int_{\partial\mathbb{R}^n_+}|V^{(j)}|^q\,dy' m j := ∫ ∂ R + n ∣ V ( j ) ∣ q d y ′ and E j : = ∫ R + n ∣ ∇ V ( j ) ∣ 2 d y E_j:=\int_{\mathbb{R}^n_+}|\nabla V^{(j)}|^2\,dy E j := ∫ R + n ∣∇ V ( j ) ∣ 2 d y . Testing the profile equations by their respective solutions gives
N g ∘ ( u ∞ ) = c ∫ ∂ M ∣ u ∞ ∣ q d σ g , E j = c m j . \mathcal{N}_g^\circ(u_\infty)=c\int_{\partial M}|u_\infty|^q\,d\sigma_g,\qquad E_j=cm_j. N g ∘ ( u ∞ ) = c ∫ ∂ M ∣ u ∞ ∣ q d σ g , E j = c m j .
Each nonzero profile has
m j ≥ ( S ∗ c ) n − 1 , (42) m_j\geq\left(\frac{S_*}{c}\right)^{n-1}, \tag*{(42)} m j ≥ ( c S ∗ ) n − 1 , (42)
so only finitely many profiles can occur.
:::
Proof. Since C E s c ∗ ( M , g ) = inf Tr u ≠ 0 J g [ u ] C_{\mathrm{Esc}}^{*}(M,g)=\inf_{\operatorname{Tr}u\ne0}\mathcal{J}_{g}[u] C Esc ∗ ( M , g ) = inf Tr u = 0 J g [ u ] and J g [ u k ] → c \mathcal{J}_{g}[u_k]\to c J g [ u k ] → c , the level satisfies c ≥ C E s c ∗ ( M , g ) > 0 c\ge C_{\mathrm{Esc}}^{*}(M,g)>0 c ≥ C Esc ∗ ( M , g ) > 0 .
The profile extraction follows Struwe’s method [43 ] and the boundary extraction lemma [11 Lemma 2.4 and (2-11)] . For a related Struwe decomposition for polyharmonic operators on compact Riemannian manifolds, with or without boundary, see [38 ] . After normalizing the boundary coefficient to one, the limiting equation has the weak form
∫ R + n ∇ V ⋅ ∇ ϕ = ∫ ∂ R + n ∣ V ∣ q − 2 V ϕ , equivalently − ∂ y n V = ∣ V ∣ q − 2 V . \int_{\mathbb{R}^{n}_{+}}\nabla V\cdot\nabla\phi=\int_{\partial\mathbb{R}^{n}_{+}}|V|^{q-2}V\phi,\qquad\text{equivalently}\quad-\partial_{y_n}V=|V|^{q-2}V. ∫ R + n ∇ V ⋅ ∇ ϕ = ∫ ∂ R + n ∣ V ∣ q − 2 V ϕ , equivalently − ∂ y n V = ∣ V ∣ q − 2 V .
The extraction applies to sequences without a sign assumption and gives a nonzero profile. The corresponding flat decomposition is proved in [ZZZ25, Theorem 5.1].
(i) Profile extraction. For u ∈ S u\in S u ∈ S set ℓ u ( ϕ ) = ∫ ∂ M ∣ u ∣ q − 2 u ϕ \ell_u(\phi)=\int_{\partial M}|u|^{q-2}u\phi ℓ u ( ϕ ) = ∫ ∂ M ∣ u ∣ q − 2 u ϕ and ψ = ϕ − ℓ u ( ϕ ) u \psi=\phi-\ell_u(\phi)u ψ = ϕ − ℓ u ( ϕ ) u . Then ψ ∈ T u S \psi\in T_uS ψ ∈ T u S and, on energy-bounded sets, ∥ ψ ∥ E ≤ C ∥ ϕ ∥ E \|\psi\|_E\le C\|\phi\|_E ∥ ψ ∥ E ≤ C ∥ ϕ ∥ E . The exact identity
N g ∘ ( u , ϕ ) − J g [ u ] ℓ u ( ϕ ) = N g ∘ ( u , ψ ) = 1 2 D ( J g ∣ S ) ( u ) [ ψ ] (43) N_g^{\circ}(u,\phi)-\mathcal{J}_g[u]\ell_u(\phi)=N_g^{\circ}(u,\psi)=\frac{1}{2}D(\mathcal{J}_g|_S)(u)[\psi] \tag*{(43)} N g ∘ ( u , ϕ ) − J g [ u ] ℓ u ( ϕ ) = N g ∘ ( u , ψ ) = 2 1 D ( J g ∣ S ) ( u ) [ ψ ] (43)
therefore gives λ k = J g [ u k ] → c \lambda_k=\mathcal{J}_g[u_k]\to c λ k = J g [ u k ] → c and
N g ∘ ( u k , ϕ ) − λ k ∫ ∂ M ∣ u k ∣ q − 2 u k ϕ d σ g = o ( 1 ) ∥ ϕ ∥ H 1 ( M ) . N_g^{\circ}(u_k,\phi)-\lambda_k\int_{\partial M}|u_k|^{q-2}u_k\phi\,d\sigma_g=o(1)\|\phi\|_{H^1(M)}. N g ∘ ( u k , ϕ ) − λ k ∫ ∂ M ∣ u k ∣ q − 2 u k ϕ d σ g = o ( 1 ) ∥ ϕ ∥ H 1 ( M ) .
Since λ k → c \lambda_k\to c λ k → c , the sequence is Palais–Smale for
I c ( u ) = 1 2 N g ∘ ( u ) − c q ∫ ∂ M ∣ u ∣ q d σ g . \mathcal{I}_c(u)=\frac{1}{2}N_g^{\circ}(u)-\frac{c}{q}\int_{\partial M}|u|^q\,d\sigma_g. I c ( u ) = 2 1 N g ∘ ( u ) − q c ∫ ∂ M ∣ u ∣ q d σ g .
After weak convergence and almost-everywhere trace convergence, the limit u ∞ u_\infty u ∞ solves the stated equation. The absolute-value estimates in the proof of [11 Lemma 2.2] give the same derivative splitting for the nonlinearity ∣ u ∣ q − 2 u |u|^{q-2}u ∣ u ∣ q − 2 u . Hence u k ( 0 ) = u k − u ∞ u_k^{(0)}=u_k-u_\infty u k ( 0 ) = u k − u ∞ is Palais–Smale for I c \mathcal{I}_c I c . The scalar-curvature and boundary mean-curvature forms are compact from H 1 ( M ) H^1(M) H 1 ( M ) to its dual: for example,
∣ ∫ M Scal g u k ( 0 ) ϕ ∣ + ∣ ∫ ∂ M H g u k ( 0 ) ϕ ∣ ≤ C ( ∥ u k ( 0 ) ∥ L 2 ( M ) + ∥ u k ( 0 ) ∥ L 2 ( ∂ M ) ) ∥ ϕ ∥ H 1 = o ( 1 ) ∥ ϕ ∥ H 1 . \left|\int_M\operatorname{Scal}_g u_k^{(0)}\phi\right|+\left|\int_{\partial M}H_g u_k^{(0)}\phi\right|\le C\left(\|u_k^{(0)}\|_{L^2(M)}+\|u_k^{(0)}\|_{L^2(\partial M)}\right)\|\phi\|_{H^1}=o(1)\|\phi\|_{H^1}. ∫ M Scal g u k ( 0 ) ϕ + ∫ ∂ M H g u k ( 0 ) ϕ ≤ C ( ∥ u k ( 0 ) ∥ L 2 ( M ) + ∥ u k ( 0 ) ∥ L 2 ( ∂ M ) ) ∥ ϕ ∥ H 1 = o ( 1 ) ∥ ϕ ∥ H 1 .
Thus u k ( 0 ) = u k − u ∞ u_k^{(0)}=u_k-u_\infty u k ( 0 ) = u k − u ∞ converges weakly to zero and is a Palais–Smale sequence for
I c ( 0 ) ( u ) = 1 2 ∫ M ∣ ∇ u ∣ 2 d V g − c q ∫ ∂ M ∣ u ∣ q d σ g . \mathcal{I}_c^{(0)}(u)=\frac{1}{2}\int_M|\nabla u|^2\,dV_g-\frac{c}{q}\int_{\partial M}|u|^q\,d\sigma_g. I c ( 0 ) ( u ) = 2 1 ∫ M ∣∇ u ∣ 2 d V g − q c ∫ ∂ M ∣ u ∣ q d σ g .
Multiplication by c 1 / ( q − 2 ) c^{1/(q-2)} c 1/ ( q − 2 ) reduces the boundary coefficient to one. We may therefore apply [11 Lemma 2.4] successively, reversing this fixed rescaling after each extraction. The extraction preserves the Palais–Smale condition after subtraction and splits the energy. Each remainder, rescaled at a previously extracted center and scale, converges weakly to zero; comparison of rescalings gives (39) . The Dirichlet tails of each fixed profile permit a diagonal choice R k j → ∞ R_k^j\to\infty R k j → ∞ with ε k j R k j → 0 \varepsilon_k^jR_k^j\to0 ε k j R k j → 0 . The Brezis–Lieb lemma [5 Theorem 1] and energy orthogonality give (4.3)–(4.4); the compact lower-order forms vanish on the rescaled profiles. An interior profile would be a finite-energy harmonic function on R n \mathbb{R}^n R n and hence zero.
(ii) Energy estimate for each profile. Testing each profile equation against V ( j ) V^{(j)} V ( j ) gives E j = c m j E_j=cm_j E j = c m j . The sharp Sobolev trace inequality bounds ∫ R + n ∣ ∇ V ∣ 2 \int_{\mathbb{R}^n_+}|\nabla V|^2 ∫ R + n ∣∇ V ∣ 2 from below by S ∗ ( ∫ ∂ R + n ∣ V ∣ q ) 2 / q S_*\left(\int_{\partial\mathbb{R}^n_+}|V|^q\right)^{2/q} S ∗ ( ∫ ∂ R + n ∣ V ∣ q ) 2/ q for every V ∈ H ˙ 1 ( R + n ) V\in\dot H^1(\mathbb{R}^n_+) V ∈ H ˙ 1 ( R + n ) , including a sign-changing V V V . Thus E j ≥ S ∗ m j 2 / q E_j\ge S_*m_j^{2/q} E j ≥ S ∗ m j 2/ q . Together with E j = c m j E_j=cm_j E j = c m j and m j > 0 m_j>0 m j > 0 this gives (42) ,
m j ≥ ( S ∗ c ) n − 1 . m_j\ge\left(\frac{S_*}{c}\right)^{n-1}. m j ≥ ( c S ∗ ) n − 1 .
Each profile therefore has boundary mass at least ( S ∗ / c ) n − 1 > 0 (S_*/c)^{n-1}>0 ( S ∗ / c ) n − 1 > 0 , and hence J ≤ ( c / S ∗ ) n − 1 J\le(c/S_*)^{n-1} J ≤ ( c / S ∗ ) n − 1 . After the last extraction the concentration function of the remainder r k r_k r k tends to zero. The vanishing case gives ∥ r k ∥ L q ( ∂ M ) → 0 \|r_k\|_{L^q(\partial M)}\to0 ∥ r k ∥ L q ( ∂ M ) → 0 . The mass and energy identities, together with E j = c m j E_j=cm_j E j = c m j and the corresponding identity for u ∞ u_\infty u ∞ , then give N g ∘ ( r k ) → 0 N_g^{\circ}(r_k)\to0 N g ∘ ( r k ) → 0 . Coercivity also yields ∥ r k ∥ H 1 ( M ) → 0 \|r_k\|_{H^1(M)}\to0 ∥ r k ∥ H 1 ( M ) → 0 . □
Compactness below the hemisphere threshold
Corollary 4.2. (Concentration measure and bubble energy.) Under the hypotheses of Proposition 4.1, pass to limiting profile centers a j = lim k x k j a_j=\lim_k x_k^j a j = lim k x k j . The boundary measures satisfy
∣ u k ∣ q d σ g ⇀ ∗ ∣ u ∞ ∣ q d σ g + ∑ a ∈ A ν a δ a , ν a = ∑ j : a j = a m j , (44) |u_k|^q\,d\sigma_g \mathrel{\mathop{\rightharpoonup}\limits^{*}} |u_\infty|^q\,d\sigma_g+\sum_{a\in A}\nu_a\delta_a,\qquad\nu_a=\sum_{j:a_j=a}m_j, \tag*{(44)} ∣ u k ∣ q d σ g ⇀ ∗ ∣ u ∞ ∣ q d σ g + a ∈ A ∑ ν a δ a , ν a = j : a j = a ∑ m j , (44) where A A A is the finite set of limiting centers. Every atom satisfies
c ν a ≥ S ∗ ν a 2 / q , 0 < ν a ≤ 1. (45) c\nu_a\ge S_*\nu_a^{2/q},\qquad0<\nu_a\le1. \tag*{(45)} c ν a ≥ S ∗ ν a 2/ q , 0 < ν a ≤ 1. (45) In particular, c < S ∗ c<S_* c < S ∗ excludes all profiles and gives strong H 1 ( M ) H^1(M) H 1 ( M ) convergence.
Proof. The strong decomposition, parameter orthogonality, and Brezis–Lieb splitting imply (44) : each continuous weight tends to its value at a j a_j a j on the shrinking support of the j j j th profile. Since E j = c m j ≥ S ∗ m j 2 / q E_j=cm_j\ge S_*m_j^{2/q} E j = c m j ≥ S ∗ m j 2/ q and 2 / q < 1 2/q<1 2/ q < 1 ,
c ν a = ∑ j : a j = a E j ≥ S ∗ ∑ j : a j = a m j 2 / q ≥ S ∗ ν a 2 / q . c\nu_a=\sum_{j:a_j=a}E_j\ge S_*\sum_{j:a_j=a}m_j^{2/q}\ge S_*\nu_a^{2/q}. c ν a = j : a j = a ∑ E j ≥ S ∗ j : a j = a ∑ m j 2/ q ≥ S ∗ ν a 2/ q .
For c < S ∗ c<S_* c < S ∗ , the lower mass bound (42) exceeds one, so there are no profiles. This recovers the measure formulation of [33 Lemma 2.3, p. 61] . □
Theorem 4.3. (Palais–Smale compactness below the hemisphere threshold.) Assume ( Y θ + ) (Y_\theta^+) ( Y θ + ) and ( B G 2 + ) (\mathrm{BG}2^+) ( BG 2 + ) . Let q = 2 ∂ ∗ q=2_\partial^* q = 2 ∂ ∗ and let ( u k ) (u_k) ( u k ) be a Palais–Smale sequence for J g \mathcal{J}_g J g subject to
∥ u k ∥ L q ( ∂ M ) = 1 \|u_k\|_{L^q(\partial M)}=1 ∥ u k ∥ L q ( ∂ M ) = 1 at a level c < S ∗ = C E s c ∗ ( S + n ) c<S_*=C_{\mathrm{Esc}}^*(S_+^n) c < S ∗ = C Esc ∗ ( S + n ) . Then ( u k ) (u_k) ( u k ) is precompact in H 1 ( M ) H^1(M) H 1 ( M ) .
In particular, if C E s c ∗ ( M , g ) < S ∗ C_{\mathrm{Esc}}^*(M,g)<S_* C Esc ∗ ( M , g ) < S ∗ , then every Palais–Smale sequence at level C E s c ∗ ( M , g ) C_{\mathrm{Esc}}^*(M,g) C Esc ∗ ( M , g ) is precompact.
Proof. The preceding proof gives J = 0 J=0 J = 0 in (38) ; the remainder converges strongly to zero in H 1 H^1 H 1 , proving the assertion. □
Corollary 4.4. (Compactness of minimizing sequences below the hemisphere threshold.) Let ( u k ) (u_k) ( u k ) be a minimizing sequence for J g \mathcal{J}_g J g , so that
∥ u k ∥ L q ( ∂ M ) = 1 , Def g E s c ( u k ) → 0. \|u_k\|_{L^q(\partial M)}=1,\qquad\operatorname{Def}_{g}^{\mathrm{Esc}}(u_k)\to0. ∥ u k ∥ L q ( ∂ M ) = 1 , Def g Esc ( u k ) → 0. If C E s c ∗ ( M , g ) < S ∗ C_{\mathrm{Esc}}^*(M,g)<S_* C Esc ∗ ( M , g ) < S ∗ , then ( u k ) (u_k) ( u k ) is precompact in H 1 ( M ) H^1(M) H 1 ( M ) . Along a subsequence,
u k → u ∞ strongly in H 1 ( M ) , ∥ u ∞ ∥ L q ( ∂ M ) = 1 , u_k\to u_\infty\quad\text{strongly in }H^1(M),\qquad\|u_\infty\|_{L^q(\partial M)}=1, u k → u ∞ strongly in H 1 ( M ) , ∥ u ∞ ∥ L q ( ∂ M ) = 1 , and u ∞ u_\infty u ∞ is an extremal.
Proof. The boundary constraint is closed in H 1 ( M ) H^1(M) H 1 ( M ) and the quotient is bounded below there. Ekeland’s principle [16 Theorem 1] gives a Palais–Smale minimizing sequence ( u ~ k ) (\widetilde{u}_k) ( u k ) with ∥ u k − u ~ k ∥ H 1 → 0 \|u_k-\widetilde{u}_k\|_{H^1}\to0 ∥ u k − u k ∥ H 1 → 0 . Theorem 4.3 makes ( u ~ k ) (\widetilde{u}_k) ( u k ) , and hence ( u k ) (u_k) ( u k ) , precompact. Along a strongly convergent subsequence, continuity of the trace and energy gives ∥ u ∞ ∥ q = 1 \|u_\infty\|_q=1 ∥ u ∞ ∥ q = 1 and J g [ u ∞ ] = C E s c ∗ ( M , g ) \mathcal{J}_g[u_\infty]=C_{\mathrm{Esc}}^*(M,g) J g [ u ∞ ] = C Esc ∗ ( M , g ) . □
Lemma 4.5. (Mass bounds for half-space profiles.) Assume ( Y θ + ) (Y_\theta^+) ( Y θ + ) and ( B G 2 + ) (\mathrm{BG}2^+) ( BG 2 + ) , let q = 2 ∂ ∗ q=2_\partial^* q = 2 ∂ ∗ , and let ( u k ) ⊂ S (u_k)\subset\mathcal{S} ( u k ) ⊂ S be a constrained Palais–Smale sequence for J g \mathcal{J}_g J g at level c > 0 c>0 c > 0 , with profile decomposition (38) . Write
β ( c ) : = ( S ∗ c ) n − 1 , V ± ( j ) : = max ( ± V ( j ) , 0 ) , m j ± : = ∫ ∂ R + n ( V ± ( j ) ) q d y ′ . \beta(c):=\left(\frac{S_*}{c}\right)^{n-1},\qquad V_\pm^{(j)}:=\max(\pm V^{(j)},0),\qquad m_j^\pm:=\int_{\partial\mathbb{R}_+^n}\left(V_\pm^{(j)}\right)^q\,dy'. β ( c ) := ( c S ∗ ) n − 1 , V ± ( j ) := max ( ± V ( j ) , 0 ) , m j ± := ∫ ∂ R + n ( V ± ( j ) ) q d y ′ . (i) Mass of a profile. If V ( j ) V^{(j)} V ( j ) does not change sign, then ∣ V ( j ) ∣ |V^{(j)}| ∣ V ( j ) ∣ is a half-space optimizer up to translation and dilation, and
m j = β ( c ) . m_j=\beta(c). m j = β ( c ) .
If V ( j ) V^{(j)} V ( j ) changes sign, then
m j + > β ( c ) , m j − > β ( c ) , hence m j > 2 β ( c ) . m_j^+ > \beta(c), \qquad m_j^- > \beta(c), \qquad\text{hence} \qquad m_j > 2\beta(c). m j + > β ( c ) , m j − > β ( c ) , hence m j > 2 β ( c ) .
In particular no profile has boundary mass in the interval ( β ( c ) , 2 β ( c ) ] (\beta(c),2\beta(c)] ( β ( c ) , 2 β ( c )] . For the critical boundary equation on the half-space, Almaraz–Wang [14 Theorem 1.4] proved that the energy of every sign-changing solution is uniformly greater than twice the bubble energy.
(ii) Zero weak limit. Suppose u ∞ ≡ 0 u_\infty\equiv0 u ∞ ≡ 0 , and let N 1 N_1 N 1 and N ± N_\pm N ± be the numbers of profiles that do not change sign and that change sign, respectively. Then
( N 1 + 2 N ± ) β ( c ) ≤ 1 , (N_1+2N_\pm)\beta(c)\le1, ( N 1 + 2 N ± ) β ( c ) ≤ 1 ,
with equality if and only if N ± = 0 N_\pm=0 N ± = 0 , in which case c = N 1 1 / ( n − 1 ) S ∗ c=N_1^{1/(n-1)}S_* c = N 1 1/ ( n − 1 ) S ∗ and no profile changes sign, with boundary mass 1 / N 1 1/N_1 1/ N 1 for each profile.
(iii) Range of possible levels. Write Λ k : = k 1 / ( n − 1 ) S ∗ \Lambda_k:=k^{1/(n-1)}S_* Λ k := k 1/ ( n − 1 ) S ∗ . There is no Palais–Smale sequence with zero weak limit at any level
c ∈ ( S ∗ , Λ 2 ) , Λ 2 = 2 1 / ( n − 1 ) S ∗ . c\in(S_*,\Lambda_2), \qquad\Lambda_2=2^{1/(n-1)}S_*. c ∈ ( S ∗ , Λ 2 ) , Λ 2 = 2 1/ ( n − 1 ) S ∗ .
At the level c = Λ 2 c=\Lambda_2 c = Λ 2 , every such sequence has exactly two half-space bubble profiles, each of a fixed sign and boundary mass 1 / 2 1/2 1/2 ; if moreover
lim inf k → ∞ ∫ ∂ M ( u k + ) q d σ g > 0 and lim inf k → ∞ ∫ ∂ M ( u k − ) q d σ g > 0 , \liminf_{k\to\infty}\int_{\partial M}(u_k^+)^q\,d\sigma_g>0
\qquad\text{and}\qquad
\liminf_{k\to\infty}\int_{\partial M}(u_k^-)^q\,d\sigma_g>0, k → ∞ lim inf ∫ ∂ M ( u k + ) q d σ g > 0 and k → ∞ lim inf ∫ ∂ M ( u k − ) q d σ g > 0 ,
the two bubbles have opposite signs. A sign-changing profile requires c > Λ 2 c>\Lambda_2 c > Λ 2 ; among the levels defined above, Λ 3 \Lambda_3 Λ 3 is the first at which such a profile is not excluded by the mass bound.
Proof. Fix a profile V = V ( j ) V=V^{(j)} V = V ( j ) and put m ± = m j ± m^\pm=m_j^\pm m ± = m j ± .
(i) If V V V has a fixed sign, the strong maximum principle and Hopf lemma give V > 0 V>0 V > 0 or V < 0 V<0 V < 0 . The Li–Zhu classification [34 Theorem 1.2] identifies ∣ V ∣ \lvert V\rvert ∣ V ∣ as a half-space optimizer. Hence
c m j = E j = S ∗ m j 2 / q , m j = ( S ∗ / c ) n − 1 = β ( c ) . cm_j=E_j=S_*m_j^{2/q}, \qquad m_j=(S_*/c)^{n-1}=\beta(c). c m j = E j = S ∗ m j 2/ q , m j = ( S ∗ / c ) n − 1 = β ( c ) .
If V V V changes sign, testing its equation with V + V_+ V + and − V − -V_- − V − gives
∫ R + n ∣ ∇ V ± ∣ 2 = c m ± ≥ S ∗ ( m ± ) 2 / q , m ± ≥ β ( c ) . \int_{\mathbb{R}^n_+}\lvert\nabla V_\pm\rvert^2=cm^\pm\ge S_*(m^\pm)^{2/q}, \qquad m^\pm\ge\beta(c). ∫ R + n ∣ ∇ V ± ∣ 2 = c m ± ≥ S ∗ ( m ± ) 2/ q , m ± ≥ β ( c ) .
Equality for m + m^+ m + would make V + V_+ V + an optimizer with strictly positive boundary trace. Since V + V − = 0 V_+V_-=0 V + V − = 0 , this would force m − = 0 m^-=0 m − = 0 , a contradiction. Thus both inequalities are strict.
(ii) When u ∞ = 0 u_\infty=0 u ∞ = 0 , the exact mass splitting gives
1 = ∑ j m j ≥ ( N 1 + 2 N ± ) β ( c ) , 1=\sum_j m_j\ge(N_1+2N_\pm)\beta(c), 1 = j ∑ m j ≥ ( N 1 + 2 N ± ) β ( c ) ,
with equality precisely when N ± = 0 N_\pm=0 N ± = 0 . Then N 1 β ( c ) = 1 N_1\beta(c)=1 N 1 β ( c ) = 1 , so each mass is 1 / N 1 1/N_1 1/ N 1 and c = Λ N 1 c=\Lambda_{N_1} c = Λ N 1 .
(iii) For S ∗ < c < Λ 2 S_*<c<\Lambda_2 S ∗ < c < Λ 2 , one has 1 / 2 < β ( c ) < 1 1/2<\beta(c)<1 1/2 < β ( c ) < 1 . A sign-changing profile would exceed the total mass, while (ii) would require the integer N 1 = 1 / β ( c ) N_1=1/\beta(c) N 1 = 1/ β ( c ) to lie in ( 1 , 2 ) (1,2) ( 1 , 2 ) . At c = Λ 2 c=\Lambda_2 c = Λ 2 , one has β ( c ) = 1 / 2 \beta(c)=1/2 β ( c ) = 1/2 , so there are exactly two profiles, each of a fixed sign and mass $1/2. If both signed boundary masses have positive limits, these profiles must have opposite signs. Finally, m j > 2 β ( c ) m_j>2\beta(c) m j > 2 β ( c ) permits a sign-changing profile only when c > Λ 2 c>\Lambda_2 c > Λ 2 ; the first level Λ k \Lambda_k Λ k not excluded by this bound is Λ 3 \Lambda_3 Λ 3 .
Lemma 4.6 (Single-bubble blow-up at the threshold). Assume n ≥ 3 n\ge3 n ≥ 3 , ( Y + δ ) (Y_+^\delta) ( Y + δ ) , and ( B G 2 + ) (BG^{2+}) ( B G 2 + ) . Let ( u k ) ⊂ S (u_k)\subset\mathcal{S} ( u k ) ⊂ S be a Palais–Smale sequence at level S ∗ S_* S ∗ that is not precompact in H 1 ( M ) H^1(M) H 1 ( M ) . After passing to a subsequence, there are σ ∈ { ± 1 } \sigma\in\{\pm1\} σ ∈ { ± 1 } , centers x k ∈ ∂ M x_k\in\partial M x k ∈ ∂ M , and scales ϵ k ↓ 0 \epsilon_k\downarrow0 ϵ k ↓ 0 such that
u k = σ T x k , ϵ k ( χ R ( ϵ k ) U ∗ ) + r k , ∥ r k ∥ H 1 ( M ) → 0. (46) u_k=\sigma T_{x_k,\epsilon_k}(\chi_{R(\epsilon_k)}U_*)+r_k,\qquad\lVert r_k\rVert_{H^1(M)}\to0. \tag*{(46)} u k = σ T x k , ϵ k ( χ R ( ϵ k ) U ∗ ) + r k , ∥ r k ∥ H 1 ( M ) → 0. (46)
Here U ∗ U_* U ∗ is the boundary-normalized optimizer (11) . The boundary concentration is a single atom of full mass, and the weak limit is zero.
Proof. By (43) , take λ k = J g [ u k ] → S ∗ \lambda_k = J_g[u_k] \to S_* λ k = J g [ u k ] → S ∗ to obtain
N g ∘ ( u k , ϕ ) − λ k ∫ ∂ M ∣ u k ∣ q − 2 u k ϕ d σ g = o ( 1 ) ∥ ϕ ∥ H 1 . (47) \mathcal{N}_g^\circ(u_k,\phi)-\lambda_k\int_{\partial M}|u_k|^{q-2}u_k\phi\,d\sigma_g=o(1)\|\phi\|_{H^1}. \tag*{(47)} N g ∘ ( u k , ϕ ) − λ k ∫ ∂ M ∣ u k ∣ q − 2 u k ϕ d σ g = o ( 1 ) ∥ ϕ ∥ H 1 . (47)
Pass first to a subsequence having no strongly convergent subsequence, and apply Proposition 4.1 . Since the remainder converges strongly to zero, there must be at least one profile. At c = S ∗ c=S_* c = S ∗ , every profile has mass at least one, while the total boundary mass is one. Thus there is exactly one profile, with mass one, and the weak limit has zero trace. Its equation and ( Y γ + ) (Y_\gamma^+) ( Y γ + ) imply u ∞ = 0 u_\infty=0 u ∞ = 0 . Lemma 4.5 (i) identifies the profile as a signed translate and dilation of U ∗ U_* U ∗ .
Absorb its fixed translation and dilation into the center and scale. We may also replace the extraction cutoff by R ( ε ) = ε − 3 / 4 R(\varepsilon)=\varepsilon^{-3/4} R ( ε ) = ε − 3/4 . The squared Dirichlet norm of the truncation error is bounded by C R 2 − n CR^{2-n} C R 2 − n . For an arbitrary admissible extraction radius, its squared L 2 ( M ) L^2(M) L 2 ( M ) norm is bounded by
C ε 2 { R , n = 3 , 1 + log R , n = 4 , 1 , n ≥ 5 , C\varepsilon^2
\begin{cases}
R, & n=3,\\
1+\log R, & n=4,\\
1, & n\ge5,
\end{cases} C ε 2 ⎩ ⎨ ⎧ R , 1 + log R , 1 , n = 3 , n = 4 , n ≥ 5 ,
which tends to zero when R → ∞ R\to\infty R → ∞ and ε R → 0 \varepsilon R\to0 εR → 0 . The fixed parameter change also has vanishing rescaled coordinate error. These changes are therefore o ( 1 ) o(1) o ( 1 ) in H 1 ( M ) H^1(M) H 1 ( M ) and give (46) . □
Theorem 4.7 (The first mixed Palais–Smale level). Assume n ≥ 3 n\ge3 n ≥ 3 , ( Y γ + ) (Y_\gamma^+) ( Y γ + ) , and ( B G 2 + ) (BG^{2+}) ( B G 2 + ) , and put
Y = C E s c ∗ ( M , g ) , c m i x = ( S ∗ n − 1 + Y n − 1 ) 1 / ( n − 1 ) , Λ 2 = 2 1 / ( n − 1 ) S ∗ . (48) Y=C_{\mathrm{Esc}}^*(M,g),\qquad c_{\mathrm{mix}}=(S_*^{n-1}+Y^{n-1})^{1/(n-1)},\qquad\Lambda_2=2^{1/(n-1)}S_*. \tag*{(48)} Y = C Esc ∗ ( M , g ) , c mix = ( S ∗ n − 1 + Y n − 1 ) 1/ ( n − 1 ) , Λ 2 = 2 1/ ( n − 1 ) S ∗ . (48) Every constrained Palais–Smale sequence at a level S ∗ < c < c m i x S_*<c<c_{\mathrm{mix}} S ∗ < c < c mix is precompact in H 1 ( M ) H^1(M) H 1 ( M ) , without a sign assumption. The endpoint c m i x c_{\mathrm{mix}} c mix admits a noncompact Palais–Smale sequence.
More precisely, if S ∗ < c < Λ 2 S_*<c<\Lambda_2 S ∗ < c < Λ 2 , a noncompact constrained Palais–Smale sequence at c c c exists if and only if
d = ( c n − 1 − S ∗ n − 1 ) 1 / ( n − 1 ) (49) d=(c^{n-1}-S_*^{n-1})^{1/(n-1)} \tag*{(49)} d = ( c n − 1 − S ∗ n − 1 ) 1/ ( n − 1 ) (49) is a critical value of J g ∣ S J_g|_{\mathcal{S}} J g ∣ S .
Proof. We have 0 < Y ≤ S ∗ 0<Y\le S_* 0 < Y ≤ S ∗ , so S ∗ < c m i x ≤ Λ 2 S_*<c_{\mathrm{mix}}\le\Lambda_2 S ∗ < c mix ≤ Λ 2 . Apply Proposition 4.1 to a noncompact sequence at a level S ∗ < c < Λ 2 S_*<c<\Lambda_2 S ∗ < c < Λ 2 , passing to a subsequence having no strongly convergent subsequence. Put β = ( S ∗ / c ) n − 1 ∈ ( 1 / 2 , 1 ) \beta=(S_*/c)^{n-1}\in(1/2,1) β = ( S ∗ / c ) n − 1 ∈ ( 1/2 , 1 ) . Lemma 4.5 excludes sign-changing profiles, and each fixed-sign profile has mass β \beta β . Thus exactly one profile occurs. The weak limit u ∞ u_\infty u ∞ has boundary mass 1 − β > 0 1-\beta>0 1 − β > 0 and satisfies the equation with multiplier c c c .
Write u ∞ = a v u_\infty=av u ∞ = a v , where a = ( 1 − β ) 1 / q a=(1-\beta)^{1/q} a = ( 1 − β ) 1/ q and v ∈ S v\in\mathcal{S} v ∈ S . Then v v v is a constrained critical point with value
J g [ v ] = c a q − 2 = c ( 1 − β ) 1 / ( n − 1 ) = d . J_g[v]=ca^{q-2}=c(1-\beta)^{1/(n-1)}=d. J g [ v ] = c a q − 2 = c ( 1 − β ) 1/ ( n − 1 ) = d .
Since d ≥ Y d\ge Y d ≥ Y , this implies c ≥ c m i x c\ge c_{\mathrm{mix}} c ≥ c mix and proves compactness in the stated interval. It also proves the necessity of (49) .
Conversely, let v ∈ S v\in\mathcal{S} v ∈ S be critical with value 0 < d < S ∗ 0<d<S_* 0 < d < S ∗ . Set
c = ( d n − 1 + S ∗ n − 1 ) 1 / ( n − 1 ) , a 0 = ( d / c ) ( n − 2 ) / 2 , a 1 = ( S ∗ / c ) ( n − 2 ) / 2 . c=(d^{n-1}+S_*^{n-1})^{1/(n-1)},\qquad a_0=(d/c)^{(n-2)/2},\qquad a_1=(S_*/c)^{(n-2)/2}. c = ( d n − 1 + S ∗ n − 1 ) 1/ ( n − 1 ) , a 0 = ( d / c ) ( n − 2 ) /2 , a 1 = ( S ∗ / c ) ( n − 2 ) /2 .
Fix x ∈ ∂ M x\in\partial M x ∈ ∂ M and put
z ε = a 0 v + a 1 u x , ε , u ε = z ε ∥ z ε ∥ q . z_\varepsilon=a_0v+a_1u_{x,\varepsilon},\qquad u_\varepsilon=\frac{z_\varepsilon}{\|z_\varepsilon\|_q}. z ε = a 0 v + a 1 u x , ε , u ε = ∥ z ε ∥ q z ε .
The bubble converges weakly to zero in H 1 H^1 H 1 , tends to zero almost everywhere, and has energy tending to S ∗ S_* S ∗ . The Brezis–Lieb identity and energy polarization give
∥ z ε ∥ q q = a 0 q + a 1 q + o ( 1 ) = 1 + o ( 1 ) , N g ∘ ( z ε ) = a 0 2 d + a 1 2 S ∗ + o ( 1 ) = c + o ( 1 ) . \|z_\varepsilon\|_q^q=a_0^q+a_1^q+o(1)=1+o(1),\qquad\mathcal{N}_g^\circ(z_\varepsilon)=a_0^2d+a_1^2S_*+o(1)=c+o(1). ∥ z ε ∥ q q = a 0 q + a 1 q + o ( 1 ) = 1 + o ( 1 ) , N g ∘ ( z ε ) = a 0 2 d + a 1 2 S ∗ + o ( 1 ) = c + o ( 1 ) .
For the derivative, critical rescaling bounds the principal metric error by C ε R C\varepsilon R C εR and the Dirichlet cutoff error by C R − ( n − 2 ) / 2 CR^{-(n-2)/2} C R − ( n − 2 ) /2 . Both tend to zero for our cutoff, in every dimension n ≥ 3 n \ge3 n ≥ 3 . The flat bubble equation therefore gives
∥ N g ∘ ( U x , ε , ⋅ ) − S ∗ ( ∫ ∂ M U x , ε q − 1 ( ⋅ ) d σ g ) ∥ H E − 1 ⟶ 0. \left\|\mathcal{N}^{\circ}_{g}(U_{x,\varepsilon},\cdot)-S_{*}\left(\int_{\partial M}U_{x,\varepsilon}^{q-1}(\cdot)\,d\sigma_{g}\right)\right\|_{H^{-1}_{E}}\longrightarrow0. N g ∘ ( U x , ε , ⋅ ) − S ∗ ( ∫ ∂ M U x , ε q − 1 ( ⋅ ) d σ g ) H E − 1 ⟶ 0.
The scalar-curvature and mean-curvature terms tend to zero by compactness of the L 2 L^{2} L 2 embeddings. Nonlinear Brezis–Lieb splitting, as in the proof of Proposition 4.1 , gives
∥ ∣ z ε ∣ q − 2 z ε − a 0 q − 1 ∣ v ∣ q − 2 v − a 1 q − 1 U x , ε q − 1 ∥ q / ( q − 1 ) ⟶ 0. \left\||z_{\varepsilon}|^{q-2}z_{\varepsilon}-a_{0}^{q-1}|v|^{q-2}v-a_{1}^{q-1}U_{x,\varepsilon}^{q-1}\right\|_{q/(q-1)}\longrightarrow0. ∣ z ε ∣ q − 2 z ε − a 0 q − 1 ∣ v ∣ q − 2 v − a 1 q − 1 U x , ε q − 1 q / ( q − 1 ) ⟶ 0.
The identities a 0 q d = c a 0 q − 1 a_{0}^{q}d=ca_{0}^{q-1} a 0 q d = c a 0 q − 1 and a 1 S ∗ = c a 1 q − 1 a_{1}S_{*}=ca_{1}^{q-1} a 1 S ∗ = c a 1 q − 1 now show that u ε u_{\varepsilon} u ε is Palais–Smale at c c c . Its nonzero concentrating mass a 1 q a_{1}^{q} a 1 q prevents strong convergence. This proves sufficiency.
If Y < S ∗ Y<S_{*} Y < S ∗ , Corollary 4.4 supplies a minimizer of value d = Y d=Y d = Y , and the preceding construction proves noncompactness at c m i x c_{\mathrm{mix}} c mix . If Y = S ∗ Y=S_{*} Y = S ∗ , take two bubbles at distinct fixed boundary points, each with amplitude 2 − 1 / q 2^{-1/q} 2 − 1/ q , and normalize their sum. The same calculation gives a noncompact Palais–Smale sequence at Λ 2 = c m i x \Lambda_{2}=c_{\mathrm{mix}} Λ 2 = c mix . □
Lyapunov–Schmidt reduction and the strict inequality
The complementary correction lowers the quadratic coefficient of the standard bubble. We identify it through the flat linearized equation and evaluate the resulting quadratic form exactly. The resulting coefficient is negative at every non-umbilic point in all dimensions n ≥ 5 n\ge5 n ≥ 5 .
The leading term of the constrained gradient
Write Sym 0 2 ( R n − 1 ) \operatorname{Sym}^{2}_{0}(\mathbb{R}^{n-1}) Sym 0 2 ( R n − 1 ) for the trace-free symmetric matrices. For S S S in this space, put ∣ S ∣ 2 = tr ( S 2 ) |S|^{2}=\operatorname{tr}(S^{2}) ∣ S ∣ 2 = tr ( S 2 ) and
r = ∣ y ′ ∣ , s = r 2 + ( 1 + t ) 2 , Q S ( y ′ ) = S a b y a y b . r=|y'|,\qquad s=r^{2}+(1+t)^{2},\qquad Q_{S}(y')=S_{ab}y_{a}y_{b}. r = ∣ y ′ ∣ , s = r 2 + ( 1 + t ) 2 , Q S ( y ′ ) = S ab y a y b .
Tangential spherical harmonics decompose functions into angular degrees. Radial functions have degree 0, translation fields have degree 1, and Q S ( y ′ ) h ( r , t ) Q_{S}(y')h(r,t) Q S ( y ′ ) h ( r , t ) has degree 2. The flat dilation and translation Jacobi fields are
Z 0 ♭ = − n − 2 2 U + − y ⋅ ∇ U + , Z a ♭ = − ∂ a U + . Z_{0}^{\flat}=-\frac{n-2}{2}U_{+}-y\cdot\nabla U_{+},\qquad Z_{a}^{\flat}=-\partial_{a}U_{+}. Z 0 ♭ = − 2 n − 2 U + − y ⋅ ∇ U + , Z a ♭ = − ∂ a U + .
Define the flat constrained complement and its Jacobi form by
X ♭ = { ϕ ∈ D 1 , 2 : ∫ ∂ R + n U + q − 1 ϕ = 0 , ∫ R + n ∇ ϕ ⋅ ∇ Z j ♭ = 0 ( 0 ≤ j ≤ n − 1 ) } , X^{\flat}=\left\{\phi\in D^{1,2}:\int_{\partial\mathbb{R}^{n}_{+}}U_{+}^{q-1}\phi=0,\quad\int_{\mathbb{R}^{n}_{+}}\nabla\phi\cdot\nabla Z_{j}^{\flat}=0\ (0\le j\le n-1)\right\}, X ♭ = { ϕ ∈ D 1 , 2 : ∫ ∂ R + n U + q − 1 ϕ = 0 , ∫ R + n ∇ ϕ ⋅ ∇ Z j ♭ = 0 ( 0 ≤ j ≤ n − 1 ) } ,
B n ( ϕ , ψ ) = ∫ R + n ∇ ϕ ⋅ ∇ ψ − ( q − 1 ) λ f l a t ∫ ∂ R + n U + q − 2 ϕ ψ . B_{n}(\phi,\psi)=\int_{\mathbb{R}^{n}_{+}}\nabla\phi\cdot\nabla\psi-(q-1)\lambda_{\mathrm{flat}}\int_{\partial\mathbb{R}^{n}_{+}}U_{+}^{q-2}\phi\psi. B n ( ϕ , ψ ) = ∫ R + n ∇ ϕ ⋅ ∇ ψ − ( q − 1 ) λ flat ∫ ∂ R + n U + q − 2 ϕ ψ .
Let X 2 X_{2} X 2 be the degree-2 subspace, and X 2 , S X_{2,S} X 2 , S its closed subspace generated by Q S h ( r , t ) Q_{S}h(r,t) Q S h ( r , t ) for a fixed S S S . Angular orthogonality places both in X ♭ X^{\flat} X ♭ . We use the cutoff bound ω ( R ) = R − ( n − 2 ) / 2 \omega(R)=R^{-(n-2)/2} ω ( R ) = R − ( n − 2 ) /2 and the metric remainder bound
η n ( ε ) = { ε 13 / 8 , n = 5 , ε 2 1 + log ( 1 / ε ) , n = 6 , ε 2 , n ≥ 7. \eta_{n}(\varepsilon)=
\begin{cases}
\varepsilon^{13/8}, & n=5,\\
\varepsilon^{2}\sqrt{1+\log(1/\varepsilon)}, & n=6,\\
\varepsilon^{2}, & n\ge7.
\end{cases} η n ( ε ) = ⎩ ⎨ ⎧ ε 13/8 , ε 2 1 + log ( 1/ ε ) , ε 2 , n = 5 , n = 6 , n ≥ 7.
With the choice R ( ε ) = ε − 3 / 4 R(\varepsilon)=\varepsilon^{-3/4} R ( ε ) = ε − 3/4 , η n ( ε ) + ω ( R ( ε ) ) = o ( ε ) \eta_{n}(\varepsilon)+\omega(R(\varepsilon))=o(\varepsilon) η n ( ε ) + ω ( R ( ε )) = o ( ε ) for n ≥ 5 n\ge5 n ≥ 5 and is O ( ε 2 ) O(\varepsilon^{2}) O ( ε 2 ) for n ≥ 8 n\ge8 n ≥ 8 .
Lemma 5.1 (Constrained gradient of a bubble: degree-2 term orthogonal to the kernel). Assume n ≥ 5 n\ge5 n ≥ 5 , ( Y ∂ + ) (Y^{+}_{\partial}) ( Y ∂ + ) , and ( B G 3 + ) (\mathrm{BG}^{3+}) ( BG 3 + ) . Use the normalized bubble of Definition 2.3 with the cutoff R ( ε ) = ε − 3 / 4 R(\varepsilon)=\varepsilon^{-3/4} R ( ε ) = ε − 3/4 , and write t ( x , ε ) = ∥ U + , x , ε ∥ q − 1 t(x,\varepsilon)=\|U_{+,x,\varepsilon}\|^{-1}_{q} t ( x , ε ) = ∥ U + , x , ε ∥ q − 1 . Its normalized parameter derivatives are
Z 0 = ε ∂ ε U x , ε , Z α = ε ∂ x α U x , ε ( 1 ≤ α ≤ n − 1 ) . Z_{0}=\varepsilon\partial_{\varepsilon}U_{x,\varepsilon},\qquad Z_{\alpha}=\varepsilon\partial_{x_{\alpha}}U_{x,\varepsilon}\quad(1\le\alpha\le n-1). Z 0 = ε ∂ ε U x , ε , Z α = ε ∂ x α U x , ε ( 1 ≤ α ≤ n − 1 ) .
Set K x , ε = span { Z 0 , … , Z n − 1 } ⊂ T U x , ε S \mathcal{K}_{x,\varepsilon}=\operatorname{span}\{Z_0,\ldots,Z_{n-1}\}\subset T_{U_{x,\varepsilon}}\mathcal{S} K x , ε = span { Z 0 , … , Z n − 1 } ⊂ T U x , ε S , and let Π x , ε \Pi_{x,\varepsilon} Π x , ε be the energy-orthogonal projection onto K x , ε ⊥ ∩ T U x , ε S \mathcal{K}_{x,\varepsilon}^{\perp}\cap T_{U_{x,\varepsilon}}\mathcal{S} K x , ε ⊥ ∩ T U x , ε S .
Then, uniformly for x ∈ ∂ M x\in\partial M x ∈ ∂ M as ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 , there exist coefficients a β ( x , ε ) ∈ R a_{\beta}(x,\varepsilon)\in\mathbb{R} a β ( x , ε ) ∈ R (β = 0 , … , n − 1 \beta=0,\ldots,n-1 β = 0 , … , n − 1 ) such that
grad S J g ( U x , ε ) = ∑ β = 0 n − 1 a β ( x , ε ) Z β ( x , ε ) + R x , ε , R x , ε = Π x , ε grad S J g ( U x , ε ) ∈ K x , ε ⊥ ∩ T U x , ε S , (50) \operatorname{grad}_{\mathcal{S}}J_g(U_{x,\varepsilon})=\sum_{\beta=0}^{n-1}a_{\beta}(x,\varepsilon)Z_{\beta}(x,\varepsilon)+R_{x,\varepsilon},\qquad R_{x,\varepsilon}=\Pi_{x,\varepsilon}\operatorname{grad}_{\mathcal{S}}J_g(U_{x,\varepsilon})\in\mathcal{K}_{x,\varepsilon}^{\perp}\cap T_{U_{x,\varepsilon}}\mathcal{S}, \tag*{(50)} grad S J g ( U x , ε ) = β = 0 ∑ n − 1 a β ( x , ε ) Z β ( x , ε ) + R x , ε , R x , ε = Π x , ε grad S J g ( U x , ε ) ∈ K x , ε ⊥ ∩ T U x , ε S , (50)
and
∥ R x , ε ∥ H 1 ( M ) ≤ C ( ε ∣ H g ( x ) ∣ + ε ∣ I I ˚ ( x ) ∣ + ε 2 + ω ( R ( ε ) ) ) , (51) \lVert R_{x,\varepsilon}\rVert_{H^1(M)}\leq C\left(\varepsilon|H_g(x)|+\varepsilon|\mathring{\mathrm{II}}(x)|+\varepsilon^2+\omega(R(\varepsilon))\right), \tag*{(51)} ∥ R x , ε ∥ H 1 ( M ) ≤ C ( ε ∣ H g ( x ) ∣ + ε ∣ II ˚ ( x ) ∣ + ε 2 + ω ( R ( ε )) ) , (51)
with ω \omega ω as above, and ∣ a β ( x , ε ) ∣ ≤ C ε |a_{\beta}(x,\varepsilon)|\leq C\varepsilon ∣ a β ( x , ε ) ∣ ≤ C ε . In particular, the constrained gradient is O ( ε ) O(\varepsilon) O ( ε ) ; when H g ( x ) = 0 H_g(x)=0 H g ( x ) = 0 , the first term in the bound disappears. At an umbilic point with zero mean curvature the bound is O ( ε 2 + ω ( R ( ε ) ) ) O(\varepsilon^2+\omega(R(\varepsilon))) O ( ε 2 + ω ( R ( ε ))) , which is O ( ε 2 ) O(\varepsilon^2) O ( ε 2 ) for n ≥ 8 n\geq8 n ≥ 8 and o ( ε ) o(\varepsilon) o ( ε ) for n = 5 , 6 , 7 n=5,6,7 n = 5 , 6 , 7 .
Proof. The normalized scale and center fields have a uniformly invertible Gram matrix by flat nondegeneracy and Fermi rescaling. Their energy projection gives (50) . Indeed, with G α β = ⟨ Z α , Z β ⟩ E G_{\alpha\beta}=\langle Z_{\alpha},Z_{\beta}\rangle_E G α β = ⟨ Z α , Z β ⟩ E ,
a β = ∑ α ( G − 1 ) β α ⟨ grad S J g ( U ) , Z α ⟩ E , ∣ a β ∣ ≤ C ε a_{\beta}=\sum_{\alpha}(G^{-1})_{\beta\alpha}\langle\operatorname{grad}_{\mathcal{S}}J_g(U),Z_{\alpha}\rangle_E,\qquad|a_{\beta}|\leq C\varepsilon a β = α ∑ ( G − 1 ) β α ⟨ grad S J g ( U ) , Z α ⟩ E , ∣ a β ∣ ≤ C ε
by the bounded normalized frame, its inverse Gram matrix, and Lemma 3.7 . It remains to estimate the differential on the complement.
The flat equation cancels the leading term. The first-order Fermi and Robin corrections give
D J g [ U x , ε ] ( ψ ) = ε H g ( x ) ℓ 0 ( ψ ) + ε I I ˚ a b ( x ) ℓ a b ( 2 ) ( ψ ) + O ( ( η n ( ε ) + ω ( R ( ε ) ) ) ∥ ψ ∥ E ) . (52) D J_g[U_{x,\varepsilon}](\psi)=\varepsilon H_g(x)\ell_0(\psi)+\varepsilon\mathring{\mathrm{II}}_{ab}(x)\ell^{(2)}_{ab}(\psi)+O\left((\eta_n(\varepsilon)+\omega(R(\varepsilon)))\lVert\psi\rVert_E\right). \tag*{(52)} D J g [ U x , ε ] ( ψ ) = ε H g ( x ) ℓ 0 ( ψ ) + ε II ˚ ab ( x ) ℓ ab ( 2 ) ( ψ ) + O ( ( η n ( ε ) + ω ( R ( ε ))) ∥ ψ ∥ E ) . (52)
Here ℓ 0 \ell_0 ℓ 0 is the radial functional from the trace of I I \mathrm{II} II and the boundary mean curvature; the trace-free contraction is the degree-2 functional generated by t I I ˚ a b ∂ a b U + t\mathring{\mathrm{II}}_{ab}\partial_{ab}U_+ t II ˚ ab ∂ ab U + . These functionals are bounded because ∣ y ∣ ∇ U + ∈ L 2 |y|\nabla U_+\in L^2 ∣ y ∣∇ U + ∈ L 2 for n ≥ 5 n\geq5 n ≥ 5 . The metric remainder follows from
ε 2 ∥ ∣ y ∣ 2 ∇ U + ∥ L 2 ( B 2 R + ) 2 ≤ C ε 2 { R 1 / 2 , n = 5 , ( 1 + log R ) 1 / 2 , n = 6 , 1 , n ≥ 7. \varepsilon^2\lVert|y|^2\nabla U_+\rVert_{L^2(B^+_{2R})}^2\leq C\varepsilon^2
\begin{cases}
R^{1/2},&n=5,\\
(1+\log R)^{1/2},&n=6,\\
1,&n\geq7.
\end{cases} ε 2 ∥ ∣ y ∣ 2 ∇ U + ∥ L 2 ( B 2 R + ) 2 ≤ C ε 2 ⎩ ⎨ ⎧ R 1/2 , ( 1 + log R ) 1/2 , 1 , n = 5 , n = 6 , n ≥ 7.
with the scalar and trace terms estimated as in Lemma 3.7 . For R = ε − 3 / 4 R=\varepsilon^{-3/4} R = ε − 3/4 this is C η n ( ε ) ≤ C ( ε 2 + ω ( R ) ) C\eta_n(\varepsilon)\leq C(\varepsilon^2+\omega(R)) C η n ( ε ) ≤ C ( ε 2 + ω ( R )) .
The radial source vanishes when H g ( x ) = 0 H_g(x)=0 H g ( x ) = 0 . The trace-free source is orthogonal to the dilation and translation fields, since
∫ S n − 2 I I ˚ a b ω a ω b d ω = 0 , ∫ S n − 2 I I ˚ a b ω a ω b ω c d ω = 0. \int_{\mathbb{S}^{n-2}}\mathring{\mathrm{II}}_{ab}\omega^a\omega^b\,d\omega=0,\qquad
\int_{\mathbb{S}^{n-2}}\mathring{\mathrm{II}}_{ab}\omega^a\omega^b\omega^c\,d\omega=0. ∫ S n − 2 II ˚ ab ω a ω b d ω = 0 , ∫ S n − 2 II ˚ ab ω a ω b ω c d ω = 0.
Consequently, on the complement,
∣ D J g [ U x , ε ] ( ψ ) ∣ ≤ C ( ε ∣ H g ( x ) ∣ + ε ∣ I I ˚ ( x ) ∣ + ε 2 + ω ( R ( ε ) ) ) ∥ ψ ∥ E . (53) |D J_g[U_{x,\varepsilon}](\psi)|\leq C\left(\varepsilon|H_g(x)|+\varepsilon|\mathring{\mathrm{II}}(x)|+\varepsilon^2+\omega(R(\varepsilon))\right)\lVert\psi\rVert_E. \tag*{(53)} ∣ D J g [ U x , ε ] ( ψ ) ∣ ≤ C ( ε ∣ H g ( x ) ∣ + ε ∣ II ˚ ( x ) ∣ + ε 2 + ω ( R ( ε )) ) ∥ ψ ∥ E . (53)
Riesz representation and energy–H 1 H^1 H 1 equivalence give (51) . Setting H g ( x ) = 0 H_g(x)=0 H g ( x ) = 0 and then I I ˚ ( x ) = 0 \mathring{\mathrm{II}}(x)=0 II ˚ ( x ) = 0 gives the stated refinements. □
Lemma 5.2 (Leading degree-2 term of the projected gradient). Assume n ≥ 5 n\geq5 n ≥ 5 and retain X b X^b X b and X 2 X_2 X 2 from the beginning of this subsection. For S ∈ Sym 0 2 ( R n − 1 ) S\in\operatorname{Sym}_0^2(\mathbb{R}^{n-1}) S ∈ Sym 0 2 ( R n − 1 ) , define
L S ( ϕ ) : = − 2 ∫ R + n y n S a b ∂ a U + ∂ b ϕ d y = 2 ∫ R + n y n S a b ∂ a b 2 U + ϕ d y , (54) L_S(\phi):=-2\int_{\mathbb{R}^n_+}y_nS^{ab}\partial_aU_+\partial_b\phi\,dy=2\int_{\mathbb{R}^n_+}y_nS^{ab}\partial^2_{ab}U_+\phi\,dy, \tag*{(54)} L S ( ϕ ) := − 2 ∫ R + n y n S ab ∂ a U + ∂ b ϕ d y = 2 ∫ R + n y n S ab ∂ ab 2 U + ϕ d y , (54) where the second identity is tangential integration by parts (y n = 0 y_n=0 y n = 0 on ∂ R + n \partial\mathbb{R}^n_+ ∂ R + n ). Let E S b ∈ X b \mathcal{E}^b_S\in X^b E S b ∈ X b be the H ˙ 1 \dot H^1 H ˙ 1 -Riesz representative of L S L_S L S .
In the setting of Lemma 5.1 , assume H g ( x ) = 0 H_g(x)=0 H g ( x ) = 0 . Set X x , ε : = K x , ε ⊥ E ∩ T U x , ε S X_{x,\varepsilon}:=K_{x,\varepsilon}^{\perp E}\cap T U_{x,\varepsilon}S X x , ε := K x , ε ⊥ E ∩ T U x , ε S , and let L S x , ε ∈ X x , ε ∗ \mathcal{L}_{S}^{x,\varepsilon}\in X_{x,\varepsilon}^{*} L S x , ε ∈ X x , ε ∗ denote the covector obtained from L S \mathcal{L}_{S} L S by the rescaled Fermi pullback, localized with the same cutoff convention as the boundary bubble. Define the restricted differential R x , ε : = D ( J g ∘ ret U x , ε ) ( 0 ) ∣ X x , ε \mathcal{R}_{x,\varepsilon}:=D(\mathcal{J}_g\circ\operatorname{ret}_{U_{x,\varepsilon}})(0)|_{X_{x,\varepsilon}} R x , ε := D ( J g ∘ ret U x , ε ) ( 0 ) ∣ X x , ε . Then, in the norm dual to the energy norm on X x , ε X_{x,\varepsilon} X x , ε ,
∥ R x , ε + 2 ε t ( x , ε ) L I I ˚ ( x ) x , ε ∥ X x , ε ∗ ≤ C ( η n ( ε ) + ω ( R ( ε ) ) ) , (55) \left\|\mathcal{R}_{x,\varepsilon}+2\varepsilon t(x,\varepsilon)\mathcal{L}_{\mathrm{\mathring{II}}(x)}^{x,\varepsilon}\right\|_{X_{x,\varepsilon}^{*}}\leq C\bigl(\eta_n(\varepsilon)+\omega(R(\varepsilon))\bigr), \tag*{(55)} R x , ε + 2 εt ( x , ε ) L II ˚ ( x ) x , ε X x , ε ∗ ≤ C ( η n ( ε ) + ω ( R ( ε )) ) , (55)
where η n \eta_n η n and ω \omega ω are defined above. The error is o ( ε ) o(\varepsilon) o ( ε ) for n ≥ 5 n\geq5 n ≥ 5 and O ( ε 2 ) O(\varepsilon^2) O ( ε 2 ) for n ≥ 8 n\geq8 n ≥ 8 . The first variation of the flat cutoff bubble is radial, so its degree-2 projection is zero. The leading source is therefore − 2 ε t ∗ L I I ˚ ( x ) x , ε -2\varepsilon t_*\mathcal{L}_{\mathrm{\mathring{II}}(x)}^{x,\varepsilon} − 2 ε t ∗ L II ˚ ( x ) x , ε for J g \mathcal{J}_g J g , or − 2 ε t ∗ − 1 L I I ˚ ( x ) x , ε -2\varepsilon t_*^{-1}\mathcal{L}_{\mathrm{\mathring{II}}(x)}^{x,\varepsilon} − 2 ε t ∗ − 1 L II ˚ ( x ) x , ε for J g / S ∗ \mathcal{J}_g/S_* J g / S ∗ . The map S ↦ E S b S\mapsto\mathcal{E}_S^b S ↦ E S b is O ( n − 1 ) O(n-1) O ( n − 1 ) -equivariant and injective on Sym 0 2 ( R n − 1 ) \operatorname{Sym}_0^2(\mathbb{R}^{n-1}) Sym 0 2 ( R n − 1 ) .
Proof. At the unnormalized bubble, differentiating the first geometric energy correction gives
4 ε ∫ t I I ˚ a b ( x ) ∂ a U + ∂ b ϕ = − 2 ε L I I ˚ ( x ) ( ϕ ) . 4\varepsilon\int t\mathring{\mathrm{II}}_{ab}(x)\partial_aU_+\partial_b\phi=-2\varepsilon\mathcal{L}_{\mathring{\mathrm{II}}(x)}(\phi). 4 ε ∫ t II ˚ ab ( x ) ∂ a U + ∂ b ϕ = − 2 ε L II ˚ ( x ) ( ϕ ) .
Under the unit boundary L q L^q L q -norm constraint the first variation is linear in the base function. At U x , ε = t ( x , ε ) U + , x , ε U_{x,\varepsilon}=t(x,\varepsilon)U_{+,x,\varepsilon} U x , ε = t ( x , ε ) U + , x , ε the source is therefore multiplied by t ( x , ε ) t(x,\varepsilon) t ( x , ε ) . The degree-2 term is unchanged by flat projection. The separate error bounds are
second metric jet: C ε 2 ℓ 2 , n ( R ) , flat cutoff: C R − α , tail of the first-order source: C ε R − ( n − 4 ) / 2 , projection: C ( ε 2 + ε R − α ) . \begin{aligned}
\text{second metric jet:}&\qquad C\varepsilon^2\ell_{2,n}(R),\\
\text{flat cutoff:}&\qquad CR^{-\alpha},\\
\text{tail of the first-order source:}&\qquad C\varepsilon R^{-(n-4)/2},\\
\text{projection:}&\qquad C(\varepsilon^2+\varepsilon R^{-\alpha}).
\end{aligned} second metric jet: flat cutoff: tail of the first-order source: projection: C ε 2 ℓ 2 , n ( R ) , C R − α , C ε R − ( n − 4 ) /2 , C ( ε 2 + ε R − α ) .
where ℓ 2 , 5 ( R ) = R 1 / 2 \ell_{2,5}(R)=R^{1/2} ℓ 2 , 5 ( R ) = R 1/2 , ℓ 2 , 6 ( R ) = log ( 2 R ) \ell_{2,6}(R)=\sqrt{\log(2R)} ℓ 2 , 6 ( R ) = log ( 2 R ) , and ℓ 2 , n ( R ) = 1 \ell_{2,n}(R)=1 ℓ 2 , n ( R ) = 1 for n ≥ 7 n\geq7 n ≥ 7 . The first two estimates follow from the weighted moment in Lemma 5.1 and (37) . The source tail is bounded by C R − α CR^{-\alpha} C R − α because ε R ≤ 1 \varepsilon R\leq1 εR ≤ 1 . For the projection, the normalized frame differs from the critically rescaled flat frame by O E ( ε + R − α ) O_E(\varepsilon+R^{-\alpha}) O E ( ε + R − α ) . This follows from the Fermi expansion, the explicit profile derivatives, and ∂ x d g ˉ ( x , z ′ ) 2 = − 2 exp − 1 ( z ′ ) \partial_{x_d}\bar g(x,z')^2=-2\exp^{-1}(z') ∂ x d g ˉ ( x , z ′ ) 2 = − 2 exp − 1 ( z ′ ) . The constraint normal is handled by the exact Riesz identity in Remark 5.5 . Angular orthogonality removes the flat moments, and the inverse Gram formula (57) gives the stated projection error. All four terms are bounded by the right-hand side of (55) .
The boundary normalization satisfies t ( x , ε ) = t ∗ + O ( ε 2 ) t(x,\varepsilon)=t_*+O(\varepsilon^2) t ( x , ε ) = t ∗ + O ( ε 2 ) by [41 Lemma 3.8] , which gives the stated leading coefficient.
Rotational symmetry gives equivariance. The explicit identity
S a b ∂ a b U + = n ( n − 2 ) A n Q S s − ( n + 2 ) / 2 S_{ab}\partial_{ab}U_+=n(n-2)A_nQ_Ss^{-(n+2)/2} S ab ∂ ab U + = n ( n − 2 ) A n Q S s − ( n + 2 ) /2
shows that L S \mathcal{L}_S L S is nonzero on degree-2 tests whenever S ≠ 0 S\neq0 S = 0 . Hence its Riesz representative is nonzero, proving injectivity. □
Remark 5.3 (The geometric source). For F x , ε ( v ) = J g ( ret U x , ε v ) F_{x,\varepsilon}(v)=\mathcal{J}_g(\operatorname{ret}_{U_{x,\varepsilon}}v) F x , ε ( v ) = J g ( ret U x , ε v ) on X x , ε X_{x,\varepsilon} X x , ε , the complementary source is the covector D X F x , ε ( 0 ) ∈ X x , ε ∗ D_XF_{x,\varepsilon}(0)\in X_{x,\varepsilon}^{*} D X F x , ε ( 0 ) ∈ X x , ε ∗ . In fixed model coordinates its leading coefficient in a minimal boundary region is
S 1 , x = − 2 t ∗ L I I ˚ ( x ) , D X F x , ε ( 0 ) = ε S 1 , x + o ( ε ) . \mathcal{S}_{1,x}=-2t_*\mathcal{L}_{\mathring{\mathrm{II}}(x)},\qquad D_XF_{x,\varepsilon}(0)=\varepsilon\mathcal{S}_{1,x}+o(\varepsilon). S 1 , x = − 2 t ∗ L II ˚ ( x ) , D X F x , ε ( 0 ) = ε S 1 , x + o ( ε ) .
For n ≥ 7 n\geq7 n ≥ 7 , at an umbilic center with the refined cutoff, the corresponding expansion starts with ε 2 S 2 , x \varepsilon^2\mathcal{S}_{2,x} ε 2 S 2 , x ; see (118) . These leading covectors are the geometric sources. The full complementary source also includes cutoff and higher-order terms. Its energy Riesz representative is a vector, defined by pairing with the energy inner product. For J g / S ∗ \mathcal{J}_g/S_* J g / S ∗ , all source covectors are divided by S ∗ S_* S ∗ .
Proposition 5.4 (Lyapunov–Schmidt correction for the degree-2 source). Assume n ≥ 5 n\geq5 n ≥ 5 . Let B n \mathcal{B}_n B n be the flat Jacobi form on the constrained complement X b X^b X b and let L S \mathcal{L}_S L S be the degree-2 source functional from
(5.5). Define δ κ 3 L S ( n ) \delta\kappa^{\mathrm{LS}}_3(n) δ κ 3 LS ( n ) by the first equality in (56) with ∣ S ∣ = 1 |S|=1 ∣ S ∣ = 1 . Its value is independent of the choice of unit S S S , is strictly positive, and for every S ∈ Sym 0 2 ( R n − 1 ) S \in\operatorname{Sym}^{2}_{0}(\mathbb{R}^{n-1}) S ∈ Sym 0 2 ( R n − 1 ) one has
δ κ 3 L S ( n ) ∣ S ∣ 2 = sup ϕ ∈ X b { 2 L S ( ϕ ) − B n ( ϕ , ϕ ) } = sup ϕ ∈ X b ∖ { 0 } L S ( ϕ ) 2 B n ( ϕ , ϕ ) . (56) \delta\kappa^{\mathrm{LS}}_3(n)|S|^2=\sup_{\phi\in X^{\mathrm{b}}}\left\{2\mathcal{L}_S(\phi)-B_n(\phi,\phi)\right\}=\sup_{\phi\in X^{\mathrm{b}}\setminus\{0\}}\frac{\mathcal{L}_S(\phi)^2}{B_n(\phi,\phi)}. \tag*{(56)} δ κ 3 LS ( n ) ∣ S ∣ 2 = ϕ ∈ X b sup { 2 L S ( ϕ ) − B n ( ϕ , ϕ ) } = ϕ ∈ X b ∖ { 0 } sup B n ( ϕ , ϕ ) L S ( ϕ ) 2 . (56)
This is the Lyapunov–Schmidt correction entering the reduced decomposition (72) . Its exact value is computed in Proposition 5.15 .
Proof. The trace inequality makes B n B_n B n continuous on X b X^{\mathrm{b}} X b , and the flat spectral gap from Lemma 3.2 gives B n ( ϕ , ϕ ) ≥ γ ∗ ∥ ϕ ∥ H ˙ 1 2 B_n(\phi,\phi)\geq\gamma_* \|\phi\|_{\dot{H}^{1}}^2 B n ( ϕ , ϕ ) ≥ γ ∗ ∥ ϕ ∥ H ˙ 1 2 . Moreover,
∣ L S ( ϕ ) ∣ ≤ 2 ∣ S ∣ ∥ y n ∇ tan U + ∥ L 2 ∥ ϕ ∥ H ˙ 1 ≤ C n ∣ S ∣ ∥ ϕ ∥ H ˙ 1 , |\mathcal{L}_S(\phi)|\leq2|S|\|y_n\nabla_{\tan}U_+\|_{L^2}\|\phi\|_{\dot{H}^{1}}\leq C_n|S|\|\phi\|_{\dot{H}^{1}}, ∣ L S ( ϕ ) ∣ ≤ 2∣ S ∣∥ y n ∇ t a n U + ∥ L 2 ∥ ϕ ∥ H ˙ 1 ≤ C n ∣ S ∣∥ ϕ ∥ H ˙ 1 ,
since y n ∇ tan U + ∈ L 2 ( R + n ) y_n\nabla_{\tan}U_+\in L^2(\mathbb{R}^n_+) y n ∇ t a n U + ∈ L 2 ( R + n ) for n ≥ 5 n\geq5 n ≥ 5 . Thus Lax–Milgram gives a unique w S ∈ X b w_S\in X^{\mathrm{b}} w S ∈ X b satisfying B n ( w S , ψ ) = L S ( ψ ) B_n(w_S,\psi)=\mathcal{L}_S(\psi) B n ( w S , ψ ) = L S ( ψ ) for every ψ ∈ X b \psi\in X^{\mathrm{b}} ψ ∈ X b . Completing the square yields
2 L S ( ϕ ) − B n ( ϕ , ϕ ) = B n ( w S , w S ) − B n ( ϕ − w S , ϕ − w S ) . 2\mathcal{L}_S(\phi)-B_n(\phi,\phi)=B_n(w_S,w_S)-B_n(\phi-w_S,\phi-w_S). 2 L S ( ϕ ) − B n ( ϕ , ϕ ) = B n ( w S , w S ) − B n ( ϕ − w S , ϕ − w S ) .
Hence the first supremum is attained uniquely at w S w_S w S and equals B n ( w S , w S ) B_n(w_S,w_S) B n ( w S , w S ) . For ϕ ≠ 0 \phi\neq0 ϕ = 0 , maximizing 2 c L S ( ϕ ) − c 2 B n ( ϕ , ϕ ) 2c\mathcal{L}_S(\phi)-c^2B_n(\phi,\phi) 2 c L S ( ϕ ) − c 2 B n ( ϕ , ϕ ) over c ∈ R c\in\mathbb{R} c ∈ R gives c = L S ( ϕ ) / B n ( ϕ , ϕ ) c=\mathcal{L}_S(\phi)/B_n(\phi,\phi) c = L S ( ϕ ) / B n ( ϕ , ϕ ) and maximum L S ( ϕ ) 2 / B n ( ϕ , ϕ ) \mathcal{L}_S(\phi)^2/B_n(\phi,\phi) L S ( ϕ ) 2 / B n ( ϕ , ϕ ) . Taking the supremum over these lines proves the second equality in (56) .
The map S ↦ w S S\mapsto w_S S ↦ w S is linear and O ( n − 1 ) O(n-1) O ( n − 1 ) -equivariant, since X b X^{\mathrm{b}} X b and B n B_n B n are invariant under tangential rotations. Therefore S ↦ B n ( w S , w S ) S\mapsto B_n(w_S,w_S) S ↦ B n ( w S , w S ) is an orthogonally invariant quadratic form on trace-free symmetric matrices. Diagonalizing S S S and permuting its eigenvalues shows that it is a linear combination of tr ( S 2 ) \operatorname{tr}(S^2) tr ( S 2 ) and ( tr S ) 2 (\operatorname{tr}S)^2 ( tr S ) 2 . The latter is zero, so the form equals δ κ 3 L S ( n ) ∣ S ∣ 2 \delta\kappa^{\mathrm{LS}}_3(n)|S|^2 δ κ 3 LS ( n ) ∣ S ∣ 2 . Lemma 5.2 gives L S ≠ 0 \mathcal{L}_S\neq0 L S = 0 for S ≠ 0 S\neq0 S = 0 , proving δ κ 3 L S ( n ) > 0 \delta\kappa^{\mathrm{LS}}_3(n)>0 δ κ 3 LS ( n ) > 0 ; when S = 0 S=0 S = 0 , w S = 0 w_S=0 w S = 0 and both suprema are zero.
The coefficients of B n B_n B n depend on y ′ y' y ′ only through ∣ y ′ ∣ |y'| ∣ y ′ ∣ . Thus distinct tangential spherical harmonics are B n B_n B n -orthogonal. Since L S \mathcal{L}_S L S pairs only with the harmonic S a b y a y b S_{ab}y_a y_b S ab y a y b , uniqueness places w S w_S w S in the closed subspace generated by S a b y a y b h ( ∣ y ′ ∣ , y n ) S_{ab}y_a y_b h(|y'|,y_n) S ab y a y b h ( ∣ y ′ ∣ , y n ) . For S ≠ 0 S\neq0 S = 0 , both suprema can therefore be restricted to this subspace.
To identify the normalization in the geometric expansion, put S = I I ˚ ( x ) S=\mathring{\mathrm{II}}(x) S = II ˚ ( x ) in a boundary region where H g ≡ 0 H_g\equiv0 H g ≡ 0 and rescale ϕ ∈ X b \phi\in X^{\mathrm{b}} ϕ ∈ X b in Fermi coordinates, multiply it by the base bubble factor t ( x , ε ) t(x,\varepsilon) t ( x , ε ) , and project it onto X x , ε X_{x,\varepsilon} X x , ε ; denote this direction by ϕ x , ε \phi_{x,\varepsilon} ϕ x , ε . Lemma 5.2 gives the leading covector − 2 ε t ∗ − 1 L S -2\varepsilon t_*^{-1}\mathcal{L}_S − 2 ε t ∗ − 1 L S for J g / S ∗ J_g/S_* J g / S ∗ at t ∗ U + t_*U_+ t ∗ U + , while the limiting quadratic form is t ∗ 2 B n t_*^2B_n t ∗ 2 B n , where t ∗ 2 = S ∗ t_*^2=S_* t ∗ 2 = S ∗ . The factor t ∗ t_* t ∗ in this rescaled direction cancels both normalization factors. Consequently, for fixed c ∈ R c\in\mathbb{R} c ∈ R ,
J g ( U x , ε + c ε ϕ x , ε ) S ∗ = 1 + ε 2 ( κ 3 b a r e ( n ) ∣ S ∣ 2 − 2 c L S ( ϕ ) + c 2 B n ( ϕ , ϕ ) ) + o ( ε 2 ) . \frac{J_g(U_{x,\varepsilon}+c\varepsilon\phi_{x,\varepsilon})}{S_*}=1+\varepsilon^2\left(\kappa^{\mathrm{bare}}_3(n)|S|^2-2c\mathcal{L}_S(\phi)+c^2B_n(\phi,\phi)\right)+o(\varepsilon^2). S ∗ J g ( U x , ε + c ε ϕ x , ε ) = 1 + ε 2 ( κ 3 bare ( n ) ∣ S ∣ 2 − 2 c L S ( ϕ ) + c 2 B n ( ϕ , ϕ ) ) + o ( ε 2 ) .
The preceding minimization lowers the quadratic coefficient by δ κ 3 L S ( n ) ∣ S ∣ 2 \delta\kappa^{\mathrm{LS}}_3(n)|S|^2 δ κ 3 LS ( n ) ∣ S ∣ 2 . The passage from this flat minimum to the parameter-dependent Lyapunov–Schmidt correction is established in Lemma 5.11 and the proof of (72) below.
Construction of the correction
The construction uses the Lyapunov–Schmidt scheme [[AMA06, Chapter 2]; a boundary Yamabe application with a geometric bubble correction is given in [27 Section 3] .
Differentiating this projection formula uses D ( G − 1 ) = − G − 1 ( D G ) G − 1 D(G^{-1})=-G^{-1}(DG)G^{-1} D ( G − 1 ) = − G − 1 ( D G ) G − 1 . The constraint normal also has the required flat limit. For a boundary-normalized bubble U U U , the quotient identity gives
N g ∘ ( U , ϕ ) = J g [ U ] ∫ ∂ M U q − 1 ϕ d σ g + 1 2 D J g [ U ] [ ϕ ] . \mathcal{N}^{\circ}_{g}(U,\phi)=\mathcal{J}_{g}[U]\int_{\partial M}U^{q-1}\phi\,d\sigma_{g}+\frac{1}{2}D\mathcal{J}_{g}[U][\phi]. N g ∘ ( U , ϕ ) = J g [ U ] ∫ ∂ M U q − 1 ϕ d σ g + 2 1 D J g [ U ] [ ϕ ] .
Thus the energy Riesz representative of the covector ϕ ↦ q ∫ U q − 1 ϕ \phi\mapsto q\int U^{q-1}\phi ϕ ↦ q ∫ U q − 1 ϕ is
q J g [ U ] U − q 2 J g [ U ] R E D J g [ U ] . \frac{q}{\mathcal{J}_{g}[U]}U-\frac{q}{2\mathcal{J}_{g}[U]}R_{E}D\mathcal{J}_{g}[U]. J g [ U ] q U − 2 J g [ U ] q R E D J g [ U ] .
where R E R_{E} R E is the energy Riesz map. The first-variation bound makes its second term O ( ε ) O(\varepsilon) O ( ε ) , or O ( ε 2 ) O(\varepsilon^{2}) O ( ε 2 ) at an umbilic point with the improved cutoff. Thus the normal converges in the global energy norm; the corresponding estimate for the parameter derivatives of the first variation gives the differentiated convergence.
Lemma 5.7 (Uniform C 1 C^{1} C 1 Lyapunov–Schmidt inversion). In the local coordinates of Remark 5.6, put
F p ( w ) = D w [ J g ( U p + I p w ) ] , R p = F p ( 0 ) , L p = D w F p ( 0 ) . F_{\mathbf{p}}(w)=D_{w}\left[\mathcal{J}_{g}(U_{\mathbf{p} }+\mathcal{I}_{\mathbf{p}}w)\right],\qquad R_{\mathbf{p}}=F_{\mathbf{p}}(0),\qquad L_{\mathbf{p}}=D_{w}F_{\mathbf{p}}(0). F p ( w ) = D w [ J g ( U p + I p w ) ] , R p = F p ( 0 ) , L p = D w F p ( 0 ) . Suppose F p F_{\mathbf{p}} F p is C 1 C^{1} C 1 jointly in ( p , w ) (\mathbf{p},w) ( p , w ) , its w w w derivative is uniformly continuous on a fixed ball, and ∥ L p − 1 ∥ ≤ C L \lVert L_{\mathbf{p}}^{-1}\rVert\leq C_{L} ∥ L p − 1 ∥ ≤ C L . For sufficiently small ∥ R p ∥ \lVert R_{\mathbf{p}}\rVert ∥ R p ∥ , the equation F p ( w ) = 0 F_{\mathbf{p}}(w)=0 F p ( w ) = 0 has a unique solution in a fixed small ball, with
∥ w p ∥ ≤ 2 C L ∥ R p ∥ . (58) \lVert w_{\mathbf{p}}\rVert\leq2C_{L}\lVert R_{\mathbf{p}}\rVert. \tag*{(58)} ∥ w p ∥ ≤ 2 C L ∥ R p ∥ . (58) The solution is C 1 C^{1} C 1 in p \mathbf{p} p and
D p w p = − ( D w F p ( w p ) ) − 1 D p F p ( w p ) . D_{\mathbf{p}}w_{\mathbf{p}}=-\left(D_{w}F_{\mathbf{p}}(w_{\mathbf{p}})\right)^{-1}D_{\mathbf{p}}F_{\mathbf{p}}(w_{\mathbf{p}}). D p w p = − ( D w F p ( w p ) ) − 1 D p F p ( w p ) . For the local coordinate representations of the quotient, the required uniform continuity follows from the C 2 C^{2} C 2 boundary trace functional, since q > 2 q>2 q > 2 .
Proof. Choose r 0 r_{0} r 0 so that ∥ L p − 1 ( D w F p ( w ) − L p ) ∥ ≤ 1 / 2 \lVert L_{\mathbf{p}}^{-1}(D_{w}F_{\mathbf{p}}(w)-L_{\mathbf{p}})\rVert\leq1/2 ∥ L p − 1 ( D w F p ( w ) − L p )∥ ≤ 1/2 for ∥ w ∥ ≤ r 0 \lVert w\rVert\leq r_{0} ∥ w ∥ ≤ r 0 . If 2 C L ∥ R p ∥ ≤ r 0 2C_{L}\lVert R_{\mathbf{p}}\rVert\leq r_{0} 2 C L ∥ R p ∥ ≤ r 0 , the map w ↦ w − L p − 1 F p ( w ) w\mapsto w-L_{\mathbf{p}}^{-1}F_{\mathbf{p}}(w) w ↦ w − L p − 1 F p ( w ) is a contraction of that ball into itself. Its fixed point satisfies (58) ; the parameter-dependent implicit-function theorem gives the derivative formula. □
Lemma 5.8 (Expansion of the Lyapunov–Schmidt functional). Fix an admissible parameter p \mathbf{p} p for a family of normalized approximate functions U p U_{\mathbf{p}} U p . Let T p : X 0 → X p \mathcal{T}_{\mathbf{p}}:X_{0}\to X_{\mathbf{p}} T p : X 0 → X p identify a fixed Hilbert space X 0 X_{0} X 0 with each complementary space X p X_{\mathbf{p}} X p . Assume that the local coordinate representations of the quotient below have uniform C 2 , γ C^{2,\gamma} C 2 , γ bounds on a fixed small ball, where 0 < γ ≤ 1 0<\gamma\leq1 0 < γ ≤ 1 (for n ≥ 5 n\geq5 n ≥ 5 , γ = q − 2 \gamma=q-2 γ = q − 2 ). In the one-bubble case, use T p = I p \mathcal{T}_{\mathbf{p}}=\mathcal{I}_{\mathbf{p}} T p = I p from Remark 5.6. Write
H p ( w ~ ) = J g ( ret U p ( T p w ~ ) ) , S p = D w ~ H p ( 0 ) , L p = D w ~ 2 H p ( 0 ) . \mathcal{H}_{\mathbf{p}}(\widetilde{w})=\mathcal{J}_{g}\bigl(\operatorname{ret}_{U_{\mathbf{p}}}(\mathcal{T}_{\mathbf{p}}\widetilde{w})\bigr),\qquad S_{\mathbf{p}}=D_{\widetilde{w}}\mathcal{H}_{\mathbf{p}}(0),\qquad\mathcal{L}_{\mathbf{p}}=D_{\widetilde{w}}^{2}\mathcal{H}_{\mathbf{p}}(0). H p ( w ) = J g ( ret U p ( T p w ) ) , S p = D w H p ( 0 ) , L p = D w 2 H p ( 0 ) . Assume L p − 1 \mathcal{L}_{\mathbf{p}}^{-1} L p − 1 is uniformly bounded. If
D w ~ H p ( w ~ p ) = 0 , ∥ w ~ p ∥ ≤ C ∥ S p ∥ , D_{\widetilde{w}}\mathcal{H}_{\mathbf{p}}(\widetilde{w}_{\mathbf{p}})=0,\qquad\lVert\widetilde{w}_{\mathbf{p}}\rVert\leq C\lVert S_{\mathbf{p}}\rVert, D w H p ( w p ) = 0 , ∥ w p ∥ ≤ C ∥ S p ∥ ,
then
w ~ p = − L p − 1 S p + O ( ∥ S p ∥ 1 + γ ) , (59) \widetilde{w}_p=-\mathcal{L}_p^{-1}S_p+O(\lVert S_p\rVert^{1+\gamma}), \tag*{(59)} w p = − L p − 1 S p + O (∥ S p ∥ 1 + γ ) , (59)
and
H p ( w ~ p ) = H p ( 0 ) − 1 2 ⟨ L p − 1 S p , S p ⟩ + O ( ∥ S p ∥ 2 + γ ) . (60) \mathcal{H}_p(\widetilde{w}_p)=\mathcal{H}_p(0)-\frac{1}{2}\langle\mathcal{L}_p^{-1}S_p,S_p\rangle+O(\lVert S_p\rVert^{2+\gamma}). \tag*{(60)} H p ( w p ) = H p ( 0 ) − 2 1 ⟨ L p − 1 S p , S p ⟩ + O (∥ S p ∥ 2 + γ ) . (60)
Proof. The inequality ∣ ∣ a ∣ γ − ∣ b ∣ γ ∣ ≤ ∣ a − b ∣ γ ||a|^\gamma-|b|^\gamma|\leq|a-b|^\gamma ∣∣ a ∣ γ − ∣ b ∣ γ ∣ ≤ ∣ a − b ∣ γ and the trace embedding give, on a fixed small ball,
∥ D 2 H p ( w ~ 1 ) − D 2 H p ( w ~ 0 ) ∥ ≤ C ∥ w ~ 1 − w ~ 0 ∥ γ . \lVert D^2\mathcal{H}_p(\widetilde{w}_1)-D^2\mathcal{H}_p(\widetilde{w}_0)\rVert\leq C\lVert\widetilde{w}_1-\widetilde{w}_0\rVert^\gamma. ∥ D 2 H p ( w 1 ) − D 2 H p ( w 0 )∥ ≤ C ∥ w 1 − w 0 ∥ γ .
Hence
D H p ( w ~ ) = S p + L p w ~ + O ( ∥ w ~ ∥ 1 + γ ) , D\mathcal{H}_p(\widetilde{w})=S_p+\mathcal{L}_p\widetilde{w}+O(\lVert\widetilde{w}\rVert^{1+\gamma}), D H p ( w ) = S p + L p w + O (∥ w ∥ 1 + γ ) ,
H p ( w ~ ) = H p ( 0 ) + ⟨ S p , w ~ ⟩ + 1 2 ⟨ L p w ~ , w ~ ⟩ + O ( ∥ w ~ ∥ 2 + γ ) . \mathcal{H}_p(\widetilde{w})=\mathcal{H}_p(0)+\langle S_p,\widetilde{w}\rangle+\frac{1}{2}\langle\mathcal{L}_p\widetilde{w},\widetilde{w}\rangle+O(\lVert\widetilde{w}\rVert^{2+\gamma}). H p ( w ) = H p ( 0 ) + ⟨ S p , w ⟩ + 2 1 ⟨ L p w , w ⟩ + O (∥ w ∥ 2 + γ ) .
Applying the first identity at w ~ p \widetilde{w}_p w p gives (59) . Substitution into the second identity gives (60) . □
Lemma 5.9 (Lyapunov–Schmidt reduction and estimates for the correction). Assume n ≥ 5 n\geq5 n ≥ 5 , ( Y ∂ + ) (Y_\partial^+) ( Y ∂ + ) , and ( B G 3 + ) (BG^{3+}) ( B G 3 + ) . For the normalized bubble U = U x , ε \mathcal{U}=\mathcal{U}_{x,\varepsilon} U = U x , ε and complement X = X x , ε X=X_{x,\varepsilon} X = X x , ε , there is a unique small solution w x , ε ∈ X w_{x,\varepsilon}\in X w x , ε ∈ X , with ∥ w x , ε ∥ ≤ c 1 ε \lVert w_{x,\varepsilon}\rVert\leq c_1\varepsilon ∥ w x , ε ∥ ≤ c 1 ε , of
D ( J g ∘ ret U ) [ w x , ε ] [ ϕ ] = 0 ( ϕ ∈ X ) . (61) D(\mathcal{J}_g\circ\operatorname{ret}_{\mathcal{U}})[w_{x,\varepsilon}][\phi]=0\qquad(\phi\in X). \tag*{(61)} D ( J g ∘ ret U ) [ w x , ε ] [ ϕ ] = 0 ( ϕ ∈ X ) . (61) It satisfies
∥ w x , ε ∥ H 1 ( M ) ≤ C ( M , g ) ε . (62) \lVert w_{x,\varepsilon}\rVert_{H^1(M)}\leq C(M,g)\varepsilon. \tag*{(62)} ∥ w x , ε ∥ H 1 ( M ) ≤ C ( M , g ) ε . (62) If additionally
H g ( x ) = 0 , (63) H_g(x)=0, \tag*{(63)} H g ( x ) = 0 , (63) then
∥ w x , ε ∥ H 1 ( M ) ≤ C ( M , g ) ( ε ∣ I I ˚ ( x ) ∣ + ε 2 + ω ( R ( ε ) ) ) . (64) \lVert w_{x,\varepsilon}\rVert_{H^1(M)}\leq C(M,g)\bigl(\varepsilon|\mathring{\mathrm{II}}(x)|+\varepsilon^2+\omega(R(\varepsilon))\bigr). \tag*{(64)} ∥ w x , ε ∥ H 1 ( M ) ≤ C ( M , g ) ( ε ∣ II ˚ ( x ) ∣ + ε 2 + ω ( R ( ε )) ) . (64) At an umbilic center this is O ( ε 2 ) O(\varepsilon^2) O ( ε 2 ) for n ≥ 8 n\geq8 n ≥ 8 and o ( ε ) o(\varepsilon) o ( ε ) for n = 5 , 6 , 7 n=5,6,7 n = 5 , 6 , 7 . The Hessian D 2 ( J g ∘ ret U ) ( 0 ) D^2(\mathcal{J}_g\circ\operatorname{ret}_{\mathcal{U}})(0) D 2 ( J g ∘ ret U ) ( 0 ) is uniformly coercive on X X X and its operator has uniformly bounded inverse. The solution is C ∞ C^\infty C ∞ in ( x , ε ) (x,\varepsilon) ( x , ε ) for positive scales, using Remark 5.6. On a boundary region with H g ≡ 0 H_g\equiv0 H g ≡ 0 , the reduced functional has the second-order expansion of Proposition 5.12 .
Proof. The boundary coordinate expansion gives
∥ U + , x , ε ∥ q q = ∥ U + ∥ q q + O ( ε 2 + R − ( n − 1 ) ) . \lVert U_{+,x,\varepsilon}\rVert_q^q=\lVert U_+\rVert_q^q+O(\varepsilon^2+R^{-(n-1)}). ∥ U + , x , ε ∥ q q = ∥ U + ∥ q q + O ( ε 2 + R − ( n − 1 ) ) .
The weighted trace moment is finite for n ≥ 5 n\geq5 n ≥ 5 , and the cutoff R ( ε ) = ε − 3 / 4 R(\varepsilon)=\varepsilon^{-3/4} R ( ε ) = ε − 3/4 has R − ( n − 1 ) = O ( ε 2 ) R^{-(n-1)}=O(\varepsilon^2) R − ( n − 1 ) = O ( ε 2 ) . Hence
t ( x , ε ) = t ∗ + O ( ε 2 ) , 0 < c ≤ t ( x , ε ) ≤ C . (65) t(x,\varepsilon)=t_*+O(\varepsilon^2),\qquad0<c\leq t(x,\varepsilon)\leq C. \tag*{(65)} t ( x , ε ) = t ∗ + O ( ε 2 ) , 0 < c ≤ t ( x , ε ) ≤ C . (65)
The supports lie in the fixed collar because their radius is 2 ε 1 / 4 2\varepsilon^{1/4} 2 ε 1/4 . Homogeneity gives J g ( ret U h ) = J g ( U + h ) \mathcal{J}_g(\operatorname{ret}_{\mathcal{U}}h)=\mathcal{J}_g(\mathcal{U}+h) J g ( ret U h ) = J g ( U + h ) . For tangent h , ϕ h,\phi h , ϕ , direct differentiation gives, also at an approximate bubble,
D 2 ( J g ∘ ret U ) ( 0 ) [ h , ϕ ] = 2 N g 0 ( h , ϕ ) − 2 ( q − 1 ) J g [ U q − 2 h ϕ ] . (66) D^2(\mathcal{J}_g\circ\operatorname{ret}_{\mathcal{U}})(0)[h,\phi]=2\mathcal{N}_g^0(h,\phi)-2(q-1)\mathcal{J}_g[\mathcal{U}^{q-2}h\phi]. \tag*{(66)} D 2 ( J g ∘ ret U ) ( 0 ) [ h , ϕ ] = 2 N g 0 ( h , ϕ ) − 2 ( q − 1 ) J g [ U q − 2 h ϕ ] . (66)
To transfer Theorem 3.6 to the normalized complement, put D 0 = ε ∂ ε D_0=\varepsilon\partial_\varepsilon D 0 = ε ∂ ε and D a = ε ∂ x a D_a=\varepsilon\partial_{x_a} D a = ε ∂ x a for 1 ≤ a ≤ n − 1 1\leq a\leq n-1 1 ≤ a ≤ n − 1 , and write U 0 = U + , x , ε U_0=U_{+,x,\varepsilon} U 0 = U + , x , ε and Z j = D j ( t U 0 ) Z_j=D_j(tU_0) Z j = D j ( t U 0 ) for the weighted parameter fields. If h ∈ X h\in X h ∈ X , then
⟨ h , D j U 0 ⟩ E = − ( D j log t ) ⟨ h , U 0 ⟩ E , ∣ ⟨ h , U 0 ⟩ E ∣ ≤ C ε ∥ h ∥ E . \langle h,D_jU_0\rangle_E=-(D_j\log t)\langle h,U_0\rangle_E,\qquad|\langle h,U_0\rangle_E|\leq C\varepsilon\lVert h\rVert_E. ⟨ h , D j U 0 ⟩ E = − ( D j log t ) ⟨ h , U 0 ⟩ E , ∣ ⟨ h , U 0 ⟩ E ∣ ≤ C ε ∥ h ∥ E .
The second estimate follows from tangency and Lemma 3.7 . The weighted derivatives of t t t are bounded. Use the full frame consisting of the common constraint normal and the Jacobi fields of U 0 U_0 U 0 . The energy projection determined by its uniformly invertible Gram matrix maps h h h into the complement associated with U 0 U_0 U 0 with error at most C ε ∥ h ∥ E C\varepsilon\|h\|_E C ε ∥ h ∥ E , preserving tangency. Since
D 2 J g [ t U 0 ] = t − 2 D 2 J g [ U 0 ] , D^2\mathcal{J}_g[tU_0]=t^{-2}D^2\mathcal{J}_g[U_0], D 2 J g [ t U 0 ] = t − 2 D 2 J g [ U 0 ] ,
the uniformly bounded Hessian loses at most C ε ∥ h ∥ E 2 C\varepsilon\|h\|_E^2 C ε ∥ h ∥ E 2 , so its positive gap persists. Use the energy Riesz map to represent this Hessian and the complementary differential by L \mathcal{L} L and R U \mathcal{R}_U R U . Define
N U ( w ) = R E ( D X F ( w ) − D X F ( 0 ) − D X 2 F ( 0 ) w ) , F = J g ∘ ret U ∣ X . \mathcal{N}_U(w)=R_E\bigl(D_XF(w)-D_XF(0)-D_X^2F(0)w\bigr),\qquad F=\mathcal{J}_g\circ\operatorname{ret}_U|_X. N U ( w ) = R E ( D X F ( w ) − D X F ( 0 ) − D X 2 F ( 0 ) w ) , F = J g ∘ ret U ∣ X .
The projected equation is
L w = − R U − N U ( w ) . \mathcal{L}w=-\mathcal{R}_U-\mathcal{N}_U(w). L w = − R U − N U ( w ) .
Lemma 3.7 gives
∥ R U ∥ H 1 ( M ) ≤ C ( M , g ) ε ; (67) \|\mathcal{R}_U\|_{H^1(M)}\leq C(M,g)\varepsilon; \tag*{(67)} ∥ R U ∥ H 1 ( M ) ≤ C ( M , g ) ε ; (67)
Lemma 5.1 gives the refined bound when H g ( x ) = 0 H_g(x)=0 H g ( x ) = 0 . Since f ( u ) = ∣ u ∣ q − 2 u f(u)=|u|^{q-2}u f ( u ) = ∣ u ∣ q − 2 u has a γ \gamma γ -Hölder derivative, the trace inequality and the uniformly bounded normalization factors imply
∥ N U ( w ) ∥ E ≤ C ∥ w ∥ H 1 1 + γ , \|\mathcal{N}_U(w)\|_E\leq C\|w\|_{H^1}^{1+\gamma}, ∥ N U ( w ) ∥ E ≤ C ∥ w ∥ H 1 1 + γ ,
∥ N U ( w ) − N U ( v ) ∥ E ≤ C ( ∥ w ∥ H 1 + ∥ v ∥ H 1 ) γ ∥ w − v ∥ H 1 . \|\mathcal{N}_U(w)-\mathcal{N}_U(v)\|_E\leq C\bigl(\|w\|_{H^1}+\|v\|_{H^1}\bigr)^\gamma\|w-v\|_{H^1}. ∥ N U ( w ) − N U ( v ) ∥ E ≤ C ( ∥ w ∥ H 1 + ∥ v ∥ H 1 ) γ ∥ w − v ∥ H 1 .
The multiplier terms obey the same bounds. Uniform C 1 C^1 C 1 inversion (Lemma 5.7) therefore gives a unique solution in a fixed small ball and
∥ w ∥ H 1 ≤ C ∥ R U ∥ E . \|w\|_{H^1}\leq C\|\mathcal{R}_U\|_E. ∥ w ∥ H 1 ≤ C ∥ R U ∥ E .
This proves both estimates for the correction and the parameter regularity. The energy expansion follows from Lemma 5.8; cutoff contributions are O ( ω ( R ) 2 + ε ω ( R ) ) = o ( ε 2 ) O(\omega(R)^2+\varepsilon\omega(R))=o(\varepsilon^2) O ( ω ( R ) 2 + ε ω ( R )) = o ( ε 2 ) . □
Definition 5.10 (Reduced functional). For the correction of Lemma 5.9, put Φ ( x , ε ) = ret U x , ε ( w x , ε ) ∈ S \Phi(x,\varepsilon)=\operatorname{ret}_{U_{x,\varepsilon}}(w_{x,\varepsilon})\in\mathcal{S} Φ ( x , ε ) = ret U x , ε ( w x , ε ) ∈ S . The reduced functional is
J 1 ( x , ε ) : = J g ( Φ ( x , ε ) ) = J g ( U x , ε + w x , ε ) , (68) \mathcal{J}_1(x,\varepsilon):=\mathcal{J}_g(\Phi(x,\varepsilon))=\mathcal{J}_g(U_{x,\varepsilon}+w_{x,\varepsilon}), \tag*{(68)} J 1 ( x , ε ) := J g ( Φ ( x , ε )) = J g ( U x , ε + w x , ε ) , (68)
the second equality following from homogeneity of the quotient.
The reduced quadratic coefficient.
Lemma 5.11 (Resolvent limit for the degree-2 source). Assume n ≥ 5 n\geq5 n ≥ 5 , ( Y ∂ + ) (Y^+_\partial) ( Y ∂ + ) , ( B G 3 + ) (\mathrm{BG3}^+) ( BG3 + ) , and H g ≡ 0 H_g\equiv0 H g ≡ 0 near x x x ; write S : = I I ˚ ( x ) S:=\mathring{\mathrm{II}}(x) S := II ˚ ( x ) . Let B x , ε B_{x,\varepsilon} B x , ε be the pulled-back Jacobi form on X x , ε X_{x,\varepsilon} X x , ε , normalized by D 2 ( J ^ ∘ ret U x , ε ) ( 0 ) [ v , ϕ ] = 2 B x , ε ( v , ϕ ) D^2(\widehat{\mathcal{J}}\circ\operatorname{ret}_{U_{x,\varepsilon}})(0)[v,\phi]=2B_{x,\varepsilon}(v,\phi) D 2 ( J ∘ ret U x , ε ) ( 0 ) [ v , ϕ ] = 2 B x , ε ( v , ϕ ) for v , ϕ ∈ X x , ε v,\phi\in X_{x,\varepsilon} v , ϕ ∈ X x , ε , where J ^ = J g / S ∗ \widehat{\mathcal{J}}=\mathcal{J}_g/S_* J = J g / S ∗ , with Hessian operator L x , ε = 2 B x , ε \mathcal{L}_{x,\varepsilon}=2B_{x,\varepsilon} L x , ε = 2 B x , ε . Write L f l a t = 2 B n \mathcal{L}_{\mathrm{flat}}=2B_n L flat = 2 B n on X ♭ X^\flat X ♭ . Model directions are rescaled without an amplitude factor, so the limiting form of B x , ε B_{x,\varepsilon} B x , ε is S ∗ − 1 B n S_*^{-1}B_n S ∗ − 1 B n . Define the degree-2 source F 1 ( x ) F_1(x) F 1 ( x ) by
⟨ F 1 ( x ) , ϕ ⟩ : = − 2 L S ( ϕ ) = 4 ∫ R + n y n S a b ∂ a U + ∂ b ϕ d y , (69) \langle F_1(x),\phi\rangle:=-2\mathcal{L}_S(\phi)=4\int_{\mathbb{R}^n_+}y_nS^{ab}\partial_aU_+\partial_b\phi\,dy, \tag*{(69)} ⟨ F 1 ( x ) , ϕ ⟩ := − 2 L S ( ϕ ) = 4 ∫ R + n y n S ab ∂ a U + ∂ b ϕ d y , (69) with L S \mathcal{L}_S L S as in (5.5), and set F ^ 1 ( x ) : = t ∗ − 1 F 1 ( x ) = S ∗ − 1 / 2 F 1 ( x ) \widehat{F}_1(x):=t_*^{-1}F_1(x)=S_*^{-1/2}F_1(x) F 1 ( x ) := t ∗ − 1 F 1 ( x ) = S ∗ − 1/2 F 1 ( x ) , where t ∗ = S ∗ 1 / 2 t_*=S_*^{1/2} t ∗ = S ∗ 1/2 . We use the same notation for the corresponding covector on X x , ε X_{x,\varepsilon} X x , ε under the rescaled Fermi identification. Then:
(a) F 1 ( x ) ∈ ( H ˙ 1 ( R + n ) ) ∗ F_1(x)\in(\dot{H}^1(\mathbb{R}^n_+))^* F 1 ( x ) ∈ ( H ˙ 1 ( R + n ) ) ∗ for n ≥ 5 n\geq5 n ≥ 5 , and the differential of the normalized functional J ^ : = J g / S ∗ \widehat{\mathcal{J}}:=\mathcal{J}_g/S_* J := J g / S ∗ , restricted to the complement and denoted by R ^ x , ε \widehat{\mathcal{R}}_{x,\varepsilon} R x , ε , satisfies ∥ ε − 1 R ^ x , ε − F ^ 1 ( x ) ∥ X x , ε ∗ → 0 \|\varepsilon^{-1}\widehat{\mathcal{R}}_{x,\varepsilon}-\widehat{F}_1(x)\|_{X^*_{x,\varepsilon}}\to0 ∥ ε − 1 R x , ε − F 1 ( x ) ∥ X x , ε ∗ → 0 , uniformly for x x x in compact subsets of a boundary region on which H g ≡ 0 H_g\equiv0 H g ≡ 0 .
(b) ⟨ F ^ 1 ( x ) , L x , ε − 1 F ^ 1 ( x ) ⟩ → ⟨ F 1 ( x ) , L f l a t − 1 F 1 ( x ) ⟩ = 2 δ κ 3 L S ( n ) ∣ S ∣ 2 \langle\widehat{F}_1(x),\mathcal{L}_{x,\varepsilon}^{-1}\widehat{F}_1(x)\rangle\to\langle F_1(x),\mathcal{L}_{\mathrm{flat}}^{-1}F_1(x)\rangle=2\delta\kappa_3^{\mathrm{LS}}(n)|S|^2 ⟨ F 1 ( x ) , L x , ε − 1 F 1 ( x )⟩ → ⟨ F 1 ( x ) , L flat − 1 F 1 ( x )⟩ = 2 δ κ 3 LS ( n ) ∣ S ∣ 2 , uniformly on compact subsets.
Proof. For (a), let f 1 b = 4 t S a b ∂ a U + f_1^b=4tS_{ab}\partial_aU_+ f 1 b = 4 t S ab ∂ a U + and f 1 n = 0 f_1^n=0 f 1 n = 0 . Then ∣ f 1 ( y ) ∣ ≤ C ( 1 + ∣ y ∣ ) 2 − n |f_1(y)|\le C(1+|y|)^{2-n} ∣ f 1 ( y ) ∣ ≤ C ( 1 + ∣ y ∣ ) 2 − n , so f 1 ∈ L 2 f_1\in L^2 f 1 ∈ L 2 for n ≥ 5 n\ge5 n ≥ 5 and F 1 = − div f 1 F_1=-\operatorname{div}f_1 F 1 = − div f 1 belongs to ( D 1 , 2 ) ∗ (D^{1,2})^* ( D 1 , 2 ) ∗ . Its tail satisfies
sup ∥ ∇ ϕ ∥ 2 ≤ 1 ∣ ∫ R + n ∖ B R + f 1 ⋅ ∇ ϕ ∣ ≤ ∥ f 1 ∥ L 2 ( R + n ∖ B R + ) = : η ( R ) ⟶ 0. (70) \sup_{\lVert\nabla\phi\rVert_2\le1}\left|\int_{\mathbb{R}^n_+\setminus B_R^+} f_1\cdot\nabla\phi\right|\le\lVert f_1\rVert_{L^2(\mathbb{R}^n_+\setminus B_R^+)}=:\eta(R)\longrightarrow0. \tag*{(70)} ∥ ∇ ϕ ∥ 2 ≤ 1 sup ∫ R + n ∖ B R + f 1 ⋅ ∇ ϕ ≤ ∥ f 1 ∥ L 2 ( R + n ∖ B R + ) =: η ( R ) ⟶ 0. (70)
Lemma 5.2 gives the error in the first-variation expansion C ( η n ( ε ) + ω ( R ( ε ) ) ) = o ( ε ) C(\eta_n(\varepsilon)+\omega(R(\varepsilon)))=o(\varepsilon) C ( η n ( ε ) + ω ( R ( ε ))) = o ( ε ) . Equation (70) also removes the localization of the source, proving (a).
For (b), solve
2 B x , ε ( v ε , ϕ ) = ⟨ F ^ 1 ( x ) , ϕ ⟩ ( ϕ ∈ X x , ε ) . (71) 2\mathcal{B}_{x,\varepsilon}(v_\varepsilon,\phi)=\langle\widehat{F}_1(x),\phi\rangle\qquad(\phi\in X_{x,\varepsilon}). \tag*{(71)} 2 B x , ε ( v ε , ϕ ) = ⟨ F 1 ( x ) , ϕ ⟩ ( ϕ ∈ X x , ε ) . (71)
Uniform coercivity bounds v ε v_\varepsilon v ε in H 1 ( M ) H^1(M) H 1 ( M ) . Its Fermi rescaling has bounded gradient norm and bounded L 2 n / ( n − 2 ) L^{2n/(n-2)} L 2 n / ( n − 2 ) norm, as in the proof of Theorem 3.6. Extract a subsequence converging weakly on every half-ball of integer radius to V ∈ D 1 , 2 ( R + n ) V\in D^{1,2}(\mathbb{R}^n_+) V ∈ D 1 , 2 ( R + n ) . Let L L L be a fixed test radius and R c ( ε ) = ε − 3 / 4 R_c(\varepsilon)=\varepsilon^{-3/4} R c ( ε ) = ε − 3/4 the moving cutoff. Eventually R c ( ε ) > 2 L R_c(\varepsilon)>2L R c ( ε ) > 2 L , so local weak convergence and (70) give
lim sup ε → 0 ∣ ⟨ F ^ 1 , v ε ⟩ − ⟨ F ^ 1 , V ⟩ ∣ ≤ C ∥ f 1 ∥ L 2 ( ∣ Y ∣ > L ) ≤ C L − ( n − 4 ) / 2 . \limsup_{\varepsilon\to0}\left|\langle\widehat{F}_1,v_\varepsilon\rangle-\langle\widehat{F}_1,V\rangle\right|\le C\lVert f_1\rVert_{L^2(|Y|>L)}\le CL^{-(n-4)/2}. ε → 0 lim sup ⟨ F 1 , v ε ⟩ − ⟨ F 1 , V ⟩ ≤ C ∥ f 1 ∥ L 2 ( ∣ Y ∣ > L ) ≤ C L − ( n − 4 ) /2 .
Letting L → ∞ L\to\infty L → ∞ proves convergence of the pairing. The normalized translation and dilation fields converge strongly to the listed Z j b Z_j^b Z j b after rescaling. The energy Riesz representative of the boundary constraint converges by Remark 5.5. Their orthogonality conditions pass to V V V , so V ∈ X b V\in X^b V ∈ X b . For ϕ ∈ X b ∩ C c ∞ ( R + n ) \phi\in X^b\cap C^\infty_c(\mathbb{R}^n_+) ϕ ∈ X b ∩ C c ∞ ( R + n ) , subtract an o ( 1 ) o(1) o ( 1 ) finite-dimensional projection from its Fermi rescaling to make it belong to X x , ε X_{x,\varepsilon} X x , ε . On its fixed support the forms converge to S ∗ − 1 B n S_*^{-1}\mathcal{B}_n S ∗ − 1 B n . Passing to the limit gives
2 S ∗ − 1 B n ( V , ϕ ) = ⟨ F ^ 1 ( x ) , ϕ ⟩ . 2S_*^{-1}\mathcal{B}_n(V,\phi)=\langle\widehat{F}_1(x),\phi\rangle. 2 S ∗ − 1 B n ( V , ϕ ) = ⟨ F 1 ( x ) , ϕ ⟩ .
Compactly supported tests in the complement are dense. To see this, approximate in the Dirichlet norm and subtract a finite linear combination of fixed compactly supported functions to make the defining moments vanish. The flat gap therefore identifies the unique limit
V = S ∗ L f l a t − 1 F ^ 1 = t ∗ L f l a t − 1 F 1 , L f l a t = 2 B n . V=S_*\mathcal{L}_{\mathrm{flat}}^{-1}\widehat{F}_1=t_*\mathcal{L}_{\mathrm{flat}}^{-1}F_1,\qquad\mathcal{L}_{\mathrm{flat}}=2\mathcal{B}_n. V = S ∗ L flat − 1 F 1 = t ∗ L flat − 1 F 1 , L flat = 2 B n .
Consequently the normalization factors cancel, and
⟨ F ^ 1 , V ⟩ = ⟨ F 1 , L f l a t − 1 F 1 ⟩ = 2 L S ( B n − 1 L S ) = 2 δ κ 3 L S ( n ) ∣ S ∣ 2 . \langle\widehat{F}_1,V\rangle=\langle F_1,\mathcal{L}_{\mathrm{flat}}^{-1}F_1\rangle=2\mathcal{L}_S(\mathcal{B}_n^{-1}\mathcal{L}_S)=2\delta\kappa_3^{\mathrm{LS}}(n)|S|^2. ⟨ F 1 , V ⟩ = ⟨ F 1 , L flat − 1 F 1 ⟩ = 2 L S ( B n − 1 L S ) = 2 δ κ 3 LS ( n ) ∣ S ∣ 2 .
Uniqueness makes every subsequence have this limit. The same argument for x j → x ∞ x_j\to x_\infty x j → x ∞ in convergent Fermi frames proves uniformity on compact sets. □
Proposition 5.12 (Reduced second-order coefficient and Lyapunov–Schmidt correction). Assume n ≥ 5 n\ge5 n ≥ 5 , ( Y ∂ + ) (Y_\partial^+) ( Y ∂ + ) , and ( B G 3 + ) (\mathrm{BG}^{3+}) ( BG 3 + ) , and use the reduction of Lemma 5.9 and Definition 5.10. Suppose H g ≡ 0 H_g\equiv0 H g ≡ 0 on a boundary region. Let δ κ 3 L S ( n ) > 0 \delta\kappa_3^{\mathrm{LS}}(n)>0 δ κ 3 LS ( n ) > 0 be the Lyapunov–Schmidt correction on the degree-2 subspace defined in (56) and set κ 3 : = κ 3 b a r e − δ κ 3 L S \kappa_3:=\kappa_3^{\mathrm{bare}}-\delta\kappa_3^{\mathrm{LS}} κ 3 := κ 3 bare − δ κ 3 LS . Then
J 1 ( x , ε ) S ∗ = 1 + R g r e d ( x ) ε 2 + o ( ε 2 ) , R g r e d ( x ) = ( κ 3 b a r e ( n ) − δ κ 3 L S ( n ) ) ∣ I I ˚ ( x ) ∣ 2 (72) \frac{\mathcal{J}_1(x,\varepsilon)}{S_*}=1+\mathcal{R}_g^{\mathrm{red}}(x)\varepsilon^2+o(\varepsilon^2),\qquad\mathcal{R}_g^{\mathrm{red}}(x)=\left(\kappa_3^{\mathrm{bare}}(n)-\delta\kappa_3^{\mathrm{LS}}(n)\right)|\mathring{\mathrm{II}}(x)|^2 \tag*{(72)} S ∗ J 1 ( x , ε ) = 1 + R g red ( x ) ε 2 + o ( ε 2 ) , R g red ( x ) = ( κ 3 bare ( n ) − δ κ 3 LS ( n ) ) ∣ II ˚ ( x ) ∣ 2 (72) uniformly for x x x on compact subsets of that region, by Proposition 2.5. If the ambient metric varies in a C m C^m C m family with m ≥ 3 m\ge3 m ≥ 3 , this coefficient depends C m − 3 C^{m-3} C m − 3 -smoothly on g g g .
Proof. Put A = I I ˚ ( x ) A=\mathring{\mathrm{II}}(x) A = II ˚ ( x ) and use the normalized functional J ^ = J g / S ∗ \widehat{\mathcal{J}}=\mathcal{J}_g/S_* J = J g / S ∗ . Lemma 5.8 gives
J 1 S ∗ = J ^ ( u ) − 1 2 ⟨ R ^ x , ε , L x , ε − 1 R ^ x , ε ⟩ + O ( ε 2 + γ ) , \frac{\mathcal{J}_1}{S_*}=\widehat{\mathcal{J}}(u)-\frac{1}{2}\langle\widehat{R}_{x,\varepsilon},\mathcal{L}_{x,\varepsilon}^{-1}\widehat{R}_{x,\varepsilon}\rangle+O(\varepsilon^{2+\gamma}), S ∗ J 1 = J ( u ) − 2 1 ⟨ R x , ε , L x , ε − 1 R x , ε ⟩ + O ( ε 2 + γ ) , where R ^ \widehat{R} R and L \mathcal{L} L are the differential and Hessian of the same normalized functional, restricted to the complement. Lemma 5.11 and uniform inversion give
R ^ x , ε = ε F ^ 1 ( x ) + o ∗ ( ε ) , \widehat{R}_{x,\varepsilon}=\varepsilon\widehat{F}_1(x)+o_*(\varepsilon), R x , ε = ε F 1 ( x ) + o ∗ ( ε ) , ⟨ R ^ x , ε , L x , ε − 1 R ^ x , ε ⟩ = 2 ε 2 δ κ 3 L S ( n ) ∣ A ∣ 2 + o ( ε 2 ) . \langle\widehat{R}_{x,\varepsilon},\mathcal{L}_{x,\varepsilon}^{-1}\widehat{R}_{x,\varepsilon}\rangle=2\varepsilon^2\delta\kappa_3^{\mathrm{LS}}(n)|A|^2+o(\varepsilon^2). ⟨ R x , ε , L x , ε − 1 R x , ε ⟩ = 2 ε 2 δ κ 3 LS ( n ) ∣ A ∣ 2 + o ( ε 2 ) .
The factor 1 / 2 1/2 1/2 therefore subtracts exactly δ κ 3 L S ( n ) ∣ A ∣ 2 \delta\kappa^{\mathrm{LS}}_3(n)|A|^2 δ κ 3 LS ( n ) ∣ A ∣ 2 from the standard-bubble coefficient. Proposition 2.5 gives the standard-bubble term and proves (72) . Uniformity follows from the uniform source and resolvent limits. The displayed formula gives the asserted metric regularity. Finally, Cauchy–Schwarz in B n \mathcal{B}_n B n gives
L S ( ϕ ) B n ( ϕ , ϕ ) ≤ δ κ 3 L S ( n ) ∣ S ∣ 2 ( 0 ≠ ϕ ∈ X ♭ ) , (73) \frac{\mathcal{L}_S(\phi)}{\mathcal{B}_n(\phi,\phi)} \leq\delta\kappa^{\mathrm{LS}}_3(n)|S|^2 \qquad(0\ne\phi\in X^\flat), \tag*{(73)} B n ( ϕ , ϕ ) L S ( ϕ ) ≤ δ κ 3 LS ( n ) ∣ S ∣ 2 ( 0 = ϕ ∈ X ♭ ) , (73)
so the exact coefficient also implies every upper bound obtained from a fixed test function.
Exact evaluation of the quadratic coefficient
Convention 5.13 (Coefficients before and after reduction). In the representative with minimal boundary, R g b a r e = κ 3 b a r e ∣ I I ˚ ∣ 2 \mathcal{R}^{\mathrm{bare}}_g=\kappa^{\mathrm{bare}}_3|\mathring{\mathrm{II}}|^2 R g bare = κ 3 bare ∣ II ˚ ∣ 2 is the standard-bubble coefficient, while R g r e d \mathcal{R}^{\mathrm{red}}_g R g red is the corrected coefficient in (72) .
We evaluate the quadratic form in (56) . Fix S ∈ Sym 0 2 ( R n − 1 ) S\in\operatorname{Sym}_0^2(\mathbb{R}^{n-1}) S ∈ Sym 0 2 ( R n − 1 ) with ∣ S ∣ = 1 |S|=1 ∣ S ∣ = 1 , and write
a = 1 + t , s = ∣ y ′ ∣ 2 + a 2 , Q = Q S ( y ′ ) . a=1+t,\qquad s=|y'|^2+a^2,\qquad Q=Q_S(y'). a = 1 + t , s = ∣ y ′ ∣ 2 + a 2 , Q = Q S ( y ′ ) .
Since ( q − 1 ) λ f l a t U + q − 2 = n / ( 1 + ∣ y ′ ∣ 2 ) (q-1)\lambda_{\mathrm{flat}}U_+^{q-2}=n/(1+|y'|^2) ( q − 1 ) λ flat U + q − 2 = n / ( 1 + ∣ y ′ ∣ 2 ) on the boundary, the boundary operator associated with B n \mathcal{B}_n B n is
R v = − ∂ t v − n 1 + ∣ y ′ ∣ 2 v . \mathcal{R}v=-\partial_t v-\frac{n}{1+|y'|^2}v. R v = − ∂ t v − 1 + ∣ y ′ ∣ 2 n v .
All functions in the following calculation have tangential spherical harmonic degree two, so the constraint and Jacobi orthogonality conditions hold automatically.
Lemma 5.14 (A harmonic Robin pairing). For n ≥ 5 n\geq5 n ≥ 5 , put b 0 = Q ( y ′ ) ( 1 + ∣ y ′ ∣ 2 ) − n / 2 b_0=Q(y')(1+|y'|^2)^{-n/2} b 0 = Q ( y ′ ) ( 1 + ∣ y ′ ∣ 2 ) − n /2 . There is a unique harmonic h 0 ∈ X 2 , S h_0\in X_{2,S} h 0 ∈ X 2 , S satisfying R h 0 = b 0 \mathcal{R}h_0=b_0 R h 0 = b 0 weakly. Moreover,
R n : = ∫ R n − 1 b 0 h 0 d y ′ = n ( n 2 − 2 n − 5 ) ( n − 2 ) ( n − 1 ) 2 ( n + 1 ) ( n − 3 ) ( n − 4 ) A n 2 . (74) R_n:=\int_{\mathbb{R}^{n-1}}b_0h_0\,dy'=\frac{n(n^2-2n-5)}{(n-2)(n-1)^2(n+1)(n-3)(n-4)A_n^2}. \tag*{(74)} R n := ∫ R n − 1 b 0 h 0 d y ′ = ( n − 2 ) ( n − 1 ) 2 ( n + 1 ) ( n − 3 ) ( n − 4 ) A n 2 n ( n 2 − 2 n − 5 ) . (74)
Proof. The datum b 0 b_0 b 0 belongs to the dual trace space for n > 4 n>4 n > 4 , since b 0 ( y ′ ) = O ( ∣ y ′ ∣ 2 − n ) b_0(y')=O(|y'|^{2-n}) b 0 ( y ′ ) = O ( ∣ y ′ ∣ 2 − n ) . Coercivity of B n \mathcal{B}_n B n on X 2 , S X_{2,S} X 2 , S gives existence and uniqueness. The equation holds against all Dirichlet tests by angular orthogonality.
Under the Cayley map to the unit ball,
ζ = 2 ( y + e n ) s − e n , h 0 ( y ) = s − ( n − 2 ) / 2 H ( ζ ) , \zeta=\frac{2(y+e_n)}{s}-e_n,\qquad h_0(y)=s^{-(n-2)/2}H(\zeta), ζ = s 2 ( y + e n ) − e n , h 0 ( y ) = s − ( n − 2 ) /2 H ( ζ ) ,
e n e_n e n is the final Euclidean coordinate vector. The function H H H is harmonic and its boundary condition is
( ∂ ν o u t − 1 ) H = f , f ( ζ ) = Q ( ζ ′ ) 2 ( 1 + ζ n ) 2 . (\partial_{\nu_{\mathrm{out}}}-1)H=f,\qquad f(\zeta)=\frac{Q(\zeta')}{2(1+\zeta_n)^2}. ( ∂ ν out − 1 ) H = f , f ( ζ ) = 2 ( 1 + ζ n ) 2 Q ( ζ ′ ) .
The change of variables gives R n = 2 − ( n − 2 ) ∫ S n − 1 f H R_n=2^{-(n-2)}\int_{\mathbb{S}^{n-1}}fH R n = 2 − ( n − 2 ) ∫ S n − 1 f H . Put z = ζ n z=\zeta_n z = ζ n and ζ ′ = 1 − z 2 θ \zeta'=\sqrt{1-z^2}\,\theta ζ ′ = 1 − z 2 θ . Let C j ( n + 2 ) / 2 C_j^{(n+2)/2} C j ( n + 2 ) /2 be the Gegenbauer polynomial of degree j j j . The relevant spherical harmonics are
( 1 − z 2 ) C j ( n + 2 ) / 2 ( z ) Q ( θ ) , j ≥ 0. (1-z^2)C_j^{(n+2)/2}(z)Q(\theta),\qquad j\geq0. ( 1 − z 2 ) C j ( n + 2 ) /2 ( z ) Q ( θ ) , j ≥ 0.
Their degrees are j + 2 j+2 j + 2 , so the harmonic Robin inverse has eigenvalue ( j + 1 ) − 1 (j+1)^{-1} ( j + 1 ) − 1 on these modes. For general scalar-flat boundary pairings, see [24 Proposition 3.3] .
Set α = ( n − 1 ) / 2 \alpha=(n-1)/2 α = ( n − 1 ) /2 , and use ( b ) j = Γ ( b + j ) / Γ ( b ) (b)_j=\Gamma(b+j)/\Gamma(b) ( b ) j = Γ ( b + j ) /Γ ( b ) for the rising factorial. The angular identity and the one-dimensional integrals are
∫ S n − 2 Q ( θ ) 2 d θ = 2 ∣ S n − 2 ∣ ( n − 1 ) ( n + 1 ) , \int_{\mathbb{S}^{n-2}}Q(\theta)^2\,d\theta=\frac{2|\mathbb{S}^{n-2}|}{(n-1)(n+1)}, ∫ S n − 2 Q ( θ ) 2 d θ = ( n − 1 ) ( n + 1 ) 2∣ S n − 2 ∣ ,
I j = ∫ − 1 1 ( 1 − z 2 ) ( n + 1 ) / 2 ( 1 + z ) − 2 C j ( n + 2 ) / 2 ( z ) d z , N j = ∫ − 1 1 ( 1 − z 2 ) ( n + 1 ) / 2 ( C j ( n + 2 ) / 2 ( z ) ) 2 d z . I_j=\int_{-1}^{1}(1-z^2)^{(n+1)/2}(1+z)^{-2}C_j^{(n+2)/2}(z)\,dz,\qquad N_j=\int_{-1}^{1}(1-z^2)^{(n+1)/2}\left(C_j^{(n+2)/2}(z)\right)^2\,dz. I j = ∫ − 1 1 ( 1 − z 2 ) ( n + 1 ) /2 ( 1 + z ) − 2 C j ( n + 2 ) /2 ( z ) d z , N j = ∫ − 1 1 ( 1 − z 2 ) ( n + 1 ) /2 ( C j ( n + 2 ) /2 ( z ) ) 2 d z .
The Gegenbauer norm formula and its Rodrigues formula, followed by integration by parts, give
I j I 0 = ( − 1 ) j ( n + 2 ) j ( 2 ) j j ! ( n + 1 ) j , N j N 0 = ( n + 2 ) j j ! n + 2 2 j + n + 2 . \frac{I_j}{I_0}=(-1)^j\frac{(n+2)_j(2)_j}{j!(n+1)_j},\qquad\frac{N_j}{N_0}=\frac{(n+2)_j}{j!}\frac{n+2}{2j+n+2}. I 0 I j = ( − 1 ) j j ! ( n + 1 ) j ( n + 2 ) j ( 2 ) j , N 0 N j = j ! ( n + 2 ) j 2 j + n + 2 n + 2 .
where
I 0 = 2 n B ( α , α + 2 ) , N 0 = 2 n + 2 B ( α + 2 , α + 2 ) . I_0=2^nB(\alpha,\alpha+2),\qquad N_0=2^{n+2}B(\alpha+2,\alpha+2). I 0 = 2 n B ( α , α + 2 ) , N 0 = 2 n + 2 B ( α + 2 , α + 2 ) .
Here B is the beta function. Thus
R n = ∣ S n − 2 ∣ 2 ( n − 1 ) ( n + 1 ) B ( α , α + 2 ) 2 B ( α + 2 , α + 2 ) ⋅ ∑ j = 0 ∞ ( n + 2 ) j ( 2 ) j 2 j ! ( n + 1 ) j 2 2 j + n + 2 ( n + 2 ) ( j + 1 ) . (75) R_n=\frac{\lvert S^{n-2}\rvert}{2(n-1)(n+1)}\frac{B(\alpha,\alpha+2)^2}{B(\alpha+2,\alpha+2)}\cdot\sum_{j=0}^{\infty}\frac{(n+2)_j(2)_j^2}{j!(n+1)_j^2}\frac{2j+n+2}{(n+2)(j+1)}. \tag*{(75)} R n = 2 ( n − 1 ) ( n + 1 ) ∣ S n − 2 ∣ B ( α + 2 , α + 2 ) B ( α , α + 2 ) 2 ⋅ j = 0 ∑ ∞ j ! ( n + 1 ) j 2 ( n + 2 ) j ( 2 ) j 2 ( n + 2 ) ( j + 1 ) 2 j + n + 2 . (75)
The summand simplifies to
n ! ( n + 1 ) ( n + 2 ) ( j + n + 1 ) ( 2 j + n + 2 ) ( j + 2 ) ( j + 3 ) ⋯ ( j + n ) . \frac{n!}{(n+1)(n+2)}\frac{(j+n+1)(2j+n+2)}{(j+2)(j+3)\cdots(j+n)}. ( n + 1 ) ( n + 2 ) n ! ( j + 2 ) ( j + 3 ) ⋯ ( j + n ) ( j + n + 1 ) ( 2 j + n + 2 ) .
Using ( ∏ k = 2 n ( j + k ) ) − 1 = B ( j + 2 , n − 1 ) / ( n − 2 ) ! \left(\prod_{k=2}^{n}(j+k)\right)^{-1}=B(j+2,n-1)/(n-2)! ( ∏ k = 2 n ( j + k ) ) − 1 = B ( j + 2 , n − 1 ) / ( n − 2 )! and summing the geometric series and its first two derivatives gives
∑ j = 0 ∞ ( j + n + 1 ) ( 2 j + n + 2 ) ∏ k = 2 n ( j + k ) = n ( n 2 − 2 n − 5 ) ( n − 2 ) ! ( n − 1 ) ( n − 3 ) ( n − 4 ) . \sum_{j=0}^{\infty}\frac{(j+n+1)(2j+n+2)}{\prod_{k=2}^{n}(j+k)}=\frac{n(n^2-2n-5)}{(n-2)!(n-1)(n-3)(n-4)}. j = 0 ∑ ∞ ∏ k = 2 n ( j + k ) ( j + n + 1 ) ( 2 j + n + 2 ) = ( n − 2 )! ( n − 1 ) ( n − 3 ) ( n − 4 ) n ( n 2 − 2 n − 5 ) .
For n > 4 n>4 n > 4 , convergence of the positive energy series gives the identity by approximation with finite spherical-harmonic sums. Substituting the last identity in (75) , and using the normalization of A n A_n A n , proves (74) . □
We shall use the following consequence of Green’s formula. Suppose p p p has tangential spherical harmonic degree two, solves − Δ p = f -\Delta p=f − Δ p = f in the half-space, and has Robin datum b = R p b=\mathcal{R}p b = R p . Let h b h_b h b be the harmonic solution with Robin datum b b b . Then B n − 1 f = p − h b \mathcal{B}_n^{-1}f=p-h_b B n − 1 f = p − h b , and
f ( B n − 1 f ) = f ( p ) − ∫ R n − 1 b p + ∫ R n − 1 b h b . (76) f(\mathcal{B}_n^{-1}f)=f(p)-\int_{\mathbb{R}^{n-1}}bp+\int_{\mathbb{R}^{n-1}}bh_b. \tag*{(76)} f ( B n − 1 f ) = f ( p ) − ∫ R n − 1 b p + ∫ R n − 1 b h b . (76)
Indeed, symmetry gives B n ( p , h b ) = ∫ b p \mathcal{B}_n(p,h_b)=\int bp B n ( p , h b ) = ∫ b p , whereas the equation for p p p gives B n ( p , h b ) = f ( h b ) + ∫ b h b \mathcal{B}_n(p,h_b)=f(h_b)+\int bh_b B n ( p , h b ) = f ( h b ) + ∫ b h b . These identities are valid by density in the finite-energy space.
Proposition 5.15 (Exact quadratic coefficients). For every n ≥ 5 n\geq5 n ≥ 5 ,
δ κ 3 L S ( n ) = ( 2 n − 1 ) ( n + 2 ) 2 ( n − 1 ) 2 ( n + 1 ) ( n − 3 ) ( n − 4 ) , (77) \delta\kappa_3^{\mathrm{LS}}(n)=\frac{(2n-1)(n+2)}{2(n-1)^2(n+1)(n-3)(n-4)}, \tag*{(77)} δ κ 3 LS ( n ) = 2 ( n − 1 ) 2 ( n + 1 ) ( n − 3 ) ( n − 4 ) ( 2 n − 1 ) ( n + 2 ) , (77) κ 3 r e d ( n ) = − ( n − 2 ) ( n 2 − 2 n − 2 ) 2 ( n − 1 ) 2 ( n + 1 ) ( n − 3 ) ( n − 4 ) < 0. (78) \kappa_3^{\mathrm{red}}(n)=-\frac{(n-2)(n^2-2n-2)}{2(n-1)^2(n+1)(n-3)(n-4)}<0. \tag*{(78)} κ 3 red ( n ) = − 2 ( n − 1 ) 2 ( n + 1 ) ( n − 3 ) ( n − 4 ) ( n − 2 ) ( n 2 − 2 n − 2 ) < 0. (78) In particular, δ κ 3 L S ( 5 ) = 21 / 128 \delta\kappa_3^{\mathrm{LS}}(5)=21/128 δ κ 3 LS ( 5 ) = 21/128 and κ 3 r e d ( 5 ) = − 13 / 128 \kappa_3^{\mathrm{red}}(5)=-13/128 κ 3 red ( 5 ) = − 13/128 .
Proof. For ∣ S ∣ = 1 |S|=1 ∣ S ∣ = 1 , let K 1 = ( n − 2 ) A n / 2 K_1=(n-2)A_n/2 K 1 = ( n − 2 ) A n /2 and set
p 1 = K 1 Q ( α − 2 ) s − n / 2 . p_1=K_1Q(\alpha-2)s^{-n/2}. p 1 = K 1 Q ( α − 2 ) s − n /2 .
Direct differentiation gives
− Δ p 1 = 2 n ( n − 2 ) A n t Q s − ( n + 2 ) / 2 , R p 1 = − K 1 b 0 . -\Delta p_1=2n(n-2)A_ntQ s^{-(n+2)/2},\qquad\mathcal{R}p_1=-K_1b_0. − Δ p 1 = 2 n ( n − 2 ) A n tQ s − ( n + 2 ) /2 , R p 1 = − K 1 b 0 .
Thus w S = p 1 + K 1 h 0 w_S=p_1+K_1h_0 w S = p 1 + K 1 h 0 solves B n ( w S , ϕ ) = L S ( ϕ ) \mathcal{B}_n(w_S,\phi)=\mathcal{L}_S(\phi) B n ( w S , ϕ ) = L S ( ϕ ) . All these functions have finite Dirichlet energy for n ≥ 5 n\geq5 n ≥ 5 . The radial beta integral
∫ 0 ∞ r m − 1 ( r 2 + a 2 ) k d r = a m − 2 k 2 B ( m / 2 , k − m / 2 ) , 0 < m < 2 k , \int_0^\infty\frac{r^{m-1}}{(r^2+a^2)^k}\,dr=\frac{a^{m-2k}}{2}B(m/2,k-m/2),\qquad0<m<2k, ∫ 0 ∞ ( r 2 + a 2 ) k r m − 1 d r = 2 a m − 2 k B ( m /2 , k − m /2 ) , 0 < m < 2 k ,
and integration over a ∈ [ 1 , ∞ ) a \in[1,\infty) a ∈ [ 1 , ∞ ) give
L S ( p 1 ) = 6 − n 2 ( n − 1 ) ( n − 3 ) ( n − 4 ) \mathcal{L}_{S}(p_{1})=\frac{6-n}{2(n-1)(n-3)(n-4)} L S ( p 1 ) = 2 ( n − 1 ) ( n − 3 ) ( n − 4 ) 6 − n
∫ R n − 1 ( − K 1 b 0 ) p 1 = n − 2 4 ( n − 1 ) ( n − 3 ) , K 1 2 R n = n ( n − 2 ) ( n 2 − 2 n − 5 ) 4 ( n − 1 ) 2 ( n + 1 ) ( n − 3 ) ( n − 4 ) . \int_{\mathbb{R}^{n-1}}(-K_{1}b_{0})p_{1}=\frac{n-2}{4(n-1)(n-3)},\qquad K_{1}^{2}R_{n}=\frac{n(n-2)(n^{2}-2n-5)}{4(n-1)^{2}(n+1)(n-3)(n-4)}. ∫ R n − 1 ( − K 1 b 0 ) p 1 = 4 ( n − 1 ) ( n − 3 ) n − 2 , K 1 2 R n = 4 ( n − 1 ) 2 ( n + 1 ) ( n − 3 ) ( n − 4 ) n ( n − 2 ) ( n 2 − 2 n − 5 ) .
Equation (76) now yields (77) . Subtracting this value from κ 3 b a r e ( n ) \kappa_{3}^{\mathrm{bare}}(n) κ 3 bare ( n ) gives (78) . Its numerator is positive for n ≥ 5 n\geq5 n ≥ 5 . At n = 5 n=5 n = 5 , this quadratic form equals 1 16 − 3 32 + 25 128 = 21 128 \frac{1}{16}-\frac{3}{32}+\frac{25}{128}=\frac{21}{128} 16 1 − 32 3 + 128 25 = 128 21 . □
Remark 5.16. (A dimension-five trial direction.) For n = 5 n=5 n = 5 , the direction ϕ M = Q S t s − 5 / 2 \phi_{M}=Q_{S}t s^{-5/2} ϕ M = Q S t s − 5/2 corresponds to Marques’ perturbation in [37 Eq. (1.3)] . The same beta integrals give
L S ( ϕ M ) 2 B 5 ( ϕ M , ϕ M ) = 1 10 , 1 16 − 1 10 = − 3 80 . \frac{\mathcal{L}_{S}(\phi_{M})^{2}}{B_{5}(\phi_{M},\phi_{M})}=\frac{1}{10},\qquad\frac{1}{16}-\frac{1}{10}=-\frac{3}{80}. B 5 ( ϕ M , ϕ M ) L S ( ϕ M ) 2 = 10 1 , 16 1 − 10 1 = − 80 3 .
The exact complementary correction gives the smaller reduced coefficient − 13 / 128 -13/128 − 13/128 .
The reduced quadratic coefficient is negative in every dimension n ≥ 5 n\geq5 n ≥ 5 . At n = 6 n=6 n = 6 the quadratic coefficient of the standard bubble vanishes; at an umbilic point with zero mean curvature its cubic coefficient also vanishes. The next weighted moment is logarithmic in that dimension, so the fourth-order analysis below assumes n ≥ 7 n\geq7 n ≥ 7 .
Strict inequality for non-umbilic boundaries
Theorem 5.17. (Strict inequality for non-umbilic boundaries.) Assume n ≥ 5 n\geq5 n ≥ 5 , ( Y ∂ + ) (Y_{\partial}^{+}) ( Y ∂ + ) , and ( B G 3 + ) (\mathrm{BG}^{3+}) ( BG 3 + ) . Suppose there exists a boundary point p ∈ ∂ M p\in\partial M p ∈ ∂ M with I I ˚ ( p ) ≠ 0 \mathring{\mathrm{II}}(p)\ne0 II ˚ ( p ) = 0 . Then
C E s c ∗ ( M , g ) < S ∗ . C_{\mathrm{Esc}}^{*}(M,g)<S_{*}. C Esc ∗ ( M , g ) < S ∗ . For n ≥ 7 n\geq7 n ≥ 7 this strict inequality is due to Escobar [18 ] , for n = 6 n=6 n = 6 to Escobar [21 ] , and for n = 5 n=5 n = 5 to Marques [37 ] . In particular, the Escobar infimum is attained by a smooth nonnegative minimizer, positive on each connected component on which it is nonzero.
Proof. Choose g ^ = ϕ 4 / ( n − 2 ) g \widehat{g}=\phi^{4/(n-2)}g g = ϕ 4/ ( n − 2 ) g with ϕ ∣ ∂ M = 1 \phi|_{\partial M}=1 ϕ ∣ ∂ M = 1 and H g ^ = 0 H_{\widehat{g}}=0 H g = 0 , using [41 Lemma 3.13] . Then I I ˚ g ^ ( p ) = I I ˚ g ( p ) ≠ 0 \mathring{\mathrm{II}}_{\widehat{g}}(p)=\mathring{\mathrm{II}}_{g}(p)\ne0 II ˚ g ( p ) = II ˚ g ( p ) = 0 . Proposition 5.12 gives
J 1 ( p , ε ) / S ∗ = 1 + κ 3 r e d ( n ) ∣ I I ˚ ( p ) ∣ 2 ε 2 + o ( ε 2 ) < 1 J_{1}(p,\varepsilon)/S_{*}=1+\kappa_{3}^{\mathrm{red}}(n)|\mathring{\mathrm{II}}(p)|^{2}\varepsilon^{2}+o(\varepsilon^{2})<1 J 1 ( p , ε ) / S ∗ = 1 + κ 3 red ( n ) ∣ II ˚ ( p ) ∣ 2 ε 2 + o ( ε 2 ) < 1
for small ε \varepsilon ε , since the reduced coefficient is negative in every dimension n ≥ 5 n\geq5 n ≥ 5 . Conformal invariance transfers this test function to g g g and proves the strict inequality. Corollary 4.4 gives attainment. Replace a minimizer by its absolute value. Cherrier’s boundedness argument [7 proof of Theorem 1, pp. 160–165] gives a bounded minimizer. Boundary regularity, the strong maximum principle, and the Hopf boundary lemma give positivity on each nonzero component. The boundary nonlinearity is smooth on the positive range, so elliptic bootstrap gives a smooth minimizer; see also [14 p. 2] . □
Differentiated one-bubble reduction
Throughout this section assume n ≥ 5 n\geq5 n ≥ 5 , ( Y ∂ + ) (Y_{\partial}^{+}) ( Y ∂ + ) , and ( B G 3 + ) (\mathrm{BG}^{3+}) ( BG 3 + ) .
The scale expansion relates the norm of the constrained gradient to the distances from the standard and corrected bubble families. The complementary gap controls the error transverse to the corrected family; the scale equation supplies the geometric obstruction away from the umbilic locus.
Retain the normalized fields and complement of Lemma 5.1, and put P x , ε = Π x , ε P_{x,\varepsilon}=\Pi_{x,\varepsilon} P x , ε = Π x , ε and ω V ( ε ) = ω ( R ( ε ) ) \omega_{\mathcal{V}}(\varepsilon)=\omega(\mathcal{R}(\varepsilon)) ω V ( ε ) = ω ( R ( ε )) .
Normalized modulation
Lemma 6.1 (Normalized one-bubble modulation). Under the hypotheses of Lemma 3.4 , let u ∈ S u \in\mathcal{S} u ∈ S be sufficiently close in H 1 ( M ) H^{1}(M) H 1 ( M ) to a normalized concentrating boundary bubble. Then there are locally unique parameters ( x , ε ) (x,\varepsilon) ( x , ε ) and w ∈ X x , ε w \in X_{x,\varepsilon} w ∈ X x , ε such that
u = ret U x , ε ( w ) , ∥ w ∥ H 1 ≤ C dist H 1 ( u , { U x , ε } ) . u=\operatorname{ret}_{U_{x,\varepsilon}}(w), \qquad\lVert w\rVert_{H^{1}} \leq C\operatorname{dist}_{H^{1}}\left(u,\{U_{x,\varepsilon}\}\right). u = ret U x , ε ( w ) , ∥ w ∥ H 1 ≤ C dist H 1 ( u , { U x , ε } ) . The parameters depend C 1 C^{1} C 1 on u u u in the local chart.
Proof. Choose a nearby base parameter ( x 0 , ε 0 ) (x^{0},\varepsilon^{0}) ( x 0 , ε 0 ) and use
ε ( s ) = e s ε 0 , x ( ξ ) = exp x 0 ∂ ( ε 0 ξ ) , Ξ ( s , ξ ) = U x ( ξ ) , ε ( s ) . \varepsilon(s)=e^{s}\varepsilon^{0}, \qquad x(\xi)=\exp_{x^{0}}^{\partial}\left(\varepsilon^{0}\xi\right), \qquad\Xi(s,\xi)=U_{x(\xi),\varepsilon(s)}. ε ( s ) = e s ε 0 , x ( ξ ) = exp x 0 ∂ ( ε 0 ξ ) , Ξ ( s , ξ ) = U x ( ξ ) , ε ( s ) .
On a fixed parameter ball, Fermi rescaling, the cutoff estimates, and the uniform Gram estimate in the proof of Lemma 3.4 give
∥ D Ξ ∥ + ∥ D 2 Ξ ∥ ≤ C , ∥ D Ξ ( 0 ) h ∥ H 1 ≥ c ∣ h ∣ . (79) \lVert D\Xi\rVert+\lVert D^{2}\Xi\rVert\leq C,\qquad\lVert D\Xi(0)h\rVert_{H^{1}}\geq c\lvert h\rvert. \tag*{(79)} ∥ D Ξ ∥ + ∥ D 2 Ξ ∥ ≤ C , ∥ D Ξ ( 0 ) h ∥ H 1 ≥ c ∣ h ∣ . (79)
For V , u ∈ S V,u\in\mathcal{S} V , u ∈ S sufficiently close, put
ν V ( u ) = u ∫ ∂ M ∣ V ∣ q − 2 V u d σ g − V . \nu_{V}(u)=\frac{u}{\int_{\partial M}\lvert V\rvert^{q-2}V u\,d\sigma_{g}}-V. ν V ( u ) = ∫ ∂ M ∣ V ∣ q − 2 V u d σ g u − V .
Then ν V ( u ) ∈ T V S \nu_{V}(u)\in T_{V}\mathcal{S} ν V ( u ) ∈ T V S and u = ret V ( ν V ( u ) ) u=\operatorname{ret}_{V}(\nu_{V}(u)) u = ret V ( ν V ( u )) . For the weighted modulation frame Z μ Z_{\mu} Z μ set
M μ ( u , z ) = ⟨ ν Ξ ( z ) ( u ) , Z μ ( z ) ⟩ E , g . \mathcal{M}_{\mu}(u,z)=\left\langle\nu_{\Xi(z)}(u),Z_{\mu}(z)\right\rangle_{E,g}. M μ ( u , z ) = ⟨ ν Ξ ( z ) ( u ) , Z μ ( z ) ⟩ E , g .
At u = Ξ ( 0 ) u=\Xi(0) u = Ξ ( 0 ) , the derivative with respect to z z z is, up to sign, the Gram matrix of the frame, uniformly invertible by (79) . The quantitative implicit-function theorem gives the decomposition, local uniqueness, and the displayed bound, exactly as in the proof of Lemma 3.4 . □ \square □
The one-bubble scale derivative
Remark 6.2 (Model rescaling and the complementary inverse). Write p = ( x , ε ) \mathbf{p}=(x,\varepsilon) p = ( x , ε ) and X + = X ♭ X_{+}=X^{\flat} X + = X ♭ , let P + ⊥ P_{+}^{\perp} P + ⊥ be the Dirichlet-orthogonal projection onto X + X_{+} X + , and put D = ε ∂ ε \mathbf{D}=\varepsilon\partial_{\varepsilon} D = ε ∂ ε at a fixed model coordinate Y Y Y . The notation D p h \mathbf{D}^{\mathrm{ph}} D ph denotes the logarithmic scale derivative at a fixed point of M M M . If V ~ ε ( Y ) = ε ( n − 2 ) / 2 V ε ( Φ x ( ε Y ) ) \widetilde{V}_{\varepsilon}(Y)=\varepsilon^{(n-2)/2}V_{\varepsilon}(\Phi_{x}(\varepsilon Y)) V ε ( Y ) = ε ( n − 2 ) /2 V ε ( Φ x ( ε Y )) , then
ε ( n − 2 ) / 2 ( D p h V ε ) ( Φ x ( ε Y ) ) = D V ~ ε − A V ~ ε , A = n − 2 2 + Y ⋅ ∇ . \varepsilon^{(n-2)/2}\left(\mathbf{D}^{\mathrm{ph}}V_{\varepsilon}\right)(\Phi_{x}(\varepsilon Y))=\mathbf{D}\widetilde{V}_{\varepsilon}-A\widetilde{V}_{\varepsilon},\qquad A=\frac{n-2}{2}+Y\cdot\nabla. ε ( n − 2 ) /2 ( D ph V ε ) ( Φ x ( ε Y )) = D V ε − A V ε , A = 2 n − 2 + Y ⋅ ∇.
Let T x , ε T_{x,\varepsilon} T x , ε denote critical Fermi rescaling, and define the bounded model map
B p z = P p T x , ε ( χ R ( ε ) z ) , z ∈ X + . \mathcal{B}_{\mathbf{p}}z=P_{\mathbf{p}}T_{x,\varepsilon}(\chi_{R(\varepsilon)}z),\qquad z\in X_{+}. B p z = P p T x , ε ( χ R ( ε ) z ) , z ∈ X + .
For R = ε − β R=\varepsilon^{-\beta} R = ε − β and z , A z ∈ D 1 , 2 z,Az\in D^{1,2} z , A z ∈ D 1 , 2 , direct differentiation gives
D p h T x , ε ( χ R z ) = T x , ε ( − χ R A z + ( β − 1 ) ( Y ⋅ ∇ χ R ) z ) . (80) \mathbf{D}^{\mathrm{ph}}T_{x,\varepsilon}(\chi_{R}z)=T_{x,\varepsilon}\left(-\chi_{R}Az+(\beta-1)(Y\cdot\nabla\chi_{R})z\right). \tag*{(80)} D ph T x , ε ( χ R z ) = T x , ε ( − χ R A z + ( β − 1 ) ( Y ⋅ ∇ χ R ) z ) . (80)
Hardy’s inequality makes the annular term tend to zero in the Dirichlet norm. For the β \beta β -cutoff family use R β ( ε ) R_{\beta}(\varepsilon) R β ( ε ) in the definition of B \mathcal{B} B . For each fixed z ∈ X + z\in X_{+} z ∈ X + , the energy norm of B p z \mathcal{B}_{\mathbf{p}}z B p z tends to ∥ ∇ z ∥ 2 \lVert\nabla z\rVert_{2} ∥ ∇ z ∥ 2 by coordinate comparison and the inverse Gram formula. The limiting complementary Hessian, with model directions rescaled without an amplitude factor, is
⟨ L 0 z , ϕ ⟩ = 2 B n ( z , ϕ ) , λ ♭ = S ∗ q / 2 . \left\langle\mathcal{L}_{0}z,\phi\right\rangle=2B_{n}(z,\phi),\qquad\lambda_{\flat}=S_{*}^{q/2}. ⟨ L 0 z , ϕ ⟩ = 2 B n ( z , ϕ ) , λ ♭ = S ∗ q /2 .
The hemisphere gap gives
B n ( z , z ) ≥ c ♭ ∥ ∇ z ∥ 2 2 , z ∈ X + . (81) B_{n}(z,z)\geq c_{\flat}\lVert\nabla z\rVert_{2}^{2},\qquad z\in X_{+}. \tag*{(81)} B n ( z , z ) ≥ c ♭ ∥ ∇ z ∥ 2 2 , z ∈ X + . (81)
We also use a complementary inverse on the fixed space H = H 1 ( M ) \mathcal{H}=H^{1}(M) H = H 1 ( M ) , identified with its dual by the energy form. If L p p h \mathcal{L}_{\mathbf{p}}^{\mathrm{ph}} L p ph is the complementary Hessian of J g ∘ ret U p \mathcal{J}_{g}\circ\operatorname{ret}_{U_{\mathbf{p}}} J g ∘ ret U p at zero, put
L p = P p L p p h P p + ( I − P p ) , G p = L p − 1 . (82) L_{\mathbf{p}}=P_{\mathbf{p}}\mathcal{L}_{\mathbf{p}}^{\mathrm{ph}}P_{\mathbf{p}}+(I-P_{\mathbf{p}}),\qquad G_{\mathbf{p}}=L_{\mathbf{p}}^{-1}. \tag*{(82)} L p = P p L p ph P p + ( I − P p ) , G p = L p − 1 . (82)
Uniform complementary coercivity and differentiation of the inverse give
∥ G p ∥ ≤ C , (83) \|\mathcal{G}_{p}\| \le C, \tag*{(83)} ∥ G p ∥ ≤ C , (83)
∂ θ G p = − G p ( ∂ θ L p ) G p . (84) \partial_{\theta}\mathcal{G}_{p}=-\mathcal{G}_{p}(\partial_{\theta}\mathcal{L}_{p})\mathcal{G}_{p}. \tag*{(84)} ∂ θ G p = − G p ( ∂ θ L p ) G p . (84)
Let S p p h ∈ X p \mathcal{S}_{p}^{\mathrm{ph}}\in X_{p} S p ph ∈ X p be the energy Riesz representative of the complementary differential. For a local identification T p : X 0 → X p \mathcal{T}_{p}:X_{0}\to X_{p} T p : X 0 → X p , put L p = T p ∗ L p p h T p \mathcal{L}_{p}=\mathcal{T}_{p}^{*}\mathcal{L}_{p}^{\mathrm{ph}}\mathcal{T}_{p} L p = T p ∗ L p ph T p and S p = T p ∗ S p p h \mathcal{S}_{p}=\mathcal{T}_{p}^{*}\mathcal{S}_{p}^{\mathrm{ph}} S p = T p ∗ S p ph . Then
⟨ L p − 1 S p , S p ⟩ = ⟨ G p S p p h , S p p h ⟩ . (85) \left\langle\mathcal{L}_{p}^{-1}\mathcal{S}_{p},\mathcal{S}_{p}\right\rangle
=\left\langle\mathcal{G}_{p}\mathcal{S}_{p}^{\mathrm{ph}},\mathcal{S}_{p}^{\mathrm{ph}}\right\rangle. \tag*{(85)} ⟨ L p − 1 S p , S p ⟩ = ⟨ G p S p ph , S p ph ⟩ . (85)
Lemma 6.3 (Differentiation after completing the square). Let a ε a_{\varepsilon} a ε be a C 1 C^{1} C 1 family of symmetric bilinear forms on a fixed Hilbert space, with ∥ a ε ∥ + ∥ D p h a ε ∥ ≤ C \|a_{\varepsilon}\|+\|\mathbf{D}^{\mathrm{ph}}a_{\varepsilon}\|\le C ∥ a ε ∥ + ∥ D ph a ε ∥ ≤ C , where D p h = ε ∂ ε \mathbf{D}^{\mathrm{ph}}=\varepsilon\partial_{\varepsilon} D ph = ε ∂ ε . Let X ε X_{\varepsilon} X ε be closed subspaces, and suppose C 1 C^{1} C 1 families v ε , V ε ∈ X ε v_{\varepsilon},V_{\varepsilon}\in X_{\varepsilon} v ε , V ε ∈ X ε satisfy
a ε ( v ε , h ) = s ε [ h ] ( h ∈ X ε ) , e ε = v ε − V ε . a_{\varepsilon}(v_{\varepsilon},h)=s_{\varepsilon}[h]\qquad(h\in X_{\varepsilon}),\qquad e_{\varepsilon}=v_{\varepsilon}-V_{\varepsilon}. a ε ( v ε , h ) = s ε [ h ] ( h ∈ X ε ) , e ε = v ε − V ε . Then
s ε [ v ε ] = 2 s ε [ V ε ] − a ε ( V ε , V ε ) + a ε ( e ε , e ε ) . (86) s_{\varepsilon}[v_{\varepsilon}]
=2s_{\varepsilon}[V_{\varepsilon}]
-a_{\varepsilon}(V_{\varepsilon},V_{\varepsilon})
+a_{\varepsilon}(e_{\varepsilon},e_{\varepsilon}). \tag*{(86)} s ε [ v ε ] = 2 s ε [ V ε ] − a ε ( V ε , V ε ) + a ε ( e ε , e ε ) . (86) If ∥ e ε ∥ = o ( ε k ) \|e_{\varepsilon}\|=o(\varepsilon^{k}) ∥ e ε ∥ = o ( ε k ) and ∥ D p h e ε ∥ = O ( ε k ) \|\mathbf{D}^{\mathrm{ph}}e_{\varepsilon}\|=O(\varepsilon^{k}) ∥ D ph e ε ∥ = O ( ε k ) , then the last term and its weighted derivative are o ( ε 2 k ) o(\varepsilon^{2k}) o ( ε 2 k ) .
Proof. Expand the last term and use the equation with h = v ε , V ε h=v_{\varepsilon},V_{\varepsilon} h = v ε , V ε . Moreover,
∣ D p h a ε ( e ε , e ε ) ∣ ≤ C ∥ e ε ∥ 2 + 2 C ∥ e ε ∥ ∥ D p h e ε ∥ = o ( ε 2 k ) . \left|\mathbf{D}^{\mathrm{ph}}a_{\varepsilon}(e_{\varepsilon},e_{\varepsilon})\right|
\le C\|e_{\varepsilon}\|^{2}
+2C\|e_{\varepsilon}\|\|\mathbf{D}^{\mathrm{ph}}e_{\varepsilon}\|
=o(\varepsilon^{2k}). D ph a ε ( e ε , e ε ) ≤ C ∥ e ε ∥ 2 + 2 C ∥ e ε ∥∥ D ph e ε ∥ = o ( ε 2 k ) .
These derivatives are taken in the fixed ambient Hilbert space on which the forms are defined. □
Proposition 6.4 (Scale derivative of the one-bubble reduced functional). Assume n ≥ 5 n\ge5 n ≥ 5 , ( Y ∂ + ) (Y_{\partial}^{+}) ( Y ∂ + ) , and ( B G 3 + ) (\mathrm{BG}^{+}_{3}) ( BG 3 + ) , and let C ⊂ ∂ M C\subset\partial M C ⊂ ∂ M be an open boundary region admitting a fixed Fermi collar, with H g = 0 H_{g}=0 H g = 0 on C C C . For
J 1 ( x , ε ) : = J g ( ret U x , ε ( w x , ε ) ) , x ∈ C , \mathcal{J}_{1}(x,\varepsilon):=\mathcal{J}_{g}\bigl(\operatorname{ret}_{\mathcal{U}_{x,\varepsilon}}(w_{x,\varepsilon})\bigr),
\qquad x\in C, J 1 ( x , ε ) := J g ( ret U x , ε ( w x , ε ) ) , x ∈ C , one has, uniformly on compact subsets of C C C ,
1 S ∗ ∂ ε J 1 ( x , ε ) = 2 R g r e d ( x ) ε + o ( ε ) . (87) \frac{1}{S_{*}}\partial_{\varepsilon}\mathcal{J}_{1}(x,\varepsilon)
=2\mathcal{R}_{g}^{\mathrm{red}}(x)\varepsilon+o(\varepsilon). \tag*{(87)} S ∗ 1 ∂ ε J 1 ( x , ε ) = 2 R g red ( x ) ε + o ( ε ) . (87)
Proof. Put α = ( n − 2 ) / 2 \alpha=(n-2)/2 α = ( n − 2 ) /2 , γ = q − 2 \gamma=q-2 γ = q − 2 , and U = U x , ε \mathcal{U}=\mathcal{U}_{x,\varepsilon} U = U x , ε , and use the first variation and Hessian restricted to the complement on M M M
S = D ( J g ∘ ret U ) ( 0 ) ∣ X x , ε , L = D 2 ( J g ∘ ret U ) ( 0 ) ∣ X x , ε . (88) \mathcal{S}=D(\mathcal{J}_{g}\circ\operatorname{ret}_{\mathcal{U}})(0)\big|_{X_{x,\varepsilon}},
\qquad
\mathcal{L}=D^{2}(\mathcal{J}_{g}\circ\operatorname{ret}_{\mathcal{U}})(0)\big|_{X_{x,\varepsilon}}. \tag*{(88)} S = D ( J g ∘ ret U ) ( 0 ) X x , ε , L = D 2 ( J g ∘ ret U ) ( 0 ) X x , ε . (88)
The Hessian on the complement is the restriction of the symmetric form
a x , ε ( h , ϕ ) = 2 N g ∘ ( h , ϕ ) − 2 ( q − 1 ) J g [ U ] ∫ ∂ M U q − 2 h ϕ d σ g . (89) a_{x,\varepsilon}(h,\phi)
=2\mathcal{N}_{g}^{\circ}(h,\phi)
-2(q-1)\mathcal{J}_{g}[\mathcal{U}]
\int_{\partial M}\mathcal{U}^{q-2}h\phi\,d\sigma_{g}. \tag*{(89)} a x , ε ( h , ϕ ) = 2 N g ∘ ( h , ϕ ) − 2 ( q − 1 ) J g [ U ] ∫ ∂ M U q − 2 h ϕ d σ g . (89)
The first model source is S 1 , x = t ∗ F 1 ( x ) = − 2 t ∗ I I ˚ ( x ) \mathcal{S}_{1,x}=t_{*}F_{1}(x)=-2t_{*}\mathring{\mathrm{II}}(x) S 1 , x = t ∗ F 1 ( x ) = − 2 t ∗ II ˚ ( x ) . Its degree-2 angular dependence gives the full moment cancellations
S 1 , x [ U + ] = S 1 , x [ Z j b ] = 0 ( 0 ≤ j ≤ n − 1 ) . (90) \mathcal{S}_{1,x}[U_{+}]=\mathcal{S}_{1,x}[Z_{j}^{b}]=0
\qquad(0\le j\le n-1). \tag*{(90)} S 1 , x [ U + ] = S 1 , x [ Z j b ] = 0 ( 0 ≤ j ≤ n − 1 ) . (90)
The differentiated standard-bubble expansion gives
S ∗ − 1 ∂ ε J g [ U ] = 2 R g b a r e ( x ) ε + o ( ε ) . (91) S_{*}^{-1}\partial_{\varepsilon}\mathcal{J}_{g}[\mathcal{U}]
=2\mathcal{R}_{g}^{\mathrm{bare}}(x)\varepsilon+o(\varepsilon). \tag*{(91)} S ∗ − 1 ∂ ε J g [ U ] = 2 R g bare ( x ) ε + o ( ε ) . (91)
This is (14) ; its proof estimates the moving cutoff and the geometric remainder before differentiating.
Let v = L − 1 S v=\mathcal{L}^{-1}\mathcal{S} v = L − 1 S and z 1 , x = L 0 − 1 S 1 , x z_{1,x}=\mathcal{L}_{0}^{-1}\mathcal{S}_{1,x} z 1 , x = L 0 − 1 S 1 , x . Write S f u l l = D J g [ U ] \mathcal{S}^{\mathrm{full}}=D\mathcal{J}_{g}[\mathcal{U}] S full = D J g [ U ] on H 1 ( M ) H^{1}(M) H 1 ( M ) , whose restriction is S \mathcal{S} S . We first estimate derivatives at a fixed point of M M M . If h ~ ε ( Y ) = ε α h ( Φ x ( ε Y ) ) \widetilde{h}_{\varepsilon}(Y)=\varepsilon^{\alpha}h(\Phi_{x}(\varepsilon Y)) h ε ( Y ) = ε α h ( Φ x ( ε Y )) and s ε [ h ] = ε k ∫ f ( Y ) h ~ ε ( Y ) d Y s_{\varepsilon}[h]=\varepsilon^{k}\int f(Y)\widetilde{h}_{\varepsilon}(Y)\,dY s ε [ h ] = ε k ∫ f ( Y ) h ε ( Y ) d Y , then integration by parts gives
D p h s ε [ h ] = ε k ∫ ( ( k + α − n ) f − Y ⋅ ∇ f ) h ~ ε d Y . (92) \mathbf{D}^{\mathrm{ph}}s_{\varepsilon}[h]
=\varepsilon^{k}\int\bigl((k+\alpha-n)f-Y\cdot\nabla f\bigr)\widetilde{h}_{\varepsilon}\,dY. \tag*{(92)} D ph s ε [ h ] = ε k ∫ ( ( k + α − n ) f − Y ⋅ ∇ f ) h ε d Y . (92)
The boundary contribution is zero because Y n = 0 Y_n=0 Y n = 0 there. For the first source, tangential integration by parts gives a scalar forcing f f f with ∣ f ∣ + ∣ Y ⋅ ∇ f ∣ ≤ C ( 1 + ∣ Y ∣ ) 1 − n |f|+|Y\cdot\nabla f|\le C(1+|Y|)^{1-n} ∣ f ∣ + ∣ Y ⋅ ∇ f ∣ ≤ C ( 1 + ∣ Y ∣ ) 1 − n . Hardy’s inequality applies since
∫ ∣ Y ∣ > 1 ∣ Y ∣ 2 ( ∣ f ∣ 2 + ∣ Y ⋅ ∇ f ∣ 2 ) d Y ≤ C ∫ 1 ∞ r 3 − 2 n d r < ∞ . \int_{|Y|>1}|Y|^2\left(|f|^2+|Y\cdot\nabla f|^2\right)\,dY\le C\int_1^\infty r^{3-2n}\,dr<\infty. ∫ ∣ Y ∣ > 1 ∣ Y ∣ 2 ( ∣ f ∣ 2 + ∣ Y ⋅ ∇ f ∣ 2 ) d Y ≤ C ∫ 1 ∞ r 3 − 2 n d r < ∞.
For the explicit metric and cutoff errors, the same differentiation and integration by parts preserve the bounds C ( η n ( ε ) + R − α ) C(\eta_n(\varepsilon)+R^{-\alpha}) C ( η n ( ε ) + R − α ) established in Lemma 5.2 . Derivatives of normalization and quotient factors are estimated by their defining integrals. Hence
∥ S f u l l ∥ + ∥ D p h S f u l l ∥ ≤ C ε , ∥ a x , ε ∥ + ∥ D p h a x , ε ∥ ≤ C . (93) \|\mathcal{S}^{\mathrm{full}}\|+\|\mathbf{D}^{\mathrm{ph}}\mathcal{S}^{\mathrm{full}}\|\le C\varepsilon,\qquad\|a_{x,\varepsilon}\|+\|\mathbf{D}^{\mathrm{ph}}a_{x,\varepsilon}\|\le C. \tag*{(93)} ∥ S full ∥ + ∥ D ph S full ∥ ≤ C ε , ∥ a x , ε ∥ + ∥ D ph a x , ε ∥ ≤ C . (93)
For the form estimate, only the boundary potential V ε = J g [ U ] U γ V_\varepsilon=\mathcal{J}_g[U]U^\gamma V ε = J g [ U ] U γ varies. Off the cutoff annulus its logarithmic derivative is bounded by C U γ CU^\gamma C U γ ; on the annulus, χ R = n R 2 / r R \chi_R=n_R^2/r_R χ R = n R 2 / r R gives the O ( R − 1 ) O(R^{-1}) O ( R − 1 ) coefficient bound in Convention 2.1. Apply (12) and D J g [ U ] = O ( ε 2 ) D\mathcal{J}_g[U]=O(\varepsilon^2) D J g [ U ] = O ( ε 2 ) from Proposition 2.5.
Let Z Z Z denote the map whose columns are the normal frame in Remark 5.5, and G = Z ∗ Z G=Z^*Z G = Z ∗ Z . Then
P = I − Z G − 1 Z ∗ , D p h G − 1 = − G − 1 ( D p h G ) G − 1 , ∥ D p h P ∥ ≤ C . P=I-ZG^{-1}Z^*,\qquad D^{\mathrm{ph}}G^{-1}=-G^{-1}(D^{\mathrm{ph}}G)G^{-1},\qquad\|D^{\mathrm{ph}}P\|\le C. P = I − Z G − 1 Z ∗ , D ph G − 1 = − G − 1 ( D ph G ) G − 1 , ∥ D ph P ∥ ≤ C .
The frame and its derivative are bounded by the explicit Jacobi fields and the exact constraint-normal identity from that remark, using (93) . The stabilized operator (82) consequently satisfies ∥ L ∥ + ∥ D p h L ∥ ≤ C \|\mathcal{L}\|+\|D^{\mathrm{ph}}\mathcal{L}\|\le C ∥ L ∥ + ∥ D ph L ∥ ≤ C . The derivative of its inverse is given by (84) , so
∥ v ∥ H 1 + ∥ D p h v ∥ H 1 ≤ C ε . (94) \|v\|_{H^1}+\|D^{\mathrm{ph}}v\|_{H^1}\le C\varepsilon. \tag*{(94)} ∥ v ∥ H 1 + ∥ D ph v ∥ H 1 ≤ C ε . (94)
Complementary coordinate derivatives obey the same bounds after the local identification of Remark 5.6.
For the solution z 1 , x z_{1,x} z 1 , x of the model linearized equation, the interior forcing f 1 , x f_{1,x} f 1 , x satisfies
∣ f 1 , x ( Y ) ∣ ≤ C ( 1 + ∣ Y ∣ ) 1 − n , ∫ ∣ Y ∣ > 1 ∣ Y ∣ 2 ∣ f 1 , x ∣ 2 d Y < ∞ , |f_{1,x}(Y)|\le C(1+|Y|)^{1-n},\qquad\int_{|Y|>1}|Y|^2|f_{1,x}|^2\,dY<\infty, ∣ f 1 , x ( Y ) ∣ ≤ C ( 1 + ∣ Y ∣ ) 1 − n , ∫ ∣ Y ∣ > 1 ∣ Y ∣ 2 ∣ f 1 , x ∣ 2 d Y < ∞ ,
and it satisfies a homogeneous Robin boundary condition by angular orthogonality. Write A r = B 2 r + ∖ B r + ‾ A_r=B^+_{2r}\setminus\overline{B^+_r} A r = B 2 r + ∖ B r + and ∂ ′ A = A ∩ ∂ R + n \partial' A=A\cap\partial\mathbb{R}^n_+ ∂ ′ A = A ∩ ∂ R + n . The unit half-annulus H 2 H^2 H 2 estimate, after dilation, gives for interior forcing f f f and boundary forcing g g g
r ∥ ∇ 2 h ∥ L 2 ( A r ) ≤ C ( ∥ ∇ h ∥ L 2 ( A r / 2 , 4 r ) + r − 1 ∥ h ∥ L 2 ( A r / 2 , 4 r ) + r ∥ f ∥ L 2 ( A r / 2 , 4 r ) + r n / 2 ∥ g ( r ⋅ ) ∥ H 1 / 2 ( ∂ ′ A 1 / 2 , 4 ) ) . (95) r\|\nabla^2h\|_{L^2(A_r)}\le C\left(\|\nabla h\|_{L^2(A_{r/2,4r})}+r^{-1}\|h\|_{L^2(A_{r/2,4r})}+r\|f\|_{L^2(A_{r/2,4r})}+r^{n/2}\|g(r\cdot)\|_{H^{1/2}(\partial' A_{1/2,4})}\right). \tag*{(95)} r ∥ ∇ 2 h ∥ L 2 ( A r ) ≤ C ( ∥∇ h ∥ L 2 ( A r /2 , 4 r ) + r − 1 ∥ h ∥ L 2 ( A r /2 , 4 r ) + r ∥ f ∥ L 2 ( A r /2 , 4 r ) + r n /2 ∥ g ( r ⋅ ) ∥ H 1/2 ( ∂ ′ A 1/2 , 4 ) ) . (95)
where A a , b = B b + ∖ B a + ‾ A_{a,b}=B^+_b\setminus\overline{B^+_a} A a , b = B b + ∖ B a + . To obtain this estimate, scale to the unit half-annulus and use tangential difference quotients in the weak equation with a compactly supported cutoff. Trace interpolation absorbs the terms containing the Robin coefficient; the remaining terms are controlled by the displayed H 1 H^1 H 1 , forcing, and boundary norms. The equation then recovers the second normal derivative. The rescaled Robin coefficient n r / ( 1 + r 2 ∣ Y ′ ∣ 2 ) nr/(1+r^2|Y'|^2) n r / ( 1 + r 2 ∣ Y ′ ∣ 2 ) and its first derivatives are uniformly bounded on the boundary half-annulus, so the constant is independent of r ≥ 1 r\ge1 r ≥ 1 . Squaring and summing over dyadic annuli, whose enlarged versions have bounded overlap, and using Hardy’s inequality proves Y ∇ 2 z 1 , x ∈ L 2 Y\nabla^2z_{1,x}\in L^2 Y ∇ 2 z 1 , x ∈ L 2 , and hence A z 1 , x ∈ D 1 , 2 \mathcal{A}z_{1,x}\in\mathcal{D}^{1,2} A z 1 , x ∈ D 1 , 2 . This estimate is uniform in x x x on the stated compact sets.
Set
V ε = ε B x , ε z 1 , x , C x = ⟨ S 1 , x , z 1 , x ⟩ . V_\varepsilon=\varepsilon\mathcal{B}_{x,\varepsilon}z_{1,x},\qquad C_x=\langle\mathcal{S}_{1,x},z_{1,x}\rangle. V ε = ε B x , ε z 1 , x , C x = ⟨ S 1 , x , z 1 , x ⟩ .
The pairings satisfy
S [ V ε ] = ε 2 C x + o ( ε 2 ) , a x , ε ( V ε , V ε ) = ε 2 C x + o ( ε 2 ) . (96) \mathcal{S}[V_\varepsilon]=\varepsilon^2C_x+o(\varepsilon^2),\qquad a_{x,\varepsilon}(V_\varepsilon,V_\varepsilon)=\varepsilon^2C_x+o(\varepsilon^2). \tag*{(96)} S [ V ε ] = ε 2 C x + o ( ε 2 ) , a x , ε ( V ε , V ε ) = ε 2 C x + o ( ε 2 ) . (96)
with weighted derivatives of both remainders o ( ε 2 ) o(\varepsilon^2) o ( ε 2 ) , and ∥ D p h V ε ∥ H 1 = O ( ε ) \lVert D^{\mathrm{ph}}V_\varepsilon\rVert_{H^1}=O(\varepsilon) ∥ D ph V ε ∥ H 1 = O ( ε ) . For the source pairing, rescale to the fixed model and differentiate its explicit coefficients. The normalized errors are bounded by
C ( η n ( ε ) + ω ( R ) ε + R − ( n − 4 ) / 2 ) = o ( 1 ) . C\left(\frac{\eta_n(\varepsilon)+\omega(R)}{\varepsilon}+R^{-(n-4)/2}\right)=o(1). C ( ε η n ( ε ) + ω ( R ) + R − ( n − 4 ) /2 ) = o ( 1 ) .
For the form pairing, the principal and scalar errors are O ( ε R ) O(\varepsilon R) O ( εR ) and O ( ( ε R ) 2 ) O((\varepsilon R)^2) O (( εR ) 2 ) , including their logarithmic derivatives. In model coordinates the boundary potential tends to V 0 = λ f l a t U + γ V_0=\lambda_{\mathrm{flat}}U_+^\gamma V 0 = λ flat U + γ . Its logarithmic scale derivative at fixed points of M M M , expressed in rescaled coordinates, tends to − V 0 − Y ′ ⋅ ∇ V 0 -V_0-Y'\cdot\nabla V_0 − V 0 − Y ′ ⋅ ∇ V 0 . Including the derivatives of both transported test functions gives the leading sum
( − V 0 − Y ′ ⋅ ∇ V 0 ) z 2 − 2 V 0 z ( α z + Y ′ ⋅ ∇ z ) = − div Y ′ ( Y ′ V 0 z 2 ) , (-V_0-Y'\cdot\nabla V_0)z^2-2V_0z(\alpha z+Y'\cdot\nabla z)=-\operatorname{div}Y'(Y'V_0z^2), ( − V 0 − Y ′ ⋅ ∇ V 0 ) z 2 − 2 V 0 z ( α z + Y ′ ⋅ ∇ z ) = − div Y ′ ( Y ′ V 0 z 2 ) ,
whose integral is zero, first for compactly supported tests and then by approximation for z , A z ∈ D 1 , 2 z,Az\in\mathcal{D}^{1,2} z , A z ∈ D 1 , 2 . The analogous Dirichlet terms cancel by dilation invariance. The coefficient errors and cutoff tails therefore tend to zero also in the differentiated fixed-profile pairing.
These assertions are unchanged by the finite-dimensional projection in B \mathcal{B} B . To see the differentiated statement, let T ε z = T x , ε ( χ R z ) T_\varepsilon z=T_{x,\varepsilon}(\chi_Rz) T ε z = T x , ε ( χ R z ) and use a normal frame ζ j \zeta_j ζ j in the inverse Gram formula. Its rescaled limit is the flat normal frame, and the logarithmic scale derivative at fixed points of M M M , expressed in rescaled coordinates, tends to − A ζ j ♭ -A\zeta_j^\flat − A ζ j ♭ . The energy Riesz representative n ε n_\varepsilon n ε of h ↦ ∫ U q − 1 h h\mapsto\int\mathcal{U}^{q-1}h h ↦ ∫ U q − 1 h satisfies the exact identity
n ε = U J g [ U ] − R ε P f u l l 2 J g [ U ] . n_\varepsilon=\frac{U}{\mathcal{J}_g[U]}-\frac{R_\varepsilon\mathcal{P}^{\mathrm{full}}}{2\mathcal{J}_g[U]}. n ε = J g [ U ] U − 2 J g [ U ] R ε P full .
By (93) and (91) , n ε − U / J g [ U ] = O H 1 ( ε ) n_\varepsilon-U/\mathcal{J}_g[U]=O_{H^1}(\varepsilon) n ε − U / J g [ U ] = O H 1 ( ε ) and D p h n ε − ( D p h U ) / J g [ U ] = O H 1 ( ε ) D^{\mathrm{ph}}n_\varepsilon-(D^{\mathrm{ph}}U)/\mathcal{J}_g[U]=O_{H^1}(\varepsilon) D ph n ε − ( D ph U ) / J g [ U ] = O H 1 ( ε ) . This proves the differentiated limit of the constraint normal; the remaining frame elements are explicit Jacobi fields. The projection coefficients of T ε z 1 , x T_\varepsilon z_{1,x} T ε z 1 , x tend to zero. Their differentiated limits also vanish, since critical dilation is a Dirichlet isometry:
∫ ∇ ( A z 1 , x ) ⋅ ∇ ζ j ♭ + ∫ ∇ z 1 , x ⋅ ∇ ( A ζ j ♭ ) = 0. \int\nabla(Az_{1,x})\cdot\nabla\zeta_j^\flat+\int\nabla z_{1,x}\cdot\nabla(A\zeta_j^\flat)=0. ∫ ∇ ( A z 1 , x ) ⋅ ∇ ζ j ♭ + ∫ ∇ z 1 , x ⋅ ∇ ( A ζ j ♭ ) = 0.
Thus the removed projection and its logarithmic scale derivative at fixed points of M M M are o H 1 ( 1 ) o_{H^1}(1) o H 1 ( 1 ) . This proves all the assertions in (96) .
Lemma 5.11 gives S [ v ] = ε 2 C x + o ( ε 2 ) \mathcal{S}[v]=\varepsilon^2C_x+o(\varepsilon^2) S [ v ] = ε 2 C x + o ( ε 2 ) . Applying (86) with the trial function V ε V_\varepsilon V ε , and using complementary coercivity, therefore gives
∥ v − V ε ∥ H 1 = o ( ε ) . \lVert v-V_\varepsilon\rVert_{H^1}=o(\varepsilon). ∥ v − V ε ∥ H 1 = o ( ε ) .
The logarithmic scale derivatives of both terms at fixed points of M M M are O ( ε ) O(\varepsilon) O ( ε ) by the preceding estimates. Lemma 6.3 now yields
⟨ L − 1 S , S ⟩ = ε 2 C x + o ( ε 2 ) , D ⟨ L − 1 S , S ⟩ = 2 ε 2 C x + o ( ε 2 ) . (97) \langle\mathcal{L}^{-1}\mathcal{S},\mathcal{S}\rangle=\varepsilon^2C_x+o(\varepsilon^2),\qquad\mathbf{D}\langle\mathcal{L}^{-1}\mathcal{S},\mathcal{S}\rangle=2\varepsilon^2C_x+o(\varepsilon^2). \tag*{(97)} ⟨ L − 1 S , S ⟩ = ε 2 C x + o ( ε 2 ) , D ⟨ L − 1 S , S ⟩ = 2 ε 2 C x + o ( ε 2 ) . (97)
It remains to differentiate the nonlinear remainder. Put γ = q − 2 \gamma=q-2 γ = q − 2 and, for a ≥ 0 a\geq0 a ≥ 0 , define
N ( a , w ) = ∣ a + w ∣ γ ( a + w ) − a 1 + γ − ( 1 + γ ) a γ w , N(a,w)=|a+w|^\gamma(a+w)-a^{1+\gamma}-(1+\gamma)a^\gamma w, N ( a , w ) = ∣ a + w ∣ γ ( a + w ) − a 1 + γ − ( 1 + γ ) a γ w ,
Q ( a , w ) = ∣ a + w ∣ 2 + γ − a 2 + γ − ( 2 + γ ) a 1 + γ w − ( 2 + γ ) ( 1 + γ ) 2 a γ w 2 . Q(a,w)=|a+w|^{2+\gamma}-a^{2+\gamma}-(2+\gamma)a^{1+\gamma}w-\frac{(2+\gamma)(1+\gamma)}{2}a^\gamma w^2. Q ( a , w ) = ∣ a + w ∣ 2 + γ − a 2 + γ − ( 2 + γ ) a 1 + γ w − 2 ( 2 + γ ) ( 1 + γ ) a γ w 2 .
Splitting into the cases ∣ w ∣ ≤ a / 2 |w|\leq a/2 ∣ w ∣ ≤ a /2 and ∣ w ∣ > a / 2 |w|>a/2 ∣ w ∣ > a /2 gives
∣ ∂ a N ( a , w ) ∣ ≤ C ∣ w ∣ 1 + γ , ∣ ∂ a Q ( a , w ) ∣ ≤ C ∣ w ∣ 2 + γ , |\partial_aN(a,w)|\leq C|w|^{1+\gamma},\qquad|\partial_aQ(a,w)|\leq C|w|^{2+\gamma}, ∣ ∂ a N ( a , w ) ∣ ≤ C ∣ w ∣ 1 + γ , ∣ ∂ a Q ( a , w ) ∣ ≤ C ∣ w ∣ 2 + γ ,
with continuous extension at a = 0 a=0 a = 0 . The explicit bubble derivative has the form D p h U = b ( Y ) U + A a n n \mathbf{D}^{\mathrm{ph}}U=b(Y)U+\mathcal{A}^{\mathrm{ann}} D ph U = b ( Y ) U + A ann , where b b b is bounded and the annular term has critical norm O ( R − α ) O(R^{-\alpha}) O ( R − α ) , α = ( n − 2 ) / 2 \alpha=(n-2)/2 α = ( n − 2 ) /2 . On the part of the annulus where η > 0 \eta>0 η > 0 , the coefficient satisfies
( η n − 2 U + ) γ − 1 ( n − 2 ) η n − 3 ( D η ) U + = ( n − 2 ) η ( D η ) U + γ . (98) (\eta^{n-2}U_+)^{\gamma-1}(n-2)\eta^{n-3}(D\eta)U_+=(n-2)\eta(D\eta)U_+^\gamma. \tag*{(98)} ( η n − 2 U + ) γ − 1 ( n − 2 ) η n − 3 ( D η ) U + = ( n − 2 ) η ( D η ) U + γ . (98)
The right-hand side extends continuously across η = 0 \eta= 0 η = 0 , and its L n − 1 L^{n-1} L n − 1 norm is O ( R − 1 ) O(R^{-1}) O ( R − 1 ) . The remaining annular term is estimated by Hölder’s inequality using ∥ A a n n ∥ q = O ( R − α ) \|A^{\mathrm{ann}}\|_{q} = O(R^{-\alpha}) ∥ A ann ∥ q = O ( R − α ) . Thus the parameter derivative of the nonlinear gradient at fixed w w w is bounded by
C ( ∥ w ∥ 1 + γ + R − α ∥ w ∥ γ + R − 1 ∥ w ∥ ) . (99) C\left(\|w\|^{1+\gamma}+R^{-\alpha}\|w\|^{\gamma}+R^{-1}\|w\|\right). \tag*{(99)} C ( ∥ w ∥ 1 + γ + R − α ∥ w ∥ γ + R − 1 ∥ w ∥ ) . (99)
Differentiating the complementary equation and using uniform inversion gives ∥ D p h w ∥ H 1 = O ( ε ) \|D^{\mathrm{ph}}w\|_{H^{1}} = O(\varepsilon) ∥ D ph w ∥ H 1 = O ( ε ) . Stationarity and the preceding pointwise estimates then give
J 1 = J g [ U ] − 1 2 ⟨ L − 1 S , S ⟩ + E , ∣ E ∣ + ∣ D E ∣ ≤ C ( ε q + R − α ε q − 1 + R − 1 ε 2 ) = o ( ε 2 ) . (100) \mathcal{J}_{1}=\mathcal{J}_{g}[U]-\frac{1}{2}\langle\mathcal{L}^{-1}\mathcal{S},\mathcal{S}\rangle+\mathcal{E},\qquad|\mathcal{E}|+|D\mathcal{E}|\leq C\left(\varepsilon^{q}+R^{-\alpha}\varepsilon^{q-1}+R^{-1}\varepsilon^{2}\right)=o(\varepsilon^{2}). \tag*{(100)} J 1 = J g [ U ] − 2 1 ⟨ L − 1 S , S ⟩ + E , ∣ E ∣ + ∣ D E ∣ ≤ C ( ε q + R − α ε q − 1 + R − 1 ε 2 ) = o ( ε 2 ) . (100)
Here 3 ( n − 2 ) / 8 + q − 1 > 2 3(n-2)/8+q-1>2 3 ( n − 2 ) /8 + q − 1 > 2 for every n ≥ 5 n\geq5 n ≥ 5 . Finally, the source normalization and (56) give
C x = 2 S ∗ δ κ L S ( n ) ∣ I I ˚ ( x ) ∣ 2 . C_{x}=2\mathcal{S}_{*}\delta_{\kappa}\mathcal{L}_{S}^{(n)}|\mathring{\mathrm{II}}(x)|^{2}. C x = 2 S ∗ δ κ L S ( n ) ∣ II ˚ ( x ) ∣ 2 .
Differentiate the last expansion, apply (97) and (91) , and subtract this correction from the standard-bubble coefficient. This proves (87) . □
Quantitative stability near a boundary bubble
Write
r g ( u ) = ∥ grad S J g ( u ) ∥ E , g , B 0 = { σ U x , ε : σ ∈ { ± 1 } } , B L S = { σ Φ ( x , ε ) : σ ∈ { ± 1 } } , r_{g}(u)=\|\operatorname{grad}_{S}\mathcal{J}_{g}(u)\|_{E,g},\qquad\mathcal{B}_{0}=\{\sigma U_{x,\varepsilon}:\sigma\in\{\pm1\}\},\qquad\mathcal{B}_{\mathrm{LS}}=\{\sigma\Phi(x,\varepsilon):\sigma\in\{\pm1\}\}, r g ( u ) = ∥ grad S J g ( u ) ∥ E , g , B 0 = { σ U x , ε : σ ∈ { ± 1 }} , B LS = { σ Φ ( x , ε ) : σ ∈ { ± 1 }} ,
where the scales are sufficiently small and the centers lie in the boundary region under consideration. Distances in this subsection are measured in the energy norm. For p = ( x , ε ) p=(x,\varepsilon) p = ( x , ε ) put
F p ( v ) = J g ( ret U p ( v ) ) , v ∈ X p . F_{p}(v)=\mathcal{J}_{g}(\operatorname{ret}_{U_{p}}(v)),\qquad v\in X_{p}. F p ( v ) = J g ( ret U p ( v )) , v ∈ X p .
The correction w p w_{p} w p of Lemma 5.9 satisfies D X F p ( w p ) = 0 D_{X}F_{p}(w_{p})=0 D X F p ( w p ) = 0 .
Lemma 6.5 (Complementary estimates and parameter derivatives). There are uniform constants c , C , ρ > 0 c,C,\rho>0 c , C , ρ > 0 such that, for sufficiently small scales and v , w p ∈ X p v,w_{p}\in X_{p} v , w p ∈ X p with ∥ v ∥ E , ∥ w p ∥ E < ρ \|v\|_{E},\|w_{p}\|_{E}<\rho ∥ v ∥ E , ∥ w p ∥ E < ρ , setting u = ret U p ( v ) u=\operatorname{ret}_{U_{p}}(v) u = ret U p ( v ) and η = v − w p \eta=v-w_{p} η = v − w p gives
∥ η ∥ E + ∥ u − Φ ( p ) ∥ E ≤ C r g ( u ) , r g ( Φ ( p ) ) ≤ C r g ( u ) , (101) \|\eta\|_{E}+\|u-\Phi(p)\|_{E}\leq Cr_{g}(u),\qquad r_{g}(\Phi(p))\leq Cr_{g}(u), \tag*{(101)} ∥ η ∥ E + ∥ u − Φ ( p ) ∥ E ≤ C r g ( u ) , r g ( Φ ( p )) ≤ C r g ( u ) , (101) c ∥ η ∥ E 2 ≤ F p ( v ) − J 1 ( p ) ≤ C ∥ η ∥ E 2 . (102) c\|\eta\|_{E}^{2}\leq F_{p}(v)-\mathcal{J}_{1}(p)\leq C\|\eta\|_{E}^{2}. \tag*{(102)} c ∥ η ∥ E 2 ≤ F p ( v ) − J 1 ( p ) ≤ C ∥ η ∥ E 2 . (102) Moreover, in local boundary coordinates,
r g ( Φ ( p ) ) ≍ ( ∣ ε ∂ ε J 1 ( p ) ∣ 2 + ε 2 ∣ ∇ ∂ J 1 ( p ) ∣ 2 ) 1 / 2 . (103) r_{g}(\Phi(p))\asymp\left(|\varepsilon\partial_{\varepsilon}\mathcal{J}_{1}(p)|^{2}+\varepsilon^{2}|\nabla_{\partial}\mathcal{J}_{1}(p)|^{2}\right)^{1/2}. \tag*{(103)} r g ( Φ ( p )) ≍ ( ∣ ε ∂ ε J 1 ( p ) ∣ 2 + ε 2 ∣ ∇ ∂ J 1 ( p ) ∣ 2 ) 1/2 . (103) The constants are uniform on compact subsets of the boundary region. The same assertions hold after multiplication by − 1 -1 − 1 .
Proof. The complementary Hessian is uniformly positive by Lemma 5.9 . Its uniform continuity on a fixed small ball gives
c ∥ h ∥ E 2 ≤ D 2 F p ( z ) [ h , h ] ≤ C ∥ h ∥ E 2 ( z , h ∈ X p , ∥ z ∥ E < 2 ρ ) . c\|h\|_{E}^{2}\leq D^{2}F_{p}(z)[h,h]\leq C\|h\|_{E}^{2}\qquad(z,h\in X_{p},\ \|z\|_{E}<2\rho). c ∥ h ∥ E 2 ≤ D 2 F p ( z ) [ h , h ] ≤ C ∥ h ∥ E 2 ( z , h ∈ X p , ∥ z ∥ E < 2 ρ ) .
Integrating along the segment from w p w_{p} w p to v v v and using D X F p ( w p ) = 0 D_{X}F_{p}(w_{p})=0 D X F p ( w p ) = 0 , we obtain
c ∥ η ∥ E 2 ≤ D X F p ( v ) [ η ] ≤ C r g ( u ) ∥ η ∥ E . c\|\eta\|_{E}^{2}\leq D_{X}F_{p}(v)[\eta]\leq Cr_{g}(u)\|\eta\|_{E}. c ∥ η ∥ E 2 ≤ D X F p ( v ) [ η ] ≤ C r g ( u ) ∥ η ∥ E .
The normalization map has uniformly bounded first and second derivatives here. This proves the first estimate in (101) . The identity
F p ( v ) − F p ( w p ) = ∫ 0 1 ( 1 − s ) D 2 F p ( w p + s η ) [ η , η ] d s F_{p}(v)-F_{p}(w_{p})=\int_{0}^{1}(1-s)D^{2}F_{p}(w_{p}+s\eta)[\eta,\eta]\,ds F p ( v ) − F p ( w p ) = ∫ 0 1 ( 1 − s ) D 2 F p ( w p + sη ) [ η , η ] d s
proves both bounds in (102) . For the gradient estimate let n ( u ) = R E ℓ u n(u)=R_{E}\ell_{u} n ( u ) = R E ℓ u , with ℓ u \ell_{u} ℓ u as in (43) . Since ⟨ n ( u ) , u ⟩ E = 1 \langle n(u),u\rangle_{E}=1 ⟨ n ( u ) , u ⟩ E = 1 , one has ∥ n ( u ) ∥ E ≥ ∥ u ∥ E − 1 \|n(u)\|_{E}\geq\|u\|_{E}^{-1} ∥ n ( u ) ∥ E ≥ ∥ u ∥ E − 1 , and the constrained gradient is exactly
grad S J g ( u ) = 2 ( u − n ( u ) ∥ n ( u ) ∥ E 2 ) . \operatorname{grad}_{S}\mathcal{J}_{g}(u)=2\left(u-\frac{n(u)}{\|n(u)\|_{E}^{2}}\right). grad S J g ( u ) = 2 ( u − ∥ n ( u ) ∥ E 2 n ( u ) ) .
The trace inequality bounds D n D_n D n on the energy-bounded neighborhood. This formula therefore makes the constrained gradient uniformly Lipschitz. Applying this to u u u and Φ ( p ) \Phi(p) Φ ( p ) proves the remaining estimate in (101) .
For (103) , let N ( z ) = z / ∥ z ∥ q N(z)=z/\lVert z\rVert_q N ( z ) = z / ∥ z ∥ q and
E p = D N ( U p + w p ) X p ⊂ T Φ ( p ) S . E_p=DN(\mathcal{U}_p+w_p)X_p\subset T_{\Phi(p)}S. E p = D N ( U p + w p ) X p ⊂ T Φ ( p ) S .
Complementary stationarity says that the constrained gradient at Φ ( p ) \Phi(p) Φ ( p ) is orthogonal to E p E_p E p . This subspace has codimension n n n in T Φ ( p ) S T_{\Phi(p)}S T Φ ( p ) S . Let P p P_p P p be the energy projection onto X p X_p X p , and let D 0 = ε ∂ ε D_0=\varepsilon\partial_\varepsilon D 0 = ε ∂ ε and D a = ε ∂ x a D_a=\varepsilon\partial_{x_a} D a = ε ∂ x a . The normalized modulation estimates and the inverse Gram formula give ∥ D α P p ∥ ≤ C \lVert D_\alpha P_p\rVert\le C ∥ D α P p ∥ ≤ C . Differentiating P p w p = w p P_pw_p=w_p P p w p = w p yields
( I − P p ) D α w p = ( D α P p ) w p = O E ( ε ) . (I-P_p)D_\alpha w_p=(D_\alpha P_p)w_p=O_E(\varepsilon). ( I − P p ) D α w p = ( D α P p ) w p = O E ( ε ) .
The term D N ( U p + w p ) P p D α w p DN(\mathcal{U}_p+w_p)P_pD_\alpha w_p D N ( U p + w p ) P p D α w p belongs to E p E_p E p . Consequently the projection of D α Φ ( p ) D_\alpha\Phi(p) D α Φ ( p ) onto E p ⊥ ∩ T Φ ( p ) S E_p^\perp\cap T_{\Phi(p)}S E p ⊥ ∩ T Φ ( p ) S equals the projection of
D N ( U p + w p ) Z α + O E ( ε ) . DN(\mathcal{U}_p+w_p)Z_\alpha+O_E(\varepsilon). D N ( U p + w p ) Z α + O E ( ε ) .
Since ∥ w p ∥ E = O ( ε ) \lVert w_p\rVert_E=O(\varepsilon) ∥ w p ∥ E = O ( ε ) , these projected vectors differ by O E ( ε ) O_E(\varepsilon) O E ( ε ) from the normalized bubble frame. Its Gram matrix has a uniformly positive flat limit by Lemma 6.1. They therefore form a uniformly nondegenerate basis of E p ⊥ ∩ T Φ ( p ) S E_p^\perp\cap T_{\Phi(p)}S E p ⊥ ∩ T Φ ( p ) S . Let K K K be the energy Gram matrix of these projected vectors and let b α = D α J 1 b_\alpha=D_\alpha\mathcal{J}_1 b α = D α J 1 . Pairing with the constrained gradient gives the exact identity
r g ( Φ ( p ) ) 2 = b T K − 1 b . r_g(\Phi(p))^2=b^{T}K^{-1}b. r g ( Φ ( p ) ) 2 = b T K − 1 b .
The uniform bounds for K K K and K − 1 K^{-1} K − 1 prove (103) . Evenness of the functional proves the assertions for both signs. □
Theorem 6.6 (Quantitative stability away from the umbilic locus). Let K ⋐ C K\Subset C K ⋐ C , where H g = 0 H_g=0 H g = 0 on the open boundary region C C C , and assume inf K ∣ I I ˚ g ∣ > 0 \inf_K\lvert\mathring{\mathrm{II}}_g\rvert>0 inf K ∣ II ˚ g ∣ > 0 . There are ρ , ε 0 , C > 0 \rho,\varepsilon_0,C>0 ρ , ε 0 , C > 0 such that, whenever normalized modulation gives
u = σ ret U x , ε ( v ) , v ∈ X x , ε , ∥ v ∥ E < ρ , x ∈ K , 0 < ε < ε 0 , σ ∈ { ± 1 } , u=\sigma\operatorname{ret}\mathcal{U}_{x,\varepsilon}(v),\qquad v\in X_{x,\varepsilon},\qquad\lVert v\rVert_E<\rho,\qquad x\in K,\qquad0<\varepsilon<\varepsilon_0,\qquad\sigma\in\{\pm1\}, u = σ ret U x , ε ( v ) , v ∈ X x , ε , ∥ v ∥ E < ρ , x ∈ K , 0 < ε < ε 0 , σ ∈ { ± 1 } , one has
∥ u − σ Φ ( x , ε ) ∥ E ≤ C r g ( u ) , (104) \lVert u-\sigma\Phi(x,\varepsilon)\rVert_E\le Cr_g(u), \tag*{(104)} ∥ u − σ Φ ( x , ε ) ∥ E ≤ C r g ( u ) , (104) ε 2 ≤ C r g ( u ) , (105) \varepsilon^2\le Cr_g(u), \tag*{(105)} ε 2 ≤ C r g ( u ) , (105) dist E ( u , B 0 ) ≤ C r g ( u ) 1 / 2 . (106) \operatorname{dist}_E(u,\mathcal{B}_0)\le Cr_g(u)^{1/2}. \tag*{(106)} dist E ( u , B 0 ) ≤ C r g ( u ) 1/2 . (106) In particular, dist E ( u , B L S ) ≤ C r g ( u ) \operatorname{dist}_E(u,\mathcal{B}_{\mathrm{LS}})\le Cr_g(u) dist E ( u , B LS ) ≤ C r g ( u ) . If J g [ u ] ≥ S ∗ \mathcal{J}_g[u]\ge S_* J g [ u ] ≥ S ∗ , then
dist E ( u , B 0 ) ≤ C r g ( u ) . (107) \operatorname{dist}_E(u,\mathcal{B}_0)\le Cr_g(u). \tag*{(107)} dist E ( u , B 0 ) ≤ C r g ( u ) . (107)
Proof. By evenness we may take σ = 1 \sigma=1 σ = 1 . The first estimate is Lemma 6.5. The exact coefficient in Proposition 5.15 is negative, so the differentiated expansion gives
c ε 2 ≤ ∣ ε ∂ ε J 1 ( x , ε ) ∣ ≤ C r g ( Φ ( x , ε ) ) ≤ C r g ( u ) . c\varepsilon^2\le\lvert\varepsilon\partial_\varepsilon\mathcal{J}_1(x,\varepsilon)\rvert\le Cr_g(\Phi(x,\varepsilon))\le Cr_g(u). c ε 2 ≤ ∣ ε ∂ ε J 1 ( x , ε )∣ ≤ C r g ( Φ ( x , ε )) ≤ C r g ( u ) .
This proves (105) . Since ∥ w x , ε ∥ E ≤ C ε \lVert w_{x,\varepsilon}\rVert_E\le C\varepsilon ∥ w x , ε ∥ E ≤ C ε ,
dist E ( u , B 0 ) ≤ C ( r g ( u ) + ε ) ≤ C r g ( u ) 1 / 2 \operatorname{dist}_E(u,\mathcal{B}_0)\le C\bigl(r_g(u)+\varepsilon\bigr)\le Cr_g(u)^{1/2} dist E ( u , B 0 ) ≤ C ( r g ( u ) + ε ) ≤ C r g ( u ) 1/2
for small r g ( u ) r_g(u) r g ( u ) . The remaining part of the neighborhood is covered by enlarging C C C . On K K K , the expansion of J 1 \mathcal{J}_1 J 1 gives J 1 ( x , ε ) ≤ S ∗ − c ε 2 \mathcal{J}_1(x,\varepsilon)\le S_*-c\varepsilon^2 J 1 ( x , ε ) ≤ S ∗ − c ε 2 . If J g [ u ] ≥ S ∗ \mathcal{J}_g[u]\ge S_* J g [ u ] ≥ S ∗ , the upper estimate in (102) implies
c ε 2 ≤ J g [ u ] − J 1 ( x , ε ) ≤ C ∥ v − w x , ε ∥ E 2 ≤ C r g ( u ) 2 . c\varepsilon^2\le\mathcal{J}_g[u]-\mathcal{J}_1(x,\varepsilon)\le C\lVert v-w_{x,\varepsilon}\rVert_E^2\le Cr_g(u)^2. c ε 2 ≤ J g [ u ] − J 1 ( x , ε ) ≤ C ∥ v − w x , ε ∥ E 2 ≤ C r g ( u ) 2 .
Thus ε ≤ C r g ( u ) \varepsilon\le Cr_g(u) ε ≤ C r g ( u ) , which proves (107) . □
Proposition 6.7 (Sharpness of the stability exponents). Suppose H g = 0 H_g=0 H g = 0 on a connected boundary component Σ \Sigma Σ and I I ˚ g \mathring{\mathrm{II}}_g II ˚ g is not identically zero there. There are corrected bubbles u ε = Φ ( x ε , ε ) u_\varepsilon=\Phi(x_\varepsilon,\varepsilon) u ε = Φ ( x ε , ε ) with centers in a compact non-umbilic subset of Σ \Sigma Σ such that
J g [ u ε ] < S ∗ , r g ( u ε ) ≍ ε 2 , dist E ( u ε , B 0 ) ≍ ε . \mathcal{J}_g[u_\varepsilon]<S_*,\qquad r_g(u_\varepsilon)\asymp\varepsilon^2,\qquad\operatorname{dist}_E(u_\varepsilon,\mathcal{B}_0)\asymp\varepsilon. J g [ u ε ] < S ∗ , r g ( u ε ) ≍ ε 2 , dist E ( u ε , B 0 ) ≍ ε . Consequently the exponent 1 / 2 1/2 1/2 in (106) cannot be increased. The exponent 1 1 1 in (107) is also optimal.
Proof. For each small ε \varepsilon ε , choose x ε x_\varepsilon x ε minimizing J 1 ( x , ε ) \mathcal{J}_1(x,\varepsilon) J 1 ( x , ε ) on the compact manifold Σ \Sigma Σ . The uniform expansion of J 1 \mathcal{J}_1 J 1 and κ 3 r e d < 0 \kappa_3^{\mathrm{red}}<0 κ 3 red < 0 imply that every limit point of x ε x_\varepsilon x ε maximizes ∣ I I ˚ ∣ 2 |\mathring{\mathrm{II}}|^2 ∣ II ˚ ∣ 2 on Σ \Sigma Σ . Hence ∣ I I ˚ ( x ε ) ∣ |\mathring{\mathrm{II}}(x_\varepsilon)| ∣ II ˚ ( x ε ) ∣ is bounded below, and J 1 ( x ε , ε ) < S ∗ \mathcal{J}_1(x_\varepsilon,\varepsilon)<S_* J 1 ( x ε , ε ) < S ∗ . At these centers ∇ ∂ J 1 = 0 \nabla_{\partial}\mathcal{J}_1=0 ∇ ∂ J 1 = 0 . The scale expansion and (103) therefore give r g ( u ε ) ≍ ε 2 r_g(u_\varepsilon)\asymp\varepsilon^2 r g ( u ε ) ≍ ε 2 .
The source estimate and injectivity of the source map in Lemma 5.2, together with compactness of the unit sphere in Sym 0 2 \operatorname{Sym}^2_0 Sym 0 2 , give ∥ D X F p ( 0 ) ∥ X p ∗ ≥ c ε \|D_XF_p(0)\|_{X_p^*}\ge c\varepsilon ∥ D X F p ( 0 ) ∥ X p ∗ ≥ c ε at these centers. Since D X F p ( w p ) = 0 D_XF_p(w_p)=0 D X F p ( w p ) = 0 and the Hessian is uniformly bounded, ∥ w p ∥ E ≥ c ε \|w_p\|_E\ge c\varepsilon ∥ w p ∥ E ≥ c ε . The reverse inequality is Lemma 5.9. Uniform normalized modulation gives
dist E ( Φ ( p ) , B 0 ) ≍ ∥ w p ∥ E ≍ ε . \operatorname{dist}_E(\Phi(p),\mathcal{B}_0)\asymp\|w_p\|_E\asymp\varepsilon. dist E ( Φ ( p ) , B 0 ) ≍ ∥ w p ∥ E ≍ ε .
Indeed, sufficiently close bubbles belong to the same normalized parameter chart, where Lemma 6.1 gives the lower estimate; bubbles outside that chart stay a fixed distance away. This proves sharpness of 1 / 2 1/2 1/2 .
To prove sharpness above S ∗ S_* S ∗ , choose h ε ∈ X x ε , ε h_\varepsilon\in X_{x_\varepsilon,\varepsilon} h ε ∈ X x ε , ε with ∥ h ε ∥ E = 1 \|h_\varepsilon\|_E=1 ∥ h ε ∥ E = 1 , fix a sufficiently large constant L L L , and put
u ~ ε = ret U x ε , ε ( w x ε , ε + L ε h ε ) . \widetilde{u}_\varepsilon=\operatorname{ret}_{U_{x_\varepsilon,\varepsilon}}(w_{x_\varepsilon,\varepsilon}+L\varepsilon h_\varepsilon). u ε = ret U x ε , ε ( w x ε , ε + L ε h ε ) .
The lower bound in (102) makes J g [ u ~ ε ] ≥ S ∗ \mathcal{J}_g[\widetilde{u}_\varepsilon]\ge S_* J g [ u ε ] ≥ S ∗ for large fixed L L L . Taking L L L also larger than twice a uniform upper bound for ε − 1 ∥ w x ε , ε ∥ E \varepsilon^{-1}\|w_{x_\varepsilon,\varepsilon}\|_E ε − 1 ∥ w x ε , ε ∥ E , modulation gives dist E ( u ~ ε , B 0 ) ≍ ε \operatorname{dist}_E(\widetilde{u}_\varepsilon,\mathcal{B}_0)\asymp\varepsilon dist E ( u ε , B 0 ) ≍ ε . Complementary coercivity gives r g ( u ~ ε ) ≥ c ε r_g(\widetilde{u}_\varepsilon)\ge c\varepsilon r g ( u ε ) ≥ c ε , and the Lipschitz estimate for the gradient gives the reverse bound. Thus no exponent larger than 1 1 1 is possible on this side. Moreover, no sign conditions are required on these sharpness sequences. □
Remark 6.8 (A quantitative obstruction to concentration). Without a lower bound on ∣ I I ˚ ∣ |\mathring{\mathrm{II}}| ∣ II ˚ ∣ , the same argument gives, uniformly on compact subsets of a minimal boundary region, ε 2 ∣ I I ˚ ( x ) ∣ 2 ≤ C r g ( u ) + o ( ε 2 ) \varepsilon^2|\mathring{\mathrm{II}}(x)|^2\le Cr_g(u)+o(\varepsilon^2) ε 2 ∣ II ˚ ( x ) ∣ 2 ≤ C r g ( u ) + o ( ε 2 ) . In particular, if a modulated concentrating sequence has r g ( u j ) = o ( ε j 2 ) r_g(u_j)=o(\varepsilon_j^2) r g ( u j ) = o ( ε j 2 ) and x j → p x_j\to p x j → p , then p p p is umbilic.
Critical points and compactness
Throughout this section assume n ≥ 5 n\ge5 n ≥ 5 , ( Y γ + ) (Y^+_\gamma) ( Y γ + ) , and ( B G 3 + ) (\mathrm{BG3}^+) ( BG3 + ) , unless stronger hypotheses are stated.
Critical points and the reduced functional
The quadratic scale coefficient excludes sufficiently small stationary scales away from the umbilic locus and vanishes on it. We use modulation and uniqueness to pass from criticality to stationarity of the reduced functional.
Proposition 7.1 (One-bubble critical points and the reduced functional). Let ( u ℓ ) ⊂ S (u_\ell)\subset\mathcal{S} ( u ℓ ) ⊂ S be constrained critical points of J g \mathcal{J}_g J g . Suppose that, after passing to a subsequence and making one fixed global sign change if necessary, there are x ℓ 0 ∈ ∂ M x_\ell^0\in\partial M x ℓ 0 ∈ ∂ M and ε ℓ ↓ 0 \varepsilon_\ell\downarrow0 ε ℓ ↓ 0 such that
∥ u ℓ − u x ℓ 0 , ε ℓ 1 ∥ H 1 ( M ) ⟶ 0. \|u_\ell-u_{x_\ell^0,\varepsilon_\ell}^{1}\|_{H^1(M)}\longrightarrow0. ∥ u ℓ − u x ℓ 0 , ε ℓ 1 ∥ H 1 ( M ) ⟶ 0. Then there are x ℓ ∈ ∂ M x_\ell\in\partial M x ℓ ∈ ∂ M and ε ℓ ↓ 0 \varepsilon_\ell\downarrow0 ε ℓ ↓ 0 such that
u ℓ = Φ ( x ℓ , ε ℓ ) , D J 1 ( x ℓ , ε ℓ ) = 0. u_\ell=\Phi(x_\ell,\varepsilon_\ell),\qquad D\mathcal{J}_1(x_\ell,\varepsilon_\ell)=0. u ℓ = Φ ( x ℓ , ε ℓ ) , D J 1 ( x ℓ , ε ℓ ) = 0. More generally , whenever a constrained critical point admits a representation u = ret U x , ε ( w ) u=\operatorname{ret}_{U_{x,\varepsilon}}(w) u = ret U x , ε ( w ) with a sufficiently small correction w ∈ X x , ε w\in X_{x,\varepsilon} w ∈ X x , ε , one has w = w x , ε w=w_{x,\varepsilon} w = w x , ε , u = Φ ( x , ε ) u=\Phi(x,\varepsilon) u = Φ ( x , ε ) , and $D\mathcal{J}_1(x,\varepsilon)=0.
Proof. For the local statement, criticality makes w w w solve (61) ; uniqueness identifies it with w x , ε w_{x,\varepsilon} w x , ε , and the chain rule on S S S gives D J 1 = 0 D\mathcal{J}_1=0 D J 1 = 0 . For the sequence, normalized modulation (Lemma 6.1 ) gives
u ℓ = ret U x ℓ , ε ℓ ( w ~ ℓ ) , w ~ ℓ ∈ X x ℓ , ε ℓ , ∥ w ~ ℓ ∥ H 1 → 0 , ε ℓ ↓ 0. u_\ell=\operatorname{ret}U_{x_\ell,\varepsilon_\ell}(\widetilde{w}_\ell),\qquad\widetilde{w}_\ell\in X_{x_\ell,\varepsilon_\ell},\qquad\lVert\widetilde{w}_\ell\rVert_{H^1}\to0,\qquad\varepsilon_\ell\downarrow0. u ℓ = ret U x ℓ , ε ℓ ( w ℓ ) , w ℓ ∈ X x ℓ , ε ℓ , ∥ w ℓ ∥ H 1 → 0 , ε ℓ ↓ 0.
For large ℓ \ell ℓ , criticality makes w ~ ℓ \widetilde{w}_\ell w ℓ solve (61) . Uniqueness gives w ~ ℓ = w x ℓ , ε ℓ \widetilde{w}_\ell=w_{x_\ell,\varepsilon_\ell} w ℓ = w x ℓ , ε ℓ , hence u ℓ = Φ ( x ℓ , ε ℓ ) u_\ell=\Phi(x_\ell,\varepsilon_\ell) u ℓ = Φ ( x ℓ , ε ℓ ) . Since Φ \Phi Φ takes values in S S S , the chain rule gives D J 1 ( x ℓ , ε ℓ ) = 0 D\mathcal{J}_1(x_\ell,\varepsilon_\ell)=0 D J 1 ( x ℓ , ε ℓ ) = 0 .
Blow-up points and compactness
Theorem 7.2 (Blow-up points at the level S ∗ S_\ast S ∗ ). Assume n ≥ 5 n\geq5 n ≥ 5 , ( Y ∂ + ) (Y_\partial^+) ( Y ∂ + ) , and ( B G 3 + ) (\mathrm{BG}3^+) ( BG 3 + ) . Let ( u ℓ ) ⊂ S (u_\ell)\subset S ( u ℓ ) ⊂ S be constrained critical points such that
J g [ u ℓ ] ⟶ S ∗ . \mathcal{J}_g[u_\ell]\longrightarrow S_\ast. J g [ u ℓ ] ⟶ S ∗ . If ( u ℓ ) (u_\ell) ( u ℓ ) is not precompact in H 1 ( M ) H^1(M) H 1 ( M ) , then, after passing to a subsequence and making one fixed global sign change,
u ℓ = Φ ( x ℓ , ε ℓ ) , ε ℓ ↓ 0 , x ℓ ⟶ p ∈ ∂ M . u_\ell=\Phi(x_\ell,\varepsilon_\ell),\qquad\varepsilon_\ell\downarrow0,\qquad x_\ell\longrightarrow p\in\partial M. u ℓ = Φ ( x ℓ , ε ℓ ) , ε ℓ ↓ 0 , x ℓ ⟶ p ∈ ∂ M . In a conformal representative with H g = 0 H_g=0 H g = 0 ,
R g r e d ( p ) = 0 , I I ˚ g ( p ) = 0 , ∇ ∂ R g r e d ( p ) = 0. \mathcal{R}^{\mathrm{red}}_g(p)=0,\qquad\mathring{\mathrm{II}}_g(p)=0,\qquad\nabla_{\partial}\mathcal{R}^{\mathrm{red}}_g(p)=0. R g red ( p ) = 0 , II ˚ g ( p ) = 0 , ∇ ∂ R g red ( p ) = 0. In particular, the limiting concentration point is umbilic in every conformal representative.
Proof. Lemma 4.6 and normalized modulation give the concentrating one-bubble representation, after one global sign change. Proposition 7.1 gives u ℓ = Φ ( x ℓ , ε ℓ ) u_\ell=\Phi(x_\ell,\varepsilon_\ell) u ℓ = Φ ( x ℓ , ε ℓ ) and D J 1 ( x ℓ , ε ℓ ) = 0 D\mathcal{J}_1(x_\ell,\varepsilon_\ell)=0 D J 1 ( x ℓ , ε ℓ ) = 0 . Extract x ℓ → p x_\ell\to p x ℓ → p .
The conformal change with boundary factor equal to one preserves the limiting center. Apply modulation again to the sequence of constrained critical points after this conformal change in the representative with H g = 0 H_g=0 H g = 0 . The scale derivative gives
0 = S ∗ − 1 ∂ ε J 1 ( x ℓ , ε ℓ ) = 2 R g r e d ( x ℓ ) ε ℓ + o ( ε ℓ ) . 0=S_\ast^{-1}\partial_\varepsilon\mathcal{J}_1(x_\ell,\varepsilon_\ell)=2\mathcal{R}^{\mathrm{red}}_g(x_\ell)\varepsilon_\ell+o(\varepsilon_\ell). 0 = S ∗ − 1 ∂ ε J 1 ( x ℓ , ε ℓ ) = 2 R g red ( x ℓ ) ε ℓ + o ( ε ℓ ) .
Division by ε ℓ \varepsilon_\ell ε ℓ gives R g r e d ( p ) = 0 \mathcal{R}^{\mathrm{red}}_g(p)=0 R g red ( p ) = 0 . Since R g r e d = κ 3 r e d ∣ I I ˚ ∣ 2 \mathcal{R}^{\mathrm{red}}_g=\kappa^{\mathrm{red}}_3\lvert\mathring{\mathrm{II}}\rvert^2 R g red = κ 3 red ∣ II ˚ ∣ 2 with κ 3 r e d < 0 \kappa^{\mathrm{red}}_3<0 κ 3 red < 0 , both I I ˚ ( p ) \mathring{\mathrm{II}}(p) II ˚ ( p ) and ∇ ∂ R g r e d ( p ) \nabla_{\partial}\mathcal{R}^{\mathrm{red}}_g(p) ∇ ∂ R g red ( p ) vanish. The umbilic locus is conformally invariant.
Lemma 7.3 (Palais–Smale sequences at level S ∗ S_\ast S ∗ with prescribed centers). Under the standing hypotheses, use the normalized bubble U x , ε U_{x,\varepsilon} U x , ε with the cutoff R ( ε ) = ε − 3 / 4 R(\varepsilon)=\varepsilon^{-3/4} R ( ε ) = ε − 3/4 . Uniformly for x ∈ ∂ M x\in\partial M x ∈ ∂ M ,
∥ grad S J g ( U x , ε ) ∥ H 1 ( M ) ≤ C ( ε + ω ∇ ( ε ) ) = O ( ε ) . (108) \lVert\operatorname{grad}_S\mathcal{J}_g(U_{x,\varepsilon})\rVert_{H^1(M)}
\leq C\bigl(\varepsilon+\omega_{\nabla}(\varepsilon)\bigr)=O(\varepsilon). \tag*{(108)} ∥ grad S J g ( U x , ε ) ∥ H 1 ( M ) ≤ C ( ε + ω ∇ ( ε ) ) = O ( ε ) . (108) and
J g ( U x , ε ) = S ∗ + o ( 1 ) . \mathcal{J}_g(U_{x,\varepsilon})=S_\ast+o(1). J g ( U x , ε ) = S ∗ + o ( 1 ) . Consequently, every p ∈ ∂ M p\in\partial M p ∈ ∂ M is the limiting center of a nonnegative Palais–Smale sequence at level S ∗ S_\ast S ∗ .
Proof. Lemma 3.7 gives the gradient estimate, and the standard bubble expansion gives J g [ U x , ε ] = S ∗ + o ( 1 ) \mathcal{J}_g[U_{x,\varepsilon}]=S_\ast+o(1) J g [ U x , ε ] = S ∗ + o ( 1 ) . Taking x = p x=p x = p and ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 proves the assertion at any prescribed point.
Remark 7.4 (Minimizing sequences of critical points and small scales). A noncompact minimizing sequence of constrained critical points has level S ∗ S_\ast S ∗ by Theorem 4.3 , so Theorem 7.2 applies. On a compact minimal boundary region where ∣ R g r e d ∣ \lvert\mathcal{R}^{\mathrm{red}}_g\rvert ∣ R g red ∣ is bounded below, the uniform scale expansion gives ∂ ε J 1 < 0 \partial_\varepsilon\mathcal{J}_1<0 ∂ ε J 1 < 0 at sufficiently small scales, excluding reduced critical points there.
Corollary 7.5 (Compactness below the first mixed level). Under the standing hypotheses, suppose that ∂ M \partial M ∂ M is nowhere umbilic. For every c 0 < c m i x c_0<c_{\mathrm{mix}} c 0 < c mix , with c m i x c_{\mathrm{mix}} c mix as in (48) , the set of normalized constrained critical points satisfying J g [ u ] ≤ c 0 \mathcal{J}_g[u]\leq c_0 J g [ u ] ≤ c 0 is compact in H 1 ( M ) H^1(M) H 1 ( M ) . In particular, every sequence of normalized constrained critical points with J g [ u ℓ ] → S ∗ \mathcal{J}_g[u_\ell]\to S_\ast J g [ u ℓ ] → S ∗ is precompact.
Proof. Pass to a subsequence whose values converge. Below S ∗ S_* S ∗ apply Theorem 4.3 ; at S ∗ S_* S ∗ , Theorem 7.2 excludes blow-up because there are no umbilic points; above S ∗ S_* S ∗ apply Theorem 4.7 . Strong convergence preserves the constraint, criticality, and the energy bound. □
Write Met m ( M ) \operatorname{Met}^{m}(M) Met m ( M ) for the Banach manifold of C m C^{m} C m Riemannian metrics. Nowhere-umbilic genericity is also recorded in [30 p. 1765] ; the finite-regularity proof is given in Appendix C.
Theorem 7.6 (Generic strict subcriticality and threshold compactness). Assume n ≥ 5 n \ge5 n ≥ 5 and fix an integer m ≥ 3 m \ge3 m ≥ 3 . Let
Met Y , B G m ( M ) : = { g ∈ Met m ( M ) : g satisfies ( Y ∂ + ) and ( B G 3 + ) } . \operatorname{Met}^{m}_{Y,BG}(M) := \left\{g \in\operatorname{Met}^{m}(M) : g\ \text{satisfies}\ (Y_{\partial}^{+})\ \text{and}\ (BG^{3+})\right\}. Met Y , B G m ( M ) := { g ∈ Met m ( M ) : g satisfies ( Y ∂ + ) and ( B G 3 + ) } . The bounds and chart radii in ( B G 3 + ) (BG^{3+}) ( B G 3 + ) may depend on the metric. Every C m C^{m} C m metric on the fixed compact manifold satisfies these bounds, and strict coercivity is open under C 2 C^{2} C 2 perturbations. Thus this is a C m C^{m} C m -open subset of Met m ( M ) \operatorname{Met}^{m}(M) Met m ( M ) . There is a C m C^{m} C m -open dense set
G ⊂ Met Y , B G m ( M ) \mathcal{G} \subset\operatorname{Met}^{m}_{Y,BG}(M) G ⊂ Met Y , B G m ( M ) such that every g ∈ G g \in\mathcal{G} g ∈ G satisfies
I I ˚ g ( x ) ≠ 0 ( x ∈ ∂ M ) , W g ≠ 0 in M ∘ . (109) \mathring{\mathrm{II}}_{g}(x) \ne0 \quad(x \in\partial M), \qquad W_{g} \ne0 \quad\text{in } M^{\circ}. \tag*{(109)} II ˚ g ( x ) = 0 ( x ∈ ∂ M ) , W g = 0 in M ∘ . (109) In particular, ( M , g ) (M,g) ( M , g ) is not conformally diffeomorphic to the round hemisphere. Moreover,
C E s c ∗ ( M , g ) < S ∗ , C E s c ∗ ( M , g ) = J g [ u g ] for some u g ∈ S . (110) C_{\mathrm{Esc}}^{*}(M,g) < S_*, \qquad C_{\mathrm{Esc}}^{*}(M,g) = J_g[u_g] \quad\text{for some } u_g \in\mathcal{S}. \tag*{(110)} C Esc ∗ ( M , g ) < S ∗ , C Esc ∗ ( M , g ) = J g [ u g ] for some u g ∈ S . (110) Every normalized Palais–Smale sequence at C E s c ∗ ( M , g ) C_{\mathrm{Esc}}^{*}(M,g) C Esc ∗ ( M , g ) is precompact in H 1 ( M ) H^{1}(M) H 1 ( M ) . For every c 0 < c m i x c_0 < c_{\mathrm{mix}} c 0 < c mix , the normalized constrained critical points with values at most c 0 c_0 c 0 form a compact subset of H 1 ( M ) H^{1}(M) H 1 ( M ) .
Proof. Appendix C proves the boundary transversality, the independent interior Weyl perturbation, and the analytic conclusions at finite C m C^{m} C m regularity. □
The fourth-order reduced coefficient
The quadratic scale equation forces blow-up points of sequences of constrained critical points at S ∗ S_* S ∗ to be umbilic. We now continue the same reduction to fourth order at umbilic points. The expansion is constructed on the umbilic locus; its explicit curvature formula requires total umbilicity in a neighborhood. Total umbilicity also permits the expansion to be applied uniformly along varying centers of sequences of constrained critical points.
Lemma 8.1 (Vanishing at umbilic centers). Assume n ≥ 6 n \ge6 n ≥ 6 , ( Y ∂ + ) (Y_{\partial}^{+}) ( Y ∂ + ) , and ( B G 4 + ) (BG^{4+}) ( B G 4 + ) , and work in a boundary region with H g = 0 H_g = 0 H g = 0 . Write Θ g \Theta_g Θ g for the cubic coefficient of the normalized standard-bubble energy J g / S ∗ J_g/S_* J g / S ∗ , as in [41 Lemma 4.11] . At every umbilic point p p p in this region,
R g b a r e ( p ) = R g r e d ( p ) = ∇ ∂ R g r e d ( p ) = Θ g ( p ) = 0. \mathfrak{R}^{\mathrm{bare}}_{g}(p) = \mathfrak{R}^{\mathrm{red}}_{g}(p) = \nabla_{\partial}\mathfrak{R}^{\mathrm{red}}_{g}(p) = \Theta_g(p) = 0. R g bare ( p ) = R g red ( p ) = ∇ ∂ R g red ( p ) = Θ g ( p ) = 0. Proof. The standard-bubble coefficients vanish by Proposition 2.5. The formula R g r e d = κ 3 r e d ∣ I I ˚ ∣ 2 \mathfrak{R}^{\mathrm{red}}_{g} = \kappa_{3}^{\mathrm{red}} \lvert\mathring{\mathrm{II}}\rvert^{2} R g red = κ 3 red ∣ II ˚ ∣ 2 from Proposition 5.12 gives the remaining equalities. □
The second-order term and the linearized equation
For n ≥ 6 n \ge6 n ≥ 6 and ( B G 4 + ) (BG^{4+}) ( B G 4 + ) , the quadratic and geometric cubic coefficients vanish at an umbilic point with zero mean curvature by Lemma 8.1 . The order-ε \varepsilon ε geometric source vanishes there as well. For n ≥ 7 n \ge7 n ≥ 7 , with the cutoff below, the complementary source is O ( ε 2 ) O(\varepsilon^{2}) O ( ε 2 ) , and the leading correction obtained by completing the square has order ε 4 \varepsilon^{4} ε 4 .
For the rest of this section assume n ≥ 7 n \ge7 n ≥ 7 and fix 4 / ( n − 2 ) < β < 1 4/(n-2) < \beta< 1 4/ ( n − 2 ) < β < 1 . We refine the cutoff to separate this contribution from the tails: R β ( ε ) = ε − β R_{\beta}(\varepsilon) = \varepsilon^{-\beta} R β ( ε ) = ε − β makes the cutoff radius on M M M tend to zero, the flat energy tail o ( ε 4 ) o(\varepsilon^{4}) o ( ε 4 ) , and the tail in the first variation o ( ε 2 ) o(\varepsilon^{2}) o ( ε 2 ) .
Lemma 8.3 (Second-order source and correction). Assume n ≥ 7 n\geq7 n ≥ 7 , ( Y ∂ + ) (Y_\partial^+) ( Y ∂ + ) , and ( B G 5 + ) (BG^{5+}) ( B G 5 + ) , and let C ⊂ ∂ M C\subset\partial M C ⊂ ∂ M be an open boundary region admitting a fixed Fermi collar, with H g = 0 H_{\mathfrak g}=0 H g = 0 on C C C . Use the β \beta β -cutoff family and objects of Remark 8.2 . In each smoothly varying local Fermi frame there is a continuous family S 2 , p ∈ X + ∗ \mathscr{S}_{2,p}\in X_+^* S 2 , p ∈ X + ∗ on the umbilic locus { p ∈ C : I I ˚ g ( p ) = 0 } \{p\in C:\mathring{\mathrm{II}}_{\mathfrak g}(p)=0\} { p ∈ C : II ˚ g ( p ) = 0 } such that
∥ ε − 2 ( B p , ε ( β ) ) ∗ S p , ε ( β ) − S 2 , p ∥ X + ∗ = o ( 1 ) . (113) \left\|\varepsilon^{-2}\left(\mathcal{B}_{p,\varepsilon}^{(\beta)}\right)^*\mathscr{S}_{p,\varepsilon}^{(\beta)}-\mathscr{S}_{2,p}\right\|_{X_+^*}=o(1). \tag*{(113)} ε − 2 ( B p , ε ( β ) ) ∗ S p , ε ( β ) − S 2 , p X + ∗ = o ( 1 ) . (113) Let A = ( n − 2 ) / 2 + Y ⋅ ∇ \mathcal{A}=(n-2)/2+Y\cdot\nabla A = ( n − 2 ) /2 + Y ⋅ ∇ and D ( A ) = { z ∈ D 1 , 2 : A z ∈ D 1 , 2 } D(\mathcal{A})=\{z\in D^{1,2}:\mathcal{A}z\in D^{1,2}\} D ( A ) = { z ∈ D 1 , 2 : A z ∈ D 1 , 2 } . For each fixed z ∈ X + ∩ D ( A ) z\in X_+\cap D(\mathcal{A}) z ∈ X + ∩ D ( A ) ,
D [ ε − 2 ⟨ S p , ε ( β ) , B p , ε ( β ) z ⟩ ] = o ( 1 ) . (114) \mathcal{D}\left[\varepsilon^{-2}\langle\mathscr{S}_{p,\varepsilon}^{(\beta)},\mathcal{B}_{p,\varepsilon}^{(\beta)}z\rangle\right]=o(1). \tag*{(114)} D [ ε − 2 ⟨ S p , ε ( β ) , B p , ε ( β ) z ⟩ ] = o ( 1 ) . (114) This last convergence is uniform for z z z in sets compact with respect to the graph norm ∥ z ∥ D 1 , 2 + ∥ A z ∥ D 1 , 2 \|z\|_{D^{1,2}}+\|\mathcal{A}z\|_{D^{1,2}} ∥ z ∥ D 1 , 2 + ∥ A z ∥ D 1 , 2 . Both statements are uniform for p p p in compact subsets of the umbilic locus. A change of tangential orthonormal frame acts by the corresponding orthogonal change of variables on X + X_+ X + and its dual.
Put z 2 , p = L 0 − 1 S 2 , p z_{2,p}=\mathscr{L}_0^{-1}\mathscr{S}_{2,p} z 2 , p = L 0 − 1 S 2 , p and v p , ε ( β ) = ( L p , ε ( β ) ) − 1 S p , ε ( β ) v_{p,\varepsilon}^{(\beta)}=(\mathscr{L}_{p,\varepsilon}^{(\beta)})^{-1}\mathscr{S}_{p,\varepsilon}^{(\beta)} v p , ε ( β ) = ( L p , ε ( β ) ) − 1 S p , ε ( β ) . Then
∥ v p , ε ( β ) − ε 2 B p , ε ( β ) z 2 , p ∥ H 1 ( M ) = o ( ε 2 ) , ∥ D p h v p , ε ( β ) ∥ H 1 ( M ) ≤ C ε 2 . (115) \left\|v_{p,\varepsilon}^{(\beta)}-\varepsilon^2\mathcal{B}_{p,\varepsilon}^{(\beta)}z_{2,p}\right\|_{H^1(M)}=o(\varepsilon^2),\qquad\left\|D^{\mathrm{ph}}v_{p,\varepsilon}^{(\beta)}\right\|_{H^1(M)}\leq C\varepsilon^2. \tag*{(115)} v p , ε ( β ) − ε 2 B p , ε ( β ) z 2 , p H 1 ( M ) = o ( ε 2 ) , D ph v p , ε ( β ) H 1 ( M ) ≤ C ε 2 . (115) Moreover,
⟨ S p , ε ( β ) , ( L p , ε ( β ) ) − 1 S p , ε ( β ) ⟩ = ε 4 ⟨ S 2 , p , L 0 − 1 S 2 , p ⟩ + e 2 , \left\langle\mathscr{S}_{p,\varepsilon}^{(\beta)},\left(\mathscr{L}_{p,\varepsilon}^{(\beta)}\right)^{-1}\mathscr{S}_{p,\varepsilon}^{(\beta)}\right\rangle
=\varepsilon^4\left\langle\mathscr{S}_{2,p},\mathscr{L}_0^{-1}\mathscr{S}_{2,p}\right\rangle+e_2, ⟨ S p , ε ( β ) , ( L p , ε ( β ) ) − 1 S p , ε ( β ) ⟩ = ε 4 ⟨ S 2 , p , L 0 − 1 S 2 , p ⟩ + e 2 , ∣ e 2 ∣ + ∣ D e 2 ∣ = o ( ε 4 ) . (116) |e_2|+|De_2|=o(\varepsilon^4). \tag*{(116)} ∣ e 2 ∣ + ∣ D e 2 ∣ = o ( ε 4 ) . (116) The estimates are uniform when β \beta β ranges over a compact subinterval of ( 4 / ( n − 2 ) , 1 ) (4/(n-2),1) ( 4/ ( n − 2 ) , 1 ) , with the same fixed cutoff profile. In particular, S 2 , p \mathscr{S}_{2,p} S 2 , p and z 2 , p z_{2,p} z 2 , p are independent of β \beta β .
Proof. Put α = ( n − 2 ) / 2 \alpha=(n-2)/2 α = ( n − 2 ) /2 and c n = ( n − 2 ) / ( 4 ( n − 1 ) ) c_n=(n-2)/(4(n-1)) c n = ( n − 2 ) / ( 4 ( n − 1 )) . All curvature tensors in the following formulas are evaluated at p p p ; barred tensors belong to the induced boundary metric, and a semicolon denotes covariant differentiation.
The source. At an umbilic point with zero mean curvature, I I ( p ) = 0 \mathrm{II}(p)=0 II ( p ) = 0 . The second-order Fermi coefficients are
b 2 ( Y ′ ) = − 1 6 Ric ‾ c d Y c Y d , d 2 ( Y ) = b 2 ( Y ′ ) − 1 2 Ric n n t 2 , b_2(Y')=-\frac{1}{6}\overline{\operatorname{Ric}}_{cd}Y_cY_d,\qquad d_2(Y)=b_2(Y')-\frac{1}{2}\operatorname{Ric}_{nn}t^2, b 2 ( Y ′ ) = − 6 1 Ric c d Y c Y d , d 2 ( Y ) = b 2 ( Y ′ ) − 2 1 Ric nn t 2 ,
A 2 a b ( Y ) = 1 3 R ‾ a c b d Y c Y d + 2 I I a b ; c Y c t + R a n b n t 2 + d 2 ( Y ) δ a b , A_2^{ab}(Y)=\frac{1}{3}\overline{R}_{acbd}Y_cY_d+2\mathrm{II}_{ab;c}Y_ct+R_{anbn}t^2+d_2(Y)\delta_{ab}, A 2 ab ( Y ) = 3 1 R a c b d Y c Y d + 2 II ab ; c Y c t + R anbn t 2 + d 2 ( Y ) δ ab ,
A 2 n n ( Y ) = d 2 ( Y ) , A 2 a n ( Y ) = 0. (117) A_2^{nn}(Y)=d_2(Y),\qquad A_2^{an}(Y)=0. \tag*{(117)} A 2 nn ( Y ) = d 2 ( Y ) , A 2 an ( Y ) = 0. (117)
Thus det g ( ε Y ) g i j ( ε Y ) = δ i j + ε 2 A 2 i j ( Y ) + O ( ε 3 ∣ Y ∣ 3 ) \sqrt{\det g(\varepsilon Y)}g^{ij}(\varepsilon Y)=\delta^{ij}+\varepsilon^2A_2^{ij}(Y)+O(\varepsilon^3|Y|^3) det g ( ε Y ) g ij ( ε Y ) = δ ij + ε 2 A 2 ij ( Y ) + O ( ε 3 ∣ Y ∣ 3 ) and the boundary density is 1 + ε 2 b 2 ( Y ′ ) + O ( ε 3 ∣ Y ′ ∣ 3 ) 1+\varepsilon^2b_2(Y')+O(\varepsilon^3|Y'|^3) 1 + ε 2 b 2 ( Y ′ ) + O ( ε 3 ∣ Y ′ ∣ 3 ) . On X + X_+ X + the source is
S 2 , p [ ϕ ] = 2 t ∗ ( ∫ R + n A 2 i j ∂ i U + ∂ j ϕ + c n Scal ( p ) ∫ R + n U + ϕ − λ f l a t ∫ ∂ R + n b 2 U + q − 1 ϕ ) . (118) \mathcal{S}_{2,p}[\phi]=2t_*\left(\int_{\mathbb{R}^n_+}A_2^{ij}\partial_iU_+\partial_j\phi+c_n\operatorname{Scal}(p)\int_{\mathbb{R}^n_+}U_+\phi-\lambda_{\mathrm{flat}}\int_{\partial\mathbb{R}^n_+}b_2U_+^{q-1}\phi\right). \tag*{(118)} S 2 , p [ ϕ ] = 2 t ∗ ( ∫ R + n A 2 ij ∂ i U + ∂ j ϕ + c n Scal ( p ) ∫ R + n U + ϕ − λ flat ∫ ∂ R + n b 2 U + q − 1 ϕ ) . (118)
Trace normalization adds only a multiple of the constraint covector, which vanishes on X + X_+ X + . The terms I I a b ; c \mathrm{II}_{ab;c} II ab ; c remain at a general umbilic point. They vanish when the boundary is totally umbilic near p p p in this representative with minimal boundary.
Since A 2 n n ( Y ′ , 0 ) = b 2 ( Y ′ ) A_2^{nn}(Y',0)=b_2(Y') A 2 nn ( Y ′ , 0 ) = b 2 ( Y ′ ) , A 2 a n = 0 A_2^{an}=0 A 2 an = 0 , and − ∂ t U + = λ f l a t U + q − 1 -\partial_tU_+=\lambda_{\mathrm{flat}}U_+^{q-1} − ∂ t U + = λ flat U + q − 1 on the boundary, integration by parts cancels the last term in (118) . Its extension to the full Dirichlet space therefore has interior forcing
f 2 , p = 2 t ∗ [ − div ( A 2 ∇ U + ) + c n Scal ( p ) U + ] , ∣ f 2 , p ∣ + ∣ Y ⋅ ∇ f 2 , p ∣ ≤ C ( 1 + ∣ Y ∣ ) 2 − n . f_{2,p}=2t_*\left[-\operatorname{div}(A_2\nabla U_+)+c_n\operatorname{Scal}(p)U_+\right],\qquad|f_{2,p}|+|Y\cdot\nabla f_{2,p}|\le C(1+|Y|)^{2-n}. f 2 , p = 2 t ∗ [ − div ( A 2 ∇ U + ) + c n Scal ( p ) U + ] , ∣ f 2 , p ∣ + ∣ Y ⋅ ∇ f 2 , p ∣ ≤ C ( 1 + ∣ Y ∣ ) 2 − n .
The model moments converge because
∥ ∣ Y ∣ 2 ∇ U + ∥ L 2 ( ∣ Y ∣ > R ) + ∥ ∣ Y ∣ U + ∥ L 2 ( ∣ Y ∣ > R ) ≤ C R − ( n − 6 ) / 2 . (119) \left\||Y|^2\nabla U_+\right\|_{L^2(|Y|>R)}+\left\||Y|U_+\right\|_{L^2(|Y|>R)}\le CR^{-(n-6)/2}. \tag*{(119)} ∣ Y ∣ 2 ∇ U + L 2 ( ∣ Y ∣ > R ) + ∥ ∣ Y ∣ U + ∥ L 2 ( ∣ Y ∣ > R ) ≤ C R − ( n − 6 ) /2 . (119)
The tail of the second-order boundary term has the stronger bound ∥ ∣ Y ′ ∣ 2 U + q − 1 ∥ L q ′ ( ∣ Y ′ ∣ > R ) ≤ C R − ( n − 4 ) / 2 \left\||Y'|^2U_+^{q-1}\right\|_{L^{q'}(|Y'|>R)}\le CR^{-(n-4)/2} ∣ Y ′ ∣ 2 U + q − 1 L q ′ ( ∣ Y ′ ∣ > R ) ≤ C R − ( n − 4 ) /2 , where q ′ = q / ( q − 1 ) q'=q/(q-1) q ′ = q / ( q − 1 ) . The next term in the interior Taylor expansion contributes C ε 3 ℓ n ( R ) C\varepsilon^3\ell_n(R) C ε 3 ℓ n ( R ) , where
ℓ n ( R ) = { R 1 / 2 , n = 7 , log ( 2 R ) , n = 8 , 1 , n ≥ 9. \ell_n(R)=
\begin{cases}
R^{1/2}, & n=7,\\
\sqrt{\log(2R)}, & n=8,\\
1, & n\ge9.
\end{cases} ℓ n ( R ) = ⎩ ⎨ ⎧ R 1/2 , log ( 2 R ) , 1 , n = 7 , n = 8 , n ≥ 9.
Indeed, the required norms are ∥ ∣ Y ∣ 3 ∇ U + ∥ L 2 ( B 2 R + ) \left\||Y|^3\nabla U_+\right\|_{L^2(B^+_{2R})} ∣ Y ∣ 3 ∇ U + L 2 ( B 2 R + ) and ∥ ∣ Y ∣ 2 U + ∥ L 2 ( B 2 R + ) \left\||Y|^2U_+\right\|_{L^2(B^+_{2R})} ∣ Y ∣ 2 U + L 2 ( B 2 R + ) . The third-order boundary term is O ( ε 3 ) O(\varepsilon^3) O ( ε 3 ) since ∣ Y ′ ∣ 3 U + q − 1 ∈ L q ′ |Y'|^3U_+^{q-1}\in L^{q'} ∣ Y ′ ∣ 3 U + q − 1 ∈ L q ′ . If C R \mathcal{C}_R C R denotes the flat cutoff covector and E p , ε , R \mathcal{E}_{p,\varepsilon,R} E p , ε , R the Taylor error, then
∥ C R ∥ + ∥ R ∂ R C R ∥ ≤ C R − α , ∥ E p , ε , R ∥ + ∥ ( ε ∂ ε − β R ∂ R ) E p , ε , R ∥ ≤ C ε 3 ℓ n ( R ) . (120) \|\mathcal{C}_R\|+\|R\partial_R\mathcal{C}_R\|\le CR^{-\alpha},\qquad\|\mathcal{E}_{p,\varepsilon,R}\|+\|(\varepsilon\partial_\varepsilon-\beta R\partial_R)\mathcal{E}_{p,\varepsilon,R}\|\le C\varepsilon^3\ell_n(R). \tag*{(120)} ∥ C R ∥ + ∥ R ∂ R C R ∥ ≤ C R − α , ∥ E p , ε , R ∥ + ∥ ( ε ∂ ε − β R ∂ R ) E p , ε , R ∥ ≤ C ε 3 ℓ n ( R ) . (120)
Consequently the normalized error, also after one logarithmic scale derivative in model coordinates, is bounded by
C ( ε − 2 R − α + R − ( n − 6 ) / 2 + ε ℓ n ( R ) ) = o ( 1 ) , R = ε − β . C\left(\varepsilon^{-2}R^{-\alpha}+R^{-(n-6)/2}+\varepsilon\ell_n(R)\right)=o(1),\qquad R=\varepsilon^{-\beta}. C ( ε − 2 R − α + R − ( n − 6 ) /2 + ε ℓ n ( R ) ) = o ( 1 ) , R = ε − β .
Here β α > 2 \beta\alpha>2 β α > 2 and β < 1 \beta<1 β < 1 . The inverse Gram matrices are uniformly bounded; the frames converge strongly, and the constraint normal is controlled by Remark 5.5 . The O ( ε ) O(\varepsilon) O ( ε ) bound for the full first variation and the Riesz identity give the strong frame convergence. Projecting the source therefore preserves this limit and proves (113) . Continuity in p p p follows from the displayed finite collection of curvature jets.
Logarithmic scale derivatives at fixed points of M M M . Apply (92) with k = 2 k=2 k = 2 . Hardy’s inequality gives
∫ ∣ Y ∣ > 1 ∣ Y ∣ 2 ( ∣ f 2 , p ∣ 2 + ∣ Y ⋅ ∇ f 2 , p ∣ 2 ) d Y ≤ C ∫ 1 ∞ r 5 − n d r < ∞ . \int_{|Y|>1}|Y|^2\left(|f_{2,p}|^2+|Y\cdot\nabla f_{2,p}|^2\right)\,dY\le C\int_1^\infty r^{5-n}\,dr<\infty. ∫ ∣ Y ∣ > 1 ∣ Y ∣ 2 ( ∣ f 2 , p ∣ 2 + ∣ Y ⋅ ∇ f 2 , p ∣ 2 ) d Y ≤ C ∫ 1 ∞ r 5 − n d r < ∞.
The explicit cutoff and Taylor expressions obey the differentiated bounds above. The constraint contribution from normalization is estimated by its scalar coefficient and the bounded constraint covector. Thus the full first variation and bilinear form on M M M satisfy
∥ S f u l l ∥ + ∥ D p h S f u l l ∥ ≤ C ε 2 , ∥ a ε ∥ + ∥ D p h a ε ∥ ≤ C . (121) \|\mathcal{S}^{\mathrm{full}}\|+\|\mathrm{D}^{\mathrm{ph}}\mathcal{S}^{\mathrm{full}}\|\le C\varepsilon^2,\qquad\|a_\varepsilon\|+\|\mathrm{D}^{\mathrm{ph}}a_\varepsilon\|\le C. \tag*{(121)} ∥ S full ∥ + ∥ D ph S full ∥ ≤ C ε 2 , ∥ a ε ∥ + ∥ D ph a ε ∥ ≤ C . (121)
As in the proof of Proposition 6.4, the Riesz identity and inverse Gram formula give bounded logarithmic scale derivatives of the projection at fixed points of M M M . Stabilized inversion yields
∥ v ∥ H 1 + ∥ D p h v ∥ H 1 ≤ C ε 2 . (122) \|v\|_{H^1}+\|\mathrm{D}^{\mathrm{ph}}v\|_{H^1}\le C\varepsilon^2. \tag*{(122)} ∥ v ∥ H 1 + ∥ D ph v ∥ H 1 ≤ C ε 2 . (122)
The solution of the linearized equation. The fixed-ball and tail argument of Lemma 5.11 , using (119) , identifies every rescaled weak limit of ε − 2 v \varepsilon^{-2}v ε − 2 v with z 2 , p = L 0 − 1 S 2 , p z_{2,p}=\mathcal{L}_0^{-1}\mathcal{S}_{2,p} z 2 , p = L 0 − 1 S 2 , p . In particular,
ε − 4 ⟨ S , v ⟩ ⟶ C p : = ⟨ S 2 , p , z 2 , p ⟩ . \varepsilon^{-4}\langle\mathcal{S},v\rangle\longrightarrow C_p:=\langle\mathcal{S}_{2,p},z_{2,p}\rangle. ε − 4 ⟨ S , v ⟩ ⟶ C p := ⟨ S 2 , p , z 2 , p ⟩ .
For the dilation regularity of z 2 , p z_{2,p} z 2 , p , extend the complementary equation to the full Dirichlet space by adding finitely many constraint and Jacobi covectors. Their Dirichlet representatives are harmonic and add only Robin boundary forcing of size O ( ( 1 + ∣ Y ′ ∣ ) − n ) O((1+|Y'|)^{-n}) O (( 1 + ∣ Y ′ ∣ ) − n ) , with the corresponding differentiated decay. These terms have finite boundary norms in (95) . The displayed Hardy integral and that estimate give Y ∇ 2 z 2 , p ∈ L 2 Y\nabla^2z_{2,p}\in L^2 Y ∇ 2 z 2 , p ∈ L 2 ; local regularity, Hardy’s inequality, and cutoff approximation then give A z 2 , p ∈ D 1 , 2 Az_{2,p}\in D^{1,2} A z 2 , p ∈ D 1 , 2 .
For a fixed z ∈ X + ∩ D ( A ) z\in X_+\cap D(A) z ∈ X + ∩ D ( A ) , the transported test and its logarithmic scale derivative at a fixed point of M M M are bounded by its graph norm, by (80) . The coefficient estimates give (114) before projection. The removed moments and their derivatives tend to zero: for each limiting normal field e e e ,
⟨ A z , e ⟩ D 1 , 2 + ⟨ z , A e ⟩ D 1 , 2 = 0 \langle Az,e\rangle_{D^{1,2}}+\langle z,Ae\rangle_{D^{1,2}}=0 ⟨ A z , e ⟩ D 1 , 2 + ⟨ z , A e ⟩ D 1 , 2 = 0
by the energy isometry of critical dilations. The exact Riesz identity handles the constraint normal on M M M with an O ( ε 2 ) O(\varepsilon^2) O ( ε 2 ) error, also after differentiation. The inverse Gram formula proves the same statement after projection. Uniform graph-norm bounds and a finite-net argument give uniformity on compact subsets of the graph-norm space.
Completing the square. Set V ε = ε 2 B p , ε ( β ) z 2 , p V_\varepsilon=\varepsilon^2\mathcal{B}^{(\beta)}_{p,\varepsilon}z_{2,p} V ε = ε 2 B p , ε ( β ) z 2 , p . The preceding source comparison and the calculation of the bilinear-form pairing in the proof of Proposition 6.4 give
S [ V ε ] = ε 4 C p + o ( ε 4 ) , a ε ( V ε , V ε ) = ε 4 C p + o ( ε 4 ) , \mathcal{S}[V_\varepsilon]=\varepsilon^4C_p+o(\varepsilon^4),\qquad a_\varepsilon(V_\varepsilon,V_\varepsilon)=\varepsilon^4C_p+o(\varepsilon^4), S [ V ε ] = ε 4 C p + o ( ε 4 ) , a ε ( V ε , V ε ) = ε 4 C p + o ( ε 4 ) ,
including o ( ε 4 ) o(\varepsilon^4) o ( ε 4 ) weighted derivatives of both errors, and ∥ D p h V ε ∥ H 1 ≤ C ε 2 \lVert\mathrm{D}^{\mathrm{ph}}V_\varepsilon\rVert_{H^1}\leq C\varepsilon^2 ∥ D ph V ε ∥ H 1 ≤ C ε 2 . Equation (86) and coercivity imply ∥ v − V ε ∥ H 1 = o ( ε 2 ) \lVert v-V_\varepsilon\rVert_{H^1}=o(\varepsilon^2) ∥ v − V ε ∥ H 1 = o ( ε 2 ) . Both logarithmic scale derivatives at fixed points of M M M are O ( ε 2 ) O(\varepsilon^2) O ( ε 2 ) , so Lemma 6.3 proves (116) . The finite-dimensional dependence on the curvature jets gives uniformity in p p p . All exponents remain positive on compact subintervals of 4 / ( n − 2 ) < β < 1 4/(n-2)<\beta<1 4/ ( n − 2 ) < β < 1 , and the limiting equation has no β \beta β dependence. □
The fourth-order expansion
Lemma 8.4 (Remainder estimate and critical points for the cutoff family). Under the hypotheses of Lemma 8.3 , let w x , ε ( β ) w^{(\beta)}_{x,\varepsilon} w x , ε ( β ) be the unique small one-bubble Lyapunov–Schmidt correction and define
Φ β ( x , ε ) : = ret U x , ε ( β ) ( w x , ε ( β ) ) , J 1 , β : = J g ∘ Φ β . \Phi_\beta(x,\varepsilon):=\operatorname{ret}_{\mathcal{U}^{(\beta)}_{x,\varepsilon}}\left(w^{(\beta)}_{x,\varepsilon}\right),\qquad\mathcal{J}_{1,\beta}:=\mathcal{J}_g\circ\Phi_\beta. Φ β ( x , ε ) := ret U x , ε ( β ) ( w x , ε ( β ) ) , J 1 , β := J g ∘ Φ β . Put
E β ( x , ε ) : = J 1 , β ( x , ε ) − J g [ U x , ε ( β ) ] + 1 2 ⟨ ( L x , ε ( β ) ) − 1 S x , ε ( β ) , S x , ε ( β ) ⟩ . \mathcal{E}_\beta(x,\varepsilon):=\mathcal{J}_{1,\beta}(x,\varepsilon)-\mathcal{J}_g[\mathcal{U}^{(\beta)}_{x,\varepsilon}]+\frac{1}{2}\left\langle(\mathcal{L}^{(\beta)}_{x,\varepsilon})^{-1}\mathcal{S}^{(\beta)}_{x,\varepsilon},\mathcal{S}^{(\beta)}_{x,\varepsilon}\right\rangle. E β ( x , ε ) := J 1 , β ( x , ε ) − J g [ U x , ε ( β ) ] + 2 1 ⟨ ( L x , ε ( β ) ) − 1 S x , ε ( β ) , S x , ε ( β ) ⟩ . If I I ˚ g ( x ) = 0 \mathring{\mathrm{II}}_g(x)=0 II ˚ g ( x ) = 0 , then
∥ S x , ε ( β ) ∥ ( X x , ε ( β ) ) ∗ + ∥ w x , ε ( β ) ∥ H 1 ( M ) ≤ C ε 2 \left\|\mathcal{S}^{(\beta)}_{x,\varepsilon}\right\|_{(\mathcal{X}^{(\beta)}_{x,\varepsilon})^*}+\left\|w^{(\beta)}_{x,\varepsilon}\right\|_{H^1(M)}\leq C\varepsilon^2 S x , ε ( β ) ( X x , ε ( β ) ) ∗ + w x , ε ( β ) H 1 ( M ) ≤ C ε 2 and, with D : = ε ∂ ε D:=\varepsilon\partial_\varepsilon D := ε ∂ ε ,
∣ E β ∣ + ∣ D E β ∣ = o ( ε 4 ) . (123) |\mathcal{E}_\beta|+|D\mathcal{E}_\beta|=o(\varepsilon^4). \tag*{(123)} ∣ E β ∣ + ∣ D E β ∣ = o ( ε 4 ) . (123) These estimates are uniform on compact subsets of the umbilic locus in C \mathcal{C} C .
Suppose that the boundary is totally umbilic in the representative satisfying H g = 0 H_g=0 H g = 0 . If ( u ℓ ) ⊂ S (u_\ell)\subset\mathcal{S} ( u ℓ ) ⊂ S is a non-precompact sequence of constrained critical points with
J g [ u ℓ ] ⟶ S ∗ , \mathcal{J}_g[u_\ell]\longrightarrow S_*, J g [ u ℓ ] ⟶ S ∗ , then, after passing to a subsequence and making one fixed global sign change, there are
x ℓ ⟶ p ∈ ∂ M , ε ℓ ↓ 0 , x_\ell\longrightarrow p\in\partial M,\qquad\varepsilon_\ell\downarrow0, x ℓ ⟶ p ∈ ∂ M , ε ℓ ↓ 0 ,
such that, for all sufficiently large ℓ \ell ℓ ,
u ℓ = Φ β ( x ℓ , ε ℓ ) , D J 1 , β ( x ℓ , ε ℓ ) = 0. (124) u_{\ell}=\Phi_{\beta}(x_{\ell},\varepsilon_{\ell}),\qquad D\mathcal{J}_{1,\beta}(x_{\ell},\varepsilon_{\ell})=0. \tag*{(124)} u ℓ = Φ β ( x ℓ , ε ℓ ) , D J 1 , β ( x ℓ , ε ℓ ) = 0. (124)
Proof. The bubble truncation error in rescaled coordinates has Dirichlet norm O ( R β − α ) = o ( ε 2 ) O(R_{\beta}^{-\alpha})=o(\varepsilon^{2}) O ( R β − α ) = o ( ε 2 ) , α = ( n − 2 ) / 2 \alpha=(n-2)/2 α = ( n − 2 ) /2 . The same bound holds after logarithmic scale differentiation and for the normalized Jacobi fields. The frame and gap estimates of Theorem 3.6, Lemma 6.1 , and Remark 5.5 therefore apply uniformly. Lemma 8.3 , followed by the inversion in Lemmas 5.7 and 5.9 , gives ∥ P ( β ) ∥ + ∥ w ( β ) ∥ H 1 ≤ C ε 2 \|\mathcal{P}^{(\beta)}\|+\|w^{(\beta)}\|_{H^{1}}\leq C\varepsilon^{2} ∥ P ( β ) ∥ + ∥ w ( β ) ∥ H 1 ≤ C ε 2 .
Suppress x , ε , β x,\varepsilon,\beta x , ε , β and put v = L − 1 P v=\mathcal{L}^{-1}\mathcal{P} v = L − 1 P . At a fixed point of M M M ,
D p h u = b ( Y ) u + A a n n , ∥ b ∥ ∞ ≤ C , ∥ A a n n ∥ E ≤ C R β − α . \mathbf{D}^{\mathrm{ph}}u=b(Y)u+\mathcal{A}^{\mathrm{ann}},\qquad\|b\|_{\infty}\leq C,\qquad\|\mathcal{A}^{\mathrm{ann}}\|_{E}\leq CR_{\beta}^{-\alpha}. D ph u = b ( Y ) u + A ann , ∥ b ∥ ∞ ≤ C , ∥ A ann ∥ E ≤ C R β − α .
Where the bubble cutoff equals one, b = D log t − α − Y ⋅ ∇ log U + b=\mathbf{D}\log t-\alpha-Y\cdot\nabla\log U_{+} b = D log t − α − Y ⋅ ∇ log U + . The estimates for N N N and Q Q Q in Proposition 6.4 control this bounded term. On the annulus, (98) gives a boundary-potential coefficient with L n − 1 L^{n-1} L n − 1 norm O ( R β − 1 ) O(R_{\beta}^{-1}) O ( R β − 1 ) . The weighted finite-rank coordinate derivatives are bounded. Thus the nonlinear gradient remainder satisfies
∥ N β ( w ) ∥ ≤ C ∥ w ∥ 1 + γ , ∥ D w N β ( w ) ∥ ≤ C ∥ w ∥ γ , \|\mathcal{N}_{\beta}(w)\|\leq C\|w\|^{1+\gamma},\qquad\|D_{w}\mathcal{N}_{\beta}(w)\|\leq C\|w\|^{\gamma}, ∥ N β ( w ) ∥ ≤ C ∥ w ∥ 1 + γ , ∥ D w N β ( w ) ∥ ≤ C ∥ w ∥ γ ,
∥ D p a r N β ( w ) ∥ ≤ C ( ∥ w ∥ 1 + γ + R β − α ∥ w ∥ γ + R β − 1 ∥ w ∥ ) . \|D^{\mathrm{par}}\mathcal{N}_{\beta}(w)\|\leq C\left(\|w\|^{1+\gamma}+R_{\beta}^{-\alpha}\|w\|^{\gamma}+R_{\beta}^{-1}\|w\|\right). ∥ D par N β ( w ) ∥ ≤ C ( ∥ w ∥ 1 + γ + R β − α ∥ w ∥ γ + R β − 1 ∥ w ∥ ) .
Here D p a r D^{\mathrm{par}} D par differentiates the parameter at fixed w w w in the complement coordinates. For ∥ w ∥ = O ( ε 2 ) \|w\|=O(\varepsilon^{2}) ∥ w ∥ = O ( ε 2 ) the three orders are
ε 2 + 2 γ , ε β α + 2 γ , ε β + 2 . \varepsilon^{2+2\gamma},\qquad\varepsilon^{\beta\alpha+2\gamma},\qquad\varepsilon^{\beta+2}. ε 2 + 2 γ , ε β α + 2 γ , ε β + 2 .
Since β α > 2 \beta\alpha>2 β α > 2 and β > 2 γ \beta>2\gamma β > 2 γ , all are O ( ε 2 + 2 γ ) O(\varepsilon^{2+2\gamma}) O ( ε 2 + 2 γ ) . Using (121) –(122) and differentiating the projected equation gives, with D \mathbf{D} D denoting the total logarithmic scale derivative in the complement coordinates,
∥ w ∥ + ∥ D w ∥ + ∥ v ∥ + ∥ D v ∥ ≤ C ε 2 . \|w\|+\|\mathbf{D}w\|+\|v\|+\|\mathbf{D}v\|\leq C\varepsilon^{2}. ∥ w ∥ + ∥ D w ∥ + ∥ v ∥ + ∥ D v ∥ ≤ C ε 2 .
The equation L ( w + v ) = − N β ( w ) \mathcal{L}(w+v)=-\mathcal{N}_{\beta}(w) L ( w + v ) = − N β ( w ) and its derivative then give
∥ w + v ∥ + ∥ D ( w + v ) ∥ ≤ C ε 2 + 2 γ . \|w+v\|+\|\mathbf{D}(w+v)\|\leq C\varepsilon^{2+2\gamma}. ∥ w + v ∥ + ∥ D ( w + v ) ∥ ≤ C ε 2 + 2 γ .
For the energy remainder after the linear and quadratic terms,
∣ Q β ( w ) ∣ ≤ C ∥ w ∥ 2 + γ , |Q_{\beta}(w)|\leq C\|w\|^{2+\gamma}, ∣ Q β ( w ) ∣ ≤ C ∥ w ∥ 2 + γ ,
∣ D Q β ( w ) ∣ ≤ C ( ∥ w ∥ 2 + γ + R β − α ∥ w ∥ 1 + γ + R β − 1 ∥ w ∥ 2 + ∥ N β ( w ) ∥ ∥ D w ∥ ) = o ( ε 4 ) . |\mathbf{D}Q_{\beta}(w)|\leq C\left(\|w\|^{2+\gamma}+R_{\beta}^{-\alpha}\|w\|^{1+\gamma}+R_{\beta}^{-1}\|w\|^{2}+\|\mathcal{N}_{\beta}(w)\|\|\mathbf{D}w\|\right)=o(\varepsilon^{4}). ∣ D Q β ( w ) ∣ ≤ C ( ∥ w ∥ 2 + γ + R β − α ∥ w ∥ 1 + γ + R β − 1 ∥ w ∥ 2 + ∥ N β ( w ) ∥∥ D w ∥ ) = o ( ε 4 ) .
The energy terms have orders ε 4 + 2 γ \varepsilon^{4+2\gamma} ε 4 + 2 γ , ε β α + 2 + 2 γ \varepsilon^{\beta\alpha+2+2\gamma} ε β α + 2 + 2 γ , and ε β + 4 \varepsilon^{\beta+4} ε β + 4 , all o ( ε 4 ) o(\varepsilon^{4}) o ( ε 4 ) . Lemma 5.8 and completing the square give
E β = 1 2 ⟨ L ( w + v ) , w + v ⟩ + Q β ( w ) . \mathcal{E}_{\beta}=\frac{1}{2}\langle\mathcal{L}(w+v),w+v\rangle+Q_{\beta}(w). E β = 2 1 ⟨ L ( w + v ) , w + v ⟩ + Q β ( w ) .
The preceding bounds and the bounded weighted scale derivative of L \mathcal{L} L prove (123) .
For the sequence of constrained critical points, Lemma 4.6 gives a single signed bubble. The o ( 1 ) o(1) o ( 1 ) cutoff comparison (112) and Lemma 6.1 give u ℓ = ret U x ℓ , ε ℓ ( β ) ( w ~ ℓ ) u_{\ell}=\operatorname{ret}U_{x_{\ell},\varepsilon_{\ell}}^{(\beta)}(\widetilde{w}_{\ell}) u ℓ = ret U x ℓ , ε ℓ ( β ) ( w ℓ ) with w ~ ℓ → 0 \widetilde{w}_{\ell}\to0 w ℓ → 0 in the complement. Criticality makes w ~ ℓ \widetilde{w}_{\ell} w ℓ solve (111) . Uniqueness in Lemma 5.9 identifies it with w x ℓ , ε ℓ ( β ) w_{x_{\ell},\varepsilon_{\ell}}^{(\beta)} w x ℓ , ε ℓ ( β ) , and the chain rule gives (124) . □
Proposition 8.5 (Fourth-order one-bubble expansion at umbilic points). Under the hypotheses and cutoff convention of Lemma 8.3 , let J 1 , β \mathcal{J}_{1,\beta} J 1 , β be the reduced functional of Lemma 8.4 . For p ∈ C p\in C p ∈ C with I I ˚ g ( p ) = 0 \mathring{\mathrm{II}}_{g}(p)=0 II ˚ g ( p ) = 0 , the limit
Ξ g 0 ( p ) : = lim ε ↓ 0 ε − 4 ( J g [ U p , ε ( β ) ] S ∗ − 1 ) \Xi_{g}^{0}(p):=\lim_{\varepsilon\downarrow0}\varepsilon^{-4}\left(\frac{J_{g}[U_{p,\varepsilon}^{(\beta)}]}{S_{*}}-1\right) Ξ g 0 ( p ) := ε ↓ 0 lim ε − 4 ( S ∗ J g [ U p , ε ( β ) ] − 1 )
exists. With S 2 , p \mathscr{S}_{2,p} S 2 , p and L 0 \mathcal{L}_0 L 0 as in Lemma 8.3, define
Ξ g r e d ( p ) : = Ξ g 0 ( p ) − 1 2 S ∗ ⟨ L 0 − 1 S 2 , p , S 2 , p ⟩ . (125) \Xi_g^{\mathrm{red}}(p) := \Xi_g^0(p) - \frac{1}{2S_*}\left\langle\mathcal{L}_0^{-1}\mathscr{S}_{2,p},\mathscr{S}_{2,p}\right\rangle. \tag*{(125)} Ξ g red ( p ) := Ξ g 0 ( p ) − 2 S ∗ 1 ⟨ L 0 − 1 S 2 , p , S 2 , p ⟩ . (125)
These coefficients are independent of β \beta β in the stated range. Uniformly on compact subsets of the umbilic locus in C \mathcal{C} C ,
J 1 , β ( p , ε ) S ∗ = 1 + Ξ g r e d ( p ) ε 4 + o ( ε 4 ) , (126) \frac{\mathcal{J}_{1,\beta}(p,\varepsilon)}{S_*}=1+\Xi_g^{\mathrm{red}}(p)\varepsilon^4+o(\varepsilon^4), \tag*{(126)} S ∗ J 1 , β ( p , ε ) = 1 + Ξ g red ( p ) ε 4 + o ( ε 4 ) , (126)
1 S ∗ ∂ ε J 1 , β ( p , ε ) = 4 Ξ g r e d ( p ) ε 3 + o ( ε 3 ) . (127) \frac{1}{S_*}\partial_\varepsilon\mathcal{J}_{1,\beta}(p,\varepsilon)=4\Xi_g^{\mathrm{red}}(p)\varepsilon^3+o(\varepsilon^3). \tag*{(127)} S ∗ 1 ∂ ε J 1 , β ( p , ε ) = 4 Ξ g red ( p ) ε 3 + o ( ε 3 ) . (127)
In particular, the third-order coefficient of the reduced functional vanishes at every such point.
Proof. For n ≥ 7 n\geq7 n ≥ 7 , the fourth gradient moment ∫ ∣ Y ∣ 4 ∣ ∇ U + ∣ 2 \int|Y|^4|\nabla U_+|^2 ∫ ∣ Y ∣ 4 ∣∇ U + ∣ 2 and fourth trace moment ∫ ∣ Y ′ ∣ 4 U + q \int|Y'|^4U_+^q ∫ ∣ Y ′ ∣ 4 U + q are finite. The fourth-order Fermi expansion has fifth-order remainder O ( ε 5 log R ) O(\varepsilon^5\log R) O ( ε 5 log R ) at n = 7 n=7 n = 7 and O ( ε 5 ) O(\varepsilon^5) O ( ε 5 ) for n ≥ 8 n\geq8 n ≥ 8 . For m ≤ 3 m\leq3 m ≤ 3 , its omitted gradient tail is
O ( ε m R β m + 2 − n ) = O ( ε β ( n − 2 ) + m ( 1 − β ) ) = o ( ε 4 ) . O(\varepsilon^mR_\beta^{m+2-n})=O(\varepsilon^{\beta(n-2)+m(1-\beta)})=o(\varepsilon^4). O ( ε m R β m + 2 − n ) = O ( ε β ( n − 2 ) + m ( 1 − β ) ) = o ( ε 4 ) .
The flat and fourth-order tails are O ( R β 2 − n ) = o ( ε 4 ) O(R_\beta^{2-n})=o(\varepsilon^4) O ( R β 2 − n ) = o ( ε 4 ) and ε 4 O ( R β 6 − n ) = o ( ε 4 ) \varepsilon^4O(R_\beta^{6-n})=o(\varepsilon^4) ε 4 O ( R β 6 − n ) = o ( ε 4 ) . Scalar and boundary terms have at least the same decay. These bounds persist after one weighted scale derivative.
The first three coefficients vanish at the umbilic point with zero mean curvature. Thus, with D = ε ∂ ε D=\varepsilon\partial_\varepsilon D = ε ∂ ε ,
J g [ U p , ε ( β ) ] S ∗ = 1 + Ξ g 0 ( p ) ε 4 + e 0 , ∣ e 0 ∣ + ∣ D e 0 ∣ = o ( ε 4 ) . \frac{\mathcal{J}_g[U_{p,\varepsilon}^{(\beta)}]}{S_*}=1+\Xi_g^0(p)\varepsilon^4+e_0,\qquad|e_0|+|De_0|=o(\varepsilon^4). S ∗ J g [ U p , ε ( β ) ] = 1 + Ξ g 0 ( p ) ε 4 + e 0 , ∣ e 0 ∣ + ∣ D e 0 ∣ = o ( ε 4 ) .
The differentiated pairing estimate (116) gives
⟨ ( L ( β ) ) − 1 S ( β ) , S ( β ) ⟩ = ε 4 ⟨ L 0 − 1 S 2 , p , S 2 , p ⟩ + e 1 , ∣ e 1 ∣ + ∣ D e 1 ∣ = o ( ε 4 ) . \left\langle\left(\mathcal{L}^{(\beta)}\right)^{-1}\mathscr{S}^{(\beta)},\mathscr{S}^{(\beta)}\right\rangle
=\varepsilon^4\left\langle\mathcal{L}_0^{-1}\mathscr{S}_{2,p},\mathscr{S}_{2,p}\right\rangle+e_1,\qquad|e_1|+|De_1|=o(\varepsilon^4). ⟨ ( L ( β ) ) − 1 S ( β ) , S ( β ) ⟩ = ε 4 ⟨ L 0 − 1 S 2 , p , S 2 , p ⟩ + e 1 , ∣ e 1 ∣ + ∣ D e 1 ∣ = o ( ε 4 ) .
Insert these identities into the quadratic expansion of Lemma 8.4, whose remaining error has the same bound. This proves (125) and (126) ; applying ε − 1 D \varepsilon^{-1}D ε − 1 D proves (127) . The tail estimates and the limiting source equation contain no β \beta β , so both coefficients are independent of the admissible cutoff exponent.
Conformal invariance
Lemma 8.6 (Conformal invariance of the reduced fourth-order coefficient). Assume n ≥ 7 n\geq7 n ≥ 7 , ( Y + γ , ( B + G ) + ) (Y_+^{\gamma},(B^{G}_+)^{+}) ( Y + γ , ( B + G ) + ) , and that the boundary is totally umbilic near p p p . Let g g g be the fixed representative with H g = 0 H_g=0 H g = 0 . Choose a globally defined conformal metric g ^ = φ 4 / ( n − 2 ) g \widehat{g}=\varphi^{4/(n-2)}g g = φ 4/ ( n − 2 ) g with H g ^ = 0 H_{\widehat{g}}=0 H g = 0 , conformal-Fermi volume density 1 + O ( ∣ x ∣ 6 ) 1+O(|x|^6) 1 + O ( ∣ x ∣ 6 ) at p p p , and φ ( p ) = 1 \varphi(p)=1 φ ( p ) = 1 , d φ ( p ) = 0 d\varphi(p)=0 d φ ( p ) = 0 . Construct the one-bubble reductions for g g g and g ^ \widehat{g} g using the same admissible β ∈ ( 4 / ( n − 2 ) , 1 ) \beta\in(4/(n-2),1) β ∈ ( 4/ ( n − 2 ) , 1 ) and the same cutoff profile. Then Ξ g r e d ( p ) = Ξ g ^ r e d ( p ) \Xi_g^{\mathrm{red}}(p)=\Xi_{\widehat{g}}^{\mathrm{red}}(p) Ξ g red ( p ) = Ξ g red ( p ) .
More precisely, let B , B ′ B,B' B , B ′ be the fourth-order coefficients in energy units of the two uncorrected families after the conformal change, let S , S ′ ∈ X + ∗ \mathscr{S},\mathscr{S}'\in X_+^* S , S ′ ∈ X + ∗ be their second-order source coefficients, and let L 0 : X + → X + ∗ \mathcal{L}_0:X_+\to X_+^* L 0 : X + → X + ∗ be the flat complementary Hessian. If, after trace normalization and modulation, the two families differ at order ε 2 \varepsilon^2 ε 2 by a ∈ X + a\in X_+ a ∈ X + , then
B ′ = B + S [ a ] + 1 2 ⟨ L 0 a , a ⟩ , S ′ = S + L 0 a , B'=B+\mathscr{S}[a]+\frac{1}{2}\langle\mathcal{L}_0a,a\rangle,\qquad\mathscr{S}'=\mathscr{S}+\mathcal{L}_0a, B ′ = B + S [ a ] + 2 1 ⟨ L 0 a , a ⟩ , S ′ = S + L 0 a , B ′ − 1 2 ⟨ L 0 − 1 S ′ , S ′ ⟩ = B − 1 2 ⟨ L 0 − 1 S , S ⟩ . (128) B'-\frac{1}{2}\left\langle\mathcal{L}_0^{-1}\mathscr{S}',\mathscr{S}'\right\rangle
=B-\frac{1}{2}\left\langle\mathcal{L}_0^{-1}\mathscr{S},\mathscr{S}\right\rangle. \tag*{(128)} B ′ − 2 1 ⟨ L 0 − 1 S ′ , S ′ ⟩ = B − 2 1 ⟨ L 0 − 1 S , S ⟩ . (128) The reduced coefficient is invariant under this change of approximate family.
Proof. By Lemma 3.1, multiplication C φ v = φ v C_\varphi v=\varphi v C φ v = φ v is an isometry of the energy spaces and preserves the boundary constraint:
⟨ C φ v , C φ w ⟩ E , g = ⟨ v , w ⟩ E , g ^ , ∥ C φ v ∥ L q ( g ) = ∥ v ∥ L q ( g ^ ) , J g [ C φ v ] = J g ^ [ v ] . (129) \langle C_\varphi v,C_\varphi w\rangle_{E,g}=\langle v,w\rangle_{E,\widehat{g}},\qquad
\|C_\varphi v\|_{L^q(g)}=\|v\|_{L^q(\widehat{g})},\qquad
\mathcal{J}_g[C_\varphi v]=\mathcal{J}_{\widehat{g}}[v]. \tag*{(129)} ⟨ C φ v , C φ w ⟩ E , g = ⟨ v , w ⟩ E , g , ∥ C φ v ∥ L q ( g ) = ∥ v ∥ L q ( g ) , J g [ C φ v ] = J g [ v ] . (129)
It also commutes with trace normalization. We compare the two bubble families on the same constraint.
Align the tangential frames at p p p . Since φ ( p ) = 1 \varphi(p)=1 φ ( p ) = 1 , d φ ( p ) = 0 d\varphi(p)=0 d φ ( p ) = 0 , and both boundaries are totally geodesic, the Fermi transition has no quadratic term:
Φ ^ p − 1 ∘ Φ p ( x ) = x + C 3 ( x ) + O ( ∣ x ∣ 4 ) , φ ( Φ p ( x ) ) = 1 + q 2 ( x ) + O ( ∣ x ∣ 3 ) , (130) \widehat{\Phi}_{p}^{-1}\circ\Phi_p(x)=x+C_3(x)+O(|x|^4),\qquad\varphi(\Phi_p(x))=1+q_2(x)+O(|x|^3), \tag*{(130)} Φ p − 1 ∘ Φ p ( x ) = x + C 3 ( x ) + O ( ∣ x ∣ 4 ) , φ ( Φ p ( x )) = 1 + q 2 ( x ) + O ( ∣ x ∣ 3 ) , (130)
where C 3 C_3 C 3 and q 2 q_2 q 2 are homogeneous of degrees three and two. Before trace normalization, critical rescaling gives
ε 2 a p 0 + o D 1 , 2 ( ε 2 ) , a p 0 = t ∗ ( q 2 U + + C 3 ⋅ ∇ U + ) . (131) \varepsilon^2a_p^0+o_{D^{1,2}}(\varepsilon^2),\qquad a_p^0=t_*(q_2U_++C_3\cdot\nabla U_+). \tag*{(131)} ε 2 a p 0 + o D 1 , 2 ( ε 2 ) , a p 0 = t ∗ ( q 2 U + + C 3 ⋅ ∇ U + ) . (131)
Its gradient is O ( ( 1 + ∣ Y ∣ ) 3 − n ) O((1+|Y|)^{3-n}) O (( 1 + ∣ Y ∣ ) 3 − n ) , and hence
∥ ∇ a p 0 ∥ L 2 ( ∣ Y ∣ > R ) ≤ C R − ( n − 6 ) / 2 . \|\nabla a_p^0\|_{L^2(|Y|>R)}\le CR^{-(n-6)/2}. ∥∇ a p 0 ∥ L 2 ( ∣ Y ∣ > R ) ≤ C R − ( n − 6 ) /2 .
The next Taylor terms have Dirichlet norm at most C ε 2 ℓ n ( R ) C\varepsilon^2\ell_n(R) C ε 2 ℓ n ( R ) , with ℓ n \ell_n ℓ n as in Lemma 8.3. Higher products are absorbed by this bound on ε R < 1 \varepsilon R<1 εR < 1 . Trace normalization and projection onto the flat complement give
a p = P + ⊥ ( a p 0 − U ∗ ( ∫ ∂ R + n U ∗ q − 1 a p 0 ) ) ∈ X + . a_p=P_+^\perp\left(a_p^0-U_*\left(\int_{\partial\mathbb{R}^n_+}U_*^{q-1}a_p^0\right)\right)\in X_+. a p = P + ⊥ ( a p 0 − U ∗ ( ∫ ∂ R + n U ∗ q − 1 a p 0 ) ) ∈ X + .
The uniform modulation estimate gives weighted parameter changes O ( ε 2 ) O(\varepsilon^2) O ( ε 2 ) , or changes in the center and scale d ∂ ( p ε , p ) + ∣ ε ~ − ε ∣ = O ( ε 3 ) d_{\partial}(p_\varepsilon,p)+|\widetilde{\varepsilon}-\varepsilon|=O(\varepsilon^3) d ∂ ( p ε , p ) + ∣ ε − ε ∣ = O ( ε 3 ) . Thus
C φ U ^ p ε , ε ~ ( β ) = ret U p , ε ( β ) ( ε 2 B p , ε ( β ) a p + o H 1 ( ε 2 ) ) . (132) \mathcal{C}_{\varphi}\widehat{U}_{p_\varepsilon,\widetilde{\varepsilon}}^{(\beta)}
=\operatorname{ret} \mathcal{U}_{p,\varepsilon}^{(\beta)}
\left(\varepsilon^2\mathcal{B}_{p,\varepsilon}^{(\beta)}a_p+o_{H^1}(\varepsilon^2)\right). \tag*{(132)} C φ U p ε , ε ( β ) = ret U p , ε ( β ) ( ε 2 B p , ε ( β ) a p + o H 1 ( ε 2 ) ) . (132)
Let Δ χ R β \Delta\chi_{R_\beta} Δ χ R β denote the difference between the two cutoff functions after the change of Fermi coordinates. On the cutoff annulus the rescaled coordinate displacement is O ( ε 2 R β 3 ) O(\varepsilon^2R_\beta^3) O ( ε 2 R β 3 ) , so its ratio to the annular radius is O ( ε 2 R β 2 ) = o ( 1 ) O(\varepsilon^2R_\beta^2)=o(1) O ( ε 2 R β 2 ) = o ( 1 ) . The mean-value theorem gives
∣ Δ χ R β ∣ ≤ C ε 2 R β 2 , ∣ ∇ Δ χ R β ∣ ≤ C ε 2 R β . |\Delta\chi_{R_\beta}|\le C\varepsilon^2R_\beta^2,\qquad|\nabla\Delta\chi_{R_\beta}|\le C\varepsilon^2R_\beta. ∣Δ χ R β ∣ ≤ C ε 2 R β 2 , ∣∇Δ χ R β ∣ ≤ C ε 2 R β .
Using U + = O ( R β 2 − n ) U_+=O(R_\beta^{2-n}) U + = O ( R β 2 − n ) and ∇ U + = O ( R β 1 − n ) \nabla U_+=O(R_\beta^{1-n}) ∇ U + = O ( R β 1 − n ) there,
∥ ∇ ( Δ χ R β U + ) ∥ 2 ≤ C ε 2 R β − ( n − 6 ) / 2 . \|\nabla(\Delta\chi_{R_\beta}U_+)\|_2\le C\varepsilon^2R_\beta^{-(n-6)/2}. ∥∇ ( Δ χ R β U + ) ∥ 2 ≤ C ε 2 R β − ( n − 6 ) /2 .
Thus, if Δ ε \Delta_\varepsilon Δ ε is the critically rescaled profile difference after normalization and modulation, the leading tail and the next Taylor terms give
∥ ε − 2 Δ ε − a p ∥ D 1 , 2 ≤ C ( R β − ( n − 6 ) / 2 + ε ℓ n ( R β ) ) + o ( 1 ) = o ( 1 ) . \|\varepsilon^{-2}\Delta_\varepsilon-a_p\|_{D^{1,2}}
\le C\left(R_\beta^{-(n-6)/2}+\varepsilon\ell_n(R_\beta)\right)+o(1)=o(1). ∥ ε − 2 Δ ε − a p ∥ D 1 , 2 ≤ C ( R β − ( n − 6 ) /2 + ε ℓ n ( R β ) ) + o ( 1 ) = o ( 1 ) .
The displayed bounds account for the cutoff-transition error and the higher-order Taylor terms. The difference e ε = ε − 2 Δ ε − χ R β a p e_\varepsilon=\varepsilon^{-2}\Delta_\varepsilon-\chi_{R_\beta}a_p e ε = ε − 2 Δ ε − χ R β a p tends to zero in the Dirichlet norm. Before the final projection its support lies in a half-ball of radius C ε R β C\varepsilon R_\beta C ε R β on M M M , so Sobolev gives
∥ T p , ε e ε ∥ L 2 ( M ) ≤ C ε R β ∥ e ε ∥ D 1 , 2 . \|T_{p,\varepsilon}e_\varepsilon\|_{L^2(M)}
\le C\varepsilon R_\beta\|e_\varepsilon\|_{D^{1,2}}. ∥ T p , ε e ε ∥ L 2 ( M ) ≤ C ε R β ∥ e ε ∥ D 1 , 2 .
The bounded energy projection and frame convergence therefore give the stated o H 1 ( ε 2 ) o_{H^1}(\varepsilon^2) o H 1 ( ε 2 ) comparison.
The Taylor expansion on the common constraint is
J g [ ret U p , ε ( β ) ( ε 2 B p , ε ( β ) z ) ] = S ∗ + ε 4 ( B + S [ z ] + 1 2 ⟨ L 0 z , z ⟩ ) + o ( ε 4 ) (133) \mathcal{J}_g\left[\operatorname{ret}\mathcal{U}_{p,\varepsilon}^{(\beta)}
\left(\varepsilon^2\mathcal{B}_{p,\varepsilon}^{(\beta)}z\right)\right]
=S_*+\varepsilon^4\left(B+\mathcal{S}[z]+\frac{1}{2}\langle\mathcal{L}_0z,z\rangle\right)+o(\varepsilon^4) \tag*{(133)} J g [ ret U p , ε ( β ) ( ε 2 B p , ε ( β ) z ) ] = S ∗ + ε 4 ( B + S [ z ] + 2 1 ⟨ L 0 z , z ⟩ ) + o ( ε 4 ) (133)
for each fixed z ∈ X + z\in X_+ z ∈ X + , uniformly on compact subsets of X + X_+ X + . Here the source converges in the model dual, the Hessian pairings converge on fixed model functions, and the C 2 , γ C^{2,\gamma} C 2 , γ Taylor error is O ( ε 4 + 2 γ ) O(\varepsilon^{4+2\gamma}) O ( ε 4 + 2 γ ) .
After conformal transport the second family has leading coordinates a p + z a_p+z a p + z in this common parametrization. This follows first for compactly supported smooth model functions by the Fermi transition above, and then for fixed z ∈ X + z\in X_+ z ∈ X + by Dirichlet approximation and the uniform bounds for rescaling and projection. The limiting normal frames agree, so projection preserves this leading coordinate comparison. The full source estimate (121) and the bounded Hessian give ∥ D J g ∥ ≤ C ε 2 \lVert D\mathcal{J}_g\rVert\le C\varepsilon^2 ∥ D J g ∥ ≤ C ε 2 throughout these O ( ε 2 ) O(\varepsilon^2) O ( ε 2 ) neighborhoods. For a profile error e = o H 1 ( ε 2 ) e=o_{H^1}(\varepsilon^2) e = o H 1 ( ε 2 ) the mean-value estimate is therefore
∣ J g [ u + e ] − J g [ u ] ∣ ≤ sup 0 ≤ s ≤ 1 ∥ D J g [ u + s e ] ∥ ∥ e ∥ H 1 = o ( ε 4 ) . \left|\mathcal{J}_g[u+e]-\mathcal{J}_g[u]\right|
\le\sup_{0\le s\le1}\left\|D\mathcal{J}_g[u+se]\right\|\lVert e\rVert_{H^1}
=o(\varepsilon^4). ∣ J g [ u + e ] − J g [ u ] ∣ ≤ 0 ≤ s ≤ 1 sup ∥ D J g [ u + se ] ∥ ∥ e ∥ H 1 = o ( ε 4 ) .
For every fixed z ∈ X + z\in X_+ z ∈ X + the common expansion gives
B ′ + S ′ [ z ] + 1 2 ⟨ L 0 z , z ⟩ = B + S [ a p + z ] + 1 2 ⟨ L 0 ( a p + z ) , a p + z ⟩ . B'+\mathcal{S}'[z]+\frac{1}{2}\langle\mathcal{L}_0z,z\rangle
=
B+\mathcal{S}[a_p+z]+\frac{1}{2}\langle\mathcal{L}_0(a_p+z),a_p+z\rangle. B ′ + S ′ [ z ] + 2 1 ⟨ L 0 z , z ⟩ = B + S [ a p + z ] + 2 1 ⟨ L 0 ( a p + z ) , a p + z ⟩ .
Comparison of the constant and linear terms gives the first two identities of (128) . Substituting S ′ = S + L 0 a p \mathcal{S}'=\mathcal{S}+\mathcal{L}_0a_p S ′ = S + L 0 a p into the quadratic form cancels the added linear and quadratic terms and proves the third. The approximation (115) identifies the minima of these quadratic expressions with the reduced coefficients. Division by S ∗ S_* S ∗ proves invariance. The parameter changes contribute no fourth-order term: ε ^ / ε = 1 + O ( ε 2 ) \widehat{\varepsilon}/\varepsilon=1+O(\varepsilon^2) ε / ε = 1 + O ( ε 2 ) , the quartic coefficient is continuous in the center, and the quadratic and cubic coefficients vanish throughout the locally totally umbilic neighborhood.
The normalized representative. Conformal Fermi coordinates were constructed by Marques [36 Proposition 3.1] . The conformal-Fermi construction used here has φ ( p ) = 1 \varphi(p)=1 φ ( p ) = 1 and tangential d φ ( p ) = 0 d\varphi(p)=0 d φ ( p ) = 0 ; see [30 Lemma 2.2] . The mean-curvature transformation gives the normal derivative zero as well. Since the first derivatives of the original Fermi metric vanish, the corrections that normalize the volume density begin in degree two. For this fixed smooth metric, apply [9 Proposition 2.2] with N = 6 N=6 N = 6 , giving a local metric g 0 g_0 g 0 with volume density 1 + O ( ∣ x ∣ 6 ) 1+O(|x|^6) 1 + O ( ∣ x ∣ 6 ) and H g 0 = O ( ∣ x ′ ∣ 5 ) H_{g_0}=O(|x'|^5) H g 0 = O ( ∣ x ′ ∣ 5 ) . Choose χ ∈ C ∞ ( [ 0 , ∞ ) ) \chi\in C^\infty([0,\infty)) χ ∈ C ∞ ([ 0 , ∞ )) equal to one near zero. In these Fermi coordinates, the factor
ψ = exp ( x n f ( x ′ ) χ ( x n ) ) , f = ( n − 2 ) H g 0 / 2 = O ( ∣ x ′ ∣ 5 ) , \psi=\exp\bigl(x_nf(x')\chi(x_n)\bigr),\qquad f=(n-2)H_{g_0}/2=O(|x'|^5), ψ = exp ( x n f ( x ′ ) χ ( x n ) ) , f = ( n − 2 ) H g 0 /2 = O ( ∣ x ′ ∣ 5 ) ,
has boundary value one and ∂ ν g 0 ψ = ( n − 2 ) H g 0 / 2 \partial_{\nu_{g_0}}\psi=(n-2)H_{g_0}/2 ∂ ν g 0 ψ = ( n − 2 ) H g 0 /2 . It makes H H H vanish and, being 1 + O ( ∣ x ∣ 6 ) 1+O(|x|^6) 1 + O ( ∣ x ∣ 6 ) , preserves the required Taylor coefficients. The resulting factor relative to g g g has zero normal derivative on the boundary. Multiply its logarithm by a smooth cutoff equal to one near p p p with zero boundary normal derivative, and extend by zero. The global positive factor preserves H = 0 H=0 H = 0 and the normalization at p p p . The energy isometry preserves strict coercivity. □
The curvature formula and its sign
Theorem 8.7 (Exact formula for the fourth-order coefficient). Assume n ≥ 7 n\ge7 n ≥ 7 , ( Y δ + ) (Y_\delta^+) ( Y δ + ) , and ( B G 5 + ) (\mathrm{BG}^{5+}) ( BG 5 + ) . Let g g g be the fixed representative with H g = 0 H_g=0 H g = 0 , and suppose that the boundary is totally umbilic near p p p . Let g ^ = φ 4 / ( n − 2 ) g \widehat{g}=\varphi^{4/(n-2)}g g = φ 4/ ( n − 2 ) g be the normalized conformal-Fermi representative from Lemma 8.6. In g ^ \widehat{g} g -Fermi coordinates at p p p , put
T a b = R ^ g , ν a v b ( p ) , W ‾ = W g ∣ ∂ M ( p ) , D n = ( n − 1 ) ( n − 3 ) ( n − 4 ) ( n − 5 ) ( n − 6 ) . T_{ab}=\widehat{R}_{g,\,\nu avb}(p),\qquad
\overline{W}=W_{g|_{\partial M}}(p),\qquad
D_n=(n-1)(n-3)(n-4)(n-5)(n-6). T ab = R g , ν a v b ( p ) , W = W g ∣ ∂ M ( p ) , D n = ( n − 1 ) ( n − 3 ) ( n − 4 ) ( n − 5 ) ( n − 6 ) . Then T T T is trace-free, and the standard-bubble coefficient is
Ξ g 0 ( p ) = 8 − n D n ∣ T ∣ 2 − b n ∣ W ‾ ∣ 2 , b n = n − 2 24 ( n − 1 ) ( n − 3 ) ( n − 5 ) ( n − 6 ) . (134) \Xi_g^0(p)=\frac{8-n}{D_n}|T|^2-b_n|\overline{W}|^2,\qquad
b_n=\frac{n-2}{24(n-1)(n-3)(n-5)(n-6)}. \tag*{(134)} Ξ g 0 ( p ) = D n 8 − n ∣ T ∣ 2 − b n ∣ W ∣ 2 , b n = 24 ( n − 1 ) ( n − 3 ) ( n − 5 ) ( n − 6 ) n − 2 . (134) The second-order source depends only on T T T ; denote it by P T \mathcal{P}_T P T . For T ≠ 0 T\ne0 T = 0 , the coefficient of the corresponding quadratic form is
κ n : = 1 2 S ∗ ⟨ L 0 − 1 P T , P T ⟩ ∣ T ∣ 2 = 3 n 3 + 30 n 2 − 79 n + 6 3 ( n − 2 ) ( n − 1 ) 2 ( n + 1 ) ( n − 3 ) ( n − 4 ) ( n − 5 ) ( n − 6 ) . (135) \kappa_n:=\frac{1}{2S_*}\frac{\langle\mathcal{L}_0^{-1}\mathcal{P}_T,\mathcal{P}_T\rangle}{|T|^2}
=
\frac{3n^3+30n^2-79n+6}{3(n-2)(n-1)^2(n+1)(n-3)(n-4)(n-5)(n-6)}. \tag*{(135)} κ n := 2 S ∗ 1 ∣ T ∣ 2 ⟨ L 0 − 1 P T , P T ⟩ = 3 ( n − 2 ) ( n − 1 ) 2 ( n + 1 ) ( n − 3 ) ( n − 4 ) ( n − 5 ) ( n − 6 ) 3 n 3 + 30 n 2 − 79 n + 6 . (135) Consequently,
Ξ g r e d ( p ) = Ξ g ^ r e d ( p ) = − e n ∣ T ∣ 2 − b n ∣ W ‾ ∣ 2 , (136) \Xi_g^{\mathrm{red}}(p)=\Xi_{\widehat{g}}^{\mathrm{red}}(p)
=-e_n|T|^2-b_n|\overline{W}|^2, \tag*{(136)} Ξ g red ( p ) = Ξ g red ( p ) = − e n ∣ T ∣ 2 − b n ∣ W ∣ 2 , (136) where
e n = 3 n 3 − 18 n 2 + 21 n + 14 3 ( n − 2 ) ( n − 1 ) 2 ( n + 1 ) ( n − 4 ) ( n − 5 ) ( n − 6 ) > 0. (137) e_n=\frac{3n^3-18n^2+21n+14}{3(n-2)(n-1)^2(n+1)(n-4)(n-5)(n-6)}>0. \tag*{(137)} e n = 3 ( n − 2 ) ( n − 1 ) 2 ( n + 1 ) ( n − 4 ) ( n − 5 ) ( n − 6 ) 3 n 3 − 18 n 2 + 21 n + 14 > 0. (137) :::`
The optimal algebraic comparison with the ambient Weyl norm is
Ξ g ^ r e d ( p ) ≤ − c ( n ) ∣ W g ( p ) ∣ g 2 , (138) \Xi^{\mathrm{red}}_{\hat g}(p) \le-c(n)|W_g(p)|^2_g, \tag*{(138)} Ξ g ^ red ( p ) ≤ − c ( n ) ∣ W g ( p ) ∣ g 2 , (138)
with
c ( n ) = 3 n 3 − 18 n 2 + 21 n + 14 12 ( n − 1 ) 2 ( n + 1 ) ( n − 3 ) ( n − 4 ) ( n − 5 ) ( n − 6 ) . (139) c(n)=\frac{3n^3-18n^2+21n+14}{12(n-1)^2(n+1)(n-3)(n-4)(n-5)(n-6)}. \tag*{(139)} c ( n ) = 12 ( n − 1 ) 2 ( n + 1 ) ( n − 3 ) ( n − 4 ) ( n − 5 ) ( n − 6 ) 3 n 3 − 18 n 2 + 21 n + 14 . (139)
Equality in (138) holds precisely when W ‾ = 0 \overline{W}=0 W = 0 . In particular,
Ξ g ^ r e d ( p ) = 0 ⟺ W g ( p ) = 0. (140) \Xi^{\mathrm{red}}_{\hat g}(p)=0 \quad\Longleftrightarrow\quad W_g(p)=0. \tag*{(140)} Ξ g ^ red ( p ) = 0 ⟺ W g ( p ) = 0. (140)
Proof. Here barred curvature tensors belong to the boundary metric, and semicolons denote covariant derivatives in the indicated directions. The conformal-Fermi identities at p p p are
Ric ‾ = 0 , Ric ( ν , ν ) = 0 , Scal ‾ = 0 , Ric ‾ a b = R ν a ν b , R ν a b c = 0 , \overline{\operatorname{Ric}}=0,\qquad\operatorname{Ric}(\nu,\nu)=0,\qquad\overline{\operatorname{Scal}}=0,\qquad\overline{\operatorname{Ric}}_{ab}=R_{\nu a\nu b},\qquad R_{\nu abc}=0, Ric = 0 , Ric ( ν , ν ) = 0 , Scal = 0 , Ric ab = R ν a ν b , R ν ab c = 0 ,
and
Scal ‾ ; a a = − 1 6 ∣ W ‾ ∣ 2 , Ric ν ν ; ν ν + 2 ∣ T ∣ 2 = 0 , T a b ; a b = − 1 2 Scal ; ν ν − ∣ T ∣ 2 ; \overline{\operatorname{Scal}}_{;aa}=-\frac{1}{6}|\overline{W}|^2,\qquad\operatorname{Ric}_{\nu\nu;\nu\nu}+2|T|^2=0,\qquad T_{ab;ab}=-\frac{1}{2}\operatorname{Scal}_{;\nu\nu}-|T|^2; Scal ; aa = − 6 1 ∣ W ∣ 2 , Ric ν ν ; ν ν + 2∣ T ∣ 2 = 0 , T ab ; ab = − 2 1 Scal ; ν ν − ∣ T ∣ 2 ;
see [9 Lemmas 2.4–2.5] .
For the uncorrected expansion, use [9 Section 3, (3-7)–(3-14)] with correction parameter A = 0 A=0 A = 0 . Fix a small chart radius δ > 0 \delta>0 δ > 0 and a smooth radial cutoff χ \chi χ equal to one on [ 0 , 1 ] [0,1] [ 0 , 1 ] and zero on [ 2 , ∞ ) [2,\infty) [ 2 , ∞ ) . In the g ^ \hat g g ^ -Fermi chart put
V ε ( x ) = ( ε ∣ x ′ ∣ 2 + ( ε + x n ) 2 ) ( n − 2 ) / 2 , ψ ε , δ = χ ( ∣ x ∣ / δ ) V ε , b ε = ∫ ∂ M ψ ε , δ q d σ g ^ . V_\varepsilon(x)=\left(\frac{\varepsilon}{|x'|^2+(\varepsilon+x_n)^2}\right)^{(n-2)/2},\qquad\psi_{\varepsilon,\delta}=\chi(|x|/\delta)V_\varepsilon,\qquad b_\varepsilon=\int_{\partial M}\psi_{\varepsilon,\delta}^{q}\,d\sigma_{\hat g}. V ε ( x ) = ( ∣ x ′ ∣ 2 + ( ε + x n ) 2 ε ) ( n − 2 ) /2 , ψ ε , δ = χ ( ∣ x ∣/ δ ) V ε , b ε = ∫ ∂ M ψ ε , δ q d σ g ^ .
The flat bubble identity and integration by parts give
∫ B δ + ∣ ∇ V ε ∣ 2 − S ∗ b ε 2 / q = O δ ( ε n − 2 ) + o ( ε 4 ) . \int_{B_\delta^+}|\nabla V_\varepsilon|^2-S_*b_\varepsilon^{2/q}=O_\delta(\varepsilon^{n-2})+o(\varepsilon^4). ∫ B δ + ∣∇ V ε ∣ 2 − S ∗ b ε 2/ q = O δ ( ε n − 2 ) + o ( ε 4 ) .
The cutoff annulus contributes O δ ( ε n − 2 ) O_\delta(\varepsilon^{n-2}) O δ ( ε n − 2 ) , and the volume density contributes O δ ( ε 5 ) O_\delta(\varepsilon^5) O δ ( ε 5 ) under the order-six normalization. The Taylor errors are O δ ( ε 5 log ( δ / ε ) ) O_\delta(\varepsilon^5\log(\delta/\varepsilon)) O δ ( ε 5 log ( δ / ε )) for n = 7 n=7 n = 7 and O δ ( ε 5 ) O_\delta(\varepsilon^5) O δ ( ε 5 ) for n ≥ 8 n\ge8 n ≥ 8 . Replacing truncated fourth moments by their full limits adds ε 4 O ( ( δ / ε ) 6 − n ) \varepsilon^4O((\delta/\varepsilon)^{6-n}) ε 4 O (( δ / ε ) 6 − n ) . These estimates also hold after D = ε ∂ ε \mathbf D=\varepsilon\partial_\varepsilon D = ε ∂ ε , so all errors are o ( ε 4 ) o(\varepsilon^4) o ( ε 4 ) for n ≥ 7 n\ge7 n ≥ 7 . Thus, with
I n : = ∫ 0 ∞ r n ( 1 + r 2 ) n d r , J n : = ∫ R + n ∣ y ′ ∣ 2 ( ∣ y ′ ∣ 2 + ( 1 + y n ) 2 ) n − 2 d y , I_n:=\int_0^\infty\frac{r^n}{(1+r^2)^n}\,dr,\qquad J_n:=\int_{\mathbb R_+^n}\frac{|y'|^2}{\left(|y'|^2+(1+y_n)^2\right)^{n-2}}\,dy, I n := ∫ 0 ∞ ( 1 + r 2 ) n r n d r , J n := ∫ R + n ( ∣ y ′ ∣ 2 + ( 1 + y n ) 2 ) n − 2 ∣ y ′ ∣ 2 d y ,
the curvature identities above give the following cancellation. If D n ′ = ( n − 3 ) ( n − 4 ) ( n − 5 ) ( n − 6 ) D'_n=(n-3)(n-4)(n-5)(n-6) D n ′ = ( n − 3 ) ( n − 4 ) ( n − 5 ) ( n − 6 ) , the three moments multiplying T a b ; a b T_{ab;ab} T ab ; ab , ∣ T ∣ 2 |T|^2 ∣ T ∣ 2 , and Scal ; ν ν \operatorname{Scal}_{;\nu\nu} Scal ; ν ν , with s = ∣ y ′ ∣ 2 + ( 1 + y n ) 2 s=|y'|^2+(1+y_n)^2 s = ∣ y ′ ∣ 2 + ( 1 + y n ) 2 , are
∫ R + n y n 2 ∣ y ′ ∣ 4 s n d y = 2 ( n + 1 ) ∣ S n − 2 ∣ I n D n ′ , \int_{\mathbb R_+^n}\frac{y_n^2|y'|^4}{s^n}\,dy=\frac{2(n+1)|S^{n-2}|I_n}{D'_n}, ∫ R + n s n y n 2 ∣ y ′ ∣ 4 d y = D n ′ 2 ( n + 1 ) ∣ S n − 2 ∣ I n ,
∫ R + n y n 4 ∣ y ′ ∣ 2 s n d y = 24 ∣ S n − 2 ∣ I n ( n − 2 ) D n ′ , ∫ R + n y n 2 s n − 2 d y = 8 ( n − 2 ) ∣ S n − 2 ∣ I n D n ′ . \int_{\mathbb R_+^n}\frac{y_n^4|y'|^2}{s^n}\,dy=\frac{24|S^{n-2}|I_n}{(n-2)D'_n},\qquad\int_{\mathbb R_+^n}\frac{y_n^2}{s^{n-2}}\,dy=\frac{8(n-2)|S^{n-2}|I_n}{D'_n}. ∫ R + n s n y n 4 ∣ y ′ ∣ 2 d y = ( n − 2 ) D n ′ 24∣ S n − 2 ∣ I n , ∫ R + n s n − 2 y n 2 d y = D n ′ 8 ( n − 2 ) ∣ S n − 2 ∣ I n .
The intrinsic scalar-curvature term is
n − 2 8 ( n − 1 ) 2 Scal ; a a J n = − n − 2 48 ( n − 1 ) 2 J n ∣ W ‾ ∣ 2 ; \frac{n-2}{8(n-1)^2}\operatorname{Scal}_{;aa}J_n=-\frac{n-2}{48(n-1)^2}J_n|\overline{W}|^2; 8 ( n − 1 ) 2 n − 2 Scal ; aa J n = − 48 ( n − 1 ) 2 n − 2 J n ∣ W ∣ 2 ;
see [9 (3-14)] . The coefficients of the three displayed moments are ( n − 2 ) 2 / ( ( n − 1 ) ( n + 1 ) ) (n-2)^2/((n-1)(n+1)) ( n − 2 ) 2 / (( n − 1 ) ( n + 1 )) , ( n − 2 ) 2 / ( 2 ( n − 1 ) ) (n-2)^2/(2(n-1)) ( n − 2 ) 2 / ( 2 ( n − 1 )) , and ( n − 2 ) / ( 8 ( n − 1 ) ) (n-2)/(8(n-1)) ( n − 2 ) / ( 8 ( n − 1 )) , respectively. Substituting T a b ; a b = − Scal ; ν ν / 2 − ∣ T ∣ 2 T_{ab;ab}=-\operatorname{Scal}_{;\nu\nu}/2-|T|^2 T ab ; ab = − Scal ; ν ν /2 − ∣ T ∣ 2 cancels Scal ; ν ν \operatorname{Scal}_{;\nu\nu} Scal ; ν ν and yields
N g ^ ∘ ( ψ ε , δ ) − S ∗ b ε 2 / q = ε 4 [ 2 ∣ S n − 2 ∣ I n ( 8 − n ) ( n − 2 ) ( n − 1 ) ( n − 3 ) ( n − 4 ) ( n − 5 ) ( n − 6 ) ∣ T ∣ 2 − n − 2 48 ( n − 1 ) 2 J n ∣ W ‾ ∣ 2 ] + E ε , δ , (141) \mathcal N_{\hat g}^{\circ}(\psi_{\varepsilon,\delta})-S_*b_\varepsilon^{2/q}
=\varepsilon^4\left[\frac{2|S^{n-2}|I_n(8-n)(n-2)}{(n-1)(n-3)(n-4)(n-5)(n-6)}|T|^2-\frac{n-2}{48(n-1)^2}J_n|\overline{W}|^2\right]+\mathcal E_{\varepsilon,\delta}, \tag*{(141)} N g ^ ∘ ( ψ ε , δ ) − S ∗ b ε 2/ q = ε 4 [ ( n − 1 ) ( n − 3 ) ( n − 4 ) ( n − 5 ) ( n − 6 ) 2∣ S n − 2 ∣ I n ( 8 − n ) ( n − 2 ) ∣ T ∣ 2 − 48 ( n − 1 ) 2 n − 2 J n ∣ W ∣ 2 ] + E ε , δ , (141)
where
∣ E ε , δ ∣ + ∣ D E ε , δ ∣ = o ( ε 4 ) . \lvert\mathcal{E}_{\varepsilon,\delta}\rvert+\lvert D\mathcal{E}_{\varepsilon,\delta}\rvert=o(\varepsilon^{4}). ∣ E ε , δ ∣ + ∣ D E ε , δ ∣ = o ( ε 4 ) .
The flat energy is 2 ( n − 2 ) ∣ S n − 2 ∣ I n 2(n-2)\lvert\mathbb{S}^{n-2}\rvert I_n 2 ( n − 2 ) ∣ S n − 2 ∣ I n , and the beta integrals give
J n ∣ S n − 2 ∣ I n = 4 ( n − 1 ) ( n − 2 ) ( n − 3 ) ( n − 5 ) ( n − 6 ) . \frac{J_n}{\lvert\mathbb{S}^{n-2}\rvert I_n}=\frac{4(n-1)(n-2)}{(n-3)(n-5)(n-6)}. ∣ S n − 2 ∣ I n J n = ( n − 3 ) ( n − 5 ) ( n − 6 ) 4 ( n − 1 ) ( n − 2 ) .
Division by the flat energy gives (134) . The fourth-order moments converge for n ≥ 7 n\geq7 n ≥ 7 . Homogeneity of the quotient and the tail estimates in Proposition 8.5 transfer the value and weighted scale derivative to the R β ( ε ) = ε − β R_\beta(\varepsilon)=\varepsilon^{-\beta} R β ( ε ) = ε − β family.
The second-order inverse-metric term is
( g a b ) 2 = 1 3 R c a d b y c y d + T a b y n 2 . (g^{ab})_2=\frac{1}{3}R^a_{\ c}{}^b_{\ d}y^c y^d+T^{ab}y_n^2. ( g ab ) 2 = 3 1 R c a d b y c y d + T ab y n 2 .
The intrinsic term vanishes after contraction with the tangential derivative of the radial bubble. The remaining second-order terms vanish in these coordinates. Hence
S T [ ϕ ] = 2 t ∗ ∫ R + n y n 2 T a b ∂ a U + ∂ b ϕ d y ( ϕ ∈ X + ) , t ∗ = S ∗ 1 / 2 . (142) \mathcal{S}_T[\phi]=2t_*\int_{\mathbb{R}^n_+}y_n^2T^{ab}\partial_aU_+\partial_b\phi\,dy\qquad(\phi\in X_+),\qquad t_*=S_*^{1/2}. \tag*{(142)} S T [ ϕ ] = 2 t ∗ ∫ R + n y n 2 T ab ∂ a U + ∂ b ϕ d y ( ϕ ∈ X + ) , t ∗ = S ∗ 1/2 . (142)
The factor t ∗ t_* t ∗ comes from the trace-normalized base U ∗ = t ∗ U + U_*=t_*U_+ U ∗ = t ∗ U + ; the model direction ϕ \phi ϕ is rescaled without this amplitude. The same integral defines S T \mathcal{S}_T S T on the whole Dirichlet space. Its continuity follows from ∥ y n 2 ∇ U + ∥ L 2 < ∞ \lVert y_n^2\nabla U_+\rVert_{L^2}<\infty ∥ y n 2 ∇ U + ∥ L 2 < ∞ , which holds for n ≥ 7 n\geq7 n ≥ 7 . The normal to the flat trace constraint is a multiple of U + U_+ U + in the Dirichlet pairing. Both U + U_+ U + and the dilation field Z 0 b Z_0^b Z 0 b are radial in y ′ y' y ′ . For either of these functions, the tangential angular integral contains
∫ S n − 2 T a b θ a θ b d θ = ∣ S n − 2 ∣ n − 1 tr T = 0. \int_{\mathbb{S}^{n-2}}T_{ab}\theta_a\theta_b\,d\theta=\frac{\lvert\mathbb{S}^{n-2}\rvert}{n-1}\operatorname{tr}T=0. ∫ S n − 2 T ab θ a θ b d θ = n − 1 ∣ S n − 2 ∣ tr T = 0.
For a translation field Z c b = − ∂ c U + Z_c^b=-\partial_cU_+ Z c b = − ∂ c U + the integrand is odd under y ′ ↦ − y ′ y'\mapsto-y' y ′ ↦ − y ′ . Thus
S T [ U + ] = S T [ Z 0 b ] = S T [ Z c b ] = 0 ( 1 ≤ c ≤ n − 1 ) . \mathcal{S}_T[U_+]=\mathcal{S}_T[Z_0^b]=\mathcal{S}_T[Z_c^b]=0\qquad(1\leq c\leq n-1). S T [ U + ] = S T [ Z 0 b ] = S T [ Z c b ] = 0 ( 1 ≤ c ≤ n − 1 ) .
These are the stated orthogonality conditions.
The equivariance follows from a change of variables. For R ∈ O ( n − 1 ) R\in O(n-1) R ∈ O ( n − 1 ) , put ( R R ϕ ) ( y ′ , y n ) = ϕ ( R − 1 y ′ , y n ) (\mathcal{R}_R\phi)(y',y_n)=\phi(R^{-1}y',y_n) ( R R ϕ ) ( y ′ , y n ) = ϕ ( R − 1 y ′ , y n ) . Then
S R T R t [ R R ϕ ] = S T [ ϕ ] . \mathcal{S}_{RTR^t}[\mathcal{R}_R\phi]=\mathcal{S}_T[\phi]. S R T R t [ R R ϕ ] = S T [ ϕ ] .
The Dirichlet form, the radial boundary potential in L 0 \mathcal{L}_0 L 0 , and the subspace X + X_+ X + are invariant under R R \mathcal{R}_R R R . Hence L 0 − 1 \mathcal{L}_0^{-1} L 0 − 1 commutes with this action, and the quadratic form
T ⟼ ⟨ L 0 − 1 S T , S T ⟩ T\longmapsto\langle\mathcal{L}_0^{-1}\mathcal{S}_T,\mathcal{S}_T\rangle T ⟼ ⟨ L 0 − 1 S T , S T ⟩
is a multiple of ∣ T ∣ 2 \lvert T\rvert^2 ∣ T ∣ 2 by the invariant quadratic-form argument in Proposition 5.4. The source is nonzero for T ≠ 0 T\neq0 T = 0 , as is seen by pairing it with a compactly supported function of the form Q T h ( ∣ y ′ ∣ , t ) Q_T h(\lvert y'\rvert,t) Q T h (∣ y ′ ∣ , t ) . Positivity of the flat form therefore makes κ n \kappa_n κ n positive. To evaluate it, it suffices by homogeneity to assume ∣ T ∣ = 1 \lvert T\rvert=1 ∣ T ∣ = 1 . Set a = 1 + t a=1+t a = 1 + t , s = ∣ y ′ ∣ 2 + a 2 s=\lvert y'\rvert^2+a^2 s = ∣ y ′ ∣ 2 + a 2 , and
f 2 ( v ) = n ( n − 2 ) A n ∫ R + n t 2 Q T s − ( n + 2 ) / 2 v d y . f_2(v)=n(n-2)A_n\int_{\mathbb{R}^n_+}t^2Q_Ts^{-(n+2)/2}v\,dy. f 2 ( v ) = n ( n − 2 ) A n ∫ R + n t 2 Q T s − ( n + 2 ) /2 v d y .
Tangential integration by parts in (8.31) gives S T = − 2 t ∗ f 2 \mathcal{S}_T=-2t_*f_2 S T = − 2 t ∗ f 2 . Since L 0 = 2 B n \mathcal{L}_0=2\mathcal{B}_n L 0 = 2 B n and t ∗ 2 = S ∗ t_*^2=S_* t ∗ 2 = S ∗ ,
κ n = f 2 ( B n − 1 f 2 ) . \kappa_n=f_2(\mathcal{B}_n^{-1}f_2). κ n = f 2 ( B n − 1 f 2 ) .
Let h 0 , b 0 , R n h_0,b_0,R_n h 0 , b 0 , R n be as in Lemma 5.14 , with Q = Q T Q=Q_T Q = Q T , and put
p 2 = A n Q T s − n / 2 [ n − 2 6 ( a 2 − 3 a + 3 ) + s 12 ] , K 2 = ( n − 3 ) A n 6 . p_2=A_nQ_Ts^{-n/2}\left[\frac{n-2}{6}(a^2-3a+3)+\frac{s}{12}\right],\qquad K_2=\frac{(n-3)A_n}{6}. p 2 = A n Q T s − n /2 [ 6 n − 2 ( a 2 − 3 a + 3 ) + 12 s ] , K 2 = 6 ( n − 3 ) A n .
Direct differentiation shows that − Δ p 2 -\Delta p_2 − Δ p 2 is the density defining f 2 f_2 f 2 and R p 2 = K 2 b 0 \mathcal{R}p_2=K_2b_0 R p 2 = K 2 b 0 . Hence
B n − 1 f 2 = p 2 − K 2 h 0 . \mathcal{B}_n^{-1}f_2=p_2-K_2h_0. B n − 1 f 2 = p 2 − K 2 h 0 .
The function p 2 p_2 p 2 has finite Dirichlet energy for n ≥ 7 n\geq7 n ≥ 7 . Define
I 0 = 1 n − 6 − 5 n − 5 + 10 n − 4 − 9 n − 3 + 3 n − 2 , I 1 = 1 n − 6 − 2 n − 5 + 1 n − 4 . I_0=\frac{1}{n-6}-\frac{5}{n-5}+\frac{10}{n-4}-\frac{9}{n-3}+\frac{3}{n-2},\qquad I_1=\frac{1}{n-6}-\frac{2}{n-5}+\frac{1}{n-4}. I 0 = n − 6 1 − n − 5 5 + n − 4 10 − n − 3 9 + n − 2 3 , I 1 = n − 6 1 − n − 5 2 + n − 4 1 .
The angular identity from Lemma 5.14 and radial beta integrals give
f 2 ( p 2 ) = n − 2 12 ( n − 1 ) I 0 + n 12 ( n − 3 ) ( n − 1 ) I 1 , f_2(p_2)=\frac{n-2}{12(n-1)}I_0+\frac{n}{12(n-3)(n-1)}I_1, f 2 ( p 2 ) = 12 ( n − 1 ) n − 2 I 0 + 12 ( n − 3 ) ( n − 1 ) n I 1 ,
∫ R n − 1 K 2 b 0 p 2 = 1 36 ( n − 1 ) + 1 36 ( n − 2 ) ( n − 5 ) , K 2 2 R n = n ( n − 3 ) ( n 2 − 2 n − 5 ) 36 ( n − 2 ) ( n − 1 ) 2 ( n + 1 ) ( n − 4 ) . \int_{\mathbb{R}^{n-1}}K_2b_0p_2=\frac{1}{36(n-1)}+\frac{1}{36(n-2)(n-5)},\qquad K_2^2R_n=\frac{n(n-3)(n^2-2n-5)}{36(n-2)(n-1)^2(n+1)(n-4)}. ∫ R n − 1 K 2 b 0 p 2 = 36 ( n − 1 ) 1 + 36 ( n − 2 ) ( n − 5 ) 1 , K 2 2 R n = 36 ( n − 2 ) ( n − 1 ) 2 ( n + 1 ) ( n − 4 ) n ( n − 3 ) ( n 2 − 2 n − 5 ) .
Substitution in (76) proves (135) . Subtracting κ n \kappa_n κ n from ( 8 − n ) / D n (8-n)/D_n ( 8 − n ) / D n gives − e n -e_n − e n and proves (136) . The numerator of e n e_n e n , written with x = n − 7 ≥ 0 x=n-7\geq0 x = n − 7 ≥ 0 , is 3 x 3 + 45 x 2 + 210 x + 308 3x^3+45x^2+210x+308 3 x 3 + 45 x 2 + 210 x + 308 , so e n > 0 e_n>0 e n > 0 .
Finally, the conformal-Fermi identities give
W ν a ν b = n − 3 n − 2 T a b , W ν a b c = 0 , ∣ W ^ ∣ 2 = ∣ W ‾ ∣ 2 + 4 ( n − 3 ) n − 2 ∣ T ∣ 2 . W_{\nu a\nu b}=\frac{n-3}{n-2}T_{ab},\qquad W_{\nu abc}=0,\qquad\lvert\widehat{W}\rvert^2=\lvert\overline{W}\rvert^2+\frac{4(n-3)}{n-2}\lvert T\rvert^2. W ν a ν b = n − 2 n − 3 T ab , W ν ab c = 0 , ∣ W ∣ 2 = ∣ W ∣ 2 + n − 2 4 ( n − 3 ) ∣ T ∣ 2 .
Thus the best algebraic constant is the smaller of b n b_n b n and e n ( n − 2 ) / ( 4 ( n − 3 ) ) e_n(n-2)/(4(n-3)) e n ( n − 2 ) / ( 4 ( n − 3 )) . The latter is strictly smaller because
b n − n − 2 4 ( n − 3 ) e n = n 2 − 3 n − 2 24 ( n − 5 ) ( n − 4 ) ( n − 1 ) 2 ( n + 1 ) > 0 ( n ≥ 7 ) . b_n-\frac{n-2}{4(n-3)}e_n=\frac{n^2-3n-2}{24(n-5)(n-4)(n-1)^2(n+1)}>0\qquad(n\geq7). b n − 4 ( n − 3 ) n − 2 e n = 24 ( n − 5 ) ( n − 4 ) ( n − 1 ) 2 ( n + 1 ) n 2 − 3 n − 2 > 0 ( n ≥ 7 ) .
This proves (139) and the equality statement. Since φ ( p ) = 1 \varphi(p)=1 φ ( p ) = 1 , the Weyl norm at p p p is unchanged. Lemma 8.6 identifies the two reduced coefficients, completing the proof. □
Corollary 8.8 (Intrinsic formula and conformal weight). Under the hypotheses of Theorem 8.7 , let E g = W g ( ν , ⋅ , ν , ⋅ ) E_g=W_g(\nu,\cdot,\nu,\cdot) E g = W g ( ν , ⋅ , ν , ⋅ ) on T p ∂ M T_p\partial M T p ∂ M , and let W ‾ g \overline{W}_g W g be the intrinsic Weyl tensor of the boundary metric. Then
Ξ g r e d ( p ) = − e n ( n − 2 ) 2 ( n − 3 ) 2 ∣ E g ( p ) ∣ g 2 − b n ∣ W ‾ g ( p ) ∣ g 2 . (143) \Xi_g^{\mathrm{red}}(p)=-e_n\frac{(n-2)^2}{(n-3)^2}\lvert E_g(p)\rvert_g^2-b_n\lvert\overline{W}_g(p)\rvert_g^2. \tag*{(143)} Ξ g red ( p ) = − e n ( n − 3 ) 2 ( n − 2 ) 2 ∣ E g ( p ) ∣ g 2 − b n ∣ W g ( p ) ∣ g 2 . (143) If g ~ = ψ 4 / ( n − 2 ) g \widetilde{g}=\psi^{4/(n-2)}g g = ψ 4/ ( n − 2 ) g is another representative with minimal boundary, then
Ξ g ~ r e d ( p ) = ψ ( p ) − 8 / ( n − 2 ) Ξ g r e d ( p ) . (144) \Xi_{\widetilde{g}}^{\mathrm{red}}(p)=\psi(p)^{-8/(n-2)}\Xi_g^{\mathrm{red}}(p). \tag*{(144)} Ξ g red ( p ) = ψ ( p ) − 8/ ( n − 2 ) Ξ g red ( p ) . (144) Proof. In the normalized conformal-Fermi representative, E = ( n − 3 ) T / ( n − 2 ) E=(n-3)T/(n-2) E = ( n − 3 ) T / ( n − 2 ) . The normalizing factor has value one at p p p , so conformal covariance identifies the Weyl tensors and their norms there with those of g g g . Substitution in (136) gives (143) . Under g ~ = ψ 4 / ( n − 2 ) g \widetilde{g}=\psi^{4/(n-2)}g g = ψ 4/ ( n − 2 ) g , both squared norms in this formula acquire the factor ψ ( p ) − 8 / ( n − 2 ) \psi(p)^{-8/(n-2)} ψ ( p ) − 8/ ( n − 2 ) . This proves (144) . □
Corollary 8.9 (Fourth-order subcriticality and compactness). Assume the hypotheses of Proposition 8.5 and work in a global representative satisfying H g = 0 H_g=0 H g = 0 .
(a) If p ∈ C p\in C p ∈ C is umbilic and Ξ g r e d ( p ) < 0 \Xi_g^{\mathrm{red}}(p)<0 Ξ g red ( p ) < 0 , then C E s c ∗ ( M , g ) < S ∗ C_{\mathrm{Esc}}^*(M,g)<S_* C Esc ∗ ( M , g ) < S ∗ .
(b) This conclusion holds in particular if the boundary is totally umbilic near p p p and W g ( p ) ≠ 0 W_g(p)\neq0 W g ( p ) = 0 .
(c) Suppose that the boundary is totally umbilic. If a normalized sequence of constrained critical points satisfies
J g [ u ℓ ] ⟶ S ∗ , J_g[u_\ell]\longrightarrow S_*, J g [ u ℓ ] ⟶ S ∗ , and is not precompact, then every limiting blow-up center p p p satisfies W g ( p ) = 0 W_g(p)=0 W g ( p ) = 0 . If inf ∂ M ∣ W g ∣ g > 0 \inf_{\partial M}\lvert W_g\rvert_g>0 inf ∂ M ∣ W g ∣ g > 0 , then, for every c 0 < c m i x c_0<c_{\mathrm{mix}} c 0 < c mix , the normalized constrained critical points with values at most c 0 c_0 c 0 form a compact subset of H 1 ( M ) H^1(M) H 1 ( M ) .
Proof. For (a), (126) gives J 1 , β ( p , ε ) < S ∗ J_{1,\beta}(p,\varepsilon)<S_* J 1 , β ( p , ε ) < S ∗ for small ε \varepsilon ε . Part (b) follows from Theorem 8.7 . For (c), Lemma 8.4 gives centers x ℓ → p x_\ell\to p x ℓ → p and scales ε ℓ ↓ 0 \varepsilon_\ell\downarrow0 ε ℓ ↓ 0 with
0 = S ∗ − 1 ∂ ε J 1 , β ( x ℓ , ε ℓ ) = 4 Ξ g r e d ( x ℓ ) ε ℓ 3 + o ( ε ℓ 3 ) . 0=S_*^{-1}\partial_\varepsilon J_{1,\beta}(x_\ell,\varepsilon_\ell)=4\Xi^{\mathrm{red}}_g(x_\ell)\varepsilon_\ell^3+o(\varepsilon_\ell^3). 0 = S ∗ − 1 ∂ ε J 1 , β ( x ℓ , ε ℓ ) = 4 Ξ g red ( x ℓ ) ε ℓ 3 + o ( ε ℓ 3 ) .
Total umbilicity makes the expansion uniform in x ℓ x_\ell x ℓ . Dividing by ε ℓ 3 \varepsilon_\ell^3 ε ℓ 3 gives Ξ g r e d ( p ) = 0 \Xi^{\mathrm{red}}_g(p)=0 Ξ g red ( p ) = 0 , hence W g ( p ) = 0 W_g(p)=0 W g ( p ) = 0 by (140) . This excludes blow-up at S ∗ S_* S ∗ if the Weyl norm is bounded away from zero. The compactness assertion below c m i x c_{\mathrm{mix}} c mix then follows from Theorems 4.3 and 4.7 , as in Corollary 7.5. □
Remark 8.10 (Boundary geometry and dimension). At a general umbilic point, tangential derivatives of I I ˚ \mathring{\mathrm{II}} II ˚ enter the second-order source. They vanish under local total umbilicity, where W g ( p ) = 0 W_g(p)=0 W g ( p ) = 0 also gives Ξ g r e d ( p ) = 0 \Xi^{\mathrm{red}}_g(p)=0 Ξ g red ( p ) = 0 . The radial integral ∫ 1 R r 5 − n d r \int_1^R r^{5-n}\,dr ∫ 1 R r 5 − n d r converges as R → ∞ R\to\infty R → ∞ for n ≥ 7 n\ge7 n ≥ 7 and grows logarithmically in dimension six.
Appendix A. Supplementary stability estimates
The coercivity and modulation results of § 3 also yield deficit estimates.
The optimizer orbit and flat trace stability
Retain F ∗ F_* F ∗ , M S \mathcal{M}_S M S , and the energy inner product from § 3 . On the hemisphere, S S S denotes the unit boundary L q L^q L q -norm constraint.
Lemma A.1 (Modulation near the optimizer manifold on S + n S^n_+ S + n ). There are 0 < δ ∗ < δ t u b e 0<\delta_*<\delta_{\mathrm{tube}} 0 < δ ∗ < δ tube and a C 1 C^1 C 1 map assigning to each u ∈ S u\in S u ∈ S with dist H 1 ( u , M S ) ≤ δ ∗ \operatorname{dist}_{H^1}(u,\mathcal{M}_S)\le\delta_* dist H 1 ( u , M S ) ≤ δ ∗ an optimizer v ♯ ∈ M S v^\sharp\in\mathcal{M}_S v ♯ ∈ M S satisfying u − v ♯ ⊥ T v ♯ M S u-v^\sharp\perp T_{v^\sharp}\mathcal{M}_S u − v ♯ ⊥ T v ♯ M S in the energy inner product. It is the unique optimizer satisfying this orthogonality and ∥ u − v ♯ ∥ H 1 < δ t u b e \lVert u-v^\sharp\rVert_{H^1}<\delta_{\mathrm{tube}} ∥ u − v ♯ ∥ H 1 < δ tube . Moreover, the parameter map is C 1 C^1 C 1 and ∥ u − v ♯ ∥ H 1 ≲ dist H 1 ( u , M S ) \lVert u-v^\sharp\rVert_{H^1}\lesssim\operatorname{dist}_{H^1}(u,\mathcal{M}_S) ∥ u − v ♯ ∥ H 1 ≲ dist H 1 ( u , M S ) .
Proof. Use the width convention
U λ , ξ ( y ′ , t ) = t ∗ λ − α U + ( ( y ′ − ξ ) / λ , t / λ ) , α = ( n − 2 ) / 2 , U_{\lambda,\xi}(y',t)=t_* \lambda^{-\alpha}U_+\left((y'-\xi)/\lambda,t/\lambda\right),\qquad\alpha=(n-2)/2, U λ , ξ ( y ′ , t ) = t ∗ λ − α U + ( ( y ′ − ξ ) / λ , t / λ ) , α = ( n − 2 ) /2 ,
and let F λ , ξ F_{\lambda,\xi} F λ , ξ be its Cayley pullback. Impose G j ( u , p ) = ⟨ u − F p , Z j ( p ) ⟩ E = 0 G_j(u,p)=\langle u-F_p,Z_j(p)\rangle_E=0 G j ( u , p ) = ⟨ u − F p , Z j ( p ) ⟩ E = 0 , where Z 0 = λ ∂ λ F p Z_0=\lambda\partial_\lambda F_p Z 0 = λ ∂ λ F p and Z a = λ ∂ ξ a F p Z_a=\lambda\partial_{\xi_a}F_p Z a = λ ∂ ξ a F p . At u = F p u=F_p u = F p , the normalized parameter Jacobian is the negative energy Gram matrix. For the Gram calculation, identify F p F_p F p and Z j Z_j Z j with their conformal images on the half-space. At ( λ , ξ ) = ( 1 , 0 ) (\lambda,\xi)=(1,0) ( λ , ξ ) = ( 1 , 0 ) the ratios Z j / U ∗ Z_j/U_* Z j / U ∗ are α \alpha α times the stereographic coordinate functions on S n − 1 S^{n-1} S n − 1 , up to signs. Since the normalized boundary measure is the uniform spherical measure and ∫ x i x j = δ i j / n \int x_i x_j=\delta_{ij}/n ∫ x i x j = δ ij / n , the Jacobi equation gives
⟨ Z i , Z j ⟩ E = ( q − 1 ) S ∗ ∫ ∂ R + n U ∗ q − 2 Z i Z j d y ′ = S ∗ ( n − 2 ) 4 δ i j . \langle Z_i,Z_j\rangle_E=(q-1)S_*\int_{\partial\mathbb{R}^n_+}U_*^{q-2}Z_iZ_j\,dy'=\frac{S_*(n-2)}{4}\delta_{ij}. ⟨ Z i , Z j ⟩ E = ( q − 1 ) S ∗ ∫ ∂ R + n U ∗ q − 2 Z i Z j d y ′ = 4 S ∗ ( n − 2 ) δ ij .
The normalized first two parameter derivatives have uniform energy bounds. The quantitative implicit-function theorem therefore gives a fixed local radius, a unique C 1 C^1 C 1 parameter choice, and the distance estimate. Conformal energy isometries make the constants uniform along the orbit; fixed energy–H 1 H^1 H 1 equivalence transfers the result to the stated norm.
For uniqueness throughout the stated neighborhood, send one candidate to F ∗ F_* F ∗ by a conformal map. When the relative translation and dilation parameters leave every compact set, the corresponding profiles converge weakly to zero in D 1 , 2 D^{1,2} D 1 , 2 , so their energy pairing with F ∗ F_* F ∗ tends to zero and their distance from it stays positive. On a compact relative parameter set, continuity and injectivity give the same separation outside the implicit-function neighborhood. Two sufficiently close candidates must therefore lie in that neighborhood and coincide. Applying the local estimate at approximate closest optimizers gives ∥ u − v ♯ ∥ H 1 ≤ C dist H 1 ( u , M S ) \lVert u-v^\sharp\rVert_{H^1}\le C\operatorname{dist}_{H^1}(u,\mathcal{M}_S) ∥ u − v ♯ ∥ H 1 ≤ C dist H 1 ( u , M S ) . Choose δ ∗ \delta_* δ ∗ small enough to place the result inside the neighborhood where uniqueness holds. □
Proposition A.2 (Model trace stability). Let n ≥ 3 n \ge3 n ≥ 3 . There are δ 0 , c ∗ > 0 \delta_0, c_* > 0 δ 0 , c ∗ > 0 , depending only on n n n , such that every positive u ∈ S u \in S u ∈ S on the round hemisphere with Def E s c ( u ) ≤ δ 0 \operatorname{Def}^{\mathrm{Esc}}(u) \le\delta_0 Def Esc ( u ) ≤ δ 0 admits the modulated optimizer v ♯ v^\sharp v ♯ of Lemma A.1 , and
u − v ♯ ⊥ T v ♯ M S , Def E s c ( u ) ≥ c ∗ ∥ u − v ♯ ∥ E , g r o u n d 2 ≍ ∥ u − v ♯ ∥ H 1 ( S + n ) 2 . u-v^\sharp\perp T_{v^\sharp}\mathcal{M}_S,\qquad\operatorname{Def}^{\mathrm{Esc}}(u) \ge c_*\lVert u-v^\sharp\rVert_{E,\mathrm{ground}}^2 \asymp\lVert u-v^\sharp\rVert_{H^1(S^n_+)}^2. u − v ♯ ⊥ T v ♯ M S , Def Esc ( u ) ≥ c ∗ ∥ u − v ♯ ∥ E , ground 2 ≍ ∥ u − v ♯ ∥ H 1 ( S + n ) 2 . In particular, arbitrarily small deficit forces arbitrarily small distance to M S \mathcal{M}_S M S .
Write Def f l a t E s c ( f ) = ∥ ∇ f ∥ 2 2 / ∥ f ∥ L q ( ∂ R + n ) 2 − S ∗ \operatorname{Def}^{\mathrm{Esc}}_{\mathrm{flat}}(f)=\lVert\nabla f\rVert_2^2/\lVert f\rVert_{L^q(\partial\mathbb{R}^n_+)}^2-S_* Def flat Esc ( f ) = ∥ ∇ f ∥ 2 2 / ∥ f ∥ L q ( ∂ R + n ) 2 − S ∗ . If f ∈ D 1 , 2 ( R + n ) f\in D^{1,2}(\mathbb{R}^n_+) f ∈ D 1 , 2 ( R + n ) is positive, has the same boundary norm as U + U_+ U + , and has sufficiently small deficit, there is c t r = c t r ( n ) > 0 c_{\mathrm{tr}}=c_{\mathrm{tr}}(n)>0 c tr = c tr ( n ) > 0 such that the corresponding optimizer V ♯ V^\sharp V ♯ with this boundary norm satisfies
Def f l a t E s c ( f ) ≥ c t r ∥ ∇ ( f − V ♯ ) ∥ L 2 ( R + n ) 2 , ∥ f − V ♯ ∥ H 1 ( B R + ) 2 ≤ C ( n ) ( 1 + R 2 ) Def f l a t E s c ( f ) ( R ≥ 1 ) . \operatorname{Def}^{\mathrm{Esc}}_{\mathrm{flat}}(f)\ge c_{\mathrm{tr}}\lVert\nabla(f-V^\sharp)\rVert_{L^2(\mathbb{R}^n_+)}^2,\qquad\lVert f-V^\sharp\rVert_{H^1(B_R^+)}^2\le C(n)(1+R^2)\operatorname{Def}^{\mathrm{Esc}}_{\mathrm{flat}}(f)\quad(R\ge1). Def flat Esc ( f ) ≥ c tr ∥ ∇ ( f − V ♯ ) ∥ L 2 ( R + n ) 2 , ∥ f − V ♯ ∥ H 1 ( B R + ) 2 ≤ C ( n ) ( 1 + R 2 ) Def flat Esc ( f ) ( R ≥ 1 ) . Proof. Ho’s trace stability theorem [29 Theorem 1.2] gives
inf V ∈ M f l a t ∥ ∇ ( f − V ) ∥ 2 2 ≤ C ( ∥ ∇ f ∥ 2 2 − S ∗ ∥ f ∥ q 2 ) = C ∥ f ∥ q 2 Def f l a t E s c ( f ) . \inf_{V\in\mathcal{M}_{\mathrm{flat}}}\lVert\nabla(f-V)\rVert_2^2\le C\bigl(\lVert\nabla f\rVert_2^2-S_*\lVert f\rVert_q^2\bigr)=C\lVert f\rVert_q^2\operatorname{Def}^{\mathrm{Esc}}_{\mathrm{flat}}(f). V ∈ M flat inf ∥ ∇ ( f − V ) ∥ 2 2 ≤ C ( ∥ ∇ f ∥ 2 2 − S ∗ ∥ f ∥ q 2 ) = C ∥ f ∥ q 2 Def flat Esc ( f ) . where M f l a t \mathcal{M}_{\mathrm{flat}} M flat is the full real-amplitude family of half-space optimizers. If V V V is an approximate closest optimizer, the trace inequality gives ∣ ∥ V ∥ q − ∥ f ∥ q ∣ ≤ C ∥ ∇ ( f − V ) ∥ 2 \lvert\lVert V\rVert_q-\lVert f\rVert_q\rvert\le C\lVert\nabla(f-V)\rVert_2 ∣∥ V ∥ q − ∥ f ∥ q ∣ ≤ C ∥ ∇ ( f − V ) ∥ 2 . Since ∥ ∇ V ∥ 2 2 = S ∗ ∥ V ∥ q 2 \lVert\nabla V\rVert_2^2=S_*\lVert V\rVert_q^2 ∥ ∇ V ∥ 2 2 = S ∗ ∥ V ∥ q 2 , rescaling V V V to the prescribed boundary norm changes the distance by at most a fixed multiple. For positive f f f and small deficit, V V V is positive. Lemma 3.1 transfers the estimate to the hemisphere, and Lemma A.1 gives the orthogonal optimizer with comparable distance. Finally,
∥ w ∥ L 2 ( B R + ) ≤ C R ∥ ∇ w ∥ L 2 ( R + n ) ( w ∈ D 1 , 2 ( R + n ) ) \lVert w\rVert_{L^2(B_R^+)}\le CR\lVert\nabla w\rVert_{L^2(\mathbb{R}^n_+)}\qquad(w\in D^{1,2}(\mathbb{R}^n_+)) ∥ w ∥ L 2 ( B R + ) ≤ C R ∥ ∇ w ∥ L 2 ( R + n ) ( w ∈ D 1 , 2 ( R + n )) by Sobolev and Hölder, proving the local H 1 H^1 H 1 estimate. □
A deficit estimate near a boundary bubble
We return to the fixed manifold ( M , g ) (M,g) ( M , g ) and its boundary constraint S S S .
Lemma A.3 (Taylor expansion under the boundary normalization). Let v ∈ S v\in S v ∈ S be nonnegative and satisfy ∥ v ∥ E , g ≤ C 0 \lVert v\rVert_{E,g}\le C_0 ∥ v ∥ E , g ≤ C 0 . For h ∈ T v S h\in T_vS h ∈ T v S with sufficiently small ∥ h ∥ H 1 \lVert h\rVert_{H^1} ∥ h ∥ H 1 , write R v ( h ) = ret v ( h ) \mathcal{R}_v(h)=\operatorname{ret}_v(h) R v ( h ) = ret v ( h ) . Uniformly on this set of base functions, and in particular for the normalized boundary bubbles v = U + , x , ε / ∥ U + , x , ε ∥ L q ( ∂ M ) v=U_{+,x,\varepsilon}/\lVert U_{+,x,\varepsilon}\rVert_{L^q(\partial M)} v = U + , x , ε / ∥ U + , x , ε ∥ L q ( ∂ M ) , 0 < ε ≤ ε 0 0<\varepsilon\le\varepsilon_0 0 < ε ≤ ε 0 , one has
J g [ R v ( h ) ] = J g [ v ] + L S [ v ] ( h ) + 1 2 Q S , v [ h ] + R v ( 3 ) ( h ) , J_g[\mathcal{R}_v(h)]=J_g[v]+L_S[v](h)+\frac{1}{2}Q_{S,v}[h]+\mathcal{R}_v^{(3)}(h), J g [ R v ( h )] = J g [ v ] + L S [ v ] ( h ) + 2 1 Q S , v [ h ] + R v ( 3 ) ( h ) , where L S [ v ] = D ( J g ∘ ret v ) ( 0 ) L_S[v]=D(J_g\circ\operatorname{ret}_v)(0) L S [ v ] = D ( J g ∘ ret v ) ( 0 ) and Q S , v = D 2 ( J g ∘ ret v ) ( 0 ) Q_{S,v}=D^2(J_g\circ\operatorname{ret}_v)(0) Q S , v = D 2 ( J g ∘ ret v ) ( 0 ) on T v S T_vS T v S , and
∣ R v ( 3 ) ( h ) ∣ ≤ C ∥ h ∥ E , g min { 3 , q } . (145) \lvert\mathcal{R}_v^{(3)}(h)\rvert\le C\lVert h\rVert_{E,g}^{\min\{3,q\}}. \tag*{(145)} ∣ R v ( 3 ) ( h )∣ ≤ C ∥ h ∥ E , g m i n { 3 , q } . (145) For an arbitrary u ∈ S u\in S u ∈ S sufficiently close to v v v , the inverse chart is
h = u ∫ ∂ M v q − 1 u d σ g − v ∈ T v S , ret v ( h ) = u . h=\frac{u}{\int_{\partial M}v^{q-1}u\,d\sigma_g}-v\in T_vS,\qquad\operatorname{ret}_v(h)=u. h = ∫ ∂ M v q − 1 u d σ g u − v ∈ T v S , ret v ( h ) = u . If w = u − v w=u-v w = u − v , then h − w = O H 1 ( ∥ w ∥ H 1 2 ) h-w=O_{H^1}(\lVert w\rVert_{H^1}^2) h − w = O H 1 (∥ w ∥ H 1 2 ) .
Proof. For 2 < q ≤ 3 2<q\le3 2 < q ≤ 3 , the trace Taylor estimate used in the proof of Lemma 5.8 gives a remainder O ( ∥ h ∥ E , g q ) O(\lVert h\rVert_{E,g}^q) O (∥ h ∥ E , g q ) . For q ≥ 3 q\ge3 q ≥ 3 , the pointwise remainder is bounded by C ( ∣ v ∣ + ∣ h ∣ ) q − 3 ∣ h ∣ 3 C(|v|+|h|)^{q-3}|h|^3 C ( ∣ v ∣ + ∣ h ∣ ) q − 3 ∣ h ∣ 3 ; Hölder’s inequality gives O ( ∥ h ∥ E , g 3 ) O(\lVert h\rVert_{E,g}^3) O (∥ h ∥ E , g 3 ) on the stated bounded set. Tangency removes the linear term in ∫ ∣ v + h ∣ q \int|v+h|^q ∫ ∣ v + h ∣ q . The quadratic energy numerator and the smooth reciprocal power of this boundary integral prove (A.1). The inverse formula is the one used in Lemma 6.1 ; applying the same boundary Taylor estimate to u = v + w ∈ S u=v+w\in S u = v + w ∈ S gives ∫ v q − 1 u = 1 + O ( ∥ w ∥ H 1 2 ) \int v^{q-1}u=1+O(\lVert w\rVert_{H^1}^2) ∫ v q − 1 u = 1 + O (∥ w ∥ H 1 2 ) and the claimed error. □
Theorem A.4 (Quantitative estimate near a boundary bubble). Assume n ≥ 5 n \ge5 n ≥ 5 , ( Y ∂ + ) (Y^+_{\partial}) ( Y ∂ + ) , and ( B G 2 + ) (\mathrm{BG}^{2+}) ( BG 2 + ) . There exist δ 0 > 0 \delta_0 > 0 δ 0 > 0 , ε 0 > 0 \varepsilon_0 > 0 ε 0 > 0 , C > 0 C > 0 C > 0 , and C g e o > 0 C_{\mathrm{geo}} > 0 C geo > 0 , depending only on ( M , g , n ) (M,g,n) ( M , g , n ) , with the following property. Suppose u ∈ S u \in\mathcal{S} u ∈ S and Def g E s c ( u ) ≤ δ 0 \operatorname{Def}^{\mathrm{Esc}}_g(u) \le\delta_0 Def g Esc ( u ) ≤ δ 0 . Assume also that u u u lies in the neighbourhood considered in Lemma 3.4 and that the corresponding boundary bubble v ♯ = a ♯ U + , x ♯ , ε ♯ v^\sharp= a^\sharp U_{+,x^\sharp,\varepsilon^\sharp} v ♯ = a ♯ U + , x ♯ , ε ♯ has 0 < ε ♯ < ε 0 0 < \varepsilon^\sharp< \varepsilon_0 0 < ε ♯ < ε 0 . Then
∥ u − v ♯ ∥ E , g 2 ≤ C Def g E s c ( u ) + C g e o ( ε ♯ ) 2 , (146) \lVert u-v^\sharp\rVert_{E,g}^{2}\le C\operatorname{Def}^{\mathrm{Esc}}_g(u)+C_{\mathrm{geo}}(\varepsilon^\sharp)^2, \tag*{(146)} ∥ u − v ♯ ∥ E , g 2 ≤ C Def g Esc ( u ) + C geo ( ε ♯ ) 2 , (146) and consequently
∥ u − v ♯ ∥ H 1 ( M ) 2 ≲ Def g E s c ( u ) + C g e o ( ε ♯ ) 2 . \lVert u-v^\sharp\rVert_{H^1(M)}^2\lesssim\operatorname{Def}^{\mathrm{Esc}}_g(u)+C_{\mathrm{geo}}(\varepsilon^\sharp)^2. ∥ u − v ♯ ∥ H 1 ( M ) 2 ≲ Def g Esc ( u ) + C geo ( ε ♯ ) 2 .
Proof. Since C E s c ∗ ≤ S ∗ C^*_{\mathrm{Esc}}\le S_* C Esc ∗ ≤ S ∗ , the small deficit bounds ∥ u ∥ E 2 ≤ S ∗ + δ 0 \lVert u\rVert_E^2\le S_*+\delta_0 ∥ u ∥ E 2 ≤ S ∗ + δ 0 . Write U = U + , x ♯ , ε ♯ U=U_{+,x^\sharp,\varepsilon^\sharp} U = U + , x ♯ , ε ♯ , e = ε ♯ e=\varepsilon^\sharp e = ε ♯ , r = u − v ♯ r=u-v^\sharp r = u − v ♯ , and λ = ∥ v ♯ ∥ q \lambda=\lVert v^\sharp\rVert_q λ = ∥ v ♯ ∥ q . By shrinking the modulation neighborhood, we may assume that r r r has small H 1 H^1 H 1 norm; ∣ λ − 1 ∣ ≤ ∥ r ∥ q |\lambda-1|\le\lVert r\rVert_q ∣ λ − 1∣ ≤ ∥ r ∥ q then bounds λ \lambda λ above and below. Put v = v ♯ / λ = t U ∈ S v=v^\sharp/\lambda=tU\in\mathcal{S} v = v ♯ / λ = t U ∈ S . Boundary orthogonality gives
h = u λ − v = r λ ∈ T v S , u = ret v ( h ) , ∥ h ∥ E ≍ ∥ r ∥ E , h=\frac{u}{\lambda}-v=\frac{r}{\lambda}\in T_v\mathcal{S},\qquad u=\operatorname{ret}_v(h),\qquad\lVert h\rVert_E\asymp\lVert r\rVert_E, h = λ u − v = λ r ∈ T v S , u = ret v ( h ) , ∥ h ∥ E ≍ ∥ r ∥ E ,
and preserves the energy orthogonality to ∂ e U \partial_eU ∂ e U and ∂ x a U \partial_{x_a}U ∂ x a U . Homogeneity and Theorem 3.6 give
D 2 ( J g ∘ ret v ) ( 0 ) [ h , h ] = t − 2 D 2 J g [ U ] [ h , h ] ≥ c ∥ h ∥ E 2 . D^2(\mathcal{J}_g\circ\operatorname{ret}_v)(0)[h,h]=t^{-2}D^2\mathcal{J}_g[U][h,h]\ge c\lVert h\rVert_E^2. D 2 ( J g ∘ ret v ) ( 0 ) [ h , h ] = t − 2 D 2 J g [ U ] [ h , h ] ≥ c ∥ h ∥ E 2 .
Lemma 3.7 bounds the linear term by C e ∥ h ∥ E Ce\lVert h\rVert_E C e ∥ h ∥ E . Lemma A.3 and Young’s inequality, with p = min { 3 , q } > 2 p=\min\{3,q\}>2 p = min { 3 , q } > 2 , therefore yield
J g [ u ] − J g [ v ] ≥ c ′ ∥ h ∥ E 2 − C e ∥ h ∥ E − C ∥ h ∥ E p ≥ c ′ 2 ∥ h ∥ E 2 − C ′ e 2 . \mathcal{J}_g[u]-\mathcal{J}_g[v]\ge c'\lVert h\rVert_E^2-Ce\lVert h\rVert_E-C\lVert h\rVert_E^p\ge\frac{c'}{2}\lVert h\rVert_E^2-C'e^2. J g [ u ] − J g [ v ] ≥ c ′ ∥ h ∥ E 2 − C e ∥ h ∥ E − C ∥ h ∥ E p ≥ 2 c ′ ∥ h ∥ E 2 − C ′ e 2 .
Now J g [ v ] ≥ C E s c ∗ ( M , g ) \mathcal{J}_g[v]\ge C^*_{\mathrm{Esc}}(M,g) J g [ v ] ≥ C Esc ∗ ( M , g ) proves (A.2); energy–H 1 H^1 H 1 equivalence proves the final estimate. □
Remark A.5 (Zero-trace comparison families). Let K 0 = ker ( Tr : H 1 ( M ) → L q ( ∂ M ) ) = H 0 1 ( M ) K_0=\ker(\operatorname{Tr}:H^1(M)\to L^q(\partial M))=H^1_0(M) K 0 = ker ( Tr : H 1 ( M ) → L q ( ∂ M )) = H 0 1 ( M ) , and let C i n t \mathcal{C}_{\mathrm{int}} C int be any nonempty subset of K 0 K_0 K 0 . Interior-supported cutoff profiles and their scalar multiples are examples. If ∥ u ∥ L q ( ∂ M ) = 1 \lVert u\rVert_{L^q(\partial M)}=1 ∥ u ∥ L q ( ∂ M ) = 1 , then for every W ∈ K 0 W\in K_0 W ∈ K 0 the trace inequality gives 1 = ∥ u − W ∥ L q ( ∂ M ) ≤ C t r ∥ u − W ∥ H 1 1=\lVert u-W\rVert_{L^q(\partial M)}\le C_{\mathrm{tr}}\lVert u-W\rVert_{H^1} 1 = ∥ u − W ∥ L q ( ∂ M ) ≤ C tr ∥ u − W ∥ H 1 . Hence
dist H 1 ( u , C i n t ) ≥ dist H 1 ( u , K 0 ) ≥ C t r − 1 . (147) \operatorname{dist}_{H^1}(u,\mathcal{C}_{\mathrm{int}})\ge\operatorname{dist}_{H^1}(u,K_0)\ge C_{\mathrm{tr}}^{-1}. \tag*{(147)} dist H 1 ( u , C int ) ≥ dist H 1 ( u , K 0 ) ≥ C tr − 1 . (147)
Interior harmonic profiles of Palais–Smale sequences vanish by Proposition 4.1.
Remark A.6 (Minimizing sequences and profile normalization). For a non-precompact minimizing sequence with Y = S ∗ Y=S_* Y = S ∗ , Ekeland’s principle and Lemma 4.6 give, after passing to a subsequence, one boundary bubble. Below S ∗ S_* S ∗ , minimizing sequences are precompact by Corollary 4.4. For the profile moments of our Dirichlet-normalized U + U_+ U + , including their relations, first-order cancellation, and the logarithmic term in dimension three, see [41 Lemma 3.9, Equation (3.16), Remark 3.10, and Lemma 4.2] .
Hemisphere comparison and coefficient consequences
Proposition B.1 (Hemisphere comparison and rigidity). Let ( M n , g ) (M^n,g) ( M n , g ) be a smooth compact connected Riemannian manifold with nonempty boundary, n ≥ 3 n\ge3 n ≥ 3 , satisfying ( Y ∂ + ) (Y^+_{\partial}) ( Y ∂ + ) . Then either ( M , g ) (M,g) ( M , g ) is conformally diffeomorphic to the round hemisphere or
C E s c ∗ ( M , g ) < S ∗ . C^*_{\mathrm{Esc}}(M,g)<S_*. C Esc ∗ ( M , g ) < S ∗ . Proof. The cases n = 3 n=3 n = 3 , and n = 4 , 5 n=4,5 n = 4 , 5 with umbilic boundary, follow from Escobar [18 ] . A non-umbilic point gives strict inequality by the Escobar–Marques comparison; see [37 pp. 383–384] and [9 pp. 1–2] . We may therefore assume n ≥ 6 n\ge6 n ≥ 6 and total umbilicity.
Put d = ⌊ ( n − 2 ) / 2 ⌋ d=\lfloor(n-2)/2\rfloor d = ⌊( n − 2 ) /2 ⌋ . If ∇ j W g ( p ) ≠ 0 \nabla^jW_g(p)\ne0 ∇ j W g ( p ) = 0 for some boundary point p p p and 0 ≤ j ≤ d − 2 0\le j\le d-2 0 ≤ j ≤ d − 2 , Chen’s local comparison gives the result [45 Theorem 2] . Otherwise fix p ∈ ∂ M p\in\partial M p ∈ ∂ M . Pass to the conformal-Fermi representative with totally geodesic boundary used in [45 Sections 2 and 4] . In these
coordinates g i n = δ i n g_{in}=\delta_{in} g in = δ in and ∂ n g a b = 0 \partial_n g_{ab}=0 ∂ n g ab = 0 on the boundary. The vanishing Weyl jets imply that the Taylor coefficients of log g \log g log g through degree d d d vanish, as in the proof of [45 Theorem 3] . Hence
g − δ = O 2 ( r d + 1 ) , Scal g = O ( r d − 1 ) , r = ∣ x ∣ , g-\delta=O_2(r^{d+1}),\qquad\operatorname{Scal}_g=O(r^{d-1}),\qquad r=|x|, g − δ = O 2 ( r d + 1 ) , Scal g = O ( r d − 1 ) , r = ∣ x ∣ ,
where T = O 2 ( r a ) T=O_2(r^a) T = O 2 ( r a ) means ∣ ∂ j T ∣ ≤ C r a − j |\partial^jT|\le Cr^{a-j} ∣ ∂ j T ∣ ≤ C r a − j for 0 ≤ j ≤ 2 0\le j\le2 0 ≤ j ≤ 2 .
Let G G G be the positive conformal Green function with singular coefficient one. It has homogeneous Neumann data away from p p p . Equation (12) of [45 ] gives F : = G − r 2 − n = O ( r d + 3 − n ) F:=G-r^{2-n}=O(r^{d+3-n}) F := G − r 2 − n = O ( r d + 3 − n ) . On a half-annulus of radius ρ \rho ρ ,
L g 0 F = − L g 0 r 2 − n = O ( r d + 1 − n ) , ∂ n F = 0 on the flat boundary . L_g^0F=-L_g^0r^{2-n}=O(r^{d+1-n}),\qquad\partial_nF=0\quad\text{on the flat boundary}. L g 0 F = − L g 0 r 2 − n = O ( r d + 1 − n ) , ∂ n F = 0 on the flat boundary .
After rescaling, F ρ ( X ) = ρ n − d − 3 F ( ρ X ) F_\rho(X)=\rho^{n-d-3}F(\rho X) F ρ ( X ) = ρ n − d − 3 F ( ρX ) has bounded supremum norm on fixed half-annuli. Smoothness and the vanishing metric jets give uniformly elliptic coefficients and a uniformly bounded C 0 , α C^{0,\alpha} C 0 , α right-hand side there, for any fixed 0 < α < 1 0<\alpha<1 0 < α < 1 . Local interior and Neumann boundary Schauder estimates on smaller half-annuli give ∣ ∂ j F ∣ ≤ C r d + 3 − n − j |\partial^jF|\le Cr^{d+3-n-j} ∣ ∂ j F ∣ ≤ C r d + 3 − n − j for j = 1 , 2 j=1,2 j = 1 , 2 . Thus G = r 2 − n ( 1 + O 2 ( r d + 1 ) ) G=r^{2-n}(1+O_2(r^{d+1})) G = r 2 − n ( 1 + O 2 ( r d + 1 )) . Inversion y = x / ∣ x ∣ 2 y=x/|x|^2 y = x /∣ x ∣ 2 now gives
g ^ : = G 4 / ( n − 2 ) g = δ + O 2 ( ∣ y ∣ − d − 1 ) , d + 1 > ( n − 2 ) / 2. \widehat g:=G^{4/(n-2)}g=\delta+O_2(|y|^{-d-1}),\qquad d+1>(n-2)/2. g := G 4/ ( n − 2 ) g = δ + O 2 ( ∣ y ∣ − d − 1 ) , d + 1 > ( n − 2 ) /2.
This metric is complete and scalar-flat, with minimal boundary and one asymptotically half-space end. Its mass is nonnegative, and zero mass forces the Euclidean half-space, by [24 Theorem 2.10] . See also [31 Theorem 1.1] and [12 Section 4] . In the nonorientable case, the connected orientation cover has two half-space ends of the same mass, so rigidity excludes zero mass. Thus either conformal compactification gives the round hemisphere or the mass is positive. In the latter case Chen’s Green-function test gives strict inequality [45 Theorem 3 and Proposition 10] .
Lemma B.2 (Equality at the hemisphere value). Suppose that the hypotheses of Proposition B.1 hold. Equality C E s c ∗ ( M , g ) = S ∗ C_{\mathrm{Esc}}^*(M,g)=S_* C Esc ∗ ( M , g ) = S ∗ forces conformal equivalence with the round hemisphere. A positive smooth extremal u u u with multiplier λ \lambda λ gives g ^ = u 4 / ( n − 2 ) g \widehat g=u^{4/(n-2)}g g = u 4/ ( n − 2 ) g with
Scal g ^ = 0 , H g ^ = 2 λ n − 2 , I I ˚ g ^ = u 2 / ( n − 2 ) I I ˚ g . \operatorname{Scal}_{\widehat g}=0,\qquad H_{\widehat g}=\frac{2\lambda}{n-2},\qquad\mathring{\mathrm{II}}_{\widehat g}=u^{2/(n-2)}\mathring{\mathrm{II}}_g. Scal g = 0 , H g = n − 2 2 λ , II ˚ g = u 2/ ( n − 2 ) II ˚ g .
Thus total umbilicity is preserved; λ = C E s c ∗ ( M , g ) \lambda=C_{\mathrm{Esc}}^*(M,g) λ = C Esc ∗ ( M , g ) if ∥ u ∥ q = 1 \|u\|_q=1 ∥ u ∥ q = 1 .
:::
Proof. Use Proposition B.1 and conformal covariance of the equations L g 0 u = 0 L_g^0u=0 L g 0 u = 0 , B g 0 u = λ u q − 1 B_g^0u=\lambda u^{q-1} B g 0 u = λ u q − 1 .
Theorem B.4 (Coefficient consequences and the Escobar threshold). Assume n ≥ 6 n\ge6 n ≥ 6 , ( Y ∂ + ) (Y_\partial^+) ( Y ∂ + ) , ( B G 4 + ) (\mathrm{BG}_4^+) ( BG 4 + ) , and H g = 0 H_g=0 H g = 0 .
A non-umbilic point gives C E s c ∗ ( M , g ) < S ∗ C_{\mathrm{Esc}}^*(M,g)<S_* C Esc ∗ ( M , g ) < S ∗ , attainment, and compactness of all normalized minimizing sequences and all normalized Palais–Smale sequences at the infimum. Either criterion R g b a r e ( x ) < 0 \mathcal{R}_g^{\mathrm{bare}}(x)<0 R g bare ( x ) < 0 or R g b a r e ( x ) = 0 ⇒ Θ g ( x ) \mathcal{R}_g^{\mathrm{bare}}(x)=0\Rightarrow\Theta_g(x) R g bare ( x ) = 0 ⇒ Θ g ( x ) implies these conclusions, as do the corresponding criteria with R g \mathcal{R}_g R g . For the standard bubble, a vanishing quadratic coefficient and a negative cubic coefficient can occur only when n = 6 n=6 n = 6 .
At every umbilic point, Lemma 8.1 gives the vanishing of the lower-order coefficients and the indicated derivative, and Proposition 2.5(ii) gives, with the stated cutoffs, the energy expansions with remainder o ( ε 3 ) o(\varepsilon^3) o ( ε 3 ) . At a non-umbilic point in dimension six, the quadratic coefficient of the standard bubble vanishes but the cubic coefficient need not vanish.
If C E s c ∗ ( M , g ) = S ∗ C_{\mathrm{Esc}}^*(M,g)=S_* C Esc ∗ ( M , g ) = S ∗ , the boundary is totally geodesic and the quantities in (ii) vanish everywhere. At every limiting concentration point p p p of a sequence of positive constrained critical points at S ∗ S_* S ∗ , one has I I ( p ) = 0 \mathrm{II}(p)=0 II ( p ) = 0 and R g b a r e ( p ) = R g r e d ( p ) = ∇ ∂ R g r e d ( p ) = Θ g ( p ) = 0 \mathcal{R}_g^{\mathrm{bare}}(p)=\mathcal{R}_g^{\mathrm{red}}(p)=\nabla_{\partial}\mathcal{R}_g^{\mathrm{red}}(p)=\Theta_g(p)=0 R g bare ( p ) = R g red ( p ) = ∇ ∂ R g red ( p ) = Θ g ( p ) = 0 . Non-umbilic concentration is excluded for all n ≥ 5 n\ge5 n ≥ 5 by Theorem 7.2.
::: (iv) If M M M is connected, C E s c ∗ ( M , g ) = S ∗ \mathcal{C}_{\mathrm{Esc}}^*(M,g)=S_* C Esc ∗ ( M , g ) = S ∗ implies conformal equivalence with the round hemisphere.
Proof. The quadratic formulas in Propositions 2.5 and 5.12 , and Θ g = 0 \Theta_g=0 Θ g = 0 at umbilic points, show that every criterion in (i) forces non-umbilicity; the case in which the standard bubble has vanishing quadratic coefficient at a non-umbilic point requires n = 6 n=6 n = 6 . Apply Theorem 5.17 and Corollary 4.4 . The results cited in (ii) use [41 Proposition 4.7, Lemma 4.11, Equations (4.25) and (B.21), and Corollary B.10] ; for the non-umbilic cubic terms see [41 Theorem B.9 and Remark B.11] . Equality forces I I ˚ ≡ 0 \mathring{\mathrm{II}}\equiv0 II ˚ ≡ 0 by (i), proving (iii) since H g = 0 H_g=0 H g = 0 ; (iv) is Proposition B.1 . □
Appendix C. Genericity at finite regularity
Throughout this appendix, the space of metrics and the notation are those of Theorem 7.6 .
Lemma C.1 (Finite-regularity conformal and coordinate estimates). Let g g g be a fixed C m C^m C m metric on the compact manifold, with m ≥ 3 m\ge3 m ≥ 3 . There is a positive C m C^m C m factor ϕ \phi ϕ with boundary value one for which H ϕ 4 / ( n − 2 ) g = 0 H_{\phi^{4/(n-2)}g}=0 H ϕ 4/ ( n − 2 ) g = 0 . At m = 3 m=3 m = 3 , the radial bubbles used above have two uniformly bounded normalized parameter derivatives in H 1 H^1 H 1 for n ≥ 5 n\ge5 n ≥ 5 . The first-order metric expansion and its radial derivative satisfy the estimates used in the quadratic reduction.
Proof. In a fixed smooth collar write ν = a ∂ t + V \nu=a\partial_t+V ν = a ∂ t + V , where V V V is tangential and a > 0 a>0 a > 0 . Prescribe
F ∣ ∂ M = 0 , ∂ t F = f : = n − 2 2 a H g ∈ C m − 1 ( ∂ M ) . F|_{\partial M}=0,\qquad\partial_tF=f:=\frac{n-2}{2a}H_g\in C^{m-1}(\partial M). F ∣ ∂ M = 0 , ∂ t F = f := 2 a n − 2 H g ∈ C m − 1 ( ∂ M ) .
In each boundary chart the jets F = 0 F=0 F = 0 , ∂ t α ∂ t F = ∂ α f \partial_t^\alpha\partial_tF=\partial^\alpha f ∂ t α ∂ t F = ∂ α f for ∣ α ∣ ≤ m − 1 |\alpha|\le m-1 ∣ α ∣ ≤ m − 1 , and all jets with at least two transverse derivatives equal to zero satisfy the Whitney compatibility conditions. The extension theorem [44 Theorem I, p. 65] , a boundary partition of unity, and a collar cutoff give a global C m C^m C m function F F F . With ϕ = e F \phi=e^F ϕ = e F , the prescribed normal derivative is ( n − 2 ) H g / 2 (n-2)H_g/2 ( n − 2 ) H g /2 , so conformal covariance gives minimal boundary and preserves coercivity.
For the coordinate estimates write the principal Fermi coefficient as A ( y ) = I + A 1 ( y ) + E ( y ) A(y)=I+A_1(y)+E(y) A ( y ) = I + A 1 ( y ) + E ( y ) , with A 1 A_1 A 1 the first jet determined by I I ( p ) \mathrm{II}(p) II ( p ) . The Jacobi and Riccati integral equations, using R m ∈ C 1 \mathrm{Rm}\in C^1 Rm ∈ C 1 and I I ∈ C 2 \mathrm{II}\in C^2 II ∈ C 2 , give
∣ E ∣ + ∣ y ⋅ ∇ E ∣ ≤ C ∣ y ∣ 2 , ∣ ∇ E ∣ ≤ C ∣ y ∣ . (148) |E|+|y\cdot\nabla E|\le C|y|^2,\qquad|\nabla E|\le C|y|. \tag*{(148)} ∣ E ∣ + ∣ y ⋅ ∇ E ∣ ≤ C ∣ y ∣ 2 , ∣∇ E ∣ ≤ C ∣ y ∣. (148)
These estimates also follow by smooth approximation in C 3 C^3 C 3 with uniform integral bounds. Scalar-curvature terms and their radial derivatives use only R m ∈ C 1 \mathrm{Rm}\in C^1 Rm ∈ C 1 .
For modulation, write the radial profile in a fixed collar using the normal distance t ( z ) t(z) t ( z ) , the boundary footpoint π ∂ ( z ) \pi_{\partial}(z) π ∂ ( z ) , and d g ˉ ( x , π ∂ ( z ) ) 2 d_{\bar g}(x,\pi_{\partial}(z))^2 d g ˉ ( x , π ∂ ( z ) ) 2 . The first-variation formula ∂ x d g ˉ ( x , z ′ ) 2 = − 2 exp x − 1 ( z ′ ) \partial_xd_{\bar g}(x,z')^2=-2\exp_{x}^{-1}(z') ∂ x d g ˉ ( x , z ′ ) 2 = − 2 exp x − 1 ( z ′ ) gives joint C 3 C^3 C 3 regularity of squared distance near the diagonal; the footpoint is C 2 C^2 C 2 . Two center derivatives and one spatial derivative therefore give the required rescaled H 1 H^1 H 1 bounds. For R = ε − 3 / 4 R=\varepsilon^{-3/4} R = ε − 3/4 the cutoff contribution to an ordinary first parameter derivative has H 1 H^1 H 1 norm O ( ε ( 3 n − 14 ) / 8 ) = o ( 1 ) O(\varepsilon^{(3n-14)/8})=o(1) O ( ε ( 3 n − 14 ) /8 ) = o ( 1 ) for n ≥ 5 n\ge5 n ≥ 5 ; normalized derivatives through order two remain bounded. These bounds give the modulation, source, and first scale-derivative estimates for each fixed C 3 C^3 C 3 metric. □
Proof of Theorem 7.6. Boundary transversality. For smooth metrics the transversality argument is due to Marques; see [10 p. 343] .
Put M = Met Y , B G m ( M ) \mathcal{M}=\operatorname{Met}^{m}_{Y,BG}(M) M = Met Y , B G m ( M ) . Consider the section ( g , x ) ↦ I I ˚ g ( x ) (g,x)\mapsto\mathring{\mathrm{II}}_g(x) ( g , x ) ↦ II ˚ g ( x ) over M × ∂ M \mathcal{M}\times\partial M M × ∂ M . At a zero ( g , x 0 ) (g,x_0) ( g , x 0 ) , choose a smooth boundary chart whose transverse coordinate vector equals ν g \nu_g ν g at x 0 x_0 x 0 , with g a n ( x 0 ) = 0 g_{an}(x_0)=0 g an ( x 0 ) = 0 and g n n ( x 0 ) = 1 g_{nn}(x_0)=1 g nn ( x 0 ) = 1 . For A ∈ Sym 0 2 ( T x 0 ∗ ∂ M , g ˉ ) A\in\operatorname{Sym}_0^2(T_{x_0}^*\partial M,\bar g) A ∈ Sym 0 2 ( T x 0 ∗ ∂ M , g ˉ ) , choose cutoffs with χ ( 0 ) = 1 \chi(0)=1 χ ( 0 ) = 1 , η ( 0 ) = 0 \eta(0)=0 η ( 0 ) = 0 , η ′ ( 0 ) = 1 \eta'(0)=1 η ′ ( 0 ) = 1 , and set
h a b : = − 2 χ ( x ′ ) η ( x n ) A a b , h a n = h n n : = 0. h_{ab}:=-2\chi(x')\eta(x_n)A_{ab},\qquad h_{an}=h_{nn}:=0. h ab := − 2 χ ( x ′ ) η ( x n ) A ab , h an = h nn := 0.
Since h ∣ ∂ M = 0 h|_{\partial M}=0 h ∣ ∂ M = 0 , the boundary metric and normal have zero variation there. At x 0 x_0 x 0 , the variation of the second fundamental form is therefore − 1 2 ∂ x n h a b = A a b -\frac{1}{2}\partial_{x_n}h_{ab}=A_{ab} − 2 1 ∂ x n h ab = A ab . Because A A A is trace-free, one has
D g I I ˚ g ( x 0 ) [ h ] = A . (149) D_g\mathring{\mathrm{II}}_g(x_0)[h]=A. \tag*{(149)} D g II ˚ g ( x 0 ) [ h ] = A . (149)
Put d = n − 1 d=n-1 d = n − 1 and N = n ( n − 1 ) / 2 − 1 N=n(n-1)/2-1 N = n ( n − 1 ) /2 − 1 . In local trace-free fiber trivializations, F ( g , x ) = I I ˚ g ( x ) F(g,x)=\mathring{\mathrm{II}}_g(x) F ( g , x ) = II ˚ g ( x ) is C m − 1 C^{m-1} C m − 1 , and the preceding variation makes D g F D_gF D g F onto at every zero. Thus Z = F − 1 ( 0 ) Z=F^{-1}(0) Z = F − 1 ( 0 ) is a C m − 1 C^{m-1} C m − 1 Banach submanifold, and the projection π : Z → M \pi:Z\to\mathcal{M} π : Z → M is Fredholm of index d − N < 0 d-N<0 d − N < 0 . The space of C m C^m C m metrics is separable, and these manifolds have countable bases. Since m − 1 ≥ 2 > max { d − N , 0 } m-1\geq2>\max\{d-N,0\} m − 1 ≥ 2 > max { d − N , 0 } , Smale’s theorem [42 Theorem 1.3] gives a residual set of regular values. Since the Fredholm index is negative, every regular fiber is empty. Therefore G 1 = { g ∈ M : I I ˚ g ≠ 0 on ∂ M } \mathcal{G}_1=\{g\in\mathcal{M}:\mathring{\mathrm{II}}_g\neq0\text{ on }\partial M\} G 1 = { g ∈ M : II ˚ g = 0 on ∂ M } is dense; compactness of ∂ M \partial M ∂ M makes it open.
An interior perturbation. Fix y 0 ∈ M ∘ y_0\in M^\circ y 0 ∈ M ∘ . In a smooth chart with g i j ( y 0 ) = δ i j g_{ij}(y_0)=\delta_{ij} g ij ( y 0 ) = δ ij , let W ≠ 0 \mathbf{W}\neq0 W = 0 be an algebraic Weyl tensor and let χ ∈ C c ∞ ( M ∘ ) \chi\in C^\infty_c(M^\circ) χ ∈ C c ∞ ( M ∘ ) be equal to one near y 0 y_0 y 0 . Write j y 0 1 h j^1_{y_0}h j y 0 1 h for the value and first derivatives of h h h at y 0 y_0 y 0 . The variation
h i j W ( x ) : = − 1 3 W i k j l x k x l χ ( x ) h^{\mathbf{W}}_{ij}(x):=-\frac{1}{3}\mathbf{W}_{ikjl}x^k x^l\chi(x) h ij W ( x ) := − 3 1 W ik j l x k x l χ ( x )
satisfies
j y 0 1 h W = 0 , D g W g ( y 0 ) [ h W ] = W . j^1_{y_0}h^{\mathbf{W}}=0,\qquad D_g\mathbf{W}_g(y_0)[h^{\mathbf{W}}]=\mathbf{W}. j y 0 1 h W = 0 , D g W g ( y 0 ) [ h W ] = W .
Only the second derivatives enter the curvature variation, because the value and first derivatives vanish. Hence G 2 = { g ∈ M : W g ( y 0 ) ≠ 0 } \mathcal{G}_2=\{g\in\mathcal{M}:\mathbf{W}_g(y_0)\neq0\} G 2 = { g ∈ M : W g ( y 0 ) = 0 } is open and dense. Taking G = G 1 ∩ G 2 \mathcal{G}=\mathcal{G}_1\cap\mathcal{G}_2 G = G 1 ∩ G 2 gives (109) . Since conformal flatness is preserved by conformal changes and diffeomorphisms, no metric in G \mathcal{G} G is conformally diffeomorphic to the round hemisphere.
Analytic consequences. Apply Lemma C.1 to pass to a representative with minimal boundary while preserving the stated regularity and coercivity. Its coordinate bounds justify the quadratic reduction, including modulation and the first scale derivative at m = 3 m=3 m = 3 . Theorem 5.17 , Corollary 4.4 , and Corollary 7.5 give strict inequality, attainment, minimizing-sequence compactness, and compactness of the sets of constrained critical points with the stated energy bounds. □
Appendix D. Supplementary one-bubble consequences
Recovery of the quadratic coefficient from two scales
Theorem D.1 (Two-scale coefficient recovery). Assume n ≥ 7 n\geq7 n ≥ 7 , ( Y ∂ + ) (Y^+_{\partial}) ( Y ∂ + ) , ( B G 4 + ) (\mathrm{BG4}^+) ( BG4 + ) , and H g = 0 H_g=0 H g = 0 on ∂ M \partial M ∂ M . For the standard normalized bubbles with the cutoff R ( ε ) = ε − 3 / 4 R(\varepsilon)=\varepsilon^{-3/4} R ( ε ) = ε − 3/4 put
Δ x ( ε ) = J g [ U x , ε ] − S ∗ S ∗ , D x ( ε ) = Δ x ( 2 ε ) − 2 Δ x ( ε ) 2 ε 2 . \Delta_x(\varepsilon)=\frac{\mathcal{J}_g[U_{x,\varepsilon}]-S_*}{S_*},\qquad D_x(\varepsilon)=\frac{\Delta_x(2\varepsilon)-2\Delta_x(\varepsilon)}{2\varepsilon^2}. Δ x ( ε ) = S ∗ J g [ U x , ε ] − S ∗ , D x ( ε ) = 2 ε 2 Δ x ( 2 ε ) − 2 Δ x ( ε ) . For all sufficiently small ε \varepsilon ε , uniformly in x ∈ ∂ M x\in\partial M x ∈ ∂ M ,
∣ D x ( ε ) − R g b a r e ( x ) ∣ ≤ C ε , ∣ Δ x ( ε ) ε − ε D x ( ε ) ∣ ≤ C ε 2 . \left|D_x(\varepsilon)-\mathfrak{R}^{\mathrm{bare}}_g(x)\right|\leq C\varepsilon,\qquad\left|\frac{\Delta_x(\varepsilon)}{\varepsilon}-\varepsilon D_x(\varepsilon)\right|\leq C\varepsilon^2. D x ( ε ) − R g bare ( x ) ≤ C ε , ε Δ x ( ε ) − ε D x ( ε ) ≤ C ε 2 . In particular, these estimates recover the quadratic coefficient of the standard bubble at the modulated center of a concentrating sequence by evaluating the standard bubbles at its scale and twice its scale.
Proof. Put ζ n = min { 2 , 3 ( n − 2 ) / 4 − 2 } ≥ 7 / 4 \zeta_n=\min\{2,3(n-2)/4-2\}\geq7/4 ζ n = min { 2 , 3 ( n − 2 ) /4 − 2 } ≥ 7/4 . The expansion in [41 Proposition 4.7 and Lemma 4.11] , with the cutoff tails of Proposition 2.5, gives
Δ x ( ε ) = R g b a r e ( x ) ε 2 + Θ g ( x ) ε 3 + O ( ε 2 + ζ n ) . \Delta_x(\varepsilon)=\mathfrak{R}^{\mathrm{bare}}_g(x)\varepsilon^2+\Theta_g(x)\varepsilon^3+O(\varepsilon^{2+\zeta_n}). Δ x ( ε ) = R g bare ( x ) ε 2 + Θ g ( x ) ε 3 + O ( ε 2 + ζ n ) . Substituting ε \varepsilon ε and 2 ε 2\varepsilon 2 ε yields
D x ( ε ) − R g b a r e ( x ) = 3 Θ g ( x ) ε + O ( ε ζ n ) , Δ x ( ε ) ε − ε D x ( ε ) = − 2 Θ g ( x ) ε 2 + O ( ε 1 + ζ n ) . D_x(\varepsilon)-\mathfrak{R}^{\mathrm{bare}}_g(x)=3\Theta_g(x)\varepsilon+O(\varepsilon^{\zeta_n}),\qquad\frac{\Delta_x(\varepsilon)}{\varepsilon}-\varepsilon D_x(\varepsilon)=-2\Theta_g(x)\varepsilon^2+O(\varepsilon^{1+\zeta_n}). D x ( ε ) − R g bare ( x ) = 3 Θ g ( x ) ε + O ( ε ζ n ) , ε Δ x ( ε ) − ε D x ( ε ) = − 2 Θ g ( x ) ε 2 + O ( ε 1 + ζ n ) . The coefficients are uniformly bounded and ζ n > 1 \zeta_n>1 ζ n > 1 . □
Strictly mean-convex Euclidean domains
Theorem D.2 (Compactness on strictly mean-convex Euclidean domains). Let Ω ⊂ R n \Omega\subset\mathbb{R}^{n} Ω ⊂ R n be a bounded smooth domain, n ≥ 4 n \ge4 n ≥ 4 , and let g = g f l a t g = g_{\mathrm{flat}} g = g flat . Assume that the boundary mean curvature with respect to the inner unit normal satisfies
H g ( x ) ≥ h 0 > 0 for all x ∈ ∂ Ω . H_{g}(x) \ge h_{0} > 0 \qquad\text{for all } x \in\partial\Omega. H g ( x ) ≥ h 0 > 0 for all x ∈ ∂ Ω. Assume moreover that ( Ω , g f l a t ) (\Omega,g_{\mathrm{flat}}) ( Ω , g flat ) is not conformally equivalent to the unit ball B n B^{n} B n (equivalently, not conformally diffeomorphic to ( S + n , g r o u n d ) (S^{n}_{+},g_{\mathrm{round}}) ( S + n , g round ) . If n ≥ 15 n \ge15 n ≥ 15 , assume in addition that the trace-free second fundamental form I I ˚ \mathring{\mathrm{II}} II ˚ of ∂ Ω \partial\Omega ∂ Ω vanishes nowhere on ∂ Ω \partial\Omega ∂ Ω .
Let u ℓ > 0 u_{\ell} > 0 u ℓ > 0 be normalized positive critical points of the Escobar quotient, i.e.
L g g u ℓ = 0 in Ω , B g g u ℓ = λ ℓ u ℓ q − 1 on ∂ Ω , ∫ ∂ Ω u ℓ q d σ = 1. L^{g}_{g}u_{\ell}=0 \text{ in }\Omega,\qquad B^{g}_{g}u_{\ell}=\lambda_{\ell}u_{\ell}^{q-1} \text{ on }\partial\Omega,\qquad\int_{\partial\Omega}u_{\ell}^{q}\,d\sigma=1. L g g u ℓ = 0 in Ω , B g g u ℓ = λ ℓ u ℓ q − 1 on ∂ Ω , ∫ ∂ Ω u ℓ q d σ = 1. Then ( u ℓ ) (u_{\ell}) ( u ℓ ) is precompact in C 2 ( Ω ‾ ) C^{2}(\overline{\Omega}) C 2 ( Ω ) , hence in H 1 ( Ω ) H^{1}(\Omega) H 1 ( Ω ) . In particular, no boundary bubbling occurs.
Proof. The trace–Poincaré inequality and H g ≥ h 0 > 0 H_{g} \ge h_{0} > 0 H g ≥ h 0 > 0 imply ( Y γ + ) (Y_{\gamma}^{+}) ( Y γ + ) . Hence μ : = C E s c ∗ ( Ω , g ) > 0 \mu:= C_{\mathrm{Esc}}^{*}(\Omega,g) > 0 μ := C Esc ∗ ( Ω , g ) > 0 . Testing the equation gives λ ℓ = J g [ u ℓ ] ≥ μ \lambda_{\ell} = J_{g}[u_{\ell}] \ge\mu λ ℓ = J g [ u ℓ ] ≥ μ . Put
c ℓ = ( λ ℓ / μ ) ( n − 2 ) / 2 , v ℓ = c ℓ u ℓ . c_{\ell}=(\lambda_{\ell}/\mu)^{(n-2)/2},\qquad v_{\ell}=c_{\ell}u_{\ell}. c ℓ = ( λ ℓ / μ ) ( n − 2 ) /2 , v ℓ = c ℓ u ℓ . Then L g g v ℓ = 0 L^{g}_{g}v_{\ell}=0 L g g v ℓ = 0 and B g g v ℓ = μ v ℓ q − 1 B^{g}_{g}v_{\ell}=\mu v_{\ell}^{q-1} B g g v ℓ = μ v ℓ q − 1 , where q = 2 ∂ ∗ q=2^{*}_{\partial} q = 2 ∂ ∗ . The fixed-coefficient compactness theorem [24 Theorem 1] applies for 4 ≤ n ≤ 14 4 \le n \le14 4 ≤ n ≤ 14 , and [10 Theorem 1.2] applies for n ≥ 15 n \ge15 n ≥ 15 under the nowhere-umbilic hypothesis. Thus a subsequence of ( v ℓ ) (v_{\ell}) ( v ℓ ) converges in C 2 ( Ω ‾ ) C^{2}(\overline{\Omega}) C 2 ( Ω ) to a positive solution. Since c ℓ = ∥ v ℓ ∥ L q ( ∂ Ω ) c_{\ell}=\|v_{\ell}\|_{L^{q}(\partial\Omega)} c ℓ = ∥ v ℓ ∥ L q ( ∂ Ω ) , these factors converge to a positive number, and u ℓ = v ℓ / c ℓ u_{\ell}=v_{\ell}/c_{\ell} u ℓ = v ℓ / c ℓ converges in C 2 ( Ω ‾ ) C^{2}(\overline{\Omega}) C 2 ( Ω ) . □
Acknowledgments
The authors thank IIT Bombay for providing ideal working conditions, and Saikat Mazumdar and Souptik Chakraborty for helpful discussions.
References
MM: Indian Institute of Technology Bombay, Powai, Maharashtra 400076, India
Email address: mathmukherjee@gmail.com, mukherjee@math.iitb.ac.in
US: Indian Institute of Technology Bombay, Powai, Maharashtra 400076, India
Email address: utsab@math.iitb.ac.in, reachutsab@gmail.com