Introduction
In this work, we consider the critical Choquard equation
A v = J μ ( ∣ v ∣ p ) ∣ v ∣ p − 2 v , v ∈ H 1 ( S n ) , (1) Av=J_{\mu}(|v|^{p})|v|^{p-2}v,\qquad v\in H^{1}(\mathbb{S}^{n}), \tag*{(1)} A v = J μ ( ∣ v ∣ p ) ∣ v ∣ p − 2 v , v ∈ H 1 ( S n ) , (1)
where
A = − Δ S n + n ( n − 2 ) 4 , J μ h ( ξ ) = ∫ S n h ( η ) ∣ ξ − η ∣ μ d V η , (2) A=-\Delta_{\mathbb{S}^{n}}+\frac{n(n-2)}{4},\qquad J_{\mu}h(\xi)=\int_{\mathbb{S}^{n}}\frac{h(\eta)}{|\xi-\eta|^{\mu}}\,dV_{\eta}, \tag*{(2)} A = − Δ S n + 4 n ( n − 2 ) , J μ h ( ξ ) = ∫ S n ∣ ξ − η ∣ μ h ( η ) d V η , (2)
and
n ≥ 3 , 0 < μ < n , p = 2 n − μ n − 2 . (3) n \ge3,\qquad0 < \mu< n,\qquad p = \frac{2n-\mu}{n-2}. \tag*{(3)} n ≥ 3 , 0 < μ < n , p = n − 2 2 n − μ . (3)
Here ∣ ξ − η ∣ |\xi-\eta| ∣ ξ − η ∣ is the chordal distance. We fix the standard product action
G = O ( k ) × O ( ℓ ) , k , ℓ ≥ 2 , k + ℓ = n + 1 (4) G = O(k) \times O(\ell),\qquad k,\ell\ge2,\qquad k+\ell=n+1 \tag*{(4)} G = O ( k ) × O ( ℓ ) , k , ℓ ≥ 2 , k + ℓ = n + 1 (4)
on S n ⊂ R k × R ℓ \mathbb{S}^{n} \subset\mathbb{R}^{k} \times\mathbb{R}^{\ell} S n ⊂ R k × R ℓ . The space of G G G -invariant functions in H 1 ( S n ) H^{1}(\mathbb{S}^{n}) H 1 ( S n ) is denoted by H H H . Under stereographic projection, (1) becomes
− Δ u ( x ) = ( ∫ R n ∣ u ( y ) ∣ p ∣ x − y ∣ μ d y ) ∣ u ( x ) ∣ p − 2 u ( x ) , u ∈ D 1 , 2 ( R n ) . (5) -\Delta u(x)=\left(\int_{\mathbb{R}^{n}}\frac{|u(y)|^{p}}{|x-y|^{\mu}}\,dy\right)|u(x)|^{p-2}u(x),\qquad u\in D^{1,2}(\mathbb{R}^{n}). \tag*{(5)} − Δ u ( x ) = ( ∫ R n ∣ x − y ∣ μ ∣ u ( y ) ∣ p d y ) ∣ u ( x ) ∣ p − 2 u ( x ) , u ∈ D 1 , 2 ( R n ) . (5)
Here D 1 , 2 ( R n ) D^{1,2}(\mathbb{R}^{n}) D 1 , 2 ( R n ) is the completion of C c ∞ ( R n ) C^{\infty}_{c}(\mathbb{R}^{n}) C c ∞ ( R n ) in the Dirichlet norm. The critical scaling is
u ( x ) ⟼ λ ( n − 2 ) / 2 u ( λ x ) , λ > 0. u(x)\longmapsto\lambda^{(n-2)/2}u(\lambda x),\qquad\lambda>0. u ( x ) ⟼ λ ( n − 2 ) /2 u ( λ x ) , λ > 0.
Both terms in the energy are invariant under this scaling, so (5) has the structure of a critical variational problem. For the variational background of the Choquard equation, see [26 , 29 ] . Positive solutions of (5) were classified in part of the parameter range in [14 , 22 ] , while the full range in (3) follows from [10 Corollary 1.5] .
A natural local counterpart is the critical equation
− Δ u = ∣ u ∣ 2 ∗ − 2 u in R n , -\Delta u=|u|^{2^{*}-2}u\quad\text{in }\mathbb{R}^{n}, − Δ u = ∣ u ∣ 2 ∗ − 2 u in R n ,
where 2 ∗ = 2 n / ( n − 2 ) 2^{*}=2n/(n-2) 2 ∗ = 2 n / ( n − 2 ) . Its conformal interpretation is described in [25 ] . The positive solutions were classified in [6 ] . The linearized analysis of the standard bubble [5 Appendix, Lemma A.1] shows that it has one negative direction and an ( n + 1 ) (n+1) ( n + 1 ) -dimensional kernel generated by conformal transformations. Ding [12 ] used product symmetries to construct an unbounded sequence of sign-changing solutions of the local critical equation. Almutairi and Gluck [1 Theorem 1.1] established the corresponding existence result for (5) for every 0 < μ < n 0<\mu<n 0 < μ < n . They used both the standard action and an action with a sign character. Under the standard action, their argument for sign change used the positive classification in dimensions 3 3 3 , 4 4 4 or in dimensions n ≥ 5 n\ge5 n ≥ 5 with μ ≤ 4 \mu\le4 μ ≤ 4 . Related variational results concern least-action nodal solutions of Choquard equations with a mass term [19 ] , including the quadratic case [18 ] . Morse-index estimates for minimax solutions based on critical groups were developed in [3 , 16 ] .
Building on these developments, the present paper studies the cohomological minimax levels in the fixed G G G -invariant space H H H . We prove that these levels are attained, establish their sharp growth behaviour, and obtain infinitely many pairs of sign-changing solutions. We further relate nodal domains to the invariant Morse index and minimax levels, and analyze the kernel and transverse negative directions of the full Hessian.
Equip H 1 ( S n ) H^{1}(\mathbb{S}^{n}) H 1 ( S n ) with the energy norm
∥ v ∥ 2 = ∫ S n ( ∣ ∇ S n v ∣ 2 + n ( n − 2 ) 4 v 2 ) d V . (6) \|v\|^{2}=\int_{\mathbb{S}^{n}}\left(|\nabla_{\mathbb{S}^{n}}v|^{2}+\frac{n(n-2)}{4}v^{2}\right)\,dV. \tag*{(6)} ∥ v ∥ 2 = ∫ S n ( ∣ ∇ S n v ∣ 2 + 4 n ( n − 2 ) v 2 ) d V . (6)
Its scalar product is denoted by ⟨ ⋅ , ⋅ ⟩ A \langle\cdot,\cdot\rangle_{A} ⟨ ⋅ , ⋅ ⟩ A , and by ⟨ ⋅ , ⋅ ⟩ H \langle\cdot,\cdot\rangle_{H} ⟨ ⋅ , ⋅ ⟩ H when restricted to H H H . Let ∥ ⋅ ∥ s \|\cdot\|_{s} ∥ ⋅ ∥ s denote the L s ( S n ) L^{s}(\mathbb{S}^{n}) L s ( S n ) norm. The exponent 2 ∗ 2^{*} 2 ∗ is as defined above. Define
D ( v ) = ∫ S n J μ ( ∣ v ∣ p ) ∣ v ∣ p d V , E ( v ) = 1 2 ∥ v ∥ 2 − 1 2 p D ( v ) . (7) D(v)=\int_{\mathbb{S}^{n}}J_{\mu}(|v|^{p})|v|^{p}\,dV,\qquad E(v)=\frac{1}{2}\|v\|^{2}-\frac{1}{2p}D(v). \tag*{(7)} D ( v ) = ∫ S n J μ ( ∣ v ∣ p ) ∣ v ∣ p d V , E ( v ) = 2 1 ∥ v ∥ 2 − 2 p 1 D ( v ) . (7)
For w ∈ S = { w ∈ H : ∥ w ∥ = 1 } w \in S = \{w \in H : \lVert w\rVert= 1\} w ∈ S = { w ∈ H : ∥ w ∥ = 1 } , let
Ψ ( w ) = max t > 0 E ( t w ) = p − 1 2 p D ( w ) − 1 / ( p − 1 ) . (8) \Psi(w) = \max_{t>0} E(tw) = \frac{p-1}{2p}\mathcal{D}(w)^{-1/(p-1)}. \tag*{(8)} Ψ ( w ) = t > 0 max E ( tw ) = 2 p p − 1 D ( w ) − 1/ ( p − 1 ) . (8)
The Z 2 \mathbb{Z}_2 Z 2 -cohomological index i i i due to Fadell and Rabinowitz [15 ] is employed. For a finite-dimensional subspace Y ⊂ H Y \subset H Y ⊂ H , let S ( Y ) = { y ∈ Y : ∥ y ∥ = 1 } S(Y) = \{y \in Y : \lVert y\rVert= 1\} S ( Y ) = { y ∈ Y : ∥ y ∥ = 1 } . The following properties of this index will be used.
If there is a continuous odd map between closed symmetric sets B → C B \to C B → C , then i ( B ) ≤ i ( C ) i(B) \le i(C) i ( B ) ≤ i ( C ) .
If B , C B, C B , C are closed symmetric sets, then i ( B ∪ C ) ≤ i ( B ) + i ( C ) i(B \cup C) \le i(B) + i(C) i ( B ∪ C ) ≤ i ( B ) + i ( C ) .
One has i ( S ( Y ) ) = dim Y i(S(Y)) = \dim Y i ( S ( Y )) = dim Y .
Every compact symmetric set has a closed symmetric neighborhood with the same index.
These properties are given in [15 Theorem 5.1] . For j ≥ 1 j \ge1 j ≥ 1 , define
Γ j = { B ⊂ S : B is compact , B = − B , i ( B ) ≥ j } . (9) \Gamma_j = \{B \subset S : B\text{ is compact},\ B=-B,\ i(B) \ge j\}. \tag*{(9)} Γ j = { B ⊂ S : B is compact , B = − B , i ( B ) ≥ j } . (9)
The associated minimax levels are then
c j = inf B ∈ Γ j sup w ∈ B Ψ ( w ) . (10) c_j = \inf_{B \in\Gamma_j}\sup_{w \in B}\Psi(w). \tag*{(10)} c j = B ∈ Γ j inf w ∈ B sup Ψ ( w ) . (10)
Throughout the paper, C > 0 C > 0 C > 0 stands for a constant whose value may vary from one estimate to another. Subscripts are used to specify its dependencies. Let
κ 0 = ∫ S n ∣ ξ − η ∣ − μ d V η , a 0 = ( n ( n − 2 ) 4 κ 0 ) 1 / ( 2 p − 2 ) . (11) \kappa_0 = \int_{\mathbb{S}^n} \lvert\xi-\eta\rvert^{-\mu}\,dV_\eta,\qquad
a_0 = \left(\frac{n(n-2)}{4\kappa_0}\right)^{1/(2p-2)}. \tag*{(11)} κ 0 = ∫ S n ∣ ξ − η ∣ − μ d V η , a 0 = ( 4 κ 0 n ( n − 2 ) ) 1/ ( 2 p − 2 ) . (11)
By rotational invariance, this integral is independent of ξ \xi ξ , and 0 < μ < n 0 < \mu< n 0 < μ < n implies 0 < κ 0 < ∞ 0 < \kappa_0 < \infty 0 < κ 0 < ∞ . The constants a 0 a_0 a 0 and − a 0 -a_0 − a 0 furnish constant-valued solutions to (1) .
Theorem 1.1. Assume (1.3) and (1.4). There are constants C 1 , C 2 > 0 C_1,C_2 > 0 C 1 , C 2 > 0 , independent of j j j , such that
C 1 j 2 p / ( p − 1 ) ≤ c j ≤ C 2 j 2 p / ( p − 1 ) , j ≥ 2. (12) C_1j^{2p/(p-1)} \le c_j \le C_2j^{2p/(p-1)},\qquad j \ge2. \tag*{(12)} C 1 j 2 p / ( p − 1 ) ≤ c j ≤ C 2 j 2 p / ( p − 1 ) , j ≥ 2. (12) Each c j c_j c j is attained by a solution in H H H . The first level is attained only by the two constant solutions ± a 0 \pm a_0 ± a 0 of (11) , and c 2 > c 1 c_2 > c_1 c 2 > c 1 . For every j ≥ 2 j \ge2 j ≥ 2 , each invariant solution of energy c j c_j c j changes sign.
For p ≥ 2 p \ge2 p ≥ 2 , let m G ( v ) m_G(v) m G ( v ) and ν G ( v ) \nu_G(v) ν G ( v ) be the index and nullity of E ′ ′ ( v ) E''(v) E ′′ ( v ) on H H H , and let m ( v ) m(v) m ( v ) and ν ( v ) \nu(v) ν ( v ) be the corresponding full-space quantities.
Remark 1.2. The growth and attainment assertions hold for the whole range 0 < μ < n 0 < \mu< n 0 < μ < n . The second-variation arguments below require p ≥ 2 p \ge2 p ≥ 2 . Indeed, for 1 < p < 2 1 < p < 2 1 < p < 2 , the second variation at a nodal solution involves the singular coefficient ∣ v ∣ p − 2 \lvert v\rvert^{p-2} ∣ v ∣ p − 2 , so the Hessian is not covered by the present framework. At p = 2 p = 2 p = 2 , this coefficient is identically one, including on the nodal set.
Theorem 1.3. Assume p ≥ 2 p \ge2 p ≥ 2 , and let v ∈ H ∖ { 0 } v \in H \setminus\{0\} v ∈ H ∖ { 0 } solve (1) . Then v ∈ C 2 , γ ( S n ) v \in C^{2,\gamma}(S^n) v ∈ C 2 , γ ( S n ) for some γ ∈ ( 0 , 1 ) \gamma\in(0,1) γ ∈ ( 0 , 1 ) . Its profile f ( t ) = v ( cos t ω , sin t θ ) f(t)=v(\cos t\,\omega,\sin t\,\theta) f ( t ) = v ( cos t ω , sin t θ ) has finitely many zeros, all simple and in ( 0 , π / 2 ) (0,\pi/2) ( 0 , π /2 ) . If these zeros are t 1 < ⋯ < t q t_1<\cdots<t_q t 1 < ⋯ < t q , then
{ v = 0 } = ⋃ i = 1 q S k − 1 ( cos t i ) × S ℓ − 1 ( sin t i ) . (13) \{v=0\}=\bigcup_{i=1}^{q} S^{k-1}(\cos t_i)\times S^{\ell-1}(\sin t_i). \tag*{(13)} { v = 0 } = i = 1 ⋃ q S k − 1 ( cos t i ) × S ℓ − 1 ( sin t i ) . (13) Let N G ( v ) N_G(v) N G ( v ) be the number of connected components of { v ≠ 0 } \{v\ne0\} { v = 0 } . Then
N G ( v ) = q + 1 ≤ m G ( v ) , c N G ( v ) ≤ E ( v ) . (14) N_G(v)=q+1\le m_G(v),\qquad c_{N_G(v)}\le E(v). \tag*{(14)} N G ( v ) = q + 1 ≤ m G ( v ) , c N G ( v ) ≤ E ( v ) . (14) For 1 ≤ j < N G ( v ) 1\le j<N_G(v) 1 ≤ j < N G ( v ) one has c j < E ( v ) c_j<E(v) c j < E ( v ) . In particular, E ( v ) = c j E(v)=c_j E ( v ) = c j implies N G ( v ) ≤ j N_G(v)\le j N G ( v ) ≤ j . The least energy of sign-changing solutions in H H H is c 2 c_2 c 2 . Every solution in H H H at c 2 c_2 c 2 has exactly two nodal domains.
The next result compares the minimax label with the index of a selected solution. Here isolation is taken in the entire critical set of Ψ \Psi Ψ .
Theorem 1.4. Assume p ≥ 2 p\ge2 p ≥ 2 . Suppose every critical point of Ψ \Psi Ψ at level c j c_j c j is isolated in the entire critical set of Ψ \Psi Ψ . Then there is a solution v j ∈ H v_j\in H v j ∈ H at c j c_j c j such that
m G ( v j ) ≤ j ≤ m G ( v j ) + ν G ( v j ) . (15) m_G(v_j)\le j\le m_G(v_j)+\nu_G(v_j). \tag*{(15)} m G ( v j ) ≤ j ≤ m G ( v j ) + ν G ( v j ) . (15) In particular, ν G ( v j ) = 0 \nu_G(v_j)=0 ν G ( v j ) = 0 implies m G ( v j ) = j m_G(v_j)=j m G ( v j ) = j .
Remark 1.5. The isolation hypothesis in Theorem 1.4 is imposed at the specified level c j c_j c j . It has not been proved along the whole minimax sequence. Under this hypothesis, Palais–Smale compactness implies that the critical set at c j c_j c j is finite and that c j c_j c j is isolated among critical values. These are consequences, rather than additional assumptions.
For the full-space Hessian, we use the angular degrees ( a , b ) (a,b) ( a , b ) of the two factors, and denote the corresponding radial index and nullity by m a , b ( v ) m_{a,b}(v) m a , b ( v ) and ν a , b ( v ) \nu_{a,b}(v) ν a , b ( v ) . These quantities are defined in § 9 . Under stereographic projection, the constant solutions a 0 a_0 a 0 and − a 0 -a_0 − a 0 correspond respectively to the standard positive bubble and its negative. Their linearized operators coincide. The nondegeneracy of the positive bubble was studied in [14 , 20 ] and established for the full parameter range in [28 ] .
Theorem 1.6. Assume p ≥ 2 p\ge2 p ≥ 2 , and let v ∈ H 1 ( S n ) ∖ { 0 } v\in H^1(S^n)\setminus\{0\} v ∈ H 1 ( S n ) ∖ { 0 } solve (1) . Set u = P v u=Pv u = P v , where P P P is the stereographic isometry in (18) . The Euclidean linearized kernel contains n + 1 n+1 n + 1 linearly independent elements
Z i = ∂ x i u ( 1 ≤ i ≤ n ) , Z 0 = n − 2 2 u + x ⋅ ∇ u . Z_i=\partial_{x_i}u\quad(1\le i\le n),\qquad Z_0=\frac{n-2}{2}u+x\cdot\nabla u. Z i = ∂ x i u ( 1 ≤ i ≤ n ) , Z 0 = 2 n − 2 u + x ⋅ ∇ u . Thus ν ( v ) ≥ n + 1 \nu(v)\ge n+1 ν ( v ) ≥ n + 1 . If v ∈ H v\in H v ∈ H is nonconstant, the tangent space to its conformal orbit is a subspace of the kernel of dimension n + 1 + k ℓ n+1+k\ell n + 1 + k ℓ . For such a nonconstant G G G -invariant solution,
m 1 , 0 ( v ) + m 0 , 1 ( v ) ≥ 1 , m ( v ) ≥ m G ( v ) + min { k , ℓ } . (16) m_{1,0}(v)+m_{0,1}(v)\ge1,\qquad m(v)\ge m_G(v)+\min\{k,\ell\}. \tag*{(16)} m 1 , 0 ( v ) + m 0 , 1 ( v ) ≥ 1 , m ( v ) ≥ m G ( v ) + min { k , ℓ } . (16) For the constant solutions, the conformal tangent space exhausts the kernel.
When k = ℓ k = \ell k = ℓ , let T ( ξ ′ , ξ ′ ′ ) = ( ξ ′ ′ , ξ ′ ) \mathcal{T}(\xi',\xi'') = (\xi'',\xi') T ( ξ ′ , ξ ′′ ) = ( ξ ′′ , ξ ′ ) and H − = { v ∈ H : v ( T ξ ) = − v ( ξ ) } H^- = \{v \in H : v(\mathcal{T}\xi) = -v(\xi)\} H − = { v ∈ H : v ( T ξ ) = − v ( ξ )} . As shown in the next theorem, a minimizer of Ψ \Psi Ψ on S ∩ H − S \cap H^- S ∩ H − gives rise to a solution via the rescaling (26) , with precise spectral and nodal properties described below.
Theorem 1.7. Suppose k = ℓ k = \ell k = ℓ . There exists 0 < μ ∗ = μ ∗ ( k ) < min { 4 , n } 0 < \mu_* = \mu_*(k) < \min\{4,n\} 0 < μ ∗ = μ ∗ ( k ) < min { 4 , n } such that for 0 < μ < μ ∗ 0 < \mu< \mu_* 0 < μ < μ ∗ , a minimizer on S ∩ H − S \cap H^- S ∩ H − yields a sign-changing solution v μ v_\mu v μ with two nodal domains. For every solution obtained from such a minimizer,
m 1 , 0 ( v μ ) = m 0 , 1 ( v μ ) = 1 , m 1 , 1 ( v μ ) = 0 , ν 1 , 0 ( v μ ) = ν 0 , 1 ( v μ ) = 1 , ν 1 , 1 ( v μ ) = 1. (17) \begin{aligned}
m_{1,0}(v_\mu) &= m_{0,1}(v_\mu) = 1,\quad m_{1,1}(v_\mu) = 0,\\
\nu_{1,0}(v_\mu) &= \nu_{0,1}(v_\mu) = 1,\quad\nu_{1,1}(v_\mu) = 1. \tag*{(17)}
\end{aligned} m 1 , 0 ( v μ ) ν 1 , 0 ( v μ ) = m 0 , 1 ( v μ ) = 1 , m 1 , 1 ( v μ ) = 0 , = ν 0 , 1 ( v μ ) = 1 , ν 1 , 1 ( v μ ) = 1. (17) In each of these three blocks the kernel is generated by the corresponding symmetry profile.
Remark 1.8. The minimization in Theorem 1.7 is performed in the proper subspace H − ⊂ H H^- \subset H H − ⊂ H , where n = 2 k − 1 n = 2k - 1 n = 2 k − 1 . Therefore, the energy of the resulting solutions is not identified with the minimax level c 2 c_2 c 2 in H H H . The nullities in (17) refer to the radial operators for fixed angular modes; after angular multiplicities are included, the corresponding full-space kernel dimensions are k k k , ℓ \ell ℓ , and k ℓ k\ell k ℓ . The theorem determines the index and nullity only in the three sectors ( 1 , 0 ) (1,0) ( 1 , 0 ) , ( 0 , 1 ) (0,1) ( 0 , 1 ) , and ( 1 , 1 ) (1,1) ( 1 , 1 ) , rather than the total full-space index and nullity.
§ 2 –§ 5 develop the conformal and invariant variational framework and prove the growth and attainment of the minimax levels. § 6 –§ 8 establish the nodal geometry, the comparison with the minimax levels, and the critical-group relation between the minimax label, index, and nullity. § 9 develops the angular decomposition of the full Hessian and proves the corresponding kernel and index results.
Conformal equivalence and invariant compactness
Let S = ( 0 , … , 0 , − 1 ) S = (0,\ldots,0,-1) S = ( 0 , … , 0 , − 1 ) and let π : S n ∖ { S } → R n \pi: S^n \setminus\{S\} \to\mathbb{R}^n π : S n ∖ { S } → R n be stereographic projection,
π ( ξ ) = ( ξ 1 , … , ξ n ) 1 + ξ n + 1 , π − 1 ( x ) = ( 2 x 1 + ∣ x ∣ 2 , 1 − ∣ x ∣ 2 1 + ∣ x ∣ 2 ) . \pi(\xi) = \frac{(\xi_1,\ldots,\xi_n)}{1+\xi_{n+1}},\qquad
\pi^{-1}(x) = \left(\frac{2x}{1+|x|^2},\frac{1-|x|^2}{1+|x|^2}\right). π ( ξ ) = 1 + ξ n + 1 ( ξ 1 , … , ξ n ) , π − 1 ( x ) = ( 1 + ∣ x ∣ 2 2 x , 1 + ∣ x ∣ 2 1 − ∣ x ∣ 2 ) .
Set Ω ( x ) = 2 / ( 1 + ∣ x ∣ 2 ) \Omega(x) = 2/(1+|x|^2) Ω ( x ) = 2/ ( 1 + ∣ x ∣ 2 ) . Define
P v ( x ) = Ω ( x ) ( n − 2 ) / 2 v ( π − 1 ( x ) ) . (18) Pv(x) = \Omega(x)^{(n-2)/2}v(\pi^{-1}(x)). \tag*{(18)} P v ( x ) = Ω ( x ) ( n − 2 ) /2 v ( π − 1 ( x )) . (18)
Proposition 2.1. The map P P P extends uniquely to an isometry from H 1 ( S n ) H^1(S^n) H 1 ( S n ) , equipped with the norm (1.6), onto D 1 , 2 ( R n ) D^{1,2}(\mathbb{R}^n) D 1 , 2 ( R n ) . For every v ∈ H 1 ( S n ) v \in H^1(S^n) v ∈ H 1 ( S n ) ,
∫ R n ∣ ∇ P v ∣ 2 d x = ∥ v ∥ 2 , ∬ R n × R n ∣ P v ( x ) ∣ p ∣ P v ( y ) ∣ p ∣ x − y ∣ μ d x d y = D ( v ) . (19) \int_{\mathbb{R}^n}|\nabla Pv|^2\,dx = \|v\|^2,\qquad
\iint_{\mathbb{R}^n\times\mathbb{R}^n}\frac{|Pv(x)|^p|Pv(y)|^p}{|x-y|^\mu}\,dx\,dy = \mathcal{D}(v). \tag*{(19)} ∫ R n ∣∇ P v ∣ 2 d x = ∥ v ∥ 2 , ∬ R n × R n ∣ x − y ∣ μ ∣ P v ( x ) ∣ p ∣ P v ( y ) ∣ p d x d y = D ( v ) . (19)
Consequently, u = P v u = Pv u = P v solves (1.5) if and only if v v v solves (1.1).
Proof. All assertions are precisely the conformal equivalence established in [1 Section 3] . □ \square □
The two singular orbits are S k − 1 × { 0 } S^{k-1} \times\{0\} S k − 1 × { 0 } and { 0 } × S ℓ − 1 \{0\} \times S^{\ell-1} { 0 } × S ℓ − 1 , of dimensions k − 1 k-1 k − 1 and ℓ − 1 \ell-1 ℓ − 1 , respectively. Every principal orbit is diffeomorphic to S k − 1 × S ℓ − 1 S^{k-1} \times S^{\ell-1} S k − 1 × S ℓ − 1 and has dimension k + ℓ − 2 = n − 1 k+\ell-2 = n-1 k + ℓ − 2 = n − 1 . Put
d = min { k − 1 , ℓ − 1 } , N = n − d = max { k , ℓ } . d = \min\{k-1,\ell-1\},\qquad N = n-d = \max\{k,\ell\}. d = min { k − 1 , ℓ − 1 } , N = n − d = max { k , ℓ } .
Then d ≥ 1 d \ge1 d ≥ 1 and 2 ≤ N < n 2 \le N < n 2 ≤ N < n .
Lemma 2.2. If N > 2 N>2 N > 2 , then
H ↪ L s ( S n ) continuously for 2 ≤ s ≤ 2 N N − 2 , H\hookrightarrow L^s(S^n)\quad\text{continuously for }2\le s\le\frac{2N}{N-2}, H ↪ L s ( S n ) continuously for 2 ≤ s ≤ N − 2 2 N , and the embedding is compact for 2 ≤ s < 2 N / ( N − 2 ) 2\le s<2N/(N-2) 2 ≤ s < 2 N / ( N − 2 ) . If N = 2 N=2 N = 2 , it is continuous and compact for every finite s ≥ 2 s\ge2 s ≥ 2 . In particular, H ↪ L 2 ∗ ( S n ) H\hookrightarrow L^{2^*}(S^n) H ↪ L 2 ∗ ( S n ) is compact, and the embedding remains compact for some s > 2 ∗ s>2^* s > 2 ∗ .
Proof. Apply the invariant Sobolev embedding in [23 Corollary 1] with minimum orbit dimension d d d . For N > 2 N>2 N > 2 , its critical exponent is 2 N / ( N − 2 ) 2N/(N-2) 2 N / ( N − 2 ) . Since N < n N<n N < n ,
2 ∗ = 2 n n − 2 < 2 N N − 2 . 2^*=\frac{2n}{n-2}<\frac{2N}{N-2}. 2 ∗ = n − 2 2 n < N − 2 2 N .
We may choose s s s strictly between these two exponents.
For N = 2 N=2 N = 2 , fix any finite s ≥ 2 s\ge2 s ≥ 2 . Choose q ∈ ( 1 , 2 ) q\in(1,2) q ∈ ( 1 , 2 ) so close to 2 that s < 2 q / ( 2 − q ) s<2q/(2-q) s < 2 q / ( 2 − q ) . The sphere has finite volume, and hence
∥ v ∥ W 1 , q ( S n ) ≤ C q ∥ v ∥ H 1 ( S n ) . \|v\|_{W^{1,q}(S^n)}\le C_q\|v\|_{H^1(S^n)}. ∥ v ∥ W 1 , q ( S n ) ≤ C q ∥ v ∥ H 1 ( S n ) .
Since s < 2 q / ( 2 − q ) s<2q/(2-q) s < 2 q / ( 2 − q ) , the invariant W 1 , q W^{1,q} W 1 , q embedding into L s L^s L s is compact. Composing it with the preceding inclusion gives the compact embedding of H H H into L s L^s L s , for every finite s s s , including s > 2 ∗ s>2^* s > 2 ∗ .
Write
q μ = 2 n 2 n − μ , p q μ = 2 ∗ . q_\mu=\frac{2n}{2n-\mu},\qquad pq_\mu=2^*. q μ = 2 n − μ 2 n , p q μ = 2 ∗ .
The HLS inequality gives a bounded map J μ : L q μ ( S n ) → L 2 n / μ ( S n ) J_\mu:L^{q_\mu}(S^n)\to L^{2n/\mu}(S^n) J μ : L q μ ( S n ) → L 2 n / μ ( S n ) , and
D ( v ) ≤ C ∥ v ∥ 2 ∗ 2 p . (20) \mathcal{D}(v)\le C\|v\|_{2^*}^{2p}. \tag*{(20)} D ( v ) ≤ C ∥ v ∥ 2 ∗ 2 p . (20)
The map v ↦ ∣ v ∣ p v\mapsto|v|^p v ↦ ∣ v ∣ p is of class C 1 : L 2 ∗ → L q μ C^1:L^{2^*}\to L^{q_\mu} C 1 : L 2 ∗ → L q μ . Its derivative at v v v is ϕ ↦ p ∣ v ∣ p − 2 v ϕ \phi\mapsto p|v|^{p-2}v\phi ϕ ↦ p ∣ v ∣ p − 2 v ϕ . Indeed, Hölder’s inequality uses ( p − 1 ) / 2 ∗ + 1 / 2 ∗ = 1 / q μ (p-1)/2^*+1/2^*=1/q_\mu ( p − 1 ) / 2 ∗ + 1/ 2 ∗ = 1/ q μ , and the continuity of this derivative follows from the power estimates below. Thus D , E ∈ C 1 ( H 1 ( S n ) , R ) \mathcal{D},\mathcal{E}\in C^1(H^1(S^n),\mathbb{R}) D , E ∈ C 1 ( H 1 ( S n ) , R ) , with
D ′ ( v ) [ ϕ ] = 2 p ∫ S n J μ ( ∣ v ∣ p ) ∣ v ∣ p − 2 v ϕ d V . (21) \mathcal{D}'(v)[\phi]=2p\int_{S^n}J_\mu(|v|^p)|v|^{p-2}v\phi\,dV. \tag*{(21)} D ′ ( v ) [ ϕ ] = 2 p ∫ S n J μ ( ∣ v ∣ p ) ∣ v ∣ p − 2 v ϕ d V . (21)
The kernel is positive, so D ( v ) > 0 \mathcal{D}(v)>0 D ( v ) > 0 for every nonzero v v v . We write H ∗ H^* H ∗ for the continuous dual of H H H .
Lemma 2.3. If v m ⇀ v v_m\rightharpoonup v v m ⇀ v in H H H , then
D ′ ( v m ) ⟶ D ′ ( v ) in H ∗ . (22) \mathcal{D}'(v_m)\longrightarrow\mathcal{D}'(v)\quad\text{in }H^*. \tag*{(22)} D ′ ( v m ) ⟶ D ′ ( v ) in H ∗ . (22)
Proof. Lemma 2.2 yields v m → v v_m\to v v m → v in L 2 ∗ ( S n ) L^{2^*}(S^n) L 2 ∗ ( S n ) . Since p q μ = 2 ∗ pq_\mu=2^* p q μ = 2 ∗ , the estimate
∥ ∣ a ∣ p − ∣ b ∣ p ∥ q μ ≤ C ( ∥ a ∥ 2 ∗ p − 1 + ∥ b ∥ 2 ∗ p − 1 ) ∥ a − b ∥ 2 ∗ \bigl\||a|^p-|b|^p\bigr\|_{q_\mu}\le C\bigl(\|a\|_{2^*}^{p-1}+\|b\|_{2^*}^{p-1}\bigr)\|a-b\|_{2^*} ∣ a ∣ p − ∣ b ∣ p q μ ≤ C ( ∥ a ∥ 2 ∗ p − 1 + ∥ b ∥ 2 ∗ p − 1 ) ∥ a − b ∥ 2 ∗
implies ∣ v m ∣ p → ∣ v ∣ p |v_m|^p\to|v|^p ∣ v m ∣ p → ∣ v ∣ p in L q μ ( S n ) L^{q_\mu}(S^n) L q μ ( S n ) . By the HLS inequality, we have
∥ J μ ( ∣ v m ∣ p − ∣ v ∣ p ) ∥ 2 n / μ ≤ C ∥ ∣ v m ∣ p − ∣ v ∣ p ∥ q μ = o ( 1 ) . (23) \|J_\mu(|v_m|^p-|v|^p)\|_{2n/\mu}\le C\||v_m|^p-|v|^p\|_{q_\mu}=o(1). \tag*{(23)} ∥ J μ ( ∣ v m ∣ p − ∣ v ∣ p ) ∥ 2 n / μ ≤ C ∥∣ v m ∣ p − ∣ v ∣ p ∥ q μ = o ( 1 ) . (23)
Define g ( s ) = ∣ s ∣ p − 2 s g(s)=|s|^{p-2}s g ( s ) = ∣ s ∣ p − 2 s . For p ≥ 2 p\ge2 p ≥ 2 , we have
∣ g ( a ) − g ( b ) ∣ ≤ C ( ∣ a ∣ + ∣ b ∣ ) p − 2 ∣ a − b ∣ , |g(a)-g(b)|\le C(|a|+|b|)^{p-2}|a-b|, ∣ g ( a ) − g ( b ) ∣ ≤ C ( ∣ a ∣ + ∣ b ∣ ) p − 2 ∣ a − b ∣ ,
while for 1 < p < 2 1<p<2 1 < p < 2 , we have
∣ g ( a ) − g ( b ) ∣ ≤ C ∣ a − b ∣ p − 1 . |g(a)-g(b)|\le C|a-b|^{p-1}. ∣ g ( a ) − g ( b ) ∣ ≤ C ∣ a − b ∣ p − 1 .
Applying Hölder’s inequality for p ≥ 2 p\ge2 p ≥ 2 and direct integration for 1 < p < 2 1<p<2 1 < p < 2 , we obtain
g ( v m ) ⟶ g ( v ) in L 2 ∗ / ( p − 1 ) ( S n ) . (24) g(v_m)\longrightarrow g(v)\quad\text{in }L^{2^*/(p-1)}(S^n). \tag*{(24)} g ( v m ) ⟶ g ( v ) in L 2 ∗ / ( p − 1 ) ( S n ) . (24)
Moreover, the sequences { J μ ( ∣ v m ∣ p ) } m ≥ 1 \{J_\mu(|v_m|^p)\}_{m\ge1} { J μ ( ∣ v m ∣ p ) } m ≥ 1 and { g ( v m ) } m ≥ 1 \{g(v_m)\}_{m\ge1} { g ( v m ) } m ≥ 1 are bounded in L 2 n / μ ( S n ) L^{2n/\mu}(S^n) L 2 n / μ ( S n ) and L 2 ∗ / ( p − 1 ) ( S n ) L^{2^*/(p-1)}(S^n) L 2 ∗ / ( p − 1 ) ( S n ) . Since
μ 2 n + p − 1 2 ∗ + 1 2 ∗ = 1 , \frac{\mu}{2n}+\frac{p-1}{2^*}+\frac{1}{2^*}=1, 2 n μ + 2 ∗ p − 1 + 2 ∗ 1 = 1 ,
the product in (21) can be estimated directly by Hölder’s inequality. For ϕ ∈ H \phi\in H ϕ ∈ H , Hölder’s inequality gives
∣ ( D ′ ( v m ) − D ′ ( v ) ) [ ϕ ] ∣ ≤ 2 p ∥ J μ ( ∣ v m ∣ p − ∣ v ∣ p ) ∥ 2 n / μ ∥ g ( v m ) ∥ 2 ∗ / ( p − 1 ) ∥ ϕ ∥ 2 ∗ + 2 p ∥ J μ ( ∣ v ∣ p ) ∥ 2 n / μ ∥ g ( v m ) − g ( v ) ∥ 2 ∗ / ( p − 1 ) ∥ ϕ ∥ 2 ∗ . \left|(D'(v_m)-D'(v))[\phi]\right|\le2p\left\|J_\mu(|v_m|^p-|v|^p)\right\|_{2n/\mu}\|g(v_m)\|_{2^*/(p-1)}\|\phi\|_{2^*}
+2p\left\|J_\mu(|v|^p)\right\|_{2n/\mu}\|g(v_m)-g(v)\|_{2^*/(p-1)}\|\phi\|_{2^*}. ∣ ( D ′ ( v m ) − D ′ ( v )) [ ϕ ] ∣ ≤ 2 p ∥ J μ ( ∣ v m ∣ p − ∣ v ∣ p ) ∥ 2 n / μ ∥ g ( v m ) ∥ 2 ∗ / ( p − 1 ) ∥ ϕ ∥ 2 ∗ + 2 p ∥ J μ ( ∣ v ∣ p ) ∥ 2 n / μ ∥ g ( v m ) − g ( v ) ∥ 2 ∗ / ( p − 1 ) ∥ ϕ ∥ 2 ∗ .
Hence, by (23) , (24) , and ∥ ϕ ∥ 2 ∗ ≤ C ∥ ϕ ∥ \|\phi\|_{2^*}\le C\|\phi\| ∥ ϕ ∥ 2 ∗ ≤ C ∥ ϕ ∥ , we obtain
∣ ( D ′ ( v m ) − D ′ ( v ) ) [ ϕ ] ∣ ≤ o ( 1 ) ∥ ϕ ∥ . \left|(D'(v_m)-D'(v))[\phi]\right|\le o(1)\|\phi\|. ∣ ( D ′ ( v m ) − D ′ ( v )) [ ϕ ] ∣ ≤ o ( 1 ) ∥ ϕ ∥.
Taking the supremum over ∥ ϕ ∥ ≤ 1 \|\phi\|\le1 ∥ ϕ ∥ ≤ 1 proves (22) . □ \square □
Lemma 2.4. The functional Ψ : S → ( 0 , ∞ ) \Psi:S\to(0,\infty) Ψ : S → ( 0 , ∞ ) is even, bounded away from zero, and of class C 1 C^1 C 1 . It satisfies the Palais–Smale condition.
Proof. Since D D D is positive on S S S , even, and of class C 1 C^1 C 1 , formula (8) gives the evenness and C 1 C^1 C 1 regularity of Ψ \Psi Ψ . For w ∈ S w\in S w ∈ S , (20) and Sobolev embedding give D ( w ) ≤ C D(w)\le C D ( w ) ≤ C . Thus
inf S Ψ ≥ p − 1 2 p C − 1 / ( p − 1 ) > 0. \inf_S\Psi\ge\frac{p-1}{2p}C^{-1/(p-1)}>0. S inf Ψ ≥ 2 p p − 1 C − 1/ ( p − 1 ) > 0. Let w m ∈ S w_m\in S w m ∈ S be a Palais–Smale sequence. After taking a subsequence,
Ψ ( w m ) → c > 0 , ∥ d ( Ψ ∣ S ) ( w m ) ∥ → 0 , w m ⇀ w in H . \Psi(w_m)\to c>0,\qquad\|d(\Psi|_S)(w_m)\|\to0,\qquad w_m\rightharpoonup w\quad\text{in }H. Ψ ( w m ) → c > 0 , ∥ d ( Ψ ∣ S ) ( w m ) ∥ → 0 , w m ⇀ w in H . By (8) ,
D ( w m ) ⟶ D c : = ( p − 1 2 p c ) p − 1 > 0. D(w_m)\longrightarrow D_c:=\left(\frac{p-1}{2pc}\right)^{p-1}>0. D ( w m ) ⟶ D c := ( 2 p c p − 1 ) p − 1 > 0. For T w S = { ϕ ∈ H : ⟨ w , ϕ ⟩ H = 0 } T_wS=\{\phi\in H:\langle w,\phi\rangle_H=0\} T w S = { ϕ ∈ H : ⟨ w , ϕ ⟩ H = 0 } , differentiation gives
d ( Ψ ∣ S ) ( w ) [ ϕ ] = − 1 2 p D ( w ) − p / ( p − 1 ) D ′ ( w ) [ ϕ ] . d(\Psi|_S)(w)[\phi]=-\frac{1}{2p}D(w)^{-p/(p-1)}D'(w)[\phi]. d ( Ψ ∣ S ) ( w ) [ ϕ ] = − 2 p 1 D ( w ) − p / ( p − 1 ) D ′ ( w ) [ ϕ ] . It follows that
sup ϕ ∈ T w m S ∥ ϕ ∥ ≤ 1 ∣ D ′ ( w m ) [ ϕ ] ∣ ⟶ 0. \sup_{\substack{\phi\in T_{w_m}S\\\|\phi\|\le1}}|D'(w_m)[\phi]|\longrightarrow0. ϕ ∈ T w m S ∥ ϕ ∥ ≤ 1 sup ∣ D ′ ( w m ) [ ϕ ] ∣ ⟶ 0.
Thus, the tangential part of D ′ ( w m ) D'(w_m) D ′ ( w m ) tends to zero. Euler’s identity gives
D ′ ( w m ) [ w m ] = 2 p D ( w m ) . D'(w_m)[w_m] = 2pD(w_m). D ′ ( w m ) [ w m ] = 2 p D ( w m ) .
The decomposition of H H H into the tangential subspace T w m S T_{w_m}S T w m S and the radial subspace span { w m } \operatorname{span}\{w_m\} span { w m } therefore yields
D ′ ( w m ) − 2 p D ( w m ) ⟨ w m , ⋅ ⟩ H ⟶ 0 in H ∗ . (25) D'(w_m)-2pD(w_m)\langle w_m,\cdot\rangle_H \longrightarrow0 \quad\text{in } H^*. \tag*{(25)} D ′ ( w m ) − 2 p D ( w m ) ⟨ w m , ⋅ ⟩ H ⟶ 0 in H ∗ . (25)
Let R : H ∗ → H \mathcal{R}:H^* \to H R : H ∗ → H be the Riesz isomorphism. Applying it to (2.8) yields
w m = 1 2 p D ( w m ) R D ′ ( w m ) + o ( 1 ) in H . w_m=\frac{1}{2pD(w_m)}\mathcal{R}D'(w_m)+o(1)\quad\text{in }H. w m = 2 p D ( w m ) 1 R D ′ ( w m ) + o ( 1 ) in H .
Lemma 2.3 now implies
w m ⟶ 1 2 p D c R D ′ ( w ) strongly in H . w_m\longrightarrow\frac{1}{2pD_c}\mathcal{R}D'(w)\quad\text{strongly in }H. w m ⟶ 2 p D c 1 R D ′ ( w ) strongly in H .
This limit equals the weak limit w w w . Consequently ∥ w ∥ = lim m → ∞ ∥ w m ∥ = 1 \lVert w\rVert=\lim_{m\to\infty}\lVert w_m\rVert=1 ∥ w ∥ = lim m → ∞ ∥ w m ∥ = 1 , and w ∈ S w\in S w ∈ S . The Palais–Smale condition follows.
Lemma 2.5. A point w ∈ S w\in S w ∈ S is a critical point of Ψ \Psi Ψ if and only if
v = t ( w ) w , t ( w ) = D ( w ) − 1 / ( 2 p − 2 ) , (26) v=t(w)w,\qquad t(w)=D(w)^{-1/(2p-2)}, \tag*{(26)} v = t ( w ) w , t ( w ) = D ( w ) − 1/ ( 2 p − 2 ) , (26) is a nonzero critical point of E ∣ H E|_H E ∣ H . Every nonzero critical point of E ∣ H E|_H E ∣ H has this form, and E ( v ) = Ψ ( w ) E(v)=\Psi(w) E ( v ) = Ψ ( w ) .
Proof. For fixed w ∈ S w\in S w ∈ S , homogeneity gives
E ( t w ) = t 2 2 − t 2 p 2 p D ( w ) , ∂ t E ( t w ) = t − t 2 p − 1 D ( w ) . E(tw)=\frac{t^2}{2}-\frac{t^{2p}}{2p}D(w),\qquad\partial_tE(tw)=t-t^{2p-1}D(w). E ( tw ) = 2 t 2 − 2 p t 2 p D ( w ) , ∂ t E ( tw ) = t − t 2 p − 1 D ( w ) .
The function t ↦ E ( t w ) t\mapsto E(tw) t ↦ E ( tw ) has a unique positive critical point at t = t ( w ) t=t(w) t = t ( w ) , where it attains its global maximum. Suppose first that d Ψ ( w ) = 0 d\Psi(w)=0 d Ψ ( w ) = 0 . The derivative formula in the preceding proof shows that D ′ ( w ) D'(w) D ′ ( w ) vanishes on T w S T_wS T w S . Its radial component is 2 p D ( w ) 2pD(w) 2 p D ( w ) . Therefore
D ′ ( w ) [ ϕ ] = 2 p D ( w ) ⟨ w , ϕ ⟩ H , ϕ ∈ H . (27) D'(w)[\phi]=2pD(w)\langle w,\phi\rangle_H,\qquad\phi\in H. \tag*{(27)} D ′ ( w ) [ ϕ ] = 2 p D ( w ) ⟨ w , ϕ ⟩ H , ϕ ∈ H . (27)
Put t = t ( w ) t=t(w) t = t ( w ) . Then t 2 p − 2 D ( w ) = 1 t^{2p-2}D(w)=1 t 2 p − 2 D ( w ) = 1 , and
E ′ ( t w ) [ ϕ ] = t ⟨ w , ϕ ⟩ H − t 2 p − 1 2 p D ′ ( w ) [ ϕ ] = t ( 1 − t 2 p − 2 D ( w ) ) ⟨ w , ϕ ⟩ H = 0. \begin{aligned}
E'(tw)[\phi]&=t\langle w,\phi\rangle_H-\frac{t^{2p-1}}{2p}D'(w)[\phi]\\
&=t\left(1-t^{2p-2}D(w)\right)\langle w,\phi\rangle_H=0.
\end{aligned} E ′ ( tw ) [ ϕ ] = t ⟨ w , ϕ ⟩ H − 2 p t 2 p − 1 D ′ ( w ) [ ϕ ] = t ( 1 − t 2 p − 2 D ( w ) ) ⟨ w , ϕ ⟩ H = 0.
Thus t w tw tw is a nonzero critical point of E ∣ H E|_H E ∣ H . Testing the equation with t w tw tw gives
∥ t w ∥ 2 = D ( t w ) , E ( t w ) = p − 1 2 p ∥ t w ∥ 2 . (28) \lVert tw\rVert^2=D(tw),\qquad E(tw)=\frac{p-1}{2p}\lVert tw\rVert^2. \tag*{(28)} ∥ tw ∥ 2 = D ( tw ) , E ( tw ) = 2 p p − 1 ∥ tw ∥ 2 . (28)
In particular,
E ( t w ) = p − 1 2 p t 2 = p − 1 2 p D ( w ) − 1 / ( p − 1 ) = Ψ ( w ) . E(tw)=\frac{p-1}{2p}t^2=\frac{p-1}{2p}D(w)^{-1/(p-1)}=\Psi(w). E ( tw ) = 2 p p − 1 t 2 = 2 p p − 1 D ( w ) − 1/ ( p − 1 ) = Ψ ( w ) .
Conversely, let v ∈ H ∖ { 0 } v \in H \setminus\{0\} v ∈ H ∖ { 0 } satisfy ( E ∣ H ) ′ ( v ) = 0 (E|_H)'(v)=0 ( E ∣ H ) ′ ( v ) = 0 . Set t = ∥ v ∥ t=\|v\| t = ∥ v ∥ and w = v / t ∈ S w=v/t\in S w = v / t ∈ S . Testing the equation with v v v gives ∥ v ∥ 2 = D ( v ) \|v\|^2=D(v) ∥ v ∥ 2 = D ( v ) . Hence
t 2 = t 2 p D ( w ) , t 2 p − 2 D ( w ) = 1 , t^2=t^{2p}D(w),\qquad t^{2p-2}D(w)=1, t 2 = t 2 p D ( w ) , t 2 p − 2 D ( w ) = 1 ,
which proves t = t ( w ) t=t(w) t = t ( w ) . For ϕ ∈ T w S \phi\in T_wS ϕ ∈ T w S ,
0 = E ′ ( t w ) [ ϕ ] = t ⟨ w , ϕ ⟩ H − t 2 p − 1 2 p D ′ ( w ) [ ϕ ] = − t 2 p − 1 2 p D ′ ( w ) [ ϕ ] . 0=E'(tw)[\phi]=t\langle w,\phi\rangle_H-\frac{t^{2p-1}}{2p}D'(w)[\phi]=-\frac{t^{2p-1}}{2p}D'(w)[\phi]. 0 = E ′ ( tw ) [ ϕ ] = t ⟨ w , ϕ ⟩ H − 2 p t 2 p − 1 D ′ ( w ) [ ϕ ] = − 2 p t 2 p − 1 D ′ ( w ) [ ϕ ] .
Thus D ′ ( w ) D'(w) D ′ ( w ) vanishes on the tangent space, and the formula for d Ψ d\Psi d Ψ gives d Ψ ( w ) = 0 d\Psi(w)=0 d Ψ ( w ) = 0 . The energy equality follows from (2.11).
The orbit interval and the Jacobi spectrum
In this section, A A A denotes the self-adjoint realization of (1.2) in L 2 ( S n ) L^2(S^n) L 2 ( S n ) , with domain H 2 ( S n ) H^2(S^n) H 2 ( S n ) . Outside the two singular orbits, write
ξ = ( cos t ω , sin t θ ) , 0 < t < π 2 , ω ∈ S k − 1 , θ ∈ S ℓ − 1 . (29) \xi=(\cos t\,\omega,\sin t\,\theta),\qquad0<t<\frac{\pi}{2},\qquad\omega\in S^{k-1},\quad\theta\in S^{\ell-1}. \tag*{(29)} ξ = ( cos t ω , sin t θ ) , 0 < t < 2 π , ω ∈ S k − 1 , θ ∈ S ℓ − 1 . (29)
The metric and volume element are
d t 2 + cos 2 t g S k − 1 + sin 2 t g S ℓ − 1 , d V = ρ ( t ) d t d V ω d V θ , dt^2+\cos^2t\,g_{S^{k-1}}+\sin^2t\,g_{S^{\ell-1}},\qquad dV=\rho(t)\,dt\,dV_\omega\,dV_\theta, d t 2 + cos 2 t g S k − 1 + sin 2 t g S ℓ − 1 , d V = ρ ( t ) d t d V ω d V θ ,
where
ρ ( t ) = ( cos t ) k − 1 ( sin t ) ℓ − 1 . (30) \rho(t)=(\cos t)^{k-1}(\sin t)^{\ell-1}. \tag*{(30)} ρ ( t ) = ( cos t ) k − 1 ( sin t ) ℓ − 1 . (30)
A G G G -invariant function has a profile v ( ξ ) = f ( t ) v(\xi)=f(t) v ( ξ ) = f ( t ) . For smooth invariant functions,
∥ v ∥ 2 2 = ∣ S k − 1 ∣ ∣ S ℓ − 1 ∣ ∫ 0 π / 2 ∣ f ( t ) ∣ 2 ρ ( t ) d t , \|v\|_2^2=|S^{k-1}||S^{\ell-1}|\int_0^{\pi/2}|f(t)|^2\rho(t)\,dt, ∥ v ∥ 2 2 = ∣ S k − 1 ∣∣ S ℓ − 1 ∣ ∫ 0 π /2 ∣ f ( t ) ∣ 2 ρ ( t ) d t ,
∥ v ∥ 2 = ∣ S k − 1 ∣ ∣ S ℓ − 1 ∣ ∫ 0 π / 2 ( ∣ f ′ ( t ) ∣ 2 + n ( n − 2 ) 4 ∣ f ( t ) ∣ 2 ) ρ ( t ) d t . (31) \|v\|^2=|S^{k-1}||S^{\ell-1}|\int_0^{\pi/2}\left(|f'(t)|^2+\frac{n(n-2)}{4}|f(t)|^2\right)\rho(t)\,dt. \tag*{(31)} ∥ v ∥ 2 = ∣ S k − 1 ∣∣ S ℓ − 1 ∣ ∫ 0 π /2 ( ∣ f ′ ( t ) ∣ 2 + 4 n ( n − 2 ) ∣ f ( t ) ∣ 2 ) ρ ( t ) d t . (31)
For v ∈ H v\in H v ∈ H , take smooth approximants in H 1 H^1 H 1 and average them over G G G . Averaging is a contraction in the energy norm. The averaged sequence is smooth, invariant, and converges to v v v , so the identities extend to H H H . On every compact interval I ⊂ ( 0 , π / 2 ) I\subset(0,\pi/2) I ⊂ ( 0 , π /2 ) , the weight is bounded above and below by positive constants. The corresponding profiles converge in H 1 ( I ) H^1(I) H 1 ( I ) . In particular, the profile of v v v has a continuous representative in the open orbit interval.
Lemma 3.1. For a C 2 C^2 C 2 , G G G -invariant function v ( ξ ) = f ( t ) v(\xi)=f(t) v ( ξ ) = f ( t ) ,
Δ S n v = ρ ( t ) − 1 ( ρ ( t ) f ′ ( t ) ) ′ = f ′ ′ ( t ) + ( ( ℓ − 1 ) cot t − ( k − 1 ) tan t ) f ′ ( t ) . (32) \Delta_{S^n}v=\rho(t)^{-1}\bigl(\rho(t)f'(t)\bigr)'=f''(t)+\bigl((\ell-1)\cot t-(k-1)\tan t\bigr)f'(t). \tag*{(32)} Δ S n v = ρ ( t ) − 1 ( ρ ( t ) f ′ ( t ) ) ′ = f ′′ ( t ) + ( ( ℓ − 1 ) cot t − ( k − 1 ) tan t ) f ′ ( t ) . (32)
Proof. All angular derivatives vanish. The divergence formula and the volume element above give
Δ S n v = 1 ρ ( t ) d d t ( ρ ( t ) f ′ ( t ) ) = f ′ ′ ( t ) + ρ ′ ( t ) ρ ( t ) f ′ ( t ) . \Delta_{S^n}v=\frac{1}{\rho(t)}\frac{d}{dt}\bigl(\rho(t)f'(t)\bigr)=f''(t)+\frac{\rho'(t)}{\rho(t)}f'(t). Δ S n v = ρ ( t ) 1 d t d ( ρ ( t ) f ′ ( t ) ) = f ′′ ( t ) + ρ ( t ) ρ ′ ( t ) f ′ ( t ) .
Differentiating (3.2), we obtain
ρ ′ ( t ) = − ( k − 1 ) ( cos t ) k − 2 ( sin t ) ℓ + ( ℓ − 1 ) ( cos t ) k ( sin t ) ℓ − 2 . \rho'(t)=-(k-1)(\cos t)^{k-2}(\sin t)^\ell+(\ell-1)(\cos t)^k(\sin t)^{\ell-2}. ρ ′ ( t ) = − ( k − 1 ) ( cos t ) k − 2 ( sin t ) ℓ + ( ℓ − 1 ) ( cos t ) k ( sin t ) ℓ − 2 .
Division by ρ ( t ) > 0 \rho(t)>0 ρ ( t ) > 0 gives
ρ ′ ( t ) ρ ( t ) = − ( k − 1 ) sin t cos t + ( ℓ − 1 ) cos t sin t = − ( k − 1 ) tan t + ( ℓ − 1 ) cot t . \frac{\rho'(t)}{\rho(t)}=-(k-1)\frac{\sin t}{\cos t}+(\ell-1)\frac{\cos t}{\sin t}=-(k-1)\tan t+(\ell-1)\cot t. ρ ( t ) ρ ′ ( t ) = − ( k − 1 ) cos t sin t + ( ℓ − 1 ) sin t cos t = − ( k − 1 ) tan t + ( ℓ − 1 ) cot t .
Substitution proves (32) .
Set α = ( ℓ − 2 ) / 2 \alpha=(\ell-2)/2 α = ( ℓ − 2 ) /2 and β = ( k − 2 ) / 2 \beta=(k-2)/2 β = ( k − 2 ) /2 . Let P m ( α , β ) P_m^{(\alpha,\beta)} P m ( α , β ) be the Jacobi polynomial of degree m m m , with weight ( 1 − x ) α ( 1 + x ) β (1-x)^\alpha(1+x)^\beta ( 1 − x ) α ( 1 + x ) β [32 ] .
Proposition 3.2. For m ≥ 0 m\geq0 m ≥ 0 , define
e m ( t ) = γ m P m ( α , β ) ( cos 2 t ) , (33) e_m(t)=\gamma_m P_m^{(\alpha,\beta)}(\cos2t), \tag*{(33)} e m ( t ) = γ m P m ( α , β ) ( cos 2 t ) , (33) where γ m > 0 \gamma_m>0 γ m > 0 is chosen so that ∥ e m ∥ 2 = 1 \lVert e_m\rVert_2=1 ∥ e m ∥ 2 = 1 . Write L G 2 ( S n ) L_G^2(S^n) L G 2 ( S n ) for the closed subspace of G G G -invariant functions in L 2 ( S n ) L^2(S^n) L 2 ( S n ) . The sequence ( e m ) m ≥ 0 (e_m)_{m\geq0} ( e m ) m ≥ 0 is an orthonormal basis of L G 2 ( S n ) L_G^2(S^n) L G 2 ( S n ) and an orthogonal basis of H H H . The functions e m e_m e m satisfy the eigenvalue relations
− Δ S n e m = 2 m ( 2 m + n − 1 ) e m , A e m = Λ m e m , (34) -\Delta_{S^n}e_m=2m(2m+n-1)e_m,\qquad Ae_m=\Lambda_m e_m, \tag*{(34)} − Δ S n e m = 2 m ( 2 m + n − 1 ) e m , A e m = Λ m e m , (34) where
Λ m = 2 m ( 2 m + n − 1 ) + n ( n − 2 ) 4 . (35) \Lambda_m=2m(2m+n-1)+\frac{n(n-2)}{4}. \tag*{(35)} Λ m = 2 m ( 2 m + n − 1 ) + 4 n ( n − 2 ) . (35) In particular,
4 m 2 ≤ Λ m ≤ C n ( m + 1 ) 2 . (36) 4m^2\leq\Lambda_m\leq C_n(m+1)^2. \tag*{(36)} 4 m 2 ≤ Λ m ≤ C n ( m + 1 ) 2 . (36)
Proof. Set x = cos 2 t x=\cos2t x = cos 2 t . Then
sin 2 t = 1 − x 2 , cos 2 t = 1 + x 2 , ∣ d x ∣ = 4 sin t cos t d t . \sin^2t=\frac{1-x}{2},\qquad\cos^2t=\frac{1+x}{2},\qquad|dx|=4\sin t\cos t\,dt. sin 2 t = 2 1 − x , cos 2 t = 2 1 + x , ∣ d x ∣ = 4 sin t cos t d t .
It follows that
ρ ( t ) d t = 1 4 ( cos t ) k − 2 ( sin t ) ℓ − 2 ∣ d x ∣ = 2 − ( k + ℓ ) / 2 ( 1 − x ) ( ℓ − 2 ) / 2 ( 1 + x ) ( k − 2 ) / 2 ∣ d x ∣ = 2 − ( k + ℓ ) / 2 ( 1 − x ) α ( 1 + x ) β ∣ d x ∣ . \begin{aligned}
\rho(t)\,dt=\frac{1}{4}(\cos t)^{k-2}(\sin t)^{\ell-2}|dx| \\
&=2^{-(k+\ell)/2}(1-x)^{(\ell-2)/2}(1+x)^{(k-2)/2}|dx| \\
&=2^{-(k+\ell)/2}(1-x)^\alpha(1+x)^\beta|dx|.
\end{aligned} ρ ( t ) d t = 4 1 ( cos t ) k − 2 ( sin t ) ℓ − 2 ∣ d x ∣ = 2 − ( k + ℓ ) /2 ( 1 − x ) ( ℓ − 2 ) /2 ( 1 + x ) ( k − 2 ) /2 ∣ d x ∣ = 2 − ( k + ℓ ) /2 ( 1 − x ) α ( 1 + x ) β ∣ d x ∣.
Thus, for every integrable F F F ,
∫ 0 π / 2 F ( cos 2 t ) ρ ( t ) d t = 2 − ( k + ℓ ) / 2 ∫ − 1 1 F ( x ) ( 1 − x ) α ( 1 + x ) β d x . \int_0^{\pi/2}F(\cos2t)\rho(t)\,dt
=2^{-(k+\ell)/2}\int_{-1}^{1}F(x)(1-x)^\alpha(1+x)^\beta\,dx. ∫ 0 π /2 F ( cos 2 t ) ρ ( t ) d t = 2 − ( k + ℓ ) /2 ∫ − 1 1 F ( x ) ( 1 − x ) α ( 1 + x ) β d x .
The correspondence
v ( ξ ) = f ( t ) ⟷ F ( x ) = f ( 1 2 arccos x ) v(\xi)=f(t)\quad\longleftrightarrow\quad F(x)=f\left(\frac{1}{2}\arccos x\right) v ( ξ ) = f ( t ) ⟷ F ( x ) = f ( 2 1 arccos x )
defines a linear bijection between L G 2 ( S n ) L_G^2(S^n) L G 2 ( S n ) and L 2 ( ( − 1 , 1 ) , ( 1 − x ) α ( 1 + x ) β d x ) L^2((-1,1),(1-x)^\alpha(1+x)^\beta\,dx) L 2 (( − 1 , 1 ) , ( 1 − x ) α ( 1 + x ) β d x ) , and the corresponding norms satisfy
∥ v ∥ 2 2 = C k , ℓ ∫ − 1 1 ∣ F ( x ) ∣ 2 ( 1 − x ) α ( 1 + x ) β d x , \lVert v\rVert_2^2=C_{k,\ell}\int_{-1}^{1}|F(x)|^2(1-x)^\alpha(1+x)^\beta\,dx, ∥ v ∥ 2 2 = C k , ℓ ∫ − 1 1 ∣ F ( x ) ∣ 2 ( 1 − x ) α ( 1 + x ) β d x ,
where
C k , ℓ = 2 − ( k + ℓ ) / 2 ∣ S k − 1 ∣ ∣ S ℓ − 1 ∣ . C_{k,\ell}=2^{-(k+\ell)/2}|S^{k-1}||S^{\ell-1}|. C k , ℓ = 2 − ( k + ℓ ) /2 ∣ S k − 1 ∣∣ S ℓ − 1 ∣.
Since α , β > − 1 \alpha,\beta>-1 α , β > − 1 , Jacobi orthogonality gives
⟨ e m , e j ⟩ L 2 = C k , ℓ γ m γ j ∫ − 1 1 P m ( α , β ) ( x ) P j ( α , β ) ( x ) ( 1 − x ) α ( 1 + x ) β d x = δ m j . \langle e_m,e_j\rangle_{L^2}=C_{k,\ell}\gamma_m\gamma_j\int_{-1}^{1}P_m^{(\alpha,\beta)}(x)P_j^{(\alpha,\beta)}(x)(1-x)^\alpha(1+x)^\beta\,dx=\delta_{mj}. ⟨ e m , e j ⟩ L 2 = C k , ℓ γ m γ j ∫ − 1 1 P m ( α , β ) ( x ) P j ( α , β ) ( x ) ( 1 − x ) α ( 1 + x ) β d x = δ mj .
If v ∈ L G 2 ( S n ) v\in L_G^2(S^n) v ∈ L G 2 ( S n ) is orthogonal to every e m e_m e m , its transformed profile F F F is orthogonal to every Jacobi polynomial. Their completeness implies F = 0 F=0 F = 0 , and hence v = 0 v=0 v = 0 . Thus ( e m ) m ≥ 0 (e_m)_{m\geq0} ( e m ) m ≥ 0 is an orthonormal basis of L G 2 ( S n ) L_G^2(S^n) L G 2 ( S n ) .
Set y ( x ) = P m ( α , β ) ( x ) y(x)=P_m^{(\alpha,\beta)}(x) y ( x ) = P m ( α , β ) ( x ) . The Jacobi differential equation [32 Eq. (4.2.4)] gives
( 1 − x 2 ) y ′ ′ + ( β − α − ( α + β + 2 ) x ) y ′ + m ( m + α + β + 1 ) y = 0. (37) (1-x^2)y''+(\beta-\alpha-(\alpha+\beta+2)x)y'+m(m+\alpha+\beta+1)y=0. \tag*{(37)} ( 1 − x 2 ) y ′′ + ( β − α − ( α + β + 2 ) x ) y ′ + m ( m + α + β + 1 ) y = 0. (37)
The chain rule gives
d d t y ( cos 2 t ) = − 2 sin 2 t y ′ ( x ) , d 2 d t 2 y ( cos 2 t ) = 4 ( 1 − x 2 ) y ′ ′ ( x ) − 4 x y ′ ( x ) . \frac{d}{dt}y(\cos2t)=-2\sin2t\,y'(x),\qquad\frac{d^2}{dt^2}y(\cos2t)=4(1-x^2)y''(x)-4xy'(x). d t d y ( cos 2 t ) = − 2 sin 2 t y ′ ( x ) , d t 2 d 2 y ( cos 2 t ) = 4 ( 1 − x 2 ) y ′′ ( x ) − 4 x y ′ ( x ) .
Substitution into (32) gives
Δ S n [ y ( cos 2 t ) ] = 4 [ ( 1 − x 2 ) y ′ ′ + ( β − α − ( α + β + 2 ) x ) y ′ ] . \Delta_{S^n}[y(\cos2t)]=4\left[(1-x^2)y''+(\beta-\alpha-(\alpha+\beta+2)x)y'\right]. Δ S n [ y ( cos 2 t )] = 4 [ ( 1 − x 2 ) y ′′ + ( β − α − ( α + β + 2 ) x ) y ′ ] .
Using (37) , the preceding identity becomes
− Δ S n [ y ( cos 2 t ) ] = 4 m ( m + α + β + 1 ) y ( cos 2 t ) . -\Delta_{S^n}[y(\cos2t)]=4m(m+\alpha+\beta+1)y(\cos2t). − Δ S n [ y ( cos 2 t )] = 4 m ( m + α + β + 1 ) y ( cos 2 t ) .
Since α + β + 1 = ( n − 1 ) / 2 \alpha+\beta+1=(n-1)/2 α + β + 1 = ( n − 1 ) /2 and e m ( t ) = γ m y ( cos 2 t ) e_m(t)=\gamma_m y(\cos2t) e m ( t ) = γ m y ( cos 2 t ) , it follows that
− Δ S n e m = 4 m ( m + n − 1 2 ) e m = 2 m ( 2 m + n − 1 ) e m . -\Delta_{S^n}e_m=4m\left(m+\frac{n-1}{2}\right)e_m=2m(2m+n-1)e_m. − Δ S n e m = 4 m ( m + 2 n − 1 ) e m = 2 m ( 2 m + n − 1 ) e m .
Moreover, writing
ξ = ( ξ ′ , ξ ′ ′ ) = ( cos t ω , sin t θ ) ∈ R k × R ℓ , \xi=(\xi',\xi'')=(\cos t\,\omega,\sin t\,\theta)\in\mathbb{R}^k\times\mathbb{R}^{\ell}, ξ = ( ξ ′ , ξ ′′ ) = ( cos t ω , sin t θ ) ∈ R k × R ℓ ,
we have
cos 2 t = cos 2 t − sin 2 t = ∣ ξ ′ ∣ 2 − ∣ ξ ′ ′ ∣ 2 . \cos2t=\cos^2t-\sin^2t=|\xi'|^2-|\xi''|^2. cos 2 t = cos 2 t − sin 2 t = ∣ ξ ′ ∣ 2 − ∣ ξ ′′ ∣ 2 .
Thus e m e_m e m is the restriction to S n S^n S n of a polynomial in the ambient coordinates. It is therefore smooth across both singular orbits, and the eigenvalue identity holds on the whole sphere. Adding the constant term in A A A gives A e m = Λ m e m Ae_m=\Lambda_m e_m A e m = Λ m e m , with Λ m \Lambda_m Λ m as in (35) . This proves (34) .
Rotations commute with A A A , and averaging over G G G is the orthogonal projection onto L G 2 ( S n ) L_G^2(S^n) L G 2 ( S n ) . Hence L G 2 ( S n ) L_G^2(S^n) L G 2 ( S n ) reduces A A A . Let A G A_G A G be the corresponding restriction. Its form domain is H H H , and the spectral theorem gives
H = { v = ∑ m ≥ 0 v ^ m e m : ∑ m ≥ 0 Λ m ∣ v ^ m ∣ 2 < ∞ } , v ^ m = ⟨ v , e m ⟩ L 2 . H=\left\{v=\sum_{m\geq0}\widehat v_m e_m:\sum_{m\geq0}\Lambda_m|\widehat v_m|^2<\infty\right\},\qquad\widehat v_m=\langle v,e_m\rangle_{L^2}. H = { v = m ≥ 0 ∑ v m e m : m ≥ 0 ∑ Λ m ∣ v m ∣ 2 < ∞ } , v m = ⟨ v , e m ⟩ L 2 .
In particular,
∥ v ∥ 2 = ∑ m ≥ 0 Λ m ∣ v ^ m ∣ 2 , ∥ v ∥ 2 2 = ∑ m ≥ 0 ∣ v ^ m ∣ 2 . (38) \|v\|^2=\sum_{m\geq0}\Lambda_m|\widehat v_m|^2,\qquad\|v\|_2^2=\sum_{m\geq0}|\widehat v_m|^2. \tag*{(38)} ∥ v ∥ 2 = m ≥ 0 ∑ Λ m ∣ v m ∣ 2 , ∥ v ∥ 2 2 = m ≥ 0 ∑ ∣ v m ∣ 2 . (38)
The identity
⟨ e m , e j ⟩ H = ⟨ A 1 / 2 e m , A 1 / 2 e j ⟩ L 2 = Λ m δ m j \langle e_m,e_j\rangle_H=\langle A^{1/2}e_m,A^{1/2}e_j\rangle_{L^2}=\Lambda_m\delta_{mj} ⟨ e m , e j ⟩ H = ⟨ A 1/2 e m , A 1/2 e j ⟩ L 2 = Λ m δ mj
proves energy orthogonality. For v ∈ H v \in H v ∈ H , its partial sums converge in H H H , since
∥ v − ∑ m = 0 M v ^ m e m ∥ 2 = ∑ m > M Λ m ∣ v ^ m ∣ 2 ⟶ 0. \left\|v-\sum_{m=0}^{M}\widehat{v}_m e_m\right\|^2=\sum_{m>M}\Lambda_m|\widehat{v}_m|^2\longrightarrow0. v − m = 0 ∑ M v m e m 2 = m > M ∑ Λ m ∣ v m ∣ 2 ⟶ 0.
Finally,
Λ m = 4 m 2 + 2 ( n − 1 ) m + n ( n − 2 ) 4 , \Lambda_m=4m^2+2(n-1)m+\frac{n(n-2)}{4}, Λ m = 4 m 2 + 2 ( n − 1 ) m + 4 n ( n − 2 ) ,
which gives (36) .
Sharp growth of the minimax levels
Throughout this section, we use the minimax family Γ j \Gamma_j Γ j defined in (9) . In addition to the HLS estimate (20) , the bound ∣ ξ − η ∣ ≤ 2 |\xi-\eta|\le2 ∣ ξ − η ∣ ≤ 2 gives
D ( w ) = ∬ S n × S n ∣ w ( ξ ) ∣ p ∣ w ( η ) ∣ p ∣ ξ − η ∣ μ d V ξ d V η ≥ 2 − μ ( ∫ S n ∣ w ∣ p d V ) 2 = 2 − μ ∥ w ∥ p 2 p . (39) D(w)=\iint_{\mathbb{S}^n\times\mathbb{S}^n}\frac{|w(\xi)|^p|w(\eta)|^p}{|\xi-\eta|^\mu}\,dV_\xi\,dV_\eta
\ge2^{-\mu}\left(\int_{\mathbb{S}^n}|w|^p\,dV\right)^2=2^{-\mu}\|w\|_p^{2p}. \tag*{(39)} D ( w ) = ∬ S n × S n ∣ ξ − η ∣ μ ∣ w ( ξ ) ∣ p ∣ w ( η ) ∣ p d V ξ d V η ≥ 2 − μ ( ∫ S n ∣ w ∣ p d V ) 2 = 2 − μ ∥ w ∥ p 2 p . (39)
For every subinterval I ⊂ [ 0 , π / 2 ] I\subset[0,\pi/2] I ⊂ [ 0 , π /2 ] , define
L ρ ( I ) = ∫ I ρ ( t ) − 1 / ( n − 1 ) d t , e f ( I ) = ∫ I ∣ f ′ ( t ) ∣ 2 ρ ( t ) d t . L_\rho(I)=\int_I\rho(t)^{-1/(n-1)}\,dt,\qquad e_f(I)=\int_I|f'(t)|^2\rho(t)\,dt. L ρ ( I ) = ∫ I ρ ( t ) − 1/ ( n − 1 ) d t , e f ( I ) = ∫ I ∣ f ′ ( t ) ∣ 2 ρ ( t ) d t .
The weight ρ \rho ρ in (30) satisfies
ρ ( t ) ≍ t ℓ − 1 ( 0 < t ≤ π / 4 ) , ρ ( t ) ≍ ( π / 2 − t ) k − 1 ( π / 4 ≤ t < π / 2 ) . \rho(t)\asymp t^{\ell-1}\quad(0<t\le\pi/4),\qquad\rho(t)\asymp(\pi/2-t)^{k-1}\quad(\pi/4\le t<\pi/2). ρ ( t ) ≍ t ℓ − 1 ( 0 < t ≤ π /4 ) , ρ ( t ) ≍ ( π /2 − t ) k − 1 ( π /4 ≤ t < π /2 ) .
Since k , ℓ < n k,\ell<n k , ℓ < n , we have
∫ 0 π / 2 ρ ( t ) − 1 / ( n − 1 ) d t < ∞ . \int_0^{\pi/2}\rho(t)^{-1/(n-1)}\,dt<\infty. ∫ 0 π /2 ρ ( t ) − 1/ ( n − 1 ) d t < ∞.
Only traces at interior points will be prescribed below. No trace condition is imposed at either singular orbit.
Lemma 4.1. Let f f f be the profile of a function in H H H . If I = [ a , b ] ⊂ [ 0 , π / 4 ] I=[a,b]\subset[0,\pi/4] I = [ a , b ] ⊂ [ 0 , π /4 ] , a < b a<b a < b , and f ( b ) = 0 f(b)=0 f ( b ) = 0 , then
∥ f ∥ L 2 ∗ ( I , ρ d t ) ≤ C L ρ ( I ) ( n − 1 ) / n e f ( I ) 1 / 2 . (40) \|f\|_{L^{2^*}(I,\rho\,dt)}\le CL_\rho(I)^{(n-1)/n}e_f(I)^{1/2}. \tag*{(40)} ∥ f ∥ L 2 ∗ ( I , ρ d t ) ≤ C L ρ ( I ) ( n − 1 ) / n e f ( I ) 1/2 . (40) Here C > 0 C>0 C > 0 depends only on n n n , k k k , and ℓ \ell ℓ . The same estimate holds on I = [ a , b ] ⊂ [ π / 4 , π / 2 ] I=[a,b]\subset[\pi/4,\pi/2] I = [ a , b ] ⊂ [ π /4 , π /2 ] with f ( a ) = 0 f(a)=0 f ( a ) = 0 .
Proof. On the left half, the profile belongs to H 1 H^1 H 1 on each compact subinterval of ( 0 , π / 2 ) (0,\pi/2) ( 0 , π /2 ) . For a < t < b a<t<b a < t < b ,
f ( t ) = − ∫ t b f ′ ( r ) d r , ∣ f ( t ) ∣ ≤ e f ( I ) 1 / 2 ( ∫ t b ρ ( r ) − 1 d r ) 1 / 2 . f(t)=-\int_t^b f'(r)\,dr,\qquad|f(t)|\le e_f(I)^{1/2}\left(\int_t^b\rho(r)^{-1}\,dr\right)^{1/2}. f ( t ) = − ∫ t b f ′ ( r ) d r , ∣ f ( t ) ∣ ≤ e f ( I ) 1/2 ( ∫ t b ρ ( r ) − 1 d r ) 1/2 .
It remains to estimate
( ∫ a b ρ ( t ) ( ∫ t b ρ ( r ) − 1 d r ) 2 ∗ / 2 d t ) 1 / 2 ∗ . \left(\int_a^b\rho(t)\left(\int_t^b\rho(r)^{-1}\,dr\right)^{2^*/2}\,dt\right)^{1/2^*}. ∫ a b ρ ( t ) ( ∫ t b ρ ( r ) − 1 d r ) 2 ∗ /2 d t 1/ 2 ∗ .
Suppose first that a ≥ b / 2 a \ge b/2 a ≥ b /2 . Then ρ ( t ) ≍ b ℓ − 1 \rho(t) \asymp b^{\ell-1} ρ ( t ) ≍ b ℓ − 1 on [ a , b ] [a,b] [ a , b ] , with constants independent of the interval. Consequently,
∫ a b ρ ( t ) ( ∫ t b ρ ( r ) − 1 d r ) 2 ∗ / 2 d t ≤ C b ( ℓ − 1 ) ( 1 − 2 ∗ / 2 ) ∫ a b ( b − t ) 2 ∗ / 2 d t ≤ C b ( ℓ − 1 ) ( 1 − 2 ∗ / 2 ) ( b − a ) 1 + 2 ∗ / 2 . \int_a^b \rho(t)\left(\int_t^b \rho(r)^{-1}\,dr\right)^{2^*/2}\,dt \le Cb^{(\ell-1)(1-2^*/2)}\int_a^b (b-t)^{2^*/2}\,dt
\le Cb^{(\ell-1)(1-2^*/2)}(b-a)^{1+2^*/2}. ∫ a b ρ ( t ) ( ∫ t b ρ ( r ) − 1 d r ) 2 ∗ /2 d t ≤ C b ( ℓ − 1 ) ( 1 − 2 ∗ /2 ) ∫ a b ( b − t ) 2 ∗ /2 d t ≤ C b ( ℓ − 1 ) ( 1 − 2 ∗ /2 ) ( b − a ) 1 + 2 ∗ /2 .
The identities
1 2 + 1 2 ∗ = n − 1 n , 1 2 ∗ − 1 2 = − 1 n \frac{1}{2}+\frac{1}{2^*}=\frac{n-1}{n}, \qquad\frac{1}{2^*}-\frac{1}{2}=-\frac{1}{n} 2 1 + 2 ∗ 1 = n n − 1 , 2 ∗ 1 − 2 1 = − n 1
give
∥ f ∥ L 2 ∗ ( I , ρ d t ) ≤ C e f ( I ) 1 / 2 ( b − a ) ( n − 1 ) / n b − ( ℓ − 1 ) / n . \|f\|_{L^{2^*}(I,\rho\,dt)}\le C e_f(I)^{1/2}(b-a)^{(n-1)/n}b^{-(\ell-1)/n}. ∥ f ∥ L 2 ∗ ( I , ρ d t ) ≤ C e f ( I ) 1/2 ( b − a ) ( n − 1 ) / n b − ( ℓ − 1 ) / n .
Also,
L ρ ( I ) ≍ ( b − a ) b − ( ℓ − 1 ) / ( n − 1 ) . L_\rho(I)\asymp(b-a)b^{-(\ell-1)/(n-1)}. L ρ ( I ) ≍ ( b − a ) b − ( ℓ − 1 ) / ( n − 1 ) .
These two estimates imply (40) .
Suppose next that a < b / 2 a<b/2 a < b /2 . Since [ b / 2 , b ] ⊂ I [b/2,b]\subset I [ b /2 , b ] ⊂ I ,
L ρ ( I ) ≥ C − 1 ∫ b / 2 b t − ( ℓ − 1 ) / ( n − 1 ) d t ≥ C − 1 b ( n − ℓ ) / ( n − 1 ) . L_\rho(I)\ge C^{-1}\int_{b/2}^b t^{-(\ell-1)/(n-1)}\,dt\ge C^{-1}b^{(n-\ell)/(n-1)}. L ρ ( I ) ≥ C − 1 ∫ b /2 b t − ( ℓ − 1 ) / ( n − 1 ) d t ≥ C − 1 b ( n − ℓ ) / ( n − 1 ) .
If ℓ > 2 \ell>2 ℓ > 2 , then
∫ t b ρ ( r ) − 1 d r ≤ C ∫ t b r 1 − ℓ d r ≤ C t 2 − ℓ . \int_t^b \rho(r)^{-1}\,dr\le C\int_t^b r^{1-\ell}\,dr\le Ct^{2-\ell}. ∫ t b ρ ( r ) − 1 d r ≤ C ∫ t b r 1 − ℓ d r ≤ C t 2 − ℓ .
Moreover,
ℓ + ( 2 − ℓ ) 2 ∗ 2 = 2 ( n − ℓ ) n − 2 > 0. \ell+\frac{(2-\ell)2^*}{2}=\frac{2(n-\ell)}{n-2}>0. ℓ + 2 ( 2 − ℓ ) 2 ∗ = n − 2 2 ( n − ℓ ) > 0.
We therefore obtain
∫ a b ρ ( t ) ( ∫ t b ρ ( r ) − 1 d r ) 2 ∗ / 2 d t ≤ C ∫ 0 b t ℓ − 1 + ( 2 − ℓ ) 2 ∗ / 2 d t ≤ C b ℓ + ( 2 − ℓ ) 2 ∗ / 2 . \int_a^b \rho(t)\left(\int_t^b \rho(r)^{-1}\,dr\right)^{2^*/2}\,dt\le C\int_0^b t^{\ell-1+(2-\ell)2^*/2}\,dt
\le Cb^{\ell+(2-\ell)2^*/2}. ∫ a b ρ ( t ) ( ∫ t b ρ ( r ) − 1 d r ) 2 ∗ /2 d t ≤ C ∫ 0 b t ℓ − 1 + ( 2 − ℓ ) 2 ∗ /2 d t ≤ C b ℓ + ( 2 − ℓ ) 2 ∗ /2 .
Taking the 2 ∗ 2^* 2 ∗ -th root gives the factor b ℓ / 2 + ( 2 − ℓ ) / 2 = b 1 − ℓ / n b^{\ell/2+(2-\ell)/2}=b^{1-\ell/n} b ℓ /2 + ( 2 − ℓ ) /2 = b 1 − ℓ / n . If ℓ = 2 \ell=2 ℓ = 2 , the power estimate is replaced by a logarithm. In this case,
∫ t b ρ ( r ) − 1 d r ≤ C log ( b / t ) \int_t^b \rho(r)^{-1}\,dr\le C\log(b/t) ∫ t b ρ ( r ) − 1 d r ≤ C log ( b / t )
and the substitution t = b u t=bu t = b u gives
∫ 0 b t ( log ( b / t ) ) 2 ∗ / 2 d t = b 2 ∫ 0 1 u ( log ( 1 / u ) ) 2 ∗ / 2 d u ≤ C b 2 . \int_0^b t(\log(b/t))^{2^*/2}\,dt=b^2\int_0^1u(\log(1/u))^{2^*/2}\,du\le Cb^2. ∫ 0 b t ( log ( b / t ) ) 2 ∗ /2 d t = b 2 ∫ 0 1 u ( log ( 1/ u ) ) 2 ∗ /2 d u ≤ C b 2 .
Its 2 ∗ 2^* 2 ∗ -th root is C b 2 / 2 ∗ = C b 1 − 2 / n Cb^{2/2^*}=Cb^{1-2/n} C b 2/ 2 ∗ = C b 1 − 2/ n . Thus both cases yield
∥ f ∥ L 2 ∗ ( I , ρ d t ) ≤ C e f ( I ) 1 / 2 b 1 − ℓ / n ≤ C e f ( I ) 1 / 2 L ρ ( I ) ( n − 1 ) / n . \|f\|_{L^{2^*}(I,\rho\,dt)}\le Ce_f(I)^{1/2}b^{1-\ell/n}\le Ce_f(I)^{1/2}L_\rho(I)^{(n-1)/n}. ∥ f ∥ L 2 ∗ ( I , ρ d t ) ≤ C e f ( I ) 1/2 b 1 − ℓ / n ≤ C e f ( I ) 1/2 L ρ ( I ) ( n − 1 ) / n .
If a = 0 a=0 a = 0 , apply the preceding estimate on [ ε , b ] [\varepsilon,b] [ ε , b ] and let ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 . Monotone convergence gives the same estimate on [ 0 , b ] [0,b] [ 0 , b ] . Reflection t ↦ π / 2 − t t\mapsto\pi/2-t t ↦ π /2 − t proves the right-half estimate, with k k k in place of ℓ \ell ℓ .
Proposition 4.2. There exist constants C 1 , C 2 > 0 C_1,C_2>0 C 1 , C 2 > 0 , independent of j j j , such that
C 1 j 2 p / ( p − 1 ) ≤ c j ≤ C 2 j 2 p / ( p − 1 ) , j ≥ 2. C_1j^{2p/(p-1)} \le c_j \le C_2j^{2p/(p-1)}, \qquad j \ge2. C 1 j 2 p / ( p − 1 ) ≤ c j ≤ C 2 j 2 p / ( p − 1 ) , j ≥ 2. Proof. Let B ∈ Γ j B \in\Gamma_j B ∈ Γ j , where j ≥ 2 j \ge2 j ≥ 2 . Choose positive integers j − j_- j − and j + j_+ j + such that
j − + j + = j , ∣ j − − j + ∣ ≤ 1. j_-+j_+=j,\qquad|j_--j_+|\le1. j − + j + = j , ∣ j − − j + ∣ ≤ 1. For the profile f f f of w ∈ B w \in B w ∈ B , put
a w ( t ) = ρ ( t ) − 1 / ( n − 1 ) + ρ ( t ) ∣ f ′ ( t ) ∣ 2 . a_w(t)=\rho(t)^{-1/(n-1)}+\rho(t)|f'(t)|^2. a w ( t ) = ρ ( t ) − 1/ ( n − 1 ) + ρ ( t ) ∣ f ′ ( t ) ∣ 2 . By (31) , the normalization ∥ w ∥ = 1 \lVert w\rVert=1 ∥ w ∥ = 1 , and the integrability of ρ − 1 / ( n − 1 ) \rho^{-1/(n-1)} ρ − 1/ ( n − 1 ) , there exists C > 0 C>0 C > 0 , independent of j j j and w w w , such that
0 < ∫ 0 π / 2 ρ ( t ) − 1 / ( n − 1 ) d t ≤ ∫ 0 π / 2 a w ( t ) d t ≤ C . 0<\int_0^{\pi/2}\rho(t)^{-1/(n-1)}\,dt\le\int_0^{\pi/2}a_w(t)\,dt\le C. 0 < ∫ 0 π /2 ρ ( t ) − 1/ ( n − 1 ) d t ≤ ∫ 0 π /2 a w ( t ) d t ≤ C . Partition [ 0 , π / 4 ] [0,\pi/4] [ 0 , π /4 ] into j − j_- j − intervals of equal a w a_w a w -mass and [ π / 4 , π / 2 ] [\pi/4,\pi/2] [ π /4 , π /2 ] into j + j_+ j + intervals in the same way. Since
a w ( t ) ≥ ρ ( t ) − 1 / ( n − 1 ) > 0 ( 0 < t < π / 2 ) , a_w(t)\ge\rho(t)^{-1/(n-1)}>0\qquad(0<t<\pi/2), a w ( t ) ≥ ρ ( t ) − 1/ ( n − 1 ) > 0 ( 0 < t < π /2 ) , the corresponding cumulative mass functions are continuous and strictly increasing. Hence all division points are uniquely determined. On each cell I I I ,
L ρ ( I ) + e f ( I ) = ∫ I a w ( t ) d t ≤ C j . (41) L_\rho(I)+e_f(I)=\int_I a_w(t)\,dt\le\frac{C}{j}. \tag*{(41)} L ρ ( I ) + e f ( I ) = ∫ I a w ( t ) d t ≤ j C . (41) Here we used min { j − , j + } ≥ j / 3 \min\{j_-,j_+\}\ge j/3 min { j − , j + } ≥ j /3 , also valid when j = 2 j=2 j = 2 .
To prove the continuity of the division points, let w m , w ∈ S w_m,w\in S w m , w ∈ S satisfy w m → w w_m\to w w m → w strongly in H H H , and denote their profiles by f m f_m f m and f f f , respectively. Then
∫ 0 π / 2 ρ ∣ ∣ f m ′ ∣ 2 − ∣ f ′ ∣ 2 ∣ d t ≤ ( ∫ 0 π / 2 ρ ∣ f m ′ − f ′ ∣ 2 d t ) 1 / 2 ( ( ∫ 0 π / 2 ρ ∣ f m ′ ∣ 2 d t ) 1 / 2 + ( ∫ 0 π / 2 ρ ∣ f ′ ∣ 2 d t ) 1 / 2 ) ⟶ 0. \int_0^{\pi/2}\rho\bigl||f'_m|^2-|f'|^2\bigr|\,dt
\le
\left(\int_0^{\pi/2}\rho|f'_m-f'|^2\,dt\right)^{1/2}
\left(\left(\int_0^{\pi/2}\rho|f'_m|^2\,dt\right)^{1/2}
+\left(\int_0^{\pi/2}\rho|f'|^2\,dt\right)^{1/2}\right)
\longrightarrow0. ∫ 0 π /2 ρ ∣ f m ′ ∣ 2 − ∣ f ′ ∣ 2 d t ≤ ( ∫ 0 π /2 ρ ∣ f m ′ − f ′ ∣ 2 d t ) 1/2 ( ∫ 0 π /2 ρ ∣ f m ′ ∣ 2 d t ) 1/2 + ( ∫ 0 π /2 ρ ∣ f ′ ∣ 2 d t ) 1/2 ⟶ 0. Hence the cumulative mass functions converge uniformly. Let t t t be a division point for w w w , and let ε > 0 \varepsilon>0 ε > 0 be small. Strict monotonicity puts the cumulative mass at t − ε t-\varepsilon t − ε below the prescribed fraction and that at t + ε t+\varepsilon t + ε above it. Uniform convergence preserves these inequalities for w m w_m w m . Its corresponding division point therefore lies between t − ε t-\varepsilon t − ε and t + ε t+\varepsilon t + ε . Thus every division point converges.
List all interior nodes, including π / 4 \pi/4 π /4 , as t 1 ( w ) < ⋯ < t j − 1 ( w ) t_1(w)<\cdots<t_{j-1}(w) t 1 ( w ) < ⋯ < t j − 1 ( w ) . They lie in ( 0 , π / 2 ) (0,\pi/2) ( 0 , π /2 ) . On each compact subinterval of ( 0 , π / 2 ) (0,\pi/2) ( 0 , π /2 ) , convergence in H H H implies convergence in the ordinary H 1 H^1 H 1 norm and hence uniform convergence. The map
B ⟶ R j − 1 , w ⟼ ( f ( t 1 ( w ) ) , … , f ( t j − 1 ( w ) ) ) B\longrightarrow\mathbb{R}^{j-1},\qquad
w\longmapsto\bigl(f(t_1(w)),\ldots,f(t_{j-1}(w))\bigr) B ⟶ R j − 1 , w ⟼ ( f ( t 1 ( w )) , … , f ( t j − 1 ( w )) ) is therefore continuous. It is odd, because a − w = a w a_{-w}=a_w a − w = a w and consequently t i ( − w ) = t i ( w ) t_i(-w)=t_i(w) t i ( − w ) = t i ( w ) . If it never vanished, its normalization would be a continuous odd map from B B B to S j − 2 S^{j-2} S j − 2 . The
index properties would then give i ( B ) ≤ j − 1 i(B) \le j-1 i ( B ) ≤ j − 1 , contrary to B ∈ Γ j B \in\Gamma_j B ∈ Γ j . We may thus choose w ∈ B w \in B w ∈ B whose profile vanishes at all nodes.
For every left-half cell, the profile vanishes at its right endpoint. For every right-half cell, it vanishes at its left endpoint. Lemma 4.1 and (41) imply
∥ f ∥ L 2 ∗ ( I , ρ d t ) ≤ C j − ( n − 1 ) / n − 1 / 2 . \|f\|_{L^{2^*}(I,\rho\,dt)} \le Cj^{-(n-1)/n-1/2}. ∥ f ∥ L 2 ∗ ( I , ρ d t ) ≤ C j − ( n − 1 ) / n − 1/2 .
The cells have disjoint interiors. Summing the 2 ∗ 2^* 2 ∗ -th powers and including the fixed sphere-area factors gives
∥ w ∥ 2 ∗ ≤ C j 1 / 2 ∗ − ( n − 1 ) / n − 1 / 2 = C j − 1 . \|w\|_{2^*} \le Cj^{1/2^*-(n-1)/n-1/2}=Cj^{-1}. ∥ w ∥ 2 ∗ ≤ C j 1/ 2 ∗ − ( n − 1 ) / n − 1/2 = C j − 1 .
It follows from (20) that
D ( w ) ≤ C j − 2 p , sup z ∈ B Ψ ( z ) ≥ Ψ ( w ) ≥ C − 1 j 2 p / ( p − 1 ) . \mathcal{D}(w) \le Cj^{-2p}, \qquad\sup_{z\in B}\Psi(z) \ge\Psi(w) \ge C^{-1}j^{2p/(p-1)}. D ( w ) ≤ C j − 2 p , z ∈ B sup Ψ ( z ) ≥ Ψ ( w ) ≥ C − 1 j 2 p / ( p − 1 ) .
Taking the infimum over B ∈ Γ j B \in\Gamma_j B ∈ Γ j proves the lower bound.
For the upper bound, fix [ a , a + b ] ⊂ ( 0 , π / 2 ) [a,a+b] \subset(0,\pi/2) [ a , a + b ] ⊂ ( 0 , π /2 ) , where b > 0 b>0 b > 0 , and choose a nonzero χ ∈ C c ∞ ( 0 , 1 ) \chi\in C^\infty_c(0,1) χ ∈ C c ∞ ( 0 , 1 ) . For z = ( z 1 , … , z j ) ∈ S j − 1 z=(z_1,\ldots,z_j)\in S^{j-1} z = ( z 1 , … , z j ) ∈ S j − 1 , set
h i = b z i 2 , a i = a + ∑ r < i h r , ∑ i = 1 j h i = b . h_i=bz_i^2,\qquad a_i=a+\sum_{r<i}h_r,\qquad\sum_{i=1}^j h_i=b. h i = b z i 2 , a i = a + r < i ∑ h r , i = 1 ∑ j h i = b .
If h i > 0 h_i>0 h i > 0 , define on [ a i , a i + h i ] [a_i,a_i+h_i] [ a i , a i + h i ]
f z ( t ) = sgn ( z i ) h i χ ( t − a i h i ) . f_z(t)=\operatorname{sgn}(z_i)h_i\chi\left(\frac{t-a_i}{h_i}\right). f z ( t ) = sgn ( z i ) h i χ ( h i t − a i ) .
Omit intervals of length zero and extend the profile by zero outside [ a , a + b ] [a,a+b] [ a , a + b ] . Since χ \chi χ vanishes near both ends, the pieces fit together in the weighted energy space.
The weight is bounded above and below on the fixed interval. Changing variables t = a i + h i u t=a_i+h_i u t = a i + h i u on each nonempty piece gives
∫ a i a i + h i ∣ f z ′ ∣ 2 ρ d t = h i ∫ 0 1 ∣ χ ′ ( u ) ∣ 2 ρ ( a i + h i u ) d u , \int_{a_i}^{a_i+h_i}|f_z'|^2\rho\,dt
=h_i\int_0^1|\chi'(u)|^2\rho(a_i+h_i u)\,du, ∫ a i a i + h i ∣ f z ′ ∣ 2 ρ d t = h i ∫ 0 1 ∣ χ ′ ( u ) ∣ 2 ρ ( a i + h i u ) d u ,
∫ a i a i + h i ∣ f z ∣ 2 ρ d t = h i 3 ∫ 0 1 ∣ χ ( u ) ∣ 2 ρ ( a i + h i u ) d u , \int_{a_i}^{a_i+h_i}|f_z|^2\rho\,dt
=h_i^3\int_0^1|\chi(u)|^2\rho(a_i+h_i u)\,du, ∫ a i a i + h i ∣ f z ∣ 2 ρ d t = h i 3 ∫ 0 1 ∣ χ ( u ) ∣ 2 ρ ( a i + h i u ) d u ,
∫ a i a i + h i ∣ f z ∣ p ρ d t = h i p + 1 ∫ 0 1 ∣ χ ( u ) ∣ p ρ ( a i + h i u ) d u . \int_{a_i}^{a_i+h_i}|f_z|^p\rho\,dt
=h_i^{p+1}\int_0^1|\chi(u)|^p\rho(a_i+h_i u)\,du. ∫ a i a i + h i ∣ f z ∣ p ρ d t = h i p + 1 ∫ 0 1 ∣ χ ( u ) ∣ p ρ ( a i + h i u ) d u .
Let v z v_z v z be the corresponding invariant function. Since ∫ 0 1 ∣ χ ′ ∣ 2 > 0 \int_0^1|\chi'|^2>0 ∫ 0 1 ∣ χ ′ ∣ 2 > 0 , these identities give
C − 1 ≤ ∥ v z ∥ ≤ C , ∥ v z ∥ p p ≥ C − 1 ∑ i = 1 j h i p + 1 ≥ C − 1 j − p . (42) C^{-1}\le\|v_z\|\le C,\qquad\|v_z\|_p^p\ge C^{-1}\sum_{i=1}^j h_i^{p+1}\ge C^{-1}j^{-p}. \tag*{(42)} C − 1 ≤ ∥ v z ∥ ≤ C , ∥ v z ∥ p p ≥ C − 1 i = 1 ∑ j h i p + 1 ≥ C − 1 j − p . (42)
The last step uses convexity and ∑ i h i = b \sum_i h_i=b ∑ i h i = b . The constants do not depend on j j j or z z z .
The map z ↦ v z z\mapsto v_z z ↦ v z is odd. It is continuous on each set where the nonzero coordinates have fixed signs, by continuity of translations and dilations of the smooth bump in H 1 H^1 H 1 . If a coordinate tends to zero, the squared energy norm of its piece is at most C ( h i + h i 3 ) C(h_i+h_i^3) C ( h i + h i 3 ) , which tends to zero. The remaining endpoints vary continuously. This proves continuity on the whole sphere, including all its coordinate faces.
By (42) , the normalized image
B j = { v z / ∥ v z ∥ : z ∈ S j − 1 } B_j = \{v_z/\lVert v_z\rVert: z \in\mathbb{S}^{j-1}\} B j = { v z / ∥ v z ∥ : z ∈ S j − 1 }
is a compact symmetric subset of S S S . The continuous odd map S j − 1 → B j \mathbb{S}^{j-1} \to B_j S j − 1 → B j gives i ( B j ) ≥ j i(B_j) \ge j i ( B j ) ≥ j .
For every w ∈ B j w \in B_j w ∈ B j , (39) and (42) yield
D ( w ) ≥ C − 1 j − 2 p , Ψ ( w ) ≤ C j 2 p / ( p − 1 ) . \mathcal{D}(w) \ge C^{-1}j^{-2p}, \qquad\Psi(w) \le Cj^{2p/(p-1)}. D ( w ) ≥ C − 1 j − 2 p , Ψ ( w ) ≤ C j 2 p / ( p − 1 ) .
Since B j ∈ Γ j B_j \in\Gamma_j B j ∈ Γ j , this proves the upper bound.
Attainment and sign change
The following minimax lemma relies on the monotonicity, subadditivity, and neighborhood properties of the cohomological index, together with the Z 2 \mathbb{Z}_2 Z 2 -equivariant deformation lemma.
Lemma 5.1. Let F F F be an even C 1 C^1 C 1 functional on the unit sphere of an infinite-dimensional Hilbert space. Assume that inf S F > 0 \inf_S F > 0 inf S F > 0 and that F F F satisfies the Palais–Smale condition. For
d j = inf B ∈ Γ j sup w ∈ B F ( w ) , d_j = \inf_{B \in\Gamma_j} \sup_{w \in B} F(w), d j = B ∈ Γ j inf w ∈ B sup F ( w ) , where Γ j \Gamma_j Γ j is defined by (9) , every finite d j d_j d j is a critical value. If d j = ⋯ = d j + s = d d_j = \cdots= d_{j+s} = d d j = ⋯ = d j + s = d , then the critical set K d K_d K d has i ( K d ) ≥ s + 1 i(K_d) \ge s + 1 i ( K d ) ≥ s + 1 , where K d = { w ∈ S : F ( w ) = d , d F ( w ) = 0 } K_d = \{w \in S : F(w) = d,\, dF(w) = 0\} K d = { w ∈ S : F ( w ) = d , d F ( w ) = 0 } .
Proof. Set F a = { w ∈ S : F ( w ) ≤ a } F^a = \{w \in S : F(w) \le a\} F a = { w ∈ S : F ( w ) ≤ a } . The Z 2 \mathbb{Z}_2 Z 2 -equivariant deformation lemma [33 Lemma 3.1] , applied on S S S to a tangential pseudogradient, shows that if d j d_j d j were not a critical value, then, for some ε > 0 \varepsilon> 0 ε > 0 , there would exist an odd homeomorphism η : S → S \eta: S \to S η : S → S such that η ( F d j + ε ) ⊂ F d j − ε \eta(F^{d_j+\varepsilon}) \subset F^{d_j-\varepsilon} η ( F d j + ε ) ⊂ F d j − ε . Choose B ∈ Γ j B \in\Gamma_j B ∈ Γ j such that sup B F < d j + ε \sup_B F < d_j + \varepsilon sup B F < d j + ε . Since η \eta η is an odd homeomorphism, η ( B ) \eta(B) η ( B ) is compact and symmetric, and i ( η ( B ) ) = i ( B ) ≥ j i(\eta(B)) = i(B) \ge j i ( η ( B )) = i ( B ) ≥ j . Thus η ( B ) ∈ Γ j \eta(B) \in\Gamma_j η ( B ) ∈ Γ j . On the other hand, sup η ( B ) F ≤ d j − ε < d j \sup_{\eta(B)} F \le d_j - \varepsilon< d_j sup η ( B ) F ≤ d j − ε < d j , which contradicts the definition of d j d_j d j . Therefore, every finite d j d_j d j is a critical value.
Suppose that d j = ⋯ = d j + s = d d_j = \cdots= d_{j+s} = d d j = ⋯ = d j + s = d and i ( K d ) ≤ s i(K_d) \le s i ( K d ) ≤ s . By the Palais–Smale condition, K d K_d K d is compact. The continuity property of the cohomological index [15 Theorem 5.1(5)] yields a symmetric open neighborhood U U U of K d K_d K d such that i ( U ‾ ) ≤ s i(\overline{U}) \le s i ( U ) ≤ s . The equivariant deformation lemma then sends F d + ε ∖ U F^{d+\varepsilon} \setminus U F d + ε ∖ U into F d − ε F^{d-\varepsilon} F d − ε for sufficiently small ε > 0 \varepsilon> 0 ε > 0 . Choose B ∈ Γ j + s B \in\Gamma_{j+s} B ∈ Γ j + s such that sup B F < d + ε \sup_B F < d + \varepsilon sup B F < d + ε . By the monotonicity and subadditivity properties of the index [15 Theorem 5.1(4),(6)] , we have
j + s ≤ i ( B ) ≤ i ( B ∖ U ) + i ( B ∩ U ‾ ) ≤ i ( B ∖ U ) + s . j+s \le i(B) \le i(B \setminus U) + i(B \cap\overline{U}) \le i(B \setminus U) + s. j + s ≤ i ( B ) ≤ i ( B ∖ U ) + i ( B ∩ U ) ≤ i ( B ∖ U ) + s .
Hence i ( B ∖ U ) ≥ j i(B \setminus U) \ge j i ( B ∖ U ) ≥ j . If η \eta η denotes the odd homeomorphism supplied by the deformation lemma, then
i ( η ( B ∖ U ) ) = i ( B ∖ U ) ≥ j . i(\eta(B \setminus U)) = i(B \setminus U) \ge j. i ( η ( B ∖ U )) = i ( B ∖ U ) ≥ j .
Thus η ( B ∖ U ) ∈ Γ j \eta(B \setminus U) \in\Gamma_j η ( B ∖ U ) ∈ Γ j , whereas
sup η ( B ∖ U ) F ≤ d − ε < d j , \sup_{\eta(B \setminus U)} F \le d - \varepsilon< d_j, η ( B ∖ U ) sup F ≤ d − ε < d j ,
contradicting the definition of d j d_j d j . Therefore, i ( K d ) ≥ s + 1 i(K_d) \ge s + 1 i ( K d ) ≥ s + 1 .
We first classify the one-signed solutions in H H H .
Lemma 5.2. Let v ∈ H ∖ { 0 } v \in H \setminus\{0\} v ∈ H ∖ { 0 } be a weak solution of (1) . If v v v does not change sign, then v = a 0 v = a_0 v = a 0 or v = − a 0 v = -a_0 v = − a 0 , where a 0 a_0 a 0 is defined in (11) . The energy of these constant solutions is
E ( a 0 ) = E ( − a 0 ) = p − 1 2 p n ( n − 2 ) 4 a 0 2 ∣ S n ∣ = c 1 . (43) E(a_0) = E(-a_0) = \frac{p-1}{2p}\frac{n(n-2)}{4}a_0^2|\mathbb{S}^n| = c_1. \tag*{(43)} E ( a 0 ) = E ( − a 0 ) = 2 p p − 1 4 n ( n − 2 ) a 0 2 ∣ S n ∣ = c 1 . (43) Moreover, the only G G G -invariant solutions at the first minimax level c 1 c_1 c 1 are a 0 a_0 a 0 and − a 0 -a_0 − a 0 .
Proof. Suppose first that v ≥ 0 v \ge0 v ≥ 0 , and set u = P v u = Pv u = P v . Proposition 2.1 shows that u u u is a nonnegative, nonzero weak solution of (5) . By [1 Theorem 2.3] , u ∈ C l o c 2 , α ( R n ) u \in C_{\mathrm{loc}}^{2,\alpha}(\mathbb{R}^n) u ∈ C loc 2 , α ( R n ) for some α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) . Since
− Δ u = ( ∫ R n ∣ u ( y ) ∣ p ∣ x − y ∣ μ d y ) u p − 1 ≥ 0 , -\Delta u = \left(\int_{\mathbb{R}^n}\frac{|u(y)|^p}{|x-y|^\mu}\,dy\right)u^{p-1} \ge0, − Δ u = ( ∫ R n ∣ x − y ∣ μ ∣ u ( y ) ∣ p d y ) u p − 1 ≥ 0 ,
the strong maximum principle [21 Theorem 3.5] and u ≢ 0 u \not\equiv0 u ≡ 0 yield u > 0 u > 0 u > 0 on R n \mathbb{R}^n R n . Apply [10 Corollary 1.5] with their s = 1 s=1 s = 1 , σ = μ \sigma=\mu σ = μ , α = 2 \alpha=2 α = 2 , m = 0 m=0 m = 0 and their two powers equal to p p p and p − 1 p-1 p − 1 . Indeed, the two critical powers in that classification become
2 n − μ n − 2 = p , n + 2 − μ n − 2 = p − 1 > 0. \frac{2n-\mu}{n-2}=p,\qquad\frac{n+2-\mu}{n-2}=p-1>0. n − 2 2 n − μ = p , n − 2 n + 2 − μ = p − 1 > 0.
The resulting Euclidean bubble, pulled back by P − 1 P^{-1} P − 1 , has the form
v ( ξ ) = C ( 1 − a ⋅ ξ ) − ( n − 2 ) / 2 , C > 0 , a ∈ R n + 1 , ∣ a ∣ < 1. (44) v(\xi)=C(1-a\cdot\xi)^{-(n-2)/2},\qquad C>0,\quad a\in\mathbb{R}^{n+1},\quad|a|<1. \tag*{(44)} v ( ξ ) = C ( 1 − a ⋅ ξ ) − ( n − 2 ) /2 , C > 0 , a ∈ R n + 1 , ∣ a ∣ < 1. (44)
The G G G -invariance of v v v yields v ( g ξ ) = v ( ξ ) v(g\xi)=v(\xi) v ( g ξ ) = v ( ξ ) for all g ∈ G g\in G g ∈ G . Substituting (44) yields
1 − a ⋅ g ξ = 1 − a ⋅ ξ , g ∈ G , ξ ∈ S n , 1-a\cdot g\xi=1-a\cdot\xi,\qquad g\in G,\quad\xi\in\mathbb{S}^n, 1 − a ⋅ g ξ = 1 − a ⋅ ξ , g ∈ G , ξ ∈ S n ,
Thus g − 1 a = a g^{-1}a=a g − 1 a = a for every g ∈ G g\in G g ∈ G . The action of O ( k ) × O ( ℓ ) O(k)\times O(\ell) O ( k ) × O ( ℓ ) on R k × R ℓ \mathbb{R}^k\times\mathbb{R}^{\ell} R k × R ℓ has no nonzero fixed vector. Hence a = 0 a=0 a = 0 , and (44) shows that v v v is constant. Write v ≡ b v\equiv b v ≡ b with b > 0 b>0 b > 0 . Then
A v = n ( n − 2 ) 4 b , J μ ( ∣ v ∣ p ) ∣ v ∣ p − 2 v = κ 0 b 2 p − 1 . Av=\frac{n(n-2)}{4}b,\qquad J_\mu(|v|^p)|v|^{p-2}v=\kappa_0b^{2p-1}. A v = 4 n ( n − 2 ) b , J μ ( ∣ v ∣ p ) ∣ v ∣ p − 2 v = κ 0 b 2 p − 1 .
Consequently, (1) reduces to
n ( n − 2 ) 4 b = κ 0 b 2 p − 1 . \frac{n(n-2)}{4}b=\kappa_0b^{2p-1}. 4 n ( n − 2 ) b = κ 0 b 2 p − 1 .
Dividing by b > 0 b>0 b > 0 gives b = a 0 b=a_0 b = a 0 . If v ≤ 0 v\le0 v ≤ 0 , then − v -v − v is a nonnegative nonzero solution. The preceding argument gives v = − a 0 v=-a_0 v = − a 0 .
For v = a 0 v=a_0 v = a 0 , testing (1) against v v v gives ∥ v ∥ 2 = D ( v ) \|v\|^2=D(v) ∥ v ∥ 2 = D ( v ) . Consequently,
E ( a 0 ) = 1 2 ∥ a 0 ∥ 2 − 1 2 p D ( a 0 ) = p − 1 2 p ∥ a 0 ∥ 2 = p − 1 2 p n ( n − 2 ) 4 a 0 2 ∣ S n ∣ . E(a_0)=\frac{1}{2}\|a_0\|^2-\frac{1}{2p}D(a_0)=\frac{p-1}{2p}\|a_0\|^2=\frac{p-1}{2p}\frac{n(n-2)}{4}a_0^2|\mathbb{S}^n|. E ( a 0 ) = 2 1 ∥ a 0 ∥ 2 − 2 p 1 D ( a 0 ) = 2 p p − 1 ∥ a 0 ∥ 2 = 2 p p − 1 4 n ( n − 2 ) a 0 2 ∣ S n ∣.
Since E E E is even, E ( − a 0 ) = E ( a 0 ) E(-a_0)=E(a_0) E ( − a 0 ) = E ( a 0 ) . It remains to prove that this common value is c 1 c_1 c 1 and to identify all invariant solutions at this level. Since every B ∈ Γ 1 B\in\Gamma_1 B ∈ Γ 1 is contained in S S S , while { w , − w } ∈ Γ 1 \{w,-w\}\in\Gamma_1 { w , − w } ∈ Γ 1 and Ψ ( − w ) = Ψ ( w ) \Psi(-w)=\Psi(w) Ψ ( − w ) = Ψ ( w ) for every w ∈ S w\in S w ∈ S , it follows from the definition of c 1 c_1 c 1 that
c 1 = inf w ∈ S Ψ ( w ) . (45) c_1=\inf_{w\in S}\Psi(w). \tag*{(45)} c 1 = w ∈ S inf Ψ ( w ) . (45)
Let S n > 0 S_n>0 S n > 0 be the best constant in
S n ∥ w ∥ L 2 ( S n ) 2 ≤ ∥ w ∥ 2 , w ∈ H 1 ( S n ) , S_n\|w\|_{L^2(S^n)}^2\le\|w\|^2,\qquad w\in H^1(S^n), S n ∥ w ∥ L 2 ( S n ) 2 ≤ ∥ w ∥ 2 , w ∈ H 1 ( S n ) ,
and let C H L S ( n , μ ) \mathrm{CHLS}(n,\mu) CHLS ( n , μ ) denote the best constant in the spherical HLS inequality. The sharp HLS inequality [27 ] , in the spherical formulation of [17 Theorem 2.2] , and the sharp Sobolev inequality [17 Theorem 3.1] and [4 ] give
D ( w ) ≤ C H L S ( n , μ ) ∥ w ∥ L 2 ∗ ( S n ) 2 p ≤ C H L S ( n , μ ) S n − p ∥ w ∥ 2 p . (46) \mathcal{D}(w) \leq\mathrm{C}_{\mathrm{HLS}}(n,\mu)\lVert w\rVert_{L^{2^*}(S^n)}^{2p} \leq\mathrm{C}_{\mathrm{HLS}}(n,\mu)S_n^{-p}\lVert w\rVert^{2p}. \tag*{(46)} D ( w ) ≤ C HLS ( n , μ ) ∥ w ∥ L 2 ∗ ( S n ) 2 p ≤ C HLS ( n , μ ) S n − p ∥ w ∥ 2 p . (46)
Set C ∗ = C H L S ( n , μ ) S n − p C_* = C_{\mathrm{HLS}}(n,\mu)S_n^{-p} C ∗ = C HLS ( n , μ ) S n − p . For w ∈ S w \in\mathcal{S} w ∈ S , (46) becomes D ( w ) ≤ C ∗ \mathcal{D}(w) \leq C_* D ( w ) ≤ C ∗ . Constant functions attain equality in both sharp inequalities. Therefore,
sup w ∈ S D ( w ) = C ∗ . \sup_{w\in\mathcal{S}} \mathcal{D}(w) = C_*. w ∈ S sup D ( w ) = C ∗ .
Equations (8) and (45) now give
c 1 = p − 1 2 p C ∗ − 1 / ( p − 1 ) . (47) c_1 = \frac{p-1}{2p}C_*^{-1/(p-1)}. \tag*{(47)} c 1 = 2 p p − 1 C ∗ − 1/ ( p − 1 ) . (47)
Let w 0 ∈ S w_0 \in\mathcal{S} w 0 ∈ S be the positive normalized constant. Then
D ( w 0 ) = C ∗ , Ψ ( w 0 ) = c 1 . \mathcal{D}(w_0) = C_*, \qquad\Psi(w_0) = c_1. D ( w 0 ) = C ∗ , Ψ ( w 0 ) = c 1 .
The radial rescaling in Lemma 2.5 transforms w 0 w_0 w 0 into the positive constant solution a 0 a_0 a 0 . Hence,
E ( a 0 ) = Ψ ( w 0 ) = c 1 . E(a_0) = \Psi(w_0) = c_1. E ( a 0 ) = Ψ ( w 0 ) = c 1 .
Finally, let v ∈ H ∖ { 0 } v \in H \setminus\{0\} v ∈ H ∖ { 0 } be any critical point of E ∣ H E|_H E ∣ H satisfying E ( v ) = c 1 E(v) = c_1 E ( v ) = c 1 . Testing the equation with v v v gives ∥ v ∥ 2 = D ( v ) \lVert v\rVert^2 = \mathcal{D}(v) ∥ v ∥ 2 = D ( v ) . Define w = v / ∥ v ∥ ∈ S w = v/\lVert v\rVert\in\mathcal{S} w = v / ∥ v ∥ ∈ S . The 2 p 2p 2 p -homogeneity of D \mathcal{D} D yields
D ( w ) = D ( v ) ∥ v ∥ 2 p = ∥ v ∥ 2 − 2 p . \mathcal{D}(w) = \frac{\mathcal{D}(v)}{\lVert v\rVert^{2p}} = \lVert v\rVert^{2-2p}. D ( w ) = ∥ v ∥ 2 p D ( v ) = ∥ v ∥ 2 − 2 p .
Therefore,
Ψ ( w ) = p − 1 2 p D ( w ) − 1 / ( p − 1 ) = p − 1 2 p ∥ v ∥ 2 = E ( v ) = c 1 . \Psi(w) = \frac{p-1}{2p}\mathcal{D}(w)^{-1/(p-1)} = \frac{p-1}{2p}\lVert v\rVert^2 = E(v) = c_1. Ψ ( w ) = 2 p p − 1 D ( w ) − 1/ ( p − 1 ) = 2 p p − 1 ∥ v ∥ 2 = E ( v ) = c 1 .
By the definition of Ψ \Psi Ψ , the identity Ψ ( w ) = E ( v ) = c 1 \Psi(w) = E(v) = c_1 Ψ ( w ) = E ( v ) = c 1 , and (47) , one has
p − 1 2 p D ( w ) − 1 / ( p − 1 ) = c 1 = p − 1 2 p C ∗ − 1 / ( p − 1 ) . \frac{p-1}{2p}\mathcal{D}(w)^{-1/(p-1)} = c_1 = \frac{p-1}{2p}C_*^{-1/(p-1)}. 2 p p − 1 D ( w ) − 1/ ( p − 1 ) = c 1 = 2 p p − 1 C ∗ − 1/ ( p − 1 ) .
Therefore, D ( w ) = C ∗ \mathcal{D}(w) = C_* D ( w ) = C ∗ . Moreover, w ∈ S w \in\mathcal{S} w ∈ S , and hence ∥ w ∥ = 1 \lVert w\rVert= 1 ∥ w ∥ = 1 . Equation (46) now gives
C ∗ = D ( w ) ≤ C H L S ( n , μ ) ∥ w ∥ L 2 ∗ ( S n ) 2 p ≤ C H L S ( n , μ ) S n − p ∥ w ∥ 2 p = C ∗ . C_* = \mathcal{D}(w) \leq C_{\mathrm{HLS}}(n,\mu)\lVert w\rVert_{L^{2^*}(S^n)}^{2p} \leq C_{\mathrm{HLS}}(n,\mu)S_n^{-p}\lVert w\rVert^{2p} = C_*. C ∗ = D ( w ) ≤ C HLS ( n , μ ) ∥ w ∥ L 2 ∗ ( S n ) 2 p ≤ C HLS ( n , μ ) S n − p ∥ w ∥ 2 p = C ∗ .
Thus equality holds throughout (46) . In particular, equality holds in the sharp Sobolev inequality. The equality cases of the Sobolev inequality [17 Theorem 3.1] and [4 ] give
w ( ξ ) = C ( 1 − a ⋅ ξ ) − ( n − 2 ) / 2 w(\xi) = C(1-a\cdot\xi)^{-(n-2)/2} w ( ξ ) = C ( 1 − a ⋅ ξ ) − ( n − 2 ) /2
up to an overall sign, where ∣ a ∣ < 1 |a| < 1 ∣ a ∣ < 1 . The G G G -invariance of w w w again implies a = 0 a = 0 a = 0 . Hence, w w w and v v v are constant. Substitution into (1) gives v = a 0 v = a_0 v = a 0 or v = − a 0 v = -a_0 v = − a 0 .
Proof of Theorem 1.1. Proposition 4.2 gives
C 1 j 2 p / ( p − 1 ) ≤ c j ≤ C 2 j 2 p / ( p − 1 ) , j ≥ 2. C_1j^{2p/(p-1)} \le c_j \le C_2j^{2p/(p-1)}, \qquad j \ge2. C 1 j 2 p / ( p − 1 ) ≤ c j ≤ C 2 j 2 p / ( p − 1 ) , j ≥ 2.
Together with Lemma 5.2, which gives c 1 = E ( a 0 ) < ∞ c_1 = E(a_0) < \infty c 1 = E ( a 0 ) < ∞ , this shows that every c j c_j c j is finite. Moreover, c j → ∞ c_j \to\infty c j → ∞ as j → ∞ j \to\infty j → ∞ . By Lemma 2.4, the functional Ψ ∈ C 1 ( S ) \Psi\in C^1(\mathcal{S}) Ψ ∈ C 1 ( S ) is even, bounded away from zero, and satisfies the Palais–Smale condition. Hence Lemma 5.1 shows that every c j c_j c j is a critical value of Ψ \Psi Ψ . Thus, for each j ≥ 1 j \ge1 j ≥ 1 , there exists w j ∈ S w_j \in\mathcal{S} w j ∈ S such that
d Ψ ( w j ) = 0 , Ψ ( w j ) = c j . d\Psi(w_j) = 0, \qquad\Psi(w_j) = c_j. d Ψ ( w j ) = 0 , Ψ ( w j ) = c j .
Set
v j = D ( w j ) − 1 / ( 2 p − 2 ) w j . v_j = D(w_j)^{-1/(2p-2)}w_j. v j = D ( w j ) − 1/ ( 2 p − 2 ) w j .
Lemma 2.5 gives
( E ∣ H ) ′ ( v j ) = 0 , E ( v j ) = Ψ ( w j ) = c j . (E|_H)'(v_j) = 0, \qquad E(v_j) = \Psi(w_j) = c_j. ( E ∣ H ) ′ ( v j ) = 0 , E ( v j ) = Ψ ( w j ) = c j .
It remains to verify criticality in the full space. Let d g dg d g be the normalized Haar measure on G G G , and let ϕ ∈ H 1 ( S n ) \phi\in H^1(\mathbb{S}^n) ϕ ∈ H 1 ( S n ) . Define
ϕ ‾ ( ξ ) = ∫ G ϕ ( g ξ ) d g . \overline{\phi}(\xi) = \int_G \phi(g\xi)\,dg. ϕ ( ξ ) = ∫ G ϕ ( g ξ ) d g .
Then ϕ ‾ ∈ H \overline{\phi} \in H ϕ ∈ H . Since v j v_j v j is G G G -invariant and the action of G G G preserves E E E ,
E ′ ( v j ) [ ϕ ] = E ′ ( v j ) [ ϕ ( g ⋅ ) ] ( g ∈ G ) . E'(v_j)[\phi] = E'(v_j)[\phi(g\cdot)] \qquad(g \in G). E ′ ( v j ) [ ϕ ] = E ′ ( v j ) [ ϕ ( g ⋅ )] ( g ∈ G ) .
Averaging this identity over G G G yields
E ′ ( v j ) [ ϕ ] = E ′ ( v j ) [ ϕ ‾ ] = 0. E'(v_j)[\phi] = E'(v_j)[\overline{\phi}] = 0. E ′ ( v j ) [ ϕ ] = E ′ ( v j ) [ ϕ ] = 0.
Therefore, v j v_j v j is a weak solution of (1.1) in the full space. By Lemma 2.5 and Lemma 5.2, the critical set of Ψ \Psi Ψ at c 1 c_1 c 1 is K c 1 = { w 0 , − w 0 } K_{c_1} = \{w_0,-w_0\} K c 1 = { w 0 , − w 0 } , where w 0 w_0 w 0 is the positive normalized constant. Hence i ( K c 1 ) = 1 i(K_{c_1}) = 1 i ( K c 1 ) = 1 . If c 2 = c 1 c_2 = c_1 c 2 = c 1 , the plateau conclusion in Lemma 5.1, applied with j = 1 j = 1 j = 1 and s = 1 s = 1 s = 1 , would give i ( K c 1 ) ≥ 2 i(K_{c_1}) \ge2 i ( K c 1 ) ≥ 2 , a contradiction. Therefore, c 2 > c 1 c_2 > c_1 c 2 > c 1 . Since Γ j + 1 ⊂ Γ j \Gamma_{j+1} \subset\Gamma_j Γ j + 1 ⊂ Γ j , the sequence ( c j ) (c_j) ( c j ) is nondecreasing. Consequently,
c j ≥ c 2 > c 1 , j ≥ 2. c_j \ge c_2 > c_1, \qquad j \ge2. c j ≥ c 2 > c 1 , j ≥ 2.
Lemma 5.2 also shows that every nonzero one-signed solution in H H H has energy c 1 c_1 c 1 . Therefore, every invariant solution at a level c j c_j c j , j ≥ 2 j \ge2 j ≥ 2 , changes sign. Finally, since ( c j ) (c_j) ( c j ) is nondecreasing and c j → ∞ c_j \to\infty c j → ∞ , it contains a strictly increasing subsequence. The corresponding critical points yield infinitely many distinct pairs of G G G -invariant sign-changing solutions.
Regularity and nodal geometry
We first prove regularity and describe the zeros of the orbit profile. The proof of Theorem 1.3 is completed in Section 7.
Lemma 6.1. Let v ∈ H 1 ( S n ) v \in H^1(\mathbb{S}^n) v ∈ H 1 ( S n ) be a weak solution of (1.1). Then v ∈ L ∞ ( S n ) v \in L^\infty(\mathbb{S}^n) v ∈ L ∞ ( S n ) . Moreover, there exists γ ∈ ( 0 , 1 ) \gamma\in(0,1) γ ∈ ( 0 , 1 ) such that
J μ ( ∣ v ∣ p ) ∈ C 0 , γ ( S n ) , v ∈ C 2 , γ ( S n ) . (48) J_\mu(|v|^p) \in C^{0,\gamma}(\mathbb{S}^n), \qquad v \in C^{2,\gamma}(\mathbb{S}^n). \tag*{(48)} J μ ( ∣ v ∣ p ) ∈ C 0 , γ ( S n ) , v ∈ C 2 , γ ( S n ) . (48)
Proof. In the chart determined by π \pi π , let u = P v u=Pv u = P v . Proposition 2.1 shows that u u u is a Euclidean weak solution. By the rotational invariance of the equation, rotating the omitted pole shows that the same argument applies in every stereographic chart. Set
I u ( x ) : = ∫ R n ∣ u ( y ) ∣ p ∣ x − y ∣ μ d y . I_u(x):=\int_{\mathbb{R}^n}\frac{|u(y)|^p}{|x-y|^\mu}\,dy. I u ( x ) := ∫ R n ∣ x − y ∣ μ ∣ u ( y ) ∣ p d y .
The regularity result [1 Theorem 2.3 and Appendix A] gives, for some α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) ,
u ∈ C l o c 2 , α ( R n ) , I u ∈ C l o c 0 , α ( R n ) . u\in C_{\mathrm{loc}}^{2,\alpha}(\mathbb{R}^n),\qquad I_u\in C_{\mathrm{loc}}^{0,\alpha}(\mathbb{R}^n). u ∈ C loc 2 , α ( R n ) , I u ∈ C loc 0 , α ( R n ) .
Moreover, the conformal change of variables gives
I u ( x ) = Ω ( x ) μ / 2 J μ ( ∣ v ∣ p ) ( π − 1 ( x ) ) . I_u(x)=\Omega(x)^{\mu/2}J_\mu(|v|^p)(\pi^{-1}(x)). I u ( x ) = Ω ( x ) μ /2 J μ ( ∣ v ∣ p ) ( π − 1 ( x )) .
Thus v v v and J μ ( ∣ v ∣ p ) J_\mu(|v|^p) J μ ( ∣ v ∣ p ) have the stated local Hölder regularity in every chart. Choose finitely many coordinate neighborhoods whose closures are contained in stereographic charts and which cover S n \mathbb{S}^n S n . On each of them, the conformal factor is smooth and bounded away from zero. Taking the minimum of their Hölder exponents gives, for some γ ∈ ( 0 , 1 ) \gamma\in(0,1) γ ∈ ( 0 , 1 ) ,
J μ ( ∣ v ∣ p ) ∈ C 0 , γ ( S n ) , v ∈ C 2 , γ ( S n ) . J_\mu(|v|^p)\in C^{0,\gamma}(\mathbb{S}^n),\qquad v\in C^{2,\gamma}(\mathbb{S}^n). J μ ( ∣ v ∣ p ) ∈ C 0 , γ ( S n ) , v ∈ C 2 , γ ( S n ) .
The compactness of S n \mathbb{S}^n S n also gives v ∈ L ∞ ( S n ) v\in L^\infty(\mathbb{S}^n) v ∈ L ∞ ( S n ) .
For the remainder of this section, assume p ≥ 2 p\ge2 p ≥ 2 and fix a nonzero G G G -invariant weak solution v ∈ H v\in H v ∈ H . For v ( ξ ) = f ( t ) v(\xi)=f(t) v ( ξ ) = f ( t ) , define
V f ( t ) = J μ ( ∣ v ∣ p ) ( cos t ω , sin t θ ) . (49) V_f(t)=J_\mu(|v|^p)(\cos t\,\omega,\sin t\,\theta). \tag*{(49)} V f ( t ) = J μ ( ∣ v ∣ p ) ( cos t ω , sin t θ ) . (49)
The value in (49) is independent of ω \omega ω and θ \theta θ . By (48) , it is continuous and nonnegative. If v ≠ 0 v\ne0 v = 0 , strict positivity of the kernel gives V f ( t ) > 0 V_f(t)>0 V f ( t ) > 0 for every t ∈ [ 0 , π / 2 ] t\in[0,\pi/2] t ∈ [ 0 , π /2 ] .
Lemma 6.2. The profile f f f of the fixed nonzero G G G -invariant weak solution v v v satisfies
− ( ρ f ′ ) ′ + n ( n − 2 ) 4 ρ f = ρ V f ∣ f ∣ p − 2 f on ( 0 , π / 2 ) , (50) -(\rho f')'+\frac{n(n-2)}{4}\rho f=\rho V_f|f|^{p-2}f\quad\text{on }(0,\pi/2), \tag*{(50)} − ( ρ f ′ ) ′ + 4 n ( n − 2 ) ρ f = ρ V f ∣ f ∣ p − 2 f on ( 0 , π /2 ) , (50) and
f ′ ( 0 ) = f ′ ( π / 2 ) = 0. (51) f'(0)=f'(\pi/2)=0. \tag*{(51)} f ′ ( 0 ) = f ′ ( π /2 ) = 0. (51)
Proof. By Lemma 3.1 ,
Δ S n v = ρ − 1 ( ρ f ′ ) ′ . \Delta_{\mathbb{S}^n}v=\rho^{-1}(\rho f')'. Δ S n v = ρ − 1 ( ρ f ′ ) ′ .
Substituting this identity and the definition of V f V_f V f into (1) gives
− ρ − 1 ( ρ f ′ ) ′ + n ( n − 2 ) 4 f = V f ∣ f ∣ p − 2 f on ( 0 , π / 2 ) . -\rho^{-1}(\rho f')'+\frac{n(n-2)}{4}f=V_f|f|^{p-2}f\quad\text{on }(0,\pi/2). − ρ − 1 ( ρ f ′ ) ′ + 4 n ( n − 2 ) f = V f ∣ f ∣ p − 2 f on ( 0 , π /2 ) .
Multiplication by ρ \rho ρ yields (50) .
It remains to verify the endpoint conditions. Fix ω ∈ S k − 1 \omega\in\mathbb{S}^{k-1} ω ∈ S k − 1 and define
F ( y ) = v ( 1 − ∣ y ∣ 2 ω , y ) F(y)=v\left(\sqrt{1-|y|^2}\,\omega,y\right) F ( y ) = v ( 1 − ∣ y ∣ 2 ω , y )
for y ∈ R ℓ y \in\mathbb{R}^{\ell} y ∈ R ℓ near 0. The O ( ℓ ) O(\ell) O ( ℓ ) -invariance of v v v gives F ( − y ) = F ( y ) F(-y) = F(y) F ( − y ) = F ( y ) . Differentiating this identity at y = 0 y = 0 y = 0 yields D F ( 0 ) = 0 DF(0) = 0 D F ( 0 ) = 0 . For every θ ∈ S ℓ − 1 \theta\in S^{\ell-1} θ ∈ S ℓ − 1 ,
f ( t ) = F ( sin t θ ) , f ′ ( t ) = D F ( sin t θ ) [ cos t θ ] . f(t) = F(\sin t\theta), \qquad f'(t) = DF(\sin t\theta)[\cos t\theta]. f ( t ) = F ( sin tθ ) , f ′ ( t ) = D F ( sin tθ ) [ cos tθ ] .
Evaluating at t = 0 t = 0 t = 0 gives f ′ ( 0 ) = D F ( 0 ) [ θ ] = 0 f'(0) = DF(0)[\theta] = 0 f ′ ( 0 ) = D F ( 0 ) [ θ ] = 0 . To treat the other endpoint, fix θ ∈ S ℓ − 1 \theta\in S^{\ell-1} θ ∈ S ℓ − 1 and define
G ( x ) = v ( x , 1 − ∣ x ∣ 2 θ ) G(x) = v\left(x,\sqrt{1-|x|^2}\theta\right) G ( x ) = v ( x , 1 − ∣ x ∣ 2 θ )
for x ∈ R k x \in\mathbb{R}^{k} x ∈ R k near 0. The O ( k ) O(k) O ( k ) -invariance of v v v gives G ( − x ) = G ( x ) G(-x) = G(x) G ( − x ) = G ( x ) , and hence D G ( 0 ) = 0 DG(0) = 0 D G ( 0 ) = 0 . For every ω ∈ S k − 1 \omega\in S^{k-1} ω ∈ S k − 1 ,
f ( t ) = G ( cos t ω ) , f ′ ( t ) = D G ( cos t ω ) [ − sin t ω ] . f(t) = G(\cos t\omega), \qquad f'(t) = DG(\cos t\omega)[-\sin t\omega]. f ( t ) = G ( cos t ω ) , f ′ ( t ) = D G ( cos t ω ) [ − sin t ω ] .
Evaluating at t = π / 2 t = \pi/2 t = π /2 gives f ′ ( π / 2 ) = D G ( 0 ) [ − ω ] = 0 f'(\pi/2) = DG(0)[-\omega] = 0 f ′ ( π /2 ) = D G ( 0 ) [ − ω ] = 0 . Therefore, (51) holds.
Lemma 6.3. Assume p ≥ 2 p \ge2 p ≥ 2 . A nonzero solution profile has no zero at 0 0 0 or π / 2 \pi/2 π /2 . Every zero in ( 0 , π / 2 ) (0,\pi/2) ( 0 , π /2 ) is simple, and the number of zeros is finite.
Proof. In the open orbit interval ρ > 0 \rho> 0 ρ > 0 , so (50) rewrites as
f ′ ′ + ρ ′ ρ f ′ − c ( t ) f = 0 , c ( t ) = n ( n − 2 ) 4 − V f ( t ) ∣ f ( t ) ∣ p − 2 . f'' + \frac{\rho'}{\rho}f' - c(t)f = 0, \qquad c(t) = \frac{n(n-2)}{4} - V_f(t)|f(t)|^{p-2}. f ′′ + ρ ρ ′ f ′ − c ( t ) f = 0 , c ( t ) = 4 n ( n − 2 ) − V f ( t ) ∣ f ( t ) ∣ p − 2 . The coefficient c c c is continuous by the regularity established above and the assumption p ≥ 2 p \ge2 p ≥ 2 , with the convention ∣ f ∣ 0 ≡ 1 |f|^0 \equiv1 ∣ f ∣ 0 ≡ 1 when p = 2 p = 2 p = 2 . Uniqueness for a linear ODE with continuous coefficients [9 Chapter 1] implies that f ( t 0 ) = f ′ ( t 0 ) = 0 f(t_0) = f'(t_0) = 0 f ( t 0 ) = f ′ ( t 0 ) = 0 at an interior point forces f ≡ 0 f \equiv0 f ≡ 0 on ( 0 , π / 2 ) (0,\pi/2) ( 0 , π /2 ) , and then on its closure by continuity. Every interior zero of a nontrivial profile is therefore simple.
We exclude endpoint zeros next. Suppose that f ( 0 ) = 0 f(0) = 0 f ( 0 ) = 0 . Equation (50) becomes
( ρ f ′ ) ′ = ρ c ( t ) f . (\rho f')' = \rho c(t)f. ( ρ f ′ ) ′ = ρ c ( t ) f . The coefficient c c c is bounded near 0, and (51) gives f ′ ( 0 ) = 0 f'(0) = 0 f ′ ( 0 ) = 0 . Integration over ( 0 , t ) (0,t) ( 0 , t ) yields
ρ ( t ) f ′ ( t ) = ∫ 0 t ρ ( s ) c ( s ) f ( s ) d s . \rho(t)f'(t) = \int_0^t \rho(s)c(s)f(s)\,ds. ρ ( t ) f ′ ( t ) = ∫ 0 t ρ ( s ) c ( s ) f ( s ) d s . For sufficiently small t > 0 t > 0 t > 0 , there exist constants C 1 , C 2 > 0 C_1,C_2 > 0 C 1 , C 2 > 0 such that
C 1 t ℓ − 1 ≤ ρ ( t ) ≤ C 2 t ℓ − 1 . C_1t^{\ell-1} \le\rho(t) \le C_2t^{\ell-1}. C 1 t ℓ − 1 ≤ ρ ( t ) ≤ C 2 t ℓ − 1 . Define M ( t ) = max 0 ≤ s ≤ t ∣ f ( s ) ∣ M(t) = \max_{0 \le s \le t}|f(s)| M ( t ) = max 0 ≤ s ≤ t ∣ f ( s ) ∣ . The preceding estimates give
∣ f ′ ( t ) ∣ ≤ C ρ ( t ) ∫ 0 t ρ ( s ) ∣ f ( s ) ∣ d s ≤ C t 1 − ℓ M ( t ) ∫ 0 t s ℓ − 1 d s ≤ C t M ( t ) . |f'(t)| \le\frac{C}{\rho(t)}\int_0^t \rho(s)|f(s)|\,ds \le Ct^{1-\ell}M(t)\int_0^t s^{\ell-1}\,ds \le CtM(t). ∣ f ′ ( t ) ∣ ≤ ρ ( t ) C ∫ 0 t ρ ( s ) ∣ f ( s ) ∣ d s ≤ C t 1 − ℓ M ( t ) ∫ 0 t s ℓ − 1 d s ≤ C tM ( t ) . Since f ( 0 ) = 0 f(0) = 0 f ( 0 ) = 0 , integration over ( 0 , s ) (0,s) ( 0 , s ) , with 0 ≤ s ≤ t 0 \le s \le t 0 ≤ s ≤ t , gives
∣ f ( s ) ∣ ≤ ∫ 0 s ∣ f ′ ( r ) ∣ d r ≤ C M ( t ) ∫ 0 s r d r ≤ C t 2 M ( t ) . |f(s)| \le\int_0^s |f'(r)|\,dr \le CM(t)\int_0^s r\,dr \le Ct^2M(t). ∣ f ( s ) ∣ ≤ ∫ 0 s ∣ f ′ ( r ) ∣ d r ≤ C M ( t ) ∫ 0 s r d r ≤ C t 2 M ( t ) . Taking the maximum over s ∈ [ 0 , t ] s \in[0,t] s ∈ [ 0 , t ] yields
M ( t ) ≤ C t 2 M ( t ) . M(t) \le Ct^2M(t). M ( t ) ≤ C t 2 M ( t ) .
Choose t > 0 t>0 t > 0 such that C t 2 < 1 Ct^2<1 C t 2 < 1 . Then M ( t ) = 0 M(t)=0 M ( t ) = 0 , and hence f ≡ 0 f\equiv0 f ≡ 0 on [ 0 , t ] [0,t] [ 0 , t ] . Interior uniqueness extends this identity to [ 0 , π / 2 ] [0,\pi/2] [ 0 , π /2 ] , which contradicts the assumption that f f f is nonzero. Therefore, f ( 0 ) ≠ 0 f(0)\ne0 f ( 0 ) = 0 .
To treat the other endpoint, set r = π / 2 − t r=\pi/2-t r = π /2 − t , f ~ ( r ) = f ( π / 2 − r ) \tilde f(r)=f(\pi/2-r) f ~ ( r ) = f ( π /2 − r ) , and ρ ~ ( r ) = ρ ( π / 2 − r ) \tilde\rho(r)=\rho(\pi/2-r) ρ ~ ( r ) = ρ ( π /2 − r ) . Near r = 0 r=0 r = 0 ,
C ~ 1 r k − 1 ≤ ρ ~ ( r ) ≤ C ~ 2 r k − 1 . \widetilde C_1r^{k-1}\le\tilde\rho(r)\le\widetilde C_2r^{k-1}. C 1 r k − 1 ≤ ρ ~ ( r ) ≤ C 2 r k − 1 .
Applying the preceding argument to f ~ \tilde f f ~ , together with f ′ ( π / 2 ) = 0 f'(\pi/2)=0 f ′ ( π /2 ) = 0 from (51) , gives f ( π / 2 ) ≠ 0 f(\pi/2)\ne0 f ( π /2 ) = 0 .
Finally, suppose that the interior zeros have an accumulation point t ∗ ∈ ( 0 , π / 2 ) t_*\in(0,\pi/2) t ∗ ∈ ( 0 , π /2 ) . For a sequence of distinct zeros t m → t ∗ t_m\to t_* t m → t ∗ , continuity gives f ( t ∗ ) = 0 f(t_*)=0 f ( t ∗ ) = 0 , while differentiability gives
f ′ ( t ∗ ) = lim m → ∞ f ( t m ) − f ( t ∗ ) t m − t ∗ = 0. f'(t_*)=\lim_{m\to\infty}\frac{f(t_m)-f(t_*)}{t_m-t_*}=0. f ′ ( t ∗ ) = m → ∞ lim t m − t ∗ f ( t m ) − f ( t ∗ ) = 0.
This contradicts the simplicity of the interior zeros. Accumulation at 0 0 0 or π / 2 \pi/2 π /2 is excluded by f ( 0 ) ≠ 0 f(0)\ne0 f ( 0 ) = 0 , f ( π / 2 ) ≠ 0 f(\pi/2)\ne0 f ( π /2 ) = 0 , and continuity. An infinite subset of the compact interval [ 0 , π / 2 ] [0,\pi/2] [ 0 , π /2 ] has an accumulation point. Therefore, f f f has only finitely many zeros.
The Hessian and nodal domain bounds
Assume p ≥ 2 p\ge2 p ≥ 2 throughout this section, and set H − 1 ( S n ) : = ( H 1 ( S n ) ) ∗ H^{-1}(S^n):=(H^1(S^n))^* H − 1 ( S n ) := ( H 1 ( S n ) ) ∗ . At p = 2 p=2 p = 2 , the convention ∣ v ∣ 0 = 1 |v|^0=1 ∣ v ∣ 0 = 1 is used in the second-variation formulas. Recall that ⟨ ⋅ , ⋅ ⟩ A \langle\cdot,\cdot\rangle_A ⟨ ⋅ , ⋅ ⟩ A denotes the inner product associated with the energy norm in (6) , and that ⟨ ⋅ , ⋅ ⟩ H \langle\cdot,\cdot\rangle_H ⟨ ⋅ , ⋅ ⟩ H is its restriction to H H H . The weak realization of the conformal Laplacian A A A defined in (2) is
A : H 1 ( S n ) ⟶ H − 1 ( S n ) , A:H^1(S^n)\longrightarrow H^{-1}(S^n), A : H 1 ( S n ) ⟶ H − 1 ( S n ) ,
characterized by
⟨ A ϕ , ψ ⟩ H − 1 , H 1 = ⟨ ϕ , ψ ⟩ A , ϕ , ψ ∈ H 1 ( S n ) . (52) \langle A\phi,\psi\rangle_{H^{-1},H^1}=\langle\phi,\psi\rangle_A,\qquad\phi,\psi\in H^1(S^n). \tag*{(52)} ⟨ A ϕ , ψ ⟩ H − 1 , H 1 = ⟨ ϕ , ψ ⟩ A , ϕ , ψ ∈ H 1 ( S n ) . (52)
The second variation is given by the following lemma.
Lemma 7.1. If p ≥ 2 p\ge2 p ≥ 2 , then D D D and E E E are of class C 2 C^2 C 2 on H 1 ( S n ) H^1(S^n) H 1 ( S n ) . For v , ϕ , ψ ∈ H 1 ( S n ) v,\phi,\psi\in H^1(S^n) v , ϕ , ψ ∈ H 1 ( S n ) ,
D ′ ′ ( v ) [ ϕ , ψ ] = 2 p ( p − 1 ) ∫ S n J μ ( ∣ v ∣ p ) ∣ v ∣ p − 2 ϕ ψ d V + 2 p 2 ∬ S n × S n ∣ v ( ξ ) ∣ p − 2 v ( ξ ) ϕ ( ξ ) ∣ v ( η ) ∣ p − 2 v ( η ) ψ ( η ) ∣ ξ − η ∣ μ d V ξ d V η . (53) D''(v)[\phi,\psi]=2p(p-1)\int_{S^n}J_\mu(|v|^p)|v|^{p-2}\phi\psi\,dV+2p^2\iint_{S^n\times S^n}\frac{|v(\xi)|^{p-2}v(\xi)\phi(\xi)|v(\eta)|^{p-2}v(\eta)\psi(\eta)}{|\xi-\eta|^\mu}\,dV_\xi dV_\eta. \tag*{(53)} D ′′ ( v ) [ ϕ , ψ ] = 2 p ( p − 1 ) ∫ S n J μ ( ∣ v ∣ p ) ∣ v ∣ p − 2 ϕ ψ d V + 2 p 2 ∬ S n × S n ∣ ξ − η ∣ μ ∣ v ( ξ ) ∣ p − 2 v ( ξ ) ϕ ( ξ ) ∣ v ( η ) ∣ p − 2 v ( η ) ψ ( η ) d V ξ d V η . (53)
Proof. Since p q μ = 2 ∗ pq_\mu=2^* p q μ = 2 ∗ , the HLS and Hölder inequalities give
∣ ∫ S n J μ ( ∣ v ∣ p ) ∣ v ∣ p − 2 ϕ ψ d V ∣ ≤ C ∥ v ∥ 2 2 p − 2 ∥ ϕ ∥ 2 ∗ ∥ ψ ∥ 2 ∗ \left|\int_{S^n}J_\mu(|v|^p)|v|^{p-2}\phi\psi\,dV\right|\le C\|v\|_2^{2p-2}\|\phi\|_{2^*}\|\psi\|_{2^*} ∫ S n J μ ( ∣ v ∣ p ) ∣ v ∣ p − 2 ϕ ψ d V ≤ C ∥ v ∥ 2 2 p − 2 ∥ ϕ ∥ 2 ∗ ∥ ψ ∥ 2 ∗
and
∣ ∫ S n J μ ( ∣ v ∣ p − 2 v ϕ ) ∣ v ∣ p − 2 v ψ d V ∣ ≤ C ∥ v ∥ 2 2 p − 2 ∥ ϕ ∥ 2 ∗ ∥ ψ ∥ 2 ∗ . \left|\int_{S^n}J_\mu(|v|^{p-2}v\phi)|v|^{p-2}v\psi\,dV\right|\le C\|v\|_2^{2p-2}\|\phi\|_{2^*}\|\psi\|_{2^*}. ∫ S n J μ ( ∣ v ∣ p − 2 v ϕ ) ∣ v ∣ p − 2 v ψ d V ≤ C ∥ v ∥ 2 2 p − 2 ∥ ϕ ∥ 2 ∗ ∥ ψ ∥ 2 ∗ .
Let B v B_v B v denote the bilinear form on the right-hand side of (53) . The estimates above give ∥ B v ∥ ≤ C ∥ v ∥ 2 2 p − 2 \|B_v\|\le C\|v\|_2^{2p-2} ∥ B v ∥ ≤ C ∥ v ∥ 2 2 p − 2 in the bilinear operator norm. We next prove the continuity of v ↦ B v v \mapsto B_v v ↦ B v . Let v m → v v_m \to v v m → v in H 1 ( S n ) H^1(S^n) H 1 ( S n ) . Then v m → v v_m \to v v m → v in L 2 ∗ ( S n ) L^{2^*}(S^n) L 2 ∗ ( S n ) , and the power maps satisfy
∣ v m ∣ p → ∣ v ∣ p in L q ( S n ) , ∣ v m ∣ p − 2 v m → ∣ v ∣ p − 2 v in L 2 ∗ / ( p − 1 ) ( S n ) . |v_m|^p \to|v|^p \quad\text{in } L^q(S^n), \qquad|v_m|^{p-2}v_m \to|v|^{p-2}v \quad\text{in } L^{2^*/(p-1)}(S^n). ∣ v m ∣ p → ∣ v ∣ p in L q ( S n ) , ∣ v m ∣ p − 2 v m → ∣ v ∣ p − 2 v in L 2 ∗ / ( p − 1 ) ( S n ) .
If p > 2 p>2 p > 2 , one also has ∣ v m ∣ p − 2 → ∣ v ∣ p − 2 |v_m|^{p-2} \to|v|^{p-2} ∣ v m ∣ p − 2 → ∣ v ∣ p − 2 in L 2 ∗ / ( p − 2 ) ( S n ) L^{2^*/(p-2)}(S^n) L 2 ∗ / ( p − 2 ) ( S n ) . For p = 2 p=2 p = 2 , this factor is identically one under the convention ∣ v ∣ 0 ≡ 1 |v|^0 \equiv1 ∣ v ∣ 0 ≡ 1 . Applying the HLS and Hölder inequalities to the differences gives ∥ B v m − B v ∥ → 0 \|B_{v_m}-B_v\|\to0 ∥ B v m − B v ∥ → 0 . The same power-map estimates also give the continuity of D ′ : H 1 ( S n ) ⟶ H − 1 ( S n ) D':H^1(S^n)\longrightarrow H^{-1}(S^n) D ′ : H 1 ( S n ) ⟶ H − 1 ( S n ) . For smooth v v v and h h h , differentiation along v + t h v+th v + t h gives
D ′ ( v + h ) − D ′ ( v ) − B v [ h , ⋅ ] = ∫ 0 1 ( B v + t h − B v ) [ h , ⋅ ] d t . D'(v+h)-D'(v)-B_v[h,\cdot]=\int_0^1(B_{v+th}-B_v)[h,\cdot]\,dt. D ′ ( v + h ) − D ′ ( v ) − B v [ h , ⋅ ] = ∫ 0 1 ( B v + t h − B v ) [ h , ⋅ ] d t .
Approximation in H 1 ( S n ) H^1(S^n) H 1 ( S n ) extends this identity to every v , h ∈ H 1 ( S n ) v,h\in H^1(S^n) v , h ∈ H 1 ( S n ) . Consequently,
∥ D ′ ( v + h ) − D ′ ( v ) − B v [ h , ⋅ ] ∥ ( H 1 ( S n ) ) ∗ ≤ sup 0 ≤ t ≤ 1 ∥ B v + t h − B v ∥ ∥ h ∥ = o ( ∥ h ∥ ) . \left\|D'(v+h)-D'(v)-B_v[h,\cdot]\right\|_{(H^1(S^n))^*}\leq\sup_{0\leq t\leq1}\|B_{v+th}-B_v\|\|h\|=o(\|h\|). ∥ D ′ ( v + h ) − D ′ ( v ) − B v [ h , ⋅ ] ∥ ( H 1 ( S n ) ) ∗ ≤ 0 ≤ t ≤ 1 sup ∥ B v + t h − B v ∥∥ h ∥ = o ( ∥ h ∥ ) .
The last relation follows from the continuity of v ↦ B v v\mapsto B_v v ↦ B v in the bilinear operator norm. Thus D ′ ′ ( v ) = B v D''(v)=B_v D ′′ ( v ) = B v and D ∈ C 2 D\in C^2 D ∈ C 2 . Formula (1.7) then gives E ∈ C 2 E\in C^2 E ∈ C 2 .
At a critical point v v v , the Hessian quadratic form is
Q v ( ϕ ) = ∥ ϕ ∥ 2 − ( p − 1 ) ∫ S n J μ ( ∣ v ∣ p ) ∣ v ∣ p − 2 ϕ 2 d V − p ∬ S n × S n ∣ v ( ξ ) ∣ p − 2 v ( ξ ) ϕ ( ξ ) ∣ v ( η ) ∣ p − 2 v ( η ) ϕ ( η ) ∣ ξ − η ∣ μ d V ξ d V η . (54) \begin{aligned}
Q_v(\phi)={}&\|\phi\|^2-(p-1)\int_{S^n}J_\mu(|v|^p)|v|^{p-2}\phi^2\,dV\\
&-p\iint_{S^n\times S^n}\frac{|v(\xi)|^{p-2}v(\xi)\phi(\xi)|v(\eta)|^{p-2}v(\eta)\phi(\eta)}{|\xi-\eta|^\mu}\,dV_\xi dV_\eta. \tag*{(54)}
\end{aligned} Q v ( ϕ ) = ∥ ϕ ∥ 2 − ( p − 1 ) ∫ S n J μ ( ∣ v ∣ p ) ∣ v ∣ p − 2 ϕ 2 d V − p ∬ S n × S n ∣ ξ − η ∣ μ ∣ v ( ξ ) ∣ p − 2 v ( ξ ) ϕ ( ξ ) ∣ v ( η ) ∣ p − 2 v ( η ) ϕ ( η ) d V ξ d V η . (54)
We use two operator realizations of the Hessian form. The energy-space realization
T v : H 1 ( S n ) ⟶ H 1 ( S n ) T_v:H^1(S^n)\longrightarrow H^1(S^n) T v : H 1 ( S n ) ⟶ H 1 ( S n )
is defined, by the Riesz representation theorem, through
⟨ T v ϕ , ψ ⟩ A = E ′ ′ ( v ) [ ϕ , ψ ] , ϕ , ψ ∈ H 1 ( S n ) . (55) \langle T_v\phi,\psi\rangle_A=E''(v)[\phi,\psi],\qquad\phi,\psi\in H^1(S^n). \tag*{(55)} ⟨ T v ϕ , ψ ⟩ A = E ′′ ( v ) [ ϕ , ψ ] , ϕ , ψ ∈ H 1 ( S n ) . (55)
Lemma 7.2. Let p ≥ 2 p\geq2 p ≥ 2 and let v ∈ H 1 ( S n ) v\in H^1(S^n) v ∈ H 1 ( S n ) be a weak solution. Formula (7.3) holds for every ϕ ∈ H 1 ( S n ) \phi\in H^1(S^n) ϕ ∈ H 1 ( S n ) . Moreover, there exists a compact self-adjoint operator K v : H 1 ( S n ) → H 1 ( S n ) K_v:H^1(S^n)\to H^1(S^n) K v : H 1 ( S n ) → H 1 ( S n ) such that
T v = I − K v . T_v=I-K_v. T v = I − K v . Consequently, v v v has finite Morse index and nullity.
Proof. Differentiating the energy (1.7) gives
E ′ ′ ( v ) [ ϕ , ψ ] = ⟨ ϕ , ψ ⟩ A − 1 2 p D ′ ′ ( v ) [ ϕ , ψ ] . E''(v)[\phi,\psi]=\langle\phi,\psi\rangle_A-\frac{1}{2p}D''(v)[\phi,\psi]. E ′′ ( v ) [ ϕ , ψ ] = ⟨ ϕ , ψ ⟩ A − 2 p 1 D ′′ ( v ) [ ϕ , ψ ] .
Lemma 7.1 therefore yields (7.3). By Lemma 6.1, for some γ ∈ ( 0 , 1 ) \gamma\in(0,1) γ ∈ ( 0 , 1 ) ,
v ∈ C 2 , γ ( S n ) , J μ ( ∣ v ∣ p ) ∈ C 0 , γ ( S n ) . v\in C^{2,\gamma}(S^n),\qquad J_\mu(|v|^p)\in C^{0,\gamma}(S^n). v ∈ C 2 , γ ( S n ) , J μ ( ∣ v ∣ p ) ∈ C 0 , γ ( S n ) .
Since S n S^n S n is compact, both functions are bounded. Hence,
a v : = ∣ v ∣ p − 2 v , V v : = J μ ( ∣ v ∣ p ) ∣ v ∣ p − 2 (56) a_v:=|v|^{p-2}v,\qquad V_v:=J_\mu(|v|^p)|v|^{p-2} \tag*{(56)} a v := ∣ v ∣ p − 2 v , V v := J μ ( ∣ v ∣ p ) ∣ v ∣ p − 2 (56)
belong to L ∞ ( S n ) L^\infty(S^n) L ∞ ( S n ) .
Let ϕ m ⇀ 0 \phi_m \rightharpoonup0 ϕ m ⇀ 0 in H 1 ( S n ) H^1(S^n) H 1 ( S n ) . The Rellich compact embedding on the compact sphere [2 ] gives
ϕ m ⟶ 0 in L 2 ( S n ) . \phi_m \longrightarrow0 \quad\text{in } L^2(S^n). ϕ m ⟶ 0 in L 2 ( S n ) .
For every ψ ∈ H 1 ( S n ) \psi\in H^1(S^n) ψ ∈ H 1 ( S n ) , Hölder’s inequality and the continuous embedding H 1 ( S n ) ↪ L 2 ( S n ) H^1(S^n) \hookrightarrow L^2(S^n) H 1 ( S n ) ↪ L 2 ( S n ) yield
∣ ∫ S n V v ϕ m ψ d V ∣ ≤ ∥ V v ∥ L ∞ ( S n ) ∥ ϕ m ∥ L 2 ( S n ) ∥ ψ ∥ L 2 ( S n ) ≤ C ∥ V v ∥ L ∞ ( S n ) ∥ ϕ m ∥ L 2 ( S n ) ∥ ψ ∥ H 1 ( S n ) . \left|\int_{S^n} V_v\phi_m\psi\,dV\right| \leq\|V_v\|_{L^\infty(S^n)}\|\phi_m\|_{L^2(S^n)}\|\psi\|_{L^2(S^n)}
\leq C\|V_v\|_{L^\infty(S^n)}\|\phi_m\|_{L^2(S^n)}\|\psi\|_{H^1(S^n)}. ∫ S n V v ϕ m ψ d V ≤ ∥ V v ∥ L ∞ ( S n ) ∥ ϕ m ∥ L 2 ( S n ) ∥ ψ ∥ L 2 ( S n ) ≤ C ∥ V v ∥ L ∞ ( S n ) ∥ ϕ m ∥ L 2 ( S n ) ∥ ψ ∥ H 1 ( S n ) .
Consequently,
sup ∥ ψ ∥ H 1 ( S n ) ≤ 1 ∣ ∫ S n V v ϕ m ψ d V ∣ ⟶ 0. (57) \sup_{\|\psi\|_{H^1(S^n)}\leq1}\left|\int_{S^n} V_v\phi_m\psi\,dV\right| \longrightarrow0. \tag*{(57)} ∥ ψ ∥ H 1 ( S n ) ≤ 1 sup ∫ S n V v ϕ m ψ d V ⟶ 0. (57)
For K ( ξ , η ) = ∣ ξ − η ∣ − μ K(\xi,\eta)=|\xi-\eta|^{-\mu} K ( ξ , η ) = ∣ ξ − η ∣ − μ , rotational invariance and (11) give
sup ξ ∫ S n K ( ξ , η ) d V η = sup η ∫ S n K ( ξ , η ) d V ξ = κ 0 . \sup_{\xi}\int_{S^n}K(\xi,\eta)\,dV_\eta=\sup_{\eta}\int_{S^n}K(\xi,\eta)\,dV_\xi=\kappa_0. ξ sup ∫ S n K ( ξ , η ) d V η = η sup ∫ S n K ( ξ , η ) d V ξ = κ 0 .
The Cauchy–Schwarz inequality with respect to the measure K ( ξ , η ) d V η K(\xi,\eta)\,dV_\eta K ( ξ , η ) d V η , followed by Fubini’s theorem, yields
∥ J μ ( g ) ∥ L 2 ( S n ) ≤ κ 0 ∥ g ∥ L 2 ( S n ) , g ∈ L 2 ( S n ) . (58) \|J_\mu(g)\|_{L^2(S^n)}\leq\kappa_0\|g\|_{L^2(S^n)},\qquad g\in L^2(S^n). \tag*{(58)} ∥ J μ ( g ) ∥ L 2 ( S n ) ≤ κ 0 ∥ g ∥ L 2 ( S n ) , g ∈ L 2 ( S n ) . (58)
Moreover,
∥ a v ϕ m ∥ L 2 ( S n ) ≤ ∥ a v ∥ L ∞ ( S n ) ∥ ϕ m ∥ L 2 ( S n ) ⟶ 0. \|a_v\phi_m\|_{L^2(S^n)}\leq\|a_v\|_{L^\infty(S^n)}\|\phi_m\|_{L^2(S^n)}\longrightarrow0. ∥ a v ϕ m ∥ L 2 ( S n ) ≤ ∥ a v ∥ L ∞ ( S n ) ∥ ϕ m ∥ L 2 ( S n ) ⟶ 0.
Applying (58) , Hölder’s inequality, and the continuous embedding H 1 ( S n ) ↪ L 2 ( S n ) H^1(S^n)\hookrightarrow L^2(S^n) H 1 ( S n ) ↪ L 2 ( S n ) , one obtains
∣ ∫ S n J μ ( a v ϕ m ) a v ψ d V ∣ ≤ ∥ J μ ( a v ϕ m ) ∥ L 2 ( S n ) ∥ a v ψ ∥ L 2 ( S n ) ≤ C ∥ a v ϕ m ∥ L 2 ( S n ) ∥ a v ∥ L ∞ ( S n ) ∥ ψ ∥ H 1 ( S n ) . \left|\int_{S^n}J_\mu(a_v\phi_m)a_v\psi\,dV\right|
\leq\|J_\mu(a_v\phi_m)\|_{L^2(S^n)}\|a_v\psi\|_{L^2(S^n)}
\leq C\|a_v\phi_m\|_{L^2(S^n)}\|a_v\|_{L^\infty(S^n)}\|\psi\|_{H^1(S^n)}. ∫ S n J μ ( a v ϕ m ) a v ψ d V ≤ ∥ J μ ( a v ϕ m ) ∥ L 2 ( S n ) ∥ a v ψ ∥ L 2 ( S n ) ≤ C ∥ a v ϕ m ∥ L 2 ( S n ) ∥ a v ∥ L ∞ ( S n ) ∥ ψ ∥ H 1 ( S n ) .
Hence,
sup ∥ ψ ∥ H 1 ( S n ) ≤ 1 ∣ ∫ S n J μ ( a v ϕ m ) a v ψ d V ∣ ⟶ 0. (59) \sup_{\|\psi\|_{H^1(S^n)}\leq1}\left|\int_{S^n}J_\mu(a_v\phi_m)a_v\psi\,dV\right|\longrightarrow0. \tag*{(59)} ∥ ψ ∥ H 1 ( S n ) ≤ 1 sup ∫ S n J μ ( a v ϕ m ) a v ψ d V ⟶ 0. (59)
By the Riesz representation theorem, define K v : H 1 ( S n ) → H 1 ( S n ) K_v:H^1(S^n)\to H^1(S^n) K v : H 1 ( S n ) → H 1 ( S n ) through the energy inner product by
⟨ K v ϕ , ψ ⟩ A = ( p − 1 ) ∫ S n V v ϕ ψ d V + p ∫ S n J μ ( a v ϕ ) a v ψ d V . \langle K_v\phi,\psi\rangle_A=(p-1)\int_{S^n}V_v\phi\psi\,dV+p\int_{S^n}J_\mu(a_v\phi)a_v\psi\,dV. ⟨ K v ϕ , ψ ⟩ A = ( p − 1 ) ∫ S n V v ϕ ψ d V + p ∫ S n J μ ( a v ϕ ) a v ψ d V .
Equations (57) and (59) imply
ϕ m ⟶ 0 in H 1 ( S n ) ⟹ K v ϕ m ⟶ 0 in H 1 ( S n ) . \phi_m\longrightarrow0\ \text{in } H^1(S^n)\quad\Longrightarrow\quad K_v\phi_m\longrightarrow0\ \text{in } H^1(S^n). ϕ m ⟶ 0 in H 1 ( S n ) ⟹ K v ϕ m ⟶ 0 in H 1 ( S n ) .
Thus, K v K_v K v is compact.
The multiplication term is symmetric in ϕ \phi ϕ and ψ \psi ψ . The symmetry of the Riesz kernel gives
∫ S n J μ ( f ) g d V = ∫ S n f J μ ( g ) d V , \int_{S^n}J_\mu(f)g\,dV=\int_{S^n}fJ_\mu(g)\,dV, ∫ S n J μ ( f ) g d V = ∫ S n f J μ ( g ) d V ,
so the nonlocal term is also symmetric. Therefore, K v K_v K v is self-adjoint with respect to the energy inner product.
By the definitions of T v T_v T v and K v K_v K v ,
⟨ T v ϕ , ψ ⟩ A = ⟨ ( I − K v ) ϕ , ψ ⟩ A , ϕ , ψ ∈ H 1 ( S n ) . \langle T_v\phi,\psi\rangle_A=\langle(I-K_v)\phi,\psi\rangle_A,\qquad\phi,\psi\in H^1(S^n). ⟨ T v ϕ , ψ ⟩ A = ⟨( I − K v ) ϕ , ψ ⟩ A , ϕ , ψ ∈ H 1 ( S n ) .
Hence, T v = I − K v T_v=I-K_v T v = I − K v . Since K v K_v K v is compact and self-adjoint, the spectral theorem shows that K v K_v K v has only finitely many eigenvalues greater than 1 1 1 , counted with multiplicity, and that its eigenspace corresponding to the eigenvalue 1 1 1 is finite-dimensional. Therefore, T v T_v T v has only finitely many negative eigenvalues and
dim ker T v < ∞ . \dim\ker T_v<\infty. dim ker T v < ∞.
Hence the Morse index and nullity of v v v are finite.
We next introduce the L 2 ( S n ) L^2(S^n) L 2 ( S n ) -realization of the same Hessian form. Using the coefficients introduced in (7.5), define
L v ϕ = A ϕ − ( p − 1 ) V v ϕ − p a v J μ ( a v ϕ ) , dom ( L v ) = H 2 ( S n ) . (60) L_v\phi=A\phi-(p-1)V_v\phi-pa_vJ_\mu(a_v\phi),\qquad\operatorname{dom}(L_v)=H^2(S^n). \tag*{(60)} L v ϕ = A ϕ − ( p − 1 ) V v ϕ − p a v J μ ( a v ϕ ) , dom ( L v ) = H 2 ( S n ) . (60)
The boundedness of a v a_v a v and V v V_v V v , together with (7.7), shows that the two lower-order terms in (7.9) define bounded operators on L 2 ( S n ) L^2(S^n) L 2 ( S n ) . Thus
L v = A + B v , B v ϕ : = − ( p − 1 ) V v ϕ − p a v J μ ( a v ϕ ) . L_v=A+B_v,\qquad B_v\phi:=-(p-1)V_v\phi-pa_vJ_\mu(a_v\phi). L v = A + B v , B v ϕ := − ( p − 1 ) V v ϕ − p a v J μ ( a v ϕ ) .
The symmetry identity used above shows that the nonlocal part of B v B_v B v is self-adjoint. Hence B v B_v B v is bounded and self-adjoint on L 2 ( S n ) L^2(S^n) L 2 ( S n ) . Consequently, L v L_v L v is self-adjoint on L 2 ( S n ) L^2(S^n) L 2 ( S n ) with dom ( L v ) = H 2 ( S n ) \operatorname{dom}(L_v)=H^2(S^n) dom ( L v ) = H 2 ( S n ) [24 Chapter V, §4, Theorem 4.3] .
The operator L v L_v L v also has compact resolvent. For a sufficiently large c > 0 c>0 c > 0 , the resolvent identity is
( L v + c ) − 1 = ( A + c ) − 1 ( I + B v ( A + c ) − 1 ) − 1 . (L_v+c)^{-1}=(A+c)^{-1}(I+B_v(A+c)^{-1})^{-1}. ( L v + c ) − 1 = ( A + c ) − 1 ( I + B v ( A + c ) − 1 ) − 1 .
The first factor is compact on L 2 ( S n ) L^2(S^n) L 2 ( S n ) , since A A A is elliptic on the compact sphere. The second factor is bounded when c c c is chosen so that ∥ B v ( A + c ) − 1 ∥ < 1 \lVert B_v(A+c)^{-1}\rVert<1 ∥ B v ( A + c ) − 1 ∥ < 1 . The product is therefore compact.
The closed bilinear form associated with L v L_v L v [24 Chapter VI, §1.6 and §2.1] has domain H 1 ( S n ) H^1(S^n) H 1 ( S n ) and is equal to E ′ ′ ( v ) E''(v) E ′′ ( v ) . More precisely, (7.4) and (7.9) give
L v = A T v : H 1 ( S n ) ⟶ H − 1 ( S n ) (61) L_v=AT_v:H^1(S^n)\longrightarrow H^{-1}(S^n) \tag*{(61)} L v = A T v : H 1 ( S n ) ⟶ H − 1 ( S n ) (61)
in the weak sense.
If T v ϕ = 0 T_v\phi=0 T v ϕ = 0 , then (7.10) gives L v ϕ = 0 L_v\phi=0 L v ϕ = 0 in H − 1 ( S n ) H^{-1}(S^n) H − 1 ( S n ) . Equivalently,
A ϕ = ( p − 1 ) V v ϕ + p a v J μ ( a v ϕ ) . A\phi=(p-1)V_v\phi+pa_vJ_\mu(a_v\phi). A ϕ = ( p − 1 ) V v ϕ + p a v J μ ( a v ϕ ) .
The right-hand side belongs to L 2 ( S n ) L^2(S^n) L 2 ( S n ) , so elliptic regularity [21 Section 9.5] yields ϕ ∈ H 2 ( S n ) \phi\in H^2(S^n) ϕ ∈ H 2 ( S n ) . Therefore, ϕ ∈ ker L v \phi\in\ker L_v ϕ ∈ ker L v . Conversely, if ϕ ∈ ker L v \phi\in\ker L_v ϕ ∈ ker L v , then
E ′ ′ ( v ) [ ϕ , ψ ] = 0 , ψ ∈ H 1 ( S n ) , E''(v)[\phi,\psi]=0,\qquad\psi\in H^1(S^n), E ′′ ( v ) [ ϕ , ψ ] = 0 , ψ ∈ H 1 ( S n ) ,
and (7.4) gives T v ϕ = 0 T_v\phi=0 T v ϕ = 0 . Hence,
ker T v = ker L v ⊂ H 2 ( S n ) . \ker T_v=\ker L_v\subset H^2(S^n). ker T v = ker L v ⊂ H 2 ( S n ) .
The common Hessian form gives the same Morse index for both realizations. The kernel identity gives the same nullity. Their numerical eigenvalues generally differ because the inner products differ. We use T v T_v T v for the Fredholm argument in the energy space and L v L_v L v for the spectral analysis in L 2 ( S n ) L^2(S^n) L 2 ( S n ) .
The nodal estimate below relies on the strict positivity of J μ J_\mu J μ .
Lemma 7.3. For every nonzero real-valued function h ∈ L 2 ( S n ) h \in L^2(S^n) h ∈ L 2 ( S n ) ,
∬ S n × S n h ( ξ ) h ( η ) ∣ ξ − η ∣ μ d V ξ d V η > 0. (62) \iint_{S^n \times S^n} \frac{h(\xi)h(\eta)}{|\xi-\eta|^\mu}\,dV_\xi\,dV_\eta> 0. \tag*{(62)} ∬ S n × S n ∣ ξ − η ∣ μ h ( ξ ) h ( η ) d V ξ d V η > 0. (62) Proof. Let h = ∑ m = 0 ∞ h m h=\sum_{m=0}^{\infty}h_m h = ∑ m = 0 ∞ h m be the orthogonal decomposition of h h h into spherical harmonics, where h m h_m h m belongs to the space of spherical harmonics of degree m m m . The Funk–Hecke formula gives
J μ h m = κ m h m , κ m = 2 n − μ π n / 2 Γ ( ( n − μ ) / 2 ) Γ ( μ / 2 ) Γ ( m + μ / 2 ) Γ ( m + n − μ / 2 ) . (63) J_\mu h_m=\kappa_m h_m,\qquad\kappa_m=2^{n-\mu}\pi^{n/2}\frac{\Gamma((n-\mu)/2)}{\Gamma(\mu/2)}\frac{\Gamma(m+\mu/2)}{\Gamma(m+n-\mu/2)}. \tag*{(63)} J μ h m = κ m h m , κ m = 2 n − μ π n /2 Γ ( μ /2 ) Γ (( n − μ ) /2 ) Γ ( m + n − μ /2 ) Γ ( m + μ /2 ) . (63) For the normalization in (63) , use [17 Corollary 4.3] in dimension n n n and with α = μ / 2 \alpha=\mu/2 α = μ /2 . Since ∣ ξ − η ∣ − μ = 2 − μ / 2 ( 1 − ξ ⋅ η ) − μ / 2 |\xi-\eta|^{-\mu}=2^{-\mu/2}(1-\xi\cdot\eta)^{-\mu/2} ∣ ξ − η ∣ − μ = 2 − μ /2 ( 1 − ξ ⋅ η ) − μ /2 , the multiplier there must be multiplied by 2 − μ / 2 2^{-\mu/2} 2 − μ /2 . For nonintegral α = μ / 2 \alpha=\mu/2 α = μ /2 , the Gamma recurrence gives
( − 1 ) m Γ ( 1 − α ) Γ ( 1 − α − m ) = Γ ( m + α ) Γ ( α ) . (-1)^m\frac{\Gamma(1-\alpha)}{\Gamma(1-\alpha-m)}=\frac{\Gamma(m+\alpha)}{\Gamma(\alpha)}. ( − 1 ) m Γ ( 1 − α − m ) Γ ( 1 − α ) = Γ ( α ) Γ ( m + α ) . The degree-zero multiplier is fixed by direct integration,
κ 0 = ∣ S n − 1 ∣ ∫ − 1 1 ( 2 − 2 t ) − μ / 2 ( 1 − t 2 ) ( n − 2 ) / 2 d t = 2 n − μ − 1 ∣ S n − 1 ∣ ∫ 0 1 u ( n − μ ) / 2 − 1 ( 1 − u ) n / 2 − 1 d u = 2 n − μ π n / 2 Γ ( ( n − μ ) / 2 ) Γ ( n − μ / 2 ) . \begin{aligned}
\kappa_0&=|S^{n-1}|\int_{-1}^{1}(2-2t)^{-\mu/2}(1-t^2)^{(n-2)/2}\,dt\\
&=2^{n-\mu-1}|S^{n-1}|\int_0^1u^{(n-\mu)/2-1}(1-u)^{n/2-1}\,du\\
&=2^{n-\mu}\pi^{n/2}\frac{\Gamma((n-\mu)/2)}{\Gamma(n-\mu/2)}.
\end{aligned} κ 0 = ∣ S n − 1 ∣ ∫ − 1 1 ( 2 − 2 t ) − μ /2 ( 1 − t 2 ) ( n − 2 ) /2 d t = 2 n − μ − 1 ∣ S n − 1 ∣ ∫ 0 1 u ( n − μ ) /2 − 1 ( 1 − u ) n /2 − 1 d u = 2 n − μ π n /2 Γ ( n − μ /2 ) Γ (( n − μ ) /2 ) . Together with the multiplier ratio from the Funk–Hecke formula, this gives (63) . Integer values of μ / 2 \mu/2 μ /2 follow by continuity. The positivity of the Gamma function on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) gives κ m > 0 \kappa_m>0 κ m > 0 for every m ≥ 0 m\geq0 m ≥ 0 . Moreover, the Gamma quotient asymptotic [13 Section 5.11(iii)] applied to (63) yields
κ m = O ( m μ − n ) as m → ∞ . \kappa_m=O(m^{\mu-n})\qquad\text{as }m\to\infty. κ m = O ( m μ − n ) as m → ∞. The condition 0 < μ < n 0<\mu<n 0 < μ < n therefore implies
sup m ≥ 0 κ m < ∞ , κ m ⟶ 0. \sup_{m\geq0}\kappa_m<\infty,\qquad\kappa_m\longrightarrow0. m ≥ 0 sup κ m < ∞ , κ m ⟶ 0. For each integer M ≥ 0 M\geq0 M ≥ 0 , set h ( M ) = ∑ m = 0 M h m h^{(M)}=\sum_{m=0}^{M}h_m h ( M ) = ∑ m = 0 M h m . The orthogonality of distinct spherical harmonic subspaces and (63) give
∫ S n J μ ( h ( M ) ) h ( M ) d V = ∑ m = 0 M κ m ∥ h m ∥ L 2 ( S n ) 2 . (64) \int_{S^n}J_\mu(h^{(M)})h^{(M)}\,dV=\sum_{m=0}^{M}\kappa_m\lVert h_m\rVert_{L^2(S^n)}^2. \tag*{(64)} ∫ S n J μ ( h ( M ) ) h ( M ) d V = m = 0 ∑ M κ m ∥ h m ∥ L 2 ( S n ) 2 . (64) :::`
By (7.7), the operator J μ : L 2 ( S n ) → L 2 ( S n ) J_{\mu}:L^{2}(\mathbb{S}^{n})\to L^{2}(\mathbb{S}^{n}) J μ : L 2 ( S n ) → L 2 ( S n ) is bounded. Moreover, the boundedness of { κ m } m ≥ 0 \{\kappa_{m}\}_{m\ge0} { κ m } m ≥ 0 ensures that the series ∑ m = 0 ∞ κ m ∥ h m ∥ L 2 ( S n ) 2 \sum_{m=0}^{\infty}\kappa_{m}\lVert h_{m}\rVert_{L^{2}(\mathbb{S}^{n})}^{2} ∑ m = 0 ∞ κ m ∥ h m ∥ L 2 ( S n ) 2 converges. Since h ( M ) → h h^{(M)}\to h h ( M ) → h in L 2 ( S n ) L^{2}(\mathbb{S}^{n}) L 2 ( S n ) , passing to the limit in (7.13) yields
∬ S n × S n h ( ξ ) h ( η ) ∣ ξ − η ∣ μ d V ξ d V η = ∫ S n J μ ( h ) h d V = ∑ m = 0 ∞ κ m ∥ h m ∥ L 2 ( S n ) 2 . \iint_{\mathbb{S}^{n}\times\mathbb{S}^{n}}\frac{h(\xi)h(\eta)}{\lvert\xi-\eta\rvert^{\mu}}\,dV_{\xi}\,dV_{\eta}=\int_{\mathbb{S}^{n}}J_{\mu}(h)h\,dV=\sum_{m=0}^{\infty}\kappa_{m}\lVert h_{m}\rVert_{L^{2}(\mathbb{S}^{n})}^{2}. ∬ S n × S n ∣ ξ − η ∣ μ h ( ξ ) h ( η ) d V ξ d V η = ∫ S n J μ ( h ) h d V = m = 0 ∑ ∞ κ m ∥ h m ∥ L 2 ( S n ) 2 .
If h ≠ 0 h\ne0 h = 0 , the completeness of the spherical harmonic decomposition implies that h m ≠ 0 h_{m}\ne0 h m = 0 for at least one m m m . Together with κ m > 0 \kappa_{m}>0 κ m > 0 , this gives
∑ m = 0 ∞ κ m ∥ h m ∥ L 2 ( S n ) 2 > 0. \sum_{m=0}^{\infty}\kappa_{m}\lVert h_{m}\rVert_{L^{2}(\mathbb{S}^{n})}^{2}>0. m = 0 ∑ ∞ κ m ∥ h m ∥ L 2 ( S n ) 2 > 0.
which proves (7.11).
For the remainder of this section, fix a nonzero G G G -invariant weak solution v ∈ H v\in H v ∈ H . The profile has finitely many simple interior zeros and no endpoint zeros. The complementary orbit intervals have constant sign. Since both sphere factors are connected, their images in S n \mathbb{S}^{n} S n are connected and separated by the zero layers. They are therefore the nodal domains Ω 1 , … , Ω s \Omega_{1},\ldots,\Omega_{s} Ω 1 , … , Ω s , and each is G G G -invariant. Put
v i = v 1 Ω i , B i j = ∬ Ω i × Ω j ∣ v ( ξ ) ∣ p ∣ v ( η ) ∣ p ∣ ξ − η ∣ μ d V ξ d V η . (65) v_{i}=v\mathbf{1}_{\Omega_{i}},\qquad B_{ij}=\iint_{\Omega_{i}\times\Omega_{j}}\frac{\lvert v(\xi)\rvert^{p}\lvert v(\eta)\rvert^{p}}{\lvert\xi-\eta\rvert^{\mu}}\,dV_{\xi}dV_{\eta}. \tag*{(65)} v i = v 1 Ω i , B ij = ∬ Ω i × Ω j ∣ ξ − η ∣ μ ∣ v ( ξ ) ∣ p ∣ v ( η ) ∣ p d V ξ d V η . (65)
Let I i I_{i} I i be the corresponding interval on which f f f has a fixed sign. The next lemma proves that the functions in (7.14) belong to H H H .
Lemma 7.4. For every i i i , one has v i ∈ H v_{i}\in H v i ∈ H and
∇ v i = 1 Ω i ∇ v almost everywhere on S n . (66) \nabla v_{i}=\mathbf{1}_{\Omega_{i}}\nabla v\quad\text{almost everywhere on }\mathbb{S}^{n}. \tag*{(66)} ∇ v i = 1 Ω i ∇ v almost everywhere on S n . (66) If i ≠ j i\ne j i = j , then v i v_{i} v i and v j v_{j} v j are orthogonal in the energy scalar product.
Proof. Let I i = ( a i , b i ) I_{i}=(a_{i},b_{i}) I i = ( a i , b i ) be the orbit interval corresponding to Ω i \Omega_{i} Ω i , and set f i = f 1 I i f_{i}=f\mathbf{1}_{I_{i}} f i = f 1 I i . By Lemma 6.3,
f ( a i ) = 0 if a i > 0 , f ( b i ) = 0 if b i < π 2 . f(a_{i})=0\ \text{if }a_{i}>0,\qquad f(b_{i})=0\ \text{if }b_{i}<\frac{\pi}{2}. f ( a i ) = 0 if a i > 0 , f ( b i ) = 0 if b i < 2 π .
For every ζ ∈ C c ∞ ( 0 , π / 2 ) \zeta\in C_{c}^{\infty}(0,\pi/2) ζ ∈ C c ∞ ( 0 , π /2 ) , integration by parts on I i I_{i} I i gives
− ∫ 0 π / 2 f i ( t ) ζ ′ ( t ) d t = − ∫ a i b i f ( t ) ζ ′ ( t ) d t = ∫ a i b i f ′ ( t ) ζ ( t ) d t − [ f ( t ) ζ ( t ) ] a i b i = ∫ I i f ′ ( t ) ζ ( t ) d t . \begin{aligned}
-\int_{0}^{\pi/2}f_{i}(t)\zeta'(t)\,dt=-\int_{a_{i}}^{b_{i}}f(t)\zeta'(t)\,dt \\
&=\int_{a_{i}}^{b_{i}}f'(t)\zeta(t)\,dt-\left[f(t)\zeta(t)\right]_{a_{i}}^{b_{i}} \\
&=\int_{I_{i}}f'(t)\zeta(t)\,dt.
\end{aligned} − ∫ 0 π /2 f i ( t ) ζ ′ ( t ) d t = − ∫ a i b i f ( t ) ζ ′ ( t ) d t = ∫ a i b i f ′ ( t ) ζ ( t ) d t − [ f ( t ) ζ ( t ) ] a i b i = ∫ I i f ′ ( t ) ζ ( t ) d t .
The boundary term vanishes at every interior endpoint by Lemma 6.3. If a i = 0 a_{i}=0 a i = 0 or b i = π / 2 b_{i}=\pi/2 b i = π /2 , it vanishes from the compact support of ζ \zeta ζ . Therefore,
f i ′ = 1 I i f ′ in D ′ ( 0 , π / 2 ) . f_{i}'=\mathbf{1}_{I_{i}}f'\quad\text{in }\mathcal{D}'(0,\pi/2). f i ′ = 1 I i f ′ in D ′ ( 0 , π /2 ) .
It follows that
∫ 0 π / 2 ( ∣ f i ′ ( t ) ∣ 2 + ∣ f i ( t ) ∣ 2 ) ρ ( t ) d t ≤ ∫ 0 π / 2 ( ∣ f ′ ( t ) ∣ 2 + ∣ f ( t ) ∣ 2 ) ρ ( t ) d t < ∞ . \int_{0}^{\pi/2} \left(|f_i'(t)|^2+|f_i(t)|^2\right)\rho(t)\,\mathrm{d}t \le\int_{0}^{\pi/2} \left(|f'(t)|^2+|f(t)|^2\right)\rho(t)\,\mathrm{d}t < \infty. ∫ 0 π /2 ( ∣ f i ′ ( t ) ∣ 2 + ∣ f i ( t ) ∣ 2 ) ρ ( t ) d t ≤ ∫ 0 π /2 ( ∣ f ′ ( t ) ∣ 2 + ∣ f ( t ) ∣ 2 ) ρ ( t ) d t < ∞.
In any interior product chart, the weak derivative identity for f i f_i f i above and the absence of angular derivatives give v i ∈ H 1 ( U ) v_i \in H^1(U) v i ∈ H 1 ( U ) for every relatively compact interior product chart U U U . Near a singular orbit, v i v_i v i either agrees with v v v or vanishes identically, so it belongs to H 1 ( U ) H^1(U) H 1 ( U ) on each corresponding singular chart U U U as well. A finite covering and a subordinate partition of unity therefore yield
v i ∈ H 1 ( S n ) . v_i \in H^1(\mathbb{S}^n). v i ∈ H 1 ( S n ) .
Both v v v and Ω i \Omega_i Ω i are G G G -invariant, so v i v_i v i is G G G -invariant. Consequently,
The distributional identity for f i ′ f_i' f i ′ , together with the vanishing of all angular derivatives, gives (7.15). The singular orbits have zero n n n -dimensional measure and therefore do not affect the almost everywhere identity.
For i ≠ j i \ne j i = j , the nodal domains Ω i \Omega_i Ω i and Ω j \Omega_j Ω j are disjoint. Hence,
v i v j = 0 and ∇ S n v i ⋅ ∇ S n v j = 0 almost everywhere on S n , v_i v_j = 0 \quad\text{and} \quad\nabla_{\mathbb{S}^n}v_i \cdot\nabla_{\mathbb{S}^n}v_j = 0 \quad\text{almost everywhere on } \mathbb{S}^n, v i v j = 0 and ∇ S n v i ⋅ ∇ S n v j = 0 almost everywhere on S n ,
where the second identity follows from (7.15). Therefore,
⟨ v i , v j ⟩ A = ∫ S n ( ∇ S n v i ⋅ ∇ S n v j + n ( n − 2 ) 4 v i v j ) d V = 0. \langle v_i,v_j\rangle_A=\int_{\mathbb{S}^n}\left(\nabla_{\mathbb{S}^n}v_i\cdot\nabla_{\mathbb{S}^n}v_j+\frac{n(n-2)}{4}v_i v_j\right)\,\mathrm{d}V=0. ⟨ v i , v j ⟩ A = ∫ S n ( ∇ S n v i ⋅ ∇ S n v j + 4 n ( n − 2 ) v i v j ) d V = 0.
Thus, v i v_i v i and v j v_j v j are orthogonal in the energy inner product.
Lemma 7.5. Assume p ≥ 2 p \ge2 p ≥ 2 . Then Q v Q_v Q v is negative definite on
V = span { v 1 , … , v s } . \mathcal{V}=\operatorname{span}\{v_1,\ldots,v_s\}. V = span { v 1 , … , v s } .
Proof. Lemma 7.4 gives v i ∈ H v_i \in H v i ∈ H , so v i v_i v i is an admissible test function in the weak formulation of (1.1). Moreover, the orthogonality established in Lemma 7.4 gives
⟨ v , v i ⟩ A = ⟨ ∑ j = 1 s v j , v i ⟩ A = ∥ v i ∥ 2 . \langle v,v_i\rangle_A=\left\langle\sum_{j=1}^{s}v_j,v_i\right\rangle_A=\|v_i\|^2. ⟨ v , v i ⟩ A = ⟨ j = 1 ∑ s v j , v i ⟩ A = ∥ v i ∥ 2 .
Testing (1.1) with v i v_i v i therefore yields
∥ v i ∥ 2 = ∫ S n J μ ( ∣ v ∣ p ) ∣ v ∣ p − 2 v v i d V = ∫ Ω i J μ ( ∣ v ∣ p ) ∣ v ∣ p d V = ∑ j = 1 s ∬ Ω i × Ω j ∣ v ( ξ ) ∣ p ∣ v ( η ) ∣ p ∣ ξ − η ∣ μ d V ξ d V η = ∑ j = 1 s B i j , (67) \begin{aligned}
\|v_i\|^2
&=\int_{\mathbb{S}^n}J_\mu(|v|^p)|v|^{p-2}vv_i\,\mathrm{d}V\\
&=\int_{\Omega_i}J_\mu(|v|^p)|v|^p\,\mathrm{d}V\\
&=\sum_{j=1}^{s}\iint_{\Omega_i\times\Omega_j}\frac{|v(\xi)|^p|v(\eta)|^p}{|\xi-\eta|^\mu}\,\mathrm{d}V_\xi\,\mathrm{d}V_\eta\\
&=\sum_{j=1}^{s}B_{ij}, \tag*{(67)}
\end{aligned} ∥ v i ∥ 2 = ∫ S n J μ ( ∣ v ∣ p ) ∣ v ∣ p − 2 v v i d V = ∫ Ω i J μ ( ∣ v ∣ p ) ∣ v ∣ p d V = j = 1 ∑ s ∬ Ω i × Ω j ∣ ξ − η ∣ μ ∣ v ( ξ ) ∣ p ∣ v ( η ) ∣ p d V ξ d V η = j = 1 ∑ s B ij , (67)
where B i j B_{ij} B ij is defined in (65) . Let ϕ = ∑ i = 1 s c i v i ∈ V \phi=\sum_{i=1}^{s}c_i v_i\in\mathcal{V} ϕ = ∑ i = 1 s c i v i ∈ V , where c 1 , … , c s ∈ R c_1,\ldots,c_s\in\mathbb{R} c 1 , … , c s ∈ R . The orthogonality of the functions v i v_i v i gives
∥ ϕ ∥ 2 = ∑ i = 1 s c i 2 ∥ v i ∥ 2 . \|\phi\|^2=\sum_{i=1}^{s}c_i^2\|v_i\|^2. ∥ ϕ ∥ 2 = i = 1 ∑ s c i 2 ∥ v i ∥ 2 .
Furthermore,
∫ S n J μ ( ∣ v ∣ ) ∣ v ∣ p − 2 ϕ 2 d V = ∑ i = 1 s c i 2 ∫ Ω i J μ ( ∣ v ∣ ) ∣ v ∣ p d V = ∑ i = 1 s c i 2 ∥ v i ∥ 2 , \int_{\mathbb{S}^n}J_\mu(|v|)|v|^{p-2}\phi^2\,dV=\sum_{i=1}^{s}c_i^2\int_{\Omega_i}J_\mu(|v|)|v|^p\,dV=\sum_{i=1}^{s}c_i^2\|v_i\|^2, ∫ S n J μ ( ∣ v ∣ ) ∣ v ∣ p − 2 ϕ 2 d V = i = 1 ∑ s c i 2 ∫ Ω i J μ ( ∣ v ∣ ) ∣ v ∣ p d V = i = 1 ∑ s c i 2 ∥ v i ∥ 2 ,
where (67) was used in the last equality. On Ω i \Omega_i Ω i , ∣ v ∣ p − 2 v ϕ = c i ∣ v ∣ p |v|^{p-2}v\phi=c_i|v|^p ∣ v ∣ p − 2 v ϕ = c i ∣ v ∣ p . Consequently,
∫ S n J μ ( ∣ v ∣ ) ∣ v ∣ p − 2 v ϕ d V = ∑ i , j = 1 s c i c j B i j . \int_{\mathbb{S}^n}J_\mu(|v|)|v|^{p-2}v\phi\,dV=\sum_{i,j=1}^{s}c_i c_jB_{ij}. ∫ S n J μ ( ∣ v ∣ ) ∣ v ∣ p − 2 v ϕ d V = i , j = 1 ∑ s c i c j B ij .
Substituting these identities into (54) gives
Q v ( ϕ ) = ( 2 − p ) ∑ i = 1 s c i 2 ∥ v i ∥ 2 − p ∑ i , j = 1 s c i c j B i j . (68) Q_v(\phi)=(2-p)\sum_{i=1}^{s}c_i^2\|v_i\|^2-p\sum_{i,j=1}^{s}c_i c_jB_{ij}. \tag*{(68)} Q v ( ϕ ) = ( 2 − p ) i = 1 ∑ s c i 2 ∥ v i ∥ 2 − p i , j = 1 ∑ s c i c j B ij . (68)
Define h = ∑ i = 1 s c i ∣ v i ∣ p h=\sum_{i=1}^{s}c_i|v_i|^p h = ∑ i = 1 s c i ∣ v i ∣ p . Lemma 6.1 gives v ∈ L ∞ ( S n ) v\in L^\infty(\mathbb{S}^n) v ∈ L ∞ ( S n ) , so h ∈ L 2 ( S n ) h\in L^2(\mathbb{S}^n) h ∈ L 2 ( S n ) . The nonzero sets of ∣ v 1 ∣ p , … , ∣ v s ∣ p |v_1|^p,\ldots,|v_s|^p ∣ v 1 ∣ p , … , ∣ v s ∣ p are pairwise disjoint, and each has positive measure. Hence, h ≠ 0 h\ne0 h = 0 whenever ( c 1 , … , c s ) ≠ 0 (c_1,\ldots,c_s)\ne0 ( c 1 , … , c s ) = 0 . By the definition of B i j B_{ij} B ij and Lemma 7.3, we obtain
∬ S n × S n h ( ξ ) h ( η ) ∣ ξ − η ∣ μ d V ξ d V η = ∑ i , j = 1 s c i c j B i j > 0 for ( c 1 , … , c s ) ≠ 0. \iint_{\mathbb{S}^n\times\mathbb{S}^n}\frac{h(\xi)h(\eta)}{|\xi-\eta|^\mu}\,dV_\xi\,dV_\eta=\sum_{i,j=1}^{s}c_i c_jB_{ij}>0\qquad\text{for }(c_1,\ldots,c_s)\ne0. ∬ S n × S n ∣ ξ − η ∣ μ h ( ξ ) h ( η ) d V ξ d V η = i , j = 1 ∑ s c i c j B ij > 0 for ( c 1 , … , c s ) = 0.
Since p ≥ 2 p\ge2 p ≥ 2 , the first term in (68) is nonpositive. The strict positivity just proved makes its second term negative. Thus
Q v ( ϕ ) < 0 for every ϕ ∈ V ∖ { 0 } . Q_v(\phi)<0\qquad\text{for every }\phi\in\mathcal{V}\setminus\{0\}. Q v ( ϕ ) < 0 for every ϕ ∈ V ∖ { 0 } .
Thus, Q v Q_v Q v is negative definite on V \mathcal{V} V . Moreover, the functions v 1 , … , v s v_1,\ldots,v_s v 1 , … , v s are linearly independent because they have pairwise disjoint nonzero supports. Hence dim V = s \dim\mathcal{V}=s dim V = s .
Proof of Theorem 1.3. Lemma 6.1 gives v ∈ C 2 , γ ( S n ) v\in C^{2,\gamma}(\mathbb{S}^n) v ∈ C 2 , γ ( S n ) for some γ ∈ ( 0 , 1 ) \gamma\in(0,1) γ ∈ ( 0 , 1 ) . Lemma 6.3 shows that the orbit profile has finitely many simple interior zeros and no endpoint zeros. By Lemma 6.3, either f f f has no zeros, in which case q = 0 q=0 q = 0 , or its zeros satisfy 0 < t 1 < ⋯ < t q < π / 2 0<t_1<\cdots<t_q<\pi/2 0 < t 1 < ⋯ < t q < π /2 . For q ≥ 1 q\ge1 q ≥ 1 , the corresponding level sets are described below. The orbit coordinates in (29) give
{ ξ ∈ S n : t ( ξ ) = t r } = S k − 1 ( cos t r ) × S ℓ − 1 ( sin t r ) . \{\xi\in\mathbb{S}^n:t(\xi)=t_r\}=\mathbb{S}^{k-1}(\cos t_r)\times\mathbb{S}^{\ell-1}(\sin t_r). { ξ ∈ S n : t ( ξ ) = t r } = S k − 1 ( cos t r ) × S ℓ − 1 ( sin t r ) .
Since v ( ξ ) = f ( t ( ξ ) ) v(\xi)=f(t(\xi)) v ( ξ ) = f ( t ( ξ )) , it follows that
{ v = 0 } = ⋃ r = 1 q S k − 1 ( cos t r ) × S ℓ − 1 ( sin t r ) , \{v=0\}=\bigcup_{r=1}^{q}\mathbb{S}^{k-1}(\cos t_r)\times\mathbb{S}^{\ell-1}(\sin t_r), { v = 0 } = r = 1 ⋃ q S k − 1 ( cos t r ) × S ℓ − 1 ( sin t r ) ,
which proves (13) .
A simple zero changes the sign of f f f . Hence, f f f has a constant sign on each of the q + 1 q+1 q + 1 complementary intervals, and the signs on adjacent intervals are opposite. For q ≥ 1 q \ge1 q ≥ 1 , denote these intervals by
I 0 = [ 0 , t 1 ) , I r = ( t r , t r + 1 ) ( 1 ≤ r ≤ q − 1 ) , I q = ( t q , π / 2 ] . I_0=[0,t_1), \qquad I_r=(t_r,t_{r+1})\quad(1\le r\le q-1), \qquad I_q=(t_q,\pi/2]. I 0 = [ 0 , t 1 ) , I r = ( t r , t r + 1 ) ( 1 ≤ r ≤ q − 1 ) , I q = ( t q , π /2 ] .
If q = 0 q=0 q = 0 , set I 0 = [ 0 , π / 2 ] I_0=[0,\pi/2] I 0 = [ 0 , π /2 ] . For each r r r , the subset of S n \mathbb{S}^n S n corresponding to t ∈ I r t\in I_r t ∈ I r is the continuous image of
I r × S k − 1 × S ℓ − 1 I_r\times\mathbb{S}^{k-1}\times\mathbb{S}^{\ell-1} I r × S k − 1 × S ℓ − 1
under the map
( t , ω , θ ) ⟼ ( cos t ω , sin t θ ) . (t,\omega,\theta)\longmapsto(\cos t\,\omega,\sin t\,\theta). ( t , ω , θ ) ⟼ ( cos t ω , sin t θ ) .
The interval I r I_r I r is connected, and S k − 1 \mathbb{S}^{k-1} S k − 1 and S ℓ − 1 \mathbb{S}^{\ell-1} S ℓ − 1 are connected under the assumption k , ℓ ≥ 2 k,\ell\ge2 k , ℓ ≥ 2 . Therefore, each corresponding subset of { v ≠ 0 } \{v\ne0\} { v = 0 } is connected. Distinct subsets are separated by the nodal layers t = t r t=t_r t = t r , and their union equals { v ≠ 0 } \{v\ne0\} { v = 0 } . Consequently, these subsets are precisely the nodal domains of v v v , and
N G ( v ) = q + 1. N_G(v)=q+1. N G ( v ) = q + 1.
For the Morse-index estimate, note that v 1 , … , v s v_1,\ldots,v_s v 1 , … , v s are nonzero and pairwise orthogonal in the energy inner product by Lemma 7.4 . They are therefore linearly independent, and dim V = s \dim\mathcal{V}=s dim V = s . By Lemma 7.5 , Q v Q_v Q v is negative definite on the s s s -dimensional subspace V ⊂ H \mathcal{V}\subset H V ⊂ H . The definition of the restricted Morse index therefore gives
s ≤ m G ( v ) . s\le m_G(v). s ≤ m G ( v ) .
Since s = N G ( v ) = q + 1 s=N_G(v)=q+1 s = N G ( v ) = q + 1 ,
q + 1 ≤ m G ( v ) , q+1\le m_G(v), q + 1 ≤ m G ( v ) ,
which proves the index inequality in (14) .
We now prove the minimax energy comparison. Put
A i = ∥ v i ∥ 2 , B = ( B i j ) 1 ≤ i , j ≤ s , 1 = ( 1 , … , 1 ) T . A_i=\lVert v_i\rVert^2,\qquad B=(B_{ij})_{1\le i,j\le s},\qquad\mathbf{1}=(1,\ldots,1)^T. A i = ∥ v i ∥ 2 , B = ( B ij ) 1 ≤ i , j ≤ s , 1 = ( 1 , … , 1 ) T .
Equation (67) states that A i = ( B 1 ) i A_i=(B\mathbf{1})_i A i = ( B 1 ) i . The proof of Lemma 7.5 also gives
y T B y > 0 for every y ∈ R s ∖ { 0 } . y^TBy>0\qquad\text{for every }y\in\mathbb{R}^s\setminus\{0\}. y T B y > 0 for every y ∈ R s ∖ { 0 } .
Hence λ min ( B ) > 0 \lambda_{\min}(B)>0 λ m i n ( B ) > 0 . For t = ( t 1 , … , t s ) ∈ R s t=(t_1,\ldots,t_s)\in\mathbb{R}^s t = ( t 1 , … , t s ) ∈ R s , set x i = ∣ t i ∣ p x_i=\lvert t_i\rvert^p x i = ∣ t i ∣ p . The disjoint supports give
∣ ∑ i t i v i ∣ p = ∑ i x i ∣ v i ∣ p , ∥ ∑ i t i v i ∥ 2 = ∑ i A i x i 2 / p . \left\lvert\sum_i t_iv_i\right\rvert^p=\sum_i x_i\lvert v_i\rvert^p,\qquad
\left\lVert\sum_i t_iv_i\right\rVert^2=\sum_i A_ix_i^{2/p}. i ∑ t i v i p = i ∑ x i ∣ v i ∣ p , i ∑ t i v i 2 = i ∑ A i x i 2/ p .
Therefore,
E ( ∑ i t i v i ) = 1 2 ∑ i A i x i 2 / p − 1 2 p x T B x . E\left(\sum_i t_iv_i\right)=\frac{1}{2}\sum_i A_ix_i^{2/p}-\frac{1}{2p}x^TBx. E ( i ∑ t i v i ) = 2 1 i ∑ A i x i 2/ p − 2 p 1 x T B x .
Since 0 < 2 / p ≤ 1 0<2/p\le1 0 < 2/ p ≤ 1 , concavity on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) gives
x i 2 / p − 1 ≤ 2 p ( x i − 1 ) . x_i^{2/p}-1\le\frac{2}{p}(x_i-1). x i 2/ p − 1 ≤ p 2 ( x i − 1 ) .
It follows that
E ( ∑ i t i v i ) − E ( v ) ≤ 1 p ∑ i A i ( x i − 1 ) − 1 2 p ( x T B x − 1 T B 1 ) = 1 p 1 T B ( x − 1 ) − 1 2 p ( x T B x − 1 T B 1 ) . E\left(\sum_i t_i v_i\right)-E(v) \le\frac{1}{p}\sum_i A_i(x_i-1)-\frac{1}{2p}\left(x^T Bx-1^T B1\right)
= \frac{1}{p}1^T B(x-1)-\frac{1}{2p}\left(x^T Bx-1^T B1\right). E ( i ∑ t i v i ) − E ( v ) ≤ p 1 i ∑ A i ( x i − 1 ) − 2 p 1 ( x T B x − 1 T B 1 ) = p 1 1 T B ( x − 1 ) − 2 p 1 ( x T B x − 1 T B 1 ) .
Completing the quadratic form yields
E ( ∑ i t i v i ) − E ( v ) ≤ − 1 2 p ( x − 1 ) T B ( x − 1 ) . (69) E\left(\sum_i t_i v_i\right)-E(v) \le-\frac{1}{2p}(x-1)^T B(x-1). \tag*{(69)} E ( i ∑ t i v i ) − E ( v ) ≤ − 2 p 1 ( x − 1 ) T B ( x − 1 ) . (69)
This inequality includes p = 2 p=2 p = 2 ; there concavity is an equality.
The energy unit sphere V ∩ S \mathcal{V}\cap\mathcal{S} V ∩ S is compact and has cohomological index s s s . For any w w w on this sphere, r w ∈ V rw\in\mathcal{V} r w ∈ V for all r > 0 r>0 r > 0 . By the radial maximization formula and (69) ,
Ψ ( w ) = max r > 0 E ( r w ) ≤ E ( v ) . \Psi(w)=\max_{r>0}E(rw)\le E(v). Ψ ( w ) = r > 0 max E ( r w ) ≤ E ( v ) .
Taking the supremum on this sphere and then the minimax infimum gives
c s ≤ sup V ∩ S Ψ ≤ E ( v ) . c_s\le\sup_{\mathcal{V}\cap\mathcal{S}}\Psi\le E(v). c s ≤ V ∩ S sup Ψ ≤ E ( v ) .
If 1 ≤ j < s 1\le j<s 1 ≤ j < s , choose j j j of the nodal pieces and let V j \mathcal{V}_j V j be their span. At least one coordinate of x x x vanishes for every vector in V j \mathcal{V}_j V j . Thus
∣ x − 1 ∣ 2 ≥ 1 , E ( ∑ i t i v i ) ≤ E ( v ) − λ min ( B ) 2 p on V j . |x-1|^2\ge1,\qquad E\left(\sum_i t_i v_i\right)\le E(v)-\frac{\lambda_{\min}(B)}{2p}\quad\text{on }\mathcal{V}_j. ∣ x − 1 ∣ 2 ≥ 1 , E ( i ∑ t i v i ) ≤ E ( v ) − 2 p λ m i n ( B ) on V j .
The sphere V j ∩ S \mathcal{V}_j\cap\mathcal{S} V j ∩ S has index j j j , so
c j ≤ sup V j ∩ S Ψ ≤ E ( v ) − λ min ( B ) 2 p < E ( v ) . c_j\le\sup_{\mathcal{V}_j\cap\mathcal{S}}\Psi\le E(v)-\frac{\lambda_{\min}(B)}{2p}<E(v). c j ≤ V j ∩ S sup Ψ ≤ E ( v ) − 2 p λ m i n ( B ) < E ( v ) .
In particular, if E ( v ) = c j E(v)=c_j E ( v ) = c j , then s > j s>j s > j would contradict this strict inequality. Hence N G ( v ) = s ≤ j N_G(v)=s\le j N G ( v ) = s ≤ j . No isolation of the critical point or critical level is used here.
Every sign-changing invariant solution has at least two nodal domains. For such a solution, monotonicity of the minimax levels gives
c 2 ≤ c s ≤ E ( v ) . c_2\le c_s\le E(v). c 2 ≤ c s ≤ E ( v ) .
Theorem 1.1 gives a sign-changing solution at c 2 c_2 c 2 . Every solution at that level satisfies 2 ≤ N G ( v ) ≤ 2 2\le N_G(v)\le2 2 ≤ N G ( v ) ≤ 2 . Thus c 2 c_2 c 2 is the least energy of sign-changing invariant solutions, and all invariant solutions at c 2 c_2 c 2 have exactly two nodal domains.
Remark 7.6. For 1 < p < 2 1<p<2 1 < p < 2 , the coefficient ∣ v ∣ p − 2 |v|^{p-2} ∣ v ∣ p − 2 can be singular at nodal zeros. The second-variation argument here applies for p ≥ 2 p\ge2 p ≥ 2 . The minimax bounds and sign change at every level c j c_j c j , j ≥ 2 j\ge2 j ≥ 2 , apply throughout the parameter range of Theorem 1.1.
Cohomological minimax labels and the Morse index
This section proves Theorem 1.4 . The index i ( B ) i(B) i ( B ) is defined using Alexander–Spanier cohomology with coefficients in Z 2 \mathbb{Z}_2 Z 2 and the characteristic class of B → B / Z 2 B \to B/\mathbb{Z}_2 B → B / Z 2 [15 ] . For local critical groups we use singular cohomology with the same coefficients. The proof below passes between these two theories on regular sublevel sets and finite complexes.
Lemma 8.1. Let F ∈ C 2 ( S ) F \in C^2(\mathcal{S}) F ∈ C 2 ( S ) be even and satisfy the Palais–Smale condition, and let d j d_j d j be defined as in Lemma 5.1. Assume d j ∈ R d_j \in\mathbb{R} d j ∈ R . Suppose that every critical point at d j d_j d j is isolated in the entire critical set of F F F . Assume that each constrained Hessian at a critical point of level d j d_j d j is represented by a self-adjoint Fredholm operator with finite Morse index. Then there exists a critical point w w w of F F F at level d j d_j d j such that
m ~ ( w ) ≤ j − 1 ≤ m ~ ( w ) + ν ~ ( w ) , (70) \widetilde{m}(w) \le j-1 \le\widetilde{m}(w)+\widetilde{\nu}(w), \tag*{(70)} m ( w ) ≤ j − 1 ≤ m ( w ) + ν ( w ) , (70) where m ~ ( w ) \widetilde{m}(w) m ( w ) and ν ~ ( w ) \widetilde{\nu}(w) ν ( w ) are the Morse index and nullity of the constrained Hessian of F F F on T w S T_w\mathcal{S} T w S .
Proof. Put c = d j c=d_j c = d j , and write K c = { F = c , d F = 0 } K_c=\{F=c,\ dF=0\} K c = { F = c , d F = 0 } . Every sequence in K c K_c K c is a Palais–Smale sequence, so K c K_c K c is compact. If it contained infinitely many distinct points, a convergent subsequence would contradict isolation at its limit. Thus K c K_c K c is a finite union of antipodal pairs. We next show that c c c is an isolated critical value. Otherwise there would be critical points w m w_m w m with
F ( w m ) ≠ c , F ( w m ) ⟶ c , d F ( w m ) = 0. F(w_m) \ne c,\qquad F(w_m) \longrightarrow c,\qquad dF(w_m)=0. F ( w m ) = c , F ( w m ) ⟶ c , d F ( w m ) = 0.
Palais–Smale gives a subsequence converging to w ∈ K c w \in K_c w ∈ K c . The points w m w_m w m are different from w w w , contradicting isolation in the full critical set. This proves isolation of the level.
Set M = S / Z 2 M=\mathcal{S}/\mathbb{Z}_2 M = S / Z 2 , let F ‾ : M → R \overline{F}:M\to\mathbb{R} F : M → R be induced by F F F , and define M a = { F ‾ ≤ a } M^a=\{\overline{F}\le a\} M a = { F ≤ a } , F a = { F ≤ a } F^a=\{F\le a\} F a = { F ≤ a } . Let ω ∈ H 1 ( M ; Z 2 ) \omega\in H^1(M;\mathbb{Z}_2) ω ∈ H 1 ( M ; Z 2 ) be the Alexander–Spanier characteristic class of the double cover S → M \mathcal{S}\to M S → M . For every regular value a a a , we have
ω j − 1 ∣ M a ≠ 0 ⟺ there exists B ∈ Γ j such that B ⊂ F a . (71) \left.\omega^{j-1}\right|_{M^a}\ne0
\quad\Longleftrightarrow\quad
\text{there exists }B\in\Gamma_j\text{ such that }B\subset F^a. \tag*{(71)} ω j − 1 M a = 0 ⟺ there exists B ∈ Γ j such that B ⊂ F a . (71)
Indeed, if B ∈ Γ j B\in\Gamma_j B ∈ Γ j and B ⊂ F a B\subset F^a B ⊂ F a , naturality of the characteristic class gives
ω B j − 1 = ι B ∗ ( ω j − 1 ∣ M a ) ≠ 0 , \omega_B^{j-1}=\iota_B^*\left(\left.\omega^{j-1}\right|_{M^a}\right)\ne0, ω B j − 1 = ι B ∗ ( ω j − 1 M a ) = 0 ,
where ι B : B / Z 2 → M a \iota_B:B/\mathbb{Z}_2\to M^a ι B : B / Z 2 → M a is induced by inclusion. This proves the implication from right to left. Conversely, suppose that α : = ω j − 1 ∣ M a ≠ 0 \alpha:=\left.\omega^{j-1}\right|_{M^a}\ne0 α := ω j − 1 M a = 0 . The regular sublevel M a M^a M a is a metrizable, locally contractible Hilbert manifold with boundary, or is empty. Hence Alexander–Spanier and singular cohomology agree naturally on M a M^a M a ; see [31 Chapter 6, Section 9, Theorem 1] . Over Z 2 \mathbb{Z}_2 Z 2 , the universal coefficient theorem yields a homology class on which the singular class corresponding to α \alpha α evaluates nontrivially. By the definition of singular homology, this homology class is represented by a finite singular cycle. Gluing its finitely many simplices along cancelling faces gives a finite Δ \Delta Δ -complex X X X and a map f : X → M a f:X\to M^a f : X → M a such that f ∗ α ≠ 0 f^*\alpha\ne0 f ∗ α = 0 . Pull back the double cover F a → M a F^a\to M^a F a → M a along f f f , and let X ~ → X \widetilde{X}\to X X → X be the resulting compact double cover. Its natural map g : X ~ → F a g:\widetilde{X}\to F^a g : X → F a is equivariant. Therefore, B : = g ( X ~ ) B:=g(\widetilde{X}) B := g ( X ) is a compact symmetric subset of F a F^a F a . If ω ~ \widetilde{\omega} ω denotes the characteristic class of the pullback cover X ~ → X \widetilde{X}\to X X → X , then naturality first gives ω ~ j − 1 = f ∗ α \widetilde{\omega}^{j-1}=f^*\alpha ω j − 1 = f ∗ α . Applying naturality again to the equivariant map g g g gives f ∗ α = g ‾ ∗ ( ω B j − 1 ) ≠ 0 f^*\alpha=\overline{g}^*(\omega_B^{j-1})\ne0 f ∗ α = g ∗ ( ω B j − 1 ) = 0 , where g ‾ : X → B / Z 2 \overline{g}:X\to B/\mathbb{Z}_2 g : X → B / Z 2 is induced by g g g . Thus ω B j − 1 ≠ 0 \omega_B^{j-1}\ne0 ω B j − 1 = 0 , so i ( B ) ≥ j i(B)\ge j i ( B ) ≥ j and B ∈ Γ j B\in\Gamma_j B ∈ Γ j . This proves (8.2).
Since c c c is isolated, choose ε > 0 \varepsilon>0 ε > 0 so that c c c is the only critical value in [ c − 2 ε , c + 2 ε ] [c-2\varepsilon,c+2\varepsilon] [ c − 2 ε , c + 2 ε ] and c ± ε c\pm\varepsilon c ± ε are regular values. The definition of d j d_j d j and (71) give
ω j − 1 ∣ M c + ε ≠ 0 , ω j − 1 ∣ M c − ε = 0. \left.\omega^{j-1}\right|_{M^{c+\varepsilon}}\ne0,\qquad\left.\omega^{j-1}\right|_{M^{c-\varepsilon}}=0. ω j − 1 M c + ε = 0 , ω j − 1 M c − ε = 0.
The exact cohomology sequence of the pair therefore yields
H j − 1 ( M c + ε , M c − ε ; Z 2 ) ≠ 0. (72) H^{j-1}\left(M^{c+\varepsilon},M^{c-\varepsilon};\mathbb{Z}_2\right)\ne0. \tag*{(72)} H j − 1 ( M c + ε , M c − ε ; Z 2 ) = 0. (72)
Let z 1 , … , z m z_1,\ldots,z_m z 1 , … , z m be the critical points of F ‾ \overline{F} F at level c c c . Choose pairwise disjoint isolating neighborhoods U r U_r U r of z r z_r z r and set
C q ( F ‾ , z r ) = H q ( M c ∩ U r , ( M c ∩ U r ) ∖ { z r } ; Z 2 ) . C^q(\overline{F},z_r)=H^q\left(M^c\cap U_r,(M^c\cap U_r)\setminus\{z_r\};\mathbb{Z}_2\right). C q ( F , z r ) = H q ( M c ∩ U r , ( M c ∩ U r ) ∖ { z r } ; Z 2 ) .
The Palais–Smale condition, the isolation of these points and the second deformation lemma localize the relative group to their local critical groups:
H q ( M c + ε , M c − ε ; Z 2 ) ≅ ⨁ r = 1 m C q ( F ‾ , z r ) . (73) H^q\left(M^{c+\varepsilon},M^{c-\varepsilon};\mathbb{Z}_2\right)\cong\bigoplus_{r=1}^{m}C^q(\overline{F},z_r). \tag*{(73)} H q ( M c + ε , M c − ε ; Z 2 ) ≅ r = 1 ⨁ m C q ( F , z r ) . (73)
This is the critical-group decomposition from the second deformation lemma and excision [7 Chapter I, Section 4] ; see also [8 Theorem III.3] for the local Gromoll–Meyer description. Hence (72) and (73) imply
C j − 1 ( F ‾ , z r ) ≠ 0 C^{j-1}(\overline{F},z_r)\ne0 C j − 1 ( F , z r ) = 0
for at least one r r r .
Choose either lift w w w of this z r z_r z r . Locally the covering map is a diffeomorphism, so the constrained Hessians at z r z_r z r and w w w have the same Morse index and nullity. The Fredholm splitting lemma and the shifting theorem give
C q ( F ‾ , z r ) ≅ C q − m ~ ( w ) ( h , 0 ) , C^q(\overline{F},z_r)\cong C^{q-\widetilde{m}(w)}(h,0), C q ( F , z r ) ≅ C q − m ( w ) ( h , 0 ) ,
where h h h is the reduced functional on the ν ~ ( w ) \widetilde{\nu}(w) ν ( w ) -dimensional null space, with an isolated critical point at 0 0 0 . Consequently these local groups vanish for
q < m ~ ( w ) or q > m ~ ( w ) + ν ~ ( w ) ; q<\widetilde{m}(w)\quad\text{or}\quad q>\widetilde{m}(w)+\widetilde{\nu}(w); q < m ( w ) or q > m ( w ) + ν ( w ) ;
see [7 Chapter I, Section 5] . Taking q = j − 1 q=j-1 q = j − 1 proves
m ~ ( w ) ≤ j − 1 ≤ m ~ ( w ) + ν ~ ( w ) , \widetilde{m}(w)\leq j-1\leq\widetilde{m}(w)+\widetilde{\nu}(w), m ( w ) ≤ j − 1 ≤ m ( w ) + ν ( w ) ,
which is (70) . □
We now compare the constrained Hessian of Ψ \Psi Ψ with the Hessian of E E E . Let w ∈ S w\in S w ∈ S be a critical point of Ψ \Psi Ψ , and set
t 0 : = D ( w ) − 1 / ( 2 p − 2 ) , v : = t 0 w , t_0:=\mathcal{D}(w)^{-1/(2p-2)},\qquad v:=t_0w, t 0 := D ( w ) − 1/ ( 2 p − 2 ) , v := t 0 w ,
as in (26) .
Lemma 8.2. At the corresponding critical points w ∈ S w\in S w ∈ S and v = t 0 w ∈ H v=t_0w\in H v = t 0 w ∈ H , one has
m G ( v ) = m ~ ( w ) + 1 , ν G ( v ) = ν ~ ( w ) . (74) m_G(v)=\widetilde{m}(w)+1,\qquad\nu_G(v)=\widetilde{\nu}(w). \tag*{(74)} m G ( v ) = m ( w ) + 1 , ν G ( v ) = ν ( w ) . (74)
Proof. For z = t ζ ∈ H ∖ { 0 } z=t\zeta\in H\setminus\{0\} z = tζ ∈ H ∖ { 0 } , where t > 0 t>0 t > 0 and ζ ∈ S \zeta\in S ζ ∈ S , the homogeneity of D \mathcal{D} D gives
E ( t ζ ) = 1 2 t 2 − 1 2 p t 2 p D ( ζ ) . (75) E(t\zeta)=\frac{1}{2}t^2-\frac{1}{2p}t^{2p}\mathcal{D}(\zeta). \tag*{(75)} E ( tζ ) = 2 1 t 2 − 2 p 1 t 2 p D ( ζ ) . (75)
The orthogonal decomposition
H = span { w } ⊕ T w S H=\operatorname{span}\{w\}\oplus T_wS H = span { w } ⊕ T w S
separates the radial and tangential directions. Let η ∈ T w S \eta\in T_wS η ∈ T w S . Differentiating (75) once in the radial direction and once in the tangential direction gives
∂ t ∂ η E ( t 0 w ) = − t 0 2 p − 1 d ( D ∣ S ) ( w ) [ η ] . \partial_t\partial_\eta E(t_0w)=-t_0^{2p-1}d(\mathcal{D}|_S)(w)[\eta]. ∂ t ∂ η E ( t 0 w ) = − t 0 2 p − 1 d ( D ∣ S ) ( w ) [ η ] .
The criticality of w w w for Ψ ∣ S \Psi|_S Ψ ∣ S , together with the definition of Ψ \Psi Ψ , implies d ( D ∣ S ) ( w ) = 0 d(\mathcal{D}|_S)(w)=0 d ( D ∣ S ) ( w ) = 0 . Hence,
∂ t ∂ η E ( t 0 w ) = 0. \partial_t\partial_\eta E(t_0w)=0. ∂ t ∂ η E ( t 0 w ) = 0.
Thus the Hessian of E E E at v v v has no mixed radial-tangential term. Differentiating (75) twice with respect to t t t and using t 0 2 p − 2 D ( w ) = 1 t_0^{2p-2}\mathcal{D}(w)=1 t 0 2 p − 2 D ( w ) = 1 , which follows from (26) , yields
∂ t 2 E ( t 0 w ) = 1 − ( 2 p − 1 ) t 0 2 p − 2 D ( w ) = − 2 ( p − 1 ) < 0. \partial_t^2E(t_0w)=1-(2p-1)t_0^{2p-2}\mathcal{D}(w)
=-2(p-1)<0. ∂ t 2 E ( t 0 w ) = 1 − ( 2 p − 1 ) t 0 2 p − 2 D ( w ) = − 2 ( p − 1 ) < 0.
Therefore, the radial direction contributes exactly one negative direction and no null direction.
The relation t 0 = D ( w ) − 1 / ( 2 p − 2 ) t_0=\mathcal{D}(w)^{-1/(2p-2)} t 0 = D ( w ) − 1/ ( 2 p − 2 ) gives t 0 2 p = D ( w ) − p / ( p − 1 ) t_0^{2p}=\mathcal{D}(w)^{-p/(p-1)} t 0 2 p = D ( w ) − p / ( p − 1 ) . At the constrained critical point w w w , the second derivative of Ψ \Psi Ψ along S S S is
Hess S Ψ ( w ) [ η , η ] = − 1 2 p D ( w ) − p / ( p − 1 ) Hess S D ( w ) [ η , η ] . (76) \operatorname{Hess}_S\Psi(w)[\eta,\eta]
=-\frac{1}{2p}\mathcal{D}(w)^{-p/(p-1)}
\operatorname{Hess}_S\mathcal{D}(w)[\eta,\eta]. \tag*{(76)} Hess S Ψ ( w ) [ η , η ] = − 2 p 1 D ( w ) − p / ( p − 1 ) Hess S D ( w ) [ η , η ] . (76)
Applying (75) along a geodesic of S S S through w w w with initial velocity η \eta η gives
Q v ( t 0 η ) = − 1 2 p t 0 2 p Hess S D ( w ) [ η , η ] . (77) Q_v(t_0\eta)=-\frac{1}{2p}t_0^{2p}\operatorname{Hess}_S\mathcal{D}(w)[\eta,\eta]. \tag*{(77)} Q v ( t 0 η ) = − 2 p 1 t 0 2 p Hess S D ( w ) [ η , η ] . (77)
The identity t 0 2 p = D ( w ) − p / ( p − 1 ) t_0^{2p}=\mathcal{D}(w)^{-p/(p-1)} t 0 2 p = D ( w ) − p / ( p − 1 ) , together with (76) and (77) , yields
Q v ( t 0 η ) = Hess S Ψ ( w ) [ η , η ] . Q_v(t_0\eta)=\operatorname{Hess}_S\Psi(w)[\eta,\eta]. Q v ( t 0 η ) = Hess S Ψ ( w ) [ η , η ] .
Since t 0 > 0 t_0>0 t 0 > 0 , the map
S t 0 : T w S ⟶ T w S , S t 0 η = t 0 η , S_{t_0}:T_wS\longrightarrow T_wS,\qquad S_{t_0}\eta=t_0\eta, S t 0 : T w S ⟶ T w S , S t 0 η = t 0 η ,
is a linear isomorphism. Polarizing the quadratic-form identity gives
E ′ ′ ( v ) [ S t 0 η , S t 0 ξ ] = Hess S Ψ ( w ) [ η , ξ ] , η , ξ ∈ T w S . E''(v)[S_{t_0}\eta,S_{t_0}\xi]
=\operatorname{Hess}_S\Psi(w)[\eta,\xi],
\qquad\eta,\xi\in T_wS. E ′′ ( v ) [ S t 0 η , S t 0 ξ ] = Hess S Ψ ( w ) [ η , ξ ] , η , ξ ∈ T w S .
Thus S t 0 S_{t_0} S t 0 carries the tangential Hessian of E E E to the constrained Hessian of Ψ \Psi Ψ . Their Morse indices and nullities agree. Adding the negative radial direction gives
m G ( v ) = m ~ ( w ) + 1 , ν G ( v ) = ν ~ ( w ) , m_G(v)=\widetilde m(w)+1,\qquad\nu_G(v)=\widetilde\nu(w), m G ( v ) = m ( w ) + 1 , ν G ( v ) = ν ( w ) ,
which proves (74) .
Proof of Theorem 1.4. Lemma 2.4 shows that Ψ \Psi Ψ is even and satisfies the Palais–Smale condition. Lemma 7.1, the positivity of D \mathcal{D} D , and (8) imply that Ψ ∈ C 2 ( S ) \Psi\in C^2(S) Ψ ∈ C 2 ( S ) . For each critical point w w w of Ψ \Psi Ψ at level c j c_j c j , set
t = D ( w ) − 1 / ( 2 p − 2 ) , v = t w . t=\mathcal{D}(w)^{-1/(2p-2)},\qquad v=tw. t = D ( w ) − 1/ ( 2 p − 2 ) , v = tw .
By Lemma 2.5, v v v is a critical point of E ∣ H E|_H E ∣ H . Since E E E is G G G -invariant, the principle of symmetric criticality [30 ] implies that v v v is a critical point of E E E on the full space H 1 ( S n ) H^1(\mathbb{S}^n) H 1 ( S n ) , and hence a G G G -invariant weak solution of (1) .
Lemma 7.2 therefore applies to v v v . Moreover, since v v v is G G G -invariant, the coefficients in the Hessian and the compact operator K v K_v K v preserve the invariant subspace H H H . Consequently, the Hessian of E ∣ H E|_H E ∣ H at v v v is represented on H H H by the identity minus a compact self-adjoint operator. Hence its representing operator is self-adjoint and Fredholm, with finite Morse index and nullity. The radial–tangential decomposition and the linear isomorphism established in Lemma 8.2 show that the constrained Hessian of Ψ \Psi Ψ on T w S T_wS T w S is also represented by a self-adjoint Fredholm operator, with finite Morse index and nullity.
By the isolation hypotheses of Theorem 1.4, every critical point of Ψ \Psi Ψ at level c j c_j c j is isolated in the full critical set of Ψ \Psi Ψ . Consequently, all the assumptions of Lemma 8.1 are satisfied with F = Ψ F=\Psi F = Ψ and d j = c j d_j=c_j d j = c j . Thus, there exists a critical point w j ∈ S w_j\in S w j ∈ S at level c j c_j c j such that
m ~ ( w j ) ≤ j − 1 ≤ m ~ ( w j ) + ν ~ ( w j ) . \widetilde{m}(w_j)\leq j-1\leq\widetilde{m}(w_j)+\widetilde{\nu}(w_j). m ( w j ) ≤ j − 1 ≤ m ( w j ) + ν ( w j ) .
Set
t j = D ( w j ) − 1 / ( 2 p − 2 ) , v j = t j w j . t_j=\mathcal{D}(w_j)^{-1/(2p-2)},\qquad v_j=t_jw_j. t j = D ( w j ) − 1/ ( 2 p − 2 ) , v j = t j w j .
By the preceding argument, v j v_j v j is a G G G -invariant weak solution of (1) with energy E ( v j ) = c j E(v_j)=c_j E ( v j ) = c j . By (74) ,
m G ( v j ) = m ~ ( w j ) + 1 , ν G ( v j ) = ν ~ ( w j ) . m_G(v_j)=\widetilde{m}(w_j)+1,\qquad\nu_G(v_j)=\widetilde{\nu}(w_j). m G ( v j ) = m ( w j ) + 1 , ν G ( v j ) = ν ( w j ) .
Substitution into the preceding inequalities yields
m G ( v j ) ≤ j ≤ m G ( v j ) + ν G ( v j ) . m_G(v_j)\leq j\leq m_G(v_j)+\nu_G(v_j). m G ( v j ) ≤ j ≤ m G ( v j ) + ν G ( v j ) .
Theorem 1.3, together with the preceding estimate, gives
N G ( v j ) ≤ m G ( v j ) ≤ j . N_G(v_j)\leq m_G(v_j)\leq j. N G ( v j ) ≤ m G ( v j ) ≤ j .
Finally, if ν G ( v j ) = 0 \nu_G(v_j)=0 ν G ( v j ) = 0 , then
m G ( v j ) ≤ j ≤ m G ( v j ) , m_G(v_j)\leq j\leq m_G(v_j), m G ( v j ) ≤ j ≤ m G ( v j ) ,
and hence m G ( v j ) = j m_G(v_j)=j m G ( v j ) = j .
Corollary 8.3. If c j = c j + 1 = ⋯ = c j + s c_j=c_{j+1}=\cdots=c_{j+s} c j = c j + 1 = ⋯ = c j + s , then the critical set at this common level has cohomological index at least s + 1 s+1 s + 1 . Consequently, a finite union of antipodal critical pairs can correspond to at most one minimax label.
Proof. Let c = c j = ⋯ = c j + s c=c_j=\cdots=c_{j+s} c = c j = ⋯ = c j + s , and denote the critical set of Ψ \Psi Ψ at level c c c by K c K_c K c . Lemma 5.1 gives i ( K c ) ≥ s + 1 i(K_c)\geq s+1 i ( K c ) ≥ s + 1 . Suppose that K c K_c K c is a nonempty finite union of antipodal pairs. Then K c / Z 2 K_c/\mathbb{Z}_2 K c / Z 2 is a nonempty finite discrete space, and hence
H 0 ( K c / Z 2 ; Z 2 ) ≠ 0 , H q ( K c / Z 2 ; Z 2 ) = 0 for q ≥ 1. H^0(K_c/\mathbb{Z}_2;\mathbb{Z}_2)\ne0,\qquad H^q(K_c/\mathbb{Z}_2;\mathbb{Z}_2)=0\quad\text{for }q\geq1. H 0 ( K c / Z 2 ; Z 2 ) = 0 , H q ( K c / Z 2 ; Z 2 ) = 0 for q ≥ 1.
The definition of the cohomological index therefore gives i ( K c ) = 1 i(K_c)=1 i ( K c ) = 1 . Combining this identity with i ( K c ) ≥ s + 1 i(K_c)\geq s+1 i ( K c ) ≥ s + 1 yields s = 0 s=0 s = 0 .
The isolation hypothesis of Theorem 1.4 is verified at c 1 c_1 c 1 in Proposition 9.5 , but remains open for the higher minimax levels. Accordingly, the asymptotic conclusions in Remark 9.6 are conditional.
Full-space Hessian and angular decomposition
We decompose the full Hessian into the angular sectors of O ( k ) × O ( ℓ ) O(k) \times O(\ell) O ( k ) × O ( ℓ ) . Each sector is determined by a radial quadratic form. We identify the symmetry-generated kernel and use the Jacobi spectrum to locate the sectors that contribute to the index and nullity. We then prove the transverse index bound and compute the three specified low blocks for minimizers odd under factor interchange when k = ℓ k = \ell k = ℓ and μ \mu μ is small.
For a , b ≥ 0 a,b \ge0 a , b ≥ 0 , let H a ( S k − 1 ) \mathcal{H}_a(\mathbb{S}^{k-1}) H a ( S k − 1 ) and H b ( S ℓ − 1 ) \mathcal{H}_b(\mathbb{S}^{\ell-1}) H b ( S ℓ − 1 ) denote the spaces of spherical harmonics of degrees a a a and b b b , respectively.
Symmetry-generated kernel modes
For a Euclidean solution u u u , the associated linearized operator is
L u ϕ = − Δ ϕ − ( p − 1 ) ( ∫ R n ∣ u ( y ) ∣ p ∣ x − y ∣ μ d y ) ∣ u ( x ) ∣ p − 2 ϕ ( x ) − p ( ∫ R n ∣ u ( y ) ∣ p − 2 u ( y ) ϕ ( y ) ∣ x − y ∣ μ d y ) ∣ u ( x ) ∣ p − 2 u ( x ) . \mathcal{L}_u\phi= -\Delta\phi-(p-1)\left(\int_{\mathbb{R}^n}\frac{|u(y)|^p}{|x-y|^\mu}\,dy\right)|u(x)|^{p-2}\phi(x)-p\left(\int_{\mathbb{R}^n}\frac{|u(y)|^{p-2}u(y)\phi(y)}{|x-y|^\mu}\,dy\right)|u(x)|^{p-2}u(x). L u ϕ = − Δ ϕ − ( p − 1 ) ( ∫ R n ∣ x − y ∣ μ ∣ u ( y ) ∣ p d y ) ∣ u ( x ) ∣ p − 2 ϕ ( x ) − p ( ∫ R n ∣ x − y ∣ μ ∣ u ( y ) ∣ p − 2 u ( y ) ϕ ( y ) d y ) ∣ u ( x ) ∣ p − 2 u ( x ) .
Proposition 9.1. Assume p ≥ 2 p \ge2 p ≥ 2 , and let v v v be a nonzero G G G -invariant solution with orbit profile f f f . Define
g 1 , 0 ( t ) = − sin t f ′ ( t ) − n − 2 2 cos t f ( t ) , g 0 , 1 ( t ) = cos t f ′ ( t ) − n − 2 2 sin t f ( t ) . (78) \begin{aligned}
g_{1,0}(t)&=-\sin t\,f'(t)-\frac{n-2}{2}\cos t\,f(t),\\
g_{0,1}(t)&=\cos t\,f'(t)-\frac{n-2}{2}\sin t\,f(t). \tag*{(78)}
\end{aligned} g 1 , 0 ( t ) g 0 , 1 ( t ) = − sin t f ′ ( t ) − 2 n − 2 cos t f ( t ) , = cos t f ′ ( t ) − 2 n − 2 sin t f ( t ) . (78) Neither radial profile is identically zero. For every Y 1 ∈ H 1 ( S k − 1 ) Y_1 \in\mathcal{H}_1(\mathbb{S}^{k-1}) Y 1 ∈ H 1 ( S k − 1 ) and Z 1 ∈ H 1 ( S ℓ − 1 ) Z_1 \in\mathcal{H}_1(\mathbb{S}^{\ell-1}) Z 1 ∈ H 1 ( S ℓ − 1 ) , the functions
g 1 , 0 ( t ) Y 1 ( ω ) , g 0 , 1 ( t ) Z 1 ( θ ) (79) g_{1,0}(t)Y_1(\omega),\qquad g_{0,1}(t)Z_1(\theta) \tag*{(79)} g 1 , 0 ( t ) Y 1 ( ω ) , g 0 , 1 ( t ) Z 1 ( θ ) (79) belong to the kernel of the full spherical Hessian. These two angular sectors therefore contribute at least k + ℓ = n + 1 k+\ell=n+1 k + ℓ = n + 1 dimensions to the full nullity.
If f f f is nonconstant, then
f ′ ( t ) Y 1 ( ω ) Z 1 ( θ ) (80) f'(t)Y_1(\omega)Z_1(\theta) \tag*{(80)} f ′ ( t ) Y 1 ( ω ) Z 1 ( θ ) (80) is a nonzero kernel field for every pair of nonzero degree-one harmonics Y 1 Y_1 Y 1 and Z 1 Z_1 Z 1 . The ( 1 , 1 ) (1,1) ( 1 , 1 ) sector consequently contributes at least k ℓ k\ell k ℓ additional dimensions to the full nullity. Hence, the tangent space K ( v ) \mathcal{K}(v) K ( v ) to the conformal orbit of v v v satisfies
dim K ( v ) = { n + 1 , f is constant , n + 1 + k ℓ , f is nonconstant . (81) \dim\mathcal{K}(v)=
\begin{cases}
n+1, & f\text{ is constant},\\
n+1+k\ell, & f\text{ is nonconstant}.
\end{cases} \tag*{(81)} dim K ( v ) = { n + 1 , n + 1 + k ℓ , f is constant , f is nonconstant . (81)
Proof. For q ∈ R n + 1 q \in\mathbb{R}^{n+1} q ∈ R n + 1 , define ζ q ( ξ ) = q ⋅ ξ \zeta_q(\xi)=q\cdot\xi ζ q ( ξ ) = q ⋅ ξ . The tangential gradient of ζ q \zeta_q ζ q is q − ζ q ξ q-\zeta_q\xi q − ζ q ξ . Taking the tangential part of its Euclidean derivative along a tangent vector gives
∇ S n 2 ζ q = − ζ q g S n . \nabla_{\mathbb{S}^n}^{2}\zeta_q=-\zeta_q g_{\mathbb{S}^n}. ∇ S n 2 ζ q = − ζ q g S n .
Hence, for X q = ∇ S n ζ q X_q=\nabla_{S^n}\zeta_q X q = ∇ S n ζ q , the Lie derivative of the metric along X q X_q X q satisfies
L X q g S n = 2 ∇ S n 2 ζ q = − 2 ζ q g S n . \mathcal{L}_{X_q}g_{S^n}=2\nabla^2_{S^n}\zeta_q=-2\zeta_qg_{S^n}. L X q g S n = 2 ∇ S n 2 ζ q = − 2 ζ q g S n .
Thus X q X_q X q is a conformal vector field, with infinitesimal length factor − ζ q -\zeta_q − ζ q . Let F s F_s F s be its flow and let F s ∗ g S n = λ s 2 g S n F_s^*g_{S^n}=\lambda_s^2g_{S^n} F s ∗ g S n = λ s 2 g S n . The metric identity gives
∂ s log λ s ( ξ ) = − ζ q ( F s ( ξ ) ) , λ 0 = 1. \partial_s\log\lambda_s(\xi)=-\zeta_q(F_s(\xi)),\qquad\lambda_0=1. ∂ s log λ s ( ξ ) = − ζ q ( F s ( ξ )) , λ 0 = 1.
For x s = F s ( ξ ) x_s=F_s(\xi) x s = F s ( ξ ) and y s = F s ( η ) y_s=F_s(\eta) y s = F s ( η ) , one has
d d s ∣ x s − y s ∣ 2 = 2 ( x s − y s ) ⋅ ( q − ( q ⋅ x s ) x s − q + ( q ⋅ y s ) y s ) = − ( ζ q ( x s ) + ζ q ( y s ) ) ∣ x s − y s ∣ 2 . \begin{aligned}
\frac{d}{ds}|x_s-y_s|^2=2(x_s-y_s)\cdot\bigl(q-(q\cdot x_s)x_s-q+(q\cdot y_s)y_s\bigr) \\
&=-(\zeta_q(x_s)+\zeta_q(y_s))|x_s-y_s|^2.
\end{aligned} d s d ∣ x s − y s ∣ 2 = 2 ( x s − y s ) ⋅ ( q − ( q ⋅ x s ) x s − q + ( q ⋅ y s ) y s ) = − ( ζ q ( x s ) + ζ q ( y s )) ∣ x s − y s ∣ 2 .
Integration and the initial value at s = 0 s=0 s = 0 give
∣ F s ( ξ ) − F s ( η ) ∣ 2 = λ s ( ξ ) λ s ( η ) ∣ ξ − η ∣ 2 , F s ∗ d V = λ s n d V . |F_s(\xi)-F_s(\eta)|^2=\lambda_s(\xi)\lambda_s(\eta)|\xi-\eta|^2,\qquad F_s^*dV=\lambda_s^n dV. ∣ F s ( ξ ) − F s ( η ) ∣ 2 = λ s ( ξ ) λ s ( η ) ∣ ξ − η ∣ 2 , F s ∗ d V = λ s n d V .
By Lemma 6.1, v ∈ C 2 , γ ( S n ) v\in C^{2,\gamma}(S^n) v ∈ C 2 , γ ( S n ) for some γ ∈ ( 0 , 1 ) \gamma\in(0,1) γ ∈ ( 0 , 1 ) . In particular, v v v is bounded on S n S^n S n , and the conformal orbit map below is differentiable as an H 1 ( S n ) H^1(S^n) H 1 ( S n ) -valued map. Put h = ( n − 2 ) / 2 h=(n-2)/2 h = ( n − 2 ) /2 and v s ( ξ ) = λ s ( ξ ) h v ( F s ( ξ ) ) v_s(\xi)=\lambda_s(\xi)^h v(F_s(\xi)) v s ( ξ ) = λ s ( ξ ) h v ( F s ( ξ )) . The conformal covariance of A A A gives ∥ v s ∥ = ∥ v ∥ \|v_s\|=\|v\| ∥ v s ∥ = ∥ v ∥ . In the nonlocal integral the exponent at each integration variable is
h p + μ 2 − n = 2 n − μ 2 + μ 2 − n = 0. hp+\frac{\mu}{2}-n=\frac{2n-\mu}{2}+\frac{\mu}{2}-n=0. h p + 2 μ − n = 2 2 n − μ + 2 μ − n = 0.
A change of variables therefore gives D ( v s ) = D ( v ) \mathcal{D}(v_s)=\mathcal{D}(v) D ( v s ) = D ( v ) , and hence E ( v s ) = E ( v ) E(v_s)=E(v) E ( v s ) = E ( v ) . The action is a linear isomorphism of the energy space for each s s s , so it maps critical points to critical points. Differentiating the conformal action on v v v gives the tangent field
Φ q = ∇ S n ζ q ⋅ ∇ S n v − n − 2 2 ζ q v . (82) \Phi_q=\nabla_{S^n}\zeta_q\cdot\nabla_{S^n}v-\frac{n-2}{2}\zeta_qv. \tag*{(82)} Φ q = ∇ S n ζ q ⋅ ∇ S n v − 2 n − 2 ζ q v . (82)
For the curve of solutions v s v_s v s defined above, we have
d d s ∣ s = 0 v s = Φ q , E ′ ( v s ) = 0. \left.\frac{d}{ds}\right|_{s=0}v_s=\Phi_q,\qquad E'(v_s)=0. d s d s = 0 v s = Φ q , E ′ ( v s ) = 0.
Differentiating the latter identity at s = 0 s=0 s = 0 and using Lemma 7.1 give
E ′ ′ ( v ) [ Φ q , ϕ ] = 0 , ϕ ∈ H 1 ( S n ) . E''(v)[\Phi_q,\phi]=0,\qquad\phi\in H^1(S^n). E ′′ ( v ) [ Φ q , ϕ ] = 0 , ϕ ∈ H 1 ( S n ) .
For 1 ≤ i ≤ k 1\leq i\leq k 1 ≤ i ≤ k , take ζ q = ξ i = cos t ω i \zeta_q=\xi_i=\cos t\,\omega_i ζ q = ξ i = cos t ω i . Then ∂ t ξ i = − sin t ω i \partial_t\xi_i=-\sin t\,\omega_i ∂ t ξ i = − sin t ω i , and (9.5) yields
Φ q = ( − sin t f ′ ( t ) − n − 2 2 cos t f ( t ) ) ω i = g 1 , 0 ( t ) ω i . \Phi_q=\left(-\sin t\,f'(t)-\frac{n-2}{2}\cos t\,f(t)\right)\omega_i=g_{1,0}(t)\omega_i. Φ q = ( − sin t f ′ ( t ) − 2 n − 2 cos t f ( t ) ) ω i = g 1 , 0 ( t ) ω i .
For 1 ≤ j ≤ ℓ 1\leq j\leq\ell 1 ≤ j ≤ ℓ , take ζ q = ξ k + j = sin t θ j \zeta_q=\xi_{k+j}=\sin t\,\theta_j ζ q = ξ k + j = sin t θ j . The identity ∂ t ξ k + j = cos t θ j \partial_t\xi_{k+j}=\cos t\,\theta_j ∂ t ξ k + j = cos t θ j gives
Φ q = ( cos t f ′ ( t ) − n − 2 2 sin t f ( t ) ) θ j = g 0 , 1 ( t ) θ j . \Phi_q=\left(\cos t\,f'(t)-\frac{n-2}{2}\sin t\,f(t)\right)\theta_j=g_{0,1}(t)\theta_j. Φ q = ( cos t f ′ ( t ) − 2 n − 2 sin t f ( t ) ) θ j = g 0 , 1 ( t ) θ j .
The coordinate functions ω 1 , … , ω k \omega_1,\ldots,\omega_k ω 1 , … , ω k and θ 1 , … , θ ℓ \theta_1,\ldots,\theta_\ell θ 1 , … , θ ℓ form bases of H 1 ( S k − 1 ) H_1(S^{k-1}) H 1 ( S k − 1 ) and H 1 ( S ℓ − 1 ) H_1(S^{\ell-1}) H 1 ( S ℓ − 1 ) , respectively. This proves the kernel assertion in (9.2).
Suppose that g 1 , 0 ≡ 0 g_{1,0} \equiv0 g 1 , 0 ≡ 0 . Then
f ′ ( t ) = − n − 2 2 cot t f ( t ) , f'(t)=-\frac{n-2}{2}\cot t f(t), f ′ ( t ) = − 2 n − 2 cot t f ( t ) ,
and integration gives
f ( t ) = C ( sin t ) − ( n − 2 ) / 2 . f(t)=C(\sin t)^{-(n-2)/2}. f ( t ) = C ( sin t ) − ( n − 2 ) /2 .
The boundedness of v v v at the singular orbit t = 0 t=0 t = 0 forces C = 0 C=0 C = 0 , contradicting v ≠ 0 v\ne0 v = 0 . Hence g 1 , 0 ≢ 0 g_{1,0}\not\equiv0 g 1 , 0 ≡ 0 . Similarly, g 0 , 1 ≡ 0 g_{0,1}\equiv0 g 0 , 1 ≡ 0 would imply
f ( t ) = C ( cos t ) − ( n − 2 ) / 2 , f(t)=C(\cos t)^{-(n-2)/2}, f ( t ) = C ( cos t ) − ( n − 2 ) /2 ,
and boundedness at t = π / 2 t=\pi/2 t = π /2 would again force C = 0 C=0 C = 0 . Therefore, g 0 , 1 ≢ 0 g_{0,1}\not\equiv0 g 0 , 1 ≡ 0 . The corresponding angular multiplicities are k k k and ℓ \ell ℓ .
For 1 ≤ i ≤ k 1\le i\le k 1 ≤ i ≤ k and 1 ≤ j ≤ ℓ 1\le j\le\ell 1 ≤ j ≤ ℓ , consider the infinitesimal rotation
R i , k + j = ξ i ∂ ξ k + j − ξ k + j ∂ ξ i . R_{i,k+j}=\xi_i\partial_{\xi_{k+j}}-\xi_{k+j}\partial_{\xi_i}. R i , k + j = ξ i ∂ ξ k + j − ξ k + j ∂ ξ i .
This vector field is tangent to S n S^n S n . From
R i , k + j ( sin 2 t ) = 2 sin t cos t ω i θ j R_{i,k+j}(\sin^2t)=2\sin t\cos t\,\omega_i\theta_j R i , k + j ( sin 2 t ) = 2 sin t cos t ω i θ j
one obtains R i , k + j t = ω i θ j R_{i,k+j}t=\omega_i\theta_j R i , k + j t = ω i θ j , and hence
R i , k + j v = f ′ ( t ) ω i θ j . R_{i,k+j}v=f'(t)\omega_i\theta_j. R i , k + j v = f ′ ( t ) ω i θ j .
Rotations preserve both the equation and the energy. Differentiating the rotated family of solutions at the identity gives
E ′ ′ ( v ) [ R i , k + j v , ϕ ] = 0 , ϕ ∈ H 1 ( S n ) . E''(v)[R_{i,k+j}v,\phi]=0,\qquad\phi\in H^1(S^n). E ′′ ( v ) [ R i , k + j v , ϕ ] = 0 , ϕ ∈ H 1 ( S n ) .
The regularity stated above shows that the fields in (9.3) belong to H 1 ( S n ) H^1(S^n) H 1 ( S n ) . If f f f is nonconstant, then f ′ ≢ 0 f'\not\equiv0 f ′ ≡ 0 . The products ω i θ j \omega_i\theta_j ω i θ j , 1 ≤ i ≤ k 1\le i\le k 1 ≤ i ≤ k , 1 ≤ j ≤ ℓ 1\le j\le\ell 1 ≤ j ≤ ℓ , form a basis of
H 1 ( S k − 1 ) ⊗ H 1 ( S ℓ − 1 ) , \mathcal{H}_1(S^{k-1})\otimes\mathcal{H}_1(S^{\ell-1}), H 1 ( S k − 1 ) ⊗ H 1 ( S ℓ − 1 ) ,
so the resulting k ℓ k\ell k ℓ kernel fields are linearly independent. The ( 1 , 1 ) (1,1) ( 1 , 1 ) sector is orthogonal to the ( 1 , 0 ) (1,0) ( 1 , 0 ) and ( 0 , 1 ) (0,1) ( 0 , 1 ) sectors.
Rotations within R k \mathbb{R}^k R k or R ℓ \mathbb{R}^{\ell} R ℓ annihilate the G G G -invariant function v v v . Every conformal vector field X X X on the round sphere is the sum of a rotational field and a gradient ∇ ζ q \nabla\zeta_q ∇ ζ q . Indeed, if L X g S n = 2 σ g S n \mathcal{L}_Xg_{S^n}=2\sigma g_{S^n} L X g S n = 2 σ g S n , differentiating the scalar curvature transformation [25 Section 2] gives − 2 ( n − 1 ) ( Δ S n + n ) σ = 0 -2(n-1)(\Delta_{S^n}+n)\sigma=0 − 2 ( n − 1 ) ( Δ S n + n ) σ = 0 , as the round scalar curvature is constant. Thus σ = − ζ q \sigma=-\zeta_q σ = − ζ q for some q ∈ R n + 1 q\in\mathbb{R}^{n+1} q ∈ R n + 1 by the first spherical eigenspace [11 Chapter 1] . The field X − ∇ ζ q X-\nabla\zeta_q X − ∇ ζ q is Killing and therefore generates rotations. Every rotational field is a linear combination of rotations within the two factors and the mixed rotations R i , k + j R_{i,k+j} R i , k + j . The within-factor rotations annihilate v v v . Thus the gradient conformal fields and the mixed rotations span K ( v ) \mathcal{K}(v) K ( v ) . When f f f is constant, the mixed rotation fields vanish, leaving n + 1 n+1 n + 1 independent directions. When f f f is nonconstant, the mixed rotations give k ℓ k\ell k ℓ additional independent directions in the orthogonal ( 1 , 1 ) (1,1) ( 1 , 1 ) sector. This proves (81) .
Angular block decomposition and endpoint domains
For nonnegative integers a , b a,b a , b , write
d k , a = dim H a ( S k − 1 ) , d ℓ , b = dim H b ( S ℓ − 1 ) d_{k,a}=\dim\mathcal{H}_a(\mathbb{S}^{k-1}),\qquad d_{\ell,b}=\dim\mathcal{H}_b(\mathbb{S}^{\ell-1}) d k , a = dim H a ( S k − 1 ) , d ℓ , b = dim H b ( S ℓ − 1 )
for the corresponding multiplicities.
Let Π a , b \Pi_{a,b} Π a , b be the orthogonal projection in the angular variables onto
H a ( S k − 1 ) ⊗ H b ( S ℓ − 1 ) . \mathcal{H}_a(\mathbb{S}^{k-1})\otimes\mathcal{H}_b(\mathbb{S}^{\ell-1}). H a ( S k − 1 ) ⊗ H b ( S ℓ − 1 ) .
The spaces
H a ( S k − 1 ) ⊗ H b ( S ℓ − 1 ) \mathcal{H}_a(\mathbb{S}^{k-1})\otimes\mathcal{H}_b(\mathbb{S}^{\ell-1}) H a ( S k − 1 ) ⊗ H b ( S ℓ − 1 )
are pairwise inequivalent irreducible representations of G = O ( k ) × O ( ℓ ) G=O(k)\times O(\ell) G = O ( k ) × O ( ℓ ) ; for a positive harmonic degree on an O ( 2 ) O(2) O ( 2 ) factor, reflection interchanges the two Fourier modes of opposite frequencies, which together form one irreducible O ( 2 ) O(2) O ( 2 ) -representation. The degree-zero space is the one-dimensional trivial representation; see [11 Chapter 1] . Since J μ J_\mu J μ is G G G -equivariant, it preserves every angular sector. Hence,
J μ Π a , b = Π a , b J μ . J_\mu\Pi_{a,b}=\Pi_{a,b}J_\mu. J μ Π a , b = Π a , b J μ .
Consequently,
Π a , b J μ Π a ′ , b ′ = 0 whenever ( a , b ) ≠ ( a ′ , b ′ ) . \Pi_{a,b}J_\mu\Pi_{a',b'}=0\qquad\text{whenever }(a,b)\ne(a',b'). Π a , b J μ Π a ′ , b ′ = 0 whenever ( a , b ) = ( a ′ , b ′ ) .
Set
λ k , a = a ( a + k − 2 ) , λ ℓ , b = b ( b + ℓ − 2 ) . \lambda_{k,a}=a(a+k-2),\qquad\lambda_{\ell,b}=b(b+\ell-2). λ k , a = a ( a + k − 2 ) , λ ℓ , b = b ( b + ℓ − 2 ) .
The radial form domain associated with the bidegree ( a , b ) (a,b) ( a , b ) is
X a , b = { g ∈ H l o c 1 ( 0 , π / 2 ) : ∫ 0 π / 2 ( ∣ g ′ ( t ) ∣ 2 ρ ( t ) + ( λ k , a cos 2 t + λ ℓ , b sin 2 t + n ( n − 2 ) 4 ) ∣ g ( t ) ∣ 2 ρ ( t ) ) d t < ∞ } . (83) \mathcal{X}_{a,b}=\left\{g\in H^1_{\mathrm{loc}}(0,\pi/2):\int_0^{\pi/2}\left(|g'(t)|^2\rho(t)+\left(\frac{\lambda_{k,a}}{\cos^2t}+\frac{\lambda_{\ell,b}}{\sin^2t}+\frac{n(n-2)}{4}\right)|g(t)|^2\rho(t)\right)\,dt<\infty\right\}. \tag*{(83)} X a , b = { g ∈ H loc 1 ( 0 , π /2 ) : ∫ 0 π /2 ( ∣ g ′ ( t ) ∣ 2 ρ ( t ) + ( cos 2 t λ k , a + sin 2 t λ ℓ , b + 4 n ( n − 2 ) ) ∣ g ( t ) ∣ 2 ρ ( t ) ) d t < ∞ } . (83)
For fixed harmonics of unit L 2 L^2 L 2 norm with respect to the standard surface measure, Y a ∈ H a ( S k − 1 ) Y_a\in\mathcal{H}_a(\mathbb{S}^{k-1}) Y a ∈ H a ( S k − 1 ) and Z b ∈ H b ( S ℓ − 1 ) Z_b\in\mathcal{H}_b(\mathbb{S}^{\ell-1}) Z b ∈ H b ( S ℓ − 1 ) , consider
ϕ ( t , ω , θ ) = g ( t ) Y a ( ω ) Z b ( θ ) , g ∈ X a , b . \phi(t,\omega,\theta)=g(t)Y_a(\omega)Z_b(\theta),\qquad g\in\mathcal{X}_{a,b}. ϕ ( t , ω , θ ) = g ( t ) Y a ( ω ) Z b ( θ ) , g ∈ X a , b .
For almost every ( t , s ) ∈ ( 0 , π / 2 ) × ( 0 , π / 2 ) (t,s)\in(0,\pi/2)\times(0,\pi/2) ( t , s ) ∈ ( 0 , π /2 ) × ( 0 , π /2 ) , define
K μ , a , b ( t , s ) = ∫ S k − 1 ∫ S ℓ − 1 ∫ S k − 1 ∫ S ℓ − 1 Y a ( ω ) Z b ( θ ) Y a ( ω ′ ) Z b ( θ ′ ) ( 2 − 2 cos t cos s ω ⋅ ω ′ − 2 sin t sin s θ ⋅ θ ′ ) μ / 2 d V θ ′ d V ω ′ d V θ d V ω . (84) K_{\mu,a,b}(t,s)=\int_{\mathbb{S}^{k-1}}\int_{\mathbb{S}^{\ell-1}}\int_{\mathbb{S}^{k-1}}\int_{\mathbb{S}^{\ell-1}}\frac{Y_a(\omega)Z_b(\theta)Y_a(\omega')Z_b(\theta')}{\left(2-2\cos t\cos s\,\omega\cdot\omega'-2\sin t\sin s\,\theta\cdot\theta'\right)^{\mu/2}}\,dV_{\theta'}\,dV_{\omega'}\,dV_\theta\,dV_\omega. \tag*{(84)} K μ , a , b ( t , s ) = ∫ S k − 1 ∫ S ℓ − 1 ∫ S k − 1 ∫ S ℓ − 1 ( 2 − 2 cos t cos s ω ⋅ ω ′ − 2 sin t sin s θ ⋅ θ ′ ) μ /2 Y a ( ω ) Z b ( θ ) Y a ( ω ′ ) Z b ( θ ′ ) d V θ ′ d V ω ′ d V θ d V ω . (84)
For fixed t ≠ s t\ne s t = s the angular integral operator commutes with G G G . Schur’s lemma on the complexified product harmonic space gives a scalar operator, so K μ , a , b K_{\mu,a,b} K μ , a , b is independent of the chosen real unit harmonics. Moreover,
K μ , a , b ( t , s ) = K μ , a , b ( s , t ) , K_{\mu,a,b}(t,s)=K_{\mu,a,b}(s,t), K μ , a , b ( t , s ) = K μ , a , b ( s , t ) ,
and the integral in (9.7) is finite whenever t ≠ s t \ne s t = s .
At a regular orbit, the angular dimension is n − 1 n-1 n − 1 . Near coincident angular points, the singular part of the integral has the form
∫ 0 ε r n − 2 − μ d r \int_{0}^{\varepsilon} r^{n-2-\mu}\,dr ∫ 0 ε r n − 2 − μ d r
which diverges when μ ≥ n − 1 \mu\ge n-1 μ ≥ n − 1 . Thus, K μ , a , b ( t , t ) K_{\mu,a,b}(t,t) K μ , a , b ( t , t ) need not be finite. The diagonal
{ ( t , s ) ∈ ( 0 , π / 2 ) 2 : t = s } \{(t,s) \in(0,\pi/2)^2 : t=s\} {( t , s ) ∈ ( 0 , π /2 ) 2 : t = s }
has two-dimensional Lebesgue measure zero. We may therefore assign arbitrary values to K μ , a , b K_{\mu,a,b} K μ , a , b there. We use its almost everywhere equivalence class, or the bounded projected operator Π a , b J μ Π a , b \Pi_{a,b}J_\mu\Pi_{a,b} Π a , b J μ Π a , b .
For the unit harmonics fixed above, define the radial operator by
( J μ , a , b g ) ( t ) = ∫ 0 π / 2 K μ , a , b ( t , s ) g ( s ) ρ ( s ) d s (J_{\mu,a,b}g)(t)=\int_{0}^{\pi/2}K_{\mu,a,b}(t,s)g(s)\rho(s)\,ds ( J μ , a , b g ) ( t ) = ∫ 0 π /2 K μ , a , b ( t , s ) g ( s ) ρ ( s ) d s
It is the bounded operator on L 2 ( ( 0 , π / 2 ) , ρ d t ) L^2((0,\pi/2),\rho\,dt) L 2 (( 0 , π /2 ) , ρ d t ) determined by
J μ ( g Y a Z b ) = ( J μ , a , b g ) Y a Z b . J_\mu(gY_aZ_b)=(J_{\mu,a,b}g)Y_aZ_b. J μ ( g Y a Z b ) = ( J μ , a , b g ) Y a Z b .
The integral formula holds almost everywhere. By (7.7) and the unit normalization of the harmonics,
∥ J μ , a , b g ∥ L 2 ( ρ d t ) ≤ κ 0 ∥ g ∥ L 2 ( ρ d t ) . \lVert J_{\mu,a,b}g\rVert_{L^2(\rho\,dt)}\le\kappa_0\lVert g\rVert_{L^2(\rho\,dt)}. ∥ J μ , a , b g ∥ L 2 ( ρ d t ) ≤ κ 0 ∥ g ∥ L 2 ( ρ d t ) .
Write
v ( cos t ω , sin t θ ) = f ( t ) , v(\cos t\,\omega,\sin t\,\theta)=f(t), v ( cos t ω , sin t θ ) = f ( t ) ,
and set V f ( t ) = J μ ( ∣ v ∣ p − 2 v ) ( cos t ω , sin t θ ) V_f(t)=J_\mu(|v|^{p-2}v)(\cos t\,\omega,\sin t\,\theta) V f ( t ) = J μ ( ∣ v ∣ p − 2 v ) ( cos t ω , sin t θ ) . Then V v ( t ) = V f ( t ) ∣ f ( t ) ∣ p − 2 V_v(t)=V_f(t)|f(t)|^{p-2} V v ( t ) = V f ( t ) ∣ f ( t ) ∣ p − 2 , with V v V_v V v as in (7.5). On X a , b \mathcal{X}_{a,b} X a , b , define
Q a , b ( g ) = ∫ 0 π / 2 [ ∣ g ′ ( t ) ∣ 2 + ( λ k , a cos 2 t + λ ℓ , b sin 2 t + n ( n − 2 ) 4 − ( p − 1 ) V f ( t ) ∣ f ( t ) ∣ p − 2 ) ∣ g ( t ) ∣ 2 ] ρ ( t ) d t − p ∫ 0 π / 2 ∫ 0 π / 2 K μ , a , b ( t , s ) ∣ f ( t ) ∣ p − 2 f ( t ) g ( t ) × ∣ f ( s ) ∣ p − 2 f ( s ) g ( s ) ρ ( t ) ρ ( s ) d t d s . (85) \begin{aligned}
Q_{a,b}(g)={}&\int_{0}^{\pi/2}\left[|g'(t)|^2+\left(\frac{\lambda_{k,a}}{\cos^2 t}+\frac{\lambda_{\ell,b}}{\sin^2 t}+\frac{n(n-2)}{4}\right.\right.\\
&\left.\left.-(p-1)V_f(t)|f(t)|^{p-2}\right)|g(t)|^2\right]\rho(t)\,dt\\
&-p\int_{0}^{\pi/2}\int_{0}^{\pi/2}K_{\mu,a,b}(t,s)|f(t)|^{p-2}f(t)g(t)\\
&\qquad\qquad\qquad\qquad\times|f(s)|^{p-2}f(s)g(s)\rho(t)\rho(s)\,dt\,ds. \tag*{(85)}
\end{aligned} Q a , b ( g ) = ∫ 0 π /2 [ ∣ g ′ ( t ) ∣ 2 + ( cos 2 t λ k , a + sin 2 t λ ℓ , b + 4 n ( n − 2 ) − ( p − 1 ) V f ( t ) ∣ f ( t ) ∣ p − 2 ) ∣ g ( t ) ∣ 2 ] ρ ( t ) d t − p ∫ 0 π /2 ∫ 0 π /2 K μ , a , b ( t , s ) ∣ f ( t ) ∣ p − 2 f ( t ) g ( t ) × ∣ f ( s ) ∣ p − 2 f ( s ) g ( s ) ρ ( t ) ρ ( s ) d t d s . (85)
The last term is understood as the quadratic form of the bounded operator Π a , b J μ Π a , b \Pi_{a,b}J_\mu\Pi_{a,b} Π a , b J μ Π a , b . Indeed, for h ( t ) = ∣ f ( t ) ∣ p − 2 f ( t ) g ( t ) h(t)=|f(t)|^{p-2}f(t)g(t) h ( t ) = ∣ f ( t ) ∣ p − 2 f ( t ) g ( t ) , the L 2 ( S n ) L^2(\mathbb{S}^n) L 2 ( S n ) -boundedness of J μ J_\mu J μ , from (7.7), gives
∣ ⟨ J μ ( h Y a Z b ) , h Y a Z b ⟩ L 2 ( S n ) ∣ ≤ κ 0 ∥ h Y a Z b ∥ L 2 ( S n ) 2 . \left|\left\langle J_\mu(hY_aZ_b),hY_aZ_b\right\rangle_{L^2(\mathbb{S}^n)}\right|\le\kappa_0\lVert hY_aZ_b\rVert_{L^2(\mathbb{S}^n)}^2. ⟨ J μ ( h Y a Z b ) , h Y a Z b ⟩ L 2 ( S n ) ≤ κ 0 ∥ h Y a Z b ∥ L 2 ( S n ) 2 .
The same estimate applied to ∣ h Y a Z b ∣ |hY_aZ_b| ∣ h Y a Z b ∣ gives
∫ S n × S n ∣ h ( t ) Y a ( ω ) Z b ( θ ) ∣ ∣ h ( s ) Y a ( ω ′ ) Z b ( θ ′ ) ∣ ∣ ξ − η ∣ μ d V ξ d V η ≤ κ 0 ∥ h Y a Z b ∥ 2 2 < ∞ . \int_{\mathbb{S}^n\times\mathbb{S}^n}
\frac{|h(t)Y_a(\omega)Z_b(\theta)||h(s)Y_a(\omega')Z_b(\theta')|}
{|\xi-\eta|^\mu}\,dV_\xi dV_\eta
\le\kappa_0\lVert hY_aZ_b\rVert_2^2<\infty. ∫ S n × S n ∣ ξ − η ∣ μ ∣ h ( t ) Y a ( ω ) Z b ( θ ) ∣∣ h ( s ) Y a ( ω ′ ) Z b ( θ ′ ) ∣ d V ξ d V η ≤ κ 0 ∥ h Y a Z b ∥ 2 2 < ∞.
Fubini’s theorem therefore yields (9.8). No finite pointwise value of K μ , a , b ( t , t ) K_{\mu,a,b}(t,t) K μ , a , b ( t , t ) is required.
The space X a , b \mathcal{X}_{a,b} X a , b is complete with respect to the weighted form norm in (83) . It is precisely the radial H 1 ( S n ) H^1(\mathbb{S}^n) H 1 ( S n ) -form domain of the ( a , b ) (a,b) ( a , b ) sector. The density at the singular orbits requires a logarithmic cutoff when the corresponding normal dimension equals two.
First replace g g g by its truncation
g T = max { − T , min { g , T } } . g_T=\max\{-T,\min\{g,T\}\}. g T = max { − T , min { g , T }} .
Then g T → g g_T\to g g T → g in the form norm as T → ∞ T\to\infty T → ∞ . It is therefore enough to consider bounded radial profiles. Near either singular orbit, let r r r denote the normal distance and let q q q denote the corresponding normal dimension, with q = ℓ q=\ell q = ℓ near t = 0 t=0 t = 0 and q = k q=k q = k near t = π / 2 t=\pi/2 t = π /2 .
If q > 2 q>2 q > 2 , take a cutoff that is zero for r ≤ ε r\leq\varepsilon r ≤ ε and one for r ≥ 2 ε r\geq2\varepsilon r ≥ 2 ε , with derivative supported in { ε < r < 2 ε } \{\varepsilon<r<2\varepsilon\} { ε < r < 2 ε } . It satisfies
∫ r q − 1 ∣ χ ε ′ ( r ) ∣ 2 d r = O ( ε q − 2 ) . \int r^{q-1}|\chi_\varepsilon'(r)|^2\,dr=O(\varepsilon^{q-2}). ∫ r q − 1 ∣ χ ε ′ ( r ) ∣ 2 d r = O ( ε q − 2 ) .
If q = 2 q=2 q = 2 , define
χ ε ( r ) = { 0 , r ≤ ε 2 , log ( r / ε 2 ) ∣ log ε ∣ , ε 2 < r < ε , 1 , r ≥ ε . \chi_\varepsilon(r)=
\begin{cases}
0, & r\leq\varepsilon^2,\\
\dfrac{\log(r/\varepsilon^2)}{|\log\varepsilon|}, & \varepsilon^2<r<\varepsilon,\\
1, & r\geq\varepsilon.
\end{cases} χ ε ( r ) = ⎩ ⎨ ⎧ 0 , ∣ log ε ∣ log ( r / ε 2 ) , 1 , r ≤ ε 2 , ε 2 < r < ε , r ≥ ε .
For this piecewise smooth cutoff,
∫ ε 2 ε r ∣ χ ε ′ ( r ) ∣ 2 d r = 1 ∣ log ε ∣ ⟶ 0. \int_{\varepsilon^2}^{\varepsilon}r|\chi_\varepsilon'(r)|^2\,dr=\frac{1}{|\log\varepsilon|}\longrightarrow0. ∫ ε 2 ε r ∣ χ ε ′ ( r ) ∣ 2 d r = ∣ log ε ∣ 1 ⟶ 0.
Smoothing the corners on comparable annuli preserves the bound C / ∣ log ε ∣ C/|\log\varepsilon| C /∣ log ε ∣ , which is sufficient for the density argument. The endpoint weights are comparable to r q − 1 d r r^{q-1}\,dr r q − 1 d r . The remaining derivative, L 2 ( S n ) L^2(\mathbb{S}^n) L 2 ( S n ) , and angular potential terms converge by dominated convergence. For bounded g g g , the derivative term satisfies
∫ ρ ∣ ( ( 1 − χ ε ) g ) ′ ∣ 2 d t ≤ 2 ∫ { χ ε ≠ 1 } ρ ∣ g ′ ∣ 2 d t + 2 ∥ g ∥ ∞ 2 ∫ ρ ∣ χ ε ′ ∣ 2 d t ⟶ 0. \int\rho\lvert((1-\chi_\varepsilon)g)'\rvert^2\,dt
\leq2\int_{\{\chi_\varepsilon\ne1\}}\rho|g'|^2\,dt
+2\|g\|_\infty^2\int\rho|\chi_\varepsilon'|^2\,dt
\longrightarrow0. ∫ ρ ∣(( 1 − χ ε ) g ) ′ ∣ 2 d t ≤ 2 ∫ { χ ε = 1 } ρ ∣ g ′ ∣ 2 d t + 2∥ g ∥ ∞ 2 ∫ ρ ∣ χ ε ′ ∣ 2 d t ⟶ 0.
The angular potential terms are bounded by the corresponding integrable terms in (83) , multiplied by ( 1 − χ ε ) 2 (1-\chi_\varepsilon)^2 ( 1 − χ ε ) 2 . They also tend to zero. Applying the cutoff at both endpoints and then smoothing in the interior proves that
C c ∞ ( 0 , π / 2 ) is dense in X a , b . C_c^\infty(0,\pi/2)\text{ is dense in }\mathcal{X}_{a,b}. C c ∞ ( 0 , π /2 ) is dense in X a , b .
The same approximation identifies X a , b \mathcal{X}_{a,b} X a , b with the radial form domain of the corresponding H 1 ( S n ) H^1(\mathbb{S}^n) H 1 ( S n ) fields across the singular orbits.
The local part of the block is the Friedrichs realization in L 2 ( ( 0 , π / 2 ) , ρ ( t ) d t ) L^2((0,\pi/2),\rho(t)\,dt) L 2 (( 0 , π /2 ) , ρ ( t ) d t ) of
A a , b g = − ρ − 1 ( ρ g ′ ) ′ + ( λ k , a cos 2 t + λ ℓ , b sin 2 t + n ( n − 2 ) 4 ) g , A_{a,b}g=-\rho^{-1}(\rho g')'
+\left(\frac{\lambda_{k,a}}{\cos^2t}
+\frac{\lambda_{\ell,b}}{\sin^2t}
+\frac{n(n-2)}{4}\right)g, A a , b g = − ρ − 1 ( ρ g ′ ) ′ + ( cos 2 t λ k , a + sin 2 t λ ℓ , b + 4 n ( n − 2 ) ) g ,
with form domain X a , b \mathcal{X}_{a,b} X a , b . The coefficients V v V_v V v and a v a_v a v are bounded. Multiplication by V v V_v V v is therefore bounded and self-adjoint on L 2 ( ρ d t ) L^2(\rho\,dt) L 2 ( ρ d t ) . Moreover, with M a v M_{a_v} M a v denoting multiplication by the radial profile of a v a_v a v ,
∥ M a v J μ , a , b M a v g ∥ L 2 ( ρ d t ) ≤ κ 0 ∥ a v ∥ ∞ 2 ∥ g ∥ L 2 ( ρ d t ) . \|M_{a_v}J_{\mu,a,b}M_{a_v}g\|_{L^2(\rho\,dt)}
\leq\kappa_0\|a_v\|_\infty^2\|g\|_{L^2(\rho\,dt)}. ∥ M a v J μ , a , b M a v g ∥ L 2 ( ρ d t ) ≤ κ 0 ∥ a v ∥ ∞ 2 ∥ g ∥ L 2 ( ρ d t ) .
The symmetry of J μ , a , b J_{\mu,a,b} J μ , a , b makes this nonlocal operator self-adjoint as well. Thus the multiplication and nonlocal terms in (9.8) are bounded self-adjoint perturbations. Therefore, the full ( a , b ) (a,b) ( a , b ) -block of L v L_v L v has the same form domain X a , b \mathcal{X}_{a,b} X a , b . The form domain prescribes the natural Friedrichs behavior at the singular endpoints. No zero trace is required when the corresponding angular label is zero. Functions in the operator domain still satisfy the regularity condition selected by this realization.
Proposition 9.2. Let p ≥ 2 p \ge2 p ≥ 2 , and let v ≠ 0 v \ne0 v = 0 be a G G G -invariant critical point of E E E on H 1 ( S n ) H^1(\mathbb{S}^n) H 1 ( S n ) . The Hessian operator commutes with the action of G = O ( k ) × O ( ℓ ) G = O(k) \times O(\ell) G = O ( k ) × O ( ℓ ) and is block diagonal with respect to the bidegrees ( a , b ) (a,b) ( a , b ) . If m a , b ( v ) m_{a,b}(v) m a , b ( v ) denotes the negative index of the radial form Q a , b Q_{a,b} Q a , b defined in (9.8), then
m ( v ) = ∑ a , b ≥ 0 d k , a d ℓ , b m a , b ( v ) . (86) m(v) = \sum_{a,b\ge0} d_{k,a}d_{\ell,b}m_{a,b}(v). \tag*{(86)} m ( v ) = a , b ≥ 0 ∑ d k , a d ℓ , b m a , b ( v ) . (86) Only finitely many summands in (9.9) are nonzero. Moreover,
m 0 , 0 ( v ) = m G ( v ) , m ( v ) ≥ m G ( v ) . (87) m_{0,0}(v) = m_G(v), \qquad m(v) \ge m_G(v). \tag*{(87)} m 0 , 0 ( v ) = m G ( v ) , m ( v ) ≥ m G ( v ) . (87) If ν a , b ( v ) \nu_{a,b}(v) ν a , b ( v ) denotes the nullity of Q a , b Q_{a,b} Q a , b , then
ν ( v ) = ∑ a , b ≥ 0 d k , a d ℓ , b ν a , b ( v ) . (88) \nu(v) = \sum_{a,b\ge0} d_{k,a}d_{\ell,b}\nu_{a,b}(v). \tag*{(88)} ν ( v ) = a , b ≥ 0 ∑ d k , a d ℓ , b ν a , b ( v ) . (88) and only finitely many summands in (9.11) are nonzero. Furthermore,
ν 1 , 0 ( v ) ≥ 1 , ν 0 , 1 ( v ) ≥ 1. (89) \nu_{1,0}(v) \ge1, \qquad\nu_{0,1}(v) \ge1. \tag*{(89)} ν 1 , 0 ( v ) ≥ 1 , ν 0 , 1 ( v ) ≥ 1. (89) If v v v is nonconstant, then ν 1 , 1 ( v ) ≥ 1 \nu_{1,1}(v) \ge1 ν 1 , 1 ( v ) ≥ 1 .
Proof. The coefficients of the local terms in (7.3) depend only on t t t . Moreover, the kernel ∣ ξ − η ∣ − μ |\xi-\eta|^{-\mu} ∣ ξ − η ∣ − μ is invariant under the diagonal action of O ( n + 1 ) O(n+1) O ( n + 1 ) , and hence under G G G . It follows that the L 2 ( S n ) L^2(\mathbb{S}^n) L 2 ( S n ) -realization L v L_v L v of the full Hessian commutes with the unitary action of G G G .
Choose orthonormal bases
{ Y a , r } r = 1 d k , a ⊂ H a ( S k − 1 ) , { Z b , s } s = 1 d ℓ , b ⊂ H b ( S ℓ − 1 ) . \{Y_{a,r}\}_{r=1}^{d_{k,a}} \subset\mathcal{H}_a(\mathbb{S}^{k-1}), \qquad\{Z_{b,s}\}_{s=1}^{d_{\ell,b}} \subset\mathcal{H}_b(\mathbb{S}^{\ell-1}). { Y a , r } r = 1 d k , a ⊂ H a ( S k − 1 ) , { Z b , s } s = 1 d ℓ , b ⊂ H b ( S ℓ − 1 ) .
The spectral decompositions on the two sphere factors [11 Chapter 1] and Fubini’s theorem yield
ϕ = ∑ a , b ≥ 0 ∑ r = 1 d k , a ∑ s = 1 d ℓ , b g a , b , r , s ( t ) Y a , r ( ω ) Z b , s ( θ ) \phi= \sum_{a,b\ge0}\sum_{r=1}^{d_{k,a}}\sum_{s=1}^{d_{\ell,b}} g_{a,b,r,s}(t)Y_{a,r}(\omega)Z_{b,s}(\theta) ϕ = a , b ≥ 0 ∑ r = 1 ∑ d k , a s = 1 ∑ d ℓ , b g a , b , r , s ( t ) Y a , r ( ω ) Z b , s ( θ )
for every ϕ ∈ H 1 ( S n ) \phi\in H^1(\mathbb{S}^n) ϕ ∈ H 1 ( S n ) . Parseval’s identity and the spherical harmonic eigenvalue equations give
∥ ϕ ∥ 2 = ∑ a , b ≥ 0 ∑ r = 1 d k , a ∑ s = 1 d ℓ , b ∫ 0 π / 2 [ ∣ g a , b , r , s ′ ∣ 2 + ( n ( n − 2 ) 4 + λ k , a cos 2 t + λ ℓ , b sin 2 t ) ∣ g a , b , r , s ∣ 2 ] ρ ( t ) d t . \|\phi\|^2 = \sum_{a,b\ge0}\sum_{r=1}^{d_{k,a}}\sum_{s=1}^{d_{\ell,b}} \int_0^{\pi/2} \left[ |g'_{a,b,r,s}|^2 + \left(\frac{n(n-2)}{4} + \frac{\lambda_{k,a}}{\cos^2 t} + \frac{\lambda_{\ell,b}}{\sin^2 t}\right)|g_{a,b,r,s}|^2 \right]\rho(t)\,\mathrm{d}t. ∥ ϕ ∥ 2 = a , b ≥ 0 ∑ r = 1 ∑ d k , a s = 1 ∑ d ℓ , b ∫ 0 π /2 [ ∣ g a , b , r , s ′ ∣ 2 + ( 4 n ( n − 2 ) + cos 2 t λ k , a + sin 2 t λ ℓ , b ) ∣ g a , b , r , s ∣ 2 ] ρ ( t ) d t .
The positivity of n ( n − 2 ) / 4 n(n-2)/4 n ( n − 2 ) /4 shows that each integral is equivalent to the weighted norm in (9.6). Consequently, g a , b , r , s ∈ X a , b g_{a,b,r,s} \in\mathcal{X}_{a,b} g a , b , r , s ∈ X a , b .
Schur’s lemma, applied after complexification, makes the Hessian matrix elements zero between distinct bidegrees and scalar in the angular variables within each bidegree. Restricting to real-valued functions gives the same decomposition. In the L 2 ( S n ) L^2(S^n) L 2 ( S n ) -realization, the restriction of L v L_v L v to the ( a , b ) (a,b) ( a , b ) -sector is
I H a ( S k − 1 ) ⊗ H b ( S ℓ − 1 ) ⊗ L a , b , I_{\mathcal{H}_a(S^{k-1})\otimes\mathcal{H}_b(S^{\ell-1})}\otimes L_{a,b}, I H a ( S k − 1 ) ⊗ H b ( S ℓ − 1 ) ⊗ L a , b ,
where L a , b L_{a,b} L a , b is the self-adjoint radial operator associated with the closed form Q a , b Q_{a,b} Q a , b on X a , b \mathcal{X}_{a,b} X a , b .
The local terms in (54) , together with the spherical harmonic eigenvalue equations, give the first integral in (85) . Applying Fubini’s theorem to the nonlocal term gives the projected kernel K μ , a , b K_{\mu,a,b} K μ , a , b defined in (84) . Hence the quadratic form associated with L a , b L_{a,b} L a , b is precisely Q a , b Q_{a,b} Q a , b .
Every negative direction of Q a , b Q_{a,b} Q a , b therefore occurs in the full Hessian with multiplicity d k , a d ℓ , b d_{k,a}d_{\ell,b} d k , a d ℓ , b . Summing these multiplicities over all bidegrees gives (86) . Lemma 7.2 shows that the negative spectral space of the full Hessian is finite-dimensional. Thus, only finitely many terms in (86) are nonzero.
The same orthogonal decomposition of the kernel gives (88) . Lemma 7.2 also shows that the full kernel is finite-dimensional, so only finitely many terms in (88) are nonzero.
The zero-degree harmonic spaces consist of constant functions. The map g ↦ g Y 0 Z 0 g\mapsto gY_0Z_0 g ↦ g Y 0 Z 0 , with Y 0 Y_0 Y 0 and Z 0 Z_0 Z 0 the unit constant harmonics, identifies X 0 , 0 \mathcal{X}_{0,0} X 0 , 0 with H H H and carries Q 0 , 0 Q_{0,0} Q 0 , 0 to Q v ∣ H Q_v|_H Q v ∣ H . Therefore, m 0 , 0 ( v ) = m G ( v ) m_{0,0}(v)=m_G(v) m 0 , 0 ( v ) = m G ( v ) . Equation (86) then gives
m ( v ) ≥ m G ( v ) , m(v)\ge m_G(v), m ( v ) ≥ m G ( v ) ,
which proves (87) .
Proposition 9.1 provides nonzero radial kernel modes in the ( 1 , 0 ) (1,0) ( 1 , 0 ) - and ( 0 , 1 ) (0,1) ( 0 , 1 ) -blocks. Therefore, (89) holds. If v v v is nonconstant, the rotation fields in Proposition 9.1 provide a nonzero radial kernel mode in the ( 1 , 1 ) (1,1) ( 1 , 1 ) -block, and hence ν 1 , 1 ( v ) ≥ 1 \nu_{1,1}(v)\ge1 ν 1 , 1 ( v ) ≥ 1 .
Jacobi spectra and localization of negative modes
Proposition 9.3. The radial Friedrichs realization of A A A in the sector ( a , b ) (a,b) ( a , b ) has the simple eigenvalues
λ m , a , b = ( a + b + 2 m ) ( a + b + 2 m + n − 1 ) + n ( n − 2 ) 4 , m ≥ 0. (90) \lambda_{m,a,b}=(a+b+2m)(a+b+2m+n-1)+\frac{n(n-2)}{4},\qquad m\ge0. \tag*{(90)} λ m , a , b = ( a + b + 2 m ) ( a + b + 2 m + n − 1 ) + 4 n ( n − 2 ) , m ≥ 0. (90) The corresponding eigenprofiles are
g m , a , b ( t ) = γ m , a , b ( cos t ) a ( sin t ) b P m ( α a , b , β a , b ) ( cos 2 t ) , (91) g_{m,a,b}(t)=\gamma_{m,a,b}(\cos t)^a(\sin t)^bP_m^{(\alpha_{a,b},\beta_{a,b})}(\cos2t), \tag*{(91)} g m , a , b ( t ) = γ m , a , b ( cos t ) a ( sin t ) b P m ( α a , b , β a , b ) ( cos 2 t ) , (91) where
α a , b = b + ℓ − 2 2 , β a , b = a + k − 2 2 , \alpha_{a,b}=b+\frac{\ell-2}{2},\qquad\beta_{a,b}=a+\frac{k-2}{2}, α a , b = b + 2 ℓ − 2 , β a , b = a + 2 k − 2 , and γ m , a , b > 0 \gamma_{m,a,b}>0 γ m , a , b > 0 is chosen so that ∥ g m , a , b ∥ L 2 ( ( 0 , π / 2 ) , ρ d t ) = 1 \lVert g_{m,a,b}\rVert_{L^2((0,\pi/2),\rho\,dt)}=1 ∥ g m , a , b ∥ L 2 (( 0 , π /2 ) , ρ d t ) = 1 . The family { g m , a , b } m ≥ 0 \{g_{m,a,b}\}_{m\ge0} { g m , a , b } m ≥ 0 is an orthonormal basis of L 2 ( ( 0 , π / 2 ) , ρ d t ) L^2((0,\pi/2),\rho\,dt) L 2 (( 0 , π /2 ) , ρ d t ) and an orthogonal basis of X a , b \mathcal{X}_{a,b} X a , b with respect to the local energy form.
The projected Riesz operator has the same eigenprofiles, with
J μ , a , b g m , a , b = κ a + b + 2 m g m , a , b . (92) J_{\mu,a,b}g_{m,a,b}=\kappa_{a+b+2m}g_{m,a,b}. \tag*{(92)} J μ , a , b g m , a , b = κ a + b + 2 m g m , a , b . (92)
Proof. Set
w a , b ( t ) = ( cos t ) a ( sin t ) b , g = w a , b h , w_{a,b}(t)=(\cos t)^a(\sin t)^b,\qquad g=w_{a,b}h, w a , b ( t ) = ( cos t ) a ( sin t ) b , g = w a , b h ,
and define
ρ a , b ( t ) = w a , b ( t ) 2 ρ ( t ) = ( cos t ) k − 1 + 2 a ( sin t ) ℓ − 1 + 2 b . \rho_{a,b}(t)=w_{a,b}(t)^2\rho(t)=(\cos t)^{k-1+2a}(\sin t)^{\ell-1+2b}. ρ a , b ( t ) = w a , b ( t ) 2 ρ ( t ) = ( cos t ) k − 1 + 2 a ( sin t ) ℓ − 1 + 2 b .
The identities
w a , b ′ w a , b = − a tan t + b cot t , λ k , a = a ( a + k − 2 ) , λ ℓ , b = b ( b + ℓ − 2 ) \frac{w'_{a,b}}{w_{a,b}}=-a\tan t+b\cot t,\qquad\lambda_{k,a}=a(a+k-2),\qquad\lambda_{\ell,b}=b(b+\ell-2) w a , b w a , b ′ = − a tan t + b cot t , λ k , a = a ( a + k − 2 ) , λ ℓ , b = b ( b + ℓ − 2 )
give
− ρ − 1 ( ρ w a , b ′ ) ′ + ( λ k , a cos 2 t + λ ℓ , b sin 2 t ) w a , b = ( a + b ) ( a + b + n − 1 ) w a , b . -\rho^{-1}(\rho w'_{a,b})'+\left(\frac{\lambda_{k,a}}{\cos^2t}+\frac{\lambda_{\ell,b}}{\sin^2t}\right)w_{a,b}=(a+b)(a+b+n-1)w_{a,b}. − ρ − 1 ( ρ w a , b ′ ) ′ + ( cos 2 t λ k , a + sin 2 t λ ℓ , b ) w a , b = ( a + b ) ( a + b + n − 1 ) w a , b .
Substituting g = w a , b h g=w_{a,b}h g = w a , b h and integrating by parts, we obtain
∫ 0 π / 2 [ ∣ g ′ ∣ 2 + ( λ k , a cos 2 t + λ ℓ , b sin 2 t + n ( n − 2 ) 4 ) g 2 ] ρ ( t ) d t = ∫ 0 π / 2 [ ∣ h ′ ∣ 2 + ( ( a + b ) ( a + b + n − 1 ) + n ( n − 2 ) 4 ) h 2 ] ρ a , b ( t ) d t . (93) \begin{aligned}
\int_0^{\pi/2}\left[|g'|^2+\left(\frac{\lambda_{k,a}}{\cos^2t}+\frac{\lambda_{\ell,b}}{\sin^2t}+\frac{n(n-2)}{4}\right)g^2\right]\rho(t)\,dt \\
&=\int_0^{\pi/2}\left[|h'|^2+\left((a+b)(a+b+n-1)+\frac{n(n-2)}{4}\right)h^2\right]\rho_{a,b}(t)\,dt. \tag*{(93)}
\end{aligned} ∫ 0 π /2 [ ∣ g ′ ∣ 2 + ( cos 2 t λ k , a + sin 2 t λ ℓ , b + 4 n ( n − 2 ) ) g 2 ] ρ ( t ) d t = ∫ 0 π /2 [ ∣ h ′ ∣ 2 + ( ( a + b ) ( a + b + n − 1 ) + 4 n ( n − 2 ) ) h 2 ] ρ a , b ( t ) d t . (93)
for every h ∈ C c ∞ ( 0 , π / 2 ) h\in C^\infty_c(0,\pi/2) h ∈ C c ∞ ( 0 , π /2 ) . Density in X a , b \mathcal{X}_{a,b} X a , b extends (9.16) to the corresponding closed form domains.
The multiplication operator
U a , b h = w a , b h U_{a,b}h=w_{a,b}h U a , b h = w a , b h
is unitary from L 2 ( ( 0 , π / 2 ) , ρ a , b d t ) L^2((0,\pi/2),\rho_{a,b}\,dt) L 2 (( 0 , π /2 ) , ρ a , b d t ) onto L 2 ( ( 0 , π / 2 ) , ρ d t ) L^2((0,\pi/2),\rho\,dt) L 2 (( 0 , π /2 ) , ρ d t ) , as
∥ U a , b h ∥ L 2 ( ( 0 , π / 2 ) , ρ d t ) 2 = ∫ 0 π / 2 ∣ h ∣ 2 ρ a , b d t . \|U_{a,b}h\|^2_{L^2((0,\pi/2),\rho\,dt)}=\int_0^{\pi/2}|h|^2\rho_{a,b}\,dt. ∥ U a , b h ∥ L 2 (( 0 , π /2 ) , ρ d t ) 2 = ∫ 0 π /2 ∣ h ∣ 2 ρ a , b d t .
Equation (9.16) also identifies the associated form domains. Under this unitary transformation, the radial operator is conjugate to
− ρ a , b − 1 ( ρ a , b h ′ ) ′ + ( a + b ) ( a + b + n − 1 ) h + n ( n − 2 ) 4 h . -\rho_{a,b}^{-1}(\rho_{a,b}h')'+(a+b)(a+b+n-1)h+\frac{n(n-2)}{4}h. − ρ a , b − 1 ( ρ a , b h ′ ) ′ + ( a + b ) ( a + b + n − 1 ) h + 4 n ( n − 2 ) h .
The weight ρ a , b \rho_{a,b} ρ a , b is obtained from the invariant radial weight by replacing k k k and ℓ \ell ℓ with k + 2 a k+2a k + 2 a and ℓ + 2 b \ell+2b ℓ + 2 b , respectively. Proposition 3.2 therefore applies with
α a , b = ℓ + 2 b − 2 2 , β a , b = k + 2 a − 2 2 . \alpha_{a,b}=\frac{\ell+2b-2}{2},\qquad\beta_{a,b}=\frac{k+2a-2}{2}. α a , b = 2 ℓ + 2 b − 2 , β a , b = 2 k + 2 a − 2 .
The derivative part has eigenfunctions
P m ( α a , b , β a , b ) ( cos 2 t ) P_m^{(\alpha_{a,b},\beta_{a,b})}(\cos2t) P m ( α a , b , β a , b ) ( cos 2 t )
and eigenvalues
2 m ( 2 m + n − 1 + 2 a + 2 b ) . 2m(2m+n-1+2a+2b). 2 m ( 2 m + n − 1 + 2 a + 2 b ) .
Adding the constant term in the conjugated operator gives
λ m , a , b = 2 m ( 2 m + n − 1 + 2 a + 2 b ) + ( a + b ) ( a + b + n − 1 ) + n ( n − 2 ) 4 = ( 2 m + a + b ) ( 2 m + a + b + n − 1 ) + n ( n − 2 ) 4 , \begin{aligned}
\lambda_{m,a,b}=2m(2m+n-1+2a+2b)+(a+b)(a+b+n-1)+\frac{n(n-2)}{4} \\
&=(2m+a+b)(2m+a+b+n-1)+\frac{n(n-2)}{4},
\end{aligned} λ m , a , b = 2 m ( 2 m + n − 1 + 2 a + 2 b ) + ( a + b ) ( a + b + n − 1 ) + 4 n ( n − 2 ) = ( 2 m + a + b ) ( 2 m + a + b + n − 1 ) + 4 n ( n − 2 ) ,
which agrees with (9.13). Applying U a , b U_{a,b} U a , b gives the profiles in (9.14).
The completeness and orthogonality of the Jacobi polynomials in the weighted Jacobi space, together with the unitary map U a , b U_{a,b} U a , b , show that { g m , a , b } m ≥ 0 \{g_{m,a,b}\}_{m\geq0} { g m , a , b } m ≥ 0 is an orthonormal basis of L 2 ( ( 0 , π / 2 ) , ρ d t ) L^2((0,\pi/2),\rho\,dt) L 2 (( 0 , π /2 ) , ρ d t ) . The spectral theorem for the radial Friedrichs operator gives orthogonal completeness in the form domain X a , b X_{a,b} X a , b . Finally, λ m + 1 , a , b > λ m , a , b \lambda_{m+1,a,b}>\lambda_{m,a,b} λ m + 1 , a , b > λ m , a , b for every m ≥ 0 m\geq0 m ≥ 0 . Together with completeness of the displayed eigenprofiles, this shows that each radial eigenspace is one-dimensional.
To identify the Riesz multipliers, extend Y a Y_a Y a and Z b Z_b Z b to homogeneous harmonic polynomials H a H_a H a and H b H_b H b on R k \mathbb{R}^k R k and R ℓ \mathbb{R}^{\ell} R ℓ . On the sphere,
( cos t ) a Y a ( ω ) = H a ( ξ ′ ) , ( sin t ) b Z b ( θ ) = H b ( ξ ′ ′ ) , cos 2 t = ∣ ξ ′ ∣ 2 − ∣ ξ ′ ′ ∣ 2 . (\cos t)^aY_a(\omega)=H_a(\xi'),\qquad(\sin t)^bZ_b(\theta)=H_b(\xi''),\qquad\cos2t=|\xi'|^2-|\xi''|^2. ( cos t ) a Y a ( ω ) = H a ( ξ ′ ) , ( sin t ) b Z b ( θ ) = H b ( ξ ′′ ) , cos 2 t = ∣ ξ ′ ∣ 2 − ∣ ξ ′′ ∣ 2 .
Thus g m , a , b Y a Z b g_{m,a,b}Y_aZ_b g m , a , b Y a Z b is smooth across both singular orbits. Its spherical Laplace eigenvalue, computed above, is ( a + b + 2 m ) ( a + b + 2 m + n − 1 ) (a+b+2m)(a+b+2m+n-1) ( a + b + 2 m ) ( a + b + 2 m + n − 1 ) . It therefore lies in the spherical harmonic space of degree a + b + 2 m a+b+2m a + b + 2 m . The multiplier formula in Lemma 7.3 gives
J μ ( g m , a , b Y a Z b ) = κ a + b + 2 m g m , a , b Y a Z b . J_\mu(g_{m,a,b}Y_aZ_b)=\kappa_{a+b+2m}g_{m,a,b}Y_aZ_b. J μ ( g m , a , b Y a Z b ) = κ a + b + 2 m g m , a , b Y a Z b .
This proves (9.15). The radial Riesz operator is therefore the same for every unit angular vector in the sector.
For a fixed sector, put g m = g m , a , b g_m=g_{m,a,b} g m = g m , a , b , and use ⟨ ⋅ , ⋅ ⟩ ρ \langle\cdot,\cdot\rangle_\rho ⟨ ⋅ , ⋅ ⟩ ρ for the scalar product in L 2 ( ( 0 , π / 2 ) , ρ d t ) L^2((0,\pi/2),\rho\,dt) L 2 (( 0 , π /2 ) , ρ d t ) . The operator associated with the radial form is
L a , b = A a , b − ( p − 1 ) V v − p a v J μ , a , b ( a v ⋅ ) . L_{a,b}=A_{a,b}-(p-1)V_v-p\,a_vJ_{\mu,a,b}(a_v\cdot). L a , b = A a , b − ( p − 1 ) V v − p a v J μ , a , b ( a v ⋅ ) .
Its polarized form has Jacobi matrix
Q a , b [ g r , g s ] = λ r , a , b δ r s − ( p − 1 ) ⟨ V v g r , g s ⟩ ρ − p ∑ m ≥ 0 κ a + b + 2 m ⟨ a v g r , g m ⟩ ρ ⟨ a v g s , g m ⟩ ρ . (94) Q_{a,b}[g_r,g_s]=\lambda_{r,a,b}\delta_{rs}-(p-1)\langle V_vg_r,g_s\rangle_\rho
-p\sum_{m\geq0}\kappa_{a+b+2m}\langle a_vg_r,g_m\rangle_\rho\langle a_vg_s,g_m\rangle_\rho. \tag*{(94)} Q a , b [ g r , g s ] = λ r , a , b δ r s − ( p − 1 ) ⟨ V v g r , g s ⟩ ρ − p m ≥ 0 ∑ κ a + b + 2 m ⟨ a v g r , g m ⟩ ρ ⟨ a v g s , g m ⟩ ρ . (94)
Indeed, expand each of a v g r a_vg_r a v g r and a v g s a_vg_s a v g s in the orthonormal Jacobi basis and use (9.15). Since a v ∈ L ∞ a_v\in L^\infty a v ∈ L ∞ and 0 < κ m ≤ κ 0 0<\kappa_m\leq\kappa_0 0 < κ m ≤ κ 0 ,
∑ m ≥ 0 κ a + b + 2 m ∣ ⟨ a v g r , g m ⟩ ρ ∣ ∣ ⟨ a v g s , g m ⟩ ρ ∣ ≤ κ 0 ∥ a v g r ∥ L 2 ( ρ d t ) ∥ a v g s ∥ L 2 ( ρ d t ) < ∞ . \sum_{m\geq0}\kappa_{a+b+2m}\lvert\langle a_vg_r,g_m\rangle_\rho\rvert\lvert\langle a_vg_s,g_m\rangle_\rho\rvert
\leq\kappa_0\lVert a_vg_r\rVert_{L^2(\rho\,dt)}\lVert a_vg_s\rVert_{L^2(\rho\,dt)}<\infty. m ≥ 0 ∑ κ a + b + 2 m ∣⟨ a v g r , g m ⟩ ρ ∣ ∣⟨ a v g s , g m ⟩ ρ ∣ ≤ κ 0 ∥ a v g r ∥ L 2 ( ρ d t ) ∥ a v g s ∥ L 2 ( ρ d t ) < ∞.
This justifies the series and passage from finite expansions to the closed form. The Riesz operator is diagonal in this basis, whereas the multiplications by a v a_v a v and V v V_v V v generally are not.
Proposition 9.4. Let p ≥ 2 p\geq2 p ≥ 2 , and let v ≠ 0 v\ne0 v = 0 be a G G G -invariant critical point of E E E on H 1 ( S n ) H^1(\mathbb{S}^n) H 1 ( S n ) . Set
M v = ( 2 p − 1 ) ∥ V v ∥ L ∞ ( S n ) , (95) M_v=(2p-1)\lVert V_v\rVert_{L^\infty(\mathbb{S}^n)}, \tag*{(95)} M v = ( 2 p − 1 ) ∥ V v ∥ L ∞ ( S n ) , (95) where V v V_v V v is defined in (7.5). Then
Q v ( ϕ ) ≥ ∥ ϕ ∥ 2 − M v ∥ ϕ ∥ L 2 ( S n ) 2 for every ϕ ∈ H 1 ( S n ) . (96) Q_v(\phi)\geq\lVert\phi\rVert^2-M_v\lVert\phi\rVert_{L^2(\mathbb{S}^n)}^2\qquad\text{for every }\phi\in H^1(\mathbb{S}^n). \tag*{(96)} Q v ( ϕ ) ≥ ∥ ϕ ∥ 2 − M v ∥ ϕ ∥ L 2 ( S n ) 2 for every ϕ ∈ H 1 ( S n ) . (96)
For every bidegree ( a , b ) (a,b) ( a , b ) ,
m a , b ( v ) ≤ # { m ≥ 0 : λ m , a , b < M v } . (97) m_{a,b}(v) \le\#\{m \ge0 : \lambda_{m,a,b} < M_v\}. \tag*{(97)} m a , b ( v ) ≤ # { m ≥ 0 : λ m , a , b < M v } . (97)
The nonpositive spectral space satisfies
m a , b ( v ) + ν a , b ( v ) ≤ # { m ≥ 0 : λ m , a , b ≤ M v } . m_{a,b}(v) + \nu_{a,b}(v) \le\#\{m \ge0 : \lambda_{m,a,b} \le M_v\}. m a , b ( v ) + ν a , b ( v ) ≤ # { m ≥ 0 : λ m , a , b ≤ M v } .
In particular,
m a , b ( v ) = 0 if ( a + b ) ( a + b + n − 1 ) + n ( n − 2 ) 4 ≥ M v . (98) m_{a,b}(v) = 0 \quad\text{if} \quad(a+b)(a+b+n-1) + \frac{n(n-2)}{4} \ge M_v. \tag*{(98)} m a , b ( v ) = 0 if ( a + b ) ( a + b + n − 1 ) + 4 n ( n − 2 ) ≥ M v . (98)
If the inequality in (98) is strict, then ν a , b ( v ) = 0 \nu_{a,b}(v)=0 ν a , b ( v ) = 0 . If N A ( s ) \mathcal{N}_A(s) N A ( s ) denotes the number of eigenvalues of A A A below s s s , counted with multiplicity, then
m ( v ) ≤ N A ( M v ) ≤ C n ( 1 + M v n / 2 ) . (99) m(v) \le\mathcal{N}_A(M_v) \le C_n(1+M_v^{n/2}). \tag*{(99)} m ( v ) ≤ N A ( M v ) ≤ C n ( 1 + M v n /2 ) . (99)
Proof. The pointwise kernel is positive, and Lemma 7.3 gives ∫ S n J μ ( a v ϕ ) a v ϕ d V ≥ 0 \int_{\mathbb{S}^n} J_\mu(a_v\phi)a_v\phi\,dV \ge0 ∫ S n J μ ( a v ϕ ) a v ϕ d V ≥ 0 . For the upper bound, apply 2 ∣ x y ∣ ≤ x 2 + y 2 2|xy| \le x^2+y^2 2∣ x y ∣ ≤ x 2 + y 2 under the double integral with
x = ∣ v ( ξ ) ∣ ( p − 2 ) / 2 ∣ v ( η ) ∣ p / 2 ϕ ( ξ ) , y = ∣ v ( η ) ∣ ( p − 2 ) / 2 ∣ v ( ξ ) ∣ p / 2 ϕ ( η ) . x=|v(\xi)|^{(p-2)/2}|v(\eta)|^{p/2}\phi(\xi), \qquad y=|v(\eta)|^{(p-2)/2}|v(\xi)|^{p/2}\phi(\eta). x = ∣ v ( ξ ) ∣ ( p − 2 ) /2 ∣ v ( η ) ∣ p /2 ϕ ( ξ ) , y = ∣ v ( η ) ∣ ( p − 2 ) /2 ∣ v ( ξ ) ∣ p /2 ϕ ( η ) .
The two square terms have the same integral after interchanging ξ \xi ξ and η \eta η . Hence
0 ≤ ∫ S n J μ ( a v ϕ ) a v ϕ d V ≤ ∫ S n J μ ( ∣ v ∣ p ) ∣ v ∣ p − 2 ϕ 2 d V = ∫ S n V v ϕ 2 d V . 0 \le\int_{\mathbb{S}^n} J_\mu(a_v\phi)a_v\phi\,dV \le\int_{\mathbb{S}^n}J_\mu(|v|^p)|v|^{p-2}\phi^2\,dV = \int_{\mathbb{S}^n}V_v\phi^2\,dV. 0 ≤ ∫ S n J μ ( a v ϕ ) a v ϕ d V ≤ ∫ S n J μ ( ∣ v ∣ p ) ∣ v ∣ p − 2 ϕ 2 d V = ∫ S n V v ϕ 2 d V .
Substituting into (54) yields
Q v ( ϕ ) ≥ ∥ ϕ ∥ 2 − ( ( p − 1 ) + p ) ∫ S n V v ϕ 2 d V ≥ ∥ ϕ ∥ 2 − M v ∥ ϕ ∥ 2 2 . \begin{aligned}
Q_v(\phi) \ge\|\phi\|^2 - ((p-1)+p)\int_{\mathbb{S}^n}V_v\phi^2\,dV \\
&\ge\|\phi\|^2-M_v\|\phi\|_2^2.
\end{aligned} Q v ( ϕ ) ≥ ∥ ϕ ∥ 2 − (( p − 1 ) + p ) ∫ S n V v ϕ 2 d V ≥ ∥ ϕ ∥ 2 − M v ∥ ϕ ∥ 2 2 .
Moreover,
M v ≤ ( 2 p − 1 ) κ 0 ∥ v ∥ ∞ 2 p − 2 . M_v \le(2p-1)\kappa_0\|v\|_\infty^{2p-2}. M v ≤ ( 2 p − 1 ) κ 0 ∥ v ∥ ∞ 2 p − 2 .
This proves (96) , with the sharper threshold M v M_v M v of (95) .
For g ∈ X a , b g \in\mathcal{X}_{a,b} g ∈ X a , b , write
∥ g ∥ A a , b 2 = ∫ 0 π / 2 [ ∣ g ′ ∣ 2 + ( λ k , a cos 2 t + λ ℓ , b sin 2 t + n ( n − 2 ) 4 ) ∣ g ∣ 2 ] ρ ( t ) d t . \|g\|_{\mathcal{A}_{a,b}}^2=\int_0^{\pi/2}\left[|g'|^2+\left(\frac{\lambda_{k,a}}{\cos^2t}+\frac{\lambda_{\ell,b}}{\sin^2t}+\frac{n(n-2)}{4}\right)|g|^2\right]\rho(t)\,dt. ∥ g ∥ A a , b 2 = ∫ 0 π /2 [ ∣ g ′ ∣ 2 + ( cos 2 t λ k , a + sin 2 t λ ℓ , b + 4 n ( n − 2 ) ) ∣ g ∣ 2 ] ρ ( t ) d t .
Fix a bidegree ( a , b ) (a,b) ( a , b ) , and define
Y a , b = span { g m , a , b : λ m , a , b ≥ M v } ‾ X a , b . \mathcal{Y}_{a,b}=\overline{\operatorname{span}\{g_{m,a,b}:\lambda_{m,a,b}\ge M_v\}}^{\mathcal{X}_{a,b}}. Y a , b = span { g m , a , b : λ m , a , b ≥ M v } X a , b .
Proposition 9.3 gives
codim X a , b Y a , b = # { m ≥ 0 : λ m , a , b < M v } . \operatorname{codim}_{\mathcal{X}_{a,b}}\mathcal{Y}_{a,b}=\#\{m\ge0:\lambda_{m,a,b}<M_v\}. codim X a , b Y a , b = # { m ≥ 0 : λ m , a , b < M v } .
For
g = ∑ λ m , a , b ≥ M v c m g m , a , b ∈ Y a , b g=\sum_{\lambda_{m,a,b}\ge M_v}c_mg_{m,a,b}\in\mathcal{Y}_{a,b} g = λ m , a , b ≥ M v ∑ c m g m , a , b ∈ Y a , b
the spectral decomposition of the local radial operator gives
∥ g ∥ A a , b 2 = ∑ λ m , a , b ≥ M v λ m , a , b ∣ c m ∣ 2 ≥ M v ∑ λ m , a , b ≥ M v ∣ c m ∣ 2 = M v ∥ g ∥ L 2 ( ( 0 , π / 2 ) , ρ d t ) 2 . \|g\|_{\mathcal{A}_{a,b}}^2=\sum_{\lambda_{m,a,b}\ge M_v}\lambda_{m,a,b}|c_m|^2
\ge M_v\sum_{\lambda_{m,a,b}\ge M_v}|c_m|^2=M_v\|g\|_{L^2((0,\pi/2),\rho\,dt)}^2. ∥ g ∥ A a , b 2 = λ m , a , b ≥ M v ∑ λ m , a , b ∣ c m ∣ 2 ≥ M v λ m , a , b ≥ M v ∑ ∣ c m ∣ 2 = M v ∥ g ∥ L 2 (( 0 , π /2 ) , ρ d t ) 2 .
The continuity of Q a , b Q_{a,b} Q a , b and (9.19) therefore imply
Q a , b ( g ) ≥ 0 for every g ∈ Y a , b . Q_{a,b}(g)\ge0\qquad\text{for every }g\in\mathcal{Y}_{a,b}. Q a , b ( g ) ≥ 0 for every g ∈ Y a , b .
Let V ⊂ X a , b \mathcal{V}\subset\mathcal{X}_{a,b} V ⊂ X a , b be a subspace on which Q a , b Q_{a,b} Q a , b is negative definite. If dim V > codim Y a , b \dim\mathcal{V}>\operatorname{codim}\mathcal{Y}_{a,b} dim V > codim Y a , b , then the finite-codimension intersection principle gives
V ∩ Y a , b ≠ { 0 } . \mathcal{V}\cap\mathcal{Y}_{a,b}\ne\{0\}. V ∩ Y a , b = { 0 } .
This contradicts the nonnegativity of Q a , b Q_{a,b} Q a , b on Y a , b \mathcal{Y}_{a,b} Y a , b . Hence,
m a , b ( v ) ≤ codim Y a , b , m_{a,b}(v)\le\operatorname{codim}\mathcal{Y}_{a,b}, m a , b ( v ) ≤ codim Y a , b ,
which proves (9.20). For the bound on the nonpositive spectral space, replace λ m , a , b ≥ M v \lambda_{m,a,b}\ge M_v λ m , a , b ≥ M v by λ m , a , b > M v \lambda_{m,a,b}>M_v λ m , a , b > M v . On the resulting closed subspace every nonzero vector satisfies
Q a , b ( g ) ≥ ∑ λ m , a , b > M v ( λ m , a , b − M v ) ∣ c m ∣ 2 > 0. Q_{a,b}(g)\ge\sum_{\lambda_{m,a,b}>M_v}(\lambda_{m,a,b}-M_v)|c_m|^2>0. Q a , b ( g ) ≥ λ m , a , b > M v ∑ ( λ m , a , b − M v ) ∣ c m ∣ 2 > 0.
It has codimension # { m : λ m , a , b ≤ M v } \#\{m:\lambda_{m,a,b}\le M_v\} # { m : λ m , a , b ≤ M v } . A larger nonpositive spectral subspace would intersect this positive subspace nontrivially, a contradiction.
Proposition 9.3 gives
λ 0 , a , b = ( a + b ) ( a + b + n − 1 ) + n ( n − 2 ) 4 . \lambda_{0,a,b}=(a+b)(a+b+n-1)+\frac{n(n-2)}{4}. λ 0 , a , b = ( a + b ) ( a + b + n − 1 ) + 4 n ( n − 2 ) .
If λ 0 , a , b ≥ M v \lambda_{0,a,b}\ge M_v λ 0 , a , b ≥ M v , then Y a , b = X a , b \mathcal{Y}_{a,b}=\mathcal{X}_{a,b} Y a , b = X a , b , and hence Q a , b ≥ 0 Q_{a,b}\ge0 Q a , b ≥ 0 on X a , b \mathcal{X}_{a,b} X a , b . This proves (9.21). If λ 0 , a , b > M v \lambda_{0,a,b}>M_v λ 0 , a , b > M v , then
Q a , b ( g ) ≥ ( λ 0 , a , b − M v ) ∥ g ∥ L 2 ( ( 0 , π / 2 ) , ρ d t ) 2 > 0 Q_{a,b}(g)\ge(\lambda_{0,a,b}-M_v)\|g\|_{L^2((0,\pi/2),\rho\,dt)}^2>0 Q a , b ( g ) ≥ ( λ 0 , a , b − M v ) ∥ g ∥ L 2 (( 0 , π /2 ) , ρ d t ) 2 > 0
for every g ≠ 0 g\ne0 g = 0 . Therefore, ν a , b ( v ) = 0 \nu_{a,b}(v)=0 ν a , b ( v ) = 0 .
Let Y \mathcal{Y} Y be the closure in H 1 ( S n ) H^1(S^n) H 1 ( S n ) of the span of all eigenfunctions of A A A whose eigenvalues are at least M v M_v M v . The spectral theorem gives
codim H 1 ( S n ) Y = N A ( M v ) . \operatorname{codim}_{H^1(S^n)}\mathcal{Y}=\mathcal{N}_A(M_v). codim H 1 ( S n ) Y = N A ( M v ) .
The spectral decomposition of A A A , together with (9.19), yields
Q v ( ϕ ) ≥ 0 for every ϕ ∈ Y . Q_v(\phi)\ge0\qquad\text{for every }\phi\in\mathcal{Y}. Q v ( ϕ ) ≥ 0 for every ϕ ∈ Y .
The same finite-codimension intersection argument then gives
m ( v ) ≤ N A ( M v ) . m(v)\le\mathcal{N}_A(M_v). m ( v ) ≤ N A ( M v ) .
To distinguish the full harmonic degree from the invariant Jacobi label, write the eigenvalues of A A A as
χ L = L ( L + n − 1 ) + n ( n − 2 ) 4 , L ≥ 0 , \chi_L=L(L+n-1)+\frac{n(n-2)}{4},\qquad L\ge0, χ L = L ( L + n − 1 ) + 4 n ( n − 2 ) , L ≥ 0 ,
with multiplicity
dim H L ( S n ) = ( n + L n ) − ( n + L − 2 n ) = O ( ( L + 1 ) n − 1 ) . \dim\mathcal{H}_L(\mathbb{S}^n)=\binom{n+L}{n}-\binom{n+L-2}{n}=O((L+1)^{n-1}). dim H L ( S n ) = ( n n + L ) − ( n n + L − 2 ) = O (( L + 1 ) n − 1 ) .
Here ( q n ) = 0 \binom{q}{n}=0 ( n q ) = 0 for nonnegative integers q < n q<n q < n . The inequality χ L < M v \chi_L<M_v χ L < M v implies L ≤ C n ( 1 + M v 1 / 2 ) L\le C_n(1+M_v^{1/2}) L ≤ C n ( 1 + M v 1/2 ) . Therefore,
N A ( M v ) ≤ C n ∑ 0 ≤ L ≤ C n ( 1 + M v 1 / 2 ) ( L + 1 ) n − 1 ≤ C n ( 1 + M v n / 2 ) , \begin{aligned}
\mathcal{N}_A(M_v)\le C_n\sum_{0\le L\le C_n(1+M_v^{1/2})}(L+1)^{n-1} \\
&\le C_n(1+M_v^{n/2}),
\end{aligned} N A ( M v ) ≤ C n 0 ≤ L ≤ C n ( 1 + M v 1/2 ) ∑ ( L + 1 ) n − 1 ≤ C n ( 1 + M v n /2 ) ,
which completes the proof of (99) .
Transverse negative directions
Proposition 9.5. Let p ≥ 2 p\ge2 p ≥ 2 . Each constant solution v = a 0 v=a_0 v = a 0 or v = − a 0 v=-a_0 v = − a 0 satisfies
m ( v ) = m G ( v ) = 1 , ν ( v ) = n + 1 , ν G ( v ) = 0. m(v)=m_G(v)=1,\qquad\nu(v)=n+1,\qquad\nu_G(v)=0. m ( v ) = m G ( v ) = 1 , ν ( v ) = n + 1 , ν G ( v ) = 0. The kernel of the full Hessian is precisely the space of degree-one spherical harmonics. Moreover, the first minimax level c 1 c_1 c 1 is isolated in the set of critical values of Ψ \Psi Ψ , and the critical set of Ψ \Psi Ψ at this level consists of one antipodal pair.
Proof. The constant state identity κ 0 a 0 2 p − 2 = n ( n − 2 ) / 4 \kappa_0a_0^{2p-2}=n(n-2)/4 κ 0 a 0 2 p − 2 = n ( n − 2 ) /4 and the Funk–Hecke multiplier formula (63) show that the L 2 ( S n ) L^2(\mathbb{S}^n) L 2 ( S n ) -realization L v L_v L v is diagonal with respect to the spherical harmonic decomposition. On the space H L ( S n ) \mathcal{H}_L(\mathbb{S}^n) H L ( S n ) of degree-L L L spherical harmonics, the corresponding eigenvalue is
q L = L ( L + n − 1 ) + ( 2 − p ) n ( n − 2 ) 4 − p n ( n − 2 ) 4 κ L κ 0 . q_L=L(L+n-1)+(2-p)\frac{n(n-2)}{4}-p\frac{n(n-2)}{4}\frac{\kappa_L}{\kappa_0}. q L = L ( L + n − 1 ) + ( 2 − p ) 4 n ( n − 2 ) − p 4 n ( n − 2 ) κ 0 κ L . To determine the Morse index and nullity, we transfer this spectral information to the energy space operator T v T_v T v . The operator T v T_v T v has eigenvalue
q L L ( L + n − 1 ) + n ( n − 2 ) / 4 \frac{q_L}{L(L+n-1)+n(n-2)/4} L ( L + n − 1 ) + n ( n − 2 ) /4 q L on the same harmonic space. The denominator is strictly positive, so the two realizations have the same signs and the same zero eigenspaces.
Formula (63) gives
κ 1 κ 0 = μ 2 n − μ . \frac{\kappa_1}{\kappa_0}=\frac{\mu}{2n-\mu}. κ 0 κ 1 = 2 n − μ μ . Using p = ( 2 n − μ ) / ( n − 2 ) p=(2n-\mu)/(n-2) p = ( 2 n − μ ) / ( n − 2 ) , direct substitution yields
q 0 = − 2 ( p − 1 ) n ( n − 2 ) 4 < 0 , q 1 = 0. q_0=-2(p-1)\frac{n(n-2)}{4}<0,\qquad q_1=0. q 0 = − 2 ( p − 1 ) 4 n ( n − 2 ) < 0 , q 1 = 0. For every L ≥ 2 L\ge2 L ≥ 2 ,
q L − q 1 = L ( L + n − 1 ) − n + p n ( n − 2 ) 4 κ 1 − κ L κ 0 . q_L-q_1=L(L+n-1)-n+p\frac{n(n-2)}{4}\frac{\kappa_1-\kappa_L}{\kappa_0}. q L − q 1 = L ( L + n − 1 ) − n + p 4 n ( n − 2 ) κ 0 κ 1 − κ L . The Laplace eigenvalues L ( L + n − 1 ) L(L+n-1) L ( L + n − 1 ) increase strictly with L L L , while the multipliers κ L \kappa_L κ L decrease strictly by (63) . Hence,
q L − q 1 > 0 for every L ≥ 2. q_L-q_1>0\qquad\text{for every }L\ge2. q L − q 1 > 0 for every L ≥ 2.
Thus, the degree-zero harmonic space is the only negative eigenspace, and the degree-one harmonic space is the entire kernel. Since dim H 0 ( S n ) = 1 \dim\mathcal{H}_0(\mathbb{S}^n)=1 dim H 0 ( S n ) = 1 and dim H 1 ( S n ) = n + 1 \dim\mathcal{H}_1(\mathbb{S}^n)=n+1 dim H 1 ( S n ) = n + 1 , it follows that
m ( v ) = 1 , ν ( v ) = n + 1. m(v)=1,\qquad\nu(v)=n+1. m ( v ) = 1 , ν ( v ) = n + 1.
Proposition 3.2 shows that the G G G -invariant spherical harmonics have even degrees L = 2 m L=2m L = 2 m . The negative degree-zero mode is G G G -invariant, whereas the degree-one kernel contains no nonzero G G G -invariant function. All invariant modes of positive even degree have positive Hessian eigenvalues. Therefore,
m G ( v ) = 1 , ν G ( v ) = 0. m_G(v)=1,\qquad\nu_G(v)=0. m G ( v ) = 1 , ν G ( v ) = 0.
Let
w 0 = a 0 ∥ a 0 ∥ ∈ S . w_0=\frac{a_0}{\lVert a_0\rVert}\in S. w 0 = ∥ a 0 ∥ a 0 ∈ S .
Lemma 5.2 , together with the radial correspondence between critical points of Ψ \Psi Ψ and nonzero critical points of E ∣ H E|_H E ∣ H , gives
{ w ∈ S : d ( Ψ ∣ S ) ( w ) = 0 , Ψ ( w ) = c 1 } = { w 0 , − w 0 } . \{w\in S:d(\Psi|_S)(w)=0,\ \Psi(w)=c_1\}=\{w_0,-w_0\}. { w ∈ S : d ( Ψ ∣ S ) ( w ) = 0 , Ψ ( w ) = c 1 } = { w 0 , − w 0 } .
By Lemma 8.2 ,
ν ~ ( w 0 ) = ν G ( a 0 ) = 0 , ν ~ ( − w 0 ) = ν G ( − a 0 ) = 0. \widetilde{\nu}(w_0)=\nu_G(a_0)=0,\qquad\widetilde{\nu}(-w_0)=\nu_G(-a_0)=0. ν ( w 0 ) = ν G ( a 0 ) = 0 , ν ( − w 0 ) = ν G ( − a 0 ) = 0.
Thus, w 0 w_0 w 0 and − w 0 -w_0 − w 0 are nondegenerate critical points of Ψ \Psi Ψ . The derivative of its gradient in a local Hilbert chart is an invertible Fredholm Hessian at each of the two points. The inverse function theorem therefore makes each critical point isolated.
Suppose that c 1 c_1 c 1 were not isolated in the set of critical values. Then there would exist distinct critical values d j → c 1 d_j\to c_1 d j → c 1 and critical points w j ∈ S w_j\in S w j ∈ S satisfying
Ψ ( w j ) = d j , d ( Ψ ∣ S ) ( w j ) = 0. \Psi(w_j)=d_j,\qquad d(\Psi|_S)(w_j)=0. Ψ ( w j ) = d j , d ( Ψ ∣ S ) ( w j ) = 0.
The sequence { w j } \{w_j\} { w j } is a Palais–Smale sequence at level c 1 c_1 c 1 . Lemma 2.4 provides a subsequence converging strongly in H H H to a critical point w w w of Ψ \Psi Ψ with Ψ ( w ) = c 1 \Psi(w)=c_1 Ψ ( w ) = c 1 . Hence, w = w 0 w=w_0 w = w 0 or w = − w 0 w=-w_0 w = − w 0 . This contradicts the local isolation of the two critical points. Therefore, c 1 c_1 c 1 is an isolated critical value.
Proof of Theorem 1.6. Let u = P v u=Pv u = P v . By Proposition 2.1, u u u is a Euclidean weak solution. Lemma 6.1 and the formula defining P P P yield u ∈ C l o c 2 , γ ( R n ) u\in C_{\mathrm{loc}}^{2,\gamma}(\mathbb{R}^n) u ∈ C loc 2 , γ ( R n ) for some γ ∈ ( 0 , 1 ) \gamma\in(0,1) γ ∈ ( 0 , 1 ) , which justifies differentiating the translation and dilation families introduced below. For λ > 0 \lambda>0 λ > 0 and a ∈ R n a\in\mathbb{R}^n a ∈ R n , define
u λ , a ( x ) = λ ( n − 2 ) / 2 u ( λ ( x − a ) ) . (100) u_{\lambda,a}(x)=\lambda^{(n-2)/2}u\bigl(\lambda(x-a)\bigr). \tag*{(100)} u λ , a ( x ) = λ ( n − 2 ) /2 u ( λ ( x − a ) ) . (100)
Set X = λ ( x − a ) X=\lambda(x-a) X = λ ( x − a ) and z = λ ( y − a ) z=\lambda(y-a) z = λ ( y − a ) . Then
− Δ u λ , a ( x ) = λ ( n + 2 ) / 2 ( − Δ u ) ( X ) . -\Delta u_{\lambda,a}(x)=\lambda^{(n+2)/2}(-\Delta u)(X). − Δ u λ , a ( x ) = λ ( n + 2 ) /2 ( − Δ u ) ( X ) .
Moreover,
∫ R n ∣ u λ , a ( y ) ∣ p ∣ x − y ∣ μ d y = λ ( n − 2 ) p / 2 + μ − n ∫ R n ∣ u ( z ) ∣ p ∣ X − z ∣ μ d z = λ μ / 2 ∫ R n ∣ u ( z ) ∣ p ∣ X − z ∣ μ d z . \int_{\mathbb{R}^n}\frac{|u_{\lambda,a}(y)|^p}{|x-y|^\mu}\,dy
=\lambda^{(n-2)p/2+\mu-n}\int_{\mathbb{R}^n}\frac{|u(z)|^p}{|X-z|^\mu}\,dz
=\lambda^{\mu/2}\int_{\mathbb{R}^n}\frac{|u(z)|^p}{|X-z|^\mu}\,dz. ∫ R n ∣ x − y ∣ μ ∣ u λ , a ( y ) ∣ p d y = λ ( n − 2 ) p /2 + μ − n ∫ R n ∣ X − z ∣ μ ∣ u ( z ) ∣ p d z = λ μ /2 ∫ R n ∣ X − z ∣ μ ∣ u ( z ) ∣ p d z .
where p = ( 2 n − μ ) / ( n − 2 ) p=(2n-\mu)/(n-2) p = ( 2 n − μ ) / ( n − 2 ) was used in the last equality. The same identity gives
∣ u λ , a ( x ) ∣ p − 2 u λ , a ( x ) = λ ( n + 2 − μ ) / 2 ∣ u ( X ) ∣ p − 2 u ( X ) . \lvert u_{\lambda,a}(x)\rvert^{p-2}u_{\lambda,a}(x)=\lambda^{(n+2-\mu)/2}\lvert u(X)\rvert^{p-2}u(X). ∣ u λ , a ( x ) ∣ p − 2 u λ , a ( x ) = λ ( n + 2 − μ ) /2 ∣ u ( X ) ∣ p − 2 u ( X ) .
Hence, both sides of the equation are multiplied by λ ( n + 2 ) / 2 \lambda^{(n+2)/2} λ ( n + 2 ) /2 , and u λ , a u_{\lambda,a} u λ , a is a weak solution for every λ > 0 \lambda>0 λ > 0 and a ∈ R n a\in\mathbb{R}^{n} a ∈ R n .
The family (9.23) is induced by smooth conformal transformations of S n \mathbb{S}^{n} S n . The C 2 , γ C^{2,\gamma} C 2 , γ -regularity of v v v gives differentiability of the corresponding parameter action in H 1 ( S n ) H^{1}(\mathbb{S}^{n}) H 1 ( S n ) , and Proposition 2.1 transfers this differentiability to D 1 , 2 ( R n ) D^{1,2}(\mathbb{R}^{n}) D 1 , 2 ( R n ) . Lemma 7.1 justifies differentiation of the weak equation with respect to the parameters. At ( λ , a ) = ( 1 , 0 ) (\lambda,a)=(1,0) ( λ , a ) = ( 1 , 0 ) ,
∂ λ u λ , 0 ∣ λ = 1 = n − 2 2 u + x ⋅ ∇ u = Z 0 , ∂ a i u 1 , a ∣ a = 0 = − ∂ x i u = − Z i . \left.\partial_{\lambda}u_{\lambda,0}\right|_{\lambda=1}=\frac{n-2}{2}u+x\cdot\nabla u=Z_{0},\qquad\left.\partial_{a_i}u_{1,a}\right|_{a=0}=-\partial_{x_i}u=-Z_i. ∂ λ u λ , 0 ∣ λ = 1 = 2 n − 2 u + x ⋅ ∇ u = Z 0 , ∂ a i u 1 , a ∣ a = 0 = − ∂ x i u = − Z i .
Consequently, Z 0 , … , Z n ∈ D 1 , 2 ( R n ) Z_{0},\ldots,Z_{n}\in D^{1,2}(\mathbb{R}^{n}) Z 0 , … , Z n ∈ D 1 , 2 ( R n ) , and differentiation of the weak equation gives
L u Z i = 0 , 0 ≤ i ≤ n . L_{u}Z_i=0,\qquad0\leq i\leq n. L u Z i = 0 , 0 ≤ i ≤ n .
To prove linear independence, suppose that
b 0 Z 0 + b ⋅ ∇ u = 0 , b = ( b 1 , … , b n ) ∈ R n . (101) b_{0}Z_{0}+b\cdot\nabla u=0,\qquad b=(b_{1},\ldots,b_{n})\in\mathbb{R}^{n}. \tag*{(101)} b 0 Z 0 + b ⋅ ∇ u = 0 , b = ( b 1 , … , b n ) ∈ R n . (101)
Assume first that b 0 = 0 b_{0}=0 b 0 = 0 . If b ≠ 0 b\neq0 b = 0 , an orthogonal change of coordinates transforms b ⋅ ∇ b\cdot\nabla b ⋅ ∇ into ∣ b ∣ ∂ x 1 \lvert b\rvert\partial_{x_{1}} ∣ b ∣ ∂ x 1 . Relation (9.24) then implies that u u u is independent of x 1 x_{1} x 1 . The Sobolev inequality gives u ∈ L 2 ∗ ( R n ) u\in L^{2^{*}}(\mathbb{R}^{n}) u ∈ L 2 ∗ ( R n ) , whereas Fubini’s theorem shows that a nonzero x 1 x_{1} x 1 -independent function cannot belong to L 2 ∗ ( R n ) L^{2^{*}}(\mathbb{R}^{n}) L 2 ∗ ( R n ) . Thus u = 0 u=0 u = 0 , contradicting the assumption that u u u is nonzero. Therefore, b = 0 b=0 b = 0 whenever b 0 = 0 b_{0}=0 b 0 = 0 .
Assume next that b 0 ≠ 0 b_{0}\neq0 b 0 = 0 , and set x 0 = − b / b 0 x_{0}=-b/b_{0} x 0 = − b / b 0 . Substitution of the definition of Z 0 Z_{0} Z 0 into (9.24) gives
( x − x 0 ) ⋅ ∇ u + n − 2 2 u = 0. (102) (x-x_{0})\cdot\nabla u+\frac{n-2}{2}u=0. \tag*{(102)} ( x − x 0 ) ⋅ ∇ u + 2 n − 2 u = 0. (102)
Along almost every ray x = x 0 + r θ x=x_{0}+r\theta x = x 0 + r θ , (9.25) becomes
r d d r u ( x 0 + r θ ) + n − 2 2 u ( x 0 + r θ ) = 0. r\frac{d}{dr}u(x_{0}+r\theta)+\frac{n-2}{2}u(x_{0}+r\theta)=0. r d r d u ( x 0 + r θ ) + 2 n − 2 u ( x 0 + r θ ) = 0.
Integration yields
u ( x 0 + r θ ) = r − ( n − 2 ) / 2 A ( θ ) . u(x_{0}+r\theta)=r^{-(n-2)/2}A(\theta). u ( x 0 + r θ ) = r − ( n − 2 ) /2 A ( θ ) .
If A A A is nonzero on a subset of S n − 1 \mathbb{S}^{n-1} S n − 1 of positive measure, then the identity ( n − 2 ) 2 ∗ / 2 = n (n-2)2^{*}/2=n ( n − 2 ) 2 ∗ /2 = n gives
∫ R n ∣ u ∣ 2 ∗ d x ≥ C ∫ 0 1 r n − 1 − ( n − 2 ) 2 ∗ / 2 d r = C ∫ 0 1 d r r = ∞ . \int_{\mathbb{R}^{n}}\lvert u\rvert^{2^{*}}\,dx\geq C\int_{0}^{1}r^{n-1-(n-2)2^{*}/2}\,dr=C\int_{0}^{1}\frac{dr}{r}=\infty. ∫ R n ∣ u ∣ 2 ∗ d x ≥ C ∫ 0 1 r n − 1 − ( n − 2 ) 2 ∗ /2 d r = C ∫ 0 1 r d r = ∞.
This contradicts u ∈ L 2 ∗ ( R n ) u\in L^{2^{*}}(\mathbb{R}^{n}) u ∈ L 2 ∗ ( R n ) . Hence, A = 0 A=0 A = 0 almost everywhere and u = 0 u=0 u = 0 , which again contradicts the nontriviality of u u u . Therefore, b 0 = 0 b_{0}=0 b 0 = 0 , and the preceding case then gives b = 0 b=0 b = 0 . Thus Z 0 , … , Z n Z_{0},\ldots,Z_{n} Z 0 , … , Z n are linearly independent.
Proposition 2.1 shows that P P P is a linear isometry preserving the full energy, including the nonlocal term. Differentiating this full energy identity twice identifies the Euclidean and spherical Hessians. Thus P − 1 Z 0 , … , P − 1 Z n P^{-1}Z_{0},\ldots,P^{-1}Z_{n} P − 1 Z 0 , … , P − 1 Z n are independent elements of the spherical kernel, and ν ( v ) ≥ n + 1 \nu(v)\geq n+1 ν ( v ) ≥ n + 1 .
For distinct coordinate labels i , j i,j i , j , write R i j = ξ i ∂ ξ j − ξ j ∂ ξ i R_{ij}=\xi_i\partial_{\xi_j}-\xi_j\partial_{\xi_i} R ij = ξ i ∂ ξ j − ξ j ∂ ξ i . The Euclidean and spherical generators are related as follows. In the stereographic chart of (18) , write ξ i = Ω x i \xi_i=\Omega x_i ξ i = Ω x i for i ≤ n i\le n i ≤ n and ξ n + 1 = ( 1 − ∣ x ∣ 2 ) / ( 1 + ∣ x ∣ 2 ) \xi_{n+1}=(1-|x|^2)/(1+|x|^2) ξ n + 1 = ( 1 − ∣ x ∣ 2 ) / ( 1 + ∣ x ∣ 2 ) . Direct differentiation gives
D π − 1 ( x ) ∂ x i = ∇ S ξ i − R i , n + 1 , D π − 1 ( x ) x = − ∇ S ξ n + 1 , D\pi^{-1}(x)\partial_{x_i}=\nabla^{\mathbb S}\xi_i-R_{i,n+1},\qquad D\pi^{-1}(x)x=-\nabla^{\mathbb S}\xi_{n+1}, D π − 1 ( x ) ∂ x i = ∇ S ξ i − R i , n + 1 , D π − 1 ( x ) x = − ∇ S ξ n + 1 ,
and
∂ x i log Ω = − ξ i , 1 + x ⋅ ∇ log Ω = ξ n + 1 . \partial_{x_i}\log\Omega=-\xi_i,\qquad1+x\cdot\nabla\log\Omega=\xi_{n+1}. ∂ x i log Ω = − ξ i , 1 + x ⋅ ∇ log Ω = ξ n + 1 .
Applying the product rule to u ( x ) = Ω ( x ) ( n − 2 ) / 2 v ( π − 1 ( x ) ) u(x)=\Omega(x)^{(n-2)/2}v(\pi^{-1}(x)) u ( x ) = Ω ( x ) ( n − 2 ) /2 v ( π − 1 ( x )) therefore gives
P − 1 Z i = Φ e i − R i , n + 1 v ( 1 ≤ i ≤ n ) , P − 1 Z 0 = − Φ e n + 1 . (103) P^{-1}Z_i=\Phi_{e_i}-R_{i,n+1}v\quad(1\le i\le n),\qquad P^{-1}Z_0=-\Phi_{e_{n+1}}. \tag*{(103)} P − 1 Z i = Φ e i − R i , n + 1 v ( 1 ≤ i ≤ n ) , P − 1 Z 0 = − Φ e n + 1 . (103)
Here e i e_i e i denotes the i i i -th coordinate vector. For an invariant nonconstant v v v , the terms R i , n + 1 v R_{i,n+1}v R i , n + 1 v with i ≤ k i\le k i ≤ k belong to the sector ( 1 , 1 ) (1,1) ( 1 , 1 ) . Thus individual pulled-back Euclidean translations need not have a pure degree-one angular profile.
If v ∈ H v\in H v ∈ H is nonconstant, Proposition 9.1 gives the dimension of the conformal orbit. The two gradient families lie in the ( 1 , 0 ) (1,0) ( 1 , 0 ) and ( 0 , 1 ) (0,1) ( 0 , 1 ) sectors, with total dimension k + ℓ = n + 1 k+\ell=n+1 k + ℓ = n + 1 . The mixed rotations occupy the ( 1 , 1 ) (1,1) ( 1 , 1 ) sector and have dimension k ℓ k\ell k ℓ . These three angular sectors are mutually orthogonal. Consequently,
ν ( v ) ≥ n + 1 + k ℓ , dim K ( v ) = n + 1 + k ℓ . \nu(v)\ge n+1+k\ell,\qquad\dim\mathcal{K}(v)=n+1+k\ell. ν ( v ) ≥ n + 1 + k ℓ , dim K ( v ) = n + 1 + k ℓ .
For the same nonconstant G G G -invariant solution v v v , it remains to prove (16) . Set F = ∣ v ∣ p F=|v|^p F = ∣ v ∣ p and use the test functions ϕ i = ξ i v \phi_i=\xi_i v ϕ i = ξ i v for 1 ≤ i ≤ n + 1 1\le i\le n+1 1 ≤ i ≤ n + 1 . Testing the equation against ξ i 2 v \xi_i^2v ξ i 2 v gives
∥ ξ i v ∥ 2 = ∫ S n V v ξ i 2 v 2 d V + ∫ S n v 2 ∣ ∇ ξ i ∣ 2 d V . \|\xi_i v\|^2=\int_{\mathbb S^n}V_v\xi_i^2v^2\,dV+\int_{\mathbb S^n}v^2|\nabla\xi_i|^2\,dV. ∥ ξ i v ∥ 2 = ∫ S n V v ξ i 2 v 2 d V + ∫ S n v 2 ∣∇ ξ i ∣ 2 d V .
Since ∑ i ξ i 2 = 1 \sum_i\xi_i^2=1 ∑ i ξ i 2 = 1 , ∑ i ∣ ∇ ξ i ∣ 2 = n \sum_i|\nabla\xi_i|^2=n ∑ i ∣∇ ξ i ∣ 2 = n , and ξ ⋅ η = 1 − 1 2 ∣ ξ − η ∣ 2 \xi\cdot\eta=1-\frac12|\xi-\eta|^2 ξ ⋅ η = 1 − 2 1 ∣ ξ − η ∣ 2 , summing (54) yields
∑ i Q v ( ξ i v ) = n ∥ v ∥ 2 2 − ( 2 p − 2 ) D ( v ) + p 2 ∬ S n × S n ∣ ξ − η ∣ 2 − μ F ( ξ ) F ( η ) d V ξ d V η . \sum_i Q_v(\xi_i v)=n\|v\|_2^2-(2p-2)D(v)+\frac{p}{2}\iint_{\mathbb S^n\times\mathbb S^n}|\xi-\eta|^{2-\mu}F(\xi)F(\eta)\,dV_\xi dV_\eta. i ∑ Q v ( ξ i v ) = n ∥ v ∥ 2 2 − ( 2 p − 2 ) D ( v ) + 2 p ∬ S n × S n ∣ ξ − η ∣ 2 − μ F ( ξ ) F ( η ) d V ξ d V η .
Applying [17 Proposition 4.1] in dimension n n n and with α = μ / 2 \alpha=\mu/2 α = μ /2 gives
∬ S n × S n F ( ξ ) ( ξ ⋅ η ) F ( η ) ∣ ξ − η ∣ μ d V ξ d V η ≥ μ 2 n − μ D ( v ) . \iint_{\mathbb S^n\times\mathbb S^n}\frac{F(\xi)(\xi\cdot\eta)F(\eta)}{|\xi-\eta|^\mu}\,dV_\xi dV_\eta\ge\frac{\mu}{2n-\mu}D(v). ∬ S n × S n ∣ ξ − η ∣ μ F ( ξ ) ( ξ ⋅ η ) F ( η ) d V ξ d V η ≥ 2 n − μ μ D ( v ) .
Since ∣ ξ − η ∣ 2 = 2 ( 1 − ξ ⋅ η ) |\xi-\eta|^2=2(1-\xi\cdot\eta) ∣ ξ − η ∣ 2 = 2 ( 1 − ξ ⋅ η ) ,
∬ S n × S n ∣ ξ − η ∣ 2 − μ F ( ξ ) F ( η ) d V ξ d V η ≤ 2 ( 1 − μ 2 n − μ ) D ( v ) = 4 ( n − μ ) 2 n − μ D ( v ) . \iint_{\mathbb S^n\times\mathbb S^n}|\xi-\eta|^{2-\mu}F(\xi)F(\eta)\,dV_\xi dV_\eta\le2\left(1-\frac{\mu}{2n-\mu}\right)D(v)=\frac{4(n-\mu)}{2n-\mu}D(v). ∬ S n × S n ∣ ξ − η ∣ 2 − μ F ( ξ ) F ( η ) d V ξ d V η ≤ 2 ( 1 − 2 n − μ μ ) D ( v ) = 2 n − μ 4 ( n − μ ) D ( v ) .
Using p = ( 2 n − μ ) / ( n − 2 ) p=(2n-\mu)/(n-2) p = ( 2 n − μ ) / ( n − 2 ) and (28) gives
∑ i Q v ( ξ i v ) ≤ n ∥ v ∥ 2 2 − 4 n − 2 ∥ v ∥ 2 2 = − 4 n − 2 ∫ S n ∣ ∇ v ∣ 2 d V . (104) \sum_i Q_v(\xi_i v)\le n\|v\|_2^2-\frac{4}{n-2}\|v\|_2^2=-\frac{4}{n-2}\int_{\mathbb S^n}|\nabla v|^2\,dV. \tag*{(104)} i ∑ Q v ( ξ i v ) ≤ n ∥ v ∥ 2 2 − n − 2 4 ∥ v ∥ 2 2 = − n − 2 4 ∫ S n ∣∇ v ∣ 2 d V . (104)
The right-hand side of (104) is negative for a nonconstant solution. The test functions belong to ( 1 , 0 ) (1,0) ( 1 , 0 ) or ( 0 , 1 ) (0,1) ( 0 , 1 ) , so at least one of these radial blocks has a negative direction. Using m 0 , 0 ( v ) = m G ( v ) m_{0,0}(v)=m_G(v) m 0 , 0 ( v ) = m G ( v ) and the angular multiplicities in (86) gives
m ( v ) ≥ m G ( v ) + k m 1 , 0 ( v ) + ℓ m 0 , 1 ( v ) ≥ m G ( v ) + min { k , ℓ } . m(v)\ge m_G(v)+k\,m_{1,0}(v)+\ell\,m_{0,1}(v)\ge m_G(v)+\min\{k,\ell\}. m ( v ) ≥ m G ( v ) + k m 1 , 0 ( v ) + ℓ m 0 , 1 ( v ) ≥ m G ( v ) + min { k , ℓ } .
This proves (16) . When k = ℓ k=\ell k = ℓ and v ( T ξ ) = ± v ( ξ ) v(T\xi)=\pm v(\xi) v ( T ξ ) = ± v ( ξ ) , the factor interchange identifies the two blocks. Both then contain a negative direction.
For v = ± a 0 v=\pm a_0 v = ± a 0 , Proposition 9.5 shows that the full kernel is exactly the ( n + 1 ) (n+1) ( n + 1 ) -dimensional conformal tangent space. This completes the proof of Theorem 1.6 .
Remark 9.6. Assume p ≥ 2 p\geq2 p ≥ 2 , and let j → ∞ j\to\infty j → ∞ through labels for which the isolation hypotheses of Theorem 1.4 hold. For each such label, choose the solution v j v_j v j supplied by that theorem. Then
0 ≤ j − m G ( v j ) ≤ ν G ( v j ) , N G ( v j ) ≤ m G ( v j ) ≤ m ( v j ) . 0\leq j-m_G(v_j)\leq\nu_G(v_j),\qquad N_G(v_j)\leq m_G(v_j)\leq m(v_j). 0 ≤ j − m G ( v j ) ≤ ν G ( v j ) , N G ( v j ) ≤ m G ( v j ) ≤ m ( v j ) .
If ν G ( v j ) = o ( j ) \nu_G(v_j)=o(j) ν G ( v j ) = o ( j ) along these labels, then m G ( v j ) / j → 1 m_G(v_j)/j\to1 m G ( v j ) / j → 1 . Consequently,
N G ( v j ) j ⟶ 1 ⟺ m G ( v j ) − N G ( v j ) = o ( j ) , \frac{N_G(v_j)}{j}\longrightarrow1\quad\Longleftrightarrow\quad m_G(v_j)-N_G(v_j)=o(j), j N G ( v j ) ⟶ 1 ⟺ m G ( v j ) − N G ( v j ) = o ( j ) ,
and
m ( v j ) j ⟶ 1 ⟺ m ( v j ) − m G ( v j ) = o ( j ) . \frac{m(v_j)}{j}\longrightarrow1\quad\Longleftrightarrow\quad m(v_j)-m_G(v_j)=o(j). j m ( v j ) ⟶ 1 ⟺ m ( v j ) − m G ( v j ) = o ( j ) .
By (86) ,
m ( v j ) − m G ( v j ) = ∑ a , b ≥ 0 ( a , b ) ≠ ( 0 , 0 ) d k , a d ℓ , b m a , b ( v j ) , m(v_j)-m_G(v_j)=\sum_{\substack{a,b\geq0\\(a,b)\neq(0,0)}}d_{k,a}d_{\ell,b}m_{a,b}(v_j), m ( v j ) − m G ( v j ) = a , b ≥ 0 ( a , b ) = ( 0 , 0 ) ∑ d k , a d ℓ , b m a , b ( v j ) ,
which is the total contribution of the noninvariant angular sectors. Moreover, Proposition 9.4 gives
m ( v j ) ≤ C ( 1 + ∥ v j ∥ L ∞ ( S n ) n ( p − 1 ) ) . m(v_j)\leq C\left(1+\|v_j\|_{L^\infty(\mathbb{S}^n)}^{n(p-1)}\right). m ( v j ) ≤ C ( 1 + ∥ v j ∥ L ∞ ( S n ) n ( p − 1 ) ) .
The low blocks for equal factors
Proof of Theorem 1.7. Let k = ℓ k=\ell k = ℓ , so n = 2 k − 1 n=2k-1 n = 2 k − 1 . Write p μ p_\mu p μ , E μ E_\mu E μ , and D μ D_\mu D μ for the parameter-dependent quantities. At μ = 0 \mu=0 μ = 0 , set p 0 = 2 ∗ p_0=2^* p 0 = 2 ∗ and J 0 h = ∫ S n h d V J_0h=\int_{\mathbb{S}^n}h\,dV J 0 h = ∫ S n h d V .
Minimizers and their limit. On S ∩ H − \mathcal{S}\cap H^- S ∩ H − , minimizing Ψ \Psi Ψ is equivalent to maximizing D μ D_\mu D μ . A normalized multiple of ∣ ξ ′ ∣ 2 − ∣ ξ ′ ′ ∣ 2 |\xi'|^2-|\xi''|^2 ∣ ξ ′ ∣ 2 − ∣ ξ ′′ ∣ 2 belongs to this sphere. A maximizing sequence has a subsequence converging weakly in H − H^- H − and strongly in L 2 ∗ L^{2^*} L 2 ∗ . The interaction converges by HLS for μ > 0 \mu>0 μ > 0 and directly for μ = 0 \mu=0 μ = 0 . The limiting mass is positive, so the limit is nonzero. If its norm were less than one, normalization would strictly increase the mass. Hence it is a maximizer on S ∩ H − \mathcal{S}\cap H^- S ∩ H − .
Rescale a maximizer by (2.9). The proof of Lemma 2.5 applies on H − H^- H − . The energy is invariant under v ( ξ ) ↦ − v ( T ξ ) v(\xi)\mapsto-v(T\xi) v ( ξ ) ↦ − v ( T ξ ) , which fixes the rescaled function. Averaging a variation under this involution and G G G reduces it to a variation in H − H^- H − . Thus the rescaled function is a full solution. In the orbit variable, T T T acts as t ↦ π / 2 − t t\mapsto\pi/2-t t ↦ π /2 − t . Put t ∗ = π / 4 t_*=\pi/4 t ∗ = π /4 . Every profile in H − H^- H − has zero trace at t ∗ t_* t ∗ . Replacing its profile by the positive absolute value on ( 0 , t ∗ ) (0,t_*) ( 0 , t ∗ ) and the negative absolute value on ( t ∗ , π / 2 ) (t_*,\pi/2) ( t ∗ , π /2 ) preserves its norm and its mass. The matching zero traces give an H 1 H^1 H 1 function. Consequently, the resulting maximizer also yields a solution.
For μ > 0 \mu>0 μ > 0 small enough that p μ > 2 p_\mu>2 p μ > 2 , the folded profile is nonnegative on the first half. An interior zero there would be simple by Lemma 6.3 and would force a sign change, a contradiction. Endpoint zeros are excluded by the same lemma. Hence the folded profile is strictly positive on [ 0 , t ∗ ) [0,t_*) [ 0 , t ∗ ) and has exactly one simple zero at t ∗ t_* t ∗ . Any original maximizer has the same absolute value as its folded profile and is continuous, so it has this sign pattern up to an overall sign. In the limit problem, any rescaled maximizer v ( 0 ) v^{(0)} v ( 0 ) solves
A v ( 0 ) = B 0 ∣ v ( 0 ) ∣ p 0 − 2 v ( 0 ) , B 0 = ∫ S n ∣ v ( 0 ) ∣ p 0 d V > 0. A v^{(0)} = B_0 |v^{(0)}|^{p_0-2}v^{(0)}, \qquad B_0 = \int_{\mathbb{S}^n} |v^{(0)}|^{p_0}\,dV > 0. A v ( 0 ) = B 0 ∣ v ( 0 ) ∣ p 0 − 2 v ( 0 ) , B 0 = ∫ S n ∣ v ( 0 ) ∣ p 0 d V > 0.
By Lemma 2.2, v ( 0 ) ∈ L R v^{(0)} \in L^R v ( 0 ) ∈ L R for some R > 2 ∗ R > 2^* R > 2 ∗ . Its potential B 0 ∣ v ( 0 ) ∣ p 0 − 2 B_0|v^{(0)}|^{p_0-2} B 0 ∣ v ( 0 ) ∣ p 0 − 2 is then in L r 0 L^{r_0} L r 0 with r 0 > n / 2 r_0 > n/2 r 0 > n /2 . Elliptic L s L^s L s estimates and Sobolev embedding give boundedness. In each subcritical step the reciprocal integrability exponent of v ( 0 ) v^{(0)} v ( 0 ) decreases by 2 / n − 1 / r 0 > 0 2/n-1/r_0>0 2/ n − 1/ r 0 > 0 . At a critical step one may take any finite exponent before applying the estimate again. Schauder regularity gives C 2 , γ C^{2,\gamma} C 2 , γ . The profile equation and Lemma 6.3 apply with the constant potential B 0 B_0 B 0 , so the same sign pattern holds at μ = 0 \mu=0 μ = 0 .
Fix R > p 0 R>p_0 R > p 0 for which the embedding H ↪ L R H\hookrightarrow L^R H ↪ L R is compact. Choose r > 1 r>1 r > 1 close enough to one that
1 r > max { p 0 R , p 0 − 1 R + 1 2 ∗ } . \frac{1}{r}>\max\left\{\frac{p_0}{R},\frac{p_0-1}{R}+\frac{1}{2^*}\right\}. r 1 > max { R p 0 , R p 0 − 1 + 2 ∗ 1 } .
Choose 0 < μ ˉ < min { 4 , n } 0<\bar{\mu}<\min\{4,n\} 0 < μ ˉ < min { 4 , n } so small that μ ˉ r ′ < n \bar{\mu}r'<n μ ˉ r ′ < n , where r ′ = r / ( r − 1 ) r'=r/(r-1) r ′ = r / ( r − 1 ) . For 0 ≤ μ ≤ μ ˉ 0\leq\mu\leq\bar{\mu} 0 ≤ μ ≤ μ ˉ , the operators J μ J_\mu J μ are uniformly bounded from L r L^r L r to L ∞ L^\infty L ∞ . In fact,
∥ J μ − J 0 ∥ L r → L ∞ ≤ ( ∫ S n ∣ ξ − η ∣ − μ − 1 ∣ r ′ d V η ) 1 / r ′ ⟶ 0 \|J_\mu-J_0\|_{L^r\to L^\infty}\leq\left(\int_{\mathbb{S}^n}|\xi-\eta|^{-\mu}-1|^{r'}\,dV_\eta\right)^{1/r'}\longrightarrow0 ∥ J μ − J 0 ∥ L r → L ∞ ≤ ( ∫ S n ∣ ξ − η ∣ − μ − 1 ∣ r ′ d V η ) 1/ r ′ ⟶ 0
by rotation invariance and dominated convergence.
Let μ i → 0 \mu_i\to0 μ i → 0 and let w i w_i w i be corresponding maximizers with the positive sign on the first half. Pass to a subsequence with w i ⇀ w w_i\rightharpoonup w w i ⇀ w in H H H and w i → w w_i\to w w i → w in L R L^R L R . We will use the following fact. If z i → z z_i\to z z i → z in L R L^R L R , s i → s > 0 s_i\to s>0 s i → s > 0 with s i ≤ s s_i\leq s s i ≤ s , and s q < R sq<R s q < R , then ∣ z i ∣ s i → ∣ z ∣ s |z_i|^{s_i}\to|z|^s ∣ z i ∣ s i → ∣ z ∣ s in L q L^q L q , also with a sign factor. Indeed, there is convergence in measure and the powers are bounded in L q + ϵ L^{q+\epsilon} L q + ϵ for small ϵ > 0 \epsilon>0 ϵ > 0 . In particular, ∣ w i ∣ p μ i → ∣ w ∣ p 0 |w_i|^{p_{\mu_i}}\to|w|^{p_0} ∣ w i ∣ p μ i → ∣ w ∣ p 0 in L r L^r L r and D μ i ( w i ) → D 0 ( w ) D_{\mu_i}(w_i)\to D_0(w) D μ i ( w i ) → D 0 ( w ) . If M 0 = max S ∩ H − D 0 M_0=\max_{S\cap H^-}D_0 M 0 = max S ∩ H − D 0 , comparison with a fixed maximizing profile at μ = 0 \mu=0 μ = 0 gives D 0 ( w ) ≥ M 0 D_0(w)\geq M_0 D 0 ( w ) ≥ M 0 . On the other hand, D 0 ( w ) ≤ M 0 ∥ w ∥ 2 p 0 ≤ M 0 D_0(w)\leq M_0\|w\|^{2p_0}\leq M_0 D 0 ( w ) ≤ M 0 ∥ w ∥ 2 p 0 ≤ M 0 . Thus w i → w w_i\to w w i → w in H H H , and w w w is a maximizer at zero. The strong L R L^R L R convergence preserves nonnegativity on the first half, and w ∈ H − w\in H^- w ∈ H − . The regularity and zero argument for the limit equation therefore give the same strict sign pattern for its rescaling v ( 0 ) v^{(0)} v ( 0 ) . The rescaled solutions converge strongly in H H H and in L R L^R L R to
v ( 0 ) = D 0 ( w ) − 1 / ( 2 p 0 − 2 ) w . v^{(0)}=D_0(w)^{-1/(2p_0-2)}w. v ( 0 ) = D 0 ( w ) − 1/ ( 2 p 0 − 2 ) w .
The three limiting radial operators. The nonlocal term of the Hessian at μ = 0 \mu=0 μ = 0 vanishes on every sector ( a , b ) ≠ ( 0 , 0 ) (a,b)\ne(0,0) ( a , b ) = ( 0 , 0 ) . The angular average of a v ( 0 ) g Y a Z b a_v^{(0)}gY_aZ_b a v ( 0 ) g Y a Z b is zero, and J 0 J_0 J 0 is integration. Thus these radial forms are local, with the common potential ( p 0 − 1 ) B 0 ∣ f ∣ p 0 − 2 (p_0-1)B_0|f|^{p_0-2} ( p 0 − 1 ) B 0 ∣ f ∣ p 0 − 2 , where f f f is the profile of v ( 0 ) v^{(0)} v ( 0 ) . The symmetry profiles from Proposition 9.1 still solve the kernel equations at μ = 0 \mu=0 μ = 0 , since its conformal computation also holds for p = p 0 p=p_0 p = p 0 and J 0 J_0 J 0 .
Consider
E f ( t ) = 1 2 f ′ ( t ) 2 + B 0 p 0 f ( t ) p 0 − n ( n − 2 ) 8 f ( t ) 2 , 0 ≤ t ≤ t ∗ . \mathcal{E}_f(t)=\frac{1}{2}f'(t)^2+\frac{B_0}{p_0}f(t)^{p_0}-\frac{n(n-2)}{8}f(t)^2,\qquad0\leq t\leq t_*. E f ( t ) = 2 1 f ′ ( t ) 2 + p 0 B 0 f ( t ) p 0 − 8 n ( n − 2 ) f ( t ) 2 , 0 ≤ t ≤ t ∗ .
On the open first half, E f ′ = − ( ρ ′ / ρ ) f ′ 2 ≤ 0 \mathcal{E}_f' =-(\rho'/\rho)f'^2\leq0 E f ′ = − ( ρ ′ / ρ ) f ′2 ≤ 0 , since k = ℓ k=\ell k = ℓ . At t ∗ t_* t ∗ , E f ( t ∗ ) = f ′ ( t ∗ ) 2 / 2 > 0 \mathcal{E}_f(t_*)=f'(t_*)^2/2>0 E f ( t ∗ ) = f ′ ( t ∗ ) 2 /2 > 0 . At any stationary point on [ 0 , t ∗ ) [0,t_*) [ 0 , t ∗ ) we therefore have
B 0 f p 0 − 2 > p 0 n ( n − 2 ) 8 > n ( n − 2 ) 4 . B_0f^{p_0-2}>\frac{p_0n(n-2)}{8}>\frac{n(n-2)}{4}. B 0 f p 0 − 2 > 8 p 0 n ( n − 2 ) > 4 n ( n − 2 ) .
The profile equation implies f ′ ′ < 0 f'' < 0 f ′′ < 0 at every interior stationary point. At zero, f ′ ( 0 ) = 0 f'(0) = 0 f ′ ( 0 ) = 0 and taking the limit in (50) gives
k f ′ ′ ( 0 ) = n ( n − 2 ) 4 f ( 0 ) − B 0 f ( 0 ) p 0 − 1 < 0. k f''(0) = \frac{n(n-2)}{4} f(0) - B_0 f(0)^{p_0-1} < 0. k f ′′ ( 0 ) = 4 n ( n − 2 ) f ( 0 ) − B 0 f ( 0 ) p 0 − 1 < 0.
Thus f ′ < 0 f' < 0 f ′ < 0 near zero. A first interior zero of f ′ f' f ′ would have f ′ ′ ≥ 0 f'' \ge0 f ′′ ≥ 0 , so f ′ < 0 f' < 0 f ′ < 0 on the first half. Oddness about t ∗ t_* t ∗ and the simple zero at t ∗ t_* t ∗ give f ′ < 0 f' < 0 f ′ < 0 on the whole open interval.
In the mixed sector, g = − f ′ > 0 g=-f' > 0 g = − f ′ > 0 is a kernel function of the local radial operator. For a compactly supported test profile ϕ \phi ϕ , integration by parts in the equation for g g g gives
Q 1 , 1 ( 0 ) ( ϕ ) = ∫ 0 π / 2 ρ g 2 ∣ ( ϕ / g ) ′ ∣ 2 d t . Q^{(0)}_{1,1}(\phi)=\int_0^{\pi/2}\rho g^2\left|(\phi/g)'\right|^2\,dt. Q 1 , 1 ( 0 ) ( ϕ ) = ∫ 0 π /2 ρ g 2 ∣ ( ϕ / g ) ′ ∣ 2 d t .
Density in the form norm gives nonnegativity. If ϕ \phi ϕ is in the kernel, approximate it by compactly supported profiles in that norm. On every compact subinterval the quotient by g g g converges in H 1 H^1 H 1 , so the last identity implies ( ϕ / g ) ′ = 0 (\phi/g)' = 0 ( ϕ / g ) ′ = 0 there. Hence the mixed block has index zero and nullity one.
In the ( 1 , 0 ) (1,0) ( 1 , 0 ) sector, the profile g 1 , 0 g_{1,0} g 1 , 0 in (9.1) has exactly one interior zero. On 0 < t < t ∗ 0<t<t_* 0 < t < t ∗ , set h = ( n − 2 ) / 2 h=(n-2)/2 h = ( n − 2 ) /2 and y = ( sin t ) h f y=(\sin t)^h f y = ( sin t ) h f . Since h = k − 3 / 2 h=k-3/2 h = k − 3/2 , n ( n − 2 ) / 4 = h ( h + 1 ) n(n-2)/4=h(h+1) n ( n − 2 ) /4 = h ( h + 1 ) , and
ρ ′ ρ = ( k − 1 ) ( cot t − tan t ) , \frac{\rho'}{\rho}=(k-1)(\cot t-\tan t), ρ ρ ′ = ( k − 1 ) ( cot t − tan t ) ,
f ′ = ( sin t ) − h ( y ′ − h cot t y ) , f'=(\sin t)^{-h}(y'-h\cot t\,y), f ′ = ( sin t ) − h ( y ′ − h cot t y ) ,
f ′ ′ = ( sin t ) − h [ y ′ ′ − 2 h cot t y ′ + ( h sin 2 t + h 2 cot 2 t ) y ] . f''=(\sin t)^{-h}\left[y''-2h\cot t\,y'+\left(\frac{h}{\sin^2t}+h^2\cot^2t\right)y\right]. f ′′ = ( sin t ) − h [ y ′′ − 2 h cot t y ′ + ( sin 2 t h + h 2 cot 2 t ) y ] .
The coefficient of y y y in the transformed linear part is
h sin 2 t + h 2 cot 2 t − h ρ ′ ρ cot t − n ( n − 2 ) 4 = h sin 2 t + ( h 2 − h ( k − 1 ) ) cot 2 t + h ( k − 1 ) − n ( n − 2 ) 4 = h 2 sin 2 t . \begin{aligned}
\frac{h}{\sin^2t}+h^2\cot^2t-h\frac{\rho'}{\rho}\cot t-\frac{n(n-2)}{4}
=\frac{h}{\sin^2t}+\left(h^2-h(k-1)\right)\cot^2t
+h(k-1)-\frac{n(n-2)}{4} \\
&=\frac{h}{2\sin^2t}.
\end{aligned} sin 2 t h + h 2 cot 2 t − h ρ ρ ′ cot t − 4 n ( n − 2 ) = sin 2 t h + ( h 2 − h ( k − 1 ) ) cot 2 t + h ( k − 1 ) − 4 n ( n − 2 ) = 2 sin 2 t h .
Here we used h 2 − h ( k − 1 ) = − h / 2 h^2-h(k-1)=-h/2 h 2 − h ( k − 1 ) = − h /2 and h ( k − 1 ) − n ( n − 2 ) / 4 = − h / 2 h(k-1)-n(n-2)/4=-h/2 h ( k − 1 ) − n ( n − 2 ) /4 = − h /2 . Also h ( p 0 − 2 ) = 2 h(p_0-2)=2 h ( p 0 − 2 ) = 2 . The equation for f f f therefore gives
y ′ ′ + ( ρ ′ ρ − 2 h cot t ) y ′ + ( h / 2 ) y + B 0 y p 0 − 1 sin 2 t = 0. y''+\left(\frac{\rho'}{\rho}-2h\cot t\right)y'
+\frac{(h/2)y+B_0y^{p_0-1}}{\sin^2t}=0. y ′′ + ( ρ ρ ′ − 2 h cot t ) y ′ + sin 2 t ( h /2 ) y + B 0 y p 0 − 1 = 0.
The function y y y is positive on this half and vanishes at both ends. Since f ( 0 ) > 0 f(0)>0 f ( 0 ) > 0 , f ′ ( 0 ) = 0 f'(0)=0 f ′ ( 0 ) = 0 , and h > 0 h>0 h > 0 ,
y ′ ( t ) = h f ( 0 ) t h − 1 + o ( t h − 1 ) > 0 y'(t)=h f(0)t^{h-1}+o(t^{h-1})>0 y ′ ( t ) = h f ( 0 ) t h − 1 + o ( t h − 1 ) > 0
for sufficiently small t > 0 t>0 t > 0 . At the other endpoint, y ′ ( t ∗ ) = ( sin t ∗ ) h f ′ ( t ∗ ) < 0 y'(t_*)=(\sin t_*)^h f'(t_*)<0 y ′ ( t ∗ ) = ( sin t ∗ ) h f ′ ( t ∗ ) < 0 . Thus y ′ y' y ′ has a zero. At each zero the transformed equation gives
y ′ ′ = − ( h / 2 ) y + B 0 y p 0 − 1 sin 2 t < 0. y''=-\frac{(h/2)y+B_0y^{p_0-1}}{\sin^2t}<0. y ′′ = − sin 2 t ( h /2 ) y + B 0 y p 0 − 1 < 0.
Every zero of y ′ y' y ′ is consequently a strict crossing from positive to negative. A second zero would require a crossing in the opposite direction in between. Hence there is exactly one stationary point. Since g 1 , 0 = − ( sin t ) 1 − h y ′ g_{1,0}=-(\sin t)^{1-h}y' g 1 , 0 = − ( sin t ) 1 − h y ′ , it has exactly one zero there, and its derivative at that point is
g 1 , 0 ′ = − ( sin t ) 1 − h y ′ ′ > 0. g'_{1,0}=-(\sin t)^{1-h}y''>0. g 1 , 0 ′ = − ( sin t ) 1 − h y ′′ > 0.
The zero is therefore simple. For t ∗ < t < π / 2 t_{*}<t<\pi/2 t ∗ < t < π /2 both f < 0 f<0 f < 0 and f ′ < 0 f'<0 f ′ < 0 , so g 1 , 0 > 0 g_{1,0}>0 g 1 , 0 > 0 there. It is also positive at t ∗ t_{*} t ∗ .
Let t 1 t_{1} t 1 be the zero of g 1 , 0 g_{1,0} g 1 , 0 . On a closed interval containing t 1 t_{1} t 1 in its interior and contained in ( 0 , π / 2 ) (0,\pi/2) ( 0 , π /2 ) , the weight ρ \rho ρ is bounded above and below by positive constants. The one-dimensional trace inequality therefore gives
∣ ϕ ( t 1 ) ∣ ≤ C ∥ ϕ ∥ X 1 , 0 . |\phi(t_{1})|\leq C\|\phi\|_{\mathcal{X}_{1,0}}. ∣ ϕ ( t 1 ) ∣ ≤ C ∥ ϕ ∥ X 1 , 0 .
This continuous linear functional is nonzero on smooth profiles supported near t 1 t_{1} t 1 . Its kernel is a closed subspace of codimension one. On this subspace, the form Q 1 , 0 ( 0 ) Q_{1,0}^{(0)} Q 1 , 0 ( 0 ) is nonnegative. Apply the previous ground state identity separately on each side of t 1 t_{1} t 1 , using ∣ g 1 , 0 ∣ |g_{1,0}| ∣ g 1 , 0 ∣ . To pass to the form domain, cut off near t 1 t_{1} t 1 . For ϕ ( t 1 ) = 0 \phi(t_{1})=0 ϕ ( t 1 ) = 0 , the derivative cost of this cutoff is bounded by
C ε − 2 ∫ ∣ t − t 1 ∣ < 2 ε ∣ ϕ ( t ) ∣ 2 d t ≤ C ∫ ∣ t − t 1 ∣ < 2 ε ∣ ϕ ′ ( t ) ∣ 2 d t ⟶ 0. C\varepsilon^{-2}\int_{|t-t_{1}|<2\varepsilon}|\phi(t)|^{2}\,dt\leq C\int_{|t-t_{1}|<2\varepsilon}|\phi'(t)|^{2}\,dt\longrightarrow0. C ε − 2 ∫ ∣ t − t 1 ∣ < 2 ε ∣ ϕ ( t ) ∣ 2 d t ≤ C ∫ ∣ t − t 1 ∣ < 2 ε ∣ ϕ ′ ( t ) ∣ 2 d t ⟶ 0.
The endpoint cutoff argument in § 9.2 then gives the required density on both subintervals. Thus m 1 , 0 ( 0 ) ≤ 1 m_{1,0}^{(0)}\leq1 m 1 , 0 ( 0 ) ≤ 1 . The mixed kernel profile belongs to X 1 , 1 ⊂ X 1 , 0 \mathcal{X}_{1,1}\subset\mathcal{X}_{1,0} X 1 , 1 ⊂ X 1 , 0 , because the ( 1 , 1 ) (1,1) ( 1 , 1 ) local form contains all the nonnegative terms defining X 1 , 0 \mathcal{X}_{1,0} X 1 , 0 . Thus f ′ f' f ′ is an admissible test in ( 1 , 0 ) (1,0) ( 1 , 0 ) , and the two local forms differ by the angular potential ( k − 1 ) / sin 2 t (k-1)/\sin^{2}t ( k − 1 ) / sin 2 t . Consequently,
Q 1 , 0 ( 0 ) ( f ′ ) = Q 1 , 1 ( 0 ) ( f ′ ) − ( k − 1 ) ∫ 0 π / 2 ∣ f ′ ∣ 2 sin 2 t ρ d t < 0 , Q_{1,0}^{(0)}(f')=Q_{1,1}^{(0)}(f')-(k-1)\int_{0}^{\pi/2}\frac{|f'|^{2}}{\sin^{2}t}\rho\,dt<0, Q 1 , 0 ( 0 ) ( f ′ ) = Q 1 , 1 ( 0 ) ( f ′ ) − ( k − 1 ) ∫ 0 π /2 sin 2 t ∣ f ′ ∣ 2 ρ d t < 0 ,
so m 1 , 0 ( 0 ) = 1 m_{1,0}^{(0)}=1 m 1 , 0 ( 0 ) = 1 .
Let ϕ \phi ϕ be another kernel profile in ( 1 , 0 ) (1,0) ( 1 , 0 ) . Near zero g 1 , 0 g_{1,0} g 1 , 0 is bounded away from zero, and the local equations give ρ g 1 , 0 2 ( ϕ / g 1 , 0 ) ′ = C \rho g_{1,0}^{2}(\phi/g_{1,0})'=C ρ g 1 , 0 2 ( ϕ / g 1 , 0 ) ′ = C for a constant C C C . The derivative of this quotient lies in L 2 ( ρ d t ) L^{2}(\rho\,dt) L 2 ( ρ d t ) near zero, since g 1 , 0 g_{1,0} g 1 , 0 is C 1 C^{1} C 1 and ϕ ∈ X 1 , 0 \phi\in\mathcal{X}_{1,0} ϕ ∈ X 1 , 0 . If C ≠ 0 C\ne0 C = 0 , its squared integral is infinite because ∫ 0 ε ρ − 1 d t = ∞ \int_{0}^{\varepsilon}\rho^{-1}\,dt=\infty ∫ 0 ε ρ − 1 d t = ∞ for k ≥ 2 k\geq2 k ≥ 2 . Hence ϕ \phi ϕ is proportional to g 1 , 0 g_{1,0} g 1 , 0 near zero, and ODE uniqueness extends this identity across the whole interval. Reflection about t ∗ t_{*} t ∗ identifies the ( 0 , 1 ) (0,1) ( 0 , 1 ) and ( 1 , 0 ) (1,0) ( 1 , 0 ) forms, so their indices and nullities agree.
Persistence for small positive μ \mu μ . Along the convergent subsequence above, set v i = v μ i v_{i}=v_{\mu_{i}} v i = v μ i and U i = J μ i ( ∣ v i ∣ p μ i ) U_{i}=J_{\mu_{i}}(|v_{i}|^{p_{\mu_{i}}}) U i = J μ i ( ∣ v i ∣ p μ i ) . The choice of r r r and the power convergence show U i → B 0 U_{i}\to B_{0} U i → B 0 uniformly. Also,
U i ∣ v i ∣ p μ i − 2 ⟶ B 0 ∣ v ( 0 ) ∣ p 0 − 2 in L r n / 2 , U_{i}|v_{i}|^{p_{\mu_{i}}-2}\longrightarrow B_{0}|v^{(0)}|^{p_{0}-2}\quad\text{in }L^{r n/2}, U i ∣ v i ∣ p μ i − 2 ⟶ B 0 ∣ v ( 0 ) ∣ p 0 − 2 in L r n /2 ,
because ( p 0 − 2 ) n / 2 = p 0 < R (p_{0}-2)n/2=p_{0}<R ( p 0 − 2 ) n /2 = p 0 < R . Thus the local Hessian term converges in bilinear form norm on H 1 H^{1} H 1 . For the nonlocal term, write b i = ∣ v i ∣ p μ i − 2 v i b_{i}=|v_{i}|^{p_{\mu_{i}}-2}v_{i} b i = ∣ v i ∣ p μ i − 2 v i . The powers converge strongly in L q L^{q} L q , where 1 / q = 1 / r − 1 / 2 ∗ > ( p 0 − 1 ) / R 1/q=1/r-1/2^{*}>(p_{0}-1)/R 1/ q = 1/ r − 1/ 2 ∗ > ( p 0 − 1 ) / R . Put b 0 = ∣ v ( 0 ) ∣ p 0 − 2 v ( 0 ) b_{0}=|v^{(0)}|^{p_{0}-2}v^{(0)} b 0 = ∣ v ( 0 ) ∣ p 0 − 2 v ( 0 ) and T i ϕ = b i ϕ T_{i}\phi=b_{i}\phi T i ϕ = b i ϕ , T 0 ϕ = b 0 ϕ T_{0}\phi=b_{0}\phi T 0 ϕ = b 0 ϕ . Hölder’s inequality and Sobolev embedding give
∥ T i − T 0 ∥ H 1 → L r ≤ C ∥ b i − b 0 ∥ L q ⟶ 0 , sup i ∥ T i ∥ H 1 → L r < ∞ . \|T_{i}-T_{0}\|_{H^{1}\to L^{r}}\leq C\|b_{i}-b_{0}\|_{L^{q}}\longrightarrow0,\qquad\sup_{i}\|T_{i}\|_{H^{1}\to L^{r}}<\infty. ∥ T i − T 0 ∥ H 1 → L r ≤ C ∥ b i − b 0 ∥ L q ⟶ 0 , i sup ∥ T i ∥ H 1 → L r < ∞.
For ϕ , ψ ∈ H 1 \phi,\psi\in H^1 ϕ , ψ ∈ H 1 , decompose the difference of the nonlocal bilinear terms as
∫ J μ i ( T i ϕ ) T i ψ d V − ∫ J 0 ( T 0 ϕ ) T 0 ψ d V = ∫ ( J μ i − J 0 ) ( T i ϕ ) T i ψ d V + ∫ J 0 ( ( T i − T 0 ) ϕ ) T i ψ d V + ∫ J 0 ( T 0 ϕ ) ( T i − T 0 ) ψ d V . \begin{aligned}
&\int J_{\mu_i}(T_i\phi)T_i\psi\,dV-\int J_0(T_0\phi)T_0\psi\,dV\\
&=\int(J_{\mu_i}-J_0)(T_i\phi)T_i\psi\,dV+\int J_0((T_i-T_0)\phi)T_i\psi\,dV\\
&\quad+\int J_0(T_0\phi)(T_i-T_0)\psi\,dV.
\end{aligned} ∫ J μ i ( T i ϕ ) T i ψ d V − ∫ J 0 ( T 0 ϕ ) T 0 ψ d V = ∫ ( J μ i − J 0 ) ( T i ϕ ) T i ψ d V + ∫ J 0 (( T i − T 0 ) ϕ ) T i ψ d V + ∫ J 0 ( T 0 ϕ ) ( T i − T 0 ) ψ d V .
The finite volume of the sphere gives ∥ z ∥ 1 ≤ C ∥ z ∥ r \|z\|_1\le C\|z\|_r ∥ z ∥ 1 ≤ C ∥ z ∥ r . Each term is therefore bounded using L r → L ∞ L^r\to L^\infty L r → L ∞ for J μ i J_{\mu_i} J μ i and J 0 J_0 J 0 . In particular, the absolute value of the displayed difference is at most
C ( ∥ J μ i − J 0 ∥ L r → L ∞ + ∥ b i − b 0 ∥ L q ) ∥ ϕ ∥ ∥ ψ ∥ . C\left(\|J_{\mu_i}-J_0\|_{L^r\to L^\infty}+\|b_i-b_0\|_{L^q}\right)\|\phi\|\|\psi\|. C ( ∥ J μ i − J 0 ∥ L r → L ∞ + ∥ b i − b 0 ∥ L q ) ∥ ϕ ∥∥ ψ ∥.
For the local term, Hölder’s inequality gives
∣ ∫ ( ( p μ i − 1 ) U i ∣ v i ∣ p μ i − 2 − ( p 0 − 1 ) B 0 ∣ v ( 0 ) ∣ p 0 − 2 ) ϕ ψ d V ∣ ≤ C ∥ ( p μ i − 1 ) U i ∣ v i ∣ p μ i − 2 − ( p 0 − 1 ) B 0 ∣ v ( 0 ) ∣ p 0 − 2 ∥ L n / 2 ∥ ϕ ∥ ∥ ψ ∥ . \left|\int\left((p_{\mu_i}-1)U_i|v_i|^{p_{\mu_i}-2}-(p_0-1)B_0|v^{(0)}|^{p_0-2}\right)\phi\psi\,dV\right|
\le C\left\|(p_{\mu_i}-1)U_i|v_i|^{p_{\mu_i}-2}-(p_0-1)B_0|v^{(0)}|^{p_0-2}\right\|_{L^{n/2}}\|\phi\|\|\psi\|. ∫ ( ( p μ i − 1 ) U i ∣ v i ∣ p μ i − 2 − ( p 0 − 1 ) B 0 ∣ v ( 0 ) ∣ p 0 − 2 ) ϕ ψ d V ≤ C ( p μ i − 1 ) U i ∣ v i ∣ p μ i − 2 − ( p 0 − 1 ) B 0 ∣ v ( 0 ) ∣ p 0 − 2 L n /2 ∥ ϕ ∥∥ ψ ∥.
The right-hand sides tend to zero. Since p μ i → p 0 p_{\mu_i}\to p_0 p μ i → p 0 , the full Hessians converge in energy operator norm.
Restrict this convergence to each of the three radial energy spaces. The limiting energy operators are compact perturbations of the identity. Their negative spectra are separated from zero, and zero is a simple isolated eigenvalue in each block. Spectral projections for the negative spectrum away from zero and for a small interval about zero have constant rank under sufficiently small norm perturbations [24 Chapter IV] . The symmetry fields are nonzero kernel elements in the three blocks at each μ i \mu_i μ i . Thus the one-dimensional spectral subspace near zero is exactly the kernel. The negative counts are therefore the limiting ones.
To obtain a threshold independent of the chosen minimizer, suppose that no μ ∗ ( k ) > 0 \mu_*(k)>0 μ ∗ ( k ) > 0 has the asserted property. We could then choose μ i → 0 \mu_i\to0 μ i → 0 and a minimizer at each μ i \mu_i μ i for which at least one of the three counts fails. Change its sign so that the profile is positive on the first half. The compactness argument above provides a subsequence converging to a minimizer of the limit problem. At that limit, each of the three zero eigenspaces is one-dimensional and the negative indices have the stated values. Norm convergence and the nonzero symmetry fields force all three counts along this subsequence to agree with those limiting values for large i i i , a contradiction. Shrinking the threshold to be less than min { 4 , n } \min\{4,n\} min { 4 , n } proves (1.17) for every minimizer in the stated parameter range. The one-dimensional kernel in each block is generated by its nonzero symmetry profile.
Acknowledgments. This work was supported by the National Natural Science Foundation of China (No. 12671143).
Data availability statement. No datasets were generated or analyzed.
Competing interests. The authors declare no competing interests.