Introduction
Let d ≥ 2 d \ge2 d ≥ 2 , 0 < s < 1 0 < s < 1 0 < s < 1 , and 1 < p < ∞ 1 < p < \infty 1 < p < ∞ . On R + d = { x ∈ R d : x d > 0 } \mathbb{R}_{+}^{d}=\{x\in\mathbb{R}^{d}:x_{d}>0\} R + d = { x ∈ R d : x d > 0 } , write
[ u ] s , p , + p = ∫ R + d ∫ R + d ∣ u ( x ) − u ( y ) ∣ p ∣ x − y ∣ d + s p d x d y . [u]_{s,p,+}^{p}=\int_{\mathbb{R}_{+}^{d}}\int_{\mathbb{R}_{+}^{d}}\frac{|u(x)-u(y)|^{p}}{|x-y|^{d+sp}}\,dx\,dy. [ u ] s , p , + p = ∫ R + d ∫ R + d ∣ x − y ∣ d + s p ∣ u ( x ) − u ( y ) ∣ p d x d y .
The sharp Hardy inequality on the half-space admits a second remainder involving the distance to the codimension-two edge { x d − 1 = x d = 0 } \{x_{d-1}=x_{d}=0\} { x d − 1 = x d = 0 } . We are concerned with the best coefficient of this remainder and the behavior of extremizing sequences.
For τ > 0 \tau>0 τ > 0 , put
H d , s , p [ u ] = [ u ] s , p , + p − D d , s , p ∫ R + d x d − s p ∣ u ∣ p d x , \mathcal{H}_{d,s,p}[u]=[u]_{s,p,+}^{p}-D_{d,s,p}\int_{\mathbb{R}_{+}^{d}}x_{d}^{-sp}|u|^{p}\,dx, H d , s , p [ u ] = [ u ] s , p , + p − D d , s , p ∫ R + d x d − s p ∣ u ∣ p d x ,
R d , s , p , τ [ u ] = ∫ R + d x d τ − s p ∣ u ( x ) ∣ p ( x d − 1 2 + x d 2 ) τ / 2 d x , \mathcal{R}_{d,s,p,\tau}[u]=\int_{\mathbb{R}_{+}^{d}}\frac{x_{d}^{\tau-sp}|u(x)|^{p}}{(x_{d-1}^{2}+x_{d}^{2})^{\tau/2}}\,dx, R d , s , p , τ [ u ] = ∫ R + d ( x d − 1 2 + x d 2 ) τ /2 x d τ − s p ∣ u ( x ) ∣ p d x ,
and
Λ d , s , p , τ = inf u ∈ C c ∞ ( R + d ) ∖ { 0 } H d , s , p [ u ] R d , s , p , τ [ u ] . \Lambda_{d,s,p,\tau}=\inf_{u\in C_{c}^{\infty}(\mathbb{R}_{+}^{d})\setminus\{0\}}\frac{\mathcal{H}_{d,s,p}[u]}{\mathcal{R}_{d,s,p,\tau}[u]}. Λ d , s , p , τ = u ∈ C c ∞ ( R + d ) ∖ { 0 } inf R d , s , p , τ [ u ] H d , s , p [ u ] .
Here D d , s , p D_{d,s,p} D d , s , p is the sharp regional half-space Hardy constant, given in (20) . Bogdan and Dyda [2 ] determined it for p = 2 p=2 p = 2 , and Frank and Seiringer [11 ] treated general powers by the nonlinear ground-state method of [10 ] .
The local second remainder goes back to Maz’ya [17 , 18 ] . Tidblom proved inequalities for a larger family of weights [23 ] . Maz’ya and Shaposhnikova identified the sharp quadratic coefficient at τ = 1 \tau=1 τ = 1 through an angular eigenvalue problem [19 ] . The fractional inequality with the same two-plane weight was established by Kijaczko and Szczukiewicz [16 ] . Their result gives a positive coefficient. The optimal coefficient and the corresponding loss of compactness remain to be determined.
Fractional Hardy–Sobolev–Maz’ya inequalities involve a Sobolev remainder; see [5 , 6 , 8 , 22 ] . Weighted versions and inequalities with flat singular sets are treated in [7 , 15 , 21 ] . Related results for Sobolev–Bregman forms can be found in [14 , 12 ] .
Banerjee, Ganguly, and Sahu [1 ] obtained full-space Hardy stability estimates by Emden–Fowler and Mellin methods, using Lorentz or Marcinkiewicz distances to virtual extremals. For the half-space problem considered here, the second sharp potential must also be subtracted. The resulting deficit is governed by a singular angular operator.
Kijaczko and Szczukiewicz proved a positive remainder in the ranges [16 ]
0 < τ < { p ( s p + 1 ) 2 ( p − 1 ) , p ≥ 2 , 2 min { s p , 1 } , 1 < p < 2. 0 < \tau< \begin{cases} \dfrac{p(sp+1)}{2(p-1)}, & p \ge2, \\ 2\min\{sp,1\}, & 1 < p < 2. \end{cases} 0 < τ < ⎩ ⎨ ⎧ 2 ( p − 1 ) p ( s p + 1 ) , 2 min { s p , 1 } , p ≥ 2 , 1 < p < 2.
Since R d , s , p , τ ≤ R d , s , p , σ \mathcal{R}_{d,s,p,\tau} \le\mathcal{R}_{d,s,p,\sigma} R d , s , p , τ ≤ R d , s , p , σ for 0 < σ ≤ τ 0 < \sigma\le\tau 0 < σ ≤ τ , monotonicity gives Λ d , s , p , τ > 0 \Lambda_{d,s,p,\tau} > 0 Λ d , s , p , τ > 0 for every d ≥ 2 d \ge2 d ≥ 2 , 0 < s < 1 0 < s < 1 0 < s < 1 , p > 1 p > 1 p > 1 , and τ > 0 \tau> 0 τ > 0 . For d = p = 2 d=p=2 d = p = 2 , the estimate used below is
[ u ] s , + 2 − D 2 , s , 2 ∫ R + 2 ∣ u ∣ 2 x 2 2 s d x ≥ c s , τ ∫ R + 2 ∣ u ∣ 2 x 2 2 s − τ ∣ x ∣ τ d x . (1) [u]_{s,+}^{2}-\mathcal{D}_{2,s,2}\int_{\mathbb{R}^{2}_{+}}\frac{|u|^{2}}{x_{2}^{2s}}\,\mathrm{d}x\ge c_{s,\tau}\int_{\mathbb{R}^{2}_{+}}\frac{|u|^{2}}{x_{2}^{2s-\tau}|x|^{\tau}}\,\mathrm{d}x. \tag*{(1)} [ u ] s , + 2 − D 2 , s , 2 ∫ R + 2 x 2 2 s ∣ u ∣ 2 d x ≥ c s , τ ∫ R + 2 x 2 2 s − τ ∣ x ∣ τ ∣ u ∣ 2 d x . (1)
where c s , τ > 0 c_{s,\tau}>0 c s , τ > 0 for 0 < τ < 1 + 2 s 0<\tau<1+2s 0 < τ < 1 + 2 s . We use this estimate to close the positive angular and cylinder forms.
Our first result reduces the best coefficient to the planar problem. The dimensional factor follows from a contraction in the transverse L p L^{p} L p variables.
Theorem 1.1 (Dimensional reduction for general powers). Let d ≥ 2 d\ge2 d ≥ 2 , 0 < s < 1 0<s<1 0 < s < 1 , p > 1 p>1 p > 1 , and τ > 0 \tau>0 τ > 0 . Then
Λ d , s , p , τ = A d , s , p Λ 2 , s , p , τ , A d , s , p = π ( d − 2 ) / 2 Γ ( 1 + s p / 2 ) Γ ( ( d + s p ) / 2 ) . (2) \Lambda_{d,s,p,\tau}=A_{d,s,p}\Lambda_{2,s,p,\tau},\qquad A_{d,s,p}=\frac{\pi^{(d-2)/2}\Gamma(1+sp/2)}{\Gamma((d+sp)/2)}. \tag*{(2)} Λ d , s , p , τ = A d , s , p Λ 2 , s , p , τ , A d , s , p = Γ (( d + s p ) /2 ) π ( d − 2 ) /2 Γ ( 1 + s p /2 ) . (2) For p = 2 p=2 p = 2 , put A d , s = A d , s , 2 A_{d,s}=A_{d,s,2} A d , s = A d , s , 2 , Λ d , s , τ = Λ d , s , 2 , τ \Lambda_{d,s,\tau}=\Lambda_{d,s,2,\tau} Λ d , s , τ = Λ d , s , 2 , τ , and Λ s , τ = Λ 2 , s , 2 , τ \Lambda_{s,\tau}=\Lambda_{2,s,2,\tau} Λ s , τ = Λ 2 , s , 2 , τ . Then
Λ d , s , τ = A d , s Λ s , τ , A d , s = π ( d − 2 ) / 2 Γ ( 1 + s ) Γ ( ( d + 2 s ) / 2 ) . (3) \Lambda_{d,s,\tau}=A_{d,s}\Lambda_{s,\tau},\qquad A_{d,s}=\frac{\pi^{(d-2)/2}\Gamma(1+s)}{\Gamma((d+2s)/2)}. \tag*{(3)} Λ d , s , τ = A d , s Λ s , τ , A d , s = Γ (( d + 2 s ) /2 ) π ( d − 2 ) /2 Γ ( 1 + s ) . (3) The quadratic quotient has no nonzero extremal in the completion of C c ∞ ( R + d ) C^\infty_c(\mathbb{R}^{d}_{+}) C c ∞ ( R + d ) under ( H d , s , 2 + R d , s , 2 , τ ) 1 / 2 (\mathcal{H}_{d,s,2}+\mathcal{R}_{d,s,2,\tau})^{1/2} ( H d , s , 2 + R d , s , 2 , τ ) 1/2 . For d = 2 d=2 d = 2 , the last assertion is Theorem 1.2 .
Theorem 2.2 further reduces the planar problem by an L p L^{p} L p contraction in the logarithmic radial variable. For every p > 1 p>1 p > 1 , it expresses Λ 2 , s , p , τ \Lambda_{2,s,p,\tau} Λ 2 , s , p , τ as an angular variational infimum. We next consider p = 2 p=2 p = 2 , for which the angular problem has a compact spectral realization.
In the planar quadratic case, write
[ u ] s , + 2 = ∫ R + 2 ∫ R + 2 ∣ u ( x ) − u ( y ) ∣ 2 ∣ x − y ∣ 2 + 2 s d x d y . (4) [u]_{s,+}^{2}=\int_{\mathbb{R}^{2}_{+}}\int_{\mathbb{R}^{2}_{+}}\frac{|u(x)-u(y)|^{2}}{|x-y|^{2+2s}}\,\mathrm{d}x\,\mathrm{d}y. \tag*{(4)} [ u ] s , + 2 = ∫ R + 2 ∫ R + 2 ∣ x − y ∣ 2 + 2 s ∣ u ( x ) − u ( y ) ∣ 2 d x d y . (4)
Then
[ u ] s , + 2 ≥ D 2 , s , 2 ∫ R + 2 x 2 − 2 s ∣ u ∣ 2 d x . [u]_{s,+}^{2}\ge\mathcal{D}_{2,s,2}\int_{\mathbb{R}^{2}_{+}}x_{2}^{-2s}|u|^{2}\,\mathrm{d}x. [ u ] s , + 2 ≥ D 2 , s , 2 ∫ R + 2 x 2 − 2 s ∣ u ∣ 2 d x .
where, in this ordered-double-integral normalization,
D 2 , s , 2 = 2 π Γ ( s + 1 2 ) Γ ( 1 + s ) ∫ 0 1 ( 1 − t ( 2 s − 1 ) / 2 ) 2 ( 1 − t ) 1 + 2 s d t . (5) \mathcal{D}_{2,s,2}=\frac{2\sqrt{\pi}\,\Gamma\left(s+\frac{1}{2}\right)}{\Gamma(1+s)}\int_{0}^{1}\frac{\left(1-t^{(2s-1)/2}\right)^{2}}{(1-t)^{1+2s}}\,\mathrm{d}t. \tag*{(5)} D 2 , s , 2 = Γ ( 1 + s ) 2 π Γ ( s + 2 1 ) ∫ 0 1 ( 1 − t ) 1 + 2 s ( 1 − t ( 2 s − 1 ) /2 ) 2 d t . (5)
In particular, D 2 , 1 / 2 , 2 = 0 \mathcal{D}_{2,1/2,2}=0 D 2 , 1/2 , 2 = 0 .
We use
H d , s = H d , s , 2 , R d , s , τ = R d , s , 2 , τ . \mathcal{H}_{d,s}=\mathcal{H}_{d,s,2},\qquad\mathcal{R}_{d,s,\tau}=\mathcal{R}_{d,s,2,\tau}. H d , s = H d , s , 2 , R d , s , τ = R d , s , 2 , τ .
For d = 2 d=2 d = 2 , put R s , τ = R 2 , s , τ \mathcal{R}_{s,\tau}=\mathcal{R}_{2,s,\tau} R s , τ = R 2 , s , τ . The planar deficit is
H s [ u ] = [ u ] s , + 2 − D 2 , s , 2 ∫ R + 2 ∣ u ∣ 2 x 2 2 s d x (6) \mathcal{H}_s[u]=[u]_{s,+}^2-\mathcal{D}_{2,s,2}\int_{\mathbb{R}_+^2}\frac{|u|^2}{x_2^{2s}}\,dx \tag*{(6)} H s [ u ] = [ u ] s , + 2 − D 2 , s , 2 ∫ R + 2 x 2 2 s ∣ u ∣ 2 d x (6)
Both quadratic terms have degree 2 s − 2 2s-2 2 s − 2 under Euclidean dilation. The energy space is the completion of C c ∞ ( R + d ) C^\infty_c(\mathbb{R}^d_+) C c ∞ ( R + d ) under ( H d , s + R d , s , τ ) 1 / 2 (\mathcal{H}_{d,s}+\mathcal{R}_{d,s,\tau})^{1/2} ( H d , s + R d , s , τ ) 1/2 . The positive remainder estimate makes this a norm. In the planar logarithmic polar coordinates, dilation becomes translation. Put
K s ( q , θ , φ ) = 2 − 1 − s ( cosh q − cos ( θ − φ ) ) 1 + s . K_s(q,\theta,\varphi)=\frac{2^{-1-s}}{(\cosh q-\cos(\theta-\varphi))^{1+s}}. K s ( q , θ , φ ) = ( cosh q − cos ( θ − φ ) ) 1 + s 2 − 1 − s .
and a s = 1 − s a_s=1-s a s = 1 − s . For ξ ∈ R \xi\in\mathbb{R} ξ ∈ R , define
b s , ξ [ ψ ] = ∫ 0 π ∫ 0 π ∫ R K s ( q , θ , φ ) × [ cosh ( a s q ) ( ∣ ψ ( θ ) ∣ 2 + ∣ ψ ( φ ) ∣ 2 ) − 2 Re ( e i ξ q ψ ( θ ) ψ ( φ ) ‾ ) ] d q d θ d φ − D 2 , s , 2 ∫ 0 π ∣ ψ ( θ ) ∣ 2 ( sin θ ) 2 s d θ (7) \begin{aligned}
\mathfrak{b}_{s,\xi}[\psi]
={}&\int_0^\pi\int_0^\pi\int_{\mathbb{R}}K_s(q,\theta,\varphi)\\
&\quad\times\left[\cosh(a_sq)(|\psi(\theta)|^2+|\psi(\varphi)|^2)
-2\operatorname{Re}(e^{i\xi q}\psi(\theta)\overline{\psi(\varphi)})\right]\,dq\,d\theta\,d\varphi\\
&\quad-\mathcal{D}_{2,s,2}\int_0^\pi\frac{|\psi(\theta)|^2}{(\sin\theta)^{2s}}\,d\theta \tag*{(7)}
\end{aligned} b s , ξ [ ψ ] = ∫ 0 π ∫ 0 π ∫ R K s ( q , θ , φ ) × [ cosh ( a s q ) ( ∣ ψ ( θ ) ∣ 2 + ∣ ψ ( φ ) ∣ 2 ) − 2 Re ( e i ξ q ψ ( θ ) ψ ( φ ) ) ] d q d θ d φ − D 2 , s , 2 ∫ 0 π ( sin θ ) 2 s ∣ ψ ( θ ) ∣ 2 d θ (7)
The form is first defined on C c ∞ ( 0 , π ) C^\infty_c(0,\pi) C c ∞ ( 0 , π ) . Its closed extension is constructed in § 4 . Let
m s , τ ( θ ) = ( sin θ ) τ − 2 s . m_{s,\tau}(\theta)=(\sin\theta)^{\tau-2s}. m s , τ ( θ ) = ( sin θ ) τ − 2 s .
and set
H s , τ = L 2 ( ( 0 , π ) , m s , τ ( θ ) d θ ) . H_{s,\tau}=L^2((0,\pi),m_{s,\tau}(\theta)\,d\theta). H s , τ = L 2 (( 0 , π ) , m s , τ ( θ ) d θ ) .
Let Y s \mathcal{Y}_s Y s be the common closed angular form domain and X s \mathcal{X}_s X s the closed cylinder domain constructed in § 3 and § 4 . We use complex forms for the Mellin transform, with angular norm
∥ ψ ∥ Y s 2 = b s , 0 [ ψ ] + ∥ ψ ∥ H s , τ 2 . \|\psi\|_{\mathcal{Y}_s}^2=\mathfrak{b}_{s,0}[\psi]+\|\psi\|_{H_{s,\tau}}^2. ∥ ψ ∥ Y s 2 = b s , 0 [ ψ ] + ∥ ψ ∥ H s , τ 2 .
The real conclusions follow by taking real and imaginary parts. Let B s ( ξ ) B_s(\xi) B s ( ξ ) be the self-adjoint operator associated with b s , ξ \mathfrak{b}_{s,\xi} b s , ξ in H s , τ H_{s,\tau} H s , τ . Lemma 4.2 shows that its resolvent is compact. We write
0 < λ 1 ( ξ ) ≤ λ 2 ( ξ ) ≤ ⋯ , λ j ( ξ ) ⟶ ∞ . 0<\lambda_1(\xi)\leq\lambda_2(\xi)\leq\cdots,\qquad\lambda_j(\xi)\longrightarrow\infty. 0 < λ 1 ( ξ ) ≤ λ 2 ( ξ ) ≤ ⋯ , λ j ( ξ ) ⟶ ∞.
In particular,
λ 1 ( ξ ) = inf ψ ∈ Y s ∖ { 0 } b s , ξ [ ψ ] ∫ 0 π m s , τ ∣ ψ ∣ 2 d θ . \lambda_1(\xi)=\inf_{\psi\in\mathcal{Y}_s\setminus\{0\}}\frac{\mathfrak{b}_{s,\xi}[\psi]}{\int_0^\pi m_{s,\tau}|\psi|^2\,d\theta}. λ 1 ( ξ ) = ψ ∈ Y s ∖ { 0 } inf ∫ 0 π m s , τ ∣ ψ ∣ 2 d θ b s , ξ [ ψ ] .
Theorem 1.2 (Sharp reduction and non-attainment). Let 0 < s < 1 0<s<1 0 < s < 1 and τ > 0 \tau>0 τ > 0 . Then
Λ s , τ = λ 1 ( 0 ) > 0. \Lambda_{s,\tau}=\lambda_1(0)>0. Λ s , τ = λ 1 ( 0 ) > 0. The eigenvalue λ 1 ( 0 ) \lambda_1(0) λ 1 ( 0 ) is simple and has a positive eigenfunction φ s , τ \varphi_{s,\tau} φ s , τ , normalized by
∫ 0 π m s , τ ( θ ) φ s , τ ( θ ) 2 d θ = 1. \int_0^\pi m_{s,\tau}(\theta)\varphi_{s,\tau}(\theta)^2\,d\theta=1. ∫ 0 π m s , τ ( θ ) φ s , τ ( θ ) 2 d θ = 1. The value Λ s , τ \Lambda_{s,\tau} Λ s , τ is not attained by a nonzero function in the completion of C c ∞ ( R + 2 ) C^\infty_c(\mathbb{R}^2_+) C c ∞ ( R + 2 ) under ( H s + R s , τ ) 1 / 2 (\mathcal{H}_s+\mathcal{R}_{s,\tau})^{1/2} ( H s + R s , τ ) 1/2 .
The homogeneous function U s , τ ( r cos θ , r sin θ ) = r s − 1 φ s , τ ( θ ) U_{s,\tau}(r\cos\theta,r\sin\theta)=r^{s-1}\varphi_{s,\tau}(\theta) U s , τ ( r cos θ , r sin θ ) = r s − 1 φ s , τ ( θ ) solves the weak equation for the full sharp potential. In Theorem 5.3 , we show that H s [ u ] − Λ s , τ R s , τ [ u ] \mathcal{H}_s[u]-\Lambda_{s,\tau}\mathcal{R}_{s,\tau}[u] H s [ u ] − Λ s , τ R s , τ [ u ] equals a positive jump form with weight U s , τ ( x ) U s , τ ( y ) U_{s,\tau}(x)U_{s,\tau}(y) U s , τ ( x ) U s , τ ( y ) . This identity extends to the completed energy space. To state the quantitative estimate, we introduce the following distance. For x = e t ( cos θ , sin θ ) x=e^t(\cos\theta,\sin\theta) x = e t ( cos θ , sin θ ) , define
w ( t , θ ) = e ( 1 − s ) t u ( e t cos θ , e t sin θ ) . w(t,\theta)=e^{(1-s)t}u(e^t\cos\theta,e^t\sin\theta). w ( t , θ ) = e ( 1 − s ) t u ( e t cos θ , e t sin θ ) .
Let w ^ ( ξ , θ ) \widehat{w}(\xi,\theta) w ( ξ , θ ) be the Fourier transform in t t t . Let P 0 P_0 P 0 be the orthogonal projection in L 2 ( ( 0 , π ) , m s , τ d θ ) L^2((0,\pi),m_{s,\tau}\,d\theta) L 2 (( 0 , π ) , m s , τ d θ ) onto span { φ s , τ } \operatorname{span}\{\varphi_{s,\tau}\} span { φ s , τ } . Set
Dist s , τ ( u ) 2 = ∫ R ∥ ( I − P 0 ) w ^ ( ξ , ⋅ ) ∥ L 2 ( m s , τ ) 2 d ξ + ∫ R min { ξ 2 , 1 } ∥ P 0 w ^ ( ξ , ⋅ ) ∥ L 2 ( m s , τ ) 2 d ξ . (8) \operatorname{Dist}_{s,\tau}(u)^2=\int_{\mathbb{R}}\left\|(I-P_0)\widehat{w}(\xi,\cdot)\right\|_{L^2(m_{s,\tau})}^2\,d\xi+\int_{\mathbb{R}}\min\{\xi^2,1\}\left\|P_0\widehat{w}(\xi,\cdot)\right\|_{L^2(m_{s,\tau})}^2\,d\xi. \tag*{(8)} Dist s , τ ( u ) 2 = ∫ R ∥ ( I − P 0 ) w ( ξ , ⋅ ) ∥ L 2 ( m s , τ ) 2 d ξ + ∫ R min { ξ 2 , 1 } ∥ P 0 w ( ξ , ⋅ ) ∥ L 2 ( m s , τ ) 2 d ξ . (8)
Its kernel in the energy space is { 0 } \{0\} { 0 } , since the formal zero mode w ( t , θ ) = c φ s , τ ( θ ) w(t,\theta)=c\varphi_{s,\tau}(\theta) w ( t , θ ) = c φ s , τ ( θ ) is not in L 2 L^2 L 2 in t t t . Since ξ 2 / ( 1 + ξ 2 ) \xi^2/(1+\xi^2) ξ 2 / ( 1 + ξ 2 ) is comparable to min { ξ 2 , 1 } \min\{\xi^2,1\} min { ξ 2 , 1 } , Plancherel also gives the intrinsic equivalent form
Dist s , τ ( u ) 2 ≍ ∥ ( I − P 0 ) w ∥ L 2 ( R ; H s , τ ) 2 + ∥ ∂ t ( 1 − ∂ t 2 ) − 1 / 2 P 0 w ∥ L 2 ( R ; H s , τ ) 2 . \operatorname{Dist}_{s,\tau}(u)^2\asymp\left\|(I-P_0)w\right\|_{L^2(\mathbb{R};H_{s,\tau})}^2+\left\|\partial_t(1-\partial_t^2)^{-1/2}P_0w\right\|_{L^2(\mathbb{R};H_{s,\tau})}^2. Dist s , τ ( u ) 2 ≍ ∥ ( I − P 0 ) w ∥ L 2 ( R ; H s , τ ) 2 + ∂ t ( 1 − ∂ t 2 ) − 1/2 P 0 w L 2 ( R ; H s , τ ) 2 .
The first term measures the component orthogonal to the angular ground state. The second measures variation in the logarithmic radius.
Theorem 1.3 (Quantitative stability). Let 0 < s < 1 0<s<1 0 < s < 1 and τ > 0 \tau>0 τ > 0 . There is γ s , τ > 0 \gamma_{s,\tau}>0 γ s , τ > 0 such that
H s [ u ] − Λ s , τ R s , τ [ u ] ≥ γ s , τ Dist s , τ ( u ) 2 (9) \mathcal{H}_s[u]-\Lambda_{s,\tau}\mathcal{R}_{s,\tau}[u]\geq\gamma_{s,\tau}\operatorname{Dist}_{s,\tau}(u)^2 \tag*{(9)} H s [ u ] − Λ s , τ R s , τ [ u ] ≥ γ s , τ Dist s , τ ( u ) 2 (9) for every u ∈ C c ∞ ( R + 2 ) u\in C^\infty_c(\mathbb{R}^2_+) u ∈ C c ∞ ( R + 2 ) . Moreover,
λ 1 ( ξ ) = Λ s , τ + κ s , τ ξ 2 + O ( ξ 4 ) as ξ → 0 , (10) \lambda_1(\xi)=\Lambda_{s,\tau}+\kappa_{s,\tau}\xi^2+O(\xi^4)\quad\text{as }\xi\to0, \tag*{(10)} λ 1 ( ξ ) = Λ s , τ + κ s , τ ξ 2 + O ( ξ 4 ) as ξ → 0 , (10) where
κ s , τ = ∫ 0 π ∫ 0 π ∫ R q 2 K s ( q , θ , ϕ ) φ s , τ ( θ ) φ s , τ ( ϕ ) d q d θ d ϕ > 0. (11) \kappa_{s,\tau}=\int_0^\pi\int_0^\pi\int_{\mathbb{R}}q^2K_s(q,\theta,\phi)\varphi_{s,\tau}(\theta)\varphi_{s,\tau}(\phi)\,dq\,d\theta\,d\phi>0. \tag*{(11)} κ s , τ = ∫ 0 π ∫ 0 π ∫ R q 2 K s ( q , θ , ϕ ) φ s , τ ( θ ) φ s , τ ( ϕ ) d q d θ d ϕ > 0. (11)
For the fixed distance in (8) , define the optimal coefficient
Γ s , τ h p = inf u ≠ 0 H s [ u ] − Λ s , τ R s , τ [ u ] Dist s , τ ( u ) 2 . \Gamma^{\mathrm{hp}}_{s,\tau}=\inf_{u\ne0}\frac{\mathcal{H}_s[u]-\Lambda_{s,\tau}\mathcal{R}_{s,\tau}[u]}{\operatorname{Dist}_{s,\tau}(u)^2}. Γ s , τ hp = u = 0 inf Dist s , τ ( u ) 2 H s [ u ] − Λ s , τ R s , τ [ u ] .
where the infimum is taken over the completed energy space. Put h ( ξ ) = min { ξ 2 , 1 } h(\xi)=\min\{\xi^2,1\} h ( ξ ) = min { ξ 2 , 1 } and Q 0 = I − P 0 Q_0=I-P_0 Q 0 = I − P 0 . For ξ ≠ 0 \xi\ne0 ξ = 0 , let
μ s , τ ( ξ ) = inf ψ ∈ Y s ∖ { 0 } b s , ξ [ ψ ] − Λ s , τ ∥ ψ ∥ L 2 ( m s , τ ) 2 ∥ Q 0 ψ ∥ L 2 ( m s , τ ) 2 + h ( ξ ) ∥ P 0 ψ ∥ L 2 ( m s , τ ) 2 . (12) \mu_{s,\tau}(\xi)=\inf_{\psi\in Y_s\setminus\{0\}}\frac{b_{s,\xi}[\psi]-\Lambda_{s,\tau}\|\psi\|_{L^2(m_{s,\tau})}^2}{\|Q_0\psi\|_{L^2(m_{s,\tau})}^2+h(\xi)\|P_0\psi\|_{L^2(m_{s,\tau})}^2}. \tag*{(12)} μ s , τ ( ξ ) = ψ ∈ Y s ∖ { 0 } inf ∥ Q 0 ψ ∥ L 2 ( m s , τ ) 2 + h ( ξ ) ∥ P 0 ψ ∥ L 2 ( m s , τ ) 2 b s , ξ [ ψ ] − Λ s , τ ∥ ψ ∥ L 2 ( m s , τ ) 2 . (12)
Theorem 1.4 (Optimal half-plane stability coefficient). Let 0 < s < 1 0<s<1 0 < s < 1 and τ > 0 \tau>0 τ > 0 . The number Γ s , τ h p \Gamma^{\mathrm{hp}}_{s,\tau} Γ s , τ hp is positive and
Γ s , τ h p = ess inf ξ ∈ R ∖ { 0 } μ s , τ ( ξ ) . (13) \Gamma^{\mathrm{hp}}_{s,\tau}=\operatorname*{ess\,inf}_{\xi\in\mathbb{R}\setminus\{0\}}\mu_{s,\tau}(\xi). \tag*{(13)} Γ s , τ hp = ξ ∈ R ∖ { 0 } ess inf μ s , τ ( ξ ) . (13) For every ξ ≠ 0 \xi\ne0 ξ = 0 , the infimum in (12) is attained. The function μ s , τ \mu_{s,\tau} μ s , τ extends continuously to zero, with the value given in (14) , and tends to infinity as ∣ ξ ∣ → ∞ |\xi|\to\infty ∣ ξ ∣ → ∞ . Consequently,
Γ s , τ h p = min ξ ∈ R μ s , τ ( ξ ) . \Gamma^{\mathrm{hp}}_{s,\tau}=\min_{\xi\in\mathbb{R}}\mu_{s,\tau}(\xi). Γ s , τ hp = ξ ∈ R min μ s , τ ( ξ ) . Set
Z s , τ = { ξ ≠ 0 : μ s , τ ( ξ ) = Γ s , τ h p } . Z_{s,\tau}=\{\xi\ne0:\mu_{s,\tau}(\xi)=\Gamma^{\mathrm{hp}}_{s,\tau}\}. Z s , τ = { ξ = 0 : μ s , τ ( ξ ) = Γ s , τ hp } . The set Z s , τ Z_{s,\tau} Z s , τ is finite, and Γ s , τ h p \Gamma^{\mathrm{hp}}_{s,\tau} Γ s , τ hp is not attained by a nonzero energy function. The variational characterization above is relative to the fixed multiplier h h h . At zero,
lim ξ → 0 μ s , τ ( ξ ) = min { κ s , τ , λ 2 ( 0 ) − λ 1 ( 0 ) } . (14) \lim_{\xi\to0}\mu_{s,\tau}(\xi)=\min\{\kappa_{s,\tau},\lambda_2(0)-\lambda_1(0)\}. \tag*{(14)} ξ → 0 lim μ s , τ ( ξ ) = min { κ s , τ , λ 2 ( 0 ) − λ 1 ( 0 )} . (14)
If R s , τ [ u n ] = 1 \mathcal{R}_{s,\tau}[u_n]=1 R s , τ [ u n ] = 1 , denote the transform of u n u_n u n by w n w_n w n . Define the axial dispersion function
ρ s , τ ( ξ ) = b s , ξ [ φ s , τ ] − Λ s , τ = 2 ∫ 0 π ∫ 0 π ∫ R ( 1 − cos ( ξ q ) ) K s ( q , θ , ϕ ) φ s , τ ( θ ) φ s , τ ( ϕ ) d q d θ d ϕ . (15) \begin{aligned}
\rho_{s,\tau}(\xi)=b_{s,\xi}[\varphi_{s,\tau}]-\Lambda_{s,\tau} \\
&=2\int_0^\pi\int_0^\pi\int_{\mathbb{R}}(1-\cos(\xi q))K_s(q,\theta,\phi)\varphi_{s,\tau}(\theta)\varphi_{s,\tau}(\phi)\,dq\,d\theta\,d\phi. \tag*{(15)}
\end{aligned} ρ s , τ ( ξ ) = b s , ξ [ φ s , τ ] − Λ s , τ = 2 ∫ 0 π ∫ 0 π ∫ R ( 1 − cos ( ξ q )) K s ( q , θ , ϕ ) φ s , τ ( θ ) φ s , τ ( ϕ ) d q d θ d ϕ . (15)
It is positive for ξ ≠ 0 \xi\ne0 ξ = 0 and ρ s , τ ( ξ ) = κ s , τ ξ 2 + O ( ξ 4 ) \rho_{s,\tau}(\xi)=\kappa_{s,\tau}\xi^2+O(\xi^4) ρ s , τ ( ξ ) = κ s , τ ξ 2 + O ( ξ 4 ) at zero. Proposition 6.5 gives Γ s , τ h p ≤ ρ s , τ ( 1 ) < κ s , τ \Gamma^{\mathrm{hp}}_{s,\tau}\le\rho_{s,\tau}(1)<\kappa_{s,\tau} Γ s , τ hp ≤ ρ s , τ ( 1 ) < κ s , τ . In particular, the quadratic dispersion coefficient does not give the best constant for this distance.
Theorem 1.5 (Energy decomposition and extremizing sequences). Let 0 < s < 1 0<s<1 0 < s < 1 and τ > 0 \tau>0 τ > 0 . The canonical decomposition in (90) satisfies the two-sided energy estimate (6.8) on X s \mathcal{X}_s X s . Let u n u_n u n belong to the completed energy space and satisfy R s , τ [ u n ] = 1 \mathcal{R}_{s,\tau}[u_n]=1 R s , τ [ u n ] = 1 . The following statements are equivalent:
(1) H s [ u n ] → Λ s , τ \mathcal{H}_s[u_n]\to\Lambda_{s,\tau} H s [ u n ] → Λ s , τ ;
(2) there are a n ∈ L 2 ( R ) a_n\in L^2(\mathbb{R}) a n ∈ L 2 ( R ) and r n ∈ X s r_n\in\mathcal{X}_s r n ∈ X s such that
w n ( t , θ ) = a n ( t ) φ s , τ ( θ ) + r n ( t , θ ) , w_n(t,\theta)=a_n(t)\varphi_{s,\tau}(\theta)+r_n(t,\theta), w n ( t , θ ) = a n ( t ) φ s , τ ( θ ) + r n ( t , θ ) , {#eq:1.16}
\lVert r_n\rVert_{\mathcal{X}_s}\to 0,\qquad \lVert a_n\rVert_{L^2(\mathbb{R})}\to 1,
Equation not extracted
(16) \int_{\mathbb{R}}\rho_{s,\tau}(\xi)|\widehat{a_n}(\xi)|^2,d\xi\to 0.
In the implication ( 1 ) ⇒ ( 2 ) , the functions a n may be chosen by the low-frequency moving projection defined in ( 6.7 ) . Thus the angular defect vanishes in the energy norm. The remaining mass disperses in the logarithmic radial variable. \text{In the implication }(1)\Rightarrow(2)\text{, the functions }a_n\text{ may be chosen by the low-frequency moving projection defined in }(6.7)\text{. Thus the angular defect vanishes in the energy norm. The remaining mass disperses in the logarithmic radial variable.} In the implication ( 1 ) ⇒ ( 2 ) , the functions a n may be chosen by the low-frequency moving projection defined in ( 6.7 ) . Thus the angular defect vanishes in the energy norm. The remaining mass disperses in the logarithmic radial variable.
When d ≥ 3 d\ge3 d ≥ 3 , the partial Fourier transform along the edge produces an additional nonnegative term. Combining this term with the planar decomposition gives the necessary and sufficient criterion of Theorem 7.3. This criterion is stated in joint Fourier–Mellin energy variables.
Our last main result concerns the limit as the fractional order tends to one. We use the fixed Hilbert space
H 1 , τ = L 2 ( ( 0 , π ) , ( sin θ ) τ − 1 d θ ) . H_{1,\tau}=L^2\left((0,\pi),(\sin\theta)^{\tau-1}\,d\theta\right). H 1 , τ = L 2 ( ( 0 , π ) , ( sin θ ) τ − 1 d θ ) .
Let l 1 , ξ l_{1,\xi} l 1 , ξ be the local angular form in § 8 , and let y 1 , τ y_{1,\tau} y 1 , τ be its positive normalized zero-frequency minimizer. The number Λ 1 , τ \Lambda_{1,\tau} Λ 1 , τ is the sharp local half-space remainder defined there.
Theorem 1.6 (Sharp local limit). Let τ > 0 \tau>0 τ > 0 . Then
lim s ↑ 1 2 ( 1 − s ) π Λ s , τ = Λ 1 , τ . \lim_{s\uparrow1}\frac{2(1-s)}{\pi}\Lambda_{s,\tau}=\Lambda_{1,\tau}. s ↑ 1 lim π 2 ( 1 − s ) Λ s , τ = Λ 1 , τ .
Let λ j ( 1 ) ( ξ ) \lambda_j^{(1)}(\xi) λ j ( 1 ) ( ξ ) be the eigenvalues of l 1 , ξ l_{1,\xi} l 1 , ξ in H 1 , τ H_{1,\tau} H 1 , τ . For every fixed j ≥ 1 j\ge1 j ≥ 1 and ξ ∈ R \xi\in\mathbb{R} ξ ∈ R ,
lim s ↑ 1 2 ( 1 − s ) π λ j ( ξ ) = λ j ( 1 ) ( ξ ) . \lim_{s\uparrow1}\frac{2(1-s)}{\pi}\lambda_j(\xi)=\lambda_j^{(1)}(\xi). s ↑ 1 lim π 2 ( 1 − s ) λ j ( ξ ) = λ j ( 1 ) ( ξ ) .
If y 1 , τ y_{1,\tau} y 1 , τ is the positive normalized minimizer in Lemma 8.2, then
( sin θ ) 1 / 2 − s φ s , τ ⟶ y 1 , τ strongly in H 1 , τ . (\sin\theta)^{1/2-s}\varphi_{s,\tau}\longrightarrow y_{1,\tau}\quad\text{strongly in }H_{1,\tau}. ( sin θ ) 1/2 − s φ s , τ ⟶ y 1 , τ strongly in H 1 , τ .
Finally,
lim s ↑ 1 2 ( 1 − s ) π κ s , τ = ∫ 0 π sin θ y 1 , τ ( θ ) 2 d θ . \lim_{s\uparrow1}\frac{2(1-s)}{\pi}\kappa_{s,\tau}=\int_0^\pi\sin\theta\,y_{1,\tau}(\theta)^2\,d\theta. s ↑ 1 lim π 2 ( 1 − s ) κ s , τ = ∫ 0 π sin θ y 1 , τ ( θ ) 2 d θ .
At τ = 2 \tau=2 τ = 2 , Proposition 8.3 gives the explicit local values
Λ 1 , 2 = 1 4 , λ 2 ( 1 ) ( 0 ) − λ 1 ( 1 ) ( 0 ) = 2 , ∫ 0 π sin θ y 1 , 2 2 d θ = 1. \Lambda_{1,2}=\frac{1}{4},\qquad\lambda_2^{(1)}(0)-\lambda_1^{(1)}(0)=2,\qquad\int_0^\pi\sin\theta\,y_{1,2}^2\,d\theta=1. Λ 1 , 2 = 4 1 , λ 2 ( 1 ) ( 0 ) − λ 1 ( 1 ) ( 0 ) = 2 , ∫ 0 π sin θ y 1 , 2 2 d θ = 1.
The local stability coefficient for the same multiplier h h h is 1 1 1 and is attained. Theorem 1.6 and Proposition 6.5 also imply that Z s , 2 \mathcal{Z}_{s,2} Z s , 2 contains nonzero frequencies for all s < 1 s<1 s < 1 sufficiently close to one. This conclusion concerns the fixed spectral distance (8) . The local limits above are at fixed frequency and index; the limit of Γ s , τ h p \Gamma^{\mathrm{hp}}_{s,\tau} Γ s , τ hp is not asserted.
Remark 1.7. The critical Hardy term prevents us from treating the angular energy and potential separately on the closed domain. We instead close the positive ground-state form. A stronger endpoint weight and interior fractional regularity then give compactness. Since ∣ x ∣ s − 1 ϕ s , τ |x|^{s-1}\phi_{s,\tau} ∣ x ∣ s − 1 ϕ s , τ does not belong to the energy space, the identity in Theorem 5.3 requires a regularization of the angular ground state.
For the stability estimate, we separate small frequencies from their complement and use the analytic first spectral projection near zero. This also gives the two-sided energy decomposition. To prove non-attainment of the best stability coefficient, we apply analytic Fredholm theory to the generalized fiber problem. This argument allows eigenvalue crossings away from zero. Finally, the local spectral limit requires a dyadic endpoint estimate uniform as s ↑ 1 s \uparrow1 s ↑ 1 ; convergence on the smooth core alone is insufficient.
§ 2 proves the dimensional and angular reductions for general powers. § 3 –§ 4 construct the quadratic closed forms and establish their compactness and frequency estimates. § 5 identifies the planar sharp constant and its full ground-state representation. § 6 proves stability and the energy decomposition. § 7 gives the higher-dimensional splitting. § 8 proves the local spectral limits. The appendices contain the spectral periodic model, angular Ritz approximation, and subcritical nonlinear coercivity.
Exact reductions for general powers
We prove the dimensional reduction first and then the angular reduction. Both follow from the reverse triangle inequality in L p L^p L p and a matching family of test functions.
Lemma 2.1 (L p L^p L p transverse reduction). Let d ≥ 3 d \ge3 d ≥ 3 , 0 < s < 1 0 < s < 1 0 < s < 1 , p > 1 p > 1 p > 1 , and τ > 0 \tau> 0 τ > 0 . Then
Λ d , s , p , τ = A d , s , p Λ 2 , s , p , τ . \Lambda_{d,s,p,\tau}=A_{d,s,p}\Lambda_{2,s,p,\tau}. Λ d , s , p , τ = A d , s , p Λ 2 , s , p , τ .
Proof. Put n = d − 2 n=d-2 n = d − 2 . Write x = ( z , X ) x=(z,X) x = ( z , X ) , where z ∈ R n z \in\mathbb{R}^n z ∈ R n and X = ( x d − 1 , x d ) ∈ R + 2 X=(x_{d-1},x_d) \in\mathbb{R}_+^2 X = ( x d − 1 , x d ) ∈ R + 2 . For u ∈ C c ∞ ( R + d ) u \in C^\infty_c(\mathbb{R}_+^d) u ∈ C c ∞ ( R + d ) , set
u X ( z ) = u ( z , X ) , U ( X ) = ∥ u X ∥ L p ( R n ) . u_X(z)=u(z,X),\qquad U(X)=\lVert u_X\rVert_{L^p(\mathbb{R}^n)}. u X ( z ) = u ( z , X ) , U ( X ) = ∥ u X ∥ L p ( R n ) .
For every h ∈ R n h \in\mathbb{R}^n h ∈ R n , translation invariance and the reverse triangle inequality give
∫ R n ∣ u ( z , X ) − u ( z − h , Y ) ∣ p d z ≥ ∣ U ( X ) − U ( Y ) ∣ p . (17) \int_{\mathbb{R}^n}|u(z,X)-u(z-h,Y)|^p\,\mathrm{d}z\geq|U(X)-U(Y)|^p. \tag*{(17)} ∫ R n ∣ u ( z , X ) − u ( z − h , Y ) ∣ p d z ≥ ∣ U ( X ) − U ( Y ) ∣ p . (17)
The beta integral gives
∫ R n d h ( ∣ X − Y ∣ 2 + ∣ h ∣ 2 ) ( d + s p ) / 2 = A d , s , p ∣ X − Y ∣ − 2 − s p , \int_{\mathbb{R}^n}\frac{\mathrm{d}h}{\left(|X-Y|^2+|h|^2\right)^{(d+sp)/2}}=A_{d,s,p}|X-Y|^{-2-sp}, ∫ R n ( ∣ X − Y ∣ 2 + ∣ h ∣ 2 ) ( d + s p ) /2 d h = A d , s , p ∣ X − Y ∣ − 2 − s p ,
A d , s , p = π n / 2 Γ ( 1 + s p / 2 ) Γ ( ( d + s p ) / 2 ) . (18) A_{d,s,p}=\frac{\pi^{n/2}\Gamma(1+sp/2)}{\Gamma((d+sp)/2)}. \tag*{(18)} A d , s , p = Γ (( d + s p ) /2 ) π n /2 Γ ( 1 + s p /2 ) . (18)
By Fubini’s theorem, (17) , and (18) ,
[ u ] s , p , + p ≥ A d , s , p [ U ] s , p , + p . (19) [u]_{s,p,+}^p\geq A_{d,s,p}[U]_{s,p,+}^p. \tag*{(19)} [ u ] s , p , + p ≥ A d , s , p [ U ] s , p , + p . (19)
The explicit sharp half-space Hardy constant is
D d , s , p = 2 π ( d − 1 ) / 2 Γ ( ( 1 + s p ) / 2 ) Γ ( ( d + s p ) / 2 ) ∫ 0 1 ∣ 1 − t ( s p − 1 ) / p ∣ p ( 1 − t ) 1 + s p d t . (20) D_{d,s,p}=\frac{2\pi^{(d-1)/2}\Gamma((1+sp)/2)}{\Gamma((d+sp)/2)}
\int_0^1\frac{|1-t^{(sp-1)/p}|^p}{(1-t)^{1+sp}}\,\mathrm{d}t. \tag*{(20)} D d , s , p = Γ (( d + s p ) /2 ) 2 π ( d − 1 ) /2 Γ (( 1 + s p ) /2 ) ∫ 0 1 ( 1 − t ) 1 + s p ∣1 − t ( s p − 1 ) / p ∣ p d t . (20)
Comparison of dimensions d d d and 2 2 2 in (20) gives
D d , s , p = A d , s , p D 2 , s , p . (21) D_{d,s,p}=A_{d,s,p}D_{2,s,p}. \tag*{(21)} D d , s , p = A d , s , p D 2 , s , p . (21)
The two weighted integrals satisfy
∫ R + d X d − s p ∣ u ∣ p d x = ∫ R + 2 X 2 − s p ∣ U ( X ) ∣ p d X , R d , s , p , τ [ u ] = R 2 , s , p , τ [ U ] . (22) \begin{aligned}
\int_{\mathbb{R}_{+}^{d}} X_d^{-sp}|u|^p\,dx &= \int_{\mathbb{R}_{+}^{2}} X_2^{-sp}|U(X)|^p\,dX,\\
\mathcal{R}_{d,s,p,\tau}[u] &= \mathcal{R}_{2,s,p,\tau}[U]. \tag*{(22)}
\end{aligned} ∫ R + d X d − s p ∣ u ∣ p d x R d , s , p , τ [ u ] = ∫ R + 2 X 2 − s p ∣ U ( X ) ∣ p d X , = R 2 , s , p , τ [ U ] . (22)
It follows from (19) –(22) that
H d , s , p [ u ] ≥ A d , s , p H 2 , s , p [ U ] . (23) \mathcal{H}_{d,s,p}[u] \ge\mathcal{A}_{d,s,p}\mathcal{H}_{2,s,p}[U]. \tag*{(23)} H d , s , p [ u ] ≥ A d , s , p H 2 , s , p [ U ] . (23)
The function U U U has compact support in the half-plane and belongs to W s , p ( R + 2 ) W^{s,p}(\mathbb{R}_{+}^{2}) W s , p ( R + 2 ) by (19) . Zero extension and mollification give U j ∈ C c ∞ ( R + 2 ) U_j \in C^\infty_c(\mathbb{R}_{+}^{2}) U j ∈ C c ∞ ( R + 2 ) , supported in a fixed compact subset, with U j → U U_j \to U U j → U in W s , p W^{s,p} W s , p . Both weights are bounded there, so the weighted integrals also converge. The planar sharp inequality therefore applies to U U U . Equation (23) gives
H d , s , p [ u ] ≥ A d , s , p Λ 2 , s , p , τ R d , s , p , τ [ u ] . \mathcal{H}_{d,s,p}[u] \ge\mathcal{A}_{d,s,p}\Lambda_{2,s,p,\tau}\mathcal{R}_{d,s,p,\tau}[u]. H d , s , p [ u ] ≥ A d , s , p Λ 2 , s , p , τ R d , s , p , τ [ u ] .
Hence Λ d , s , p ≥ A d , s , p Λ 2 , s , p , τ \Lambda_{d,s,p} \ge\mathcal{A}_{d,s,p}\Lambda_{2,s,p,\tau} Λ d , s , p ≥ A d , s , p Λ 2 , s , p , τ .
For the reverse inequality, choose g ∈ C c ∞ ( R n ) g \in C^\infty_c(\mathbb{R}^n) g ∈ C c ∞ ( R n ) and v ∈ C c ∞ ( R + 2 ) v \in C^\infty_c(\mathbb{R}_{+}^{2}) v ∈ C c ∞ ( R + 2 ) with ∥ g ∥ L p = 1 \|g\|_{L^p}=1 ∥ g ∥ L p = 1 . Put
g R ( z ) = R − n / p g ( z / R ) , u R ( z , X ) = g R ( z ) v ( X ) . g_R(z)=R^{-n/p}g(z/R),\qquad u_R(z,X)=g_R(z)v(X). g R ( z ) = R − n / p g ( z / R ) , u R ( z , X ) = g R ( z ) v ( X ) .
The weighted integrals factor. In particular,
R d , s , p , τ [ u R ] = R 2 , s , p , τ [ v ] . (24) \mathcal{R}_{d,s,p,\tau}[u_R]=\mathcal{R}_{2,s,p,\tau}[v]. \tag*{(24)} R d , s , p , τ [ u R ] = R 2 , s , p , τ [ v ] . (24)
On the measure space with measure
d μ = d z d h d X d Y ( ∣ X − Y ∣ 2 + ∣ h ∣ 2 ) ( d + s p ) / 2 , d\mu=\frac{dz\,dh\,dX\,dY}{\left(|X-Y|^2+|h|^2\right)^{(d+sp)/2}}, d μ = ( ∣ X − Y ∣ 2 + ∣ h ∣ 2 ) ( d + s p ) /2 d z d h d X d Y ,
write
u R ( z , X ) − u R ( z − h , Y ) = A R + B R , u_R(z,X)-u_R(z-h,Y)=A_R+B_R, u R ( z , X ) − u R ( z − h , Y ) = A R + B R ,
A R = g R ( z ) ( v ( X ) − v ( Y ) ) , A_R=g_R(z)(v(X)-v(Y)), A R = g R ( z ) ( v ( X ) − v ( Y )) ,
B R = v ( Y ) ( g R ( z ) − g R ( z − h ) ) . B_R=v(Y)(g_R(z)-g_R(z-h)). B R = v ( Y ) ( g R ( z ) − g R ( z − h )) .
Equation (18) gives
∥ A R ∥ L p ( d μ ) p = A d , s , p [ v ] s , p , + p . (25) \|A_R\|_{L^p(d\mu)}^p=\mathcal{A}_{d,s,p}[v]_{s,p,+}^p. \tag*{(25)} ∥ A R ∥ L p ( d μ ) p = A d , s , p [ v ] s , p , + p . (25)
For h ≠ 0 h\ne0 h = 0 , integration in X X X gives
∫ R + 2 d X ( ∣ X − Y ∣ 2 + ∣ h ∣ 2 ) ( d + s p ) / 2 ≤ C d , s , p ∣ h ∣ − n − s p . \int_{\mathbb{R}_{+}^{2}}\frac{dX}{\left(|X-Y|^2+|h|^2\right)^{(d+sp)/2}}\le C_{d,s,p}|h|^{-n-sp}. ∫ R + 2 ( ∣ X − Y ∣ 2 + ∣ h ∣ 2 ) ( d + s p ) /2 d X ≤ C d , s , p ∣ h ∣ − n − s p .
Consequently,
∥ B R ∥ L p ( d μ ) p ≤ C d , s , p ∥ v ∥ L p ( R + 2 ) p ∫ R n ∫ R n ∣ g R ( z ) − g R ( z − h ) ∣ p ∣ h ∣ n + s p d h d z = C d , s , p R − s p ∥ v ∥ L p ( R + 2 ) p [ g ] W s , p ( R n ) p . (26) \begin{aligned}
\|B_R\|_{L^p(d\mu)}^p
&\le C_{d,s,p}\|v\|_{L^p(\mathbb{R}_{+}^{2})}^p
\int_{\mathbb{R}^n}\int_{\mathbb{R}^n}\frac{|g_R(z)-g_R(z-h)|^p}{|h|^{n+sp}}\,dh\,dz\\
&=C_{d,s,p}R^{-sp}\|v\|_{L^p(\mathbb{R}_{+}^{2})}^p[g]_{W^{s,p}(\mathbb{R}^n)}^p. \tag*{(26)}
\end{aligned} ∥ B R ∥ L p ( d μ ) p ≤ C d , s , p ∥ v ∥ L p ( R + 2 ) p ∫ R n ∫ R n ∣ h ∣ n + s p ∣ g R ( z ) − g R ( z − h ) ∣ p d h d z = C d , s , p R − s p ∥ v ∥ L p ( R + 2 ) p [ g ] W s , p ( R n ) p . (26)
Minkowski’s inequality and (25) –(26) give
∣ [ u R ] s , p , + − A d , s , p 1 / p [ v ] s , p , + ∣ ≤ ∥ B R ∥ L p ( d μ ) ⟶ 0. \left|[u_R]_{s,p,+}-\mathcal{A}_{d,s,p}^{1/p}[v]_{s,p,+}\right|\le\|B_R\|_{L^p(d\mu)}\longrightarrow0. [ u R ] s , p , + − A d , s , p 1/ p [ v ] s , p , + ≤ ∥ B R ∥ L p ( d μ ) ⟶ 0.
Together with (21) , this proves
H d , s , p [ u R ] ⟶ A d , s , p H 2 , s , p [ v ] . \mathcal{H}_{d,s,p}[u_R]\longrightarrow\mathcal{A}_{d,s,p}\mathcal{H}_{2,s,p}[v]. H d , s , p [ u R ] ⟶ A d , s , p H 2 , s , p [ v ] .
Divide by (24) , let R → ∞ R\to\infty R → ∞ , and minimize over v v v . This proves the reduction formula. □ \square □
Theorem 2.2 (Angular reduction for general powers). Let 0 < s < 1 0 < s < 1 0 < s < 1 , p > 1 p > 1 p > 1 , and τ > 0 \tau> 0 τ > 0 . Put
α s , p = 2 − s p p , K s , p ( q , θ , φ ) = 2 − 1 − s p / 2 ( cosh q − cos ( θ − φ ) ) 1 + s p / 2 . \alpha_{s,p} = \frac{2-sp}{p}, \qquad\mathcal{K}_{s,p}(q,\theta,\varphi) = \frac{2^{-1-sp/2}}{(\cosh q-\cos(\theta-\varphi))^{1+sp/2}}. α s , p = p 2 − s p , K s , p ( q , θ , φ ) = ( cosh q − cos ( θ − φ ) ) 1 + s p /2 2 − 1 − s p /2 . For ψ ∈ C c ∞ ( 0 , π ) \psi\in C^\infty_c(0,\pi) ψ ∈ C c ∞ ( 0 , π ) , define
A s , p [ ψ ] = ∫ 0 π ∫ 0 π ∫ R K s , p ( q , θ , φ ) ∣ e − α s , p q / 2 ψ ( θ ) − e α s , p q / 2 ψ ( φ ) ∣ p d q d θ d φ , \mathcal{A}_{s,p}[\psi] = \int_0^\pi\int_0^\pi\int_{\mathbb{R}} \mathcal{K}_{s,p}(q,\theta,\varphi)\left|e^{-\alpha_{s,p}q/2}\psi(\theta)-e^{\alpha_{s,p}q/2}\psi(\varphi)\right|^p\,dq\,d\theta\,d\varphi, A s , p [ ψ ] = ∫ 0 π ∫ 0 π ∫ R K s , p ( q , θ , φ ) e − α s , p q /2 ψ ( θ ) − e α s , p q /2 ψ ( φ ) p d q d θ d φ , B s , p [ ψ ] = A s , p [ ψ ] − D 2 , s , p ∫ 0 π ( sin θ ) − s p ∣ ψ ∣ p d θ . \mathcal{B}_{s,p}[\psi] = \mathcal{A}_{s,p}[\psi]-D_{2,s,p}\int_0^\pi(\sin\theta)^{-sp}|\psi|^p\,d\theta. B s , p [ ψ ] = A s , p [ ψ ] − D 2 , s , p ∫ 0 π ( sin θ ) − s p ∣ ψ ∣ p d θ . Then
Λ 2 , s , p , τ = inf ψ ∈ C c ∞ ( 0 , π ) ∖ { 0 } B s , p [ ψ ] ∫ 0 π ( sin θ ) τ − s p ∣ ψ ∣ p d θ . (27) \Lambda_{2,s,p,\tau}=\inf_{\psi\in C^\infty_c(0,\pi)\setminus\{0\}}\frac{\mathcal{B}_{s,p}[\psi]}{\int_0^\pi(\sin\theta)^{\tau-sp}|\psi|^p\,d\theta}. \tag*{(27)} Λ 2 , s , p , τ = ψ ∈ C c ∞ ( 0 , π ) ∖ { 0 } inf ∫ 0 π ( sin θ ) τ − s p ∣ ψ ∣ p d θ B s , p [ ψ ] . (27) The infimum is unchanged if ψ ≥ 0 \psi\ge0 ψ ≥ 0 is required. For p = 2 p=2 p = 2 , one has B s , 2 = b s , 0 \mathcal{B}_{s,2}=b_{s,0} B s , 2 = b s , 0 on the angular core. In particular, (27) extends the zero-frequency variational formula to every p > 1 p>1 p > 1 .
Proof. For u ∈ C c ∞ ( R + 2 ) u\in C^\infty_c(\mathbb{R}^2_+) u ∈ C c ∞ ( R + 2 ) , put
w ( t , θ ) = e α s , p t u ( e t cos θ , e t sin θ ) , V ( θ ) = ∥ w ( ⋅ , θ ) ∥ L p ( R ) . w(t,\theta)=e^{\alpha_{s,p}t}u(e^t\cos\theta,e^t\sin\theta),\qquad V(\theta)=\|w(\cdot,\theta)\|_{L^p(\mathbb{R})}. w ( t , θ ) = e α s , p t u ( e t cos θ , e t sin θ ) , V ( θ ) = ∥ w ( ⋅ , θ ) ∥ L p ( R ) . The polar distance identity gives
[ u ] s , p , + p = ∫ R ∫ 0 π ∫ 0 π K s , p ( q , θ , φ ) ∫ R ∣ e − α s , p q / 2 w ( t , θ ) − e α s , p q / 2 w ( t − q , φ ) ∣ p d t d θ d φ d q . (28) [u]_{s,p,+}^p=\int_{\mathbb{R}}\int_0^\pi\int_0^\pi\mathcal{K}_{s,p}(q,\theta,\varphi)\int_{\mathbb{R}}\left|e^{-\alpha_{s,p}q/2}w(t,\theta)-e^{\alpha_{s,p}q/2}w(t-q,\varphi)\right|^p\,dt\,d\theta\,d\varphi\,dq. \tag*{(28)} [ u ] s , p , + p = ∫ R ∫ 0 π ∫ 0 π K s , p ( q , θ , φ ) ∫ R e − α s , p q /2 w ( t , θ ) − e α s , p q /2 w ( t − q , φ ) p d t d θ d φ d q . (28) Both weighted integrals transform without a radial weight:
∫ R + 2 x 2 − s p ∣ u ∣ p d x = ∫ 0 π ( sin θ ) − s p V ( θ ) p d θ , \int_{\mathbb{R}^2_+}x_2^{-sp}|u|^p\,dx=\int_0^\pi(\sin\theta)^{-sp}V(\theta)^p\,d\theta, ∫ R + 2 x 2 − s p ∣ u ∣ p d x = ∫ 0 π ( sin θ ) − s p V ( θ ) p d θ , R 2 , s , p , τ [ u ] = ∫ 0 π ( sin θ ) τ − s p V ( θ ) p d θ . \mathcal{R}_{2,s,p,\tau}[u]=\int_0^\pi(\sin\theta)^{\tau-sp}V(\theta)^p\,d\theta. R 2 , s , p , τ [ u ] = ∫ 0 π ( sin θ ) τ − s p V ( θ ) p d θ . Translation invariance of the L p ( R ) L^p(\mathbb{R}) L p ( R ) norm and its reverse triangle inequality imply
∥ e − α s , p q / 2 w ( ⋅ , θ ) − e α s , p q / 2 w ( ⋅ − q , φ ) ∥ p ≥ ∣ e − α s , p q / 2 V ( θ ) − e α s , p q / 2 V ( φ ) ∣ . \left\|e^{-\alpha_{s,p}q/2}w(\cdot,\theta)-e^{\alpha_{s,p}q/2}w(\cdot-q,\varphi)\right\|_p\ge\left|e^{-\alpha_{s,p}q/2}V(\theta)-e^{\alpha_{s,p}q/2}V(\varphi)\right|. e − α s , p q /2 w ( ⋅ , θ ) − e α s , p q /2 w ( ⋅ − q , φ ) p ≥ e − α s , p q /2 V ( θ ) − e α s , p q /2 V ( φ ) . Consequently,
H 2 , s , p [ u ] ≥ B s , p [ V ] . (29) \mathcal{H}_{2,s,p}[u]\ge\mathcal{B}_{s,p}[V]. \tag*{(29)} H 2 , s , p [ u ] ≥ B s , p [ V ] . (29) We first check that the angular variational infimum applies to V V V . The map θ ↦ w ( ⋅ , θ ) \theta\mapsto w(\cdot,\theta) θ ↦ w ( ⋅ , θ ) is smooth with values in L p ( R ) L^p(\mathbb{R}) L p ( R ) . The norm is Lipschitz. Thus V V V is Lipschitz and has compact support in some I ⋐ ( 0 , π ) I\Subset(0,\pi) I ⋐ ( 0 , π ) . For functions supported in a fixed such interval,
A s , p [ f ] ≤ C I ( ∥ f ∥ L p ( R ) p + [ f ] W s , p ( R ) p ) . (30) \mathcal{A}_{s,p}[f]\le C_I\left(\|f\|_{L^p(\mathbb{R})}^p+[f]^p_{W^{s,p}(\mathbb{R})}\right). \tag*{(30)} A s , p [ f ] ≤ C I ( ∥ f ∥ L p ( R ) p + [ f ] W s , p ( R ) p ) . (30) where f f f is extended by zero. Indeed, for ∣ q ∣ < 1 |q|<1 ∣ q ∣ < 1 , split the drifted difference into f ( θ ) − f ( φ ) f(\theta)-f(\varphi) f ( θ ) − f ( φ ) and an error bounded by C ∣ q ∣ ( ∣ f ( θ ) ∣ + ∣ f ( φ ) ∣ ) C|q|(|f(\theta)|+|f(\varphi)|) C ∣ q ∣ ( ∣ f ( θ ) ∣ + ∣ f ( φ ) ∣ ) . Integration in q q q of the first kernel gives C ∣ θ − φ ∣ − 1 − s p C|\theta-\varphi|^{-1-sp} C ∣ θ − φ ∣ − 1 − s p . Integration in the angular difference of the error kernel gives C ∣ q ∣ p − 1 − s p C|q|^{p-1-sp} C ∣ q ∣ p − 1 − s p , which is integrable since s < 1 s<1 s < 1 . For ∣ q ∣ ≥ 1 |q|\ge1 ∣ q ∣ ≥ 1 ,
K s , p ( q , θ , φ ) e p ∣ α s , p q ∣ / 2 ≤ C e − min { 2 , s p } ∣ q ∣ . \mathcal{K}_{s,p}(q,\theta,\varphi)e^{p|\alpha_{s,p}q|/2}\le Ce^{-\min\{2,sp\}|q|}. K s , p ( q , θ , φ ) e p ∣ α s , p q ∣/2 ≤ C e − m i n { 2 , s p } ∣ q ∣ . This proves (30) . The p p p th root of A s , p \mathcal{A}_{s,p} A s , p is the norm of a linear difference map, so the same bound proves its continuity under W s , p W^{s,p} W s , p approximation. Mollification therefore approximates V V V by angular core functions, with convergence of A s , p \mathcal{A}_{s,p} A s , p and both weighted integrals. The
angular infimum in (27) , denoted temporarily by L L L , applies to V V V . Equation (29) gives Λ 2 , s , p , τ ≥ L \Lambda_{2,s,p,\tau} \ge L Λ 2 , s , p , τ ≥ L .
For the reverse inequality, fix a nonzero angular core function ψ \psi ψ . Choose χ ∈ C c ∞ ( R ) \chi\in C^\infty_c(\mathbb{R}) χ ∈ C c ∞ ( R ) with ∥ χ ∥ p = 1 \|\chi\|_p = 1 ∥ χ ∥ p = 1 , and put
χ R ( t ) = R − 1 / p χ ( t / R ) , w R ( t , θ ) = χ R ( t ) ψ ( θ ) . \chi_R(t) = R^{-1/p}\chi(t/R), \qquad w_R(t,\theta) = \chi_R(t)\psi(\theta). χ R ( t ) = R − 1/ p χ ( t / R ) , w R ( t , θ ) = χ R ( t ) ψ ( θ ) .
Its inverse polar transform belongs to C c ∞ ( R + 2 ) C^\infty_c(\mathbb{R}^2_+) C c ∞ ( R + 2 ) . The Hardy and remainder integrals factor exactly. On the measure space with measure K s , p d t d q d θ d φ K_{s,p}\,dt\,dq\,d\theta\,d\varphi K s , p d t d q d θ d φ , write
e − α s , p q / 2 w R ( t , θ ) − e α s , p q / 2 w R ( t − q , φ ) = A R + B R , e^{-\alpha_{s,p}q/2}w_R(t,\theta)-e^{\alpha_{s,p}q/2}w_R(t-q,\varphi)=A_R+B_R, e − α s , p q /2 w R ( t , θ ) − e α s , p q /2 w R ( t − q , φ ) = A R + B R ,
A R = χ R ( t ) ( e − α s , p q / 2 ψ ( θ ) − e α s , p q / 2 ψ ( φ ) ) , A_R=\chi_R(t)\left(e^{-\alpha_{s,p}q/2}\psi(\theta)-e^{\alpha_{s,p}q/2}\psi(\varphi)\right), A R = χ R ( t ) ( e − α s , p q /2 ψ ( θ ) − e α s , p q /2 ψ ( φ ) ) ,
B R = e α s , p q / 2 ψ ( φ ) ( χ R ( t ) − χ R ( t − q ) ) . B_R=e^{\alpha_{s,p}q/2}\psi(\varphi)\left(\chi_R(t)-\chi_R(t-q)\right). B R = e α s , p q /2 ψ ( φ ) ( χ R ( t ) − χ R ( t − q ) ) .
Then ∥ A R ∥ p p = A s , p [ ψ ] \|A_R\|_p^p=A_{s,p}[\psi] ∥ A R ∥ p p = A s , p [ ψ ] and
∥ χ R − χ R ( ⋅ − q ) ∥ p ≤ C min { ∣ q ∣ / R , 1 } . \|\chi_R-\chi_R(\mathord{\cdot}-q)\|_p\le C\min\{|q|/R,1\}. ∥ χ R − χ R ( ⋅ − q ) ∥ p ≤ C min { ∣ q ∣/ R , 1 } .
For ∣ q ∣ < 1 |q|<1 ∣ q ∣ < 1 , integration of K s , p K_{s,p} K s , p in θ \theta θ is bounded by C ∣ q ∣ − 1 − s p C|q|^{-1-sp} C ∣ q ∣ − 1 − s p . The preceding exponential estimate controls the tail. It follows that
∥ B R ∥ p p ≤ C ψ R p ( ∫ 0 1 q p − 1 − s p d q + ∫ 1 ∞ q p e − min { 2 , s p } q d q ) ⟶ 0. \|B_R\|_p^p\le\frac{C\psi}{R^p}\left(\int_0^1q^{p-1-sp}\,dq+\int_1^\infty q^p e^{-\min\{2,sp\}q}\,dq\right)\longrightarrow0. ∥ B R ∥ p p ≤ R p C ψ ( ∫ 0 1 q p − 1 − s p d q + ∫ 1 ∞ q p e − m i n { 2 , s p } q d q ) ⟶ 0.
Minkowski's inequality in (28) now gives
H 2 , s , p [ u R ] ⟶ B s , p [ ψ ] . \mathcal{H}_{2,s,p}[u_R]\longrightarrow\mathcal{B}_{s,p}[\psi]. H 2 , s , p [ u R ] ⟶ B s , p [ ψ ] .
Divide by the factored remainder and minimize over ψ \psi ψ . This proves the opposite inequality.
Finally, replacing ψ \psi ψ by ∣ ψ ∣ |\psi| ∣ ψ ∣ decreases A s , p A_{s,p} A s , p and preserves the weighted integrals. The modulus is a compactly supported W 1 , p W^{1,p} W 1 , p function. Nonnegative mollification and (30) give nonnegative angular core approximants. This proves the last infimum assertion. For p = 2 p=2 p = 2 , expand the square and use the evenness of K s , 2 = K s K_{s,2}=K_s K s , 2 = K s in q q q . The resulting expression is (7) with ξ = 0 \xi=0 ξ = 0 . □
Remark 2.3. The variational reduction holds for all p > 1 p>1 p > 1 . Attainment in the nonlinear critical angular completion is left open. The spectral realization, energy decomposition, and local limits below concern p = 2 p=2 p = 2 .
Logarithmic polar coordinates and the closed energy
Write x ∈ R + 2 x\in\mathbb{R}^2_+ x ∈ R + 2 as
x = e t ( cos θ , sin θ ) , t ∈ R , 0 < θ < π . (31) x=e^t(\cos\theta,\sin\theta),\qquad t\in\mathbb{R},\quad0<\theta<\pi. \tag*{(31)} x = e t ( cos θ , sin θ ) , t ∈ R , 0 < θ < π . (31)
The Jacobian is d x = e 2 t d t d θ dx=e^{2t}\,dt\,d\theta d x = e 2 t d t d θ . For y = e t ′ ( cos φ , sin φ ) y=e^{t'}(\cos\varphi,\sin\varphi) y = e t ′ ( cos φ , sin φ ) , we have
∣ x − y ∣ 2 = 2 e t + t ′ ( cosh ( t − t ′ ) − cos ( θ − φ ) ) . (32) |x-y|^2=2e^{t+t'}\left(\cosh(t-t')-\cos(\theta-\varphi)\right). \tag*{(32)} ∣ x − y ∣ 2 = 2 e t + t ′ ( cosh ( t − t ′ ) − cos ( θ − φ ) ) . (32)
Definition 3.1. For u ∈ C c ∞ ( R + 2 ) u\in C^\infty_c(\mathbb{R}^2_+) u ∈ C c ∞ ( R + 2 ) , its logarithmic polar transform is
w ( t , θ ) = e ( 1 − s ) t u ( e t cos θ , e t sin θ ) . (33) w(t,\theta)=e^{(1-s)t}u(e^t\cos\theta,e^t\sin\theta). \tag*{(33)} w ( t , θ ) = e ( 1 − s ) t u ( e t cos θ , e t sin θ ) . (33)
Lemma 3.2 (Exact transform). Let u u u and w w w be related by (33) . Then
R s , τ [ u ] = ∫ R ∫ 0 π m s , τ ( θ ) ∣ w ( t , θ ) ∣ 2 d θ d t . (34) \mathcal{R}_{s,\tau}[u]=\int_{\mathbb{R}}\int_0^\pi m_{s,\tau}(\theta)|w(t,\theta)|^2\,d\theta\,dt. \tag*{(34)} R s , τ [ u ] = ∫ R ∫ 0 π m s , τ ( θ ) ∣ w ( t , θ ) ∣ 2 d θ d t . (34) Moreover,
H s [ u ] = q s [ w ] , (35) \mathcal{H}_s[u]=q_s[w], \tag*{(35)} H s [ u ] = q s [ w ] , (35)
where
q s [ w ] = ∫ R 2 ∫ 0 π ∫ 0 π K s ( t − t ′ , θ , φ ) ∣ e − a s ( t − t ′ ) / 2 w ( t , θ ) − e a s ( t − t ′ ) / 2 w ( t ′ , φ ) ∣ 2 d θ d φ d t d t ′ − D 2 , s , 2 ∫ R ∫ 0 π ∣ w ( t , θ ) ∣ 2 ( sin θ ) 2 s d θ d t . (36) q_s[w]=\int_{\mathbb{R}^2}\int_0^\pi\int_0^\pi K_s(t-t',\theta,\varphi)\left|e^{-a_s(t-t')/2}w(t,\theta)-e^{a_s(t-t')/2}w(t',\varphi)\right|^2\,d\theta\,d\varphi\,dt\,dt'
-D_{2,s,2}\int_{\mathbb{R}}\int_0^\pi\frac{|w(t,\theta)|^2}{(\sin\theta)^{2s}}\,d\theta\,dt. \tag*{(36)} q s [ w ] = ∫ R 2 ∫ 0 π ∫ 0 π K s ( t − t ′ , θ , φ ) e − a s ( t − t ′ ) /2 w ( t , θ ) − e a s ( t − t ′ ) /2 w ( t ′ , φ ) 2 d θ d φ d t d t ′ − D 2 , s , 2 ∫ R ∫ 0 π ( sin θ ) 2 s ∣ w ( t , θ ) ∣ 2 d θ d t . (36)
Proof. By (33) ,
u ( e t cos θ , e t sin θ ) = e − ( 1 − s ) t w ( t , θ ) . u(e^t\cos\theta,e^t\sin\theta)=e^{-(1-s)t}w(t,\theta). u ( e t cos θ , e t sin θ ) = e − ( 1 − s ) t w ( t , θ ) .
Since x 2 = e t sin θ x_2=e^t\sin\theta x 2 = e t sin θ and ∣ x ∣ = e t |x|=e^t ∣ x ∣ = e t , substitution gives
e − 2 ( 1 − s ) t ∣ w ∣ 2 e 2 t e ( 2 s − 2 ) t ( sin θ ) 2 s − 2 e τ t = m s , τ ( θ ) ∣ w ∣ 2 . \frac{e^{-2(1-s)t}|w|^2e^{2t}}{e^{(2s-2)t}(\sin\theta)^{2s-2}e^{\tau t}}=m_{s,\tau}(\theta)|w|^2. e ( 2 s − 2 ) t ( sin θ ) 2 s − 2 e τ t e − 2 ( 1 − s ) t ∣ w ∣ 2 e 2 t = m s , τ ( θ ) ∣ w ∣ 2 .
This proves (34) . The Hardy term is
∫ R + 2 ∣ u ∣ 2 x 2 2 s d x = ∫ R ∫ 0 π ∣ w ∣ 2 ( sin θ ) 2 s d θ d t . (37) \int_{\mathbb{R}_+^2}\frac{|u|^2}{x_2^{2s}}\,dx=\int_{\mathbb{R}}\int_0^\pi\frac{|w|^2}{(\sin\theta)^{2s}}\,d\theta\,dt. \tag*{(37)} ∫ R + 2 x 2 2 s ∣ u ∣ 2 d x = ∫ R ∫ 0 π ( sin θ ) 2 s ∣ w ∣ 2 d θ d t . (37)
For the energy, (32) gives
e 2 t + 2 t ′ 2 1 + s e ( 1 + s ) ( t + t ′ ) ( cosh ( t − t ′ ) − cos ( θ − φ ) ) 1 + s = e a s ( t + t ′ ) K s ( t − t ′ , θ , φ ) . \frac{e^{2t+2t'}}{2^{1+s}e^{(1+s)(t+t')}(\cosh(t-t')-\cos(\theta-\varphi))^{1+s}}=e^{a_s(t+t')}K_s(t-t',\theta,\varphi). 2 1 + s e ( 1 + s ) ( t + t ′ ) ( cosh ( t − t ′ ) − cos ( θ − φ ) ) 1 + s e 2 t + 2 t ′ = e a s ( t + t ′ ) K s ( t − t ′ , θ , φ ) .
Furthermore,
e a s ( t + t ′ ) / 2 ( u ( x ) − u ( y ) ) = e − a s ( t − t ′ ) / 2 w ( t , θ ) − e a s ( t − t ′ ) / 2 w ( t ′ , φ ) . e^{a_s(t+t')/2}(u(x)-u(y))=e^{-a_s(t-t')/2}w(t,\theta)-e^{a_s(t-t')/2}w(t',\varphi). e a s ( t + t ′ ) /2 ( u ( x ) − u ( y )) = e − a s ( t − t ′ ) /2 w ( t , θ ) − e a s ( t − t ′ ) /2 w ( t ′ , φ ) .
Thus (36) and (37) imply (35) .
The following positive representation allows us to close the Hardy deficit. On the completed space, the two terms in (6) need not be finite separately.
Lemma 3.3 (Ground-state representation). Set
ω s ( x ) = x 2 s − 1 / 2 , v ( x ) = ω s ( x ) − 1 u ( x ) . \omega_s(x)=x_2^{s-1/2},\qquad v(x)=\omega_s(x)^{-1}u(x). ω s ( x ) = x 2 s − 1/2 , v ( x ) = ω s ( x ) − 1 u ( x ) . For u ∈ C c ∞ ( R + 2 ) u\in C^\infty_c(\mathbb{R}_+^2) u ∈ C c ∞ ( R + 2 ) ,
H s [ u ] = ∫ R + 2 ∫ R + 2 ∣ v ( x ) − v ( y ) ∣ 2 ω s ( x ) ω s ( y ) ∣ x − y ∣ 2 + 2 s d x d y . (38) \mathcal{H}_s[u]=\int_{\mathbb{R}_+^2}\int_{\mathbb{R}_+^2}\frac{|v(x)-v(y)|^2\omega_s(x)\omega_s(y)}{|x-y|^{2+2s}}\,dx\,dy. \tag*{(38)} H s [ u ] = ∫ R + 2 ∫ R + 2 ∣ x − y ∣ 2 + 2 s ∣ v ( x ) − v ( y ) ∣ 2 ω s ( x ) ω s ( y ) d x d y . (38) Consequently, if q = t − t ′ q=t-t' q = t − t ′ and z ( t , θ ) = ( sin θ ) 1 / 2 − s w ( t , θ ) z(t,\theta)=(\sin\theta)^{1/2-s}w(t,\theta) z ( t , θ ) = ( sin θ ) 1/2 − s w ( t , θ ) , then
q s [ w ] = ∫ R 2 ∫ 0 π ∫ 0 π K s ( q , θ , φ ) ( sin θ sin φ ) s − 1 / 2 ∣ e − q / 4 z ( t , θ ) − e q / 4 z ( t ′ , φ ) ∣ 2 d θ d φ d t d t ′ . (39) q_s[w]=\int_{\mathbb{R}^2}\int_0^\pi\int_0^\pi K_s(q,\theta,\varphi)(\sin\theta\sin\varphi)^{s-1/2}\left|e^{-q/4}z(t,\theta)-e^{q/4}z(t',\varphi)\right|^2\,d\theta\,d\varphi\,dt\,dt'. \tag*{(39)} q s [ w ] = ∫ R 2 ∫ 0 π ∫ 0 π K s ( q , θ , φ ) ( sin θ sin φ ) s − 1/2 e − q /4 z ( t , θ ) − e q /4 z ( t ′ , φ ) 2 d θ d φ d t d t ′ . (39) Both right-hand sides are nonnegative.
Proof. We first prove that ω s \omega_s ω s satisfies
2 p.v. ∫ R + 2 ω s ( x ) − ω s ( y ) ∣ x − y ∣ 2 + 2 s d y = D 2 , s , 2 x 2 − 2 s ω s ( x ) . (40) 2\operatorname{p.v.}\int_{\mathbb{R}_+^2}\frac{\omega_s(x)-\omega_s(y)}{|x-y|^{2+2s}}\,dy=D_{2,s,2}x_2^{-2s}\omega_s(x). \tag*{(40)} 2 p.v. ∫ R + 2 ∣ x − y ∣ 2 + 2 s ω s ( x ) − ω s ( y ) d y = D 2 , s , 2 x 2 − 2 s ω s ( x ) . (40)
Integration in the tangential variable gives
∫ R d z ( z 2 + a 2 ) 1 + s = π Γ ( s + 1 2 ) Γ ( 1 + s ) ∣ a ∣ − 1 − 2 s . \int_{\mathbb{R}}\frac{dz}{(z^2+a^2)^{1+s}}=\frac{\sqrt{\pi}\,\Gamma\left(s+\frac{1}{2}\right)}{\Gamma(1+s)}|a|^{-1-2s}. ∫ R ( z 2 + a 2 ) 1 + s d z = Γ ( 1 + s ) π Γ ( s + 2 1 ) ∣ a ∣ − 1 − 2 s .
To justify this reduction at the singularity, fix r 0 < x 2 / 2 r_0 < x_2/2 r 0 < x 2 /2 and subtract ∇ ω s ( x ) ⋅ ( y − x ) \nabla\omega_s(x)\cdot(y-x) ∇ ω s ( x ) ⋅ ( y − x ) on B r 0 ( x ) B_{r_0}(x) B r 0 ( x ) . The remainder is O ( ∣ x − y ∣ 2 ) O(|x-y|^2) O ( ∣ x − y ∣ 2 ) , and its integral over B ε ( x ) B_\varepsilon(x) B ε ( x ) is O ( ε 2 − 2 s ) O(\varepsilon^{2-2s}) O ( ε 2 − 2 s ) . The linear term vanishes under both the Euclidean cutoff and the symmetric one-dimensional cutoff. Outside B r 0 ( x ) B_{r_0}(x) B r 0 ( x ) , the integrals converge absolutely. Hence the two principal values agree. Put b = s − 1 2 b=s-\frac{1}{2} b = s − 2 1 and
C s = π Γ ( s + 1 2 ) Γ ( 1 + s ) . C_s=\frac{\sqrt{\pi}\,\Gamma\left(s+\frac{1}{2}\right)}{\Gamma(1+s)}. C s = Γ ( 1 + s ) π Γ ( s + 2 1 ) .
After integration and the substitution y 2 = x 2 r y_2=x_2r y 2 = x 2 r , the left-hand side of (40) , divided by two, equals
C s x 2 b − 2 s p.v. ∫ 0 ∞ 1 − r b ∣ 1 − r ∣ 1 + 2 s d r . C_sx_2^{b-2s}\operatorname{p.v.}\int_0^\infty\frac{1-r^b}{|1-r|^{1+2s}}\,dr. C s x 2 b − 2 s p.v. ∫ 0 ∞ ∣1 − r ∣ 1 + 2 s 1 − r b d r .
Split the integral at one and invert the variable in the second part. Since 2 s − 1 = 2 b 2s-1=2b 2 s − 1 = 2 b , we obtain
p.v. ∫ 0 ∞ 1 − r b ∣ 1 − r ∣ 1 + 2 s d r = ∫ 0 1 1 − 2 r b + r 2 b ( 1 − r ) 1 + 2 s d r = ∫ 0 1 ( 1 − r b ) 2 ( 1 − r ) 1 + 2 s d r . \begin{aligned}
\operatorname{p.v.}\int_0^\infty\frac{1-r^b}{|1-r|^{1+2s}}\,dr=\int_0^1\frac{1-2r^b+r^{2b}}{(1-r)^{1+2s}}\,dr \\
&=\int_0^1\frac{(1-r^b)^2}{(1-r)^{1+2s}}\,dr.
\end{aligned} p.v. ∫ 0 ∞ ∣1 − r ∣ 1 + 2 s 1 − r b d r = ∫ 0 1 ( 1 − r ) 1 + 2 s 1 − 2 r b + r 2 b d r = ∫ 0 1 ( 1 − r ) 1 + 2 s ( 1 − r b ) 2 d r .
The cutoff error caused by inversion is O ( ε 2 − 2 s ) O(\varepsilon^{2-2s}) O ( ε 2 − 2 s ) . The numerator has a double zero at one, and 2 b = 2 s − 1 > − 1 2b=2s-1>-1 2 b = 2 s − 1 > − 1 . Thus the last integral is finite. Formula (5) now gives (40) . For the truncated energies, use the identity
∣ ω s ( x ) v ( x ) − ω s ( y ) v ( y ) ∣ 2 − ω s ( x ) ω s ( y ) ∣ v ( x ) − v ( y ) ∣ 2 = ω s ( x ) ( ω s ( x ) − ω s ( y ) ) ∣ v ( x ) ∣ 2 + ω s ( y ) ( ω s ( y ) − ω s ( x ) ) ∣ v ( y ) ∣ 2 . \begin{aligned}
|\omega_s(x)v(x)-\omega_s(y)v(y)|^2-\omega_s(x)\omega_s(y)|v(x)-v(y)|^2 \\
&=\omega_s(x)(\omega_s(x)-\omega_s(y))|v(x)|^2+\omega_s(y)(\omega_s(y)-\omega_s(x))|v(y)|^2.
\end{aligned} ∣ ω s ( x ) v ( x ) − ω s ( y ) v ( y ) ∣ 2 − ω s ( x ) ω s ( y ) ∣ v ( x ) − v ( y ) ∣ 2 = ω s ( x ) ( ω s ( x ) − ω s ( y )) ∣ v ( x ) ∣ 2 + ω s ( y ) ( ω s ( y ) − ω s ( x )) ∣ v ( y ) ∣ 2 .
Interchanging x x x and y y y in the second term, the difference of the truncated energies becomes
2 ∫ R + 2 ∣ v ( x ) ∣ 2 ω s ( x ) ∫ y ∈ R + 2 ∣ x − y ∣ > ε ω s ( x ) − ω s ( y ) ∣ x − y ∣ 2 + 2 s d y d x . 2\int_{\mathbb{R}^2_+}|v(x)|^2\omega_s(x)\int_{\substack{y\in\mathbb{R}^2_+\\|x-y|>\varepsilon}}\frac{\omega_s(x)-\omega_s(y)}{|x-y|^{2+2s}}\,dy\,dx. 2 ∫ R + 2 ∣ v ( x ) ∣ 2 ω s ( x ) ∫ y ∈ R + 2 ∣ x − y ∣ > ε ∣ x − y ∣ 2 + 2 s ω s ( x ) − ω s ( y ) d y d x .
Let K = supp u K=\operatorname{supp}u K = supp u and choose r 0 < 1 2 dist ( K , ∂ R + 2 ) r_0<\frac{1}{2}\operatorname{dist}(K,\partial\mathbb{R}^2_+) r 0 < 2 1 dist ( K , ∂ R + 2 ) . For x ∈ K x\in K x ∈ K , put h = y − x h=y-x h = y − x . For ε < r 0 \varepsilon<r_0 ε < r 0 , oddness gives
∫ ε < ∣ h ∣ < r 0 ω s ( x ) − ω s ( x + h ) ∣ h ∣ 2 + 2 s d h = − ∫ ε < ∣ h ∣ < r 0 ω s ( x + h ) − ω s ( x ) − ∇ ω s ( x ) ⋅ h ∣ h ∣ 2 + 2 s d h . \int_{\varepsilon<|h|<r_0}\frac{\omega_s(x)-\omega_s(x+h)}{|h|^{2+2s}}\,dh
=-\int_{\varepsilon<|h|<r_0}\frac{\omega_s(x+h)-\omega_s(x)-\nabla\omega_s(x)\cdot h}{|h|^{2+2s}}\,dh. ∫ ε < ∣ h ∣ < r 0 ∣ h ∣ 2 + 2 s ω s ( x ) − ω s ( x + h ) d h = − ∫ ε < ∣ h ∣ < r 0 ∣ h ∣ 2 + 2 s ω s ( x + h ) − ω s ( x ) − ∇ ω s ( x ) ⋅ h d h .
The last integral is bounded uniformly on K K K by
C ∫ 0 r 0 r 1 − 2 s d r < ∞ . C\int_0^{r_0}r^{1-2s}\,dr<\infty. C ∫ 0 r 0 r 1 − 2 s d r < ∞.
On ∣ x − y ∣ ≥ r 0 |x-y|\ge r_0 ∣ x − y ∣ ≥ r 0 , absolute integrability follows from s − 1 2 > − 1 s-\frac{1}{2}>-1 s − 2 1 > − 1 at the boundary and s − 1 2 < 2 s s-\frac{1}{2}<2s s − 2 1 < 2 s at infinity, uniformly for x ∈ K x\in K x ∈ K . The same Taylor estimate shows that the truncated potentials converge uniformly on K K K . We may therefore let ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 in the preceding identity. By (40) , the potential term converges to the Hardy term; the positive energy converges by monotone convergence. This proves (38) .
Finally,
v ( e t cos θ , e t sin θ ) = e − t / 2 ( sin θ ) 1 / 2 − s w ( t , θ ) = e − t / 2 z ( t , θ ) . v(e^t\cos\theta,e^t\sin\theta)=e^{-t/2}(\sin\theta)^{1/2-s}w(t,\theta)=e^{-t/2}z(t,\theta). v ( e t cos θ , e t sin θ ) = e − t /2 ( sin θ ) 1/2 − s w ( t , θ ) = e − t /2 z ( t , θ ) .
Furthermore,
ω s ( x ) ω s ( y ) d x d y ∣ x − y ∣ 2 + 2 s = e ( t + t ′ ) / 2 K s ( q , θ , φ ) ( sin θ sin φ ) s − 1 / 2 d t d t ′ d θ d φ . \frac{\omega_s(x)\omega_s(y)\,dx\,dy}{|x-y|^{2+2s}}
=e^{(t+t')/2}K_s(q,\theta,\varphi)(\sin\theta\sin\varphi)^{s-1/2}\,dt\,dt'\,d\theta\,d\varphi. ∣ x − y ∣ 2 + 2 s ω s ( x ) ω s ( y ) d x d y = e ( t + t ′ ) /2 K s ( q , θ , φ ) ( sin θ sin φ ) s − 1/2 d t d t ′ d θ d φ .
Factoring e − ( t + t ′ ) / 4 e^{-(t+t')/4} e − ( t + t ′ ) /4 from v ( x ) − v ( y ) v(x)-v(y) v ( x ) − v ( y ) cancels the radial weight and gives (39) . □
Lemma 3.4 (Local kernel comparison). If ∣ q ∣ ≤ 1 |q| \le1 ∣ q ∣ ≤ 1 and ∣ θ − φ ∣ ≤ 1 |\theta-\varphi| \le1 ∣ θ − φ ∣ ≤ 1 , then
c ( q 2 + ∣ θ − φ ∣ 2 ) ≤ cosh q − cos ( θ − φ ) ≤ C ( q 2 + ∣ θ − φ ∣ 2 ) . (41) c(q^2+|\theta-\varphi|^2)\le\cosh q-\cos(\theta-\varphi)\le C(q^2+|\theta-\varphi|^2). \tag*{(41)} c ( q 2 + ∣ θ − φ ∣ 2 ) ≤ cosh q − cos ( θ − φ ) ≤ C ( q 2 + ∣ θ − φ ∣ 2 ) . (41) If ∣ q ∣ ≥ 1 |q|\ge1 ∣ q ∣ ≥ 1 , then
K s ( q , θ , φ ) ≤ C e − ( 1 + s ) ∣ q ∣ , K_s(q,\theta,\varphi)\le Ce^{-(1+s)|q|}, K s ( q , θ , φ ) ≤ C e − ( 1 + s ) ∣ q ∣ , K s ( q , θ , φ ) e ( 1 − s ) ∣ q ∣ ≤ C e − 2 s ∣ q ∣ , (42) K_s(q,\theta,\varphi)e^{(1-s)|q|}\le Ce^{-2s|q|}, \tag*{(42)} K s ( q , θ , φ ) e ( 1 − s ) ∣ q ∣ ≤ C e − 2 s ∣ q ∣ , (42) K s ( q , θ , φ ) e ∣ q ∣ / 2 ≤ C e − ( s + 1 / 2 ) ∣ q ∣ . K_s(q,\theta,\varphi)e^{|q|/2}\le Ce^{-(s+1/2)|q|}. K s ( q , θ , φ ) e ∣ q ∣/2 ≤ C e − ( s + 1/2 ) ∣ q ∣ . Consequently, if I ′ ⋐ I ⋐ ( 0 , π ) I'\Subset I\Subset(0,\pi) I ′ ⋐ I ⋐ ( 0 , π ) , then
[ w ] H s ( R × I ′ ) 2 ≤ C I ′ , I ( q s [ w ] + ∥ w ∥ L 2 ( R × I ) 2 ) (43) [w]^2_{H^s(\mathbb{R}\times I')}\le C_{I',I}\left(q_s[w]+\|w\|^2_{L^2(\mathbb{R}\times I)}\right) \tag*{(43)} [ w ] H s ( R × I ′ ) 2 ≤ C I ′ , I ( q s [ w ] + ∥ w ∥ L 2 ( R × I ) 2 ) (43) for w ∈ C c ∞ ( R × I ) w\in C^\infty_c(\mathbb{R}\times I) w ∈ C c ∞ ( R × I ) . Here
[ w ] H s ( R × I ′ ) 2 = ∫ R × I ′ ∫ R × I ′ ∣ w ( t , θ ) − w ( t ′ , φ ) ∣ 2 ( ( t − t ′ ) 2 + ( θ − φ ) 2 ) 1 + s d t d θ d t ′ d φ . [w]^2_{H^s(\mathbb{R}\times I')}=\int_{\mathbb{R}\times I'}\int_{\mathbb{R}\times I'}\frac{|w(t,\theta)-w(t',\varphi)|^2}{((t-t')^2+(\theta-\varphi)^2)^{1+s}}\,dt\,d\theta\,dt'\,d\varphi. [ w ] H s ( R × I ′ ) 2 = ∫ R × I ′ ∫ R × I ′ (( t − t ′ ) 2 + ( θ − φ ) 2 ) 1 + s ∣ w ( t , θ ) − w ( t ′ , φ ) ∣ 2 d t d θ d t ′ d φ . Proof. Taylor’s formula gives
cosh q − 1 = q 2 2 + O ( q 4 ) , 1 − cos α = α 2 2 + O ( α 4 ) . \cosh q-1=\frac{q^2}{2}+O(q^4),\qquad1-\cos\alpha=\frac{\alpha^2}{2}+O(\alpha^4). cosh q − 1 = 2 q 2 + O ( q 4 ) , 1 − cos α = 2 α 2 + O ( α 4 ) . Since cosh q − cos α = ( cosh q − 1 ) + ( 1 − cos α ) \cosh q-\cos\alpha=(\cosh q-1)+(1-\cos\alpha) cosh q − cos α = ( cosh q − 1 ) + ( 1 − cos α ) , (41) follows. For ∣ q ∣ ≥ 1 |q|\ge1 ∣ q ∣ ≥ 1 , cosh q − cos α ≥ c e ∣ q ∣ \cosh q-\cos\alpha\ge ce^{|q|} cosh q − cos α ≥ c e ∣ q ∣ . This proves (42) , including the estimate with the drift factor.
For θ , φ ∈ I ′ \theta,\varphi\in I' θ , φ ∈ I ′ , absorb the angular weight in (39) into the square. The coefficients are smooth and bounded above and below. For ∣ q ∣ < 1 |q|<1 ∣ q ∣ < 1 , the mean-value theorem gives
∣ e − q / 4 ( sin φ / sin θ ) ( s − 1 / 2 ) / 2 w ( t , θ ) − e q / 4 ( sin θ / sin φ ) ( s − 1 / 2 ) / 2 w ( t − q , φ ) ∣ 2 \left|e^{-q/4}(\sin\varphi/\sin\theta)^{(s-1/2)/2}w(t,\theta)-e^{q/4}(\sin\theta/\sin\varphi)^{(s-1/2)/2}w(t-q,\varphi)\right|^2 e − q /4 ( sin φ / sin θ ) ( s − 1/2 ) /2 w ( t , θ ) − e q /4 ( sin θ / sin φ ) ( s − 1/2 ) /2 w ( t − q , φ ) 2 ≥ 1 2 ∣ w ( t , θ ) − w ( t − q , φ ) ∣ 2 − C ( q 2 + ∣ θ − φ ∣ 2 ) ( ∣ w ( t , θ ) ∣ 2 + ∣ w ( t − q , φ ) ∣ 2 ) . \ge\frac{1}{2}|w(t,\theta)-w(t-q,\varphi)|^2-C(q^2+|\theta-\varphi|^2)(|w(t,\theta)|^2+|w(t-q,\varphi)|^2). ≥ 2 1 ∣ w ( t , θ ) − w ( t − q , φ ) ∣ 2 − C ( q 2 + ∣ θ − φ ∣ 2 ) ( ∣ w ( t , θ ) ∣ 2 + ∣ w ( t − q , φ ) ∣ 2 ) . By (41) , the error is integrable near the origin, since
∫ q 2 + α 2 < 1 ( q 2 + α 2 ) − s d q d α < ∞ . \int_{q^2+\alpha^2<1}(q^2+\alpha^2)^{-s}\,dq\,d\alpha<\infty. ∫ q 2 + α 2 < 1 ( q 2 + α 2 ) − s d q d α < ∞. Outside this neighborhood, the flat seminorm is bounded by an L 2 ( R × I ) L^2(\mathbb{R}\times I) L 2 ( R × I ) term. Restricting the positive cylinder integral to R 2 × ( I ′ ) 2 \mathbb{R}^2\times(I')^2 R 2 × ( I ′ ) 2 proves (43) . □
The closed deficit space
For ρ > 0 \rho>0 ρ > 0 , the normalized dilation and its transform are
u ρ ( x ) = ρ 1 − s u ( ρ x ) , w ρ ( t , θ ) = w ( t + log ρ , θ ) . u^\rho(x)=\rho^{1-s}u(\rho x),\qquad w^\rho(t,\theta)=w(t+\log\rho,\theta). u ρ ( x ) = ρ 1 − s u ( ρ x ) , w ρ ( t , θ ) = w ( t + log ρ , θ ) .
Both H s \mathcal{H}_s H s and R s , τ \mathcal{R}_{s,\tau} R s , τ are invariant under this normalization.
Lemma 3.5 (Core coercivity). Let C = C c ∞ ( R × ( 0 , π ) ) \mathcal{C}=C^\infty_c(\mathbb{R}\times(0,\pi)) C = C c ∞ ( R × ( 0 , π )) . The image under (33) of C c ∞ ( R + 2 ) C^\infty_c(\mathbb{R}^2_+) C c ∞ ( R + 2 ) is C \mathcal{C} C . The form q s q_s q s is nonnegative and closable. For every σ > 0 \sigma>0 σ > 0 ,
q s [ w ] ≥ c s , σ ∫ R ∫ 0 π ( sin θ ) σ − 2 s ∣ w ∣ 2 d θ d t (44) q_s[w]\ge c_{s,\sigma}\int_{\mathbb{R}}\int_0^\pi(\sin\theta)^{\sigma-2s}|w|^2\,d\theta\,dt \tag*{(44)} q s [ w ] ≥ c s , σ ∫ R ∫ 0 π ( sin θ ) σ − 2 s ∣ w ∣ 2 d θ d t (44) with c s , σ > 0 c_{s,\sigma}>0 c s , σ > 0 .
Proof. The map in (31) is a diffeomorphism from R × ( 0 , π ) \mathbb{R}\times(0,\pi) R × ( 0 , π ) onto R + 2 \mathbb{R}^2_+ R + 2 and preserves compact support in both directions. Nonnegativity follows from (39) .
Fix σ > 0 \sigma>0 σ > 0 and choose 0 < σ 0 < min { σ , 1 + 2 s } 0<\sigma_0<\min\{\sigma,1+2s\} 0 < σ 0 < min { σ , 1 + 2 s } . By [16 Theorem 1] , (1) holds with c s , σ 0 > 0 c_{s,\sigma_0}>0 c s , σ 0 > 0 in the ordered-integral normalization of (4) . Since m s , σ ≤ m s , σ 0 m_{s,\sigma}\le m_{s,\sigma_0} m s , σ ≤ m s , σ 0 , Lemma 3.2 gives (44) . For complex w w w , apply the inequality to its real and imaginary parts.
Let d ν s d\nu_s d ν s be the positive measure in (39) , including the kernel and angular factors, and put
( T w ) ( t , t ′ , θ , φ ) = e − ( t − t ′ ) / 4 ( sin θ ) 1 / 2 − s w ( t , θ ) − e ( t − t ′ ) / 4 ( sin φ ) 1 / 2 − s w ( t ′ , φ ) . (Tw)(t,t',\theta,\varphi)=e^{-(t-t')/4}(\sin\theta)^{1/2-s}w(t,\theta)-e^{(t-t')/4}(\sin\varphi)^{1/2-s}w(t',\varphi). ( T w ) ( t , t ′ , θ , φ ) = e − ( t − t ′ ) /4 ( sin θ ) 1/2 − s w ( t , θ ) − e ( t − t ′ ) /4 ( sin φ ) 1/2 − s w ( t ′ , φ ) .
Then q s [ w ] = ∥ T w ∥ L 2 ( d ν s ) q_s[w]=\|Tw\|_{L^2(d\nu_s)} q s [ w ] = ∥ T w ∥ L 2 ( d ν s ) on C \mathcal{C} C . Suppose that w n → 0 w_n\to0 w n → 0 in L 2 ( m s , τ ) L^2(m_{s,\tau}) L 2 ( m s , τ ) and T w n → F Tw_n\to F T w n → F in L 2 ( d ν s ) L^2(d\nu_s) L 2 ( d ν s ) . The measure d ν s d\nu_s d ν s is σ \sigma σ -finite and absolutely continuous off the null diagonal. Passing to subsequences gives w n → 0 w_n\to0 w n → 0 almost everywhere and T w n → F Tw_n\to F T w n → F d ν s d\nu_s d ν s -almost everywhere. Product-null sets are d ν s d\nu_s d ν s -null. The formula for T T T therefore gives F = 0 F=0 F = 0 . Thus T T T , and hence q s q_s q s , is closable. □
Denote the closed form domain by X s \mathcal{X}_s X s . It is the completion of C \mathcal{C} C under
∥ w ∥ X s 2 = q s [ w ] + ∫ R ∫ 0 π m s , τ ∣ w ∣ 2 d θ d t . \|w\|_{\mathcal{X}_s}^2=q_s[w]+\int_{\mathbb{R}}\int_0^\pi m_{s,\tau}|w|^2\,d\theta\,dt. ∥ w ∥ X s 2 = q s [ w ] + ∫ R ∫ 0 π m s , τ ∣ w ∣ 2 d θ d t .
By (44) , this norm is equivalent to q s 1 / 2 q_s^{1/2} q s 1/2 for every τ > 0 \tau>0 τ > 0 . Thus the abstract completion is independent of τ \tau τ . Its representatives also agree: a common core sequence converges in the stronger weight corresponding to any 0 < σ < min { τ 1 , τ 2 } 0<\sigma<\min\{\tau_1,\tau_2\} 0 < σ < min { τ 1 , τ 2 } . Consequently, X s ↪ L 2 ( m s , τ ) \mathcal{X}_s\hookrightarrow L^2(m_{s,\tau}) X s ↪ L 2 ( m s , τ ) continuously for every τ > 0 \tau>0 τ > 0 . By Lemma 3.2 and density, the logarithmic polar map extends to an isometric isomorphism from the energy completion in Theorem 1.2 onto X s \mathcal{X}_s X s .
Mellin fibers and angular compactness
For w ∈ C w\in\mathcal{C} w ∈ C , take the unitary Fourier transform in t t t :
w ^ ( ξ , θ ) = 1 2 π ∫ R e − i t ξ w ( t , θ ) d t . \widehat{w}(\xi,\theta)=\frac{1}{\sqrt{2\pi}}\int_{\mathbb{R}}e^{-it\xi}w(t,\theta)\,dt. w ( ξ , θ ) = 2 π 1 ∫ R e − i t ξ w ( t , θ ) d t .
Put H s , τ = L 2 ( ( 0 , π ) , m s , τ d θ ) H_{s,\tau}=L^2((0,\pi),m_{s,\tau}\,d\theta) H s , τ = L 2 (( 0 , π ) , m s , τ d θ ) . For ψ ∈ C ∞ ( 0 , π ) \psi\in C^\infty(0,\pi) ψ ∈ C ∞ ( 0 , π ) , the positive fiber form is
g s , ξ [ ψ ] = ∫ 0 π ∫ 0 π ∫ R K s ( q , θ , φ ) ( sin θ sin φ ) s − 1 / 2 ∣ e − q / 4 ( sin θ ) 1 / 2 − s ψ ( θ ) − e q / 4 − i ξ q ( sin φ ) 1 / 2 − s ψ ( φ ) ∣ 2 d q d θ d φ . (45) g_{s,\xi}[\psi]=\int_0^\pi\int_0^\pi\int_{\mathbb{R}}K_s(q,\theta,\varphi)(\sin\theta\sin\varphi)^{s-1/2}
\left|e^{-q/4}(\sin\theta)^{1/2-s}\psi(\theta)-e^{q/4-i\xi q}(\sin\varphi)^{1/2-s}\psi(\varphi)\right|^2\,dq\,d\theta\,d\varphi. \tag*{(45)} g s , ξ [ ψ ] = ∫ 0 π ∫ 0 π ∫ R K s ( q , θ , φ ) ( sin θ sin φ ) s − 1/2 e − q /4 ( sin θ ) 1/2 − s ψ ( θ ) − e q /4 − i ξ q ( sin φ ) 1/2 − s ψ ( φ ) 2 d q d θ d φ . (45)
We compare the two expressions with a common kernel cutoff. The calculation below uses (40) and gives
g s , ξ [ ψ ] = b s , ξ [ ψ ] . (46) g_{s,\xi}[\psi]=b_{s,\xi}[\psi]. \tag*{(46)} g s , ξ [ ψ ] = b s , ξ [ ψ ] . (46)
Put b = s − 1 2 b=s-\frac{1}{2} b = s − 2 1 , and let K s ε K_s^\varepsilon K s ε be K s K_s K s multiplied by the indicator of q 2 + ∣ θ − φ ∣ 2 > ε 2 q^2+|\theta-\varphi|^2>\varepsilon^2 q 2 + ∣ θ − φ ∣ 2 > ε 2 . Expansion of the positive square gives
g s , ξ ε [ ψ ] = ∫ K s ε { e − q / 2 ( sin φ sin θ ) b ∣ ψ ( θ ) ∣ 2 + e q / 2 ( sin θ sin φ ) b ∣ ψ ( φ ) ∣ 2 − 2 Re ( e i ξ q ψ ( θ ) ψ ( φ ) ‾ ) } d q d θ d φ . g_{s,\xi}^\varepsilon[\psi]=\int K_s^\varepsilon\left\{e^{-q/2}\left(\frac{\sin\varphi}{\sin\theta}\right)^b|\psi(\theta)|^2+e^{q/2}\left(\frac{\sin\theta}{\sin\varphi}\right)^b|\psi(\varphi)|^2-2\operatorname{Re}\left(e^{i\xi q}\psi(\theta)\overline{\psi(\varphi)}\right)\right\}\,dq\,d\theta\,d\varphi. g s , ξ ε [ ψ ] = ∫ K s ε { e − q /2 ( sin θ sin φ ) b ∣ ψ ( θ ) ∣ 2 + e q /2 ( sin φ sin θ ) b ∣ ψ ( φ ) ∣ 2 − 2 Re ( e i ξ q ψ ( θ ) ψ ( φ ) ) } d q d θ d φ .
The transformed ground-state equation reads
2 lim ε → 0 ∫ 0 π ∫ R K s ε ( q , θ , φ ) [ e − q / 2 ( sin φ sin θ ) b − cosh ( a s q ) ] d q d φ = − D 2 , s , 2 ( sin θ ) − 2 s . (47) 2\lim_{\varepsilon\to0}\int_0^\pi\int_{\mathbb{R}}K_s^\varepsilon(q,\theta,\varphi)\left[e^{-q/2}\left(\frac{\sin\varphi}{\sin\theta}\right)^b-\cosh(a_s q)\right]\,dq\,d\varphi=-D_{2,s,2}(\sin\theta)^{-2s}. \tag*{(47)} 2 ε → 0 lim ∫ 0 π ∫ R K s ε ( q , θ , φ ) [ e − q /2 ( sin θ sin φ ) b − cosh ( a s q ) ] d q d φ = − D 2 , s , 2 ( sin θ ) − 2 s . (47)
Indeed, take x = ( cos θ , sin θ ) x=(\cos\theta,\sin\theta) x = ( cos θ , sin θ ) in (40) and put y = e − q ( cos φ , sin φ ) y=e^{-q}(\cos\varphi,\sin\varphi) y = e − q ( cos φ , sin φ ) . Before symmetrization in q q q , the two terms in brackets are e − q / 2 ( sin φ / sin θ ) b e^{-q/2}(\sin\varphi/\sin\theta)^b e − q /2 ( sin φ / sin θ ) b and e − a s q e^{-a_s q} e − a s q . The evenness of K s K_s K s changes the latter into cosh ( a s q ) \cosh(a_s q) cosh ( a s q ) . This proves (47) .
The angular cutoff agrees with the Euclidean principal value in (40) . Indeed, put α = θ − φ \alpha=\theta-\varphi α = θ − φ and r = ( q 2 + α 2 ) 1 / 2 r=(q^2+\alpha^2)^{1/2} r = ( q 2 + α 2 ) 1/2 . For fixed θ ∈ ( 0 , π ) \theta\in(0,\pi) θ ∈ ( 0 , π ) ,
∣ x − y ∣ 2 = 2 e − q ( cosh q − cos α ) = r 2 + O ( r 3 ) . |x-y|^2=2e^{-q}(\cosh q-\cos\alpha)=r^2+O(r^3). ∣ x − y ∣ 2 = 2 e − q ( cosh q − cos α ) = r 2 + O ( r 3 ) .
The numerator in the transformed principal-value integral is a linear function of ( q , α ) (q,\alpha) ( q , α ) plus O ( r 2 ) O(r^2) O ( r 2 ) . The linear part has zero integral over every disk centered at the origin because K s K_s K s is even in q q q and α \alpha α . The symmetric difference between the pulled-back Euclidean cutoff and the disk r > ε r>\varepsilon r > ε is contained in an annulus of width O ( ε 2 ) O(\varepsilon^2) O ( ε 2 ) . Its contribution from the linear part is O ( ε 2 − 2 s ) O(\varepsilon^{2-2s}) O ( ε 2 − 2 s ) . The quadratic remainder on r < 2 ε r<2\varepsilon r < 2 ε is also O ( ε 2 − 2 s ) O(\varepsilon^{2-2s}) O ( ε 2 − 2 s ) . Both tend to zero because s < 1 s<1 s < 1 . Thus the two principal values agree. In the first two diagonal terms of g s , ξ ε g_{s,\xi}^{\varepsilon} g s , ξ ε , interchange ( q , θ , ϕ ) (q,\theta,\phi) ( q , θ , ϕ ) with ( − q , ϕ , θ ) (-q,\phi,\theta) ( − q , ϕ , θ ) in one copy. Their sum has twice the first displayed coefficient. Apply (47) . The cross term already equals the cross term in (7) . Letting ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 proves (46) . Thus (7) is understood as the limit of the complete truncated expression. Formula (45) is used on the completed domain.
Proposition 4.1 (Direct-integral identity). For w ∈ C w\in\mathcal{C} w ∈ C ,
q s [ w ] = ∫ R b s , ξ [ w ^ ( ξ , ⋅ ) ] d ξ . (48) q_s[w]=\int_{\mathbb{R}}b_{s,\xi}[\widehat{w}(\xi,\cdot)]\,d\xi. \tag*{(48)} q s [ w ] = ∫ R b s , ξ [ w ( ξ , ⋅ )] d ξ . (48) Also,
∫ R ∫ 0 π m s , τ ∣ w ∣ 2 = ∫ R ∫ 0 π m s , τ ∣ w ^ ∣ 2 . (49) \int_{\mathbb{R}}\int_0^\pi m_{s,\tau}|w|^2=\int_{\mathbb{R}}\int_0^\pi m_{s,\tau}|\widehat{w}|^2. \tag*{(49)} ∫ R ∫ 0 π m s , τ ∣ w ∣ 2 = ∫ R ∫ 0 π m s , τ ∣ w ∣ 2 . (49) Proof. Let q = t − t ′ q=t-t' q = t − t ′ . Delete q 2 + ∣ θ − ϕ ∣ 2 < ε 2 q^2+|\theta-\phi|^2<\varepsilon^2 q 2 + ∣ θ − ϕ ∣ 2 < ε 2 from the positive form (39) . Denote the result by q s ε [ w ] q_s^\varepsilon[w] q s ε [ w ] . Denote by g s , ξ ε g_{s,\xi}^\varepsilon g s , ξ ε the same truncation of (45) . The truncated kernel is locally bounded, and its q q q -tail is exponentially integrable. Because w ∈ C w\in\mathcal{C} w ∈ C , its angular support is separated from the endpoints. In a diagonal term, only one angular variable can leave this support. The remaining factor is a power ( sin θ ) s − 1 / 2 (\sin\theta)^{s-1/2} ( sin θ ) s − 1/2 . It is integrable because s − 1 / 2 > − 1 s-1/2>-1 s − 1/2 > − 1 . Estimate (42) controls the factors e ± q / 2 e^{\pm q/2} e ± q /2 . Thus the truncated expression is absolutely integrable. Fubini’s theorem and Plancherel’s theorem in t t t give
q s ε [ w ] = ∫ R g s , ξ ε [ w ^ ( ξ , ⋅ ) ] d ξ . q_s^\varepsilon[w]=\int_{\mathbb{R}}g_{s,\xi}^\varepsilon[\widehat{w}(\xi,\cdot)]\,d\xi. q s ε [ w ] = ∫ R g s , ξ ε [ w ( ξ , ⋅ )] d ξ . By monotone convergence,
q s [ w ] = ∫ R g s , ξ [ w ^ ( ξ , ⋅ ) ] d ξ . q_s[w]=\int_{\mathbb{R}}g_{s,\xi}[\widehat{w}(\xi,\cdot)]\,d\xi. q s [ w ] = ∫ R g s , ξ [ w ( ξ , ⋅ )] d ξ . Use (46) to obtain (48) . Weighted Plancherel gives (49) .
Lemma 4.2 (Closed angular forms). For every ξ ∈ R \xi\in\mathbb{R} ξ ∈ R , the closure of b s , ξ b_{s,\xi} b s , ξ is a closed form in H s , τ H_{s,\tau} H s , τ . All fibers have the same form domain, denoted by Y s \mathcal{Y}_s Y s . There is c > 0 c>0 c > 0 , independent of ξ \xi ξ , such that
b s , ξ [ ψ ] ≥ c ∫ 0 π m s , τ ∣ ψ ∣ 2 d θ . (50) b_{s,\xi}[\psi]\ge c\int_0^\pi m_{s,\tau}|\psi|^2\,d\theta. \tag*{(50)} b s , ξ [ ψ ] ≥ c ∫ 0 π m s , τ ∣ ψ ∣ 2 d θ . (50) The embedding
Y s ↪ L 2 ( ( 0 , π ) , m s , τ d θ ) (51) \mathcal{Y}_s\hookrightarrow L^2((0,\pi),m_{s,\tau}\,d\theta) \tag*{(51)} Y s ↪ L 2 (( 0 , π ) , m s , τ d θ ) (51) is compact.
Proof. Let T ξ ψ T_\xi\psi T ξ ψ denote the expression between absolute-value signs in (45) . Suppose that ψ n → 0 \psi_n\to0 ψ n → 0 in H s , τ H_{s,\tau} H s , τ and T ξ ψ n → F T_\xi\psi_n\to F T ξ ψ n → F in the target kernel space. Pass to subsequences which converge almost everywhere. The target measure is σ \sigma σ -finite and is absolutely continuous off the null diagonal. The pointwise formula for T ξ T_\xi T ξ gives F = 0 F=0 F = 0 . Thus T ξ T_\xi T ξ is closable. The Cauchy-sequence criterion then shows that b s , ξ = ∥ T ξ ⋅ ∥ 2 b_{s,\xi}=\lVert T_\xi\cdot\rVert^2 b s , ξ = ∥ T ξ ⋅ ∥ 2 is closable.
Fix ξ 0 ∈ R \xi_0\in\mathbb{R} ξ 0 ∈ R and ψ ∈ C c ∞ ( 0 , π ) \psi\in C^\infty_c(0,\pi) ψ ∈ C c ∞ ( 0 , π ) . Let χ ∈ C c ∞ ( R ) \chi\in C^\infty_c(\mathbb{R}) χ ∈ C c ∞ ( R ) satisfy ∥ χ ∥ 2 = 1 \lVert\chi\rVert_2=1 ∥ χ ∥ 2 = 1 . Put
w R ( t , θ ) = R − 1 / 2 χ ( t / R ) e i ξ 0 t ψ ( θ ) . w_R(t,\theta)=R^{-1/2}\chi(t/R)e^{i\xi_0t}\psi(\theta). w R ( t , θ ) = R − 1/2 χ ( t / R ) e i ξ 0 t ψ ( θ ) . Then w R ∈ C w_R\in\mathcal{C} w R ∈ C and
w ^ R ( ξ , θ ) = R 1 / 2 χ ^ ( R ( ξ − ξ 0 ) ) ψ ( θ ) . \widehat{w}_R(\xi,\theta)=R^{1/2}\widehat{\chi}(R(\xi-\xi_0))\psi(\theta). w R ( ξ , θ ) = R 1/2 χ ( R ( ξ − ξ 0 )) ψ ( θ ) . Apply (44) with σ = τ \sigma=\tau σ = τ . Proposition 4.1 yields
∫ R R ∣ χ ^ ( R ( ξ − ξ 0 ) ) ∣ 2 b s , ξ [ ψ ] d ξ ≥ c s , τ ∫ 0 π m s , τ ∣ ψ ∣ 2 . \int_{\mathbb{R}}R|\widehat{\chi}(R(\xi-\xi_0))|^2b_{s,\xi}[\psi]\,d\xi\ge c_{s,\tau}\int_0^\pi m_{s,\tau}|\psi|^2. ∫ R R ∣ χ ( R ( ξ − ξ 0 )) ∣ 2 b s , ξ [ ψ ] d ξ ≥ c s , τ ∫ 0 π m s , τ ∣ ψ ∣ 2 .
By Lemma 3.4, subtraction of the zero fiber and integration in θ − φ \theta-\varphi θ − φ give
∣ b s , η [ ψ ] − b s , 0 [ ψ ] ∣ ≤ C ψ { ∣ η ∣ 2 , ∣ η ∣ ≤ 1 , ∣ η ∣ 2 s , ∣ η ∣ > 1. \left|b_{s,\eta}[\psi]-b_{s,0}[\psi]\right|\le C_{\psi}
\begin{cases}
|\eta|^2, & |\eta|\le1,\\
|\eta|^{2s}, & |\eta|>1.
\end{cases} ∣ b s , η [ ψ ] − b s , 0 [ ψ ] ∣ ≤ C ψ { ∣ η ∣ 2 , ∣ η ∣ 2 s , ∣ η ∣ ≤ 1 , ∣ η ∣ > 1.
The local integral is ∫ 0 1 min { η 2 q 2 , 1 } q − 1 − 2 s d q \int_0^1\min\{\eta^2q^2,1\}q^{-1-2s}\,dq ∫ 0 1 min { η 2 q 2 , 1 } q − 1 − 2 s d q , and the tail is exponential. The same cancellation proves continuity in η \eta η . Thus
b s , ξ 0 + ζ / R [ ψ ] ≤ C ψ , ξ 0 ( 1 + ∣ ζ ∣ 2 s ) , R ≥ 1. b_{s,\xi_0+\zeta/R}[\psi]\le C_{\psi,\xi_0}(1+|\zeta|^{2s}),\qquad R\ge1. b s , ξ 0 + ζ / R [ ψ ] ≤ C ψ , ξ 0 ( 1 + ∣ ζ ∣ 2 s ) , R ≥ 1.
Change variables to ζ = R ( ξ − ξ 0 ) \zeta=R(\xi-\xi_0) ζ = R ( ξ − ξ 0 ) and apply dominated convergence against ∣ χ ^ ( ζ ) ∣ 2 |\widehat{\chi}(\zeta)|^2 ∣ χ ( ζ ) ∣ 2 . This proves (50) on the core. With any σ > 0 \sigma>0 σ > 0 , it gives
b s , ξ [ ψ ] ≥ c s , σ ∫ 0 π m s , σ ∣ ψ ∣ 2 d θ . (52) b_{s,\xi}[\psi]\ge c_{s,\sigma}\int_0^\pi m_{s,\sigma}|\psi|^2\,d\theta. \tag*{(52)} b s , ξ [ ψ ] ≥ c s , σ ∫ 0 π m s , σ ∣ ψ ∣ 2 d θ . (52)
Choose 0 < τ 0 < min { τ , 2 s } 0<\tau_0<\min\{\tau,2s\} 0 < τ 0 < min { τ , 2 s } . Since m s , τ 0 ≥ 1 m_{s,\tau_0}\ge1 m s , τ 0 ≥ 1 on ( 0 , π ) (0,\pi) ( 0 , π ) , we have
∥ ψ ∥ L 2 ( 0 , π ) 2 ≤ ∫ 0 π m s , τ 0 ∣ ψ ∣ 2 d θ ≤ C b s , ξ [ ψ ] . (53) \|\psi\|_{L^2(0,\pi)}^2\le\int_0^\pi m_{s,\tau_0}|\psi|^2\,d\theta\le Cb_{s,\xi}[\psi]. \tag*{(53)} ∥ ψ ∥ L 2 ( 0 , π ) 2 ≤ ∫ 0 π m s , τ 0 ∣ ψ ∣ 2 d θ ≤ C b s , ξ [ ψ ] . (53)
For α ∈ ( − π , π ) \alpha\in(-\pi,\pi) α ∈ ( − π , π ) , set
H s , ξ ( α ) = 2 ∫ R ( 1 − cos ( ξ q ) ) K s ( q , α ) d q , H_{s,\xi}(\alpha)=2\int_{\mathbb{R}}(1-\cos(\xi q))K_s(q,\alpha)\,dq, H s , ξ ( α ) = 2 ∫ R ( 1 − cos ( ξ q )) K s ( q , α ) d q ,
where K s ( q , α ) = 2 − 1 − s ( cosh q − cos α ) − 1 − s K_s(q,\alpha)=2^{-1-s}(\cosh q-\cos\alpha)^{-1-s} K s ( q , α ) = 2 − 1 − s ( cosh q − cos α ) − 1 − s . For ∣ q ∣ < 1 |q|<1 ∣ q ∣ < 1 , Lemma 3.4 gives
∫ − π π K s ( q , α ) d α ≤ C ∣ q ∣ 1 − 2 s . \int_{-\pi}^{\pi}K_s(q,\alpha)\,d\alpha\le C|q|^{1-2s}. ∫ − π π K s ( q , α ) d α ≤ C ∣ q ∣ 1 − 2 s .
For ∣ q ∣ ≥ 1 |q|\ge1 ∣ q ∣ ≥ 1 , the kernel has an exponential tail. It follows that
∥ H s , ξ ∥ L 1 ( − π , π ) ≤ C { ξ 2 , ∣ ξ ∣ ≤ 1 , ∣ ξ ∣ 2 s , ∣ ξ ∣ ≥ 1. (54) \|H_{s,\xi}\|_{L^1(-\pi,\pi)}\le C
\begin{cases}
\xi^2, & |\xi|\le1,\\
|\xi|^{2s}, & |\xi|\ge1.
\end{cases} \tag*{(54)} ∥ H s , ξ ∥ L 1 ( − π , π ) ≤ C { ξ 2 , ∣ ξ ∣ 2 s , ∣ ξ ∣ ≤ 1 , ∣ ξ ∣ ≥ 1. (54)
Young’s inequality therefore gives
∣ b s , ξ [ ψ ] − b s , 0 [ ψ ] ∣ ≤ ∥ H s , ξ ∥ 1 ∥ ψ ∥ L 2 ( 0 , π ) 2 . (55) |b_{s,\xi}[\psi]-b_{s,0}[\psi]|\le\|H_{s,\xi}\|_1\|\psi\|_{L^2(0,\pi)}^2. \tag*{(55)} ∣ b s , ξ [ ψ ] − b s , 0 [ ψ ] ∣ ≤ ∥ H s , ξ ∥ 1 ∥ ψ ∥ L 2 ( 0 , π ) 2 . (55)
Use (53) first with ξ = 0 \xi=0 ξ = 0 and then with the given ξ \xi ξ . The graph norms of the closures of b s , ξ b_{s,\xi} b s , ξ and b s , 0 b_{s,0} b s , 0 are equivalent for each fixed ξ \xi ξ . Hence all closed fibers have one domain Y s \mathcal{Y}_s Y s .
For a core sequence ψ n → ψ \psi_n\to\psi ψ n → ψ in Y s \mathcal{Y}_s Y s , (52) makes ψ n \psi_n ψ n Cauchy in L 2 ( m s , σ ) L^2(m_{s,\sigma}) L 2 ( m s , σ ) . Its limit equals ψ \psi ψ almost everywhere. Thus (52) and (53) hold on the closed domain.
For two base exponents τ 1 , τ 2 > 0 \tau_1,\tau_2>0 τ 1 , τ 2 > 0 , choose 0 < σ < min { τ 1 , τ 2 , 2 s } 0<\sigma<\min\{\tau_1,\tau_2,2s\} 0 < σ < min { τ 1 , τ 2 , 2 s } . By (52) , both completions embed into L 2 ( m s , σ ) L^2(m_{s,\sigma}) L 2 ( m s , σ ) with the same representatives. The form controls both base norms, so the completions coincide. Hence Y s \mathcal{Y}_s Y s is independent of τ \tau τ .
For I ′ ⋐ I ⋐ ( 0 , π ) I'\Subset I\Subset(0,\pi) I ′ ⋐ I ⋐ ( 0 , π ) and ψ ∈ C 0 ∞ ( 0 , π ) \psi\in C^\infty_0(0,\pi) ψ ∈ C 0 ∞ ( 0 , π ) , choose η ∈ C c ∞ ( I ) \eta\in C^\infty_c(I) η ∈ C c ∞ ( I ) with η = 1 \eta=1 η = 1 on I ′ I' I ′ . The angular multipliers in (45) are smooth and bounded above and below on I I I . For ∣ q ∣ < 1 |q|<1 ∣ q ∣ < 1 and θ , φ ∈ I \theta,\varphi\in I θ , φ ∈ I ,
∣ ψ ( θ ) − ψ ( φ ) ∣ 2 ≤ C ∣ e − q / 4 ( sin θ ) 1 / 2 − s ψ ( θ ) − e q / 4 ( sin φ ) 1 / 2 − s ψ ( φ ) ∣ 2 + C ( q 2 + ∣ θ − φ ∣ 2 ) ( ∣ ψ ( θ ) ∣ 2 + ∣ ψ ( φ ) ∣ 2 ) . |\psi(\theta)-\psi(\varphi)|^2\le C\left|e^{-q/4}(\sin\theta)^{1/2-s}\psi(\theta)-e^{q/4}(\sin\varphi)^{1/2-s}\psi(\varphi)\right|^2+C(q^2+|\theta-\varphi|^2)(|\psi(\theta)|^2+|\psi(\varphi)|^2). ∣ ψ ( θ ) − ψ ( φ ) ∣ 2 ≤ C e − q /4 ( sin θ ) 1/2 − s ψ ( θ ) − e q /4 ( sin φ ) 1/2 − s ψ ( φ ) 2 + C ( q 2 + ∣ θ − φ ∣ 2 ) ( ∣ ψ ( θ ) ∣ 2 + ∣ ψ ( φ ) ∣ 2 ) .
Put b = s − 1 / 2 b=s-1/2 b = s − 1/2 . In the positive square, write
η ( θ ) A ( θ ) − η ( φ ) B ( φ ) = η ( θ ) ( A ( θ ) − B ( φ ) ) + ( η ( θ ) − η ( φ ) ) B ( φ ) . \eta(\theta)A(\theta)-\eta(\varphi)B(\varphi)=\eta(\theta)(A(\theta)-B(\varphi))+(\eta(\theta)-\eta(\varphi))B(\varphi). η ( θ ) A ( θ ) − η ( φ ) B ( φ ) = η ( θ ) ( A ( θ ) − B ( φ )) + ( η ( θ ) − η ( φ )) B ( φ ) .
The first term is controlled by C g s , 0 [ ψ ] C g_{s,0}[\psi] C g s , 0 [ ψ ] . For the second term, split ∣ θ − φ ∣ < c I |\theta-\varphi|<c_I ∣ θ − φ ∣ < c I and its complement. On the first set, a nonzero cutoff difference forces both angular variables into a fixed compact subinterval of ( 0 , π ) (0,\pi) ( 0 , π ) . The Lipschitz estimate for η \eta η and (41) give an error bounded by C ∥ ψ ∥ 2 2 C\|\psi\|_{2}^{2} C ∥ ψ ∥ 2 2 . On the complement, the angular kernel is bounded after integration in q q q . The sine factors in (45) give the bound
C ∫ 0 π { 1 + ( sin φ ) − ( s − 1 / 2 ) + } ∣ ψ ( φ ) ∣ 2 d φ . C\int_{0}^{\pi}\left\{1+(\sin\varphi)^{-(s-1/2)_{+}}\right\}|\psi(\varphi)|^{2}\,d\varphi. C ∫ 0 π { 1 + ( sin φ ) − ( s − 1/2 ) + } ∣ ψ ( φ ) ∣ 2 d φ .
Set σ ∗ = min { 2 s , s + 1 / 2 } \sigma_{*}=\min\{2s,s+1/2\} σ ∗ = min { 2 s , s + 1/2 } . Then m s , σ ∗ = ( sin φ ) − ( s − 1 / 2 ) + m_{s,\sigma_{*}}=(\sin\varphi)^{-(s-1/2)_{+}} m s , σ ∗ = ( sin φ ) − ( s − 1/2 ) + . Estimate (52) , with σ = σ ∗ \sigma=\sigma_{*} σ = σ ∗ , yields
g s , 0 [ η ψ ] ≤ C I ′ , I b s , 0 [ ψ ] . \mathfrak{g}_{s,0}[\eta\psi]\leq C_{I',I}\mathfrak{b}_{s,0}[\psi]. g s , 0 [ η ψ ] ≤ C I ′ , I b s , 0 [ ψ ] .
The error in the preceding pointwise comparison is integrable because
∫ q 2 + α 2 < 1 ( q 2 + α 2 ) − s d q d α < ∞ . \int_{q^{2}+\alpha^{2}<1}(q^{2}+\alpha^{2})^{-s}\,dq\,d\alpha<\infty. ∫ q 2 + α 2 < 1 ( q 2 + α 2 ) − s d q d α < ∞.
Finally,
∫ − 1 1 ( q 2 + α 2 ) − 1 − s d q ≥ c ∣ α ∣ − 1 − 2 s \int_{-1}^{1}(q^{2}+\alpha^{2})^{-1-s}\,dq\geq c|\alpha|^{-1-2s} ∫ − 1 1 ( q 2 + α 2 ) − 1 − s d q ≥ c ∣ α ∣ − 1 − 2 s
for 0 < ∣ α ∣ < 1 / 2 0<|\alpha|<1/2 0 < ∣ α ∣ < 1/2 . Therefore
∥ ψ ∥ H s ( I ′ ) 2 ≤ C I ′ , I ( b s , 0 [ ψ ] + ∥ ψ ∥ L 2 ( 0 , π ) 2 ) (56) \|\psi\|_{H^{s}(I')}^{2}\leq C_{I',I}\left(\mathfrak{b}_{s,0}[\psi]+\|\psi\|_{L^{2}(0,\pi)}^{2}\right) \tag*{(56)} ∥ ψ ∥ H s ( I ′ ) 2 ≤ C I ′ , I ( b s , 0 [ ψ ] + ∥ ψ ∥ L 2 ( 0 , π ) 2 ) (56)
The same estimates applied to a core sequence show that η ψ n → η ψ \eta\psi_{n}\to\eta\psi η ψ n → η ψ in Y s \mathcal{Y}_{s} Y s . Lower semicontinuity gives (56) for every ψ ∈ Y s \psi\in\mathcal{Y}_{s} ψ ∈ Y s .
For compactness, estimate (52) with σ = τ 0 \sigma=\tau_{0} σ = τ 0 gives
∫ 0 π ( sin θ ) τ 0 − 2 s ∣ ψ ∣ 2 d θ ≤ C b s , 0 [ ψ ] . \int_{0}^{\pi}(\sin\theta)^{\tau_{0}-2s}|\psi|^{2}\,d\theta\leq C\mathfrak{b}_{s,0}[\psi]. ∫ 0 π ( sin θ ) τ 0 − 2 s ∣ ψ ∣ 2 d θ ≤ C b s , 0 [ ψ ] .
Let { ψ n } \{\psi_{n}\} { ψ n } be bounded in Y s \mathcal{Y}_{s} Y s . On [ δ , π − δ ] [\delta,\pi-\delta] [ δ , π − δ ] , (56) gives a uniform H s H^{s} H s bound. The Rellich theorem yields a subsequence convergent in L 2 ( [ δ , π − δ ] ) L^{2}([\delta,\pi-\delta]) L 2 ([ δ , π − δ ]) . At the endpoints,
∫ ( 0 , δ ) ∪ ( π − δ , π ) m s , τ ∣ ψ n ∣ 2 d θ ≤ C δ τ − τ 0 ∫ 0 π ( sin θ ) τ 0 − 2 s ∣ ψ n ∣ 2 d θ ≤ C δ τ − τ 0 . (57) \int_{(0,\delta)\cup(\pi-\delta,\pi)}m_{s,\tau}|\psi_{n}|^{2}\,d\theta
\leq C\delta^{\tau-\tau_{0}}\int_{0}^{\pi}(\sin\theta)^{\tau_{0}-2s}|\psi_{n}|^{2}\,d\theta
\leq C\delta^{\tau-\tau_{0}}. \tag*{(57)} ∫ ( 0 , δ ) ∪ ( π − δ , π ) m s , τ ∣ ψ n ∣ 2 d θ ≤ C δ τ − τ 0 ∫ 0 π ( sin θ ) τ 0 − 2 s ∣ ψ n ∣ 2 d θ ≤ C δ τ − τ 0 . (57)
Together with interior convergence, this uniform endpoint bound proves (51) by a diagonal argument. □
The angular space is the closure of the core for the nonnegative ground-state form. We always use this closure, hence its Friedrichs operator realization. In particular, no unweighted boundary trace condition is imposed separately. The two terms in the Hardy difference need not be finite on Y s \mathcal{Y}_{s} Y s .
The associated self-adjoint operator is defined by
D ( B s ( ξ ) ) = { ψ ∈ Y s : there exists f ∈ H s , τ such that D(B_{s}(\xi))=\{\psi\in\mathcal{Y}_{s}:\text{ there exists }f\in H_{s,\tau}\text{ such that} D ( B s ( ξ )) = { ψ ∈ Y s : there exists f ∈ H s , τ such that
b s , ξ ( ψ , η ) = ⟨ f , η ⟩ H s , τ for every η ∈ Y s } , \mathfrak{b}_{s,\xi}(\psi,\eta)=\langle f,\eta\rangle_{H_{s,\tau}}\text{ for every }\eta\in\mathcal{Y}_{s}\}, b s , ξ ( ψ , η ) = ⟨ f , η ⟩ H s , τ for every η ∈ Y s } ,
B s ( ξ ) ψ = f . (58) B_{s}(\xi)\psi=f. \tag*{(58)} B s ( ξ ) ψ = f . (58)
Density of the form domain makes f f f unique. This is the operator used in all eigenvalue, resolvent, and spectral projection statements below.
Lemma 4.3 (Continuity of the angular resolvents). The map ξ ↦ ( B s ( ξ ) + 1 ) − 1 \xi\mapsto(B_{s}(\xi)+1)^{-1} ξ ↦ ( B s ( ξ ) + 1 ) − 1 is continuous from R \mathbb{R} R to L ( H s , τ , Y s ) \mathcal{L}(H_{s,\tau},\mathcal{Y}_{s}) L ( H s , τ , Y s ) . In particular it is continuous in operator norm on H s , τ H_{s,\tau} H s , τ , and every ordered eigenvalue λ j ( ξ ) \lambda_{j}(\xi) λ j ( ξ ) is continuous and even on R \mathbb{R} R .
Proof. Fix a compact frequency interval K K K . The kernel estimate leading to (54) also gives
∣ b s , ξ ( u , v ) − b s , η ( u , v ) ∣ ≤ ω K ( ∣ ξ − η ∣ ) ∥ u ∥ L 2 ( 0 , π ) ∥ v ∥ L 2 ( 0 , π ) , ω K ( r ) ⟶ 0. |\mathfrak{b}_{s,\xi}(u,v)-\mathfrak{b}_{s,\eta}(u,v)|
\leq\omega_{K}(|\xi-\eta|)\|u\|_{L^{2}(0,\pi)}\|v\|_{L^{2}(0,\pi)},\qquad\omega_{K}(r)\longrightarrow0. ∣ b s , ξ ( u , v ) − b s , η ( u , v ) ∣ ≤ ω K ( ∣ ξ − η ∣ ) ∥ u ∥ L 2 ( 0 , π ) ∥ v ∥ L 2 ( 0 , π ) , ω K ( r ) ⟶ 0.
Indeed, for ξ , η ∈ K \xi,\eta\in K ξ , η ∈ K and ∣ q ∣ < 1 |q|<1 ∣ q ∣ < 1 , the mean-value formula gives
∣ cos ( ξ q ) − cos ( η q ) ∣ ≤ C K ∣ ξ − η ∣ q 2 . |\cos(\xi q)-\cos(\eta q)| \le C_K|\xi-\eta|q^2. ∣ cos ( ξ q ) − cos ( η q ) ∣ ≤ C K ∣ ξ − η ∣ q 2 .
For ∣ q ∣ ≥ 1 |q|\ge1 ∣ q ∣ ≥ 1 , the difference is bounded by 2. Hence
∫ − π π ∫ R ∣ cos ( ξ q ) − cos ( η q ) ∣ K s ( q , α ) d q d α ≤ C K ∣ ξ − η ∣ ∫ 0 1 q 1 − 2 s d q + C ∫ 1 ∞ min { ∣ ξ − η ∣ q , 2 } e − ( 1 + s ) q d q ⟶ 0. \int_{-\pi}^{\pi}\int_{\mathbb{R}}|\cos(\xi q)-\cos(\eta q)|K_s(q,\alpha)\,dq\,d\alpha
\le C_K|\xi-\eta|\int_0^1 q^{1-2s}\,dq+C\int_1^\infty\min\{|\xi-\eta|q,2\}e^{-(1+s)q}\,dq\longrightarrow0. ∫ − π π ∫ R ∣ cos ( ξ q ) − cos ( η q ) ∣ K s ( q , α ) d q d α ≤ C K ∣ ξ − η ∣ ∫ 0 1 q 1 − 2 s d q + C ∫ 1 ∞ min { ∣ ξ − η ∣ q , 2 } e − ( 1 + s ) q d q ⟶ 0.
Young’s inequality applied to the polarized cosine difference proves the asserted modulus of continuity.
Let A ξ : Y s → Y s ∗ A_\xi:\mathcal{Y}_s\to\mathcal{Y}_s^* A ξ : Y s → Y s ∗ be the map of b s , ξ + ⟨ ⋅ , ⋅ ⟩ H s , τ b_{s,\xi}+\langle\cdot,\cdot\rangle_{H_{s,\tau}} b s , ξ + ⟨ ⋅ , ⋅ ⟩ H s , τ . The graph norms are uniformly equivalent on K K K by (55) and (53) , applied also at the frequency ξ \xi ξ . Thus A ξ A_\xi A ξ is a coercive isomorphism, with inverse uniformly bounded on K K K . The preceding estimate and (53) imply
∥ A ξ − A η ∥ Y s → Y s ∗ ⟶ 0 as ξ − η → 0 , ξ , η ∈ K . \|A_\xi-A_\eta\|_{\mathcal{Y}_s\to\mathcal{Y}_s^*}\longrightarrow0\quad\text{as }\xi-\eta\to0,\qquad\xi,\eta\in K. ∥ A ξ − A η ∥ Y s → Y s ∗ ⟶ 0 as ξ − η → 0 , ξ , η ∈ K .
The inverse identity
A ξ − 1 − A η − 1 = − A ξ − 1 ( A ξ − A η ) A η − 1 A_\xi^{-1}-A_\eta^{-1}=-A_\xi^{-1}(A_\xi-A_\eta)A_\eta^{-1} A ξ − 1 − A η − 1 = − A ξ − 1 ( A ξ − A η ) A η − 1
gives continuity from H s , τ H_{s,\tau} H s , τ to Y s \mathcal{Y}_s Y s after restriction. By (58) , the restricted inverse is ( B s ( ξ ) + 1 ) − 1 (B_s(\xi)+1)^{-1} ( B s ( ξ ) + 1 ) − 1 . Compactness of the resolvents and the min–max principle give continuity of their ordered eigenvalues, hence of λ j ( ξ ) \lambda_j(\xi) λ j ( ξ ) . The cosine formula makes B s ( − ξ ) = B s ( ξ ) B_s(-\xi)=B_s(\xi) B s ( − ξ ) = B s ( ξ ) , which proves evenness. □
Proposition 4.4 (Closed direct integral). The closure of q s q_s q s has domain
{ w ∈ L 2 ( R ; H s , τ ) : w ^ ( ξ , ⋅ ) ∈ Y s for a.e. ξ , ∫ R b s , ξ [ w ^ ( ξ , ⋅ ) ] d ξ < ∞ } . \left\{w\in L^2(\mathbb{R};H_{s,\tau}):\widehat{w}(\xi,\cdot)\in\mathcal{Y}_s\ \text{for a.e. }\xi,\ \int_{\mathbb{R}}b_{s,\xi}[\widehat{w}(\xi,\cdot)]\,d\xi<\infty\right\}. { w ∈ L 2 ( R ; H s , τ ) : w ( ξ , ⋅ ) ∈ Y s for a.e. ξ , ∫ R b s , ξ [ w ( ξ , ⋅ )] d ξ < ∞ } . For every function in this domain,
q s [ w ] = ∫ R b s , ξ [ w ^ ( ξ , ⋅ ) ] d ξ . (59) q_s[w]=\int_{\mathbb{R}}b_{s,\xi}[\widehat{w}(\xi,\cdot)]\,d\xi. \tag*{(59)} q s [ w ] = ∫ R b s , ξ [ w ( ξ , ⋅ )] d ξ . (59) Proof. Let the right-hand side define q ~ s \widetilde{q}_s q s . On ∣ ξ ∣ ≤ N |\xi|\le N ∣ ξ ∣ ≤ N , (54) , (55) , and (53) give
C N − 1 ( b s , 0 [ v ] + ∥ v ∥ H s , τ 2 ) ≤ b s , ξ [ v ] + ∥ v ∥ H s , τ 2 ≤ C N ( b s , 0 [ v ] + ∥ v ∥ H s , τ 2 ) . (60) C_N^{-1}\left(b_{s,0}[v]+\|v\|_{H_{s,\tau}}^2\right)\le b_{s,\xi}[v]+\|v\|_{H_{s,\tau}}^2\le C_N\left(b_{s,0}[v]+\|v\|_{H_{s,\tau}}^2\right). \tag*{(60)} C N − 1 ( b s , 0 [ v ] + ∥ v ∥ H s , τ 2 ) ≤ b s , ξ [ v ] + ∥ v ∥ H s , τ 2 ≤ C N ( b s , 0 [ v ] + ∥ v ∥ H s , τ 2 ) . (60) The common form domain is separable. By (55) and polarization, ξ ↦ b s , ξ ( u , v ) \xi\mapsto b_{s,\xi}(u,v) ξ ↦ b s , ξ ( u , v ) is continuous for u , v ∈ Y s u,v\in\mathcal{Y}_s u , v ∈ Y s . A countable dense subset of the angular core therefore defines a measurable field of form spaces. On each bounded frequency interval, (60) identifies this field with the constant space Y s \mathcal{Y}_s Y s . Finite-energy sections are strongly measurable, and common-core simple sections are dense.
Let { f n } \{f_n\} { f n } be Cauchy for the integrated graph norm. Choose a subsequence, still denoted by { f n } \{f_n\} { f n } , such that
∥ f n + 1 − f n ∥ g r a p h ≤ 2 − n . \|f_{n+1}-f_n\|_{\mathrm{graph}}\le2^{-n}. ∥ f n + 1 − f n ∥ graph ≤ 2 − n . On [ − N , N ] [-N,N] [ − N , N ] , Cauchy–Schwarz gives
∫ − N N ( b s , ξ [ f n + 1 ( ξ ) − f n ( ξ ) ] + ∥ f n + 1 ( ξ ) − f n ( ξ ) ∥ H s , τ 2 ) 1 / 2 d ξ ≤ ( 2 N ) 1 / 2 2 − n . \int_{-N}^{N}\left(b_{s,\xi}[f_{n+1}(\xi)-f_n(\xi)]+\|f_{n+1}(\xi)-f_n(\xi)\|_{H_{s,\tau}}^2\right)^{1/2}\,d\xi\le(2N)^{1/2}2^{-n}. ∫ − N N ( b s , ξ [ f n + 1 ( ξ ) − f n ( ξ )] + ∥ f n + 1 ( ξ ) − f n ( ξ ) ∥ H s , τ 2 ) 1/2 d ξ ≤ ( 2 N ) 1/2 2 − n . Fubini’s theorem shows that the sum of the pointwise fiber graph increments is finite for almost every ξ ∈ [ − N , N ] \xi\in[-N,N] ξ ∈ [ − N , N ] . Diagonalize in N N N . Fiber completeness gives a pointwise limit f ( ξ ) ∈ Y s f(\xi)\in\mathcal{Y}_s f ( ξ ) ∈ Y s for almost every ξ \xi ξ . Fatou’s lemma yields
∥ f n − f ∥ g r a p h 2 ≤ lim inf m → ∞ ∥ f n − f m ∥ g r a p h 2 . \|f_n-f\|_{\mathrm{graph}}^2\le\liminf_{m\to\infty}\|f_n-f_m\|_{\mathrm{graph}}^2. ∥ f n − f ∥ graph 2 ≤ m → ∞ lim inf ∥ f n − f m ∥ graph 2 . The right-hand side tends to zero. Thus the integrated form is closed.
To identify a core, truncate a finite-energy section to ∣ ξ ∣ ≤ N |\xi| \le N ∣ ξ ∣ ≤ N . By (60) , Bochner simple functions with values in Y s \mathcal{Y}_s Y s are dense. Since C c ∞ ( 0 , π ) C^\infty_c(0,\pi) C c ∞ ( 0 , π ) is the common form core, their values may be replaced by angular core functions. By (60) , their scalar coefficients may next be approximated in L 2 ( − N , N ) L^2(-N,N) L 2 ( − N , N ) by functions in C c ∞ ( R ξ ) C^\infty_c(\mathbb{R}_\xi) C c ∞ ( R ξ ) .
It remains to obtain compact support in t t t . Consider one term g ( ξ ) ψ ( θ ) g(\xi)\psi(\theta) g ( ξ ) ψ ( θ ) , where g ∈ C c ∞ ( R ) g \in C^\infty_c(\mathbb{R}) g ∈ C c ∞ ( R ) and ψ ∈ C c ∞ ( 0 , π ) \psi\in C^\infty_c(0,\pi) ψ ∈ C c ∞ ( 0 , π ) . Its inverse Fourier transform g ˇ \check{g} g ˇ is a Schwartz function. Choose χ ∈ C c ∞ ( R ) \chi\in C^\infty_c(\mathbb{R}) χ ∈ C c ∞ ( R ) with χ = 1 \chi=1 χ = 1 near zero. Then χ ( t / R ) g ˇ ( t ) → g ˇ ( t ) \chi(t/R)\check{g}(t) \to\check{g}(t) χ ( t / R ) g ˇ ( t ) → g ˇ ( t ) in H s ( R ) H^s(\mathbb{R}) H s ( R ) . Indeed, use dominated convergence for the L 2 L^2 L 2 term. For the difference quotient, split ∣ t − t ′ ∣ < 1 |t-t'|<1 ∣ t − t ′ ∣ < 1 and ∣ t − t ′ ∣ ≥ 1 |t-t'|\ge1 ∣ t − t ′ ∣ ≥ 1 . The Lipschitz bound for χ ( t / R ) \chi(t/R) χ ( t / R ) controls the first part. The Schwartz decay of g ˇ \check{g} g ˇ controls both parts. For the fixed angular core function, (54) and (55) give
b s , ξ [ ψ ] + ∥ ψ ∥ H s , τ 2 ≤ C ψ ( 1 + ∣ ξ ∣ 2 s ) . b_{s,\xi}[\psi]+\|\psi\|_{H_{s,\tau}}^2\le C_\psi(1+|\xi|^{2s}). b s , ξ [ ψ ] + ∥ ψ ∥ H s , τ 2 ≤ C ψ ( 1 + ∣ ξ ∣ 2 s ) .
Plancherel’s theorem therefore gives convergence in the integrated graph norm. Each product χ ( t / R ) g ˇ ( t ) ψ ( θ ) \chi(t/R)\check{g}(t)\psi(\theta) χ ( t / R ) g ˇ ( t ) ψ ( θ ) belongs to C \mathcal{C} C . Thus C \mathcal{C} C is a form core for q ~ s \widetilde{q}_s q s . Proposition 4.1 shows that the two forms agree on this core. Their closures are equal. □
We now establish the modulus inequality and the positivity of the first eigenfunction.
Lemma 4.5 (Modulus stability of the closed angular form). For every \psi\in\mathcal{Y}_s\), *one has* |\psi|\in\mathcal{Y}_s$ and
b s , 0 [ ∣ ψ ∣ ] ≤ b s , 0 [ ψ ] . b_{s,0}[|\psi|]\le b_{s,0}[\psi]. b s , 0 [ ∣ ψ ∣ ] ≤ b s , 0 [ ψ ] . Moreover, there are η j ∈ C c ∞ ( 0 , π ) \eta_j\in C^\infty_c(0,\pi) η j ∈ C c ∞ ( 0 , π ) such that η j ≥ 0 \eta_j\ge0 η j ≥ 0 and η j → ∣ ψ ∣ \eta_j\to|\psi| η j → ∣ ψ ∣ in Y s \mathcal{Y}_s Y s .
Proof. Let I ⋐ I 1 ⋐ ( 0 , π ) I\Subset I_1\Subset(0,\pi) I ⋐ I 1 ⋐ ( 0 , π ) . If f ∈ C c ∞ ( I ) f\in C^\infty_c(I) f ∈ C c ∞ ( I ) and f f f is extended by zero to R \mathbb{R} R , then
b s , 0 [ f ] ≤ C I { ∥ f ∥ L 2 ( R ) 2 + ∫ R ∫ R ∣ f ( θ ) − f ( φ ) ∣ 2 ∣ θ − φ ∣ 1 + 2 s d θ d φ } . b_{s,0}[f]\le C_I\left\{\|f\|_{L^2(\mathbb{R})}^2+\int_{\mathbb{R}}\int_{\mathbb{R}}\frac{|f(\theta)-f(\varphi)|^2}{|\theta-\varphi|^{1+2s}}\,d\theta\,d\varphi\right\}. b s , 0 [ f ] ≤ C I { ∥ f ∥ L 2 ( R ) 2 + ∫ R ∫ R ∣ θ − φ ∣ 1 + 2 s ∣ f ( θ ) − f ( φ ) ∣ 2 d θ d φ } .
To prove this, use (45) and put b = s − 1 2 b=s-\frac12 b = s − 2 1 . On ∣ q ∣ + ∣ θ − φ ∣ < δ I |q|+|\theta-\varphi|<\delta_I ∣ q ∣ + ∣ θ − φ ∣ < δ I , a nonzero term forces both angular variables into I 1 I_1 I 1 . The functions ( sin θ ) − b (\sin\theta)^{-b} ( sin θ ) − b and e − q / 4 e^{-q/4} e − q /4 are smooth there. The mean-value theorem gives
∣ e − q / 4 ( sin θ ) − b f ( θ ) − e q / 4 ( sin φ ) − b f ( φ ) ∣ 2 ≤ C I ∣ f ( θ ) − f ( φ ) ∣ 2 + C I ( q 2 + ∣ θ − φ ∣ 2 ) ( ∣ f ( θ ) ∣ 2 + ∣ f ( φ ) ∣ 2 ) . \left|e^{-q/4}(\sin\theta)^{-b}f(\theta)-e^{q/4}(\sin\varphi)^{-b}f(\varphi)\right|^2
\le C_I|f(\theta)-f(\varphi)|^2+C_I(q^2+|\theta-\varphi|^2)(|f(\theta)|^2+|f(\varphi)|^2). e − q /4 ( sin θ ) − b f ( θ ) − e q /4 ( sin φ ) − b f ( φ ) 2 ≤ C I ∣ f ( θ ) − f ( φ ) ∣ 2 + C I ( q 2 + ∣ θ − φ ∣ 2 ) ( ∣ f ( θ ) ∣ 2 + ∣ f ( φ ) ∣ 2 ) .
The sine factor outside the square is bounded on I 1 I_1 I 1 . By Lemma 3.4, integration in q q q of the first term gives C I ∣ θ − φ ∣ − 1 − 2 s C_I|\theta-\varphi|^{-1-2s} C I ∣ θ − φ ∣ − 1 − 2 s . The second term is integrable because
∫ q 2 + α 2 < 1 ( q 2 + α 2 ) − s d q d α < ∞ . \int_{q^2+\alpha^2<1}(q^2+\alpha^2)^{-s}\,dq\,d\alpha<\infty. ∫ q 2 + α 2 < 1 ( q 2 + α 2 ) − s d q d α < ∞.
On the complementary set, expand the square. At least one function value has its angular variable in I I I . If ∣ q ∣ ≥ δ I / 2 |q|\ge\delta_I/2 ∣ q ∣ ≥ δ I /2 , the exponential tail in (42) absorbs e ∣ q ∣ / 2 e^{|q|/2} e ∣ q ∣/2 . If ∣ θ − φ ∣ ≥ δ I / 2 |\theta-\varphi|\ge\delta_I/2 ∣ θ − φ ∣ ≥ δ I /2 , the kernel is bounded after integration in q q q . The remaining endpoint factor is a fixed power of sine. Its worst power is ( sin ϑ ) s − 1 / 2 (\sin\vartheta)^{s-1/2} ( sin ϑ ) s − 1/2 , which is integrable because s − 1 2 > − 1 s-\frac12>-1 s − 2 1 > − 1 . This bounds the complementary part by C I ∥ f ∥ L 2 2 C_I\|f\|_{L^2}^2 C I ∥ f ∥ L 2 2 . The local upper estimate follows.
Now let ψ ∈ C c ∞ ( 0 , π ) \psi\in C^\infty_c(0,\pi) ψ ∈ C c ∞ ( 0 , π ) . The function ∣ ψ ∣ |\psi| ∣ ψ ∣ , extended by zero, belongs to H 1 ( R ) H^1(\mathbb{R}) H 1 ( R ) . Indeed, the modulus map is Lipschitz and ∣ ( ∣ ψ ∣ ′ ) ∣ ≤ ∣ ψ ′ ∣ |(|\psi|')|\le|\psi'| ∣ ( ∣ ψ ∣ ′ ) ∣ ≤ ∣ ψ ′ ∣ almost everywhere. Let ρ ϵ \rho_\epsilon ρ ϵ be a nonnegative Friedrichs mollifier and put η ϵ = ρ ϵ ∗ ∣ ψ ∣ \eta_\epsilon=\rho_\epsilon*|\psi| η ϵ = ρ ϵ ∗ ∣ ψ ∣ . For small ϵ \epsilon ϵ , η ϵ \eta_\epsilon η ϵ is nonnegative and belongs to C c ∞ ( 0 , π ) C^\infty_c(0,\pi) C c ∞ ( 0 , π ) . It converges to ∣ ψ ∣ |\psi| ∣ ψ ∣ in H s ( R ) H^s(\mathbb{R}) H s ( R ) . Apply the local upper estimate to η ϵ − η δ \eta_\epsilon-\eta_\delta η ϵ − η δ . The family { η ϵ } \{\eta_\epsilon\} { η ϵ } is Cauchy in Y s \mathcal{Y}_s Y s . Its limit in H s , τ H_{s,\tau} H s , τ is ∣ ψ ∣ |\psi| ∣ ψ ∣ . Hence ∣ ψ ∣ ∈ Y s |\psi|\in\mathcal{Y}_s ∣ ψ ∣ ∈ Y s .
For A , B > 0 A,B>0 A , B > 0 and z , w ∈ C z,w\in\mathbb{C} z , w ∈ C , ∣ A ∣ z ∣ − B ∣ w ∣ ∣ ≤ ∣ A z − B w ∣ |A|z|-B|w||\le|Az-Bw| ∣ A ∣ z ∣ − B ∣ w ∣∣ ≤ ∣ A z − B w ∣ . Apply this to (45) . A subsequence of η ϵ \eta_\epsilon η ϵ and its image under the closed difference operator converges almost everywhere, so the closed form agrees with that integral at ∣ ψ ∣ |\psi| ∣ ψ ∣ . Hence
b s , 0 [ ∣ ψ ∣ ] ≤ b s , 0 [ ψ ] b_{s,0}[|\psi|]\le b_{s,0}[\psi] b s , 0 [ ∣ ψ ∣ ] ≤ b s , 0 [ ψ ]
for every core function ψ \psi ψ . The preceding mollifiers also give its nonnegative core approximation.
Finally, let ψ ∈ Y s \psi\in\mathcal{Y}_s ψ ∈ Y s . Choose core functions ψ n → ψ \psi_n \to\psi ψ n → ψ in Y s \mathcal{Y}_s Y s . Then ∣ ψ n ∣ → ∣ ψ ∣ |\psi_n| \to|\psi| ∣ ψ n ∣ → ∣ ψ ∣ in H s , τ H_{s,\tau} H s , τ . The core inequality shows that { ∣ ψ n ∣ } \{|\psi_n|\} { ∣ ψ n ∣ } is bounded in Y s \mathcal{Y}_s Y s . A weakly convergent subsequence has weak limit ∣ ψ ∣ |\psi| ∣ ψ ∣ , because its strong H s , τ H_{s,\tau} H s , τ limit is ∣ ψ ∣ |\psi| ∣ ψ ∣ . Weak lower semicontinuity gives
b s , 0 [ ∣ ψ ∣ ] ≤ lim inf n → ∞ b s , 0 [ ∣ ψ n ∣ ] ≤ lim n → ∞ b s , 0 [ ψ n ] = b s , 0 [ ψ ] . \mathfrak{b}_{s,0}[|\psi|] \le\liminf_{n\to\infty} \mathfrak{b}_{s,0}[|\psi_n|] \le\lim_{n\to\infty} \mathfrak{b}_{s,0}[\psi_n] = \mathfrak{b}_{s,0}[\psi]. b s , 0 [ ∣ ψ ∣ ] ≤ n → ∞ lim inf b s , 0 [ ∣ ψ n ∣ ] ≤ n → ∞ lim b s , 0 [ ψ n ] = b s , 0 [ ψ ] .
Mazur’s lemma gives convex combinations of the nonnegative functions ∣ ψ n ∣ |\psi_n| ∣ ψ n ∣ which converge strongly to ∣ ψ ∣ |\psi| ∣ ψ ∣ in Y s \mathcal{Y}_s Y s . Approximate the finitely many terms in each combination by the nonnegative smooth approximants constructed above. The resulting functions are nonnegative, belong to C c ∞ ( 0 , π ) C^\infty_c(0,\pi) C c ∞ ( 0 , π ) , and converge to ∣ ψ ∣ |\psi| ∣ ψ ∣ in Y s \mathcal{Y}_s Y s . □
The next lemma gives simplicity and positivity at zero frequency.
Lemma 4.6 (Positive simple ground state). The eigenvalue λ 1 ( 0 ) \lambda_1(0) λ 1 ( 0 ) is simple. Its eigenfunction can be chosen strictly positive almost everywhere in ( 0 , π ) (0,\pi) ( 0 , π ) .
Proof. Lemma 4.5 permits replacement of every form-domain function by its modulus. Fix α > 0 \alpha> 0 α > 0 . The shifted resolvent of B s ( 0 ) B_s(0) B s ( 0 ) is positivity preserving. Indeed, for F ≥ 0 F \ge0 F ≥ 0 , the vector ( B s ( 0 ) + α ) − 1 F (B_s(0)+\alpha)^{-1}F ( B s ( 0 ) + α ) − 1 F uniquely minimizes
b s , 0 [ v ] + α ∥ v ∥ H s , τ 2 − 2 Re ⟨ F , v ⟩ H s , τ . \mathfrak{b}_{s,0}[v] + \alpha\lVert v\rVert_{H_{s,\tau}}^2 - 2\operatorname{Re}\langle F,v\rangle_{H_{s,\tau}}. b s , 0 [ v ] + α ∥ v ∥ H s , τ 2 − 2 Re ⟨ F , v ⟩ H s , τ .
Replacing v v v by ∣ v ∣ |v| ∣ v ∣ cannot increase this functional. Uniqueness gives a nonnegative minimizer.
For irreducibility, the jump kernel is
J s ( θ , φ ) = 2 ∫ R K s ( q , θ , φ ) d q > 0 J_s(\theta,\varphi)=2\int_{\mathbb{R}}K_s(q,\theta,\varphi)\,\mathrm{d}q>0 J s ( θ , φ ) = 2 ∫ R K s ( q , θ , φ ) d q > 0
for θ ≠ φ \theta\ne\varphi θ = φ . Suppose that a measurable set A A A and its complement both have positive measure and that the ideal L 2 ( A ) L^2(A) L 2 ( A ) is invariant. We first construct h ∈ Y s h\in\mathcal{Y}_s h ∈ Y s with h > 0 h>0 h > 0 almost everywhere. Choose nonnegative core functions h n h_n h n which are positive on [ 1 / n , π − 1 / n ] [1/n,\pi-1/n] [ 1/ n , π − 1/ n ] . Choose c n > 0 c_n>0 c n > 0 so rapidly decreasing that ∑ n c n ∥ h n ∥ Y s < ∞ \sum_n c_n\lVert h_n\rVert_{\mathcal{Y}_s}<\infty ∑ n c n ∥ h n ∥ Y s < ∞ . Then h = ∑ n c n h n h=\sum_n c_nh_n h = ∑ n c n h n has the required properties. The invariant-ideal criterion for closed forms gives f = 1 A h ∈ Y s f=1_Ah\in\mathcal{Y}_s f = 1 A h ∈ Y s , g = 1 A c h ∈ Y s g=1_{A^c}h\in\mathcal{Y}_s g = 1 A c h ∈ Y s , and b s , 0 ( f , g ) = 0 \mathfrak{b}_{s,0}(f,g)=0 b s , 0 ( f , g ) = 0 . Both functions are nonzero. We compute the mixed term using the positive representation (45) , rather than separating the possibly infinite Hardy terms. Let g s , 0 ε \mathfrak{g}^{\varepsilon}_{s,0} g s , 0 ε be its kernel truncation. Since f g = 0 fg=0 f g = 0 almost everywhere, expansion gives
g s , 0 ε [ f + g ] − g s , 0 ε [ f ] − g s , 0 ε [ g ] = − 4 ∫ A ∫ A c ∫ R K s ε ( q , θ , φ ) f ( θ ) g ( φ ) d q d θ d φ . \mathfrak{g}^{\varepsilon}_{s,0}[f+g]-\mathfrak{g}^{\varepsilon}_{s,0}[f]-\mathfrak{g}^{\varepsilon}_{s,0}[g]
=-4\int_A\int_{A^c}\int_{\mathbb{R}}K_s^\varepsilon(q,\theta,\varphi)f(\theta)g(\varphi)\,\mathrm{d}q\,\mathrm{d}\theta\,\mathrm{d}\varphi. g s , 0 ε [ f + g ] − g s , 0 ε [ f ] − g s , 0 ε [ g ] = − 4 ∫ A ∫ A c ∫ R K s ε ( q , θ , φ ) f ( θ ) g ( φ ) d q d θ d φ .
Each positive energy on the left converges to its finite closed-form value. Monotone convergence on the nonnegative integral on the right therefore proves that this integral is finite. Polarization gives
b s , 0 ( f , g ) = − ∫ A ∫ A c J s ( θ , φ ) f ( θ ) g ( φ ) d θ d φ < 0. \mathfrak{b}_{s,0}(f,g)=-\int_A\int_{A^c}J_s(\theta,\varphi)f(\theta)g(\varphi)\,\mathrm{d}\theta\,\mathrm{d}\varphi<0. b s , 0 ( f , g ) = − ∫ A ∫ A c J s ( θ , φ ) f ( θ ) g ( φ ) d θ d φ < 0.
This contradicts the assumed decomposition. Thus the form is irreducible.
For completeness, the invariant-ideal criterion used above states that, for a closed symmetric positivity-preserving form, an orthogonal band projection P A P_A P A commutes with one shifted resolvent if and only if P A Y s ⊂ Y s P_A\mathcal{Y}_s\subset\mathcal{Y}_s P A Y s ⊂ Y s and b s , 0 ( P A v , ( I − P A ) v ) = 0 \mathfrak{b}_{s,0}(P_Av,(I-P_A)v)=0 b s , 0 ( P A v , ( I − P A ) v ) = 0 for every v ∈ Y s v\in\mathcal{Y}_s v ∈ Y s . The forward implication follows by inserting resolvent vectors in the weak resolvent equation; these vectors are form-dense. The reverse implication follows by splitting that equation into the two bands. Thus the preceding calculation excludes every nontrivial invariant band of each shifted resolvent.
Write R α = ( B s ( 0 ) + α ) − 1 R_\alpha=(B_s(0)+\alpha)^{-1} R α = ( B s ( 0 ) + α ) − 1 . Fix β > 0 \beta>0 β > 0 and choose α > β \alpha>\beta α > β . All these resolvents preserve positivity, and ∥ R α ∥ ≤ α − 1 \lVert R_\alpha\rVert\leq\alpha^{-1} ∥ R α ∥ ≤ α − 1 . The resolvent identity therefore gives the norm-convergent Neumann series
R β = ∑ n = 1 ∞ ( α − β ) n − 1 R α n . R_\beta=\sum_{n=1}^{\infty}(\alpha-\beta)^{n-1}R_\alpha^n. R β = n = 1 ∑ ∞ ( α − β ) n − 1 R α n .
For F ≥ 0 F\geq0 F ≥ 0 , F ≠ 0 F\neq0 F = 0 , let A A A be the union, up to null sets, of the sets where R α n F > 0 R_\alpha^nF>0 R α n F > 0 , n ≥ 1 n\geq1 n ≥ 1 . Choose positive coefficients c n c_n c n so that H = ∑ n ≥ 1 c n R α n F H=\sum_{n\geq1}c_nR_\alpha^nF H = ∑ n ≥ 1 c n R α n F converges in H s , τ H_{s,\tau} H s , τ . Then H > 0 H>0 H > 0 almost everywhere on A A A , and R α H R_\alpha H R α H vanishes outside A A A . For every u ≥ 0 u\geq0 u ≥ 0 supported in A A A , the functions min { u , j H } \min\{u,jH\} min { u , j H } converge to u u u in H s , τ H_{s,\tau} H s , τ . Their images under R α R_\alpha R α vanish outside A A A , since they are bounded above by j R α H jR_\alpha H j R α H . Hence L 2 ( A , m s , τ d θ ) L^2(A,m_{s,\tau}\,d\theta) L 2 ( A , m s , τ d θ ) is invariant under R α R_\alpha R α . Self-adjointness makes it a reducing band. Irreducibility and R α F ≠ 0 R_\alpha F\neq0 R α F = 0 imply that A A A has full measure. Consequently, for every G ≥ 0 G\geq0 G ≥ 0 , G ≠ 0 G\neq0 G = 0 , there is n ≥ 1 n\geq1 n ≥ 1 with ⟨ G , R α n F ⟩ H s , τ > 0 \langle G,R_\alpha^nF\rangle_{H_{s,\tau}}>0 ⟨ G , R α n F ⟩ H s , τ > 0 . The Neumann series gives ⟨ G , R β F ⟩ H s , τ > 0 \langle G,R_\beta F\rangle_{H_{s,\tau}}>0 ⟨ G , R β F ⟩ H s , τ > 0 . Thus R β F > 0 R_\beta F>0 R β F > 0 almost everywhere, and every shifted resolvent is positivity improving.
The compact embedding (51) makes R α R_\alpha R α compact. The Perron–Frobenius theorem for a compact positivity-improving self-adjoint operator now gives a strictly positive, simple eigenvector at its spectral radius. This vector is the first eigenfunction of B s ( 0 ) B_s(0) B s ( 0 ) .
Lemma 4.7 (Diamagnetic comparison). For ψ ∈ Y s \psi\in\mathcal{Y}_s ψ ∈ Y s ,
b s , ξ [ ψ ] ≥ b s , 0 [ ∣ ψ ∣ ] . (61) b_{s,\xi}[\psi]\geq b_{s,0}[|\psi|]. \tag*{(61)} b s , ξ [ ψ ] ≥ b s , 0 [ ∣ ψ ∣ ] . (61) Consequently,
λ 1 ( ξ ) ≥ λ 1 ( 0 ) . (62) \lambda_1(\xi)\geq\lambda_1(0). \tag*{(62)} λ 1 ( ξ ) ≥ λ 1 ( 0 ) . (62) The inequality in (62) is strict when $\xi\neq 0.
Proof. For a core function, truncate the kernel. The mixed term satisfies
Re ( e i ξ q ψ ( θ ) ψ ( φ ) ‾ ) ≤ ∣ ψ ( θ ) ∣ ∣ ψ ( φ ) ∣ . \operatorname{Re}\left(e^{\mathrm{i}\xi q}\psi(\theta)\overline{\psi(\varphi)}\right)\leq|\psi(\theta)||\psi(\varphi)|. Re ( e i ξ q ψ ( θ ) ψ ( φ ) ) ≤ ∣ ψ ( θ ) ∣∣ ψ ( φ ) ∣.
The diagonal terms are unchanged by the modulus. Integration, monotone convergence in the positive representations, and core approximation give (61) . Taking the infimum proves (62) .
Assume equality for some ξ ≠ 0 \xi\neq0 ξ = 0 . Compactness gives a minimizer ψ \psi ψ . Then ∣ ψ ∣ |\psi| ∣ ψ ∣ minimizes the zero fiber. By Lemma 4.6, ∣ ψ ∣ > 0 |\psi|>0 ∣ ψ ∣ > 0 almost everywhere. Equality in the truncated phase inequality, followed by monotone convergence, implies
Re ( e i ξ q ψ ( θ ) ψ ( φ ) ‾ ) = ∣ ψ ( θ ) ∣ ∣ ψ ( φ ) ∣ . \operatorname{Re}\left(e^{\mathrm{i}\xi q}\psi(\theta)\overline{\psi(\varphi)}\right)=|\psi(\theta)||\psi(\varphi)|. Re ( e i ξ q ψ ( θ ) ψ ( φ ) ) = ∣ ψ ( θ ) ∣∣ ψ ( φ ) ∣.
for almost every ( q , θ , φ ) (q,\theta,\varphi) ( q , θ , φ ) . The defect is nonnegative. Since ∣ ψ ∣ > 0 |\psi|>0 ∣ ψ ∣ > 0 almost everywhere, Fubini’s theorem gives a pair ( θ , φ ) (\theta,\varphi) ( θ , φ ) for which the equality holds for almost every q q q . It forces e i ξ q e^{\mathrm{i}\xi q} e i ξ q to have a fixed phase, which is impossible when ξ ≠ 0 \xi\neq0 ξ = 0 . □
Lemma 4.8 (Interior oscillatory estimate). Let 0 < τ 0 < min { τ , 2 s } 0<\tau_0<\min\{\tau,2s\} 0 < τ 0 < min { τ , 2 s } . For 0 < δ < 1 / 8 0<\delta<1/8 0 < δ < 1/8 , set I δ = ( δ , π − δ ) I_\delta=(\delta,\pi-\delta) I δ = ( δ , π − δ ) . Then
∣ ξ ∣ 2 s ∫ I 2 δ ∣ ψ ∣ 2 d θ ≤ C δ − 2 ( b s , ξ [ ψ ] + ∫ 0 π m s , τ 0 ∣ ψ ∣ 2 d θ ) . (63) |\xi|^{2s}\int_{I_{2\delta}}|\psi|^2\,d\theta\leq C\delta^{-2}\left(b_{s,\xi}[\psi]+\int_0^\pi m_{s,\tau_0}|\psi|^2\,d\theta\right). \tag*{(63)} ∣ ξ ∣ 2 s ∫ I 2 δ ∣ ψ ∣ 2 d θ ≤ C δ − 2 ( b s , ξ [ ψ ] + ∫ 0 π m s , τ 0 ∣ ψ ∣ 2 d θ ) . (63) for every ψ ∈ Y s \psi\in\mathcal{Y}_s ψ ∈ Y s . The constant is independent of δ \delta δ , ξ \xi ξ , and ψ \psi ψ .
Proof. For ψ ∈ C c ∞ ( 0 , π ) \psi\in C^\infty_c(0,\pi) ψ ∈ C c ∞ ( 0 , π ) , put b = s − 1 2 b=s-\frac12 b = s − 2 1 . Choose η δ ∈ C c ∞ ( I 3 δ / 2 ) \eta_\delta\in C^\infty_c(I_{3\delta/2}) η δ ∈ C c ∞ ( I 3 δ /2 ) such that η δ = 1 \eta_\delta=1 η δ = 1 on I 2 δ I_{2\delta} I 2 δ and ∣ η δ ′ ∣ ≤ C δ − 1 |\eta_\delta'|\leq C\delta^{-1} ∣ η δ ′ ∣ ≤ C δ − 1 . Extend f = η δ ψ f=\eta_\delta\psi f = η δ ψ by zero to R \mathbb{R} R . Absorbing the angular weight in (45) into the square gives
S s , ξ ( q , θ , φ ; ψ ) : = ∣ e − q / 4 ( sin φ sin θ ) b / 2 ψ ( θ ) − e q / 4 − i ξ q ( sin θ sin φ ) b / 2 ψ ( φ ) ∣ 2 . S_{s,\xi}(q,\theta,\varphi;\psi):=\left|e^{-q/4}\left(\frac{\sin\varphi}{\sin\theta}\right)^{b/2}\psi(\theta)-e^{q/4-\mathrm{i}\xi q}\left(\frac{\sin\theta}{\sin\varphi}\right)^{b/2}\psi(\varphi)\right|^2. S s , ξ ( q , θ , φ ; ψ ) := e − q /4 ( sin θ sin φ ) b /2 ψ ( θ ) − e q /4 − i ξ q ( sin φ sin θ ) b /2 ψ ( φ ) 2 .
If θ , φ ∈ I 5 δ / 4 \theta,\varphi\in I_{5\delta/4} θ , φ ∈ I 5 δ /4 and q 2 + ∣ θ − φ ∣ 2 < ( δ / 4 ) 2 q^2+|\theta-\varphi|^2<(\delta/4)^2 q 2 + ∣ θ − φ ∣ 2 < ( δ /4 ) 2 , the sine ratios are bounded above and below by absolute constants. The mean-value theorem gives
∣ e − q / 4 ( sin φ sin θ ) b / 2 − 1 ∣ + ∣ e q / 4 ( sin θ sin φ ) b / 2 − 1 ∣ ≤ C s δ − 1 ( ∣ q ∣ + ∣ θ − φ ∣ ) . \left|e^{-q/4}\left(\frac{\sin\varphi}{\sin\theta}\right)^{b/2}-1\right|+\left|e^{q/4}\left(\frac{\sin\theta}{\sin\varphi}\right)^{b/2}-1\right|\le C_s\delta^{-1}(|q|+|\theta-\varphi|). e − q /4 ( sin θ sin φ ) b /2 − 1 + e q /4 ( sin φ sin θ ) b /2 − 1 ≤ C s δ − 1 ( ∣ q ∣ + ∣ θ − φ ∣ ) .
It follows from ∣ A + B ∣ 2 ≤ 2 ∣ A ∣ 2 + 2 ∣ B ∣ 2 |A+B|^2\le2|A|^2+2|B|^2 ∣ A + B ∣ 2 ≤ 2∣ A ∣ 2 + 2∣ B ∣ 2 that
∣ ψ ( θ ) − e − i ξ q ψ ( φ ) ∣ 2 ≤ C s S s , ξ ( q , θ , φ ; ψ ) + C s δ − 2 ( q 2 + ∣ θ − φ ∣ 2 ) ( ∣ ψ ( θ ) ∣ 2 + ∣ ψ ( φ ) ∣ 2 ) . (64) |\psi(\theta)-e^{-i\xi q}\psi(\varphi)|^2\le C_sS_{s,\xi}(q,\theta,\varphi;\psi)+C_s\delta^{-2}(q^2+|\theta-\varphi|^2)(|\psi(\theta)|^2+|\psi(\varphi)|^2). \tag*{(64)} ∣ ψ ( θ ) − e − i ξ q ψ ( φ ) ∣ 2 ≤ C s S s , ξ ( q , θ , φ ; ψ ) + C s δ − 2 ( q 2 + ∣ θ − φ ∣ 2 ) ( ∣ ψ ( θ ) ∣ 2 + ∣ ψ ( φ ) ∣ 2 ) . (64)
Use also the cutoff identity
η δ ( θ ) ψ ( θ ) − e − i ξ q η δ ( φ ) ψ ( φ ) = η δ ( θ ) [ ψ ( θ ) − e − i ξ q ψ ( φ ) ] + e − i ξ q [ η δ ( θ ) − η δ ( φ ) ] ψ ( φ ) \eta_\delta(\theta)\psi(\theta)-e^{-i\xi q}\eta_\delta(\varphi)\psi(\varphi)=\eta_\delta(\theta)[\psi(\theta)-e^{-i\xi q}\psi(\varphi)]+e^{-i\xi q}[\eta_\delta(\theta)-\eta_\delta(\varphi)]\psi(\varphi) η δ ( θ ) ψ ( θ ) − e − i ξ q η δ ( φ ) ψ ( φ ) = η δ ( θ ) [ ψ ( θ ) − e − i ξ q ψ ( φ )] + e − i ξ q [ η δ ( θ ) − η δ ( φ )] ψ ( φ )
and ∣ η δ ( θ ) − η δ ( φ ) ∣ ≤ C δ − 1 ∣ θ − φ ∣ |\eta_\delta(\theta)-\eta_\delta(\varphi)|\le C\delta^{-1}|\theta-\varphi| ∣ η δ ( θ ) − η δ ( φ ) ∣ ≤ C δ − 1 ∣ θ − φ ∣ . On the ball q 2 + ∣ θ − φ ∣ 2 < ( δ / 4 ) 2 q^2+|\theta-\varphi|^2<(\delta/4)^2 q 2 + ∣ θ − φ ∣ 2 < ( δ /4 ) 2 , any pair for which the cutoff term is nonzero lies in I 5 δ / 4 I_{5\delta/4} I 5 δ /4 . Thus (3.11), (45) , and (64) apply. The error integral is bounded by
C δ − 2 ( ∫ 0 δ / 4 r 1 − 2 s d r ) ∥ ψ ∥ 2 2 ≤ C δ − 2 s ∥ ψ ∥ 2 2 . C\delta^{-2}\left(\int_0^{\delta/4}r^{1-2s}\,dr\right)\|\psi\|_2^2\le C\delta^{-2s}\|\psi\|_2^2. C δ − 2 ( ∫ 0 δ /4 r 1 − 2 s d r ) ∥ ψ ∥ 2 2 ≤ C δ − 2 s ∥ ψ ∥ 2 2 .
On the complement of this ball, use ∣ f ( θ ) − e − i ξ q f ( φ ) ∣ 2 ≤ 2 ( ∣ f ( θ ) ∣ 2 + ∣ f ( φ ) ∣ 2 ) |f(\theta)-e^{-i\xi q}f(\varphi)|^2\le2(|f(\theta)|^2+|f(\varphi)|^2) ∣ f ( θ ) − e − i ξ q f ( φ ) ∣ 2 ≤ 2 ( ∣ f ( θ ) ∣ 2 + ∣ f ( φ ) ∣ 2 ) . Direct integration gives another C δ − 2 s ∥ ψ ∥ 2 2 C\delta^{-2s}\|\psi\|_2^2 C δ − 2 s ∥ ψ ∥ 2 2 bound. Since δ − 2 s ≤ δ − 2 \delta^{-2s}\le\delta^{-2} δ − 2 s ≤ δ − 2 , we have proved
∫ R 2 ∫ R ∣ f ( θ ) − e − i ξ q f ( φ ) ∣ 2 ( q 2 + ∣ θ − φ ∣ 2 ) 1 + s d q d θ d φ ≤ C δ − 2 ( b s , ξ [ ψ ] + ∥ ψ ∥ L 2 ( 0 , π ) 2 ) . (65) \int_{\mathbb{R}^2}\int_{\mathbb{R}}\frac{|f(\theta)-e^{-i\xi q}f(\varphi)|^2}{(q^2+|\theta-\varphi|^2)^{1+s}}\,dq\,d\theta\,d\varphi\le C\delta^{-2}\left(b_{s,\xi}[\psi]+\|\psi\|_{L^2(0,\pi)}^2\right). \tag*{(65)} ∫ R 2 ∫ R ( q 2 + ∣ θ − φ ∣ 2 ) 1 + s ∣ f ( θ ) − e − i ξ q f ( φ ) ∣ 2 d q d θ d φ ≤ C δ − 2 ( b s , ξ [ ψ ] + ∥ ψ ∥ L 2 ( 0 , π ) 2 ) . (65)
Let f ^ \widehat f f be the unitary Fourier transform on R \mathbb{R} R . Put α = θ − φ \alpha=\theta-\varphi α = θ − φ and apply Plancherel in θ \theta θ . We use the identity
∫ R 2 1 − cos ( ξ q + ζ α ) ( q 2 + α 2 ) 1 + s d q d α = c s ( ξ 2 + ζ 2 ) s , \int_{\mathbb{R}^2}\frac{1-\cos(\xi q+\zeta\alpha)}{(q^2+\alpha^2)^{1+s}}\,dq\,d\alpha=c_s(\xi^2+\zeta^2)^s, ∫ R 2 ( q 2 + α 2 ) 1 + s 1 − cos ( ξ q + ζ α ) d q d α = c s ( ξ 2 + ζ 2 ) s ,
where c s > 0 c_s>0 c s > 0 . Plancherel’s theorem gives, after absorbing the harmless factor two into c s c_s c s ,
∫ R 2 ∫ R ∣ f ( θ ) − e − i ξ q f ( φ ) ∣ 2 ( q 2 + ∣ θ − φ ∣ 2 ) 1 + s d q d θ d φ = c s ∫ R ( ξ 2 + ζ 2 ) s ∣ f ^ ( ζ ) ∣ 2 d ζ ≥ c s ∣ ξ ∣ 2 s ∥ f ∥ 2 2 . \int_{\mathbb{R}^2}\int_{\mathbb{R}}\frac{|f(\theta)-e^{-i\xi q}f(\varphi)|^2}{(q^2+|\theta-\varphi|^2)^{1+s}}\,dq\,d\theta\,d\varphi=c_s\int_{\mathbb{R}}(\xi^2+\zeta^2)^s|\widehat f(\zeta)|^2\,d\zeta\ge c_s|\xi|^{2s}\|f\|_2^2. ∫ R 2 ∫ R ( q 2 + ∣ θ − φ ∣ 2 ) 1 + s ∣ f ( θ ) − e − i ξ q f ( φ ) ∣ 2 d q d θ d φ = c s ∫ R ( ξ 2 + ζ 2 ) s ∣ f ( ζ ) ∣ 2 d ζ ≥ c s ∣ ξ ∣ 2 s ∥ f ∥ 2 2 .
Since η δ = 1 \eta_\delta=1 η δ = 1 on I 2 δ I_{2\delta} I 2 δ , ∥ f ∥ 2 2 ≥ ∫ I 2 δ ∣ ψ ∣ 2 \|f\|_2^2\ge\int_{I_{2\delta}}|\psi|^2 ∥ f ∥ 2 2 ≥ ∫ I 2 δ ∣ ψ ∣ 2 . Finally, (53) controls the last L 2 L^2 L 2 term in (65) . This proves (63) for core functions. Approximate an arbitrary ψ ∈ Y s \psi\in\mathcal{Y}_s ψ ∈ Y s by core functions ψ n \psi_n ψ n . Apply (65) to ψ n − ψ m \psi_n-\psi_m ψ n − ψ m . It shows that η δ ψ n \eta_\delta\psi_n η δ ψ n is Cauchy in the flat Fourier form. Its L 2 L^2 L 2 limit is η δ ψ \eta_\delta\psi η δ ψ . Passing to the limit in the Fourier identity and using lower semicontinuity proves the general case. □
Lemma 4.9 (Large-frequency coercivity). There are c , C > 0 c,C>0 c , C > 0 and β > 0 \beta>0 β > 0 such that
b s , ξ [ ψ ] ≥ ( c ∣ ξ ∣ β − C ) ∫ 0 π m s , τ ∣ ψ ∣ 2 d θ (66) b_{s,\xi}[\psi]\ge(c|\xi|^\beta-C)\int_0^\pi m_{s,\tau}|\psi|^2\,d\theta \tag*{(66)} b s , ξ [ ψ ] ≥ ( c ∣ ξ ∣ β − C ) ∫ 0 π m s , τ ∣ ψ ∣ 2 d θ (66) for every ψ ∈ Y s \psi\in\mathcal{Y}_s ψ ∈ Y s . Hence
λ 1 ( ξ ) ⟶ ∞ as ∣ ξ ∣ ⟶ ∞ . (67) \lambda_1(\xi)\longrightarrow\infty\quad\text{as }|\xi|\longrightarrow\infty. \tag*{(67)} λ 1 ( ξ ) ⟶ ∞ as ∣ ξ ∣ ⟶ ∞. (67)
Proof. Choose 0 < τ 0 < min { τ , 2 s } 0<\tau_0<\min\{\tau,2s\} 0 < τ 0 < min { τ , 2 s } . By (52) and (61) ,
∫ 0 π m s , τ 0 ∣ ψ ∣ 2 d θ ≤ C b s , 0 [ ∣ ψ ∣ ] ≤ C b s , ξ [ ψ ] . \int_0^\pi m_{s,\tau_0}|\psi|^2\,d\theta\le Cb_{s,0}[|\psi|]\le Cb_{s,\xi}[\psi]. ∫ 0 π m s , τ 0 ∣ ψ ∣ 2 d θ ≤ C b s , 0 [ ∣ ψ ∣ ] ≤ C b s , ξ [ ψ ] .
On the boundary part,
∫ ( 0 , 2 δ ) ∪ ( π − 2 δ , π ) m s , τ ∣ ψ ∣ 2 d θ ≤ C δ τ − τ 0 b s , ξ [ ψ ] . (68) \int_{(0,2\delta)\cup(\pi-2\delta,\pi)}m_{s,\tau}|\psi|^2\,d\theta\le C\delta^{\tau-\tau_0}b_{s,\xi}[\psi]. \tag*{(68)} ∫ ( 0 , 2 δ ) ∪ ( π − 2 δ , π ) m s , τ ∣ ψ ∣ 2 d θ ≤ C δ τ − τ 0 b s , ξ [ ψ ] . (68)
On I 2 δ I_{2\delta} I 2 δ , m s , τ ≤ C δ − ∣ τ − 2 s ∣ m_{s,\tau}\le C\delta^{-\lvert\tau-2s\rvert} m s , τ ≤ C δ − ∣ τ − 2 s ∣ . Combine (63) and (68) . We obtain
∫ 0 π m s , τ ∣ ψ ∣ 2 d θ ≤ C ( ∣ ξ ∣ − 2 s δ − 2 − ∣ τ − 2 s ∣ + δ τ − τ 0 ) b s , ξ [ ψ ] . \int_0^\pi m_{s,\tau}\lvert\psi\rvert^2\,d\theta\le C\left(\lvert\xi\rvert^{-2s}\delta^{-2-\lvert\tau-2s\rvert}+\delta^{\tau-\tau_0}\right)b_{s,\xi}[\psi]. ∫ 0 π m s , τ ∣ ψ ∣ 2 d θ ≤ C ( ∣ ξ ∣ − 2 s δ − 2 − ∣ τ − 2 s ∣ + δ τ − τ 0 ) b s , ξ [ ψ ] .
Put η = τ − τ 0 \eta=\tau-\tau_0 η = τ − τ 0 and M = ∣ τ − 2 s ∣ M=\lvert\tau-2s\rvert M = ∣ τ − 2 s ∣ . Choose
ε = 2 s 2 + M + η , δ = ∣ ξ ∣ − ε . \varepsilon=\frac{2s}{2+M+\eta},\qquad\delta=\lvert\xi\rvert^{-\varepsilon}. ε = 2 + M + η 2 s , δ = ∣ ξ ∣ − ε .
For large ∣ ξ ∣ \lvert\xi\rvert ∣ ξ ∣ , we have δ < 1 / 8 \delta<1/8 δ < 1/8 . Both coefficients in the preceding estimate are bounded by C ∣ ξ ∣ − β C\lvert\xi\rvert^{-\beta} C ∣ ξ ∣ − β , where
β = 2 s ( τ − τ 0 ) 2 + ∣ τ − 2 s ∣ + τ − τ 0 > 0. \beta=\frac{2s(\tau-\tau_0)}{2+\lvert\tau-2s\rvert+\tau-\tau_0}>0. β = 2 + ∣ τ − 2 s ∣ + τ − τ 0 2 s ( τ − τ 0 ) > 0.
Thus
b s , ξ [ ψ ] ≥ c ∣ ξ ∣ β ∥ ψ ∥ H s , τ 2 b_{s,\xi}[\psi]\ge c\lvert\xi\rvert^\beta\lVert\psi\rVert_{H_{s,\tau}}^2 b s , ξ [ ψ ] ≥ c ∣ ξ ∣ β ∥ ψ ∥ H s , τ 2
for large ∣ ξ ∣ \lvert\xi\rvert ∣ ξ ∣ . Adjusting the lower bound on a bounded frequency interval gives (66) . Taking the infimum gives (67) . □
Kernel moments and frequency growth
Lemma 4.10 (Moment bounds). For every integer j ≥ 1 j\ge1 j ≥ 1 , the form
∫ 0 π ∫ 0 π ∫ R ∣ q ∣ 2 j K s ( q , θ , φ ) ∣ ψ ( θ ) ∣ ∣ ψ ( φ ) ∣ d q d θ d φ (69) \int_0^\pi\int_0^\pi\int_{\mathbb{R}}\lvert q\rvert^{2j}K_s(q,\theta,\varphi)\lvert\psi(\theta)\rvert\lvert\psi(\varphi)\rvert\,dq\,d\theta\,d\varphi \tag*{(69)} ∫ 0 π ∫ 0 π ∫ R ∣ q ∣ 2 j K s ( q , θ , φ ) ∣ ψ ( θ )∣ ∣ ψ ( φ )∣ d q d θ d φ (69) is finite on Y s \mathcal{Y}_s Y s and is bounded by C j ( b s , 0 [ ψ ] + ∥ ψ ∥ L 2 ( m s , τ ) 2 ) C_j\bigl(b_{s,0}[\psi]+\lVert\psi\rVert_{L^2(m_{s,\tau})}^2\bigr) C j ( b s , 0 [ ψ ] + ∥ ψ ∥ L 2 ( m s , τ ) 2 ) . For ψ , χ ∈ Y s \psi,\chi\in\mathcal{Y}_s ψ , χ ∈ Y s and ∣ Im z ∣ < 1 + s \lvert\operatorname{Im}z\rvert<1+s ∣ Im z ∣ < 1 + s , define
d s , z ( ψ , χ ) = ∫ 0 π ∫ 0 π ∫ R ( 1 − cos ( z q ) ) K s ( q , θ , φ ) ( ψ ( θ ) χ ( φ ) ‾ + ψ ( φ ) χ ( θ ) ‾ ) d q d θ d φ d_{s,z}(\psi,\chi)=\int_0^\pi\int_0^\pi\int_{\mathbb{R}}(1-\cos(zq))K_s(q,\theta,\varphi)\bigl(\psi(\theta)\overline{\chi(\varphi)}+\psi(\varphi)\overline{\chi(\theta)}\bigr)\,dq\,d\theta\,d\varphi d s , z ( ψ , χ ) = ∫ 0 π ∫ 0 π ∫ R ( 1 − cos ( z q )) K s ( q , θ , φ ) ( ψ ( θ ) χ ( φ ) + ψ ( φ ) χ ( θ ) ) d q d θ d φ and b s , z = b s , 0 + d s , z b_{s,z}=b_{s,0}+d_{s,z} b s , z = b s , 0 + d s , z . Then z ↦ d s , z z\mapsto d_{s,z} z ↦ d s , z is a holomorphic family of bounded forms on Y s \mathcal{Y}_s Y s in this strip. It is even. For ∣ z ∣ \lvert z\rvert ∣ z ∣ sufficiently small, b s , z b_{s,z} b s , z is a closed sectorial form with common domain Y s \mathcal{Y}_s Y s . The resulting family is of type (a). Moreover,
b s , z = b s , 0 + z 2 c s , 2 + O Y s ( z 4 ) as z → 0. (70) b_{s,z}=b_{s,0}+z^2c_{s,2}+O_{\mathcal{Y}_s}(z^4)\quad\text{as }z\to0. \tag*{(70)} b s , z = b s , 0 + z 2 c s , 2 + O Y s ( z 4 ) as z → 0. (70) Here
c s , 2 ( ψ , χ ) = 1 2 ∫ 0 π ∫ 0 π ∫ R q 2 K s ( q , θ , φ ) ( ψ ( θ ) χ ( φ ) ‾ + ψ ( φ ) χ ( θ ) ‾ ) d q d θ d φ . (71) c_{s,2}(\psi,\chi)=\frac12\int_0^\pi\int_0^\pi\int_{\mathbb{R}}q^2K_s(q,\theta,\varphi)\bigl(\psi(\theta)\overline{\chi(\varphi)}+\psi(\varphi)\overline{\chi(\theta)}\bigr)\,dq\,d\theta\,d\varphi. \tag*{(71)} c s , 2 ( ψ , χ ) = 2 1 ∫ 0 π ∫ 0 π ∫ R q 2 K s ( q , θ , φ ) ( ψ ( θ ) χ ( φ ) + ψ ( φ ) χ ( θ ) ) d q d θ d φ . (71)
Proof. For ∣ q ∣ ≥ 1 \lvert q\rvert\ge1 ∣ q ∣ ≥ 1 , use (42) and ∥ ψ ∥ L 1 ( 0 , π ) ≤ π ∥ ψ ∥ L 2 ( 0 , π ) \lVert\psi\rVert_{L^1(0,\pi)}\le\sqrt{\pi}\lVert\psi\rVert_{L^2(0,\pi)} ∥ ψ ∥ L 1 ( 0 , π ) ≤ π ∥ ψ ∥ L 2 ( 0 , π ) . Put α = θ − φ \alpha=\theta-\varphi α = θ − φ . For ∣ q ∣ < 1 \lvert q\rvert<1 ∣ q ∣ < 1 and ∣ α ∣ ≤ 1 \lvert\alpha\rvert\le1 ∣ α ∣ ≤ 1 , use (41) . On 1 ≤ ∣ α ∣ ≤ π 1\le\lvert\alpha\rvert\le\pi 1 ≤ ∣ α ∣ ≤ π , the kernel is uniformly bounded because cosh q − cos α ≥ 1 − cos 1 > 0 \cosh q-\cos\alpha\ge1-\cos1>0 cosh q − cos α ≥ 1 − cos 1 > 0 . Integration in q q q of the local bound gives
∫ − 1 1 ∣ q ∣ 2 j ( q 2 + α 2 ) 1 + s d q ≤ C j { 1 + ∣ α ∣ 2 j − 1 − 2 s , 2 j − 1 − 2 s < 0 , 1 + ∣ log ∣ α ∣ ∣ , 2 j − 1 − 2 s = 0 , 1 , 2 j − 1 − 2 s > 0. (72) \int_{-1}^{1}\frac{\lvert q\rvert^{2j}}{(q^2+\alpha^2)^{1+s}}\,dq\le C_j
\begin{cases}
1+\lvert\alpha\rvert^{2j-1-2s},&2j-1-2s<0,\\
1+\lvert\log\lvert\alpha\rvert\rvert,&2j-1-2s=0,\\
1,&2j-1-2s>0.
\end{cases} \tag*{(72)} ∫ − 1 1 ( q 2 + α 2 ) 1 + s ∣ q ∣ 2 j d q ≤ C j ⎩ ⎨ ⎧ 1 + ∣ α ∣ 2 j − 1 − 2 s , 1 + ∣ log ∣ α ∣∣ , 1 , 2 j − 1 − 2 s < 0 , 2 j − 1 − 2 s = 0 , 2 j − 1 − 2 s > 0. (72)
Indeed, split at ∣ q ∣ = ∣ α ∣ \lvert q\rvert=\lvert\alpha\rvert ∣ q ∣ = ∣ α ∣ and integrate q 2 j − 2 − 2 s q^{2j-2-2s} q 2 j − 2 − 2 s on the outer part. All three kernels are integrable in α \alpha α , because 2 j − 1 − 2 s > − 1 2j-1-2s>-1 2 j − 1 − 2 s > − 1 . Young’s inequality and (53) give the left-hand side of (69) ≤ C j ∥ ψ ∥ L 2 ( 0 , π ) 2 ≤ C j b s , 0 [ ψ ] \le C_j\lVert\psi\rVert_{L^2(0,\pi)}^2\le C_jb_{s,0}[\psi] ≤ C j ∥ ψ ∥ L 2 ( 0 , π ) 2 ≤ C j b s , 0 [ ψ ] .
For analyticity, let K \mathcal{K} K be a compact subset of { z : ∣ Im z ∣ < 1 + s } \{z:\lvert\operatorname{Im}z\rvert<1+s\} { z : ∣ Im z ∣ < 1 + s } . Put
r = max z ∈ K ∣ Im z ∣ < 1 + s . r=\max_{z\in\mathcal{K}}\lvert\operatorname{Im}z\rvert<1+s. r = z ∈ K max ∣ Im z ∣ < 1 + s .
For z ∈ K z\in\mathcal{K} z ∈ K and ∣ q ∣ ≤ 1 \lvert q\rvert\le1 ∣ q ∣ ≤ 1 ,
∣ 1 − cos ( z q ) ∣ ≤ C K q 2 . \lvert1-\cos(zq)\rvert\le C_{\mathcal{K}}q^2. ∣ 1 − cos ( z q )∣ ≤ C K q 2 .
For ∣ q ∣ ≥ 1 |q| \ge1 ∣ q ∣ ≥ 1 ,
∣ 1 − cos ( z q ) ∣ ≤ C K e r ∣ q ∣ . |1-\cos(zq)| \le C_K e^{r|q|}. ∣1 − cos ( z q ) ∣ ≤ C K e r ∣ q ∣ .
The first bound is controlled by (72) with j = 1 j=1 j = 1 . The second is integrable because K s ( q , θ , φ ) ≤ C e − ( 1 + s ) ∣ q ∣ K_s(q,\theta,\varphi) \le C e^{-(1+s)|q|} K s ( q , θ , φ ) ≤ C e − ( 1 + s ) ∣ q ∣ . For every integer ℓ ≥ 1 \ell\ge1 ℓ ≥ 1 , the tail of the ℓ \ell ℓ th z z z -derivative is bounded by C K , ℓ ∣ q ∣ ℓ e r ∣ q ∣ C_{K,\ell}|q|^\ell e^{r|q|} C K , ℓ ∣ q ∣ ℓ e r ∣ q ∣ , uniformly on K \mathcal{K} K . After multiplication by K s K_s K s , the polynomial factor is absorbed by the positive exponential margin 1 + s − r 1+s-r 1 + s − r , giving an integrable tail. Near q = 0 q=0 q = 0 , even derivatives of positive order contain q ℓ cos ( z q ) q^\ell\cos(zq) q ℓ cos ( z q ) , whereas odd derivatives contain q ℓ sin ( z q ) = O ( ∣ q ∣ ℓ + 1 ) q^\ell\sin(zq)=O(|q|^{\ell+1}) q ℓ sin ( z q ) = O ( ∣ q ∣ ℓ + 1 ) . Each has at least quadratic cancellation. Differentiation under the integral is therefore valid in bounded-form norm on Y s \mathcal{Y}_s Y s . For z z z near zero, the moment estimate and ∣ 1 − cos ( z q ) ∣ ≤ C ∣ z ∣ 2 q 2 e ∣ Im z ∣ ∣ q ∣ |1-\cos(zq)| \le C|z|^2q^2e^{|\operatorname{Im}z||q|} ∣1 − cos ( z q ) ∣ ≤ C ∣ z ∣ 2 q 2 e ∣ Im z ∣∣ q ∣ give
∣ b s , z [ ψ ] ∣ ≤ C ∣ z ∣ 2 ( b s , 0 [ ψ ] + ∥ ψ ∥ H s , τ 2 ) ≤ C ′ ∣ z ∣ 2 b s , 0 [ ψ ] . |\mathfrak{b}_{s,z}[\psi]| \le C|z|^2\left(b_{s,0}[\psi]+\|\psi\|_{H_{s,\tau}}^2\right) \le C'|z|^2b_{s,0}[\psi]. ∣ b s , z [ ψ ] ∣ ≤ C ∣ z ∣ 2 ( b s , 0 [ ψ ] + ∥ ψ ∥ H s , τ 2 ) ≤ C ′ ∣ z ∣ 2 b s , 0 [ ψ ] .
By (50) , the relative bound is less than one for small ∣ z ∣ |z| ∣ z ∣ . Hence b s , z b_{s,z} b s , z is closed and sectorial on Y s \mathcal{Y}_s Y s , of type (a); the associated operators are of type (B).
For z z z near zero, Taylor’s formula gives
1 − cos ( z q ) = z 2 q 2 2 + O ( ∣ z ∣ 4 q 4 e ∣ Im z ∣ ∣ q ∣ ) . 1-\cos(zq)=\frac{z^2q^2}{2}+O\left(|z|^4q^4e^{|\operatorname{Im}z||q|}\right). 1 − cos ( z q ) = 2 z 2 q 2 + O ( ∣ z ∣ 4 q 4 e ∣ Im z ∣∣ q ∣ ) .
Use (72) with j = 2 j=2 j = 2 for ∣ q ∣ ≤ 1 |q|\le1 ∣ q ∣ ≤ 1 and the exponential tail for ∣ q ∣ ≥ 1 |q|\ge1 ∣ q ∣ ≥ 1 . This proves (70) with (71) . Evenness follows directly from the cosine. □
Lemma 4.11 (Two-sided fiber growth). There is C ≥ 1 C \ge1 C ≥ 1 such that
b s , ξ [ ψ ] ≤ C ( 1 + ∣ ξ ∣ 2 s ) ( b s , 0 [ ψ ] + ∥ ψ ∥ L 2 ( m s , τ ) 2 ) . (73) b_{s,\xi}[\psi] \le C(1+|\xi|^{2s})\left(b_{s,0}[\psi]+\|\psi\|_{L^2(m_{s,\tau})}^2\right). \tag*{(73)} b s , ξ [ ψ ] ≤ C ( 1 + ∣ ξ ∣ 2 s ) ( b s , 0 [ ψ ] + ∥ ψ ∥ L 2 ( m s , τ ) 2 ) . (73) For ψ = ϕ s , τ \psi=\phi_{s,\tau} ψ = ϕ s , τ ,
c ( 1 + ∣ ξ ∣ 2 s ) − C ≤ b s , ξ [ ϕ s , τ ] ≤ C ( 1 + ∣ ξ ∣ 2 s ) . (74) c(1+|\xi|^{2s})-C \le b_{s,\xi}[\phi_{s,\tau}] \le C(1+|\xi|^{2s}). \tag*{(74)} c ( 1 + ∣ ξ ∣ 2 s ) − C ≤ b s , ξ [ ϕ s , τ ] ≤ C ( 1 + ∣ ξ ∣ 2 s ) . (74)
Proof. By (54) , (55) , and (53) ,
∣ b s , ξ [ ψ ] − b s , 0 [ ψ ] ∣ ≤ C min { ξ 2 , ∣ ξ ∣ 2 s } ∥ ψ ∥ L 2 ( 0 , π ) 2 ≤ C ( 1 + ∣ ξ ∣ 2 s ) b s , 0 [ ψ ] . |b_{s,\xi}[\psi]-b_{s,0}[\psi]| \le C\min\{\xi^2,|\xi|^{2s}\}\|\psi\|_{L^2(0,\pi)}^2 \le C(1+|\xi|^{2s})b_{s,0}[\psi]. ∣ b s , ξ [ ψ ] − b s , 0 [ ψ ] ∣ ≤ C min { ξ 2 , ∣ ξ ∣ 2 s } ∥ ψ ∥ L 2 ( 0 , π ) 2 ≤ C ( 1 + ∣ ξ ∣ 2 s ) b s , 0 [ ψ ] .
This proves (73) and the upper bound in (74) .
For the lower bound, choose 0 < τ 0 < min { τ , 2 s } 0<\tau_0<\min\{\tau,2s\} 0 < τ 0 < min { τ , 2 s } and fix 0 < δ < 1 / 8 0<\delta<1/8 0 < δ < 1/8 so small that
∫ I 2 δ ϕ s , τ 2 d θ > 0. \int_{I_{2\delta}}\phi_{s,\tau}^2\,d\theta>0. ∫ I 2 δ ϕ s , τ 2 d θ > 0.
This is possible because ϕ s , τ > 0 \phi_{s,\tau}>0 ϕ s , τ > 0 almost everywhere. Apply (63) to ϕ s , τ \phi_{s,\tau} ϕ s , τ . The term with m s , τ 0 m_{s,\tau_0} m s , τ 0 is finite by (52) . Rearranging gives
b s , ξ [ ϕ s , τ ] ≥ C − 1 δ 2 ∣ ξ ∣ 2 s ∫ I 2 δ ϕ s , τ 2 d θ − ∫ 0 π m s , τ 0 ϕ s , τ 2 d θ . b_{s,\xi}[\phi_{s,\tau}] \ge C^{-1}\delta^2|\xi|^{2s}\int_{I_{2\delta}}\phi_{s,\tau}^2\,d\theta-\int_0^\pi m_{s,\tau_0}\phi_{s,\tau}^2\,d\theta. b s , ξ [ ϕ s , τ ] ≥ C − 1 δ 2 ∣ ξ ∣ 2 s ∫ I 2 δ ϕ s , τ 2 d θ − ∫ 0 π m s , τ 0 ϕ s , τ 2 d θ .
Hence b s , ξ [ ϕ s , τ ] ≥ c ∣ ξ ∣ 2 s − C b_{s,\xi}[\phi_{s,\tau}] \ge c|\xi|^{2s}-C b s , ξ [ ϕ s , τ ] ≥ c ∣ ξ ∣ 2 s − C . Positivity handles bounded frequencies. □
The sharp constant and its ground-state representation
Lemma 5.1 (Logarithmic cutoff). Let χ ∈ C ∞ ( R ) \chi\in C^\infty(\mathbb{R}) χ ∈ C ∞ ( R ) satisfy ∥ χ ∥ 2 = 1 \|\chi\|_2=1 ∥ χ ∥ 2 = 1 . Let ψ ∈ Y s \psi\in\mathcal{Y}_s ψ ∈ Y s and
w R ( t , θ ) = R − 1 / 2 χ ( t / R ) ψ ( θ ) . w_R(t,\theta)=R^{-1/2}\chi(t/R)\psi(\theta). w R ( t , θ ) = R − 1/2 χ ( t / R ) ψ ( θ ) . Then
∥ w R ∥ L 2 ( m s , τ ) 2 = ∥ ψ ∥ L 2 ( m s , τ ) 2 , q s [ w R ] ⟶ b s , 0 [ ψ ] . \|w_R\|_{L^2(m_{s,\tau})}^2=\|\psi\|_{L^2(m_{s,\tau})}^2,\qquad q_s[w_R]\longrightarrow b_{s,0}[\psi]. ∥ w R ∥ L 2 ( m s , τ ) 2 = ∥ ψ ∥ L 2 ( m s , τ ) 2 , q s [ w R ] ⟶ b s , 0 [ ψ ] . If ψ ∈ C ∞ ( 0 , π ) \psi\in C^\infty(0,\pi) ψ ∈ C ∞ ( 0 , π ) , then w R w_R w R is the transform of a function in C ∞ ( R + 2 ) C^\infty(\mathbb{R}_+^2) C ∞ ( R + 2 ) .
Proof. The change of variables t = R r t=Rr t = R r gives the norm identity. Also,
w ^ R ( ξ , θ ) = R 1 / 2 χ ^ ( R ξ ) ψ ( θ ) . \widehat{w}_R(\xi,\theta)=R^{1/2}\widehat{\chi}(R\xi)\psi(\theta). w R ( ξ , θ ) = R 1/2 χ ( R ξ ) ψ ( θ ) .
By (59) ,
q s [ w R ] = ∫ R R ∣ χ ^ ( R ξ ) ∣ 2 b s , ξ [ ψ ] d ξ . q_s[w_R]=\int_{\mathbb{R}}R|\widehat{\chi}(R\xi)|^2b_{s,\xi}[\psi]\,d\xi. q s [ w R ] = ∫ R R ∣ χ ( R ξ ) ∣ 2 b s , ξ [ ψ ] d ξ .
Putting ζ = R ξ \zeta=R\xi ζ = R ξ , we have
q s [ w R ] = ∫ R ∣ χ ^ ( ζ ) ∣ 2 b s , ζ / R [ ψ ] d ζ . q_s[w_R]=\int_{\mathbb{R}}|\widehat{\chi}(\zeta)|^2b_{s,\zeta/R}[\psi]\,d\zeta. q s [ w R ] = ∫ R ∣ χ ( ζ ) ∣ 2 b s , ζ / R [ ψ ] d ζ .
For R ≥ 1 R\geq1 R ≥ 1 , (55) and (54) give
b s , ζ / R [ ψ ] ≤ C ( 1 + ∣ ζ ∣ 2 s ) ( b s , 0 [ ψ ] + ∥ ψ ∥ L 2 ( m s , τ ) 2 ) . b_{s,\zeta/R}[\psi]\leq C(1+|\zeta|^{2s})\left(b_{s,0}[\psi]+\|\psi\|_{L^2(m_{s,\tau})}^2\right). b s , ζ / R [ ψ ] ≤ C ( 1 + ∣ ζ ∣ 2 s ) ( b s , 0 [ ψ ] + ∥ ψ ∥ L 2 ( m s , τ ) 2 ) .
Since χ ^ \widehat{\chi} χ is rapidly decreasing and the fiber form is continuous at zero, dominated convergence proves the energy limit. If ψ ∈ C c ∞ ( 0 , π ) \psi\in C^\infty_c(0,\pi) ψ ∈ C c ∞ ( 0 , π ) , the inverse transform has compact support in the half-plane by (31) .
Proof of Theorem 1.2. Let u ∈ C c ∞ ( R + 2 ) u\in C^\infty_c(\mathbb{R}^2_+) u ∈ C c ∞ ( R + 2 ) and let w w w be its transform. Proposition 4.1 and Lemma 4.7 yield
H s [ u ] = ∫ b s , ξ [ w ^ ] d ξ ≥ λ 1 ( 0 ) ∫ ∥ w ^ ( ξ , ⋅ ) ∥ L 2 ( m s , τ ) 2 d ξ = λ 1 ( 0 ) R s , τ [ u ] . \mathcal{H}_s[u]=\int b_{s,\xi}[\widehat{w}]\,d\xi\geq\lambda_1(0)\int\|\widehat{w}(\xi,\cdot)\|_{L^2(m_{s,\tau})}^2\,d\xi=\lambda_1(0)\mathcal{R}_{s,\tau}[u]. H s [ u ] = ∫ b s , ξ [ w ] d ξ ≥ λ 1 ( 0 ) ∫ ∥ w ( ξ , ⋅ ) ∥ L 2 ( m s , τ ) 2 d ξ = λ 1 ( 0 ) R s , τ [ u ] .
Hence Λ s , τ ≥ λ 1 ( 0 ) \Lambda_{s,\tau}\geq\lambda_1(0) Λ s , τ ≥ λ 1 ( 0 ) . By the definition of Y s \mathcal{Y}_s Y s , choose normalized ψ j ∈ C c ∞ ( 0 , π ) \psi_j\in C^\infty_c(0,\pi) ψ j ∈ C c ∞ ( 0 , π ) with
ψ j ⟶ ϕ s , τ in Y s . \psi_j\longrightarrow\phi_{s,\tau}\quad\text{in }\mathcal{Y}_s. ψ j ⟶ ϕ s , τ in Y s .
For each j j j , the cutoffs in Lemma 5.1 correspond to half-plane test functions. Letting first R → ∞ R\to\infty R → ∞ and then j → ∞ j\to\infty j → ∞ , we obtain
Λ s , τ ≤ lim j → ∞ b s , 0 [ ψ j ] = λ 1 ( 0 ) . \Lambda_{s,\tau}\leq\lim_{j\to\infty}b_{s,0}[\psi_j]=\lambda_1(0). Λ s , τ ≤ j → ∞ lim b s , 0 [ ψ j ] = λ 1 ( 0 ) .
The positive lower bound follows from (50) . Lemmas 4.2 and 4.6 give the unique normalized positive angular minimizer.
Suppose that a nonzero w ∈ X s w\in\mathcal{X}_s w ∈ X s attains the quotient. Then
0 = ∫ R ( b s , ξ [ w ^ ( ξ , ⋅ ) ] − λ 1 ( 0 ) ∥ w ^ ( ξ , ⋅ ) ∥ L 2 ( m s , τ ) 2 ) d ξ . 0=\int_{\mathbb{R}}\left(b_{s,\xi}[\widehat{w}(\xi,\cdot)]-\lambda_1(0)\|\widehat{w}(\xi,\cdot)\|_{L^2(m_{s,\tau})}^2\right)\,d\xi. 0 = ∫ R ( b s , ξ [ w ( ξ , ⋅ )] − λ 1 ( 0 ) ∥ w ( ξ , ⋅ ) ∥ L 2 ( m s , τ ) 2 ) d ξ .
For almost every ξ \xi ξ , the integrand is at least
( λ 1 ( ξ ) − λ 1 ( 0 ) ) ∥ w ^ ( ξ , ⋅ ) ∥ L 2 ( m s , τ ) 2 ≥ 0. \left(\lambda_1(\xi)-\lambda_1(0)\right)\|\widehat{w}(\xi,\cdot)\|_{L^2(m_{s,\tau})}^2\geq0. ( λ 1 ( ξ ) − λ 1 ( 0 ) ) ∥ w ( ξ , ⋅ ) ∥ L 2 ( m s , τ ) 2 ≥ 0.
By the strict inequality in Lemma 4.7, w ^ ( ξ , ⋅ ) = 0 \widehat{w}(\xi,\cdot)=0 w ( ξ , ⋅ ) = 0 for almost every ξ ≠ 0 \xi\neq0 ξ = 0 . Since w ^ \widehat{w} w is an L 2 L^2 L 2 function, it follows that w = 0 w=0 w = 0 . This contradiction proves non-attainment.
Corollary 5.2 (Angular Euler–Lagrange equation). For every ψ ∈ Y s \psi\in\mathcal{Y}_s ψ ∈ Y s ,
b s , 0 ( ϕ s , τ , ψ ) = Λ s , τ ∫ 0 π m s , τ ϕ s , τ ψ ‾ d θ . (75) b_{s,0}(\phi_{s,\tau},\psi)=\Lambda_{s,\tau}\int_0^\pi m_{s,\tau}\phi_{s,\tau}\overline{\psi}\,d\theta. \tag*{(75)} b s , 0 ( ϕ s , τ , ψ ) = Λ s , τ ∫ 0 π m s , τ ϕ s , τ ψ d θ . (75)
where b s , 0 ( ⋅ , ⋅ ) b_{s,0}(\cdot,\cdot) b s , 0 ( ⋅ , ⋅ ) is the polarization of (7) . The same equation characterizes ϕ s , τ \phi_{s,\tau} ϕ s , τ among normalized nonnegative functions in Y s \mathcal{Y}_s Y s .
Proof. For ψ ∈ Y s \psi\in\mathcal{Y}_s ψ ∈ Y s and r ∈ R r\in\mathbb{R} r ∈ R , minimality gives
b s , 0 [ ϕ s , τ + r ψ ] − Λ s , τ ∥ ϕ s , τ + r ψ ∥ H s , τ 2 ≥ 0. b_{s,0}[\phi_{s,\tau}+r\psi]-\Lambda_{s,\tau}\|\phi_{s,\tau}+r\psi\|_{\mathcal{H}_{s,\tau}}^2\geq0. b s , 0 [ ϕ s , τ + r ψ ] − Λ s , τ ∥ ϕ s , τ + r ψ ∥ H s , τ 2 ≥ 0.
Differentiating at r = 0 r=0 r = 0 gives the real part of (75) . Replacing ψ \psi ψ by i ψ i\psi i ψ gives the imaginary part. Conversely, testing the same equation for a normalized nonnegative ϕ \phi ϕ against ϕ \phi ϕ shows that ϕ \phi ϕ is a minimizer. Lemma 4.6 then gives ϕ = ϕ s , τ \phi=\phi_{s,\tau} ϕ = ϕ s , τ .
Theorem 5.3 (Ground-state representation at the sharp level). Let 0 < s < 1 0<s<1 0 < s < 1 and τ > 0 \tau>0 τ > 0 . For x = r ( cos θ , sin θ ) ∈ R + 2 x=r(\cos\theta,\sin\theta)\in\mathbb{R}^{2}_{+} x = r ( cos θ , sin θ ) ∈ R + 2 , put
U s , τ ( x ) = r s − 1 ϕ s , τ ( θ ) . U_{s,\tau}(x)=r^{s-1}\phi_{s,\tau}(\theta). U s , τ ( x ) = r s − 1 ϕ s , τ ( θ ) . This positive homogeneous function satisfies the weak equation
∬ R + 2 × R + 2 ( U s , τ ( x ) − U s , τ ( y ) ) ( η ( x ) ‾ − η ( y ) ‾ ) ∣ x − y ∣ 2 + 2 s d x d y = ∫ R + 2 ( D 2 , s , 2 x 2 − 2 s + Λ s , τ x 2 τ − 2 s ∣ x ∣ − τ ) U s , τ ( x ) η ( x ) ‾ d x . (76) \begin{aligned}
\iint_{\mathbb{R}^{2}_{+}\times\mathbb{R}^{2}_{+}}\frac{\left(U_{s,\tau}(x)-U_{s,\tau}(y)\right)\left(\overline{\eta(x)}-\overline{\eta(y)}\right)}{|x-y|^{2+2s}}\,dx\,dy
&=\int_{\mathbb{R}^{2}_{+}}\left(D_{2,s,2}x_{2}^{-2s}+\Lambda_{s,\tau}x_{2}^{\tau-2s}|x|^{-\tau}\right)U_{s,\tau}(x)\overline{\eta(x)}\,dx . \tag*{(76)}
\end{aligned} ∬ R + 2 × R + 2 ∣ x − y ∣ 2 + 2 s ( U s , τ ( x ) − U s , τ ( y ) ) ( η ( x ) − η ( y ) ) d x d y = ∫ R + 2 ( D 2 , s , 2 x 2 − 2 s + Λ s , τ x 2 τ − 2 s ∣ x ∣ − τ ) U s , τ ( x ) η ( x ) d x . (76) for η ∈ C c ∞ ( R + 2 ) \eta\in C^{\infty}_{c}(\mathbb{R}^{2}_{+}) η ∈ C c ∞ ( R + 2 ) . For every core function u u u ,
H s [ u ] − Λ s , τ R s , τ [ u ] = ∬ R + 2 × R + 2 U s , τ ( x ) U s , τ ( y ) ∣ x − y ∣ 2 + 2 s ∣ u ( x ) U s , τ ( x ) − u ( y ) U s , τ ( y ) ∣ 2 d x d y . (77) \mathcal{H}_{s}[u]-\Lambda_{s,\tau}\mathcal{R}_{s,\tau}[u]=\iint_{\mathbb{R}^{2}_{+}\times\mathbb{R}^{2}_{+}}\frac{U_{s,\tau}(x)U_{s,\tau}(y)}{|x-y|^{2+2s}}\left|\frac{u(x)}{U_{s,\tau}(x)}-\frac{u(y)}{U_{s,\tau}(y)}\right|^{2}\,dx\,dy . \tag*{(77)} H s [ u ] − Λ s , τ R s , τ [ u ] = ∬ R + 2 × R + 2 ∣ x − y ∣ 2 + 2 s U s , τ ( x ) U s , τ ( y ) U s , τ ( x ) u ( x ) − U s , τ ( y ) u ( y ) 2 d x d y . (77) The identity extends to the completed energy space, with the left-hand side interpreted as the closed sharp deficit. In cylinder coordinates it reads
q s [ w ] − Λ s , τ ∥ w ∥ L 2 ( R ; H s , τ ) 2 = ∫ R 2 ∫ 0 π ∫ 0 π K s ( t − t ′ , θ , φ ) ϕ s , τ ( θ ) ϕ s , τ ( φ ) × ∣ w ( t , θ ) ϕ s , τ ( θ ) − w ( t ′ , φ ) ϕ s , τ ( φ ) ∣ 2 d θ d φ d t d t ′ . (78) \begin{aligned}
q_s[w]-\Lambda_{s,\tau}\|w\|_{L^{2}(\mathbb{R};H_{s,\tau})}^{2}
&=\int_{\mathbb{R}^{2}}\int_{0}^{\pi}\int_{0}^{\pi}K_s(t-t',\theta,\varphi)\phi_{s,\tau}(\theta)\phi_{s,\tau}(\varphi)\\
&\qquad{}\times\left|\frac{w(t,\theta)}{\phi_{s,\tau}(\theta)}-\frac{w(t',\varphi)}{\phi_{s,\tau}(\varphi)}\right|^{2}\,d\theta\,d\varphi\,dt\,dt' . \tag*{(78)}
\end{aligned} q s [ w ] − Λ s , τ ∥ w ∥ L 2 ( R ; H s , τ ) 2 = ∫ R 2 ∫ 0 π ∫ 0 π K s ( t − t ′ , θ , φ ) ϕ s , τ ( θ ) ϕ s , τ ( φ ) × ϕ s , τ ( θ ) w ( t , θ ) − ϕ s , τ ( φ ) w ( t ′ , φ ) 2 d θ d φ d t d t ′ . (78) Proof. Write ϕ = ϕ s , τ \phi=\phi_{s,\tau} ϕ = ϕ s , τ , Λ = Λ s , τ \Lambda=\Lambda_{s,\tau} Λ = Λ s , τ , and a s = 1 − s a_s=1-s a s = 1 − s . All double integrals below are ordered integrals. Put
J s ( θ , φ ) = ∫ R K s ( q , θ , φ ) d q , J_s(\theta,\varphi)=\int_{\mathbb{R}}K_s(q,\theta,\varphi)\,dq, J s ( θ , φ ) = ∫ R K s ( q , θ , φ ) d q , W s ( θ ) = 2 ∫ 0 π ∫ R ( cosh ( a s q ) − 1 ) K s ( q , θ , φ ) d q d φ , W_s(\theta)=2\int_{0}^{\pi}\int_{\mathbb{R}}(\cosh(a_sq)-1)K_s(q,\theta,\varphi)\,dq\,d\varphi, W s ( θ ) = 2 ∫ 0 π ∫ R ( cosh ( a s q ) − 1 ) K s ( q , θ , φ ) d q d φ , V s ( θ ) = W s ( θ ) − D 2 , s , 2 ( sin θ ) − 2 s . V_s(\theta)=W_s(\theta)-D_{2,s,2}(\sin\theta)^{-2s}. V s ( θ ) = W s ( θ ) − D 2 , s , 2 ( sin θ ) − 2 s . For ∣ q ∣ < 1 |q|<1 ∣ q ∣ < 1 , integration in φ \varphi φ gives
∫ 0 π K s ( q , θ , φ ) d φ ≤ C s ∣ q ∣ − 1 − 2 s , ∣ cosh ( a s q ) − 1 ∣ ≤ C s q 2 . \int_{0}^{\pi}K_s(q,\theta,\varphi)\,d\varphi\le C_s|q|^{-1-2s},\qquad|\cosh(a_sq)-1|\le C_sq^2. ∫ 0 π K s ( q , θ , φ ) d φ ≤ C s ∣ q ∣ − 1 − 2 s , ∣ cosh ( a s q ) − 1∣ ≤ C s q 2 . For ∣ q ∣ ≥ 1 |q|\ge1 ∣ q ∣ ≥ 1 , the product is at most C s e − 2 s ∣ q ∣ C_se^{-2s|q|} C s e − 2 s ∣ q ∣ . Consequently,
0 ≤ W s ( θ ) ≤ C s ( ∫ 0 1 q 1 − 2 s d q + ∫ 1 ∞ e − 2 s q d q ) < ∞ 0\le W_s(\theta)\le C_s\left(\int_{0}^{1}q^{1-2s}\,dq+\int_{1}^{\infty}e^{-2sq}\,dq\right)<\infty 0 ≤ W s ( θ ) ≤ C s ( ∫ 0 1 q 1 − 2 s d q + ∫ 1 ∞ e − 2 s q d q ) < ∞ uniformly in θ \theta θ . In particular, V s − Λ m s , τ ∈ L ∞ ( I ) V_s-\Lambda m_{s,\tau}\in L^{\infty}(I) V s − Λ m s , τ ∈ L ∞ ( I ) for every I ⋐ ( 0 , π ) I\Subset(0,\pi) I ⋐ ( 0 , π ) .
For w ∈ C w\in\mathcal{C} w ∈ C , define
j s [ w ] = ∫ R 2 ∫ 0 π ∫ 0 π K s ( t − t ′ , θ , φ ) ∣ w ( t , θ ) − w ( t ′ , φ ) ∣ 2 d θ d φ d t d t ′ . j_s[w]=\int_{\mathbb{R}^{2}}\int_{0}^{\pi}\int_{0}^{\pi}K_s(t-t',\theta,\varphi)|w(t,\theta)-w(t',\varphi)|^2\,d\theta\,d\varphi\,dt\,dt'. j s [ w ] = ∫ R 2 ∫ 0 π ∫ 0 π K s ( t − t ′ , θ , φ ) ∣ w ( t , θ ) − w ( t ′ , φ ) ∣ 2 d θ d φ d t d t ′ . Near the diagonal, the smooth difference is bounded by C ( ∣ t − t ′ ∣ + ∣ θ − φ ∣ ) C(|t-t'|+|\theta-\varphi|) C ( ∣ t − t ′ ∣ + ∣ θ − φ ∣ ) . Its energy is therefore bounded by C ∫ 0 1 r 1 − 2 s d r C\int_{0}^{1}r^{1-2s}\,dr C ∫ 0 1 r 1 − 2 s d r . Off the diagonal, the compact support of w w w and the exponential kernel tail give an integrable majorant. Thus j s [ w ] < ∞ j_s[w]<\infty j s [ w ] < ∞ . To expand (36) , first use
K s δ ( q , θ , φ ) = K s ( q , θ , φ ) 1 { q 2 + ∣ θ − φ ∣ 2 > δ 2 } . (79) K_s^\delta(q,\theta,\varphi)=K_s(q,\theta,\varphi)\mathbf{1}_{\{q^2+|\theta-\varphi|^2>\delta^2\}}. \tag*{(79)} K s δ ( q , θ , φ ) = K s ( q , θ , φ ) 1 { q 2 + ∣ θ − φ ∣ 2 > δ 2 } . (79)
In the coefficient of ∣ w ( t , θ ) ∣ 2 |w(t,\theta)|^2 ∣ w ( t , θ ) ∣ 2 , symmetry in q q q changes e − a s q e^{-a_sq} e − a s q into cosh ( a s q ) \cosh(a_sq) cosh ( a s q ) . Interchanging ( t , θ ) (t,\theta) ( t , θ ) and ( t ′ , φ ) (t',\varphi) ( t ′ , φ ) in the other diagonal term gives
∫ K s δ ∣ e − a s ( t − t ′ ) / 2 w ( t , θ ) − e a s ( t − t ′ ) / 2 w ( t ′ , φ ) ∣ 2 = ∫ K s δ ∣ w ( t , θ ) − w ( t ′ , φ ) ∣ 2 + ∫ R ∫ 0 π W s δ ( θ ) ∣ w ( t , θ ) ∣ 2 d θ d t . \int K_s^\delta\left|e^{-a_s(t-t')/2}w(t,\theta)-e^{a_s(t-t')/2}w(t',\varphi)\right|^2
= \int K_s^\delta|w(t,\theta)-w(t',\varphi)|^2+\int_{\mathbb{R}}\int_0^\pi W_s^\delta(\theta)|w(t,\theta)|^2\,d\theta\,dt. ∫ K s δ e − a s ( t − t ′ ) /2 w ( t , θ ) − e a s ( t − t ′ ) /2 w ( t ′ , φ ) 2 = ∫ K s δ ∣ w ( t , θ ) − w ( t ′ , φ ) ∣ 2 + ∫ R ∫ 0 π W s δ ( θ ) ∣ w ( t , θ ) ∣ 2 d θ d t .
where W s δ W_s^\delta W s δ has the same cutoff. The preceding bounds permit δ ↓ 0 \delta\downarrow0 δ ↓ 0 and prove
q s [ w ] = j s [ w ] + ∫ R ∫ 0 π V s ( θ ) ∣ w ( t , θ ) ∣ 2 d θ d t . (80) q_s[w]=j_s[w]+\int_{\mathbb{R}}\int_0^\pi V_s(\theta)|w(t,\theta)|^2\,d\theta\,dt. \tag*{(80)} q s [ w ] = j s [ w ] + ∫ R ∫ 0 π V s ( θ ) ∣ w ( t , θ ) ∣ 2 d θ d t . (80)
Integrals without displayed limits in this proof are over the four cylinder variables, unless their variables indicate otherwise.
We next record the localized angular polarization:
b s , 0 ( f , η ) = ∫ 0 π ∫ 0 π J s ( θ , φ ) ( f ( θ ) − f ( φ ) ) ( η ( θ ) ‾ − η ( φ ) ‾ ) d θ d φ + ∫ 0 π V s f η ‾ d θ . (81) \mathfrak{b}_{s,0}(f,\eta)=\int_0^\pi\int_0^\pi J_s(\theta,\varphi)(f(\theta)-f(\varphi))(\overline{\eta(\theta)}-\overline{\eta(\varphi)})\,d\theta\,d\varphi+\int_0^\pi V_sf\overline{\eta}\,d\theta. \tag*{(81)} b s , 0 ( f , η ) = ∫ 0 π ∫ 0 π J s ( θ , φ ) ( f ( θ ) − f ( φ )) ( η ( θ ) − η ( φ ) ) d θ d φ + ∫ 0 π V s f η d θ . (81)
for f ∈ Y s f\in\mathcal{Y}_s f ∈ Y s and η ∈ H s ( 0 , π ) \eta\in H^s(0,\pi) η ∈ H s ( 0 , π ) with compact support. For core functions, expand (7) at ξ = 0 \xi=0 ξ = 0 with K s δ K_s^\delta K s δ and use the preceding symmetry. To pass to the stated domains, choose intervals supp η ⊂ I ⋐ I 1 ⋐ ( 0 , π ) \operatorname{supp}\eta\subset I\Subset I_1\Subset(0,\pi) supp η ⊂ I ⋐ I 1 ⋐ ( 0 , π ) . The local kernel bound gives
J s ( θ , φ ) ≤ C s ∣ θ − φ ∣ − 1 − 2 s . J_s(\theta,\varphi)\le C_s|\theta-\varphi|^{-1-2s}. J s ( θ , φ ) ≤ C s ∣ θ − φ ∣ − 1 − 2 s .
Hence the part with both variables in I 1 I_1 I 1 is bounded in absolute value by
C I 1 ∥ f ∥ H s ( I 1 ) ∥ η ∥ H s ( I 1 ) . C_{I_1}\|f\|_{H^s(I_1)}\|\eta\|_{H^s(I_1)}. C I 1 ∥ f ∥ H s ( I 1 ) ∥ η ∥ H s ( I 1 ) .
If one variable is outside I 1 I_1 I 1 and the integrand is nonzero, the other belongs to I I I ; the variables are separated. This part is bounded by
C I , I 1 ( ∥ f ∥ L 2 ( 0 , π ) + ∥ f ∥ L 2 ( I ) ) ∥ η ∥ L 2 ( I ) . C_{I,I_1}\left(\|f\|_{L^2(0,\pi)}+\|f\|_{L^2(I)}\right)\|\eta\|_{L^2(I)}. C I , I 1 ( ∥ f ∥ L 2 ( 0 , π ) + ∥ f ∥ L 2 ( I ) ) ∥ η ∥ L 2 ( I ) .
The potential term is bounded by ∥ V s ∥ L ∞ ( I ) ∥ f ∥ L 2 ( I ) ∥ η ∥ L 2 ( I ) \|V_s\|_{L^\infty(I)}\|f\|_{L^2(I)}\|\eta\|_{L^2(I)} ∥ V s ∥ L ∞ ( I ) ∥ f ∥ L 2 ( I ) ∥ η ∥ L 2 ( I ) . Use (56) and core approximation for f f f . For η \eta η , zero extension and mollification converge in H s H^s H s with support in I 1 I_1 I 1 ; the local upper estimate in Lemma 4.5 gives convergence in Y s \mathcal{Y}_s Y s . This proves (81) , including absolute convergence of its mixed integral. We do not split the angular form this way for two arbitrary elements of Y s \mathcal{Y}_s Y s .
Fix w ∈ C w\in\mathcal{C} w ∈ C and put
N ( θ ) = ∫ R ∣ w ( t , θ ) ∣ 2 d t , Φ ε = ϕ + ε , η ε = N Φ ε , ε > 0. N(\theta)=\int_{\mathbb{R}}|w(t,\theta)|^2\,dt,\qquad\Phi_\varepsilon=\phi+\varepsilon,\qquad\eta_\varepsilon=\frac{N}{\Phi_\varepsilon},\qquad\varepsilon>0. N ( θ ) = ∫ R ∣ w ( t , θ ) ∣ 2 d t , Φ ε = ϕ + ε , η ε = Φ ε N , ε > 0.
Here N ∈ C c ∞ ( 0 , π ) N\in C^\infty_c(0,\pi) N ∈ C c ∞ ( 0 , π ) . On a compact interval containing its support, ϕ ∈ H s \phi\in H^s ϕ ∈ H s and ϕ ≥ 0 \phi\ge0 ϕ ≥ 0 . For a , b ≥ 0 a,b\ge0 a , b ≥ 0 ,
∣ 1 a + ε − 1 b + ε ∣ ≤ ε − 2 ∣ a − b ∣ , 1 a + ε ≤ ε − 1 . \left|\frac{1}{a+\varepsilon}-\frac{1}{b+\varepsilon}\right|\le\varepsilon^{-2}|a-b|,\qquad\frac{1}{a+\varepsilon}\le\varepsilon^{-1}. a + ε 1 − b + ε 1 ≤ ε − 2 ∣ a − b ∣ , a + ε 1 ≤ ε − 1 .
Multiplication by N N N therefore gives
∥ η ε ∥ H s ( R ) ≤ C N ( ε − 1 + ε − 2 ∥ ϕ ∥ H s ( I 1 ) ) , \|\eta_\varepsilon\|_{H^s(\mathbb{R})}\le C_N\left(\varepsilon^{-1}+\varepsilon^{-2}\|\phi\|_{H^s(I_1)}\right), ∥ η ε ∥ H s ( R ) ≤ C N ( ε − 1 + ε − 2 ∥ ϕ ∥ H s ( I 1 ) ) ,
where η ε \eta_\varepsilon η ε is extended by zero. Thus η ε ∈ Y s \eta_\varepsilon\in\mathcal{Y}_s η ε ∈ Y s and is an admissible real test in (75) . By (81) ,
∫ 0 π ∫ 0 π J s ( θ , φ ) ( ϕ ( θ ) − ϕ ( φ ) ) ( η ε ( θ ) − η ε ( φ ) ) d θ d φ = − ∫ 0 π ( V s − Λ m s , τ ) ϕ N ϕ + ε d θ . \int_0^\pi\int_0^\pi J_s(\theta,\varphi)(\phi(\theta)-\phi(\varphi))(\eta_\varepsilon(\theta)-\eta_\varepsilon(\varphi))\,d\theta\,d\varphi
=-\int_0^\pi(V_s-\Lambda m_{s,\tau})\frac{\phi N}{\phi+\varepsilon}\,d\theta. ∫ 0 π ∫ 0 π J s ( θ , φ ) ( ϕ ( θ ) − ϕ ( φ )) ( η ε ( θ ) − η ε ( φ )) d θ d φ = − ∫ 0 π ( V s − Λ m s , τ ) ϕ + ε ϕN d θ .
For A , A ′ > 0 A,A'>0 A , A ′ > 0 and z , z ′ ∈ C z,z'\in\mathbb{C} z , z ′ ∈ C , direct expansion gives
∣ z − z ′ ∣ 2 − A A ′ ∣ z / A − z ′ / A ′ ∣ 2 = ( A − A ′ ) ( ∣ z ∣ 2 A − ∣ z ′ ∣ 2 A ′ ) . |z-z'|^2-AA'\left|z/A-z'/A'\right|^2=(A-A')\left(\frac{|z|^2}{A}-\frac{|z'|^2}{A'}\right). ∣ z − z ′ ∣ 2 − A A ′ ∣ z / A − z ′ / A ′ ∣ 2 = ( A − A ′ ) ( A ∣ z ∣ 2 − A ′ ∣ z ′ ∣ 2 ) .
Apply this identity with A = Φ ε ( θ ) A=\Phi_\varepsilon(\theta) A = Φ ε ( θ ) , A ′ = Φ ε ( φ ) A'=\Phi_\varepsilon(\varphi) A ′ = Φ ε ( φ ) and integrate against K s δ K_s^\delta K s δ . Integration in t , t ′ t,t' t , t ′ of the right-hand side gives exactly
∫ 0 π ∫ 0 π J s δ ( ϕ ( θ ) − ϕ ( φ ) ) ( η ε ( θ ) − η ε ( φ ) ) d θ d φ . \int_0^\pi\int_0^\pi\mathcal{J}_s^\delta(\phi(\theta)-\phi(\varphi))(\eta_\varepsilon(\theta)-\eta_\varepsilon(\varphi))\,d\theta\,d\varphi. ∫ 0 π ∫ 0 π J s δ ( ϕ ( θ ) − ϕ ( φ )) ( η ε ( θ ) − η ε ( φ )) d θ d φ .
Indeed, in the first summand put q = t − t ′ q=t-t' q = t − t ′ and integrate ∣ w ( t , θ ) ∣ 2 |w(t,\theta)|^2 ∣ w ( t , θ ) ∣ 2 in t t t ; in the second integrate ∣ w ( t ′ , φ ) ∣ 2 |w(t',\varphi)|^2 ∣ w ( t ′ , φ ) ∣ 2 in t ′ t' t ′ . The mixed angular integral converges absolutely by the estimates above. Monotone convergence applies to each of the two nonnegative cylinder energies. Letting δ ↓ 0 \delta\downarrow0 δ ↓ 0 yields
I ε : = ∫ K s Φ ε ( θ ) Φ ε ( φ ) ∣ w ( t , θ ) Φ ε ( θ ) − w ( t ′ , φ ) Φ ε ( φ ) ∣ 2 = j s [ w ] + ∫ 0 π ( V s − Λ m s , τ ) ϕ ϕ + ε N d θ . (82) \begin{aligned}
I_\varepsilon&:=\int K_s\Phi_\varepsilon(\theta)\Phi_\varepsilon(\varphi)\left|\frac{w(t,\theta)}{\Phi_\varepsilon(\theta)}-\frac{w(t',\varphi)}{\Phi_\varepsilon(\varphi)}\right|^2\\
&=j_s[w]+\int_0^\pi(V_s-\Lambda m_{s,\tau})\frac{\phi}{\phi+\varepsilon}N\,d\theta. \tag*{(82)}
\end{aligned} I ε := ∫ K s Φ ε ( θ ) Φ ε ( φ ) Φ ε ( θ ) w ( t , θ ) − Φ ε ( φ ) w ( t ′ , φ ) 2 = j s [ w ] + ∫ 0 π ( V s − Λ m s , τ ) ϕ + ε ϕ N d θ . (82)
Since ϕ > 0 \phi>0 ϕ > 0 almost everywhere and the potential is bounded on supp N \operatorname{supp}N supp N , dominated convergence on the right and (5.5) give
lim ε ↓ 0 I ε = j s [ w ] + ∫ 0 π ( V s − Λ m s , τ ) N d θ = q s [ w ] − Λ ∥ w ∥ L 2 ( m s , τ ) 2 . \lim_{\varepsilon\downarrow0}I_\varepsilon=j_s[w]+\int_0^\pi(V_s-\Lambda m_{s,\tau})N\,d\theta=q_s[w]-\Lambda\|w\|_{L^2(m_{s,\tau})}^2. ε ↓ 0 lim I ε = j s [ w ] + ∫ 0 π ( V s − Λ m s , τ ) N d θ = q s [ w ] − Λ∥ w ∥ L 2 ( m s , τ ) 2 .
Denote the integral on the right of (78) by I 0 I_0 I 0 . Fatou’s lemma gives I 0 ≤ lim inf ε ↓ 0 I ε < ∞ I_0\leq\liminf_{\varepsilon\downarrow0}I_\varepsilon<\infty I 0 ≤ lim inf ε ↓ 0 I ε < ∞ . To prove convergence to I 0 I_0 I 0 , put r = A ′ / A r=A'/A r = A ′ / A and r ε = ( A ′ + ε ) / ( A + ε ) r_\varepsilon=(A'+\varepsilon)/(A+\varepsilon) r ε = ( A ′ + ε ) / ( A + ε ) . There is γ ε ∈ [ 0 , 1 ] \gamma_\varepsilon\in[0,1] γ ε ∈ [ 0 , 1 ] such that log r ε = γ ε log r \log r_\varepsilon=\gamma_\varepsilon\log r log r ε = γ ε log r . The function
F ( t ) = e t ∣ z ∣ 2 + e − t ∣ z ′ ∣ 2 − 2 Re ( z z ′ ‾ ) F(t)=e^t|z|^2+e^{-t}|z'|^2-2\operatorname{Re}(z\overline{z'}) F ( t ) = e t ∣ z ∣ 2 + e − t ∣ z ′ ∣ 2 − 2 Re ( z z ′ )
is nonnegative and convex. Hence
( A + ε ) ( A ′ + ε ) ∣ z / ( A + ε ) − z ′ / ( A ′ + ε ) ∣ 2 = F ( log r ε ) ≤ γ ε F ( log r ) + ( 1 − γ ε ) F ( 0 ) ≤ A A ′ ∣ z / A − z ′ / A ′ ∣ 2 + ∣ z − z ′ ∣ 2 . \begin{aligned}
(A+\varepsilon)(A'+\varepsilon)\left|z/(A+\varepsilon)-z'/(A'+\varepsilon)\right|^2
&=F(\log r_\varepsilon)\\
&\leq\gamma_\varepsilon F(\log r)+(1-\gamma_\varepsilon)F(0)\\
&\leq AA'\left|z/A-z'/A'\right|^2+|z-z'|^2.
\end{aligned} ( A + ε ) ( A ′ + ε ) ∣ z / ( A + ε ) − z ′ / ( A ′ + ε ) ∣ 2 = F ( log r ε ) ≤ γ ε F ( log r ) + ( 1 − γ ε ) F ( 0 ) ≤ A A ′ ∣ z / A − z ′ / A ′ ∣ 2 + ∣ z − z ′ ∣ 2 .
The majorant is integrable against K s K_s K s , since its integral is I 0 + j s [ w ] I_0+j_s[w] I 0 + j s [ w ] . Dominated convergence gives I ε → I 0 I_\varepsilon\to I_0 I ε → I 0 in (82) and proves (78) .
We now verify the weak equation. For v ∈ C v\in\mathcal{C} v ∈ C , set η 0 ( θ ) = ∫ R v ( t , θ ) d t \eta_0(\theta)=\int_{\mathbb{R}}v(t,\theta)\,dt η 0 ( θ ) = ∫ R v ( t , θ ) d t . Using a common cutoff in the stationary-cylinder pairing, Fubini gives
∫ K s ( ϕ ( θ ) − ϕ ( φ ) ) ( v ( t , θ ) ‾ − v ( t ′ , φ ) ‾ ) + ∫ R ∫ 0 π V s ϕ ‾ v d θ d t = b s , 0 ( ϕ , η 0 ) = Λ ∫ R ∫ 0 π m s , τ ϕ ‾ v d θ d t . \int K_s(\phi(\theta)-\phi(\varphi))(\overline{v(t,\theta)}-\overline{v(t',\varphi)})+\int_{\mathbb{R}}\int_0^\pi V_s\overline{\phi}v\,d\theta\,dt
=b_{s,0}(\phi,\eta_0)=\Lambda\int_{\mathbb{R}}\int_0^\pi m_{s,\tau}\overline{\phi}v\,d\theta\,dt. ∫ K s ( ϕ ( θ ) − ϕ ( φ )) ( v ( t , θ ) − v ( t ′ , φ ) ) + ∫ R ∫ 0 π V s ϕ v d θ d t = b s , 0 ( ϕ , η 0 ) = Λ ∫ R ∫ 0 π m s , τ ϕ v d θ d t .
Absolute convergence follows locally from the H s H^s H s bound and Cauchy–Schwarz; on separated angular pairs it follows from ϕ ∈ L 2 ( 0 , π ) \phi\in L^2(0,\pi) ϕ ∈ L 2 ( 0 , π ) . The same bounds justify removal of the cutoff. For a physical test η ∈ C c ∞ ( R + 2 ) \eta\in C^\infty_c(\mathbb{R}^2_+) η ∈ C c ∞ ( R + 2 ) , take
v ( t , θ ) = e ( 1 − s ) t η ( e t cos θ , e t sin θ ) . v(t,\theta)=e^{(1-s)t}\eta(e^t\cos\theta,e^t\sin\theta). v ( t , θ ) = e ( 1 − s ) t η ( e t cos θ , e t sin θ ) .
The stationary profile corresponds to U ( x ) = r s − 1 ϕ ( θ ) U(x)=r^{s-1}\phi(\theta) U ( x ) = r s − 1 ϕ ( θ ) . Polarization of the calculation proving (35) , with the same cutoff, identifies the preceding pairing with the physical energy pairing minus the Hardy term. Also,
∫ R + 2 x 2 − 2 s U η ‾ d x = ∫ R ∫ 0 π ( sin θ ) − 2 s φ ‾ v d θ d t , \int_{\mathbb{R}^2_+} x_2^{-2s}U\overline{\eta}\,dx=\int_{\mathbb{R}}\int_0^\pi(\sin\theta)^{-2s}\overline{\varphi}v\,d\theta\,dt, ∫ R + 2 x 2 − 2 s U η d x = ∫ R ∫ 0 π ( sin θ ) − 2 s φ v d θ d t ,
∫ R + 2 x 2 τ − 2 s ∣ x ∣ − τ U η ‾ d x = ∫ R ∫ 0 π m s , τ φ ‾ v d θ d t . \int_{\mathbb{R}^2_+}x_2^{\tau-2s}|x|^{-\tau}U\overline{\eta}\,dx=\int_{\mathbb{R}}\int_0^\pi m_{s,\tau}\overline{\varphi}v\,d\theta\,dt. ∫ R + 2 x 2 τ − 2 s ∣ x ∣ − τ U η d x = ∫ R ∫ 0 π m s , τ φ v d θ d t .
For completeness, the physical mixed energy is absolutely convergent. On a compact neighborhood of supp η \operatorname{supp}\eta supp η , the smooth polar change of variables and φ ∈ H l o c s \varphi\in H^s_{\mathrm{loc}} φ ∈ H loc s give U ∈ H s U\in H^s U ∈ H s . Cauchy–Schwarz bounds the local mixed integral. If x x x is in the test support and ∣ y ∣ |y| ∣ y ∣ is small or large, its remaining part is bounded by
C η [ 1 + ∥ φ ∥ L 1 ( 0 , π ) ( ∫ 0 1 r s d r + ∫ 1 ∞ r − 2 − s d r ) ] < ∞ . C_\eta\left[1+\|\varphi\|_{L^1(0,\pi)}\left(\int_0^1r^s\,dr+\int_1^\infty r^{-2-s}\,dr\right)\right]<\infty. C η [ 1 + ∥ φ ∥ L 1 ( 0 , π ) ( ∫ 0 1 r s d r + ∫ 1 ∞ r − 2 − s d r ) ] < ∞.
The separated bounded region is controlled by U ∈ L l o c 1 U\in L^1_{\mathrm{loc}} U ∈ L loc 1 . Thus the cutoff may also be removed in physical coordinates. This proves (76) .
Finally, u / U = w / φ u/U=w/\varphi u / U = w / φ , and (32) gives
U ( x ) U ( y ) ∣ x − y ∣ 2 + 2 s d x d y = K s ( t − t ′ , θ , φ ) φ ( θ ) φ ( φ ) d t d t ′ d θ d φ . \frac{U(x)U(y)}{|x-y|^{2+2s}}\,dx\,dy=K_s(t-t',\theta,\varphi)\varphi(\theta)\varphi(\varphi)\,dt\,dt'\,d\theta\,d\varphi. ∣ x − y ∣ 2 + 2 s U ( x ) U ( y ) d x d y = K s ( t − t ′ , θ , φ ) φ ( θ ) φ ( φ ) d t d t ′ d θ d φ .
Substitution in (78) proves (77) on the core. To pass to the completion, let w n → w w_n\to w w n → w in X s \mathcal{X}_s X s and define
( D w ) ( t , t ′ , θ , φ ) = φ ( θ ) φ ( φ ) ( w ( t , θ ) φ ( θ ) − w ( t ′ , φ ) φ ( φ ) ) . (Dw)(t,t',\theta,\varphi)=\sqrt{\varphi(\theta)\varphi(\varphi)}\left(\frac{w(t,\theta)}{\varphi(\theta)}-\frac{w(t',\varphi)}{\varphi(\varphi)}\right). ( D w ) ( t , t ′ , θ , φ ) = φ ( θ ) φ ( φ ) ( φ ( θ ) w ( t , θ ) − φ ( φ ) w ( t ′ , φ ) ) .
Use the kernel measure K s d t d t ′ d θ d φ K_s\,dt\,dt'\,d\theta\,d\varphi K s d t d t ′ d θ d φ . The core identity gives
∥ D ( w n − w m ) ∥ L 2 ( K s ) 2 = q s [ w n − w m ] − Λ ∥ w n − w m ∥ L 2 ( m s , τ ) 2 ≤ q s [ w n − w m ] ⟶ 0. \|D(w_n-w_m)\|_{L^2(K_s)}^2=q_s[w_n-w_m]-\Lambda\|w_n-w_m\|_{L^2(m_{s,\tau})}^2\le q_s[w_n-w_m]\longrightarrow0. ∥ D ( w n − w m ) ∥ L 2 ( K s ) 2 = q s [ w n − w m ] − Λ∥ w n − w m ∥ L 2 ( m s , τ ) 2 ≤ q s [ w n − w m ] ⟶ 0.
Thus D w n → F Dw_n\to F D w n → F in the kernel space. A subsequence converges almost everywhere both in the base space and in the kernel space. Since φ > 0 \varphi>0 φ > 0 almost everywhere and the kernel measure is absolutely continuous off the null diagonal, its pointwise limit is D w Dw D w . Therefore F = D w F=Dw F = D w . Taking norms and using continuity of the closed form,
∥ D w ∥ L 2 ( K s ) 2 = lim n → ∞ ∥ D w n ∥ L 2 ( K s ) 2 = q s [ w ] − Λ ∥ w ∥ L 2 ( m s , τ ) 2 . \|Dw\|_{L^2(K_s)}^2=\lim_{n\to\infty}\|Dw_n\|_{L^2(K_s)}^2=q_s[w]-\Lambda\|w\|_{L^2(m_{s,\tau})}^2. ∥ D w ∥ L 2 ( K s ) 2 = n → ∞ lim ∥ D w n ∥ L 2 ( K s ) 2 = q s [ w ] − Λ∥ w ∥ L 2 ( m s , τ ) 2 .
This proves both completed-space identities.
The next proposition identifies the angular stability constant with the spectral gap. The coefficient in (9) also depends on variation in the logarithmic radial variable.
Proposition 5.4 (Attainment of the angular stability constant). Define
Γ s , τ a n g = inf ψ ∈ Y s ∖ { 0 } ∫ 0 π m s , τ ψ φ s , τ d θ = 0 b s , 0 [ ψ ] − Λ s , τ ∥ ψ ∥ L 2 ( m s , τ ) 2 ∥ ψ ∥ L 2 ( m s , τ ) 2 . \Gamma^{\mathrm{ang}}_{s,\tau}=\inf_{\substack{\psi\in Y_s\setminus\{0\}\\ \int_0^\pi m_{s,\tau}\psi\varphi_{s,\tau}\,d\theta=0}}\frac{b_{s,0}[\psi]-\Lambda_{s,\tau}\|\psi\|_{L^2(m_{s,\tau})}^2}{\|\psi\|_{L^2(m_{s,\tau})}^2}. Γ s , τ ang = ψ ∈ Y s ∖ { 0 } ∫ 0 π m s , τ ψ φ s , τ d θ = 0 inf ∥ ψ ∥ L 2 ( m s , τ ) 2 b s , 0 [ ψ ] − Λ s , τ ∥ ψ ∥ L 2 ( m s , τ ) 2 . Then
Γ s , τ a n g = λ 2 ( 0 ) − λ 1 ( 0 ) > 0. \Gamma^{\mathrm{ang}}_{s,\tau}=\lambda_2(0)-\lambda_1(0)>0. Γ s , τ ang = λ 2 ( 0 ) − λ 1 ( 0 ) > 0. The infimum is attained. Equality in
b s , 0 [ ψ ] − Λ s , τ ∥ ψ ∥ H s , τ 2 ≥ Γ s , τ a n g dist H s , τ ( ψ , span { φ s , τ } ) 2 (83) b_{s,0}[\psi]-\Lambda_{s,\tau}\|\psi\|_{H_{s,\tau}}^2\ge\Gamma^{\mathrm{ang}}_{s,\tau}\operatorname{dist}_{H_{s,\tau}}\left(\psi,\operatorname{span}\{\varphi_{s,\tau}\}\right)^2 \tag*{(83)} b s , 0 [ ψ ] − Λ s , τ ∥ ψ ∥ H s , τ 2 ≥ Γ s , τ ang dist H s , τ ( ψ , span { φ s , τ } ) 2 (83) holds precisely when the orthogonal component of ψ \psi ψ belongs to the second eigenspace of B s ( 0 ) B_s(0) B s ( 0 ) .
Proof. Write
ψ = c ϕ s , τ + ψ ⊥ , ∫ 0 π m s , τ ψ ⊥ ϕ s , τ d θ = 0. \psi=c\phi_{s,\tau}+\psi^\perp,\qquad\int_0^\pi m_{s,\tau}\psi^\perp\phi_{s,\tau}\,d\theta=0. ψ = c ϕ s , τ + ψ ⊥ , ∫ 0 π m s , τ ψ ⊥ ϕ s , τ d θ = 0.
By (75) , the left-hand side of (83) equals
b s , 0 [ ψ ⊥ ] − λ 1 ( 0 ) ∥ ψ ⊥ ∥ H s , τ 2 . \mathfrak{b}_{s,0}[\psi^\perp]-\lambda_1(0)\|\psi^\perp\|_{H_{s,\tau}}^2. b s , 0 [ ψ ⊥ ] − λ 1 ( 0 ) ∥ ψ ⊥ ∥ H s , τ 2 .
The min–max principle and (51) give the attained value λ 2 ( 0 ) − λ 1 ( 0 ) \lambda_2(0)-\lambda_1(0) λ 2 ( 0 ) − λ 1 ( 0 ) . It is positive by Lemma 4.6. Expanding ψ ⊥ = ∑ j ≥ 2 c j e j \psi^\perp=\sum_{j\geq2}c_je_j ψ ⊥ = ∑ j ≥ 2 c j e j in an orthonormal eigenbasis, we have
b s , 0 [ ψ ⊥ ] − λ 1 ( 0 ) ∥ ψ ⊥ ∥ 2 = ∑ j ≥ 2 ( λ j ( 0 ) − λ 1 ( 0 ) ) ∣ c j ∣ 2 . \mathfrak{b}_{s,0}[\psi^\perp]-\lambda_1(0)\|\psi^\perp\|^2=\sum_{j\geq2}(\lambda_j(0)-\lambda_1(0))|c_j|^2. b s , 0 [ ψ ⊥ ] − λ 1 ( 0 ) ∥ ψ ⊥ ∥ 2 = j ≥ 2 ∑ ( λ j ( 0 ) − λ 1 ( 0 )) ∣ c j ∣ 2 .
Equality holds if and only if c j = 0 c_j=0 c j = 0 whenever λ j ( 0 ) > λ 2 ( 0 ) \lambda_j(0)>\lambda_2(0) λ j ( 0 ) > λ 2 ( 0 ) .
Spectral stability and energy decomposition
Proposition 6.1 (Dispersion at the threshold). The first eigenvalue is even and real analytic for ∣ ξ ∣ |\xi| ∣ ξ ∣ small. It satisfies
λ 1 ′ ( 0 ) = 0 , 1 2 λ 1 ′ ′ ( 0 ) = κ s , τ , \lambda_1'(0)=0,\qquad\frac{1}{2}\lambda_1''(0)=\kappa_{s,\tau}, λ 1 ′ ( 0 ) = 0 , 2 1 λ 1 ′′ ( 0 ) = κ s , τ , where κ s , τ \kappa_{s,\tau} κ s , τ is given by (11) .
Proof. By Lemma 4.10, the forms b s , z \mathfrak{b}_{s,z} b s , z are a holomorphic family of type (a) on the common domain Y s \mathcal{Y}_s Y s . The associated operators form a family of type (B) [13 Chapter VII, Section 4] . Since λ 1 ( 0 ) \lambda_1(0) λ 1 ( 0 ) is simple and isolated, a Riesz projection of rank one gives analytic branches λ 1 ( z ) \lambda_1(z) λ 1 ( z ) and ϕ ( z ) \phi(z) ϕ ( z ) near zero. For real ξ \xi ξ , normalize by
∥ ϕ ( ξ ) ∥ H s , τ = 1 , ⟨ ϕ ( ξ ) , ϕ s , τ ⟩ H s , τ > 0. \|\phi(\xi)\|_{H_{s,\tau}}=1,\qquad\langle\phi(\xi),\phi_{s,\tau}\rangle_{H_{s,\tau}}>0. ∥ ϕ ( ξ ) ∥ H s , τ = 1 , ⟨ ϕ ( ξ ) , ϕ s , τ ⟩ H s , τ > 0. The real form depends on ξ \xi ξ through cos ( ξ q ) \cos(\xi q) cos ( ξ q ) . Uniqueness of the simple branch gives λ 1 ( − ξ ) = λ 1 ( ξ ) \lambda_1(-\xi)=\lambda_1(\xi) λ 1 ( − ξ ) = λ 1 ( ξ ) and ϕ ( − ξ ) = ϕ ( ξ ) \phi(-\xi)=\phi(\xi) ϕ ( − ξ ) = ϕ ( ξ ) . Equation (70) gives b s , 0 ′ = 0 \mathfrak{b}'_{s,0}=0 b s , 0 ′ = 0 in L ( Y s , Y s ∗ ) \mathcal{L}(\mathcal{Y}_s,\mathcal{Y}_s^*) L ( Y s , Y s ∗ ) and b s , 0 ′ ′ = 2 c s , 2 \mathfrak{b}''_{s,0}=2c_{s,2} b s , 0 ′′ = 2 c s , 2 . The form version of the Feynman–Hellmann formula therefore yields
λ 1 ′ ( 0 ) = 0 , λ 1 ′ ′ ( 0 ) = b s , 0 ′ ′ ( ϕ s , τ , ϕ s , τ ) . \lambda_1'(0)=0,\qquad\lambda_1''(0)=\mathfrak{b}''_{s,0}(\phi_{s,\tau},\phi_{s,\tau}). λ 1 ′ ( 0 ) = 0 , λ 1 ′′ ( 0 ) = b s , 0 ′′ ( ϕ s , τ , ϕ s , τ ) . Here the derivative of the eigenvector can be checked directly. Differentiating the weak eigenvalue equation at zero gives
( B s ( 0 ) − Λ s , τ ) ϕ ′ ( 0 ) = 0 in Y s ∗ . (B_s(0)-\Lambda_{s,\tau})\phi'(0)=0\quad\text{in }\mathcal{Y}_s^*. ( B s ( 0 ) − Λ s , τ ) ϕ ′ ( 0 ) = 0 in Y s ∗ . The real normalization and phase condition give ⟨ ϕ ′ ( 0 ) , ϕ s , τ ⟩ H s , τ = 0 \langle\phi'(0),\phi_{s,\tau}\rangle_{H_{s,\tau}}=0 ⟨ ϕ ′ ( 0 ) , ϕ s , τ ⟩ H s , τ = 0 . Simplicity implies ϕ ′ ( 0 ) = 0 \phi'(0)=0 ϕ ′ ( 0 ) = 0 . Differentiating a second time and pairing with ϕ s , τ \phi_{s,\tau} ϕ s , τ cancels the terms containing ϕ ′ ′ ( 0 ) \phi''(0) ϕ ′′ ( 0 ) by (75) . Thus only b s , 0 ′ ′ ( ϕ s , τ , ϕ s , τ ) \mathfrak{b}''_{s,0}(\phi_{s,\tau},\phi_{s,\tau}) b s , 0 ′′ ( ϕ s , τ , ϕ s , τ ) remains. By (71) ,
λ 1 ′ ′ ( 0 ) = 2 ∫ 0 π ∫ 0 π ∫ R q 2 K s ( q , θ , ϕ ) ϕ s , τ ( θ ) ϕ s , τ ( ϕ ) d q d θ d ϕ . \lambda_1''(0)=2\int_0^\pi\int_0^\pi\int_{\mathbb{R}}q^2K_s(q,\theta,\phi)\phi_{s,\tau}(\theta)\phi_{s,\tau}(\phi)\,dq\,d\theta\,d\phi. λ 1 ′′ ( 0 ) = 2 ∫ 0 π ∫ 0 π ∫ R q 2 K s ( q , θ , ϕ ) ϕ s , τ ( θ ) ϕ s , τ ( ϕ ) d q d θ d ϕ . This integral is finite by Lemma 4.10 and positive because K s > 0 K_s>0 K s > 0 and ϕ s , τ > 0 \phi_{s,\tau}>0 ϕ s , τ > 0 almost everywhere. Even analyticity gives the O ( ξ 4 ) O(\xi^4) O ( ξ 4 ) remainder.
Lemma 6.2 (Axial dispersion function). The function ρ s , τ \rho_{s,\tau} ρ s , τ in (15) is even and real analytic. It satisfies
ρ s , τ ( ξ ) = κ s , τ ξ 2 − ξ 4 12 ∫ 0 π ∫ 0 π ∫ R q 4 K s ( q , θ , ϕ ) ϕ s , τ ( θ ) ϕ s , τ ( ϕ ) d q d θ d ϕ + O ( ξ 6 ) (84) \rho_{s,\tau}(\xi)=\kappa_{s,\tau}\xi^2-\frac{\xi^4}{12}\int_0^\pi\int_0^\pi\int_{\mathbb{R}}q^4K_s(q,\theta,\phi)\phi_{s,\tau}(\theta)\phi_{s,\tau}(\phi)\,dq\,d\theta\,d\phi+O(\xi^6) \tag*{(84)} ρ s , τ ( ξ ) = κ s , τ ξ 2 − 12 ξ 4 ∫ 0 π ∫ 0 π ∫ R q 4 K s ( q , θ , ϕ ) ϕ s , τ ( θ ) ϕ s , τ ( ϕ ) d q d θ d ϕ + O ( ξ 6 ) (84) at zero. Moreover,
ρ s , τ ( ξ ) > 0 ( ξ ≠ 0 ) , c ∣ ξ ∣ 2 s − C ≤ ρ s , τ ( ξ ) ≤ C ( 1 + ∣ ξ ∣ 2 s ) . (85) \rho_{s,\tau}(\xi)>0\quad(\xi\ne0),\qquad c|\xi|^{2s}-C\leq\rho_{s,\tau}(\xi)\leq C(1+|\xi|^{2s}). \tag*{(85)} ρ s , τ ( ξ ) > 0 ( ξ = 0 ) , c ∣ ξ ∣ 2 s − C ≤ ρ s , τ ( ξ ) ≤ C ( 1 + ∣ ξ ∣ 2 s ) . (85)
Proof. First use truncated kernels in (15) . The integrand is nonnegative and, for ξ ≠ 0 \xi\ne0 ξ = 0 , positive for almost every ( q , θ , ϕ ) (q,\theta,\phi) ( q , θ , ϕ ) . This proves strict positivity. By Lemma 4.10 , Taylor expansion under the integral is valid. Since
2 ( 1 − cos ( ξ q ) ) = ξ 2 q 2 − ξ 4 q 4 12 + O ( ξ 6 q 6 ) , 2(1-\cos(\xi q))=\xi^2q^2-\frac{\xi^4q^4}{12}+O(\xi^6q^6), 2 ( 1 − cos ( ξ q )) = ξ 2 q 2 − 12 ξ 4 q 4 + O ( ξ 6 q 6 ) ,
we obtain (84) and real analyticity. Finally,
ρ s , τ ( ξ ) = b s , ξ [ ϕ s , τ ] − b s , 0 [ ϕ s , τ ] . \rho_{s,\tau}(\xi)=b_{s,\xi}[\phi_{s,\tau}]-b_{s,0}[\phi_{s,\tau}]. ρ s , τ ( ξ ) = b s , ξ [ ϕ s , τ ] − b s , 0 [ ϕ s , τ ] .
The two estimates in (85) follow from (74) and b s , 0 [ ϕ s , τ ] = Λ s , τ b_{s,0}[\phi_{s,\tau}]=\Lambda_{s,\tau} b s , 0 [ ϕ s , τ ] = Λ s , τ .
Let P ( ξ ) P(\xi) P ( ξ ) be the first spectral projection for ∣ ξ ∣ |\xi| ∣ ξ ∣ small.
Lemma 6.3 (Uniform threshold estimate). There are ξ 0 , c 0 , c 1 > 0 \xi_0,c_0,c_1>0 ξ 0 , c 0 , c 1 > 0 such that
b s , ξ [ ψ ] − Λ s , τ ∥ ψ ∥ L 2 ( m s , τ ) 2 ≥ c 0 ξ 2 ∥ P ( ξ ) ψ ∥ L 2 ( m s , τ ) 2 + c 1 ∥ ( I − P ( ξ ) ) ψ ∥ L 2 ( m s , τ ) 2 (86) b_{s,\xi}[\psi]-\Lambda_{s,\tau}\|\psi\|_{L^2(m_{s,\tau})}^2\ge c_0\xi^2\|P(\xi)\psi\|_{L^2(m_{s,\tau})}^2+c_1\|(I-P(\xi))\psi\|_{L^2(m_{s,\tau})}^2 \tag*{(86)} b s , ξ [ ψ ] − Λ s , τ ∥ ψ ∥ L 2 ( m s , τ ) 2 ≥ c 0 ξ 2 ∥ P ( ξ ) ψ ∥ L 2 ( m s , τ ) 2 + c 1 ∥ ( I − P ( ξ )) ψ ∥ L 2 ( m s , τ ) 2 (86) for ∣ ξ ∣ ≤ ξ 0 |\xi|\le\xi_0 ∣ ξ ∣ ≤ ξ 0 . Moreover,
∥ P ( ξ ) − P 0 ∥ H s , τ → H s , τ + ∥ P ( ξ ) − P 0 ∥ H s , τ → Y s ≤ C ξ 2 (87) \|P(\xi)-P_0\|_{H_{s,\tau}\to H_{s,\tau}}+\|P(\xi)-P_0\|_{H_{s,\tau}\to\mathcal{Y}_s}\le C\xi^2 \tag*{(87)} ∥ P ( ξ ) − P 0 ∥ H s , τ → H s , τ + ∥ P ( ξ ) − P 0 ∥ H s , τ → Y s ≤ C ξ 2 (87)
Proof. Put g 0 = λ 2 ( 0 ) − λ 1 ( 0 ) > 0 g_0=\lambda_2(0)-\lambda_1(0)>0 g 0 = λ 2 ( 0 ) − λ 1 ( 0 ) > 0 . By continuity and Proposition 6.1 , choose ξ 0 > 0 \xi_0>0 ξ 0 > 0 such that
λ 1 ( ξ ) − Λ s , τ ≥ 1 2 κ s , τ ξ 2 , λ 2 ( ξ ) − Λ s , τ ≥ 1 2 g 0 . \lambda_1(\xi)-\Lambda_{s,\tau}\ge\frac{1}{2}\kappa_{s,\tau}\xi^2,\qquad\lambda_2(\xi)-\Lambda_{s,\tau}\ge\frac{1}{2}g_0. λ 1 ( ξ ) − Λ s , τ ≥ 2 1 κ s , τ ξ 2 , λ 2 ( ξ ) − Λ s , τ ≥ 2 1 g 0 .
The spectral theorem gives (86) . The Riesz formula
P ( ξ ) = − 1 2 π i ∫ Γ ( B s ( ξ ) − z ) − 1 d z P(\xi)=-\frac{1}{2\pi i}\int_\Gamma(B_s(\xi)-z)^{-1}\,dz P ( ξ ) = − 2 π i 1 ∫ Γ ( B s ( ξ ) − z ) − 1 d z
around the first eigenvalue gives P ′ ( 0 ) = 0 P'(0)=0 P ′ ( 0 ) = 0 , since the family is even. To estimate this integral in the form norm, define A ξ ( z ) : Y s → Y s ∗ A_\xi(z):\mathcal{Y}_s\to\mathcal{Y}_s^* A ξ ( z ) : Y s → Y s ∗ by
⟨ A ξ ( z ) u , v ⟩ = b s , ξ ( u , v ) − z ⟨ u , v ⟩ H s , τ . \langle A_\xi(z)u,v\rangle=b_{s,\xi}(u,v)-z\langle u,v\rangle_{H_{s,\tau}}. ⟨ A ξ ( z ) u , v ⟩ = b s , ξ ( u , v ) − z ⟨ u , v ⟩ H s , τ .
For c > 0 c>0 c > 0 , the map A 0 ( − c ) A_0(-c) A 0 ( − c ) is a coercive isomorphism. The map J J J induced by the H s , τ H_{s,\tau} H s , τ inner product is compact by (51) , and
A 0 ( z ) = A 0 ( − c ) { I − ( z + c ) A 0 ( − c ) − 1 J } . A_0(z)=A_0(-c)\{I-(z+c)A_0(-c)^{-1}J\}. A 0 ( z ) = A 0 ( − c ) { I − ( z + c ) A 0 ( − c ) − 1 J } .
Thus A 0 ( z ) A_0(z) A 0 ( z ) is Fredholm of index zero. For z ∈ Γ z\in\Gamma z ∈ Γ , its kernel is zero, so it is invertible with a bound uniform on Γ \Gamma Γ . Equation (70) and a Neumann series give the same bound for R ξ ( z ) = A ξ ( z ) − 1 R_\xi(z)=A_\xi(z)^{-1} R ξ ( z ) = A ξ ( z ) − 1 when ∣ ξ ∣ |\xi| ∣ ξ ∣ is small. The restriction of R ξ ( z ) : Y s ∗ → Y s R_\xi(z):\mathcal{Y}_s^*\to\mathcal{Y}_s R ξ ( z ) : Y s ∗ → Y s to H s , τ H_{s,\tau} H s , τ is the operator resolvent. Moreover,
∥ A ξ ( z ) − A 0 ( z ) ∥ Y s → Y s ∗ ≤ C ξ 2 . \|A_\xi(z)-A_0(z)\|_{\mathcal{Y}_s\to\mathcal{Y}_s^*}\le C\xi^2. ∥ A ξ ( z ) − A 0 ( z ) ∥ Y s → Y s ∗ ≤ C ξ 2 .
The resolvent identity in these spaces is
R ξ ( z ) − R 0 ( z ) = − R ξ ( z ) ( A ξ ( z ) − A 0 ( z ) ) R 0 ( z ) . R_\xi(z)-R_0(z)=-R_\xi(z)(A_\xi(z)-A_0(z))R_0(z). R ξ ( z ) − R 0 ( z ) = − R ξ ( z ) ( A ξ ( z ) − A 0 ( z )) R 0 ( z ) .
The right-hand side is O ( ξ 2 ) O(\xi^2) O ( ξ 2 ) from H s , τ H_{s,\tau} H s , τ to Y s \mathcal{Y}_s Y s , uniformly on Γ \Gamma Γ . Integration proves (87) . For real ξ \xi ξ near zero, take
ϕ ( ξ ) = P ( ξ ) ϕ s , τ ∥ P ( ξ ) ϕ s , τ ∥ H s , τ . \phi(\xi)=\frac{P(\xi)\phi_{s,\tau}}{\|P(\xi)\phi_{s,\tau}\|_{H_{s,\tau}}}. ϕ ( ξ ) = ∥ P ( ξ ) ϕ s , τ ∥ H s , τ P ( ξ ) ϕ s , τ .
Consequently, ∥ ϕ ( ξ ) − ϕ s , τ ∥ Y s ≤ C ξ 2 \|\phi(\xi)-\phi_{s,\tau}\|_{\mathcal{Y}_s}\le C\xi^2 ∥ ϕ ( ξ ) − ϕ s , τ ∥ Y s ≤ C ξ 2 .
Lemma 6.4 (Gap away from zero). For every ξ 0 > 0 \xi_0>0 ξ 0 > 0 , there is c ( ξ 0 ) > 0 c(\xi_0)>0 c ( ξ 0 ) > 0 such that
b s , ξ [ ψ ] − Λ s , τ ∥ ψ ∥ L 2 ( m s , τ ) 2 ≥ c ( ξ 0 ) ∥ ψ ∥ L 2 ( m s , τ ) 2 (88) b_{s,\xi}[\psi]-\Lambda_{s,\tau}\|\psi\|_{L^2(m_{s,\tau})}^2\ge c(\xi_0)\|\psi\|_{L^2(m_{s,\tau})}^2 \tag*{(88)} b s , ξ [ ψ ] − Λ s , τ ∥ ψ ∥ L 2 ( m s , τ ) 2 ≥ c ( ξ 0 ) ∥ ψ ∥ L 2 ( m s , τ ) 2 (88) when ∣ ξ ∣ ≥ ξ 0 |\xi|\ge\xi_0 ∣ ξ ∣ ≥ ξ 0 .
Proof. By Lemmas 4.3 and 4.7 , λ 1 ( ξ ) − Λ s , τ \lambda_1(\xi)-\Lambda_{s,\tau} λ 1 ( ξ ) − Λ s , τ has a positive minimum on ξ 0 ≤ ∣ ξ ∣ ≤ R \xi_0\leq|\xi|\leq R ξ 0 ≤ ∣ ξ ∣ ≤ R . For large R R R , Lemma 4.9 gives a positive bound on ∣ ξ ∣ ≥ R |\xi|\geq R ∣ ξ ∣ ≥ R . The variational definition of λ 1 ( ξ ) \lambda_1(\xi) λ 1 ( ξ ) proves (88) . □ \square □
Proof of Theorem 1.3 . Apply Lemma 6.3 to w ^ ( ξ , ⋅ ) \widehat w(\xi,\cdot) w ( ξ , ⋅ ) for ∣ ξ ∣ ≤ ξ 0 |\xi|\leq\xi_0 ∣ ξ ∣ ≤ ξ 0 . By (87) , the two squares with P ( ξ ) P(\xi) P ( ξ ) control
ξ 2 ∥ P 0 w ^ ∥ L 2 ( m s , τ ) 2 + ∥ ( I − P 0 ) w ^ ∥ L 2 ( m s , τ ) 2 . \xi^2\lVert P_0\widehat w\rVert_{L^2(m_{s,\tau})}^2+\lVert(I-P_0)\widehat w\rVert_{L^2(m_{s,\tau})}^2. ξ 2 ∥ P 0 w ∥ L 2 ( m s , τ ) 2 + ∥( I − P 0 ) w ∥ L 2 ( m s , τ ) 2 .
Indeed, put P = P ( ξ ) P=P(\xi) P = P ( ξ ) and Q = I − P Q=I-P Q = I − P . Then
ξ 2 ∥ P 0 f ∥ 2 ≤ C ξ 2 ∥ P f ∥ 2 + C ξ 6 ∥ Q f ∥ 2 , \xi^2\lVert P_0f\rVert^2\leq C\xi^2\lVert Pf\rVert^2+C\xi^6\lVert Qf\rVert^2, ξ 2 ∥ P 0 f ∥ 2 ≤ C ξ 2 ∥ P f ∥ 2 + C ξ 6 ∥ Q f ∥ 2 ,
∥ Q 0 f ∥ 2 ≤ C ∥ Q f ∥ 2 + C ξ 4 ∥ P f ∥ 2 . \lVert Q_0f\rVert^2\leq C\lVert Qf\rVert^2+C\xi^4\lVert Pf\rVert^2. ∥ Q 0 f ∥ 2 ≤ C ∥ Q f ∥ 2 + C ξ 4 ∥ P f ∥ 2 .
Take ξ 0 ≤ 1 \xi_0\leq1 ξ 0 ≤ 1 . Adding these inequalities gives
ξ 2 ∥ P 0 f ∥ 2 + ∥ Q 0 f ∥ 2 ≤ C ( ξ 2 ∥ P ( ξ ) f ∥ 2 + ∥ ( I − P ( ξ ) ) f ∥ 2 ) ≤ C D ξ [ f ] , \xi^2\lVert P_0f\rVert^2+\lVert Q_0f\rVert^2\leq C\bigl(\xi^2\lVert P(\xi)f\rVert^2+\lVert(I-P(\xi))f\rVert^2\bigr)\leq CD_\xi[f], ξ 2 ∥ P 0 f ∥ 2 + ∥ Q 0 f ∥ 2 ≤ C ( ξ 2 ∥ P ( ξ ) f ∥ 2 + ∥( I − P ( ξ )) f ∥ 2 ) ≤ C D ξ [ f ] ,
where D ξ [ f ] = b s , ξ [ f ] − Λ s , τ ∥ f ∥ 2 D_\xi[f]=b_{s,\xi}[f]-\Lambda_{s,\tau}\lVert f\rVert^2 D ξ [ f ] = b s , ξ [ f ] − Λ s , τ ∥ f ∥ 2 . For ∣ ξ ∣ ≥ ξ 0 |\xi|\geq\xi_0 ∣ ξ ∣ ≥ ξ 0 , Lemma 6.4 gives
h ( ξ ) ∥ P 0 f ∥ 2 + ∥ Q 0 f ∥ 2 ≤ ∥ f ∥ 2 ≤ c ( ξ 0 ) − 1 D ξ [ f ] . h(\xi)\lVert P_0f\rVert^2+\lVert Q_0f\rVert^2\leq\lVert f\rVert^2\leq c(\xi_0)^{-1}D_\xi[f]. h ( ξ ) ∥ P 0 f ∥ 2 + ∥ Q 0 f ∥ 2 ≤ ∥ f ∥ 2 ≤ c ( ξ 0 ) − 1 D ξ [ f ] .
Integrate with f = w ^ ( ξ , ⋅ ) f=\widehat w(\xi,\cdot) f = w ( ξ , ⋅ ) and use (59) and (8) . This proves (9) ; density extends the estimate from the core to X s \mathcal X_s X s . Proposition 6.1 gives (10) . □ \square □
Proof of Theorem 1.4 . Step 1: fixed-frequency minimizers. For ξ ≠ 0 \xi\neq0 ξ = 0 , write
d ξ [ ψ ] = b s , ξ [ ψ ] − Λ s , τ ∥ ψ ∥ H s , τ 2 d_\xi[\psi]=b_{s,\xi}[\psi]-\Lambda_{s,\tau}\lVert\psi\rVert_{H_{s,\tau}}^2 d ξ [ ψ ] = b s , ξ [ ψ ] − Λ s , τ ∥ ψ ∥ H s , τ 2
n ξ [ ψ ] = ∥ Q 0 ψ ∥ H s , τ 2 + h ( ξ ) ∥ P 0 ψ ∥ H s , τ 2 . n_\xi[\psi]=\lVert Q_0\psi\rVert_{H_{s,\tau}}^2+h(\xi)\lVert P_0\psi\rVert_{H_{s,\tau}}^2. n ξ [ ψ ] = ∥ Q 0 ψ ∥ H s , τ 2 + h ( ξ ) ∥ P 0 ψ ∥ H s , τ 2 .
For fixed ξ ≠ 0 \xi\neq0 ξ = 0 , the form n ξ n_\xi n ξ is equivalent to the squared H s , τ H_{s,\tau} H s , τ norm. The proof of Theorem 1.3 gives d ξ ≥ c n ξ d_\xi\geq cn_\xi d ξ ≥ c n ξ . A minimizing sequence for (12) , normalized by n ξ [ ψ j ] = 1 n_\xi[\psi_j]=1 n ξ [ ψ j ] = 1 , satisfies
b s , ξ [ ψ j ] = d ξ [ ψ j ] + Λ s , τ ∥ ψ j ∥ H s , τ 2 b_{s,\xi}[\psi_j]=d_\xi[\psi_j]+\Lambda_{s,\tau}\lVert\psi_j\rVert_{H_{s,\tau}}^2 b s , ξ [ ψ j ] = d ξ [ ψ j ] + Λ s , τ ∥ ψ j ∥ H s , τ 2
with a bounded right-hand side. Hence a subsequence converges weakly in Y s \mathcal Y_s Y s and strongly in H s , τ H_{s,\tau} H s , τ by (51) . Lower semicontinuity gives a minimizer, and μ s , τ ( ξ ) ≥ c > 0 \mu_{s,\tau}(\xi)\geq c>0 μ s , τ ( ξ ) ≥ c > 0 .
Step 2: continuity away from zero and coercivity at infinity. Let ξ j → ξ ≠ 0 \xi_j\to\xi\neq0 ξ j → ξ = 0 . Testing with a minimizer at ξ \xi ξ gives upper semicontinuity. For the lower bound, normalize minimizers at ξ j \xi_j ξ j by n ξ j [ ψ j ] = 1 n_{\xi_j}[\psi_j]=1 n ξ j [ ψ j ] = 1 . These norms are uniformly equivalent to ∥ ⋅ ∥ H s , τ 2 \lVert\cdot\rVert_{H_{s,\tau}}^2 ∥ ⋅ ∥ H s , τ 2 near ξ \xi ξ . A fixed trial function bounds μ s , τ ( ξ j ) \mu_{s,\tau}(\xi_j) μ s , τ ( ξ j ) , so ψ j \psi_j ψ j is bounded in Y s \mathcal Y_s Y s . For every compact frequency interval K K K , the proof of (55) gives
∣ b s , ξ [ ψ ] − b s , η [ ψ ] ∣ ≤ ω K ( ∣ ξ − η ∣ ) ∥ ψ ∥ L 2 ( 0 , π ) 2 , ω K ( r ) ⟶ 0. \lvert b_{s,\xi}[\psi]-b_{s,\eta}[\psi]\rvert\leq\omega_K(\lvert\xi-\eta\rvert)\lVert\psi\rVert_{L^2(0,\pi)}^2,\qquad\omega_K(r)\longrightarrow0. ∣ b s , ξ [ ψ ] − b s , η [ ψ ]∣ ≤ ω K (∣ ξ − η ∣) ∥ ψ ∥ L 2 ( 0 , π ) 2 , ω K ( r ) ⟶ 0.
After passage to a subsequence, ψ j ⇀ ψ \psi_j\rightharpoonup\psi ψ j ⇀ ψ in Y s \mathcal Y_s Y s and ψ j → ψ \psi_j\to\psi ψ j → ψ in H s , τ H_{s,\tau} H s , τ . Continuity of h h h gives n ξ [ ψ ] = 1 n_\xi[\psi]=1 n ξ [ ψ ] = 1 . The two-frequency estimate and (53) give
∣ b s , ξ j [ ψ j ] − b s , ξ [ ψ j ] ∣ ⟶ 0. \lvert b_{s,\xi_j}[\psi_j]-b_{s,\xi}[\psi_j]\rvert\longrightarrow0. ∣ b s , ξ j [ ψ j ] − b s , ξ [ ψ j ]∣ ⟶ 0.
Weak lower semicontinuity now yields
μ s , τ ( ξ ) ≤ d ξ [ ψ ] ≤ lim inf j → ∞ μ s , τ ( ξ j ) . \mu_{s,\tau}(\xi)\leq d_\xi[\psi]\leq\liminf_{j\to\infty}\mu_{s,\tau}(\xi_j). μ s , τ ( ξ ) ≤ d ξ [ ψ ] ≤ j → ∞ lim inf μ s , τ ( ξ j ) .
Thus μ s , τ \mu_{s,\tau} μ s , τ is continuous away from zero. For ∣ ξ ∣ ≥ 1 |\xi|\geq1 ∣ ξ ∣ ≥ 1 , n ξ [ ψ ] = ∥ ψ ∥ H s , τ 2 n_\xi[\psi]=\lVert\psi\rVert_{H_{s,\tau}}^2 n ξ [ ψ ] = ∥ ψ ∥ H s , τ 2 , and
μ s , τ ( ξ ) = λ 1 ( ξ ) − Λ s , τ ⟶ ∞ \mu_{s,\tau}(\xi)=\lambda_1(\xi)-\Lambda_{s,\tau}\longrightarrow\infty μ s , τ ( ξ ) = λ 1 ( ξ ) − Λ s , τ ⟶ ∞
as ∣ ξ ∣ → ∞ |\xi|\to\infty ∣ ξ ∣ → ∞ .
Step 3: the direct-integral minimum. By the closed direct-integral identity,
q s [ w ] − Λ s , τ ∥ w ∥ L 2 ( m s , τ ) 2 Dist s , τ ( u ) 2 = ∫ R d ξ [ w ^ ( ξ ) ] d ξ ∫ R n ξ [ w ^ ( ξ ) ] d ξ . \frac{q_s[w]-\Lambda_{s,\tau}\lVert w\rVert_{L^2(m_{s,\tau})}^2}{\operatorname{Dist}_{s,\tau}(u)^2}
=
\frac{\int_{\mathbb R}d_\xi[\widehat w(\xi)]\,d\xi}{\int_{\mathbb R}n_\xi[\widehat w(\xi)]\,d\xi}. Dist s , τ ( u ) 2 q s [ w ] − Λ s , τ ∥ w ∥ L 2 ( m s , τ ) 2 = ∫ R n ξ [ w ( ξ )] d ξ ∫ R d ξ [ w ( ξ )] d ξ .
This quotient is at least ess inf μ s , τ \operatorname*{ess\,inf}\mu_{s,\tau} ess inf μ s , τ . Conversely, choose ξ ∗ ≠ 0 \xi_* \ne0 ξ ∗ = 0 with μ s , τ ( ξ ∗ ) < ess inf μ s , τ + ϵ \mu_{s,\tau}(\xi_*) < \operatorname*{ess\,inf}\mu_{s,\tau}+\epsilon μ s , τ ( ξ ∗ ) < ess inf μ s , τ + ϵ . Let ψ ∗ \psi_* ψ ∗ be its minimizing vector, normalized by n ξ ∗ [ ψ ∗ ] = 1 n_{\xi_*}[\psi_*]=1 n ξ ∗ [ ψ ∗ ] = 1 . Choose χ ∈ C c ∞ ( ( − 1 , 1 ) ) \chi\in C_c^\infty((-1,1)) χ ∈ C c ∞ (( − 1 , 1 )) with ∥ χ ∥ 2 = 1 \|\chi\|_2=1 ∥ χ ∥ 2 = 1 and set
w ^ δ ( ξ , θ ) = δ − 1 / 2 χ ( ( ξ − ξ ∗ ) / δ ) ψ ∗ ( θ ) , 0 < δ < ∣ ξ ∗ ∣ / 2. \widehat w_\delta(\xi,\theta)=\delta^{-1/2}\chi((\xi-\xi_*)/\delta)\psi_*(\theta), \qquad0<\delta<|\xi_*|/2. w δ ( ξ , θ ) = δ − 1/2 χ (( ξ − ξ ∗ ) / δ ) ψ ∗ ( θ ) , 0 < δ < ∣ ξ ∗ ∣/2.
Uniform graph equivalence on its bounded frequency support shows that w δ ∈ X s w_\delta\in\mathcal{X}_s w δ ∈ X s . Changing variables to ξ = ξ ∗ + δ r \xi=\xi_*+\delta r ξ = ξ ∗ + δ r gives
∫ R ∂ ξ [ w ^ δ ( ξ ) ] d ξ = ∫ − 1 1 ∣ χ ( r ) ∣ 2 ∂ ξ ∗ + δ r [ ψ ∗ ] d r ⟶ μ s , τ ( ξ ∗ ) , \int_{\mathbb{R}}\partial_\xi[\widehat w_\delta(\xi)]\,d\xi=\int_{-1}^{1}|\chi(r)|^2\partial_{\xi_*+\delta r}[\psi_*]\,dr\longrightarrow\mu_{s,\tau}(\xi_*), ∫ R ∂ ξ [ w δ ( ξ )] d ξ = ∫ − 1 1 ∣ χ ( r ) ∣ 2 ∂ ξ ∗ + δ r [ ψ ∗ ] d r ⟶ μ s , τ ( ξ ∗ ) ,
∫ R n ξ [ w ^ δ ( ξ ) ] d ξ ⟶ 1. \int_{\mathbb{R}}n_\xi[\widehat w_\delta(\xi)]\,d\xi\longrightarrow1. ∫ R n ξ [ w δ ( ξ )] d ξ ⟶ 1.
Both limits follow from the form continuity in Step 2. For sufficiently small δ \delta δ the quotient is less than ess inf μ s , τ + 2 ϵ \operatorname*{ess\,inf}\mu_{s,\tau}+2\epsilon ess inf μ s , τ + 2 ϵ . The distance is bounded by the base norm, so core approximation preserves this quotient whenever its denominator is nonzero. If real test functions are required, use the sum of this packet and its reflected complex conjugate at − ξ ∗ -\xi_* − ξ ∗ ; their frequency supports are disjoint. This proves (13) ; positivity follows from Theorem 1.3 .
Step 4: exclusion of flat minimizing bands. If a nonzero w w w attains the global quotient, then
∂ ξ [ w ^ ( ξ ) ] − Γ s , τ h , p n ξ [ w ^ ( ξ ) ] \partial_\xi[\widehat w(\xi)]-\Gamma_{s,\tau}^{\mathrm{h},p}n_\xi[\widehat w(\xi)] ∂ ξ [ w ( ξ )] − Γ s , τ h , p n ξ [ w ( ξ )]
is nonnegative and has zero integral. Thus w ^ \widehat w w is supported in Z s , τ Z_{s,\tau} Z s , τ , which must have positive measure. We prove that this set is finite.
Write H = H s , τ H=H_{s,\tau} H = H s , τ , Y = Y s Y=\mathcal{Y}_s Y = Y s , P = P 0 P=P_0 P = P 0 , Q = I − P Q=I-P Q = I − P , and Λ = Λ s , τ \Lambda=\Lambda_{s,\tau} Λ = Λ s , τ . For u , v ∈ Y u,v\in Y u , v ∈ Y , put
⟨ J u , v ⟩ = ⟨ u , v ⟩ H , ⟨ J P u , v ⟩ = ⟨ P u , v ⟩ H , ⟨ J Q u , v ⟩ = ⟨ Q u , v ⟩ H . \langle Ju,v\rangle=\langle u,v\rangle_H,\qquad\langle JPu,v\rangle=\langle Pu,v\rangle_H,\qquad\langle JQu,v\rangle=\langle Qu,v\rangle_H. ⟨ J u , v ⟩ = ⟨ u , v ⟩ H , ⟨ J P u , v ⟩ = ⟨ P u , v ⟩ H , ⟨ J Q u , v ⟩ = ⟨ Q u , v ⟩ H .
These maps are compact by Y ↪ H Y\hookrightarrow H Y ↪ H . We also have Y ↪ L 2 ( 0 , π ) Y\hookrightarrow L^2(0,\pi) Y ↪ L 2 ( 0 , π ) compactly. Indeed, choose 0 < τ 0 < min { τ , 2 s } 0<\tau_0<\min\{\tau,2s\} 0 < τ 0 < min { τ , 2 s } and a bounded sequence v n v_n v n in Y Y Y . Equation (56) and Rellich’s theorem give compactness on [ δ , π − δ ] [\delta,\pi-\delta] [ δ , π − δ ] . For E δ = ( 0 , δ ) ∪ ( π − δ , π ) E_\delta=(0,\delta)\cup(\pi-\delta,\pi) E δ = ( 0 , δ ) ∪ ( π − δ , π ) , (52) gives
∫ E δ ∣ v n ∣ 2 d θ ≤ C δ 2 s − τ 0 ∫ E δ m s , τ 0 ∣ v n ∣ 2 d θ ≤ C δ 2 s − τ 0 . \int_{E_\delta}|v_n|^2\,d\theta\le C\delta^{2s-\tau_0}\int_{E_\delta}m_{s,\tau_0}|v_n|^2\,d\theta\le C\delta^{2s-\tau_0}. ∫ E δ ∣ v n ∣ 2 d θ ≤ C δ 2 s − τ 0 ∫ E δ m s , τ 0 ∣ v n ∣ 2 d θ ≤ C δ 2 s − τ 0 .
A diagonal argument gives compactness. Let S = { z ∈ C : ∣ Im z ∣ < 1 + s } \mathcal{S}=\{z\in\mathbb{C}:|\operatorname{Im}z|<1+s\} S = { z ∈ C : ∣ Im z ∣ < 1 + s } . For z ∈ S z\in\mathcal{S} z ∈ S and almost every α ∈ ( − π , π ) \alpha\in(-\pi,\pi) α ∈ ( − π , π ) , define
D z ( α ) = ∫ R ( 1 − cos ( z q ) ) K s ( q , α ) d q , D_z(\alpha)=\int_{\mathbb{R}}(1-\cos(zq))K_s(q,\alpha)\,dq, D z ( α ) = ∫ R ( 1 − cos ( z q )) K s ( q , α ) d q ,
and define B ( z ) : Y → Y ∗ B(z):Y\to Y^* B ( z ) : Y → Y ∗ by
⟨ ( B ( z ) − B ( 0 ) ) u , v ⟩ = ∫ 0 π ∫ 0 π D z ( θ − φ ) ( u ( θ ) v ( φ ) + u ( φ ) v ( θ ) ) d θ d φ . \langle(B(z)-B(0))u,v\rangle=\int_0^\pi\int_0^\pi D_z(\theta-\varphi)\bigl(u(\theta)v(\varphi)+u(\varphi)v(\theta)\bigr)\,d\theta\,d\varphi. ⟨( B ( z ) − B ( 0 )) u , v ⟩ = ∫ 0 π ∫ 0 π D z ( θ − φ ) ( u ( θ ) v ( φ ) + u ( φ ) v ( θ ) ) d θ d φ .
For real z = ξ z=\xi z = ξ , this is the form map of b s , ξ b_{s,\xi} b s , ξ . For K ⋐ S \mathcal{K}\Subset\mathcal{S} K ⋐ S , put r = sup z ∈ K ∣ Im z ∣ < 1 + s r=\sup_{z\in\mathcal{K}}|\operatorname{Im}z|<1+s r = sup z ∈ K ∣ Im z ∣ < 1 + s . For every integer k ≥ 0 k\ge0 k ≥ 0 ,
∣ ∂ z k ( 1 − cos ( z q ) ) ∣ ≤ C K , k q 2 ( ∣ q ∣ ≤ 1 ) , ∣ ∂ z k ( 1 − cos ( z q ) ) ∣ ≤ C K , k ( 1 + ∣ q ∣ k ) e r ∣ q ∣ ( ∣ q ∣ ≥ 1 ) . |\partial_z^k(1-\cos(zq))|\le C_{\mathcal{K},k}q^2\quad(|q|\le1),\qquad|\partial_z^k(1-\cos(zq))|\le C_{\mathcal{K},k}(1+|q|^k)e^{r|q|}\quad(|q|\ge1). ∣ ∂ z k ( 1 − cos ( z q )) ∣ ≤ C K , k q 2 ( ∣ q ∣ ≤ 1 ) , ∣ ∂ z k ( 1 − cos ( z q )) ∣ ≤ C K , k ( 1 + ∣ q ∣ k ) e r ∣ q ∣ ( ∣ q ∣ ≥ 1 ) .
Since
∫ q 2 + α 2 < 1 q 2 ( q 2 + α 2 ) 1 + s d q d α < ∞ , \int_{q^2+\alpha^2<1}\frac{q^2}{(q^2+\alpha^2)^{1+s}}\,dq\,d\alpha<\infty, ∫ q 2 + α 2 < 1 ( q 2 + α 2 ) 1 + s q 2 d q d α < ∞ ,
the exponential tail of K s K_s K s gives holomorphy of z ↦ D z z\mapsto D_z z ↦ D z in L 1 ( − π , π ) L^1(-\pi,\pi) L 1 ( − π , π ) . Young’s inequality yields
∣ ⟨ ( B ( z ) − B ( 0 ) ) u , v ⟩ ∣ ≤ 2 ∥ D z ∥ L 1 ( − π , π ) ∥ u ∥ L 2 ( 0 , π ) ∥ v ∥ L 2 ( 0 , π ) . |\langle(B(z)-B(0))u,v\rangle|\le2\|D_z\|_{L^1(-\pi,\pi)}\|u\|_{L^2(0,\pi)}\|v\|_{L^2(0,\pi)}. ∣ ⟨( B ( z ) − B ( 0 )) u , v ⟩ ∣ ≤ 2∥ D z ∥ L 1 ( − π , π ) ∥ u ∥ L 2 ( 0 , π ) ∥ v ∥ L 2 ( 0 , π ) .
Thus B ( z ) B(z) B ( z ) is holomorphic in L ( Y , Y ∗ ) \mathcal{L}(Y,Y^*) L ( Y , Y ∗ ) , and B ( z ) − B ( 0 ) B(z)-B(0) B ( z ) − B ( 0 ) is compact by the preceding L 2 L^2 L 2 embedding. Fix c > 0 c>0 c > 0 and put
F < ( z ) = B ( z ) − Λ J − c ( J Q + z 2 J P ) , \mathcal{F}_{<}(z)=B(z)-\Lambda J-c(J_Q+z^2J_P), F < ( z ) = B ( z ) − Λ J − c ( J Q + z 2 J P ) ,
F > ( z ) = B ( z ) − ( Λ + c ) J . \mathcal{F}_{>}(z)=B(z)-(\Lambda+c)J. F > ( z ) = B ( z ) − ( Λ + c ) J .
By the form norm and Lax–Milgram, B ( 0 ) + J : Y → Y ∗ B(0)+J:Y\to Y^* B ( 0 ) + J : Y → Y ∗ is an isomorphism. Moreover,
( B ( 0 ) + J ) − 1 F < ( z ) = I + ( B ( 0 ) + J ) − 1 { B ( z ) − B ( 0 ) − ( Λ + 1 ) J − c J Q − c z 2 J P } , (B(0)+J)^{-1}\mathcal{F}_{<}(z)=I+(B(0)+J)^{-1}\{B(z)-B(0)-(\Lambda+1)J-cJ_Q-cz^2J_P\}, ( B ( 0 ) + J ) − 1 F < ( z ) = I + ( B ( 0 ) + J ) − 1 { B ( z ) − B ( 0 ) − ( Λ + 1 ) J − c J Q − c z 2 J P } ,
( B ( 0 ) + J ) − 1 F > ( z ) = I + ( B ( 0 ) + J ) − 1 { B ( z ) − B ( 0 ) − ( Λ + c + 1 ) J } . (B(0)+J)^{-1}\mathcal{F}_{>}(z)=I+(B(0)+J)^{-1}\{B(z)-B(0)-(\Lambda+c+1)J\}. ( B ( 0 ) + J ) − 1 F > ( z ) = I + ( B ( 0 ) + J ) − 1 { B ( z ) − B ( 0 ) − ( Λ + c + 1 ) J } .
Both pencils are Fredholm of index zero. By the analytic Fredholm theorem [13 Chapter VII] , each singular set is discrete in S \mathcal{S} S provided the pencil has one invertible point. By (67) , choose R > 1 R>1 R > 1 with
λ 1 ( R ) − Λ > c . \lambda_1(R)-\Lambda>c. λ 1 ( R ) − Λ > c .
The variational principle gives ker F > ( R ) = { 0 } \ker\mathcal{F}_{>}(R)=\{0\} ker F > ( R ) = { 0 } , hence F > ( R ) \mathcal{F}_{>}(R) F > ( R ) is invertible. For R > 1 R>1 R > 1 , F < ( R ) \mathcal{F}_{<}(R) F < ( R ) is the analytic continuation of the low-frequency pencil; the actual multiplier in the stability quotient is h ( R ) = 1 h(R)=1 h ( R ) = 1 . Let A R A_R A R be associated with d R \mathfrak{d}_R d R and set T R = A R − c Q T_R=A_R-cQ T R = A R − c Q . Its form on Y Y Y is t R [ u , v ] = d R ( u , v ) − c ⟨ Q u , v ⟩ H t_R[u,v]=\mathfrak{d}_R(u,v)-c\langle Qu,v\rangle_H t R [ u , v ] = d R ( u , v ) − c ⟨ Q u , v ⟩ H , and
t R [ v ] ≥ ( λ 1 ( R ) − Λ − c ) ∥ v ∥ H 2 . t_R[v]\geq(\lambda_1(R)-\Lambda-c)\lVert v\rVert_H^2. t R [ v ] ≥ ( λ 1 ( R ) − Λ − c ) ∥ v ∥ H 2 .
Thus T R T_R T R is positive and invertible. Since F < ( R ) \mathcal{F}_{<}(R) F < ( R ) is the form map of T R − c R 2 P T_R-cR^2P T R − c R 2 P , the rank-one resolvent identity gives
ker F < ( R ) ≠ { 0 } ⟺ c R 2 ⟨ ϕ s , τ , T R − 1 ϕ s , τ ⟩ H = 1. \ker\mathcal{F}_{<}(R)\ne\{0\}\quad\Longleftrightarrow\quad cR^2\langle\phi_{s,\tau},T_R^{-1}\phi_{s,\tau}\rangle_H=1. ker F < ( R ) = { 0 } ⟺ c R 2 ⟨ ϕ s , τ , T R − 1 ϕ s , τ ⟩ H = 1.
Indeed, a nonzero kernel vector v v v satisfies
v = c R 2 ⟨ v , ϕ s , τ ⟩ H T R − 1 ϕ s , τ . v=cR^2\langle v,\phi_{s,\tau}\rangle_H T_R^{-1}\phi_{s,\tau}. v = c R 2 ⟨ v , ϕ s , τ ⟩ H T R − 1 ϕ s , τ .
The coefficient ⟨ v , ϕ s , τ ⟩ H \langle v,\phi_{s,\tau}\rangle_H ⟨ v , ϕ s , τ ⟩ H is nonzero. Taking the scalar product with ϕ s , τ \phi_{s,\tau} ϕ s , τ proves necessity; inserting T R − 1 ϕ s , τ T_R^{-1}\phi_{s,\tau} T R − 1 ϕ s , τ proves sufficiency. Since ϕ s , τ ∈ Y \phi_{s,\tau}\in Y ϕ s , τ ∈ Y , Cauchy–Schwarz gives
1 = ∣ ⟨ T R − 1 / 2 ϕ s , τ , T R 1 / 2 ϕ s , τ ⟩ H ∣ 2 ≤ ⟨ ϕ s , τ , T R − 1 ϕ s , τ ⟩ H t R [ ϕ s , τ ] . 1=\left|\langle T_R^{-1/2}\phi_{s,\tau},T_R^{1/2}\phi_{s,\tau}\rangle_H\right|^2\leq\langle\phi_{s,\tau},T_R^{-1}\phi_{s,\tau}\rangle_Ht_R[\phi_{s,\tau}]. 1 = ⟨ T R − 1/2 ϕ s , τ , T R 1/2 ϕ s , τ ⟩ H 2 ≤ ⟨ ϕ s , τ , T R − 1 ϕ s , τ ⟩ H t R [ ϕ s , τ ] .
Since Q ϕ s , τ = 0 Q\phi_{s,\tau}=0 Q ϕ s , τ = 0 and ∥ ϕ s , τ ∥ H = 1 \lVert\phi_{s,\tau}\rVert_H=1 ∥ ϕ s , τ ∥ H = 1 ,
t R [ ϕ s , τ ] = d R [ ϕ s , τ ] = ρ s , τ ( R ) . t_R[\phi_{s,\tau}]=\mathfrak{d}_R[\phi_{s,\tau}]=\rho_{s,\tau}(R). t R [ ϕ s , τ ] = d R [ ϕ s , τ ] = ρ s , τ ( R ) .
By (85) ,
c R 2 ⟨ ϕ s , τ , T R − 1 ϕ s , τ ⟩ H ≥ c R 2 ρ s , τ ( R ) ≥ c R 2 C ( 1 + R 2 s ) ⟶ ∞ . cR^2\langle\phi_{s,\tau},T_R^{-1}\phi_{s,\tau}\rangle_H\geq\frac{cR^2}{\rho_{s,\tau}(R)}\geq\frac{cR^2}{C(1+R^{2s})}\longrightarrow\infty. c R 2 ⟨ ϕ s , τ , T R − 1 ϕ s , τ ⟩ H ≥ ρ s , τ ( R ) c R 2 ≥ C ( 1 + R 2 s ) c R 2 ⟶ ∞.
For larger R R R , the left-hand side exceeds one. The rank-one criterion and index zero imply that F < ( R ) \mathcal{F}_{<}(R) F < ( R ) is invertible.
Take c = Γ s , τ h p c=\Gamma^{\mathrm{hp}}_{s,\tau} c = Γ s , τ hp . A fiber minimizer at a point of Z s , τ Z_{s,\tau} Z s , τ belongs to ker F < ( ξ ) \ker\mathcal{F}_{<}(\xi) ker F < ( ξ ) for 0 < ∣ ξ ∣ < 1 0<|\xi|<1 0 < ∣ ξ ∣ < 1 , and to ker F > ( ξ ) \ker\mathcal{F}_{>}(\xi) ker F > ( ξ ) for ∣ ξ ∣ > 1 |\xi|>1 ∣ ξ ∣ > 1 . By Step 2, μ s , τ ( ξ ) ⟶ ∞ \mu_{s,\tau}(\xi)\longrightarrow\infty μ s , τ ( ξ ) ⟶ ∞ as ∣ ξ ∣ → ∞ |\xi|\to\infty ∣ ξ ∣ → ∞ , so Z s , τ Z_{s,\tau} Z s , τ lies in a compact real frequency segment. Both singular sets have no accumulation point in S \mathcal{S} S , and hence each meets this compact segment in only finitely many points. Adding the junction points ± 1 \pm1 ± 1 therefore still gives a finite set. Thus Z s , τ Z_{s,\tau} Z s , τ is finite and has measure zero. The necessary condition above proves nonattainment.
Step 5: the zero-frequency endpoint. It remains to prove (14) . Put g 0 = λ 2 ( 0 ) − λ 1 ( 0 ) g_0=\lambda_2(0)-\lambda_1(0) g 0 = λ 2 ( 0 ) − λ 1 ( 0 ) . For small ξ \xi ξ , decompose ψ = P ( ξ ) ψ + ( I − P ( ξ ) ) ψ = p + q \psi=P(\xi)\psi+(I-P(\xi))\psi=p+q ψ = P ( ξ ) ψ + ( I − P ( ξ )) ψ = p + q . The spectral theorem gives
d ξ [ ψ ] ≥ ( κ s , τ ξ 2 + O ( ξ 4 ) ) ∥ p ∥ 2 + ( g 0 + o ( 1 ) ) ∥ q ∥ 2 . \mathfrak{d}_{\xi}[\psi]\geq(\kappa_{s,\tau}\xi^2+O(\xi^4))\lVert p\rVert^2+(g_0+o(1))\lVert q\rVert^2. d ξ [ ψ ] ≥ ( κ s , τ ξ 2 + O ( ξ 4 )) ∥ p ∥ 2 + ( g 0 + o ( 1 )) ∥ q ∥ 2 .
By (87) ,
( 1 − C ∣ ξ ∣ ) ( ξ 2 ∥ p ∥ 2 + ∥ q ∥ 2 ) ≤ n ξ [ p + q ] ≤ ( 1 + C ∣ ξ ∣ ) ( ξ 2 ∥ p ∥ 2 + ∥ q ∥ 2 ) . (1-C|\xi|)(\xi^2\lVert p\rVert^2+\lVert q\rVert^2)\leq n_{\xi}[p+q]\leq(1+C|\xi|)(\xi^2\lVert p\rVert^2+\lVert q\rVert^2). ( 1 − C ∣ ξ ∣ ) ( ξ 2 ∥ p ∥ 2 + ∥ q ∥ 2 ) ≤ n ξ [ p + q ] ≤ ( 1 + C ∣ ξ ∣ ) ( ξ 2 ∥ p ∥ 2 + ∥ q ∥ 2 ) .
Indeed, Q 0 p = O ( ξ 2 ) ∥ p ∥ Q_{0}p=O(\xi^{2})\lVert p\rVert Q 0 p = O ( ξ 2 ) ∥ p ∥ and P 0 q = O ( ξ 2 ) ∥ q ∥ P_{0}q=O(\xi^{2})\lVert q\rVert P 0 q = O ( ξ 2 ) ∥ q ∥ . The mixed term is bounded by
C ξ 2 ∥ p ∥ ∥ q ∥ ≤ C ∣ ξ ∣ ( ξ 2 ∥ p ∥ 2 + ∥ q ∥ 2 ) . C\xi^{2}\lVert p\rVert\lVert q\rVert\le C\lvert\xi\rvert(\xi^{2}\lVert p\rVert^{2}+\lVert q\rVert^{2}). C ξ 2 ∥ p ∥ ∥ q ∥ ≤ C ∣ ξ ∣ ( ξ 2 ∥ p ∥ 2 + ∥ q ∥ 2 ) .
The remaining errors are smaller. Hence
lim inf ξ → 0 μ s , τ ( ξ ) ≥ min { κ s , τ , g 0 } . \liminf_{\xi\to0}\mu_{s,\tau}(\xi)\ge\min\{\kappa_{s,\tau},g_{0}\}. ξ → 0 lim inf μ s , τ ( ξ ) ≥ min { κ s , τ , g 0 } .
For the first eigenvector, (87) and Proposition 6.1 give
n ξ [ ϕ ( ξ ) ] = ξ 2 + O ( ξ 4 ) , d ξ [ ϕ ( ξ ) ] = κ s , τ ξ 2 + O ( ξ 4 ) . n_{\xi}[\phi(\xi)]=\xi^{2}+O(\xi^{4}),\qquad d_{\xi}[\phi(\xi)]=\kappa_{s,\tau}\xi^{2}+O(\xi^{4}). n ξ [ ϕ ( ξ )] = ξ 2 + O ( ξ 4 ) , d ξ [ ϕ ( ξ )] = κ s , τ ξ 2 + O ( ξ 4 ) .
Its quotient tends to κ s , τ \kappa_{s,\tau} κ s , τ . Let e 2 e_{2} e 2 be a normalized second eigenvector of the zero fiber. Then P 0 e 2 = 0 P_{0}e_{2}=0 P 0 e 2 = 0 , so n ξ [ e 2 ] = 1 n_{\xi}[e_{2}]=1 n ξ [ e 2 ] = 1 . The form expansion gives
d ξ [ e 2 ] = g 0 + O ( ξ 2 ) . d_{\xi}[e_{2}]=g_{0}+O(\xi^{2}). d ξ [ e 2 ] = g 0 + O ( ξ 2 ) .
Its quotient tends to g 0 g_{0} g 0 . This gives the second upper bound and proves (14) . Together with Step 2, this defines a continuous extension of μ s , τ \mu_{s,\tau} μ s , τ to R \mathbb{R} R which is coercive at infinity. Its infimum is therefore a minimum. Continuity also turns the essential infimum in (13) into this same minimum. □
Proposition 6.5 (Strict comparison and nonzero minimizing frequencies). For every 0 < s < 1 0<s<1 0 < s < 1 and τ > 0 \tau>0 τ > 0 ,
0 < Γ s , τ h p ≤ λ 1 ( 1 ) − Λ s , τ ≤ ρ s , τ ( 1 ) < κ s , τ . (89) 0<\Gamma^{\mathrm{hp}}_{s,\tau}\le\lambda_{1}(1)-\Lambda_{s,\tau}\le\rho_{s,\tau}(1)<\kappa_{s,\tau}. \tag*{(89)} 0 < Γ s , τ hp ≤ λ 1 ( 1 ) − Λ s , τ ≤ ρ s , τ ( 1 ) < κ s , τ . (89) If λ 2 ( 0 ) − λ 1 ( 0 ) ≥ κ s , τ \lambda_{2}(0)-\lambda_{1}(0)\ge\kappa_{s,\tau} λ 2 ( 0 ) − λ 1 ( 0 ) ≥ κ s , τ , then Z s , τ Z_{s,\tau} Z s , τ is nonempty and contains a pair of opposite nonzero frequencies. In particular, this conclusion holds for τ = 2 \tau=2 τ = 2 and all s < 1 s<1 s < 1 sufficiently close to one.
Proof. Since h ( 1 ) = 1 h(1)=1 h ( 1 ) = 1 , the variational formula gives μ s , τ ( 1 ) = λ 1 ( 1 ) − Λ s , τ \mu_{s,\tau}(1)=\lambda_{1}(1)-\Lambda_{s,\tau} μ s , τ ( 1 ) = λ 1 ( 1 ) − Λ s , τ . Testing its Rayleigh quotient with ϕ s , τ \phi_{s,\tau} ϕ s , τ gives the second upper bound. The strict scalar inequality 2 ( 1 − cos q ) < q 2 2(1-\cos q)<q^{2} 2 ( 1 − cos q ) < q 2 for q ≠ 0 q\ne0 q = 0 , together with positivity of K s K_{s} K s and ϕ s , τ \phi_{s,\tau} ϕ s , τ , gives
ρ s , τ ( 1 ) < ∫ 0 π ∫ 0 π ∫ R q 2 K s ( q , θ , ϕ ) ϕ s , τ ( θ ) ϕ s , τ ( ϕ ) d q d θ d ϕ = κ s , τ . \rho_{s,\tau}(1)<\int_{0}^{\pi}\int_{0}^{\pi}\int_{\mathbb{R}}q^{2}K_{s}(q,\theta,\phi)\phi_{s,\tau}(\theta)\phi_{s,\tau}(\phi)\,dq\,d\theta\,d\phi=\kappa_{s,\tau}. ρ s , τ ( 1 ) < ∫ 0 π ∫ 0 π ∫ R q 2 K s ( q , θ , ϕ ) ϕ s , τ ( θ ) ϕ s , τ ( ϕ ) d q d θ d ϕ = κ s , τ .
The moment integral is finite. This proves (89) .
Under the gap assumption, (14) gives μ s , τ ( 0 ) = κ s , τ > Γ s , τ h p \mu_{s,\tau}(0)=\kappa_{s,\tau}>\Gamma^{\mathrm{hp}}_{s,\tau} μ s , τ ( 0 ) = κ s , τ > Γ s , τ hp . Continuity and coercivity of μ s , τ \mu_{s,\tau} μ s , τ force a minimizer away from zero. Evenness gives its opposite, and Theorem 1.4 makes the set finite. For τ = 2 \tau=2 τ = 2 , Proposition 8.3 and Theorem 1.6 give
2 ( 1 − s ) π { λ 2 ( 0 ) − λ 1 ( 0 ) } ⟶ 2 , 2 ( 1 − s ) π κ s , 2 ⟶ 1. \frac{2(1-s)}{\pi}\{\lambda_{2}(0)-\lambda_{1}(0)\}\longrightarrow2,\qquad\frac{2(1-s)}{\pi}\kappa_{s,2}\longrightarrow1. π 2 ( 1 − s ) { λ 2 ( 0 ) − λ 1 ( 0 )} ⟶ 2 , π 2 ( 1 − s ) κ s , 2 ⟶ 1.
The gap assumption therefore holds for s s s sufficiently close to one. □
Remark 6.6. The minimizing frequencies refer to the fixed distance with multiplier h ( ξ ) = min { ξ 2 , 1 } h(\xi)=\min\{\xi^{2},1\} h ( ξ ) = min { ξ 2 , 1 } . The strict inequality (89) shows that its global coefficient is smaller than the quadratic dispersion coefficient. It does not locate the minimizing frequencies explicitly.
Corollary 6.7 (Sharp axial order). Let 0 < α < 2 0<\alpha<2 0 < α < 2 . There is no c > 0 c>0 c > 0 such that the left-hand side of (9) controls
∫ R min { ∣ ξ ∣ α , 1 } ∥ P 0 w ^ ( ξ , ⋅ ) ∥ H s , τ 2 d ξ \int_{\mathbb{R}}\min\{|\xi|^{\alpha},1\}\lVert P_{0}\widehat{w}(\xi,\cdot)\rVert_{H_{s,\tau}}^{2}\,d\xi ∫ R min { ∣ ξ ∣ α , 1 } ∥ P 0 w ( ξ , ⋅ ) ∥ H s , τ 2 d ξ for every test function. Thus the quadratic power in the axial term of (8) is optimal.
Proof. Take
w ^ ε ( ξ , θ ) = ε − 1 / 2 η ( ξ / ε ) ϕ s , τ ( θ ) , \widehat{w}_{\varepsilon}(\xi,\theta)=\varepsilon^{-1/2}\eta(\xi/\varepsilon)\phi_{s,\tau}(\theta), w ε ( ξ , θ ) = ε − 1/2 η ( ξ / ε ) ϕ s , τ ( θ ) ,
where η ∈ C c ∞ ( ( − 1 , 1 ) ) \eta\in C^\infty_c((-1,1)) η ∈ C c ∞ (( − 1 , 1 )) and ∥ η ∥ 2 = 1 \|\eta\|_2=1 ∥ η ∥ 2 = 1 . Equation (84) gives
q s [ w ε ] − Λ s , τ ∥ w ε ∥ L 2 ( R ; H s , τ ) 2 = κ s , τ ε 2 ∫ ζ 2 ∣ η ( ζ ) ∣ 2 d ζ + O ( ε 4 ) . q_s[w_\varepsilon]-\Lambda_{s,\tau}\|w_\varepsilon\|_{L^2(\mathbb{R};H_{s,\tau})}^2=\kappa_{s,\tau}\varepsilon^2\int\zeta^2|\eta(\zeta)|^2\,d\zeta+O(\varepsilon^4). q s [ w ε ] − Λ s , τ ∥ w ε ∥ L 2 ( R ; H s , τ ) 2 = κ s , τ ε 2 ∫ ζ 2 ∣ η ( ζ ) ∣ 2 d ζ + O ( ε 4 ) .
A direct change of variables gives
∫ min { ∣ ξ ∣ α , 1 } ∥ P 0 w ^ ε ( ξ ) ∥ 2 d ξ = ε α ∫ ∣ ζ ∣ α ∣ η ( ζ ) ∣ 2 d ζ \int\min\{|\xi|^\alpha,1\}\|P_0\widehat{w}_\varepsilon(\xi)\|^2\,d\xi=\varepsilon^\alpha\int|\zeta|^\alpha|\eta(\zeta)|^2\,d\zeta ∫ min { ∣ ξ ∣ α , 1 } ∥ P 0 w ε ( ξ ) ∥ 2 d ξ = ε α ∫ ∣ ζ ∣ α ∣ η ( ζ ) ∣ 2 d ζ
for small ε \varepsilon ε . Since α < 2 \alpha<2 α < 2 , this is larger in order than the preceding deficit. No positive uniform coefficient is possible. The displayed functions belong to the completed form domain. For each ε \varepsilon ε , core density gives v ε ∈ C v_\varepsilon\in\mathcal C v ε ∈ C such that
∥ v ε − w ε ∥ X s = o ( ε 2 ) . \|v_\varepsilon-w_\varepsilon\|_{X_s}=o(\varepsilon^2). ∥ v ε − w ε ∥ X s = o ( ε 2 ) .
The form Cauchy–Schwarz inequality and the uniform boundedness of ∥ w ε ∥ X s \|w_\varepsilon\|_{X_s} ∥ w ε ∥ X s show that the deficit of v ε v_\varepsilon v ε differs from that of w ε w_\varepsilon w ε by o ( ε 2 ) o(\varepsilon^2) o ( ε 2 ) . The multiplier min { ∣ ξ ∣ α , 1 } \min\{|\xi|^\alpha,1\} min { ∣ ξ ∣ α , 1 } is bounded by one. Plancherel’s theorem therefore shows that the corresponding multiplier integrals differ by o ( ε 2 ) = o ( ε α ) o(\varepsilon^2)=o(\varepsilon^\alpha) o ( ε 2 ) = o ( ε α ) . Their quotient still tends to zero. The inverse logarithmic polar transform of v ε v_\varepsilon v ε belongs to C c ∞ ( R + 2 ) C^\infty_c(\mathbb{R}_+^2) C c ∞ ( R + 2 ) . This proves the assertion on the test core. □
Energy decomposition and extremizing sequences
Fix δ > 0 \delta>0 δ > 0 such that the analytic projection P ( ξ ) P(\xi) P ( ξ ) and the normalized eigenfunction ϕ ( ξ ) \phi(\xi) ϕ ( ξ ) are defined for ∣ ξ ∣ < 4 δ |\xi|<4\delta ∣ ξ ∣ < 4 δ . Let χ 0 ∈ C c ∞ ( ( − 2 δ , 2 δ ) ) \chi_0\in C^\infty_c((-2\delta,2\delta)) χ 0 ∈ C c ∞ (( − 2 δ , 2 δ )) be even and satisfy 0 ≤ χ 0 ≤ 1 0\leq\chi_0\leq1 0 ≤ χ 0 ≤ 1 and χ 0 = 1 \chi_0=1 χ 0 = 1 on [ − δ , δ ] [-\delta,\delta] [ − δ , δ ] . For w ∈ X s w\in X_s w ∈ X s , define
c ( ξ ) = ⟨ w ^ ( ξ , ⋅ ) , ϕ ( ξ ) ⟩ H s , τ , ∣ ξ ∣ < 2 δ , c ( ξ ) = 0 , ∣ ξ ∣ ≥ 2 δ , a ^ ( ξ ) = χ 0 ( ξ ) c ( ξ ) , r ^ ( ξ , θ ) = w ^ ( ξ , θ ) − a ^ ( ξ ) ϕ s , τ ( θ ) . (90) \begin{aligned}
c(\xi)&=\langle\widehat{w}(\xi,\cdot),\phi(\xi)\rangle_{H_{s,\tau}}, &&|\xi|<2\delta,\\
c(\xi)&=0, &&|\xi|\geq2\delta,\\
\widehat{a}(\xi)&=\chi_0(\xi)c(\xi),\\
\widehat{r}(\xi,\theta)&=\widehat{w}(\xi,\theta)-\widehat{a}(\xi)\phi_{s,\tau}(\theta). \tag*{(90)}
\end{aligned} c ( ξ ) c ( ξ ) a ( ξ ) r ( ξ , θ ) = ⟨ w ( ξ , ⋅ ) , ϕ ( ξ ) ⟩ H s , τ , = 0 , = χ 0 ( ξ ) c ( ξ ) , = w ( ξ , θ ) − a ( ξ ) ϕ s , τ ( θ ) . ∣ ξ ∣ < 2 δ , ∣ ξ ∣ ≥ 2 δ , (90)
The cutoff restricts the moving projection to a neighborhood of zero.
Proposition 6.8 (Two-sided energy decomposition). There is C ≥ 1 C\geq1 C ≥ 1 , depending on s s s , τ \tau τ , and the fixed cutoff, such that for every w ∈ X s w\in X_s w ∈ X s ,
C − 1 ( ∥ r ∥ X s 2 + ∫ R ρ s , τ ( ξ ) ∣ a ^ ( ξ ) ∣ 2 d ξ ) ≤ q s [ w ] − Λ s , τ ∥ w ∥ L 2 ( m s , τ ) 2 C^{-1}\left(\|r\|_{X_s}^2+\int_{\mathbb R}\rho_{s,\tau}(\xi)|\widehat{a}(\xi)|^2\,d\xi\right)\leq q_s[w]-\Lambda_{s,\tau}\|w\|_{L^2(m_{s,\tau})}^2 C − 1 ( ∥ r ∥ X s 2 + ∫ R ρ s , τ ( ξ ) ∣ a ( ξ ) ∣ 2 d ξ ) ≤ q s [ w ] − Λ s , τ ∥ w ∥ L 2 ( m s , τ ) 2 ≤ C ( ∥ r ∥ X s 2 + ∫ R ρ s , τ ( ξ ) ∣ a ^ ( ξ ) ∣ 2 d ξ ) . \leq C\left(\|r\|_{X_s}^2+\int_{\mathbb R}\rho_{s,\tau}(\xi)|\widehat{a}(\xi)|^2\,d\xi\right). ≤ C ( ∥ r ∥ X s 2 + ∫ R ρ s , τ ( ξ ) ∣ a ( ξ ) ∣ 2 d ξ ) . If q s [ w ] ≤ C 0 q_s[w]\leq C_0 q s [ w ] ≤ C 0 and ∥ w ∥ L 2 ( m s , τ ) = 1 \|w\|_{L^2(m_{s,\tau})}=1 ∥ w ∥ L 2 ( m s , τ ) = 1 , then ∥ a ∥ 2 + ∥ r ∥ X s \|a\|_2+\|r\|_{X_s} ∥ a ∥ 2 + ∥ r ∥ X s is bounded by a constant depending only on C 0 C_0 C 0 and the fixed parameters.
Proof. Write
D ξ [ f ] = b s , ξ [ f ] − Λ s , τ ∥ f ∥ H s , τ 2 . D_\xi[f]=b_{s,\xi}[f]-\Lambda_{s,\tau}\|f\|_{H_{s,\tau}}^2. D ξ [ f ] = b s , ξ [ f ] − Λ s , τ ∥ f ∥ H s , τ 2 .
We divide the proof into three frequency ranges. If ∣ ξ ∣ < δ |\xi|<\delta ∣ ξ ∣ < δ , then
w ^ = c ( ξ ) ϕ ( ξ ) + z ( ξ ) , z ( ξ ) = ( I − P ( ξ ) ) w ^ . \widehat{w}=c(\xi)\phi(\xi)+z(\xi),\qquad z(\xi)=(I-P(\xi))\widehat{w}. w = c ( ξ ) ϕ ( ξ ) + z ( ξ ) , z ( ξ ) = ( I − P ( ξ )) w .
The two terms are orthogonal for both the Hilbert norm and the fiber operator. Hence
D ξ [ w ^ ] = ( λ 1 ( ξ ) − Λ s , τ ) ∣ c ( ξ ) ∣ 2 + D ξ [ z ( ξ ) ] . (91) D_\xi[\widehat{w}]=(\lambda_1(\xi)-\Lambda_{s,\tau})|c(\xi)|^2+D_\xi[z(\xi)]. \tag*{(91)} D ξ [ w ] = ( λ 1 ( ξ ) − Λ s , τ ) ∣ c ( ξ ) ∣ 2 + D ξ [ z ( ξ )] . (91)
The spectral gap above λ 1 ( ξ ) \lambda_1(\xi) λ 1 ( ξ ) gives
b s , ξ [ z ] ≤ C D ξ [ z ] . b_{s,\xi}[z]\leq CD_\xi[z]. b s , ξ [ z ] ≤ C D ξ [ z ] .
Equations (87) and (10) give
∥ ϕ ( ξ ) − ϕ s , τ ∥ Y s ≤ C ξ 2 , ξ 2 ∣ c ( ξ ) ∣ 2 ≤ C D ξ [ w ^ ] . \|\phi(\xi)-\phi_{s,\tau}\|_{Y_s}\le C\xi^2,\qquad\xi^2|c(\xi)|^2\le CD_\xi[\widehat{w}]. ∥ ϕ ( ξ ) − ϕ s , τ ∥ Y s ≤ C ξ 2 , ξ 2 ∣ c ( ξ ) ∣ 2 ≤ C D ξ [ w ] .
Since r ^ = z + c ( ϕ ( ξ ) − ϕ s , τ ) \widehat{r}=z+c(\phi(\xi)-\phi_{s,\tau}) r = z + c ( ϕ ( ξ ) − ϕ s , τ ) in this region, the form triangle inequality gives
b s , ξ [ r ^ ] 1 / 2 ≤ b s , ξ [ z ] 1 / 2 + ∣ c ( ξ ) ∣ b s , ξ [ ϕ ( ξ ) − ϕ s , τ ] 1 / 2 . b_{s,\xi}[\widehat{r}]^{1/2}\le b_{s,\xi}[z]^{1/2}+|c(\xi)|b_{s,\xi}[\phi(\xi)-\phi_{s,\tau}]^{1/2}. b s , ξ [ r ] 1/2 ≤ b s , ξ [ z ] 1/2 + ∣ c ( ξ ) ∣ b s , ξ [ ϕ ( ξ ) − ϕ s , τ ] 1/2 .
Uniform graph equivalence and (87) imply
b s , ξ [ ϕ ( ξ ) − ϕ s , τ ] ≤ C ξ 4 . b_{s,\xi}[\phi(\xi)-\phi_{s,\tau}]\le C\xi^4. b s , ξ [ ϕ ( ξ ) − ϕ s , τ ] ≤ C ξ 4 .
Combining these estimates with ρ s , τ ( ξ ) ≤ C ξ 2 \rho_{s,\tau}(\xi)\le C\xi^2 ρ s , τ ( ξ ) ≤ C ξ 2 and (91) , we obtain
b s , ξ [ r ^ ] + ρ s , τ ( ξ ) ∣ a ^ ( ξ ) ∣ 2 ≤ C D ξ [ w ^ ] . (92) b_{s,\xi}[\widehat{r}]+\rho_{s,\tau}(\xi)|\widehat{a}(\xi)|^2\le CD_\xi[\widehat{w}]. \tag*{(92)} b s , ξ [ r ] + ρ s , τ ( ξ ) ∣ a ( ξ ) ∣ 2 ≤ C D ξ [ w ] . (92)
On δ ≤ ∣ ξ ∣ ≤ 2 δ \delta\le|\xi|\le2\delta δ ≤ ∣ ξ ∣ ≤ 2 δ , Lemma 6.4 gives
∥ w ^ ∥ H s , τ 2 ≤ C D ξ [ w ^ ] , b s , ξ [ w ^ ] ≤ C D ξ [ w ^ ] . \|\widehat{w}\|_{H_{s,\tau}}^2\le CD_\xi[\widehat{w}],\qquad b_{s,\xi}[\widehat{w}]\le CD_\xi[\widehat{w}]. ∥ w ∥ H s , τ 2 ≤ C D ξ [ w ] , b s , ξ [ w ] ≤ C D ξ [ w ] .
By Cauchy–Schwarz and ∥ ϕ ( ξ ) ∥ H s , τ = 1 \|\phi(\xi)\|_{H_{s,\tau}}=1 ∥ ϕ ( ξ ) ∥ H s , τ = 1 ,
∣ a ^ ( ξ ) ∣ ≤ ∥ w ^ ( ξ ) ∥ H s , τ . |\widehat{a}(\xi)|\le\|\widehat{w}(\xi)\|_{H_{s,\tau}}. ∣ a ( ξ ) ∣ ≤ ∥ w ( ξ ) ∥ H s , τ .
Moreover, b s , ξ [ ϕ s , τ ] b_{s,\xi}[\phi_{s,\tau}] b s , ξ [ ϕ s , τ ] and ρ s , τ ( ξ ) \rho_{s,\tau}(\xi) ρ s , τ ( ξ ) are bounded on this compact interval. Thus
b s , ξ [ r ^ ] ≤ 2 b s , ξ [ w ^ ] + 2 ∣ a ^ ∣ 2 b s , ξ [ ϕ s , τ ] ≤ C D ξ [ w ^ ] , b_{s,\xi}[\widehat{r}]\le2b_{s,\xi}[\widehat{w}]+2|\widehat{a}|^2b_{s,\xi}[\phi_{s,\tau}]\le CD_\xi[\widehat{w}], b s , ξ [ r ] ≤ 2 b s , ξ [ w ] + 2∣ a ∣ 2 b s , ξ [ ϕ s , τ ] ≤ C D ξ [ w ] ,
ρ s , τ ( ξ ) ∣ a ^ ∣ 2 ≤ C ∥ w ^ ∥ H s , τ 2 ≤ C D ξ [ w ^ ] . \rho_{s,\tau}(\xi)|\widehat{a}|^2\le C\|\widehat{w}\|_{H_{s,\tau}}^2\le CD_\xi[\widehat{w}]. ρ s , τ ( ξ ) ∣ a ∣ 2 ≤ C ∥ w ∥ H s , τ 2 ≤ C D ξ [ w ] .
This proves (92) in the middle range.
If ∣ ξ ∣ > 2 δ |\xi|>2\delta ∣ ξ ∣ > 2 δ , then a ^ = 0 \widehat{a}=0 a = 0 and r ^ = w ^ \widehat{r}=\widehat{w} r = w . Lemma 6.4 again gives
b s , ξ [ r ^ ] = D ξ [ w ^ ] + Λ s , τ ∥ w ^ ∥ H s , τ 2 ≤ C D ξ [ w ^ ] . b_{s,\xi}[\widehat{r}]=D_\xi[\widehat{w}]+\Lambda_{s,\tau}\|\widehat{w}\|_{H_{s,\tau}}^2\le CD_\xi[\widehat{w}]. b s , ξ [ r ] = D ξ [ w ] + Λ s , τ ∥ w ∥ H s , τ 2 ≤ C D ξ [ w ] .
Integrate the three estimates and use (59) . Uniform fiber coercivity adds ∥ r ∥ L 2 ( R ; H s , τ ) 2 \|r\|_{L^2(\mathbb{R};H_{s,\tau})}^2 ∥ r ∥ L 2 ( R ; H s , τ ) 2 to the left-hand side. This proves the lower bound in (6.8). For the upper bound, use the nonnegative quadratic form
E s , τ [ v ] = q s [ v ] − Λ s , τ ∥ v ∥ L 2 ( m s , τ ) 2 . E_{s,\tau}[v]=q_s[v]-\Lambda_{s,\tau}\|v\|_{L^2(m_{s,\tau})}^2. E s , τ [ v ] = q s [ v ] − Λ s , τ ∥ v ∥ L 2 ( m s , τ ) 2 .
The canonical a ^ \widehat{a} a is supported in a fixed bounded frequency interval. Hence a ϕ s , τ ∈ X s a\phi_{s,\tau}\in X_s a ϕ s , τ ∈ X s , and the direct-integral identity gives
E s , τ [ a ϕ s , τ ] = ∫ R ρ s , τ ( ξ ) ∣ a ^ ( ξ ) ∣ 2 d ξ . E_{s,\tau}[a\phi_{s,\tau}]=\int_{\mathbb{R}}\rho_{s,\tau}(\xi)|\widehat{a}(\xi)|^2\,d\xi. E s , τ [ a ϕ s , τ ] = ∫ R ρ s , τ ( ξ ) ∣ a ( ξ ) ∣ 2 d ξ .
Cauchy–Schwarz for E s , τ E_{s,\tau} E s , τ yields
E s , τ [ w ] ≤ 2 E s , τ [ r ] + 2 E s , τ [ a ϕ s , τ ] ≤ 2 ∥ r ∥ X s 2 + 2 ∫ R ρ s , τ ( ξ ) ∣ a ^ ( ξ ) ∣ 2 d ξ . E_{s,\tau}[w]\le2E_{s,\tau}[r]+2E_{s,\tau}[a\phi_{s,\tau}]
\le2\|r\|_{X_s}^2+2\int_{\mathbb{R}}\rho_{s,\tau}(\xi)|\widehat{a}(\xi)|^2\,d\xi. E s , τ [ w ] ≤ 2 E s , τ [ r ] + 2 E s , τ [ a ϕ s , τ ] ≤ 2∥ r ∥ X s 2 + 2 ∫ R ρ s , τ ( ξ ) ∣ a ( ξ ) ∣ 2 d ξ .
This proves the upper bound. The boundedness assertion follows from the lower bound, Plancherel’s identity, and w = a ϕ s , τ + r w=a\phi_{s,\tau}+r w = a ϕ s , τ + r . □
Proof of Theorem 1.5. The two-sided estimate is Proposition 6.8. Assume (1). The transformed sequence is normalized in L 2 ( m s , τ ) L^2(m_{s,\tau}) L 2 ( m s , τ ) , and its deficit tends to zero. Apply Proposition 6.8. We obtain (1.16) and (16) . Since w n = a n ϕ s , τ + r n w_n=a_n\phi_{s,\tau}+r_n w n = a n ϕ s , τ + r n and ∥ ϕ s , τ ∥ H s , τ = 1 \|\phi_{s,\tau}\|_{H_{s,\tau}}=1 ∥ ϕ s , τ ∥ H s , τ = 1 ,
∣ ∥ a n ∥ 2 2 − 1 ∣ ≤ ∥ r n ∥ L 2 ( R ; H s , τ ) . \bigl|\|a_n\|_2^2-1\bigr|\le\|r_n\|_{L^2(\mathbb{R};H_{s,\tau})}. ∥ a n ∥ 2 2 − 1 ≤ ∥ r n ∥ L 2 ( R ; H s , τ ) .
Hence ∥ a n ∥ 2 → 1 \|a_n\|_2\to1 ∥ a n ∥ 2 → 1 . This proves (2).
Conversely, assume (2). Since a n ϕ s , τ = w n − r n ∈ X s a_n\phi_{s,\tau}=w_n-r_n\in X_s a n ϕ s , τ = w n − r n ∈ X s , the direct-integral identity gives
q s [ a n ϕ s , τ ] − Λ s , τ ∥ a n ∥ 2 2 = ∫ R ρ s , τ ( ξ ) ∣ a n ^ ( ξ ) ∣ 2 d ξ . (93) q_s[a_n\phi_{s,\tau}]-\Lambda_{s,\tau}\|a_n\|_2^2=\int_{\mathbb{R}}\rho_{s,\tau}(\xi)|\widehat{a_n}(\xi)|^2\,d\xi. \tag*{(93)} q s [ a n ϕ s , τ ] − Λ s , τ ∥ a n ∥ 2 2 = ∫ R ρ s , τ ( ξ ) ∣ a n ( ξ ) ∣ 2 d ξ . (93)
By (16) and ∥ a n ∥ 2 → 1 \lVert a_n\rVert_2 \to1 ∥ a n ∥ 2 → 1 , the functions a n ϕ s , τ a_n\phi_{s,\tau} a n ϕ s , τ are bounded in X s \mathcal{X}_s X s . The form Cauchy–Schwarz inequality gives
∣ q s [ w n ] − q s [ a n ϕ s , τ ] ∣ ≤ q s [ r n ] + 2 q s [ a n ϕ s , τ ] 1 / 2 q s [ r n ] 1 / 2 = o ( 1 ) . \lvert q_s[w_n]-q_s[a_n\phi_{s,\tau}]\rvert\le q_s[r_n]+2q_s[a_n\phi_{s,\tau}]^{1/2}q_s[r_n]^{1/2}=o(1). ∣ q s [ w n ] − q s [ a n ϕ s , τ ]∣ ≤ q s [ r n ] + 2 q s [ a n ϕ s , τ ] 1/2 q s [ r n ] 1/2 = o ( 1 ) .
Equation (93) now proves (1). □
Corollary 6.9 (No compact extremizing sequence). Let { u n } \{u_n\} { u n } be normalized and extremizing, and let w n w_n w n be its logarithmic polar transform. For every R > 0 R>0 R > 0 ,
sup t 0 ∈ R ∫ t 0 − R t 0 + R ∫ 0 π m s , τ ( θ ) ∣ w n ( t , θ ) ∣ 2 d θ d t ⟶ 0. \sup_{t_0\in\mathbb{R}}\int_{t_0-R}^{t_0+R}\int_0^\pi m_{s,\tau}(\theta)\lvert w_n(t,\theta)\rvert^2\,d\theta\,dt\longrightarrow0. t 0 ∈ R sup ∫ t 0 − R t 0 + R ∫ 0 π m s , τ ( θ ) ∣ w n ( t , θ ) ∣ 2 d θ d t ⟶ 0.
Consequently, no choice of normalized dilations of { u n } \{u_n\} { u n } admits a strongly convergent subsequence in the remainder norm.
Proof. Use the decomposition w n = a n ϕ s , τ + r n w_n=a_n\phi_{s,\tau}+r_n w n = a n ϕ s , τ + r n from Theorem 1.5 . We have ∥ r n ∥ L 2 ( R ; H s , τ ) → 0 \lVert r_n\rVert_{L^2(\mathbb{R};H_{s,\tau})}\to0 ∥ r n ∥ L 2 ( R ; H s , τ ) → 0 and ∥ a n ∥ 2 → 1 \lVert a_n\rVert_2\to1 ∥ a n ∥ 2 → 1 . For δ > 0 \delta>0 δ > 0 , split a n ^ = 1 { ∣ ξ ∣ ≤ δ } a n ^ + 1 { ∣ ξ ∣ > δ } a n ^ \widehat{a_n}=\mathbf{1}_{\{|\xi|\le\delta\}}\widehat{a_n}+\mathbf{1}_{\{|\xi|>\delta\}}\widehat{a_n} a n = 1 { ∣ ξ ∣ ≤ δ } a n + 1 { ∣ ξ ∣ > δ } a n and denote the inverse transforms by a n l o w a_n^{\mathrm{low}} a n low and a n h i g h a_n^{\mathrm{high}} a n high . Cauchy–Schwarz in frequency gives
∥ a n l o w ∥ ∞ 2 ≤ C δ ∥ a n ∥ 2 2 . \lVert a_n^{\mathrm{low}}\rVert_\infty^2\le C\delta\lVert a_n\rVert_2^2. ∥ a n low ∥ ∞ 2 ≤ C δ ∥ a n ∥ 2 2 .
Since ρ s , τ \rho_{s,\tau} ρ s , τ is continuous, positive away from zero, and coercive at infinity,
∥ a n h i g h ∥ 2 2 ≤ ( inf ∣ ξ ∣ ≥ δ ρ s , τ ( ξ ) ) − 1 ∫ R ρ s , τ ( ξ ) ∣ a n ^ ( ξ ) ∣ 2 d ξ ⟶ 0. \lVert a_n^{\mathrm{high}}\rVert_2^2\le\left(\inf_{|\xi|\ge\delta}\rho_{s,\tau}(\xi)\right)^{-1}\int_{\mathbb{R}}\rho_{s,\tau}(\xi)\lvert\widehat{a_n}(\xi)\rvert^2\,d\xi\longrightarrow0. ∥ a n high ∥ 2 2 ≤ ( ∣ ξ ∣ ≥ δ inf ρ s , τ ( ξ ) ) − 1 ∫ R ρ s , τ ( ξ ) ∣ a n ( ξ ) ∣ 2 d ξ ⟶ 0.
It follows that
lim sup n → ∞ sup t 0 ∈ R ∫ t 0 − R t 0 + R ∣ a n ( t ) ∣ 2 d t ≤ C R δ . \limsup_{n\to\infty}\sup_{t_0\in\mathbb{R}}\int_{t_0-R}^{t_0+R}\lvert a_n(t)\rvert^2\,dt\le CR\delta. n → ∞ lim sup t 0 ∈ R sup ∫ t 0 − R t 0 + R ∣ a n ( t ) ∣ 2 d t ≤ C R δ .
Let δ ↓ 0 \delta\downarrow0 δ ↓ 0 and use ∥ ϕ s , τ ∥ H s , τ = 1 \lVert\phi_{s,\tau}\rVert_{H_{s,\tau}}=1 ∥ ϕ s , τ ∥ H s , τ = 1 and ∥ r n ∥ L 2 ( R ; H s , τ ) → 0 \lVert r_n\rVert_{L^2(\mathbb{R};H_{s,\tau})}\to0 ∥ r n ∥ L 2 ( R ; H s , τ ) → 0 . This proves local vanishing. Normalized dilations translate t t t . A nonzero strong limit after such translations would have positive mass on a bounded interval, which is impossible. □
Remark 6.10. Theorem 1.5 fixes the limiting angular profile and controls the error in the energy norm. The remaining loss of compactness lies in the logarithmic radial amplitude.
Higher-dimensional splitting and extremizing sequences
We now set p = 2 p=2 p = 2 and retain the notation x = ( z , X ) x=(z,X) x = ( z , X ) and n = d − 2 n=d-2 n = d − 2 . Let u ^ ( ζ , X ) \widehat{u}(\zeta,X) u ( ζ , X ) denote the unitary Fourier transform in z z z . For f ∈ C c ∞ ( R + 2 ) f\in C^\infty_c(\mathbb{R}^2_+) f ∈ C c ∞ ( R + 2 ) , set
E d , s , ζ [ f ] = ∫ R + 2 ∫ R + 2 ∫ R n ∣ f ( X ) − e − i ζ ⋅ h f ( Y ) ∣ 2 ( ∣ X − Y ∣ 2 + ∣ h ∣ 2 ) ( d + 2 s ) / 2 d h d X d Y (94) \mathcal{E}_{d,s,\zeta}[f]=\int_{\mathbb{R}^2_+}\int_{\mathbb{R}^2_+}\int_{\mathbb{R}^n}\frac{\lvert f(X)-e^{-i\zeta\cdot h}f(Y)\rvert^2}{\left(\lvert X-Y\rvert^2+\lvert h\rvert^2\right)^{(d+2s)/2}}\,dh\,dX\,dY \tag*{(94)} E d , s , ζ [ f ] = ∫ R + 2 ∫ R + 2 ∫ R n ( ∣ X − Y ∣ 2 + ∣ h ∣ 2 ) ( d + 2 s ) /2 ∣ f ( X ) − e − i ζ ⋅ h f ( Y ) ∣ 2 d h d X d Y (94)
Lemma 7.1 (Transverse Fourier identity). For every ζ ∈ R n \zeta\in\mathbb{R}^n ζ ∈ R n , there is a nonnegative quadratic form T d , s , ζ \mathcal{T}_{d,s,\zeta} T d , s , ζ such that
E d , s , ζ [ f ] = A d , s [ f ] s , + 2 + T d , s , ζ [ f ] . (95) \mathcal{E}_{d,s,\zeta}[f]=\mathcal{A}_{d,s}[f]_{s,+}^2+\mathcal{T}_{d,s,\zeta}[f]. \tag*{(95)} E d , s , ζ [ f ] = A d , s [ f ] s , + 2 + T d , s , ζ [ f ] . (95) It satisfies
0 ≤ T d , s , ζ [ f ] ≤ C d , s ∣ ζ ∣ 2 s ∥ f ∥ L 2 ( R + 2 ) 2 . (96) 0\le\mathcal{T}_{d,s,\zeta}[f]\le C_{d,s}\lvert\zeta\rvert^{2s}\lVert f\rVert_{L^2(\mathbb{R}^2_+)}^2. \tag*{(96)} 0 ≤ T d , s , ζ [ f ] ≤ C d , s ∣ ζ ∣ 2 s ∥ f ∥ L 2 ( R + 2 ) 2 . (96) If ζ ≠ 0 \zeta\ne0 ζ = 0 , then T d , s , ζ [ f ] = 0 \mathcal{T}_{d,s,\zeta}[f]=0 T d , s , ζ [ f ] = 0 only when f = 0 f=0 f = 0 . Moreover,
H d , s [ u ] = ∫ R n { A d , s H s [ u ^ ( ζ , ⋅ ) ] + T d , s , ζ [ u ^ ( ζ , ⋅ ) ] } d ζ . (97) \mathcal{H}_{d,s}[u]=\int_{\mathbb{R}^n}\left\{\mathcal{A}_{d,s}\mathcal{H}_s[\widehat{u}(\zeta,\cdot)]+\mathcal{T}_{d,s,\zeta}[\widehat{u}(\zeta,\cdot)]\right\}\,d\zeta. \tag*{(97)} H d , s [ u ] = ∫ R n { A d , s H s [ u ( ζ , ⋅ )] + T d , s , ζ [ u ( ζ , ⋅ )] } d ζ . (97) :::`
Proof. Put ν = ( d + 2 s ) / 2 \nu=(d+2s)/2 ν = ( d + 2 s ) /2 . For a > 0 a>0 a > 0 , define
ℓ a ( ζ ) = ∫ R n e − i h ⋅ ζ ( ∣ h ∣ 2 + a 2 ) ν d h . \ell_a(\zeta)=\int_{\mathbb{R}^n}\frac{e^{-ih\cdot\zeta}}{(|h|^2+a^2)^\nu}\,dh. ℓ a ( ζ ) = ∫ R n ( ∣ h ∣ 2 + a 2 ) ν e − ih ⋅ ζ d h .
The Laplace representation and the Gaussian Fourier transform give
ℓ a ( ζ ) = π n / 2 Γ ( ν ) ∫ 0 ∞ t s exp ( − a 2 t − ∣ ζ ∣ 2 4 t ) d t . (98) \ell_a(\zeta)=\frac{\pi^{n/2}}{\Gamma(\nu)}\int_0^\infty t^s\exp\left(-a^2t-\frac{|\zeta|^2}{4t}\right)\,dt. \tag*{(98)} ℓ a ( ζ ) = Γ ( ν ) π n /2 ∫ 0 ∞ t s exp ( − a 2 t − 4 t ∣ ζ ∣ 2 ) d t . (98)
In particular,
ℓ a ( 0 ) = π n / 2 Γ ( 1 + s ) Γ ( ν ) a − 2 − 2 s = A d , s a − 2 − 2 s . \ell_a(0)=\frac{\pi^{n/2}\Gamma(1+s)}{\Gamma(\nu)}a^{-2-2s}=A_{d,s}a^{-2-2s}. ℓ a ( 0 ) = Γ ( ν ) π n /2 Γ ( 1 + s ) a − 2 − 2 s = A d , s a − 2 − 2 s .
Plancherel’s theorem applied to the z z z variable gives
∫ R + d ∫ R + d ∣ u ( x ) − u ( y ) ∣ 2 ∣ x − y ∣ d + 2 s d x d y = ∫ R n E d , s , ζ [ u ^ ( ζ , ⋅ ) ] d ζ . \int_{\mathbb{R}^d_+}\int_{\mathbb{R}^d_+}\frac{|u(x)-u(y)|^2}{|x-y|^{d+2s}}\,dx\,dy=\int_{\mathbb{R}^n}\mathcal{E}_{d,s,\zeta}[\widehat{u}(\zeta,\cdot)]\,d\zeta. ∫ R + d ∫ R + d ∣ x − y ∣ d + 2 s ∣ u ( x ) − u ( y ) ∣ 2 d x d y = ∫ R n E d , s , ζ [ u ( ζ , ⋅ )] d ζ .
To justify this identity, first restrict to ∣ X − Y ∣ > ϵ |X-Y|>\epsilon ∣ X − Y ∣ > ϵ . The kernel in h h h is then integrable, and Fubini and Plancherel apply. Monotone convergence as ϵ ↓ 0 \epsilon\downarrow0 ϵ ↓ 0 gives the full positive energy.
Expansion of the square in (94) yields
E d , s , ζ [ f ] = ∫ R + 2 ∫ R + 2 { ℓ ∣ X − Y ∣ ( 0 ) ( ∣ f ( X ) ∣ 2 + ∣ f ( Y ) ∣ 2 ) − 2 ℓ ∣ X − Y ∣ ( ζ ) Re ( f ( X ) f ( Y ) ‾ ) } d X d Y . \mathcal{E}_{d,s,\zeta}[f]=\int_{\mathbb{R}^2_+}\int_{\mathbb{R}^2_+}\left\{\ell_{|X-Y|}(0)\left(|f(X)|^2+|f(Y)|^2\right)-2\ell_{|X-Y|}(\zeta)\operatorname{Re}\left(f(X)\overline{f(Y)}\right)\right\}\,dX\,dY. E d , s , ζ [ f ] = ∫ R + 2 ∫ R + 2 { ℓ ∣ X − Y ∣ ( 0 ) ( ∣ f ( X ) ∣ 2 + ∣ f ( Y ) ∣ 2 ) − 2 ℓ ∣ X − Y ∣ ( ζ ) Re ( f ( X ) f ( Y ) ) } d X d Y .
With the same cutoff and ℓ ∣ X − Y ∣ ( 0 ) = A d , s ∣ X − Y ∣ − 2 − 2 s \ell_{|X-Y|}(0)=A_{d,s}|X-Y|^{-2-2s} ℓ ∣ X − Y ∣ ( 0 ) = A d , s ∣ X − Y ∣ − 2 − 2 s , subtraction of the zero fiber gives
T d , s , ζ [ f ] = 2 ∫ R + 2 ∫ R + 2 L ζ ( X − Y ) Re ( f ( X ) f ( Y ) ‾ ) d X d Y , (99) \mathcal{T}_{d,s,\zeta}[f]=2\int_{\mathbb{R}^2_+}\int_{\mathbb{R}^2_+}L_\zeta(X-Y)\operatorname{Re}\left(f(X)\overline{f(Y)}\right)\,dX\,dY, \tag*{(99)} T d , s , ζ [ f ] = 2 ∫ R + 2 ∫ R + 2 L ζ ( X − Y ) Re ( f ( X ) f ( Y ) ) d X d Y , (99)
where L ζ ( Z ) = ℓ ∣ Z ∣ ( 0 ) − ℓ ∣ Z ∣ ( ζ ) L_\zeta(Z)=\ell_{|Z|}(0)-\ell_{|Z|}(\zeta) L ζ ( Z ) = ℓ ∣ Z ∣ ( 0 ) − ℓ ∣ Z ∣ ( ζ ) .
By (98) , L ζ ≥ 0 L_\zeta\ge0 L ζ ≥ 0 . The bound 1 − cos t ≤ min { t 2 / 2 , 2 } 1-\cos t\le\min\{t^2/2,2\} 1 − cos t ≤ min { t 2 /2 , 2 } gives
0 ≤ L ζ ( Z ) ≤ C d , s min { ∣ ζ ∣ 2 ∣ Z ∣ − 2 s , ∣ Z ∣ − 2 − 2 s } . 0\le L_\zeta(Z)\le C_{d,s}\min\left\{|\zeta|^2|Z|^{-2s},|Z|^{-2-2s}\right\}. 0 ≤ L ζ ( Z ) ≤ C d , s min { ∣ ζ ∣ 2 ∣ Z ∣ − 2 s , ∣ Z ∣ − 2 − 2 s } .
For this estimate, scale h = ∣ Z ∣ r h=|Z|r h = ∣ Z ∣ r in the defining integral; both comparison integrals are finite. Since 0 < s < 1 0<s<1 0 < s < 1 , we have L ζ ∈ L 1 ( R 2 ) L_\zeta\in L^1(\mathbb{R}^2) L ζ ∈ L 1 ( R 2 ) . Thus (99) is absolutely convergent for f ∈ L 2 ( R + 2 ) f\in L^2(\mathbb{R}^2_+) f ∈ L 2 ( R + 2 ) and is the limit of the cutoff calculation.
Extend f f f by zero to R 2 \mathbb{R}^2 R 2 . The multiplier can be computed directly from (98) . With the unitary Fourier transform in two variables,
F Z → κ ( e − t ∣ Z ∣ 2 ) = 1 2 t e − ∣ κ ∣ 2 / ( 4 t ) . \mathcal{F}_{Z\to\kappa}\left(e^{-t|Z|^2}\right)=\frac{1}{2t}e^{-|\kappa|^2/(4t)}. F Z → κ ( e − t ∣ Z ∣ 2 ) = 2 t 1 e − ∣ κ ∣ 2 / ( 4 t ) .
For a ≥ 0 a\ge0 a ≥ 0 and b > 0 b>0 b > 0 , the change of variables r = 1 / t r=1/t r = 1/ t and integration by parts give
∫ 0 ∞ t s − 1 ( e − a / t − e − ( a + b ) / t ) d t = ∫ 0 ∞ r − s − 1 ( e − a r − e − ( a + b ) r ) d r = Γ ( 1 − s ) s { ( a + b ) s − a s } . \int_0^\infty t^{s-1}\left(e^{-a/t}-e^{-(a+b)/t}\right)\,dt=\int_0^\infty r^{-s-1}\left(e^{-ar}-e^{-(a+b)r}\right)\,dr
=\frac{\Gamma(1-s)}{s}\left\{(a+b)^s-a^s\right\}. ∫ 0 ∞ t s − 1 ( e − a / t − e − ( a + b ) / t ) d t = ∫ 0 ∞ r − s − 1 ( e − a r − e − ( a + b ) r ) d r = s Γ ( 1 − s ) { ( a + b ) s − a s } .
The boundary terms vanish: the difference is O ( r ) O(r) O ( r ) at zero and is bounded at infinity. The formula remains valid at a = 0 a=0 a = 0 because 0 < s < 1 0<s<1 0 < s < 1 . The L 1 L^1 L 1 bound for L ζ L_\zeta L ζ permits Fubini in the Gaussian representation. Taking a = ∣ κ ∣ 2 / 4 a=|\kappa|^2/4 a = ∣ κ ∣ 2 /4 and b = ∣ ζ ∣ 2 / 4 b=|\zeta|^2/4 b = ∣ ζ ∣ 2 /4 gives
F Z → κ L ζ = c d , s ′ ( ( ∣ κ ∣ 2 + ∣ ζ ∣ 2 ) s − ∣ κ ∣ 2 s ) , c d , s ′ > 0. (100) \mathcal{F}_{Z\to\kappa}L_\zeta=c'_{d,s}\left((|\kappa|^2+|\zeta|^2)^s-|\kappa|^{2s}\right),\qquad c'_{d,s}>0. \tag*{(100)} F Z → κ L ζ = c d , s ′ ( ( ∣ κ ∣ 2 + ∣ ζ ∣ 2 ) s − ∣ κ ∣ 2 s ) , c d , s ′ > 0. (100)
The cutoff calculation gives (100) in distributions. Since L ζ ∈ L 1 L_\zeta\in L^1 L ζ ∈ L 1 , Plancherel in X X X gives
T d , s , ζ [ f ] = c d , s ′ ′ ∫ R 2 ( ( ∣ κ ∣ 2 + ∣ ζ ∣ 2 ) s − ∣ κ ∣ 2 s ) ∣ F X f ~ ( κ ) ∣ 2 d κ . \mathcal{T}_{d,s,\zeta}[f]=c''_{d,s}\int_{\mathbb{R}^2}\left((|\kappa|^2+|\zeta|^2)^s-|\kappa|^{2s}\right)|\mathcal{F}_X\widetilde{f}(\kappa)|^2\,d\kappa. T d , s , ζ [ f ] = c d , s ′′ ∫ R 2 ( ( ∣ κ ∣ 2 + ∣ ζ ∣ 2 ) s − ∣ κ ∣ 2 s ) ∣ F X f ( κ ) ∣ 2 d κ .
where f ~ \widetilde{f} f is the zero extension. The multiplier is nonnegative. It is strictly positive if ζ ≠ 0 \zeta\ne0 ζ = 0 . The concavity inequality ( a + b ) s − a s ≤ b s (a+b)^s-a^s \le b^s ( a + b ) s − a s ≤ b s proves (96) . It also proves the strict zero statement and (95) .
By (21) , D d , s , 2 = A d , s D 2 , s , 2 \mathcal{D}_{d,s,2}=A_{d,s}\mathcal{D}_{2,s,2} D d , s , 2 = A d , s D 2 , s , 2 . The Hardy term also splits by Plancherel. Subtracting it and integrating (95) in ζ \zeta ζ proves (97) .
Proof of Theorem 1.1 . Lemma 2.1 proves (2) for d ≥ 3 d\ge3 d ≥ 3 . For d = 2 d=2 d = 2 the identity is immediate. Fix p > 1 p>1 p > 1 and τ > 0 \tau>0 τ > 0 . Choose
0 < σ < min { τ , τ ∗ ( s , p ) } , τ ∗ ( s , p ) = { 2 min { s p , 1 } , 1 < p < 2 , p ( s p + 1 ) 2 ( p − 1 ) , p ≥ 2. 0<\sigma<\min\{\tau,\tau_*(s,p)\},\qquad\tau_*(s,p)=\begin{cases}2\min\{sp,1\},&1<p<2,\\ \dfrac{p(sp+1)}{2(p-1)},&p\ge2.\end{cases} 0 < σ < min { τ , τ ∗ ( s , p )} , τ ∗ ( s , p ) = ⎩ ⎨ ⎧ 2 min { s p , 1 } , 2 ( p − 1 ) p ( s p + 1 ) , 1 < p < 2 , p ≥ 2.
The theorem of Kijaczko and Szczukiewicz gives Λ d , s , p , σ > 0 \Lambda_{d,s,p,\sigma}>0 Λ d , s , p , σ > 0 . Since R d , s , p , τ ≤ R d , s , p , σ \mathcal{R}_{d,s,p,\tau}\le\mathcal{R}_{d,s,p,\sigma} R d , s , p , τ ≤ R d , s , p , σ , one has Λ d , s , p , τ ≥ Λ d , s , p , σ > 0 \Lambda_{d,s,p,\tau}\ge\Lambda_{d,s,p,\sigma}>0 Λ d , s , p , τ ≥ Λ d , s , p , σ > 0 . For p = 2 p=2 p = 2 , the sharp identity follows from the same reduction. The planar non-attainment is Theorem 1.2 . Suppose d ≥ 3 d\ge3 d ≥ 3 . Plancherel gives
R d , s , τ [ u ] = ∫ R d − 2 R s , τ [ u ^ ( ζ , ⋅ ) ] d ζ . (101) \mathcal{R}_{d,s,\tau}[u]=\int_{\mathbb{R}^{d-2}}\mathcal{R}_{s,\tau}[\widehat{u}(\zeta,\cdot)]\,d\zeta. \tag*{(101)} R d , s , τ [ u ] = ∫ R d − 2 R s , τ [ u ( ζ , ⋅ )] d ζ . (101)
We first show that the energy completion embeds into the weighted remainder space. Put ρ ( x ) = X 2 τ − 2 s ∣ X ∣ − τ \rho(x)=X_2^{\tau-2s}|X|^{-\tau} ρ ( x ) = X 2 τ − 2 s ∣ X ∣ − τ , ω ( x ) = x d s − 1 / 2 \omega(x)=x_d^{s-1/2} ω ( x ) = x d s − 1/2 , and d ν = ∣ x − y ∣ − d − 2 s d x d y d\nu=|x-y|^{-d-2s}\,dx\,dy d ν = ∣ x − y ∣ − d − 2 s d x d y on R + d × R + d \mathbb{R}_+^d\times\mathbb{R}_+^d R + d × R + d . Define on the core
( S u ) ( x , y ) = ω ( x ) ω ( y ) ( u ( x ) ω ( x ) − u ( y ) ω ( y ) ) . (Su)(x,y)=\sqrt{\omega(x)\omega(y)}\left(\frac{u(x)}{\omega(x)}-\frac{u(y)}{\omega(y)}\right). ( S u ) ( x , y ) = ω ( x ) ω ( y ) ( ω ( x ) u ( x ) − ω ( y ) u ( y ) ) .
The p = 2 p=2 p = 2 identity in Lemma C.1 , applied to real and imaginary parts, gives H d , s [ u ] = ∥ S u ∥ L 2 ( d ν ) 2 \mathcal{H}_{d,s}[u]=\|Su\|_{L^2(d\nu)}^2 H d , s [ u ] = ∥ S u ∥ L 2 ( d ν ) 2 . Suppose u j → 0 u_j\to0 u j → 0 in L 2 ( ρ d x ) L^2(\rho\,dx) L 2 ( ρ d x ) and S u j → F Su_j\to F S u j → F in L 2 ( d ν ) L^2(d\nu) L 2 ( d ν ) . Since ρ \rho ρ is positive and finite almost everywhere, subsequences give u j → 0 u_j\to0 u j → 0 Lebesgue-almost everywhere and S u j → F Su_j\to F S u j → F d ν d\nu d ν -almost everywhere. The measure d ν d\nu d ν is σ \sigma σ -finite and absolutely continuous off the null diagonal, so the displayed difference formula gives F = 0 F=0 F = 0 . Thus S S S is closable, and the completion under ( H d , s + R d , s , τ ) 1 / 2 (\mathcal{H}_{d,s}+\mathcal{R}_{d,s,\tau})^{1/2} ( H d , s + R d , s , τ ) 1/2 embeds faithfully into L 2 ( ρ d x ) L^2(\rho\,dx) L 2 ( ρ d x ) .
We now extend (101) and (97) to this closed energy space. Let { u j } \{u_j\} { u j } be a Cauchy sequence in the core. Apply (97) and (101) to u j − u k u_j-u_k u j − u k . Since every term is nonnegative,
min { A d , s , 1 } ∫ R d − 2 ( H s [ u j ^ − u k ^ ] + R s , τ [ u j ^ − u k ^ ] ) d ζ + ∫ R d − 2 T d , s , ζ [ u j ^ − u k ^ ] d ζ ≤ H d , s [ u j − u k ] + R d , s , τ [ u j − u k ] . \min\{A_{d,s},1\}\int_{\mathbb{R}^{d-2}}\left(\mathcal{H}_s[\widehat{u_j}-\widehat{u_k}]+\mathcal{R}_{s,\tau}[\widehat{u_j}-\widehat{u_k}]\right)\,d\zeta
+\int_{\mathbb{R}^{d-2}}\mathcal{T}_{d,s,\zeta}[\widehat{u_j}-\widehat{u_k}]\,d\zeta
\le\mathcal{H}_{d,s}[u_j-u_k]+\mathcal{R}_{d,s,\tau}[u_j-u_k]. min { A d , s , 1 } ∫ R d − 2 ( H s [ u j − u k ] + R s , τ [ u j − u k ] ) d ζ + ∫ R d − 2 T d , s , ζ [ u j − u k ] d ζ ≤ H d , s [ u j − u k ] + R d , s , τ [ u j − u k ] .
The Fourier slices converge in the direct integral of the planar closed energy space, and the transverse terms converge in their seminorm. The partial Fourier transform is unitary on
L 2 ( R d − 2 ; L 2 ( R + 2 , X 2 τ − 2 s ∣ X ∣ − τ d X ) ) . L^2\left(\mathbb{R}^{d-2};L^2\left(\mathbb{R}_+^2,X_2^{\tau-2s}|X|^{-\tau}\,dX\right)\right). L 2 ( R d − 2 ; L 2 ( R + 2 , X 2 τ − 2 s ∣ X ∣ − τ d X ) ) .
which is the remainder space. It identifies the slice limit with the limit of u j u_j u j . Both splitting identities pass to the limit; the transverse term is defined by seminorm completion.
Suppose that a nonzero completed energy function attains the sharp coefficient. Subtract the sharp right-hand side from (97) . Equations (101) and Theorem 1.2 give
∫ R d − 2 { A d , s ( H s [ u ^ ( ζ ) ] − Λ s , τ R s , τ [ u ^ ( ζ ) ] ) + T d , s , ζ [ u ^ ( ζ ) ] } d ζ = 0. \int_{\mathbb{R}^{d-2}}\left\{A_{d,s}\left(\mathcal{H}_s[\widehat{u}(\zeta)]-\Lambda_{s,\tau}\mathcal{R}_{s,\tau}[\widehat{u}(\zeta)]\right)+\mathcal{T}_{d,s,\zeta}[\widehat{u}(\zeta)]\right\}\,d\zeta=0. ∫ R d − 2 { A d , s ( H s [ u ( ζ )] − Λ s , τ R s , τ [ u ( ζ )] ) + T d , s , ζ [ u ( ζ )] } d ζ = 0.
Both summands are nonnegative. By Theorem 1.2 , the first can vanish only for the zero planar slice. Hence u ^ ( ζ , ⋅ ) = 0 \widehat{u}(\zeta,\cdot)=0 u ( ζ , ⋅ ) = 0 almost everywhere. Plancherel in the remainder space gives u = 0 u=0 u = 0 , a contradiction.
Corollary 7.2 (Lifted stability). Let d ≥ 3 d\ge3 d ≥ 3 . For every u ∈ C c ∞ ( R + d ) u\in C^\infty_c(\mathbb{R}_+^d) u ∈ C c ∞ ( R + d ) ,
H d , s [ u ] − Λ d , s , τ R d , s , τ [ u ] ≥ A d , s γ s , τ ∫ R d − 2 Dist s , τ ( u ^ ( ζ , ⋅ ) ) 2 d ζ + ∫ R d − 2 T d , s , ζ [ u ^ ( ζ ) ] d ζ . \mathcal{H}_{d,s}[u]-\Lambda_{d,s,\tau}\mathcal{R}_{d,s,\tau}[u]\ge A_{d,s}\gamma_{s,\tau}\int_{\mathbb{R}^{d-2}}\operatorname{Dist}_{s,\tau}(\widehat{u}(\zeta,\cdot))^2\,d\zeta+\int_{\mathbb{R}^{d-2}}\mathcal{T}_{d,s,\zeta}[\widehat{u}(\zeta)]\,d\zeta. H d , s [ u ] − Λ d , s , τ R d , s , τ [ u ] ≥ A d , s γ s , τ ∫ R d − 2 Dist s , τ ( u ( ζ , ⋅ ) ) 2 d ζ + ∫ R d − 2 T d , s , ζ [ u ( ζ )] d ζ .
The estimate extends to the closed energy space, with the transverse term defined by completion as in the proof of Theorem 1.1 .
Proof. Use (3) and (97) , and integrate Theorem 1.3 over the planar Fourier slices. The distance seminorm is bounded by the planar energy, so it is continuous under the direct-integral convergence above. The transverse seminorm is continuous by its definition. This also proves the estimate on the closed space. □
Theorem 7.3 (Joint Fourier–Mellin energy decomposition). Let d ≥ 3 d \ge3 d ≥ 3 , 0 < s < 1 0 < s < 1 0 < s < 1 , and τ > 0 \tau> 0 τ > 0 . For a function in the closed energy space, let w ζ ( t , θ ) w_\zeta(t,\theta) w ζ ( t , θ ) be the planar logarithmic polar transform of u ^ ( ζ , X ) \widehat{u}(\zeta,X) u ( ζ , X ) . Apply the fixed decomposition (90) to each slice:
w ζ ( t , θ ) = a ζ ( t ) ϕ s , τ ( θ ) + r ζ ( t , θ ) . w_\zeta(t,\theta)=a_\zeta(t)\phi_{s,\tau}(\theta)+r_\zeta(t,\theta). w ζ ( t , θ ) = a ζ ( t ) ϕ s , τ ( θ ) + r ζ ( t , θ ) . Define
S d , s , τ [ u ] = ∫ R d − 2 { ∥ r ζ ∥ X s 2 + ∫ R ρ s , τ ( ξ ) ∣ a ζ ^ ( ξ ) ∣ 2 d ξ } d ζ + ∫ R d − 2 T d , s , ζ [ u ^ ( ζ , ⋅ ) ] d ζ . S_{d,s,\tau}[u]=\int_{\mathbb{R}^{d-2}}\left\{\|r_\zeta\|_{\mathcal{X}_s}^2+\int_{\mathbb{R}}\rho_{s,\tau}(\xi)|\widehat{a_\zeta}(\xi)|^2\,d\xi\right\}\,d\zeta+\int_{\mathbb{R}^{d-2}}\mathcal{T}_{d,s,\zeta}[\widehat{u}(\zeta,\cdot)]\,d\zeta. S d , s , τ [ u ] = ∫ R d − 2 { ∥ r ζ ∥ X s 2 + ∫ R ρ s , τ ( ξ ) ∣ a ζ ( ξ ) ∣ 2 d ξ } d ζ + ∫ R d − 2 T d , s , ζ [ u ( ζ , ⋅ )] d ζ . There are c , C > 0 c,C>0 c , C > 0 such that
c S d , s , τ [ u ] ≤ H d , s [ u ] − Λ d , s , τ R d , s , τ [ u ] ≤ C S d , s , τ [ u ] . (102) cS_{d,s,\tau}[u]\le\mathcal{H}_{d,s}[u]-\Lambda_{d,s,\tau}\mathcal{R}_{d,s,\tau}[u]\le CS_{d,s,\tau}[u]. \tag*{(102)} c S d , s , τ [ u ] ≤ H d , s [ u ] − Λ d , s , τ R d , s , τ [ u ] ≤ C S d , s , τ [ u ] . (102) Consequently, a sequence with R d , s , τ [ u n ] = 1 \mathcal{R}_{d,s,\tau}[u_n]=1 R d , s , τ [ u n ] = 1 is extremizing if and only if all three nonnegative terms defining S d , s , τ [ u n ] S_{d,s,\tau}[u_n] S d , s , τ [ u n ] tend to zero. In that case,
∫ R d − 2 ∥ a n , ζ ∥ L 2 ( R ) 2 d ζ ⟶ 1. \int_{\mathbb{R}^{d-2}}\|a_{n,\zeta}\|_{L^2(\mathbb{R})}^2\,d\zeta\longrightarrow1. ∫ R d − 2 ∥ a n , ζ ∥ L 2 ( R ) 2 d ζ ⟶ 1.
Proof. Subtract the sharp remainder in (97) and use (101) . The result is
H d , s [ u ] − Λ d , s , τ R d , s , τ [ u ] = A d , s ∫ R d − 2 { q s [ w ζ ] − Λ s , τ ∥ w ζ ∥ L 2 ( m s , τ ) 2 } d ζ + ∫ R d − 2 T d , s , ζ [ u ^ ( ζ , ⋅ ) ] d ζ . \mathcal{H}_{d,s}[u]-\Lambda_{d,s,\tau}\mathcal{R}_{d,s,\tau}[u]
=A_{d,s}\int_{\mathbb{R}^{d-2}}\left\{q_s[w_\zeta]-\Lambda_{s,\tau}\|w_\zeta\|_{L^2(m_{s,\tau})}^2\right\}\,d\zeta+\int_{\mathbb{R}^{d-2}}\mathcal{T}_{d,s,\zeta}[\widehat{u}(\zeta,\cdot)]\,d\zeta. H d , s [ u ] − Λ d , s , τ R d , s , τ [ u ] = A d , s ∫ R d − 2 { q s [ w ζ ] − Λ s , τ ∥ w ζ ∥ L 2 ( m s , τ ) 2 } d ζ + ∫ R d − 2 T d , s , ζ [ u ( ζ , ⋅ )] d ζ .
Integrate Proposition 6.8 in ζ \zeta ζ . The canonical maps w ↦ a w\mapsto a w ↦ a and w ↦ r w\mapsto r w ↦ r are bounded linear maps on the relevant form spaces by its proof; their low-frequency coefficients are measurable. They therefore preserve strong measurability of the Fourier slices. The two-sided planar estimate and the exact transverse splitting give (102) , first on the core and then by the completed identities. Nonnegativity proves the sequence criterion. Finally, in the Hilbert space L 2 ( R d − 2 × R ; H s , τ ) L^2(\mathbb{R}^{d-2}\times\mathbb{R};H_{s,\tau}) L 2 ( R d − 2 × R ; H s , τ ) ,
∣ ( ∫ ∥ a n , ζ ∥ 2 d ζ ) 1 / 2 − 1 ∣ ≤ ( ∫ ∥ r n , ζ ∥ L 2 ( m s , τ ) 2 d ζ ) 1 / 2 ⟶ 0. \left|\left(\int\|a_{n,\zeta}\|^2\,d\zeta\right)^{1/2}-1\right|
\le\left(\int\|r_{n,\zeta}\|_{L^2(m_{s,\tau})}^2\,d\zeta\right)^{1/2}\longrightarrow0. ( ∫ ∥ a n , ζ ∥ 2 d ζ ) 1/2 − 1 ≤ ( ∫ ∥ r n , ζ ∥ L 2 ( m s , τ ) 2 d ζ ) 1/2 ⟶ 0.
This proves the last assertion. □
Remark 7.4 (The higher-dimensional angular singularity). For the original two-plane remainder, the spherical weight is
m d , s , τ ( ω ) = ω d τ − 2 s ( ω d − 1 2 + ω d 2 ) − τ / 2 , ω ∈ S + d − 1 . m_{d,s,\tau}(\omega)=\omega_d^{\tau-2s}(\omega_{d-1}^2+\omega_d^2)^{-\tau/2},\qquad\omega\in S_+^{d-1}. m d , s , τ ( ω ) = ω d τ − 2 s ( ω d − 1 2 + ω d 2 ) − τ /2 , ω ∈ S + d − 1 . Put r = ( ω d − 1 2 + ω d 2 ) 1 / 2 r=(\omega_{d-1}^2+\omega_d^2)^{1/2} r = ( ω d − 1 2 + ω d 2 ) 1/2 and ω d = r sin ϑ \omega_d=r\sin\vartheta ω d = r sin ϑ . Then
m d , s , τ ( ω ) = r − 2 s ( sin ϑ ) τ − 2 s . m_{d,s,\tau}(\omega)=r^{-2s}(\sin\vartheta)^{\tau-2s}. m d , s , τ ( ω ) = r − 2 s ( sin ϑ ) τ − 2 s . The set
Σ = { ω ∈ S d − 1 : ω d − 1 = ω d = 0 } ≃ S d − 3 \Sigma=\{\omega\in S^{d-1}:\omega_{d-1}=\omega_d=0\}\simeq S^{d-3} Σ = { ω ∈ S d − 1 : ω d − 1 = ω d = 0 } ≃ S d − 3 has codimension two. In a local chart near Σ \Sigma Σ , write Z = ( z , y ) Z=(z,y) Z = ( z , y ) , with y ∈ R + 2 y\in\mathbb{R}_+^2 y ∈ R + 2 normal and z z z tangential to Σ \Sigma Σ . The chart metric and density are uniformly comparable to the Euclidean ones. Choose nonzero smooth functions χ \chi χ and η \eta η with compact support and
supp χ ⊂ { y : 1 < ∣ y ∣ < 2 , y 2 > c ∣ y ∣ } , c > 0. \operatorname{supp}\chi\subset\{y:1<|y|<2,\ y_2>c|y|\},\qquad c>0. supp χ ⊂ { y : 1 < ∣ y ∣ < 2 , y 2 > c ∣ y ∣ } , c > 0.
Set
ψ ε ( z , y ) = ε s − 1 η ( z ) χ ( y / ε ) . \psi_\varepsilon(z,y)=\varepsilon^{s-1}\eta(z)\chi(y/\varepsilon). ψ ε ( z , y ) = ε s − 1 η ( z ) χ ( y / ε ) .
For d = 3 d=3 d = 3 , the tangential factor and all tangential integrations below are omitted. Direct scaling gives
∫ m d , s , τ ∣ ψ ε ∣ 2 d ω ≲ 1 , ∥ ψ ε ∥ 2 2 = O ( ε 2 s ) . \int m_{d,s,\tau}|\psi_\varepsilon|^2\,d\omega\lesssim1,\qquad\|\psi_\varepsilon\|_2^2=O(\varepsilon^{2s}). ∫ m d , s , τ ∣ ψ ε ∣ 2 d ω ≲ 1 , ∥ ψ ε ∥ 2 2 = O ( ε 2 s ) .
Indeed, the first scaling factor is ε 2 s − 2 ε 2 − 2 s = 1 \varepsilon^{2s-2}\varepsilon^{2-2s}=1 ε 2 s − 2 ε 2 − 2 s = 1 . Writing the Fourier variables as ( k , ξ ) (k,\xi) ( k , ξ ) , we have
ψ ε ^ ( k , ξ ) = ε s + 1 η ^ ( k ) χ ^ ( ε ξ ) . \widehat{\psi_\varepsilon}(k,\xi)=\varepsilon^{s+1}\widehat{\eta}(k)\widehat{\chi}(\varepsilon\xi). ψ ε ( k , ξ ) = ε s + 1 η ( k ) χ ( ε ξ ) .
The change of variables ν = ε ξ \nu=\varepsilon\xi ν = ε ξ yields
∥ ψ ε ∥ H ˙ s 2 = C ∫ ( ε 2 ∣ k ∣ 2 + ∣ ν ∣ 2 ) s ∣ η ^ ( k ) ∣ 2 ∣ χ ^ ( ν ) ∣ 2 d k d ν ≤ C . \|\psi_\varepsilon\|_{\dot H^s}^2=C\int\left(\varepsilon^2|k|^2+|\nu|^2\right)^s|\widehat{\eta}(k)|^2|\widehat{\chi}(\nu)|^2\,dk\,d\nu\leq C. ∥ ψ ε ∥ H ˙ s 2 = C ∫ ( ε 2 ∣ k ∣ 2 + ∣ ν ∣ 2 ) s ∣ η ( k ) ∣ 2 ∣ χ ( ν ) ∣ 2 d k d ν ≤ C .
Here H ˙ s \dot H^s H ˙ s is the homogeneous Gagliardo seminorm in the chart. Extend the functions by zero outside it. Since ω d ≍ ∣ y ∣ ≍ ε \omega_d\asymp|y|\asymp\varepsilon ω d ≍ ∣ y ∣ ≍ ε on their support, the quantities
∥ ψ ε ∥ H ˙ s 2 + ∫ S + d − 1 ω d − 2 s ∣ ψ ε ∣ 2 d ω + ∥ ψ ε ∥ 2 2 \|\psi_\varepsilon\|_{\dot H^s}^2+\int_{\mathbb{S}^{d-1}_+}\omega_d^{-2s}|\psi_\varepsilon|^2\,d\omega+\|\psi_\varepsilon\|_2^2 ∥ ψ ε ∥ H ˙ s 2 + ∫ S + d − 1 ω d − 2 s ∣ ψ ε ∣ 2 d ω + ∥ ψ ε ∥ 2 2
are bounded. However, ψ ε → 0 \psi_\varepsilon\to0 ψ ε → 0 almost everywhere and their L 2 ( m d , s , τ d ω ) L^2(m_{d,s,\tau}\,d\omega) L 2 ( m d , s , τ d ω ) norms stay away from zero. Thus these local regularity and Hardy bounds do not give a compact weighted embedding. Moreover,
m d , s , τ m d , s , τ 0 = ( sin ϑ ) τ − τ 0 \frac{m_{d,s,\tau}}{m_{d,s,\tau_0}}=(\sin\vartheta)^{\tau-\tau_0} m d , s , τ 0 m d , s , τ = ( sin ϑ ) τ − τ 0
provides no small factor as r → 0 r\to0 r → 0 , unlike the planar endpoint estimate (57) . The planar compactness argument therefore does not extend to this spherical weight. The transverse Fourier splitting in Corollary 7.2 retains this additional channel.
Replacing the two-plane radius by ∣ x ∣ |x| ∣ x ∣ gives the weaker weight ω d τ − 2 s \omega_d^{\tau-2s} ω d τ − 2 s and removes the factor r − 2 s r^{-2s} r − 2 s . For the original remainder, Theorem 7.3 gives the sequence criterion in joint Fourier–Mellin energy variables. A physical-space profile decomposition would additionally have to resolve translations along the edge and their scales relative to planar dilations.
The local limit
For d ≥ 2 d\geq2 d ≥ 2 and p > 1 p>1 p > 1 , put
K d , p = 1 p ∫ S d − 1 ∣ ω d ∣ p d ω = 2 π ( d − 1 ) / 2 Γ ( ( p + 1 ) / 2 ) p Γ ( ( d + p ) / 2 ) . K_{d,p}=\frac{1}{p}\int_{\mathbb{S}^{d-1}}|\omega_d|^p\,d\omega=\frac{2\pi^{(d-1)/2}\Gamma((p+1)/2)}{p\Gamma((d+p)/2)}. K d , p = p 1 ∫ S d − 1 ∣ ω d ∣ p d ω = p Γ (( d + p ) /2 ) 2 π ( d − 1 ) /2 Γ (( p + 1 ) /2 ) .
Define
H d , 1 , p [ u ] = ∫ R + d ∣ ∇ u ∣ p d x − ( p − 1 p ) p ∫ R + d ∣ u ∣ p x d p d x , \mathcal{H}_{d,1,p}[u]=\int_{\mathbb{R}^d_+}|\nabla u|^p\,dx-\left(\frac{p-1}{p}\right)^p\int_{\mathbb{R}^d_+}\frac{|u|^p}{x_d^p}\,dx, H d , 1 , p [ u ] = ∫ R + d ∣∇ u ∣ p d x − ( p p − 1 ) p ∫ R + d x d p ∣ u ∣ p d x ,
R d , 1 , p , τ [ u ] = ∫ R + d x d τ − p ∣ u ( x ) ∣ p ( x d − 1 2 + x d 2 ) τ / 2 d x . \mathcal{R}_{d,1,p,\tau}[u]=\int_{\mathbb{R}^d_+}\frac{x_d^{\tau-p}|u(x)|^p}{(x_{d-1}^2+x_d^2)^{\tau/2}}\,dx. R d , 1 , p , τ [ u ] = ∫ R + d ( x d − 1 2 + x d 2 ) τ /2 x d τ − p ∣ u ( x ) ∣ p d x .
Proposition 8.1 (Limit on the smooth core). Let d ≥ 2 d\geq2 d ≥ 2 , p > 1 p>1 p > 1 , and τ > 0 \tau>0 τ > 0 . For every u ∈ C c ∞ ( R + d ) u\in C^\infty_c(\mathbb{R}^d_+) u ∈ C c ∞ ( R + d ) ,
lim s ↑ 1 ( 1 − s ) H d , s , p [ u ] = K d , p H d , 1 , p [ u ] \lim_{s\uparrow1}(1-s)\mathcal{H}_{d,s,p}[u]=K_{d,p}\mathcal{H}_{d,1,p}[u] s ↑ 1 lim ( 1 − s ) H d , s , p [ u ] = K d , p H d , 1 , p [ u ] and
lim s ↑ 1 R d , s , p , τ [ u ] = R d , 1 , p , τ [ u ] . \lim_{s\uparrow1}\mathcal{R}_{d,s,p,\tau}[u]=\mathcal{R}_{d,1,p,\tau}[u]. s ↑ 1 lim R d , s , p , τ [ u ] = R d , 1 , p , τ [ u ] . Consequently,
lim sup s ↑ 1 ( 1 − s ) Λ d , s , p , τ K d , p ≤ inf u ∈ C c ∞ ( R + d ) ∖ { 0 } H d , 1 , p [ u ] R d , 1 , p , τ [ u ] . \limsup_{s\uparrow1}\frac{(1-s)\Lambda_{d,s,p,\tau}}{K_{d,p}}\leq\inf_{u\in C^\infty_c(\mathbb{R}^d_+)\setminus\{0\}}\frac{\mathcal{H}_{d,1,p}[u]}{\mathcal{R}_{d,1,p,\tau}[u]}. s ↑ 1 lim sup K d , p ( 1 − s ) Λ d , s , p , τ ≤ u ∈ C c ∞ ( R + d ) ∖ { 0 } inf R d , 1 , p , τ [ u ] H d , 1 , p [ u ] .
Proof. The Bourgain–Brezis–Mironescu formula [3 ] (see also [20 formula (3)] ) gives
lim s ↑ 1 ( 1 − s ) [ u ] s , p , + p = K d , p ∫ R + d ∣ ∇ u ∣ p d x . \lim_{s\uparrow1}(1-s)[u]_{s,p,+}^{p}=K_{d,p}\int_{\mathbb{R}_{+}^{d}}|\nabla u|^{p}\,dx. s ↑ 1 lim ( 1 − s ) [ u ] s , p , + p = K d , p ∫ R + d ∣∇ u ∣ p d x .
The support of u u u has positive distance from the boundary. Thus the regional form has the same local contribution; its off-diagonal part vanishes after multiplication by 1 − s 1-s 1 − s .
For the Hardy constant, set in (20)
b s = s p − 1 p , ε s = p ( 1 − s ) . b_s=\frac{sp-1}{p},\qquad\varepsilon_s=p(1-s). b s = p s p − 1 , ε s = p ( 1 − s ) .
After r = 1 − t r=1-t r = 1 − t , its integral is
∫ 0 1 ∣ 1 − ( 1 − r ) b s r ∣ p r − 1 + ε s d r . \int_{0}^{1}\left|\frac{1-(1-r)^{b_s}}{r}\right|^{p}r^{-1+\varepsilon_s}\,dr. ∫ 0 1 r 1 − ( 1 − r ) b s p r − 1 + ε s d r .
The expression in absolute values is uniformly bounded for s s s near one and tends to ( p − 1 ) / p (p-1)/p ( p − 1 ) / p as ( s , r ) → ( 1 , 0 ) (s,r)\to(1,0) ( s , r ) → ( 1 , 0 ) . The Abelian limit
lim ε ↓ 0 ε ∫ 0 1 F ( r ) r − 1 + ε d r = F ( 0 ) \lim_{\varepsilon\downarrow0}\varepsilon\int_{0}^{1}F(r)r^{-1+\varepsilon}\,dr=F(0) ε ↓ 0 lim ε ∫ 0 1 F ( r ) r − 1 + ε d r = F ( 0 )
follows by splitting at δ \delta δ , using continuity at zero on ( 0 , δ ) (0,\delta) ( 0 , δ ) and boundedness on ( δ , 1 ) (\delta,1) ( δ , 1 ) . For the present family, put F s ( r ) = ∣ [ 1 − ( 1 − r ) b s ] / r ∣ p F_s(r)=\left|[1-(1-r)^{b_s}]/r\right|^p F s ( r ) = [ 1 − ( 1 − r ) b s ] / r p and F s ( 0 ) = ∣ b s ∣ p F_s(0)=|b_s|^p F s ( 0 ) = ∣ b s ∣ p . For every fixed δ > 0 \delta>0 δ > 0 ,
∣ ε s ∫ 0 1 F s ( r ) r − 1 + ε s d r − F s ( 0 ) ∣ ≤ sup 0 < r < δ ∣ F s ( r ) − F s ( 0 ) ∣ + 2 sup 0 < r < 1 ∣ F s ( r ) ∣ ( 1 − δ ε s ) . \left|\varepsilon_s\int_{0}^{1}F_s(r)r^{-1+\varepsilon_s}\,dr-F_s(0)\right|
\leq\sup_{0<r<\delta}|F_s(r)-F_s(0)|
+2\sup_{0<r<1}|F_s(r)|(1-\delta^{\varepsilon_s}). ε s ∫ 0 1 F s ( r ) r − 1 + ε s d r − F s ( 0 ) ≤ 0 < r < δ sup ∣ F s ( r ) − F s ( 0 ) ∣ + 2 0 < r < 1 sup ∣ F s ( r ) ∣ ( 1 − δ ε s ) .
The second term tends to zero as s ↑ 1 s\uparrow1 s ↑ 1 . Taylor expansion at r = 0 r=0 r = 0 , uniformly for b s b_s b s in a compact subset of ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) , makes the first term tend to zero as δ ↓ 0 \delta\downarrow0 δ ↓ 0 . Since ε s = p ( 1 − s ) \varepsilon_s=p(1-s) ε s = p ( 1 − s ) , this gives
lim s ↑ 1 ( 1 − s ) ∫ 0 1 ∣ 1 − t ( s p − 1 ) / p ∣ p ( 1 − t ) 1 + s p d t = 1 p ( p − 1 p ) p . \lim_{s\uparrow1}(1-s)\int_{0}^{1}\frac{|1-t^{(sp-1)/p}|^{p}}{(1-t)^{1+sp}}\,dt
=\frac{1}{p}\left(\frac{p-1}{p}\right)^p. s ↑ 1 lim ( 1 − s ) ∫ 0 1 ( 1 − t ) 1 + s p ∣1 − t ( s p − 1 ) / p ∣ p d t = p 1 ( p p − 1 ) p .
Also,
2 π ( d − 1 ) / 2 Γ ( ( 1 + p ) / 2 ) Γ ( ( d + p ) / 2 ) = ∫ S d − 1 ∣ ω d ∣ p d ω . 2\pi^{(d-1)/2}\frac{\Gamma((1+p)/2)}{\Gamma((d+p)/2)}
=\int_{\mathbb{S}^{d-1}}|\omega_d|^p\,d\omega. 2 π ( d − 1 ) /2 Γ (( d + p ) /2 ) Γ (( 1 + p ) /2 ) = ∫ S d − 1 ∣ ω d ∣ p d ω .
Thus
lim s ↑ 1 ( 1 − s ) D d , s , p = K d , p ( p − 1 p ) p . \lim_{s\uparrow1}(1-s)D_{d,s,p}=K_{d,p}\left(\frac{p-1}{p}\right)^p. s ↑ 1 lim ( 1 − s ) D d , s , p = K d , p ( p p − 1 ) p .
Dominated convergence on supp u \operatorname{supp}u supp u proves both weighted limits. Test the optimal quotient with a fixed u u u and then take the infimum to obtain the upper bound. □
For p = 2 p=2 p = 2 , we also prove convergence of the sharp constants. We first take d = 2 d=2 d = 2 , so K 2 , 2 = π / 2 K_{2,2}=\pi/2 K 2 , 2 = π /2 . Set
H 1 [ u ] = ∫ R + 2 ∣ ∇ u ∣ 2 d x − 1 4 ∫ R + 2 ∣ u ∣ 2 x 2 2 d x , \mathcal{H}_1[u]=\int_{\mathbb{R}_{+}^{2}}|\nabla u|^2\,dx-\frac14\int_{\mathbb{R}_{+}^{2}}\frac{|u|^2}{x_2^2}\,dx, H 1 [ u ] = ∫ R + 2 ∣∇ u ∣ 2 d x − 4 1 ∫ R + 2 x 2 2 ∣ u ∣ 2 d x ,
R 1 , τ [ u ] = ∫ R + 2 x 2 τ − 2 ∣ u ( x ) ∣ 2 ∣ x ∣ τ d x , \mathcal{R}_{1,\tau}[u]=\int_{\mathbb{R}_{+}^{2}}\frac{x_2^{\tau-2}|u(x)|^2}{|x|^\tau}\,dx, R 1 , τ [ u ] = ∫ R + 2 ∣ x ∣ τ x 2 τ − 2 ∣ u ( x ) ∣ 2 d x ,
and let Λ 1 , τ \Lambda_{1,\tau} Λ 1 , τ be the infimum of H 1 [ u ] / R 1 , τ [ u ] \mathcal{H}_1[u]/\mathcal{R}_{1,\tau}[u] H 1 [ u ] / R 1 , τ [ u ] over nonzero functions in C c ∞ ( R + 2 ) C_c^\infty(\mathbb{R}_{+}^{2}) C c ∞ ( R + 2 ) . Put
H 1 , τ = L 2 ( ( 0 , π ) , ( sin θ ) τ − 1 d θ ) H_{1,\tau}=L^2((0,\pi),(\sin\theta)^{\tau-1}\,d\theta) H 1 , τ = L 2 (( 0 , π ) , ( sin θ ) τ − 1 d θ )
and, on C c ∞ ( 0 , π ) C_c^\infty(0,\pi) C c ∞ ( 0 , π ) , define
l 1 , ξ [ y ] = ∫ 0 π sin θ { ∣ y ′ ∣ 2 + ( ξ 2 + 1 4 ) ∣ y ∣ 2 } d θ . \mathfrak{l}_{1,\xi}[y]=\int_{0}^{\pi}\sin\theta\left\{|y'|^2+\left(\xi^2+\frac14\right)|y|^2\right\}\,d\theta. l 1 , ξ [ y ] = ∫ 0 π sin θ { ∣ y ′ ∣ 2 + ( ξ 2 + 4 1 ) ∣ y ∣ 2 } d θ .
Let V τ \mathcal{V}_{\tau} V τ be the closure of the angular core under ( ℓ 1 , 0 [ y ] + ∥ y ∥ H τ 1 2 ) 1 / 2 (\ell_{1,0}[y]+\|y\|_{H^1_{\tau}}^2)^{1/2} ( ℓ 1 , 0 [ y ] + ∥ y ∥ H τ 1 2 ) 1/2 .
Lemma 8.2 (The local angular form). For every τ > 0 \tau>0 τ > 0 , the embedding V τ ↪ H 1 , τ \mathcal{V}_{\tau}\hookrightarrow H_{1,\tau} V τ ↪ H 1 , τ is compact. Moreover,
Λ 1 , τ = min y ∈ V τ ∖ { 0 } ℓ 1 , 0 [ y ] ∥ y ∥ H τ 1 2 . \Lambda_{1,\tau}=\min_{y\in\mathcal{V}_{\tau}\setminus\{0\}}\frac{\ell_{1,0}[y]}{\|y\|_{H^1_{\tau}}^2}. Λ 1 , τ = y ∈ V τ ∖ { 0 } min ∥ y ∥ H τ 1 2 ℓ 1 , 0 [ y ] . The minimum is positive and is attained by a strictly positive function, unique up to a constant factor.
Proof. Write ψ ( θ ) = ( sin θ ) 1 / 2 y ( θ ) \psi(\theta)=(\sin\theta)^{1/2}y(\theta) ψ ( θ ) = ( sin θ ) 1/2 y ( θ ) for an angular core function. Direct differentiation and one integration by parts give
∫ 0 π ( ∣ ψ ′ ∣ 2 − ∣ ψ ∣ 2 4 sin 2 θ ) d θ = ∫ 0 π sin θ ( ∣ y ′ ∣ 2 + 1 4 ∣ y ∣ 2 ) d θ . \int_0^\pi\left(|\psi'|^2-\frac{|\psi|^2}{4\sin^2\theta}\right)\,d\theta=\int_0^\pi\sin\theta\left(|y'|^2+\frac{1}{4}|y|^2\right)\,d\theta. ∫ 0 π ( ∣ ψ ′ ∣ 2 − 4 sin 2 θ ∣ ψ ∣ 2 ) d θ = ∫ 0 π sin θ ( ∣ y ′ ∣ 2 + 4 1 ∣ y ∣ 2 ) d θ .
Also,
∫ 0 π ( sin θ ) τ − 2 ∣ ψ ∣ 2 d θ = ∥ y ∥ H τ 1 2 . \int_0^\pi(\sin\theta)^{\tau-2}|\psi|^2\,d\theta=\|y\|_{H^1_{\tau}}^2. ∫ 0 π ( sin θ ) τ − 2 ∣ ψ ∣ 2 d θ = ∥ y ∥ H τ 1 2 .
For compactness, use
t = log tan θ 2 , sin θ = sech t , d θ = sech t d t . t=\log\tan\frac{\theta}{2},\qquad\sin\theta=\operatorname{sech}t,\qquad d\theta=\operatorname{sech}t\,dt. t = log tan 2 θ , sin θ = sech t , d θ = sech t d t .
If Y ( t ) = y ( θ ( t ) ) Y(t)=y(\theta(t)) Y ( t ) = y ( θ ( t )) , then
∫ 0 π sin θ ∣ y ′ ∣ 2 d θ = ∫ R ∣ Y ′ ∣ 2 d t , ∫ 0 π sin θ ∣ y ∣ 2 d θ = ∫ R sech 2 t ∣ Y ∣ 2 d t , \int_0^\pi\sin\theta|y'|^2\,d\theta=\int_{\mathbb{R}}|Y'|^2\,dt,\qquad\int_0^\pi\sin\theta|y|^2\,d\theta=\int_{\mathbb{R}}\operatorname{sech}^2t|Y|^2\,dt, ∫ 0 π sin θ ∣ y ′ ∣ 2 d θ = ∫ R ∣ Y ′ ∣ 2 d t , ∫ 0 π sin θ ∣ y ∣ 2 d θ = ∫ R sech 2 t ∣ Y ∣ 2 d t ,
whereas
∥ y ∥ H τ 1 2 = ∫ R sech τ t ∣ Y ∣ 2 d t . \|y\|_{H^1_{\tau}}^2=\int_{\mathbb{R}}\operatorname{sech}^{\tau}t|Y|^2\,dt. ∥ y ∥ H τ 1 2 = ∫ R sech τ t ∣ Y ∣ 2 d t .
Identifying y y y with Y Y Y under this change of variables, the form domain is
V τ = { Y ∈ H l o c 1 ( R ) : Y ′ ∈ L 2 ( R ) , sech t Y ∈ L 2 ( R ) , sech τ / 2 t Y ∈ L 2 ( R ) } . \mathcal{V}_{\tau}=\left\{Y\in H^1_{\mathrm{loc}}(\mathbb{R}):Y'\in L^2(\mathbb{R}),\ \operatorname{sech}tY\in L^2(\mathbb{R}),\ \operatorname{sech}^{\tau/2}tY\in L^2(\mathbb{R})\right\}. V τ = { Y ∈ H loc 1 ( R ) : Y ′ ∈ L 2 ( R ) , sech t Y ∈ L 2 ( R ) , sech τ /2 t Y ∈ L 2 ( R ) } .
For the reverse inclusion, let Y Y Y belong to the set on the right. Since Y ′ ∈ L 2 ( R ) Y'\in L^2(\mathbb{R}) Y ′ ∈ L 2 ( R ) ,
Y ( t ) = o ( ∣ t ∣ 1 / 2 ) ( ∣ t ∣ → ∞ ) . Y(t)=o(|t|^{1/2})\qquad(|t|\to\infty). Y ( t ) = o ( ∣ t ∣ 1/2 ) ( ∣ t ∣ → ∞ ) .
Indeed, for t > T > 0 t>T>0 t > T > 0 ,
∣ Y ( t ) ∣ ≤ ∣ Y ( T ) ∣ + ( t − T ) 1 / 2 ∥ Y ′ ∥ L 2 ( T , ∞ ) . |Y(t)|\leq|Y(T)|+(t-T)^{1/2}\|Y'\|_{L^2(T,\infty)}. ∣ Y ( t ) ∣ ≤ ∣ Y ( T ) ∣ + ( t − T ) 1/2 ∥ Y ′ ∥ L 2 ( T , ∞ ) .
Divide by t 1 / 2 t^{1/2} t 1/2 and let first t → ∞ t\to\infty t → ∞ , then T → ∞ T\to\infty T → ∞ . The other end is identical. Choose χ ∈ C c ∞ ( R ) \chi\in C^\infty_c(\mathbb{R}) χ ∈ C c ∞ ( R ) satisfying χ = 1 \chi=1 χ = 1 on [ − 1 , 1 ] [-1,1] [ − 1 , 1 ] and supp χ ⊂ [ − 2 , 2 ] \operatorname{supp}\chi\subset[-2,2] supp χ ⊂ [ − 2 , 2 ] . Put Y R ( t ) = χ ( t / R ) Y ( t ) Y_R(t)=\chi(t/R)Y(t) Y R ( t ) = χ ( t / R ) Y ( t ) . The two weighted L 2 L^2 L 2 errors tend to zero by dominated convergence. The derivative error away from the cutoff annulus also tends to zero. On that annulus,
∫ R < ∣ t ∣ < 2 R ∣ Y ( t ) ∣ 2 R 2 d t = o ( 1 ) . \int_{R<|t|<2R}\frac{|Y(t)|^2}{R^2}\,dt=o(1). ∫ R < ∣ t ∣ < 2 R R 2 ∣ Y ( t ) ∣ 2 d t = o ( 1 ) .
Thus Y R → Y Y_R\to Y Y R → Y in the graph norm. Mollification proves density of the transformed core and the domain description.
The trace inequality on [ − 1 , 1 ] [-1,1] [ − 1 , 1 ] gives
∣ Y ( 0 ) ∣ 2 ≤ C { ∫ R ∣ Y ′ ∣ 2 d t + ∫ R sech 2 t ∣ Y ∣ 2 d t } . |Y(0)|^2\leq C\left\{\int_{\mathbb{R}}|Y'|^2\,dt+\int_{\mathbb{R}}\operatorname{sech}^2t|Y|^2\,dt\right\}. ∣ Y ( 0 ) ∣ 2 ≤ C { ∫ R ∣ Y ′ ∣ 2 d t + ∫ R sech 2 t ∣ Y ∣ 2 d t } .
The fundamental theorem of calculus and Cauchy–Schwarz give
∣ Y ( t ) ∣ 2 ≤ C ( 1 + ∣ t ∣ ) { ∫ R ∣ Y ′ ∣ 2 d t + ∫ R sech 2 t ∣ Y ∣ 2 d t } . |Y(t)|^2\leq C(1+|t|)\left\{\int_{\mathbb{R}}|Y'|^2\,dt+\int_{\mathbb{R}}\operatorname{sech}^2t|Y|^2\,dt\right\}. ∣ Y ( t ) ∣ 2 ≤ C ( 1 + ∣ t ∣ ) { ∫ R ∣ Y ′ ∣ 2 d t + ∫ R sech 2 t ∣ Y ∣ 2 d t } .
Integration against sech τ t \operatorname{sech}^{\tau}t sech τ t also gives
∥ y ∥ H 1 , τ 2 ≤ C τ l 1 , 0 [ y ] . \|y\|_{H^1,\tau}^2 \le C_\tau l_{1,0}[y]. ∥ y ∥ H 1 , τ 2 ≤ C τ l 1 , 0 [ y ] .
In particular,
∫ ∣ t ∣ > T sech τ t ∣ Y ( t ) ∣ 2 d t ≤ C τ ( 1 + T ) e − τ T l 1 , 0 [ y ] . \int_{\lvert t\rvert>T}\operatorname{sech}^{\tau}t\lvert Y(t)\rvert^2\,dt \le C_\tau(1+T)e^{-\tau T}l_{1,0}[y]. ∫ ∣ t ∣ > T sech τ t ∣ Y ( t ) ∣ 2 d t ≤ C τ ( 1 + T ) e − τ T l 1 , 0 [ y ] .
For a form-bounded sequence, Rellich compactness gives a subsequence convergent in L 2 ( − T , T ) L^2(-T,T) L 2 ( − T , T ) for each integer T T T . The displayed tail bound makes this subsequence Cauchy in L 2 ( sech τ t d t ) L^2(\operatorname{sech}^{\tau}t\,dt) L 2 ( sech τ t d t ) . This proves compactness in H τ 1 H^1_\tau H τ 1 .
For x = e t ( cos θ , sin θ ) x=e^t(\cos\theta,\sin\theta) x = e t ( cos θ , sin θ ) , put w ( t , θ ) = u ( e t cos θ , e t sin θ ) w(t,\theta)=u(e^t\cos\theta,e^t\sin\theta) w ( t , θ ) = u ( e t cos θ , e t sin θ ) . The local polar identities and Fourier transformation in t t t give
H 1 [ u ] = ∫ R l 1 , ξ [ y ( ξ , ⋅ ) ] d ξ , R 1 , τ [ u ] = ∫ R ∥ y ( ξ , ⋅ ) ∥ H 1 , τ 2 d ξ , \mathcal{H}_1[u]=\int_{\mathbb{R}}l_{1,\xi}[y(\xi,\cdot)]\,d\xi,\qquad R_{1,\tau}[u]=\int_{\mathbb{R}}\|y(\xi,\cdot)\|_{H^1,\tau}^2\,d\xi, H 1 [ u ] = ∫ R l 1 , ξ [ y ( ξ , ⋅ )] d ξ , R 1 , τ [ u ] = ∫ R ∥ y ( ξ , ⋅ ) ∥ H 1 , τ 2 d ξ ,
where w ^ ( ξ , θ ) = ( sin θ ) 1 / 2 y ( ξ , θ ) \widehat{w}(\xi,\theta)=(\sin\theta)^{1/2}y(\xi,\theta) w ( ξ , θ ) = ( sin θ ) 1/2 y ( ξ , θ ) . Since
l 1 , ξ [ y ] = l 1 , 0 [ y ] + ξ 2 ∫ 0 π sin θ ∣ y ∣ 2 d θ , l_{1,\xi}[y]=l_{1,0}[y]+\xi^2\int_0^\pi\sin\theta\lvert y\rvert^2\,d\theta, l 1 , ξ [ y ] = l 1 , 0 [ y ] + ξ 2 ∫ 0 π sin θ ∣ y ∣ 2 d θ ,
the quotient is bounded below by the zero-fiber minimum. A logarithmic cutoff, as in Lemma 5.1 , gives equality.
Compactness gives a minimizer. Taking its modulus and applying the one-dimensional maximum principle gives positivity. For simplicity, let Y 1 , Y 2 Y_1,Y_2 Y 1 , Y 2 be two positive minimizers in the t t t variable. They satisfy
− Y j ′ ′ + 1 4 sech 2 t Y j = Λ 1 , τ sech τ t Y j on R . -Y_j''+\frac{1}{4}\operatorname{sech}^2t\,Y_j=\Lambda_{1,\tau}\operatorname{sech}^{\tau}t\,Y_j\quad\text{on }\mathbb{R}. − Y j ′′ + 4 1 sech 2 t Y j = Λ 1 , τ sech τ t Y j on R .
Their Wronskian
W ( t ) = Y 1 ′ ( t ) Y 2 ( t ) − Y 1 ( t ) Y 2 ′ ( t ) W(t)=Y_1'(t)Y_2(t)-Y_1(t)Y_2'(t) W ( t ) = Y 1 ′ ( t ) Y 2 ( t ) − Y 1 ( t ) Y 2 ′ ( t )
is constant. The form bounds give Y j ′ ∈ L 2 ( R ) Y_j'\in L^2(\mathbb{R}) Y j ′ ∈ L 2 ( R ) and ∣ Y j ( t ) ∣ ≤ C ( 1 + ∣ t ∣ ) 1 / 2 \lvert Y_j(t)\rvert\le C(1+\lvert t\rvert)^{1/2} ∣ Y j ( t )∣ ≤ C ( 1 + ∣ t ∣ ) 1/2 . There is a sequence t n → + ∞ t_n\to+\infty t n → + ∞ such that
t n { ∣ Y 1 ′ ( t n ) ∣ 2 + ∣ Y 2 ′ ( t n ) ∣ 2 } ⟶ 0. t_n\{\lvert Y_1'(t_n)\rvert^2+\lvert Y_2'(t_n)\rvert^2\}\longrightarrow0. t n {∣ Y 1 ′ ( t n ) ∣ 2 + ∣ Y 2 ′ ( t n ) ∣ 2 } ⟶ 0.
Otherwise their squared derivatives would dominate c / t c/t c / t at infinity, contrary to integrability. Hence W ( t n ) → 0 W(t_n)\to0 W ( t n ) → 0 , so W = 0 W=0 W = 0 and Y 1 / Y 2 Y_1/Y_2 Y 1 / Y 2 is constant. □
Proposition 8.3 (The local benchmark at τ = 2 \tau=2 τ = 2 ). For τ = 2 \tau=2 τ = 2 , the local angular eigenvalues are
λ j ( 1 ) ( ξ ) = ( j − 1 ) j + 1 4 + ξ 2 , j ≥ 1. (103) \lambda_j^{(1)}(\xi)=(j-1)j+\frac{1}{4}+\xi^2,\qquad j\ge1. \tag*{(103)} λ j ( 1 ) ( ξ ) = ( j − 1 ) j + 4 1 + ξ 2 , j ≥ 1. (103) The positive normalized zero-fiber ground state in H 2 1 H^1_2 H 2 1 is y 1 , 2 = 2 − 1 / 2 y_{1,2}=2^{-1/2} y 1 , 2 = 2 − 1/2 . Hence
Λ 1 , 2 = 1 4 , λ 2 ( 1 ) ( 0 ) − λ 1 ( 1 ) ( 0 ) = 2 , ∫ 0 π sin θ y 1 , 2 2 d θ = 1. \Lambda_{1,2}=\frac{1}{4},\qquad\lambda_2^{(1)}(0)-\lambda_1^{(1)}(0)=2,\qquad\int_0^\pi\sin\theta\,y_{1,2}^2\,d\theta=1. Λ 1 , 2 = 4 1 , λ 2 ( 1 ) ( 0 ) − λ 1 ( 1 ) ( 0 ) = 2 , ∫ 0 π sin θ y 1 , 2 2 d θ = 1. For the local cylinder form and the same multiplier h ( ξ ) = min { ξ 2 , 1 } h(\xi)=\min\{\xi^2,1\} h ( ξ ) = min { ξ 2 , 1 } , the optimal stability coefficient is 1 and is attained by nonzero finite-energy functions.
Proof. Put z = cos θ z=\cos\theta z = cos θ and y ( θ ) = Y ( z ) y(\theta)=Y(z) y ( θ ) = Y ( z ) . Then
∥ y ∥ H 1 , 2 2 = ∫ − 1 1 ∣ Y ∣ 2 d z , l 1 , ξ [ y ] = ∫ − 1 1 { ( 1 − z 2 ) ∣ Y ′ ∣ 2 + ( 1 4 + ξ 2 ) ∣ Y ∣ 2 } d z . \|y\|_{H^1,2}^2=\int_{-1}^1\lvert Y\rvert^2\,dz,\qquad l_{1,\xi}[y]=\int_{-1}^1\left\{(1-z^2)\lvert Y'\rvert^2+\left(\frac{1}{4}+\xi^2\right)\lvert Y\rvert^2\right\}\,dz. ∥ y ∥ H 1 , 2 2 = ∫ − 1 1 ∣ Y ∣ 2 d z , l 1 , ξ [ y ] = ∫ − 1 1 { ( 1 − z 2 ) ∣ Y ′ ∣ 2 + ( 4 1 + ξ 2 ) ∣ Y ∣ 2 } d z . The Legendre polynomials belong to the closed form domain: under t = log tan ( θ / 2 ) t=\log\tan(\theta/2) t = log tan ( θ /2 ) they are bounded with square-integrable derivative, so the cutoff argument in Lemma 8.2 applies. Their differential equation and integration by parts show that their eigenvalues are n ( n + 1 ) + 1 / 4 + ξ 2 n(n+1)+1/4+\xi^2 n ( n + 1 ) + 1/4 + ξ 2 , n ≥ 0 n\ge0 n ≥ 0 . Completeness of the Legendre basis in L 2 ( − 1 , 1 ) L^2(-1,1) L 2 ( − 1 , 1 ) proves (103) .
To specify the local stability statement, write y ^ ( ξ , θ ) \widehat{y}(\xi,\theta) y ( ξ , θ ) in this orthonormal basis of H 1 , 2 H_{1,2} H 1 , 2 and let c n ( ξ ) c_n(\xi) c n ( ξ ) be its coefficients. The local sharp deficit and distance are respectively
∫ R ∑ n ≥ 0 { n ( n + 1 ) + ξ 2 } ∣ c n ( ξ ) ∣ 2 d ξ \int_{\mathbb{R}}\sum_{n\ge0}\{n(n+1)+\xi^2\}|c_n(\xi)|^2\,d\xi ∫ R n ≥ 0 ∑ { n ( n + 1 ) + ξ 2 } ∣ c n ( ξ ) ∣ 2 d ξ
and
∫ R { h ( ξ ) ∣ c 0 ( ξ ) ∣ 2 + ∑ n ≥ 1 ∣ c n ( ξ ) ∣ 2 } d ξ . \int_{\mathbb{R}}\left\{h(\xi)|c_0(\xi)|^2+\sum_{n\ge1}|c_n(\xi)|^2\right\}\,d\xi. ∫ R { h ( ξ ) ∣ c 0 ( ξ ) ∣ 2 + n ≥ 1 ∑ ∣ c n ( ξ ) ∣ 2 } d ξ .
The first is at least the second. Equality holds precisely when c n = 0 c_n=0 c n = 0 for n ≥ 1 n\ge1 n ≥ 1 and c 0 c_0 c 0 vanishes outside [ − 1 , 1 ] [-1,1] [ − 1 , 1 ] , up to null sets. Any nonzero c 0 ∈ L 2 ( − 1 , 1 ) c_0\in L^2(-1,1) c 0 ∈ L 2 ( − 1 , 1 ) defines such a finite-energy function. Thus the coefficient 1 is sharp and attained. □
For s < 1 s<1 s < 1 , use the ground-state conjugation
ψ ( θ ) = ( sin θ ) s − 1 / 2 y ( θ ) . \psi(\theta)=(\sin\theta)^{s-1/2}y(\theta). ψ ( θ ) = ( sin θ ) s − 1/2 y ( θ ) .
It gives the exact identity
∥ ψ ∥ H s , τ 2 = ∥ y ∥ H 1 , τ 2 . \|\psi\|_{H_{s,\tau}}^2=\|y\|_{H_{1,\tau}}^2. ∥ ψ ∥ H s , τ 2 = ∥ y ∥ H 1 , τ 2 .
Define
a s , ξ [ y ] = 2 ( 1 − s ) π b s , ξ [ ( sin θ ) s − 1 / 2 y ] . a_{s,\xi}[y]=\frac{2(1-s)}{\pi}b_{s,\xi}[(\sin\theta)^{s-1/2}y]. a s , ξ [ y ] = π 2 ( 1 − s ) b s , ξ [( sin θ ) s − 1/2 y ] .
We use its closure in H 1 , τ H_{1,\tau} H 1 , τ , with domain
D ( a s , ξ ) = { y ∈ H 1 , τ : ( sin θ ) s − 1 / 2 y ∈ Y s } . D(a_{s,\xi})=\{y\in H_{1,\tau}:(\sin\theta)^{s-1/2}y\in\mathcal{Y}_s\}. D ( a s , ξ ) = { y ∈ H 1 , τ : ( sin θ ) s − 1/2 y ∈ Y s } .
The conjugation is unitary from H 1 , τ H_{1,\tau} H 1 , τ onto H s , τ H_{s,\tau} H s , τ . Thus the form is closed with compact resolvent. By the positive fiber representation,
a s , ξ [ y ] = 2 ( 1 − s ) π ∫ 0 π ∫ 0 π ∫ R K s ( q , θ , φ ) ( sin θ sin φ ) s − 1 / 2 ∣ e − q / 4 y ( θ ) − e q / 4 − i ξ q y ( φ ) ∣ 2 d q d θ d φ . a_{s,\xi}[y]=\frac{2(1-s)}{\pi}\int_0^\pi\int_0^\pi\int_{\mathbb{R}}K_s(q,\theta,\varphi)(\sin\theta\sin\varphi)^{s-1/2}\left|e^{-q/4}y(\theta)-e^{q/4-i\xi q}y(\varphi)\right|^2\,dq\,d\theta\,d\varphi. a s , ξ [ y ] = π 2 ( 1 − s ) ∫ 0 π ∫ 0 π ∫ R K s ( q , θ , φ ) ( sin θ sin φ ) s − 1/2 e − q /4 y ( θ ) − e q /4 − i ξ q y ( φ ) 2 d q d θ d φ .
Lemma 8.4 (Uniform endpoint estimate). *Fix s 0 ∈ ( 1 / 2 , 1 ) s_0\in(1/2,1) s 0 ∈ ( 1/2 , 1 ) . For every σ > 0 \sigma>0 σ > 0 , there are C s 0 , σ , τ > 0 C_{s_0,\sigma,\tau}>0 C s 0 , σ , τ > 0 and δ 0 ∈ ( 0 , 1 / 8 ) \delta_0\in(0,1/8) δ 0 ∈ ( 0 , 1/8 ) , depending only on s 0 s_0 s 0 , σ \sigma σ , and τ \tau τ , such that, for s 0 < s < 1 s_0<s<1 s 0 < s < 1 , 0 < δ < δ 0 0<\delta<\delta_0 0 < δ < δ 0 , and y ∈ D ( a s , 0 ) y\in D(a_{s,0}) y ∈ D ( a s , 0 ) ,
∫ ( 0 , δ ) ∪ ( π − δ , π ) ( sin θ ) σ − 1 ∣ y ( θ ) ∣ 2 d θ ≤ C s 0 , σ , τ δ σ / 2 { a s , 0 [ y ] + ∥ y ∥ H 1 , τ 2 } . (104) \int_{(0,\delta)\cup(\pi-\delta,\pi)}(\sin\theta)^{\sigma-1}|y(\theta)|^2\,d\theta\le C_{s_0,\sigma,\tau}\delta^{\sigma/2}\{a_{s,0}[y]+\|y\|_{H_{1,\tau}}^2\}. \tag*{(104)} ∫ ( 0 , δ ) ∪ ( π − δ , π ) ( sin θ ) σ − 1 ∣ y ( θ ) ∣ 2 d θ ≤ C s 0 , σ , τ δ σ /2 { a s , 0 [ y ] + ∥ y ∥ H 1 , τ 2 } . (104) The constants are independent of s s s , δ \delta δ , and y y y . Here σ \sigma σ may differ from the base-space exponent τ \tau τ .
Proof. It is enough to work near zero. Set
I k = ( 2 − k − 1 , 2 − k ) , k ≥ k 0 . I_k=(2^{-k-1},2^{-k}),\qquad k\ge k_0. I k = ( 2 − k − 1 , 2 − k ) , k ≥ k 0 .
Let E s , k [ y ] E_{s,k}[y] E s , k [ y ] be the contribution to a s , 0 [ y ] a_{s,0}[y] a s , 0 [ y ] from
θ , φ ∈ I k ∪ I k + 1 , ∣ q ∣ < 2 − k − 3 . \theta,\varphi\in I_k\cup I_{k+1},\qquad|q|<2^{-k-3}. θ , φ ∈ I k ∪ I k + 1 , ∣ q ∣ < 2 − k − 3 .
These regions have bounded overlap. Hence
∑ k ≥ k 0 E s , k [ y ] ≤ C a s , 0 [ y ] . \sum_{k\ge k_0}E_{s,k}[y]\le Ca_{s,0}[y]. k ≥ k 0 ∑ E s , k [ y ] ≤ C a s , 0 [ y ] .
On the region defining E s , k E_{s,k} E s , k ,
K s ( q , θ , φ ) ≍ ( q 2 + ∣ θ − φ ∣ 2 ) − 1 − s K_s(q,\theta,\varphi)\asymp(q^2+|\theta-\varphi|^2)^{-1-s} K s ( q , θ , φ ) ≍ ( q 2 + ∣ θ − φ ∣ 2 ) − 1 − s
and
( sin θ sin φ ) s − 1 / 2 ≍ 2 − k ( 2 s − 1 ) . (\sin\theta\sin\varphi)^{s-1/2}\asymp2^{-k(2s-1)}. ( sin θ sin φ ) s − 1/2 ≍ 2 − k ( 2 s − 1 ) .
Make the changes
q = 2 − k Q , θ = 2 − k Θ , φ = 2 − k Φ . q=2^{-k}Q,\qquad\theta=2^{-k}\Theta,\qquad\varphi=2^{-k}\Phi. q = 2 − k Q , θ = 2 − k Θ , φ = 2 − k Φ.
All dyadic powers cancel. Put J = ( 1 / 4 , 1 ) J=(1/4,1) J = ( 1/4 , 1 ) and f J = ∣ J ∣ − 1 ∫ J f f_J=|J|^{-1}\int_J f f J = ∣ J ∣ − 1 ∫ J f . The needed estimate is
1 ∣ J ∣ ∫ J ∣ f − f J ∣ 2 d Θ ≤ C ( 1 − s ) ∫ J ∫ J ∫ ∣ Q ∣ < 1 / 8 ∣ f ( Θ ) − f ( Φ ) ∣ 2 ( Q 2 + ∣ Θ − Φ ∣ 2 ) 1 + s d Q d Θ d Φ , \frac{1}{|J|}\int_J |f-f_J|^2\,d\Theta\le C(1-s)\int_J\int_J\int_{|Q|<1/8}\frac{|f(\Theta)-f(\Phi)|^2}{(Q^2+|\Theta-\Phi|^2)^{1+s}}\,dQ\,d\Theta\,d\Phi, ∣ J ∣ 1 ∫ J ∣ f − f J ∣ 2 d Θ ≤ C ( 1 − s ) ∫ J ∫ J ∫ ∣ Q ∣ < 1/8 ( Q 2 + ∣Θ − Φ ∣ 2 ) 1 + s ∣ f ( Θ ) − f ( Φ ) ∣ 2 d Q d Θ d Φ ,
with C C C independent of s ∈ ( s 0 , 1 ) s\in(s_0,1) s ∈ ( s 0 , 1 ) . In fact, if ∣ Θ − Φ ∣ < 1 / 16 |\Theta-\Phi|<1/16 ∣Θ − Φ∣ < 1/16 , integration over ∣ Q ∣ < ∣ Θ − Φ ∣ |Q|<|\Theta-\Phi| ∣ Q ∣ < ∣Θ − Φ∣ gives
∫ ∣ Q ∣ < 1 / 8 d Q ( Q 2 + ∣ Θ − Φ ∣ 2 ) 1 + s ≥ c ∣ Θ − Φ ∣ − 1 − 2 s . \int_{|Q|<1/8}\frac{dQ}{(Q^2+|\Theta-\Phi|^2)^{1+s}}\ge c|\Theta-\Phi|^{-1-2s}. ∫ ∣ Q ∣ < 1/8 ( Q 2 + ∣Θ − Φ ∣ 2 ) 1 + s d Q ≥ c ∣Θ − Φ ∣ − 1 − 2 s .
The constant c c c is uniform. If the Poincaré bound fails, choose s n ∈ ( s 0 , 1 ) s_n\in(s_0,1) s n ∈ ( s 0 , 1 ) and f n f_n f n with ( f n ) J = 0 (f_n)_J=0 ( f n ) J = 0 , ∥ f n ∥ 2 = 1 \|f_n\|_2=1 ∥ f n ∥ 2 = 1 , and normalized truncated seminorm tending to zero. For s n → s ˉ < 1 s_n\to\bar{s}<1 s n → s ˉ < 1 , one has 1 − s n ≥ c > 0 1-s_n\ge c>0 1 − s n ≥ c > 0 for large n n n . Since s n > s 0 s_n>s_0 s n > s 0 , the truncated H s 0 H^{s_0} H s 0 seminorm is bounded by the truncated H s n H^{s_n} H s n seminorm up to a fixed constant. Distances larger than 1 / 16 1/16 1/16 contribute at most C ∥ f n ∥ 2 2 C\|f_n\|_2^2 C ∥ f n ∥ 2 2 . Rellich compactness gives f n → f f_n\to f f n → f in L 2 ( J ) L^2(J) L 2 ( J ) and, along a subsequence, almost everywhere. Fatou’s lemma gives
∫ Θ , Φ ∈ J ∣ Θ − Φ ∣ < 1 / 16 ∣ f ( Θ ) − f ( Φ ) ∣ 2 ∣ Θ − Φ ∣ 1 + 2 s ˉ d Θ d Φ = 0. \int_{\substack{\Theta,\Phi\in J\\|\Theta-\Phi|<1/16}}\frac{|f(\Theta)-f(\Phi)|^2}{|\Theta-\Phi|^{1+2\bar{s}}}\,d\Theta\,d\Phi=0. ∫ Θ , Φ ∈ J ∣Θ − Φ∣ < 1/16 ∣Θ − Φ ∣ 1 + 2 s ˉ ∣ f ( Θ ) − f ( Φ ) ∣ 2 d Θ d Φ = 0.
Thus f f f is constant on each interval of length less than $1/16; overlapping such intervals shows that it is constant on J J J . Strong convergence gives f J = 0 f_J=0 f J = 0 and ∥ f ∥ 2 = 1 \|f\|_2=1 ∥ f ∥ 2 = 1 , a contradiction. For the second case, put r ∗ = 1 / 16 r_* = 1/16 r ∗ = 1/16 and
ρ s ( h ) = 1 − s r ∗ 2 − 2 s ∣ h ∣ 1 − 2 s 1 { ∣ h ∣ < r ∗ } . \rho_s(h)=\frac{1-s}{r_*^{2-2s}}|h|^{1-2s}\mathbf{1}_{\{|h|<r_*\}}. ρ s ( h ) = r ∗ 2 − 2 s 1 − s ∣ h ∣ 1 − 2 s 1 { ∣ h ∣ < r ∗ } .
Direct integration gives
∫ R ρ s ( h ) d h = 1 , ∫ ∣ h ∣ > η ρ s ( h ) d h = 1 − ( η r ∗ ) 2 − 2 s ⟶ 0 ( 0 < η < r ∗ ) . \int_{\mathbb{R}}\rho_s(h)\,dh=1,\qquad\int_{|h|>\eta}\rho_s(h)\,dh=1-\left(\frac{\eta}{r_*}\right)^{2-2s}\longrightarrow0\qquad(0<\eta<r_*). ∫ R ρ s ( h ) d h = 1 , ∫ ∣ h ∣ > η ρ s ( h ) d h = 1 − ( r ∗ η ) 2 − 2 s ⟶ 0 ( 0 < η < r ∗ ) .
Since s > s 0 > 1 / 2 s>s_0>1/2 s > s 0 > 1/2 , it is nonincreasing in ∣ h ∣ |h| ∣ h ∣ . Hence
inf 1 / 2 ≤ t ≤ 1 ρ s ( t h ) = ρ s ( h ) , inf 1 / 2 ≤ t ≤ 1 ∫ R ρ s ( t h ) d h = 1. \inf_{1/2\le t\le1}\rho_s(th)=\rho_s(h),\qquad\inf_{1/2\le t\le1}\int_{\mathbb{R}}\rho_s(th)\,dh=1. 1/2 ≤ t ≤ 1 inf ρ s ( t h ) = ρ s ( h ) , 1/2 ≤ t ≤ 1 inf ∫ R ρ s ( t h ) d h = 1.
The additional one-dimensional condition in [20 Theorem 1.3] is thus satisfied. The normalized truncated seminorm above bounds
∫ J ∫ J ∣ f n ( Θ ) − f n ( Φ ) ∣ 2 ∣ Θ − Φ ∣ 2 ρ s n ( Θ − Φ ) d Θ d Φ , \int_J\int_J\frac{|f_n(\Theta)-f_n(\Phi)|^2}{|\Theta-\Phi|^2}\rho_{s_n}(\Theta-\Phi)\,d\Theta\,d\Phi, ∫ J ∫ J ∣Θ − Φ ∣ 2 ∣ f n ( Θ ) − f n ( Φ ) ∣ 2 ρ s n ( Θ − Φ ) d Θ d Φ ,
because r ∗ 2 − 2 s n r_*^{2-2s_n} r ∗ 2 − 2 s n stays bounded away from zero. After rescaling J J J , that theorem gives f n → f f_n\to f f n → f strongly in L 2 ( J ) L^2(J) L 2 ( J ) and f ′ = 0 f'=0 f ′ = 0 , again contradicting the normalization.
Let
a k = 1 ∣ I k ∣ ∫ I k y ( θ ) d θ , v k = 1 ∣ I k ∣ ∫ I k ∣ y − a k ∣ 2 d θ . a_k=\frac{1}{|I_k|}\int_{I_k}y(\theta)\,d\theta,\qquad v_k=\frac{1}{|I_k|}\int_{I_k}|y-a_k|^2\,d\theta. a k = ∣ I k ∣ 1 ∫ I k y ( θ ) d θ , v k = ∣ I k ∣ 1 ∫ I k ∣ y − a k ∣ 2 d θ .
For the drift factors, use
∣ y ( θ ) − y ( φ ) ∣ 2 ≤ 2 ∣ e − q / 4 y ( θ ) − e q / 4 y ( φ ) ∣ 2 + C q 2 ( ∣ y ( θ ) ∣ 2 + ∣ y ( φ ) ∣ 2 ) . |y(\theta)-y(\varphi)|^2\le2|e^{-q/4}y(\theta)-e^{q/4}y(\varphi)|^2+Cq^2\bigl(|y(\theta)|^2+|y(\varphi)|^2\bigr). ∣ y ( θ ) − y ( φ ) ∣ 2 ≤ 2∣ e − q /4 y ( θ ) − e q /4 y ( φ ) ∣ 2 + C q 2 ( ∣ y ( θ ) ∣ 2 + ∣ y ( φ ) ∣ 2 ) .
On the fixed interval J J J , the normalized integral of the drift kernel is uniformly bounded:
sup s 0 < s < 1 ( 1 − s ) ∫ J ∫ J ∫ ∣ Q ∣ < 1 / 8 Q 2 ( Q 2 + ∣ Θ − Φ ∣ 2 ) 1 + s d Q d Θ d Φ < ∞ . \sup_{s_0<s<1}(1-s)\int_J\int_J\int_{|Q|<1/8}\frac{Q^2}{(Q^2+|\Theta-\Phi|^2)^{1+s}}\,dQ\,d\Theta\,d\Phi<\infty. s 0 < s < 1 sup ( 1 − s ) ∫ J ∫ J ∫ ∣ Q ∣ < 1/8 ( Q 2 + ∣Θ − Φ ∣ 2 ) 1 + s Q 2 d Q d Θ d Φ < ∞.
Its integral operator is also uniformly bounded on L 2 ( J ) L^2(J) L 2 ( J ) , by Schur’s test. Rescaling therefore bounds the drift term by C 4 − k ( v k + v k + 1 + ∣ a k ∣ 2 + ∣ a k + 1 ∣ 2 ) C4^{-k}(v_k+v_{k+1}+|a_k|^2+|a_{k+1}|^2) C 4 − k ( v k + v k + 1 + ∣ a k ∣ 2 + ∣ a k + 1 ∣ 2 ) . Let b k b_k b k be the average of y y y on I k ∪ I k + 1 I_k\cup I_{k+1} I k ∪ I k + 1 . The variance identity gives
∫ I k ∪ I k + 1 ∣ y − b k ∣ 2 = ∣ I k ∣ v k + ∣ I k + 1 ∣ v k + 1 + ∣ I k ∣ ∣ I k + 1 ∣ ∣ I k ∣ + ∣ I k + 1 ∣ ∣ a k − a k + 1 ∣ 2 . \int_{I_k\cup I_{k+1}} |y-b_k|^2 = |I_k|v_k+|I_{k+1}|v_{k+1}+\frac{|I_k||I_{k+1}|}{|I_k|+|I_{k+1}|}|a_k-a_{k+1}|^2. ∫ I k ∪ I k + 1 ∣ y − b k ∣ 2 = ∣ I k ∣ v k + ∣ I k + 1 ∣ v k + 1 + ∣ I k ∣ + ∣ I k + 1 ∣ ∣ I k ∣∣ I k + 1 ∣ ∣ a k − a k + 1 ∣ 2 .
Divide by ∣ I k ∣ + ∣ I k + 1 ∣ |I_k|+|I_{k+1}| ∣ I k ∣ + ∣ I k + 1 ∣ . The two length ratios are 2 / 3 2/3 2/3 and 1 / 3 1/3 1/3 . The rescaled Poincaré estimate therefore controls v k + v k + 1 + ∣ a k + 1 − a k ∣ 2 v_k+v_{k+1}+|a_{k+1}-a_k|^2 v k + v k + 1 + ∣ a k + 1 − a k ∣ 2 . After increasing k 0 k_0 k 0 , the two variance terms are absorbed into the left-hand side, and we obtain
v k + ∣ a k + 1 − a k ∣ 2 ≤ C E s , k [ y ] + C 4 − k ( ∣ a k ∣ 2 + ∣ a k + 1 ∣ 2 ) . v_k+|a_{k+1}-a_k|^2\leq CE_{s,k}[y]+C4^{-k}(|a_k|^2+|a_{k+1}|^2). v k + ∣ a k + 1 − a k ∣ 2 ≤ C E s , k [ y ] + C 4 − k ( ∣ a k ∣ 2 + ∣ a k + 1 ∣ 2 ) .
These constants are independent of k 0 k_0 k 0 . For fixed k 0 k_0 k 0 ,
∣ a k 0 ∣ 2 ≤ C 0 ( k 0 , τ ) ∥ y ∥ H τ 1 2 , |a_{k_0}|^2\leq C_0(k_0,\tau)\|y\|_{H^1_\tau}^2, ∣ a k 0 ∣ 2 ≤ C 0 ( k 0 , τ ) ∥ y ∥ H τ 1 2 ,
since the weight is bounded above and below on I k 0 I_{k_0} I k 0 .
Put
E s [ y ] = a s , 0 [ y ] + ∥ y ∥ H τ 1 2 . E_s[y]=a_{s,0}[y]+\|y\|_{H^1_\tau}^2. E s [ y ] = a s , 0 [ y ] + ∥ y ∥ H τ 1 2 .
Telescoping the means and applying Cauchy–Schwarz give, for k 0 ≤ k ≤ N k_0\leq k\leq N k 0 ≤ k ≤ N ,
∣ a k ∣ 2 1 + k ≤ C 0 ( k 0 , τ ) E s [ y ] + C 1 ∑ j = k 0 k − 1 { E s , j [ y ] + 4 − j ( ∣ a j ∣ 2 + ∣ a j + 1 ∣ 2 ) } . \frac{|a_k|^2}{1+k}\leq C_0(k_0,\tau)E_s[y]+C_1\sum_{j=k_0}^{k-1}\left\{E_{s,j}[y]+4^{-j}(|a_j|^2+|a_{j+1}|^2)\right\}. 1 + k ∣ a k ∣ 2 ≤ C 0 ( k 0 , τ ) E s [ y ] + C 1 j = k 0 ∑ k − 1 { E s , j [ y ] + 4 − j ( ∣ a j ∣ 2 + ∣ a j + 1 ∣ 2 ) } .
Here C 1 C_1 C 1 is independent of s s s , k 0 k_0 k 0 , and N N N . Set
M N = max k 0 ≤ k ≤ N ∣ a k ∣ 2 1 + k . M_N=\max_{k_0\leq k\leq N}\frac{|a_k|^2}{1+k}. M N = k 0 ≤ k ≤ N max 1 + k ∣ a k ∣ 2 .
Using ∑ j ≥ k 0 E s , j [ y ] ≤ C E s [ y ] \sum_{j\geq k_0}E_{s,j}[y]\leq CE_s[y] ∑ j ≥ k 0 E s , j [ y ] ≤ C E s [ y ] , and increasing the constants without introducing k 0 k_0 k 0 dependence in C 1 C_1 C 1 , we obtain
M N ≤ C 0 ( k 0 , τ ) E s [ y ] + C 1 M N ∑ j ≥ k 0 4 − j ( 1 + j ) . M_N\leq C_0(k_0,\tau)E_s[y]+C_1M_N\sum_{j\geq k_0}4^{-j}(1+j). M N ≤ C 0 ( k 0 , τ ) E s [ y ] + C 1 M N j ≥ k 0 ∑ 4 − j ( 1 + j ) .
Choose k 0 k_0 k 0 after fixing C 1 C_1 C 1 , so large that
C 1 ∑ j ≥ k 0 4 − j ( 1 + j ) ≤ 1 2 . C_1\sum_{j\geq k_0}4^{-j}(1+j)\leq\frac{1}{2}. C 1 j ≥ k 0 ∑ 4 − j ( 1 + j ) ≤ 2 1 .
Then
M N ≤ 2 C 0 ( k 0 , τ ) E s [ y ] M_N\leq2C_0(k_0,\tau)E_s[y] M N ≤ 2 C 0 ( k 0 , τ ) E s [ y ]
uniformly in N N N . Consequently,
∣ a k ∣ 2 + v k ≤ C ( 1 + k ) E s [ y ] , k ≥ k 0 . |a_k|^2+v_k\leq C(1+k)E_s[y],\qquad k\geq k_0. ∣ a k ∣ 2 + v k ≤ C ( 1 + k ) E s [ y ] , k ≥ k 0 .
Let 2 − K − 1 < δ ≤ 2 − K 2^{-K-1}<\delta\leq2^{-K} 2 − K − 1 < δ ≤ 2 − K . For every σ > 0 \sigma>0 σ > 0 ,
∫ 0 δ θ σ − 1 ∣ y ∣ 2 d θ ≤ C ∑ k ≥ K 2 − k σ ( ∣ a k ∣ 2 + v k ) ≤ C E s [ y ] ∑ k ≥ K 2 − k σ ( 1 + k ) ≤ C s 0 , σ , τ δ σ / 2 E s [ y ] . \int_0^\delta\theta^{\sigma-1}|y|^2\,d\theta\leq C\sum_{k\geq K}2^{-k\sigma}(|a_k|^2+v_k)\leq CE_s[y]\sum_{k\geq K}2^{-k\sigma}(1+k)\leq C_{s_0,\sigma,\tau}\delta^{\sigma/2}E_s[y]. ∫ 0 δ θ σ − 1 ∣ y ∣ 2 d θ ≤ C k ≥ K ∑ 2 − k σ ( ∣ a k ∣ 2 + v k ) ≤ C E s [ y ] k ≥ K ∑ 2 − k σ ( 1 + k ) ≤ C s 0 , σ , τ δ σ /2 E s [ y ] .
For the other endpoint put y ~ ( θ ) = y ( π − θ ) \widetilde y(\theta)=y(\pi-\theta) y ( θ ) = y ( π − θ ) . The sine weight and K s ( q , θ , φ ) K_s(q,\theta,\varphi) K s ( q , θ , φ ) are invariant under simultaneous reflection, so the same estimate applies to y ~ \widetilde y y . For a fixed s s s , take core approximants y j → y y_j\to y y j → y in the graph norm of a s , 0 a_{s,0} a s , 0 . Applying the estimate to y j − y l y_j-y_l y j − y l shows that they are Cauchy in the endpoint weighted norm for every σ > 0 \sigma>0 σ > 0 . Their local L 2 L^2 L 2 limits identify that weighted limit with y y y . Letting j → ∞ j\to\infty j → ∞ proves the estimate on the closed domain.
Lemma 8.5 (Interior compactness and the covariant lower bound). Let s n ↑ 1 s_n \uparrow1 s n ↑ 1 and let ξ n \xi_n ξ n be bounded. If a s n , ξ n [ y n ] + ∥ y n ∥ H 1 , τ 2 a_{s_n,\xi_n}[y_n]+\|y_n\|_{H^1,\tau}^2 a s n , ξ n [ y n ] + ∥ y n ∥ H 1 , τ 2 is bounded, then y n y_n y n has a subsequence converging in L 2 ( J ) L^2(J) L 2 ( J ) for every J ⋐ ( 0 , π ) J\Subset(0,\pi) J ⋐ ( 0 , π ) . If also ξ n → ξ \xi_n\to\xi ξ n → ξ and y n → y y_n\to y y n → y in L 2 ( J ′ ) L^2(J') L 2 ( J ′ ) , where J ⋐ J ′ ⋐ ( 0 , π ) J\Subset J'\Subset(0,\pi) J ⋐ J ′ ⋐ ( 0 , π ) , then the localized lower bound (105) holds.
Proof. For compactness, let J ⋐ J ′ ⋐ ( 0 , π ) J\Subset J'\Subset(0,\pi) J ⋐ J ′ ⋐ ( 0 , π ) and ∣ ξ ∣ ≤ M |\xi|\le M ∣ ξ ∣ ≤ M . For ∣ q ∣ + ∣ θ − φ ∣ |q|+|\theta-\varphi| ∣ q ∣ + ∣ θ − φ ∣ small and θ , φ ∈ J ′ \theta,\varphi\in J' θ , φ ∈ J ′ ,
∣ y ( θ ) − y ( φ ) ∣ 2 ≤ 2 ∣ e − q / 4 y ( θ ) − e q / 4 − i ξ q y ( φ ) ∣ 2 + C M q 2 ( ∣ y ( θ ) ∣ 2 + ∣ y ( φ ) ∣ 2 ) . |y(\theta)-y(\varphi)|^2\le2\left|e^{-q/4}y(\theta)-e^{q/4-i\xi q}y(\varphi)\right|^2+CMq^2\left(|y(\theta)|^2+|y(\varphi)|^2\right). ∣ y ( θ ) − y ( φ ) ∣ 2 ≤ 2 e − q /4 y ( θ ) − e q /4 − i ξ q y ( φ ) 2 + C M q 2 ( ∣ y ( θ ) ∣ 2 + ∣ y ( φ ) ∣ 2 ) .
After multiplication by the normalized kernel and integration, the last term is bounded by C J ′ , M ∥ y ∥ L 2 ( J ′ ) 2 C_{J',M}\|y\|_{L^2(J')}^2 C J ′ , M ∥ y ∥ L 2 ( J ′ ) 2 . Indeed, write h = φ − θ h=\varphi-\theta h = φ − θ and fix 0 < R < dist ( J , ∂ J ′ ) 0<R<\operatorname{dist}(J,\partial J') 0 < R < dist ( J , ∂ J ′ ) . Then
( 1 − s ) ∫ q 2 + h 2 < R 2 q 2 q 2 + h 2 1 − s d q d h ≤ C . (1-s)\int_{q^2+h^2<R^2}\frac{q^2}{q^2+h^2}^{1-s}\,dq\,dh\le C. ( 1 − s ) ∫ q 2 + h 2 < R 2 q 2 + h 2 q 2 1 − s d q d h ≤ C .
For ∣ h ∣ < R / 4 |h|<R/4 ∣ h ∣ < R /4 , restrict the q q q integration to ∣ q ∣ < ∣ h ∣ |q|<|h| ∣ q ∣ < ∣ h ∣ . The resulting kernel is bounded below by c ( 1 − s ) ∣ h ∣ − 1 − 2 s c(1-s)|h|^{-1-2s} c ( 1 − s ) ∣ h ∣ − 1 − 2 s . The one-dimensional compactness theorem applies with radius $R/4: the normalized kernels have unit mass, concentrate at zero, and satisfy the dilation condition. It gives compactness in L 2 ( J ) L^2(J) L 2 ( J ) . This proves interior compactness.
For the lower limit, let J ⋐ J ′ ⋐ ( 0 , π ) J\Subset J'\Subset(0,\pi) J ⋐ J ′ ⋐ ( 0 , π ) , let s n ↑ 1 s_n\uparrow1 s n ↑ 1 and ξ n → ξ \xi_n\to\xi ξ n → ξ , and suppose that y n → y y_n\to y y n → y in L 2 ( J ′ ) L^2(J') L 2 ( J ′ ) . Then
lim inf n → ∞ a s n , ξ n [ y n ] ≥ ∫ J sin θ { ∣ y ′ ∣ 2 + ( ξ 2 + 1 4 ) ∣ y ∣ 2 } d θ . (105) \liminf_{n\to\infty}a_{s_n,\xi_n}[y_n]\ge\int_J\sin\theta\left\{|y'|^2+\left(\xi^2+\frac14\right)|y|^2\right\}\,d\theta. \tag*{(105)} n → ∞ lim inf a s n , ξ n [ y n ] ≥ ∫ J sin θ { ∣ y ′ ∣ 2 + ( ξ 2 + 4 1 ) ∣ y ∣ 2 } d θ . (105)
Here the right-hand side is + ∞ +\infty + ∞ unless y ∈ H 1 ( J ) y\in H^1(J) y ∈ H 1 ( J ) .
Write ( q , h ) = r ω (q,h)=r\omega ( q , h ) = r ω , where ω = ( ω 1 , ω 2 ) ∈ S 1 \omega=(\omega_1,\omega_2)\in S^1 ω = ( ω 1 , ω 2 ) ∈ S 1 . Set
d ν n ( r ) = 2 ( 1 − s n ) r 1 − 2 s n 1 ( 0 , R ) ( r ) d r d\nu_n(r)=2(1-s_n)r^{1-2s_n}\mathbf{1}_{(0,R)}(r)\,dr d ν n ( r ) = 2 ( 1 − s n ) r 1 − 2 s n 1 ( 0 , R ) ( r ) d r
and
A n ( θ , r , ω ) = r 2 + 2 s n K s n ( r ω 1 , θ , θ + r ω 2 ) ( sin θ sin ( θ + r ω 2 ) ) s n − 1 / 2 , A_n(\theta,r,\omega)=r^{2+2s_n}K_{s_n}(r\omega_1,\theta,\theta+r\omega_2)(\sin\theta\sin(\theta+r\omega_2))^{s_n-1/2}, A n ( θ , r , ω ) = r 2 + 2 s n K s n ( r ω 1 , θ , θ + r ω 2 ) ( sin θ sin ( θ + r ω 2 ) ) s n − 1/2 ,
V n ( θ , r , ω ) = 1 r { e − r ω 1 / 4 y n ( θ ) − e r ω 1 / 4 − i ξ n r ω 1 y n ( θ + r ω 2 ) } . V_n(\theta,r,\omega)=\frac1r\left\{e^{-r\omega_1/4}y_n(\theta)-e^{r\omega_1/4-i\xi_nr\omega_1}y_n(\theta+r\omega_2)\right\}. V n ( θ , r , ω ) = r 1 { e − r ω 1 /4 y n ( θ ) − e r ω 1 /4 − i ξ n r ω 1 y n ( θ + r ω 2 ) } .
Restricting the positive integral defining a s n , ξ n a_{s_n,\xi_n} a s n , ξ n to θ , θ + r ω 2 ∈ J ′ \theta,\theta+r\omega_2\in J' θ , θ + r ω 2 ∈ J ′ and 0 < r < R 0<r<R 0 < r < R gives
a s n , ξ n [ y n ] ≥ 1 π ∫ 0 R ∫ S 1 ∫ J ′ 1 J ′ ( θ + r ω 2 ) A n ∣ V n ∣ 2 d θ d ω d ν n ( r ) . (106) a_{s_n,\xi_n}[y_n]\ge\frac1\pi\int_0^R\int_{S^1}\int_{J'}\mathbf{1}_{J'}(\theta+r\omega_2)A_n|V_n|^2\,d\theta\,d\omega\,d\nu_n(r). \tag*{(106)} a s n , ξ n [ y n ] ≥ π 1 ∫ 0 R ∫ S 1 ∫ J ′ 1 J ′ ( θ + r ω 2 ) A n ∣ V n ∣ 2 d θ d ω d ν n ( r ) . (106)
The local kernel expansion gives, uniformly on compact angular subintervals,
r 2 + 2 s n K s n ( r ω 1 , θ , θ + r ω 2 ) = 1 + O ( r 2 ) . r^{2+2s_n}K_{s_n}(r\omega_1,\theta,\theta+r\omega_2)=1+O(r^2). r 2 + 2 s n K s n ( r ω 1 , θ , θ + r ω 2 ) = 1 + O ( r 2 ) .
Moreover, ν n ( ( 0 , R ) ) = R 2 − 2 s n → 1 \nu_n((0,R))=R^{2-2s_n}\to1 ν n (( 0 , R )) = R 2 − 2 s n → 1 and ν n ( ( η , R ) ) → 0 \nu_n((\eta,R))\to0 ν n (( η , R )) → 0 for each η > 0 \eta>0 η > 0 . Thus ν n \nu_n ν n concentrates at zero.
Let F ∈ C ∞ ( J × S 1 ) F\in C^\infty(J\times S^1) F ∈ C ∞ ( J × S 1 ) and extend it by zero in the angular variable. From ∣ V ∣ 2 ≥ 2 Re ( V F ‾ ) − ∣ F ∣ 2 |V|^2\ge2\operatorname{Re}(V\overline{F})-|F|^2 ∣ V ∣ 2 ≥ 2 Re ( V F ) − ∣ F ∣ 2 and (106) ,
lim inf n → ∞ a s n , ξ n [ y n ] ≥ 1 π lim n → ∞ ∫ 0 R ∫ S 1 ∫ J ′ A n { 2 Re ( V n F ‾ ) − ∣ F ∣ 2 } d θ d ω d ν n ( r ) . (107) \liminf_{n\to\infty}a_{s_n,\xi_n}[y_n]\ge\frac1\pi\lim_{n\to\infty}\int_0^R\int_{S^1}\int_{J'}A_n\{2\operatorname{Re}(V_n\overline{F})-|F|^2\}\,d\theta\,d\omega\,d\nu_n(r). \tag*{(107)} n → ∞ lim inf a s n , ξ n [ y n ] ≥ π 1 n → ∞ lim ∫ 0 R ∫ S 1 ∫ J ′ A n { 2 Re ( V n F ) − ∣ F ∣ 2 } d θ d ω d ν n ( r ) . (107)
On the support of F F F , the indicator is one for small r r r . The part away from zero has vanishing ν n \nu_n ν n -mass.
The ∣ F ∣ 2 |F|^2 ∣ F ∣ 2 term converges to
∫ S 1 ∫ J sin θ ∣ F ( θ , ω ) ∣ 2 d θ d ω . \int_{S^1}\int_J\sin\theta\,|F(\theta,\omega)|^2\,d\theta\,d\omega. ∫ S 1 ∫ J sin θ ∣ F ( θ , ω ) ∣ 2 d θ d ω .
For the linear term, change θ + r ω 2 \theta+r\omega_2 θ + r ω 2 to θ \theta θ in its second summand. The coefficient of y n ( θ ) y_n(\theta) y n ( θ ) is
1 r [ e − r ω 1 / 4 A n ( θ , r , ω ) F ( θ , ω ) ‾ − e r ω 1 / 4 − i ξ n r ω 1 A n ( θ − r ω 2 , r , ω ) F ( θ − r ω 2 , ω ) ‾ ] . \frac{1}{r}\left[e^{-r\omega_1/4}A_n(\theta,r,\omega)\overline{F(\theta,\omega)}-e^{r\omega_1/4-i\xi_n r\omega_1}A_n(\theta-r\omega_2,r,\omega)\overline{F(\theta-r\omega_2,\omega)}\right]. r 1 [ e − r ω 1 /4 A n ( θ , r , ω ) F ( θ , ω ) − e r ω 1 /4 − i ξ n r ω 1 A n ( θ − r ω 2 , r , ω ) F ( θ − r ω 2 , ω ) ] .
Taylor expansion of the smooth test factor, together with s n → 1 s_n\to1 s n → 1 , gives, uniformly on its support as r ↓ 0 r\downarrow0 r ↓ 0 ,
ω 2 ∂ θ ( sin θ F ‾ ) + ( − 1 2 + i ξ ) ω 1 sin θ F ‾ + o ( 1 ) . \omega_2\partial_\theta(\sin\theta\,\overline{F})+\left(-\frac{1}{2}+i\xi\right)\omega_1\sin\theta\,\overline{F}+o(1). ω 2 ∂ θ ( sin θ F ) + ( − 2 1 + i ξ ) ω 1 sin θ F + o ( 1 ) .
To see the first coefficient explicitly, put β n = s n − 1 / 2 \beta_n=s_n-1/2 β n = s n − 1/2 . Then
( sin θ sin ( θ + r ω 2 ) ) β n = sin 2 β n θ + β n r ω 2 sin 2 β n − 1 θ cos θ + O ( r 2 ) . (\sin\theta\sin(\theta+r\omega_2))^{\beta_n}=\sin^{2\beta_n}\theta+\beta_n r\omega_2\sin^{2\beta_n-1}\theta\cos\theta+O(r^2). ( sin θ sin ( θ + r ω 2 ) ) β n = sin 2 β n θ + β n r ω 2 sin 2 β n − 1 θ cos θ + O ( r 2 ) .
The same factor based at θ − r ω 2 \theta-r\omega_2 θ − r ω 2 has the opposite linear term. The part r ≥ η r\ge\eta r ≥ η is harmless because ν n ( ( η , R ) ) → 0 \nu_n((\eta,R))\to0 ν n (( η , R )) → 0 . Hence the strong L 2 ( J ′ ) L^2(J') L 2 ( J ′ ) convergence of y n y_n y n implies that the linear term in (107) converges, in distributional notation, to
∫ S 1 ∫ J sin θ { − ω 2 y ′ + ( − 1 2 + i ξ ) ω 1 y } F ‾ d θ d ω . \int_{\mathbb{S}^1}\int_J\sin\theta\left\{-\omega_2y'+\left(-\frac{1}{2}+i\xi\right)\omega_1y\right\}\overline{F}\,d\theta\,d\omega. ∫ S 1 ∫ J sin θ { − ω 2 y ′ + ( − 2 1 + i ξ ) ω 1 y } F d θ d ω .
To justify the regularity implicit in the duality step, write
L ( θ , ω ) = − ω 2 y ′ + ( − 1 2 + i ξ ) ω 1 y L(\theta,\omega)=-\omega_2y'+\left(-\frac{1}{2}+i\xi\right)\omega_1y L ( θ , ω ) = − ω 2 y ′ + ( − 2 1 + i ξ ) ω 1 y
in distributions. If the energy liminf is finite, optimizing the dual inequality over scalar multiples of any test F F F shows that L L L defines a bounded functional on L 2 ( J × S 1 , sin θ d θ d ω ) L^2(J\times\mathbb{S}^1,\sin\theta\,d\theta\,d\omega) L 2 ( J × S 1 , sin θ d θ d ω ) . It is therefore represented by an L 2 L^2 L 2 function. Since
∫ S 1 ω 2 L ( θ , ω ) d ω = − π y ′ \int_{\mathbb{S}^1}\omega_2L(\theta,\omega)\,d\omega=-\pi y' ∫ S 1 ω 2 L ( θ , ω ) d ω = − π y ′
and sin θ \sin\theta sin θ is bounded below on J J J , one obtains y ′ ∈ L 2 ( J ) y'\in L^2(J) y ′ ∈ L 2 ( J ) . Density of smooth tests and sup F ( 2 Re ⟨ L , F ⟩ − ∥ F ∥ 2 ) = ∥ L ∥ 2 \sup_F(2\operatorname{Re}\langle L,F\rangle-\lVert F\rVert^2)=\lVert L\rVert^2 sup F ( 2 Re ⟨ L , F ⟩ − ∥ F ∥ 2 ) = ∥ L ∥ 2 now give
lim inf n → ∞ a s n , ξ n [ y n ] ≥ 1 π ∫ S 1 ∫ J sin θ ∣ − ω 2 y ′ + ( − 1 2 + i ξ ) ω 1 y ∣ 2 d θ d ω . \liminf_{n\to\infty}a_{s_n,\xi_n}[y_n]\ge\frac{1}{\pi}\int_{\mathbb{S}^1}\int_J\sin\theta\left|-\omega_2y'+\left(-\frac{1}{2}+i\xi\right)\omega_1y\right|^2\,d\theta\,d\omega. n → ∞ lim inf a s n , ξ n [ y n ] ≥ π 1 ∫ S 1 ∫ J sin θ − ω 2 y ′ + ( − 2 1 + i ξ ) ω 1 y 2 d θ d ω .
Since ∫ S 1 ω 1 ω 2 d ω = 0 \int_{\mathbb{S}^1}\omega_1\omega_2\,d\omega=0 ∫ S 1 ω 1 ω 2 d ω = 0 and
∫ S 1 ω 1 2 d ω = ∫ S 1 ω 2 2 d ω = π , \int_{\mathbb{S}^1}\omega_1^2\,d\omega=\int_{\mathbb{S}^1}\omega_2^2\,d\omega=\pi, ∫ S 1 ω 1 2 d ω = ∫ S 1 ω 2 2 d ω = π ,
this is (105) .
Lemma 8.6 (Compact convergence of the angular fibers). Fix τ > 0 \tau>0 τ > 0 . Let s n ↑ 1 s_n\uparrow1 s n ↑ 1 and ξ n → ξ \xi_n\to\xi ξ n → ξ . The forms a s n , ξ n a_{s_n,\xi_n} a s n , ξ n converge to l 1 , ξ l_{1,\xi} l 1 , ξ in the compact Mosco sense on H 1 , τ H_{1,\tau} H 1 , τ . More precisely:
if
sup n { a s n , ξ n [ y n ] + ∥ y n ∥ H 1 , τ 2 } < ∞ , \sup_n\{a_{s_n,\xi_n}[y_n]+\lVert y_n\rVert_{H_{1,\tau}}^2\}<\infty, n sup { a s n , ξ n [ y n ] + ∥ y n ∥ H 1 , τ 2 } < ∞ , then a subsequence of y n y_n y n converges strongly in H 1 , τ H_{1,\tau} H 1 , τ ;
if y n ⇀ y y_n\rightharpoonup y y n ⇀ y in H 1 , τ H_{1,\tau} H 1 , τ , then
l 1 , ξ [ y ] ≤ lim inf n → ∞ a s n , ξ n [ y n ] ; l_{1,\xi}[y]\le\liminf_{n\to\infty}a_{s_n,\xi_n}[y_n]; l 1 , ξ [ y ] ≤ n → ∞ lim inf a s n , ξ n [ y n ] ; for each y ∈ V τ y\in\mathcal{V}_\tau y ∈ V τ , there are y n → y y_n\to y y n → y strongly in H 1 , τ H_{1,\tau} H 1 , τ such that
lim sup n → ∞ a s n , ξ n [ y n ] ≤ l 1 , ξ [ y ] . \limsup_{n\to\infty}a_{s_n,\xi_n}[y_n]\le l_{1,\xi}[y]. n → ∞ lim sup a s n , ξ n [ y n ] ≤ l 1 , ξ [ y ] .
Proof. Let y ∈ C c ∞ ( 0 , π ) y \in C^\infty_c(0,\pi) y ∈ C c ∞ ( 0 , π ) and put α = φ − θ \alpha=\varphi-\theta α = φ − θ . Uniformly on compact angular subintervals,
K s ( q , θ , θ + α ) = ( q 2 + α 2 ) − 1 − s ( 1 + O ( q 2 + α 2 ) ) . K_s(q,\theta,\theta+\alpha)=(q^2+\alpha^2)^{-1-s}\left(1+O(q^2+\alpha^2)\right). K s ( q , θ , θ + α ) = ( q 2 + α 2 ) − 1 − s ( 1 + O ( q 2 + α 2 ) ) .
Also, ( sin θ sin ( θ + α ) ) s − 1 / 2 → sin θ (\sin\theta\sin(\theta+\alpha))^{s-1/2}\to\sin\theta ( sin θ sin ( θ + α ) ) s − 1/2 → sin θ . Taylor’s formula gives
e − q / 4 y ( θ ) − e q / 4 − i ξ q y ( θ + α ) = − α y ′ ( θ ) + ( − 1 2 + i ξ ) q y ( θ ) + O ( q 2 + α 2 ) . e^{-q/4}y(\theta)-e^{q/4-i\xi q}y(\theta+\alpha)=-\alpha y'(\theta)+\left(-\frac{1}{2}+i\xi\right)qy(\theta)+O(q^2+\alpha^2). e − q /4 y ( θ ) − e q /4 − i ξ q y ( θ + α ) = − α y ′ ( θ ) + ( − 2 1 + i ξ ) q y ( θ ) + O ( q 2 + α 2 ) .
The mixed q α q\alpha q α term is odd. Furthermore,
lim s ↑ 1 2 ( 1 − s ) π ∫ q 2 + α 2 < δ 2 q 2 ( q 2 + α 2 ) 1 + s d q d α = 1. \lim_{s\uparrow1}\frac{2(1-s)}{\pi}\int_{q^2+\alpha^2<\delta^2}\frac{q^2}{(q^2+\alpha^2)^{1+s}}\,dq\,d\alpha=1. s ↑ 1 lim π 2 ( 1 − s ) ∫ q 2 + α 2 < δ 2 ( q 2 + α 2 ) 1 + s q 2 d q d α = 1.
The same formula holds with q 2 q^2 q 2 replaced by α 2 \alpha^2 α 2 . The normalized integral outside a fixed neighborhood of the origin tends to zero. The exponential q q q -tail is uniform for s s s near one. Hence
a s , ξ [ y ] ⟶ l 1 , ξ [ y ] a_{s,\xi}[y]\longrightarrow l_{1,\xi}[y] a s , ξ [ y ] ⟶ l 1 , ξ [ y ]
locally uniformly for bounded ξ \xi ξ .
For part (1), the diamagnetic inequality and Lemma 4.5 give a s n , 0 [ ∣ y n ∣ ] ≤ a s n , ξ n [ y n ] a_{s_n,0}[|y_n|]\le a_{s_n,\xi_n}[y_n] a s n , 0 [ ∣ y n ∣ ] ≤ a s n , ξ n [ y n ] . Apply Lemma 8.4 with σ = τ \sigma=\tau σ = τ . The weighted endpoint mass tends to zero uniformly as δ ↓ 0 \delta\downarrow0 δ ↓ 0 . Lemma 8.5 gives compactness on every interior interval. A diagonal argument proves strong compactness in H 1 , τ H_{1,\tau} H 1 , τ .
For part (2), take a subsequence attaining the finite liminf; the infinite case is immediate. Weak convergence bounds the norms, so part (1) gives strong convergence in H 1 , τ H_{1,\tau} H 1 , τ , hence locally in L 2 ( 0 , π ) L^2(0,\pi) L 2 ( 0 , π ) , to the same limit y y y . Apply Lemma 8.5 with J = ( δ , π − δ ) J=(\delta,\pi-\delta) J = ( δ , π − δ ) and let δ ↓ 0 \delta\downarrow0 δ ↓ 0 . The finite bound gives y ′ ∈ L 2 ( ( 0 , π ) , sin θ d θ ) y'\in L^2((0,\pi),\sin\theta\,d\theta) y ′ ∈ L 2 (( 0 , π ) , sin θ d θ ) and y ∈ L 2 ( ( 0 , π ) , sin θ d θ ) y\in L^2((0,\pi),\sin\theta\,d\theta) y ∈ L 2 (( 0 , π ) , sin θ d θ ) . The weak limit already belongs to H 1 , τ H_{1,\tau} H 1 , τ . The closed-form domain description in the proof of Lemma 8.2 therefore places y y y in V τ \mathcal{V}_\tau V τ . It also gives
l 1 , ξ [ y ] ≤ lim inf n → ∞ a s n , ξ n [ y n ] . l_{1,\xi}[y]\le\liminf_{n\to\infty}a_{s_n,\xi_n}[y_n]. l 1 , ξ [ y ] ≤ n → ∞ lim inf a s n , ξ n [ y n ] .
To prove part (3), approximate y ∈ V τ y\in\mathcal{V}_\tau y ∈ V τ in the local graph norm by y ( j ) ∈ C c ∞ ( 0 , π ) y^{(j)}\in C^\infty_c(0,\pi) y ( j ) ∈ C c ∞ ( 0 , π ) . For fixed j j j , the core calculation gives
a s n , ξ n [ y ( j ) ] ⟶ l 1 , ξ [ y ( j ) ] a_{s_n,\xi_n}[y^{(j)}]\longrightarrow l_{1,\xi}[y^{(j)}] a s n , ξ n [ y ( j ) ] ⟶ l 1 , ξ [ y ( j ) ]
as n → ∞ n\to\infty n → ∞ . Choose strictly increasing integers N j ≥ j N_j\ge j N j ≥ j such that
∣ a s n , ξ n [ y ( j ) ] − l 1 , ξ [ y ( j ) ] ∣ ≤ 1 j for all n ≥ N j . \left|a_{s_n,\xi_n}[y^{(j)}]-l_{1,\xi}[y^{(j)}]\right|\le\frac{1}{j}\quad\text{for all }n\ge N_j. a s n , ξ n [ y ( j ) ] − l 1 , ξ [ y ( j ) ] ≤ j 1 for all n ≥ N j .
Set j ( n ) = max { j : N j ≤ n } j(n)=\max\{j:N_j\le n\} j ( n ) = max { j : N j ≤ n } for n ≥ N 1 n\ge N_1 n ≥ N 1 and y n = y ( j ( n ) ) y_n=y^{(j(n))} y n = y ( j ( n )) ; define the finitely many earlier terms arbitrarily in the core. Then j ( n ) → ∞ j(n)\to\infty j ( n ) → ∞ . Then y n → y y_n\to y y n → y strongly in H 1 , τ H_{1,\tau} H 1 , τ and
lim sup n → ∞ a s n , ξ n [ y n ] ≤ l 1 , ξ [ y ] . \limsup_{n\to\infty}a_{s_n,\xi_n}[y_n]\le l_{1,\xi}[y]. n → ∞ lim sup a s n , ξ n [ y n ] ≤ l 1 , ξ [ y ] .
This proves part (3). □
Lemma 8.7 (Convergence of the second-moment form). Fix s 0 ∈ ( 1 / 2 , 1 ) s_0\in(1/2,1) s 0 ∈ ( 1/2 , 1 ) . For s 0 < s < 1 s_0<s<1 s 0 < s < 1 and y , z ∈ C c ∞ ( 0 , π ) y,z\in C^\infty_c(0,\pi) y , z ∈ C c ∞ ( 0 , π ) , define
t s [ y , z ] = 2 ( 1 − s ) π ∫ 0 π ∫ 0 π ∫ R q 2 K s ( q , θ , φ ) ( sin θ sin φ ) s − 1 / 2 y ( θ ) z ( φ ) ‾ d q d θ d φ . t_s[y,z]=\frac{2(1-s)}{\pi}\int_0^\pi\int_0^\pi\int_{\mathbb{R}}q^2K_s(q,\theta,\varphi)(\sin\theta\sin\varphi)^{s-1/2}y(\theta)\overline{z(\varphi)}\,dq\,d\theta\,d\varphi. t s [ y , z ] = π 2 ( 1 − s ) ∫ 0 π ∫ 0 π ∫ R q 2 K s ( q , θ , φ ) ( sin θ sin φ ) s − 1/2 y ( θ ) z ( φ ) d q d θ d φ . There is C = C ( s 0 ) > 0 C=C(s_0)>0 C = C ( s 0 ) > 0 such that
∣ t s [ y , z ] ∣ ≤ C ( ∫ 0 π sin θ ∣ y ∣ 2 d θ ) 1 / 2 ( ∫ 0 π sin θ ∣ z ∣ 2 d θ ) 1 / 2 . |t_s[y,z]|\le C\left(\int_0^\pi\sin\theta\,|y|^2\,d\theta\right)^{1/2}\left(\int_0^\pi\sin\theta\,|z|^2\,d\theta\right)^{1/2}. ∣ t s [ y , z ] ∣ ≤ C ( ∫ 0 π sin θ ∣ y ∣ 2 d θ ) 1/2 ( ∫ 0 π sin θ ∣ z ∣ 2 d θ ) 1/2 .
Thus t s t_s t s extends uniquely to a bounded form on L 2 ( ( 0 , π ) , sin θ d θ ) L^2((0,\pi),\sin\theta\,d\theta) L 2 (( 0 , π ) , sin θ d θ ) . Suppose that s n ↑ 1 s_n \uparrow1 s n ↑ 1 , that y n → y y_n \to y y n → y in L l o c 2 ( 0 , π ) L^2_{\mathrm{loc}}(0,\pi) L loc 2 ( 0 , π ) , and that
sup n { a s n , 0 [ y n ] + ∥ y n ∥ H 1 , τ 2 } < ∞ . \sup_n\{a_{s_n,0}[y_n]+\|y_n\|_{H^1,\tau}^2\}<\infty. n sup { a s n , 0 [ y n ] + ∥ y n ∥ H 1 , τ 2 } < ∞.
Then
t s n [ y n , y n ] ⟶ ∫ 0 π sin θ ∣ y ( θ ) ∣ 2 d θ . t_{s_n}[y_n,y_n]\longrightarrow\int_0^\pi\sin\theta\,|y(\theta)|^2\,d\theta. t s n [ y n , y n ] ⟶ ∫ 0 π sin θ ∣ y ( θ ) ∣ 2 d θ .
Proof. Put
L s ( α ) = 2 ( 1 − s ) π ∫ R q 2 K s ( q , α ) d q . L_s(\alpha)=\frac{2(1-s)}{\pi}\int_{\mathbb{R}}q^2K_s(q,\alpha)\,dq. L s ( α ) = π 2 ( 1 − s ) ∫ R q 2 K s ( q , α ) d q .
The local kernel comparison and the exponential tail give
0 ≤ L s ( α ) ≤ C ( 1 − s ) ( ∣ α ∣ 1 − 2 s + 1 ) , 0 < ∣ α ∣ < π , 0\leq L_s(\alpha)\leq C(1-s)(|\alpha|^{1-2s}+1),\qquad0<|\alpha|<\pi, 0 ≤ L s ( α ) ≤ C ( 1 − s ) ( ∣ α ∣ 1 − 2 s + 1 ) , 0 < ∣ α ∣ < π ,
with a constant uniform for s 0 < s < 1 s_0<s<1 s 0 < s < 1 . Indeed, for 0 < ∣ α ∣ < 1 0<|\alpha|<1 0 < ∣ α ∣ < 1 , scale q = ∣ α ∣ r q=|\alpha|r q = ∣ α ∣ r in the local part. The remaining q q q -integral is bounded by a constant times ∣ α ∣ 1 − 2 s |\alpha|^{1-2s} ∣ α ∣ 1 − 2 s . The tail is uniform. Consequently,
sup s 0 < s < 1 ∥ L s ∥ L 1 ( − π , π ) < ∞ . \sup_{s_0<s<1}\|L_s\|_{L^1(-\pi,\pi)}<\infty. s 0 < s < 1 sup ∥ L s ∥ L 1 ( − π , π ) < ∞.
Set
f ( θ ) = ( sin θ ) 1 / 2 y ( θ ) , g ( θ ) = ( sin θ ) 1 / 2 z ( θ ) . f(\theta)=(\sin\theta)^{1/2}y(\theta),\qquad g(\theta)=(\sin\theta)^{1/2}z(\theta). f ( θ ) = ( sin θ ) 1/2 y ( θ ) , g ( θ ) = ( sin θ ) 1/2 z ( θ ) .
The kernel acting on unweighted L 2 ( 0 , π ) L^2(0,\pi) L 2 ( 0 , π ) is
L s ( θ , φ ) = L s ( θ − φ ) ( sin θ sin φ ) s − 1 . L_s(\theta,\varphi)=L_s(\theta-\varphi)(\sin\theta\sin\varphi)^{s-1}. L s ( θ , φ ) = L s ( θ − φ ) ( sin θ sin φ ) s − 1 .
We apply the weighted Schur test with p ∗ ( θ ) = ( sin θ ) − 1 / 2 p_*(\theta)=(\sin\theta)^{-1/2} p ∗ ( θ ) = ( sin θ ) − 1/2 . Let 0 < θ < ε / 4 0<\theta<\varepsilon/4 0 < θ < ε /4 , where 0 < ε < 1 / 8 0<\varepsilon<1/8 0 < ε < 1/8 is fixed. If θ / 2 < φ < 2 θ \theta/2<\varphi<2\theta θ /2 < φ < 2 θ , both sine factors are comparable to θ \theta θ . Since
( 1 − s ) θ 2 s − 2 ∫ ∣ h ∣ < θ ∣ h ∣ 1 − 2 s d h ≤ C , (1-s)\theta^{2s-2}\int_{|h|<\theta}|h|^{1-2s}\,dh\leq C, ( 1 − s ) θ 2 s − 2 ∫ ∣ h ∣ < θ ∣ h ∣ 1 − 2 s d h ≤ C ,
the kernel estimate gives
∫ θ / 2 2 θ L s ( θ , φ ) p ∗ ( φ ) p ∗ ( θ ) d φ ≤ C . \int_{\theta/2}^{2\theta}L_s(\theta,\varphi)\frac{p_*(\varphi)}{p_*(\theta)}\,d\varphi\leq C. ∫ θ /2 2 θ L s ( θ , φ ) p ∗ ( θ ) p ∗ ( φ ) d φ ≤ C .
If 0 < φ ≤ θ / 2 0<\varphi\leq\theta/2 0 < φ ≤ θ /2 , then ∣ θ − φ ∣ ≍ θ |\theta-\varphi|\asymp\theta ∣ θ − φ ∣ ≍ θ , and the singular part is bounded by
C ( 1 − s ) θ 1 / 2 − s φ s − 3 / 2 . C(1-s)\theta^{1/2-s}\varphi^{s-3/2}. C ( 1 − s ) θ 1/2 − s φ s − 3/2 .
Its integral is at most C ( 1 − s ) / ( s − 1 / 2 ) C(1-s)/(s-1/2) C ( 1 − s ) / ( s − 1/2 ) . If 2 θ ≤ φ 2\theta\leq\varphi 2 θ ≤ φ and both variables are near zero, the singular part is bounded by
C ( 1 − s ) θ s − 1 / 2 φ − s − 1 / 2 . C(1-s)\theta^{s-1/2}\varphi^{-s-1/2}. C ( 1 − s ) θ s − 1/2 φ − s − 1/2 .
Its integral is bounded by C ( 1 − s ) / ( s − 1 / 2 ) C(1-s)/(s-1/2) C ( 1 − s ) / ( s − 1/2 ) . For the nonsingular term in L s L_s L s , the entire Schur integral is bounded by
C ( 1 − s ) ( sin θ ) s − 1 / 2 ∫ 0 π ( sin φ ) s − 3 / 2 d φ ≤ C s 0 . C(1-s)(\sin\theta)^{s-1/2}\int_0^\pi(\sin\varphi)^{s-3/2}\,d\varphi\leq C_{s_0}. C ( 1 − s ) ( sin θ ) s − 1/2 ∫ 0 π ( sin φ ) s − 3/2 d φ ≤ C s 0 .
Indeed, s > s 0 > 1 / 2 s>s_0>1/2 s > s 0 > 1/2 and 0 < sin θ ≤ 1 0<\sin\theta\leq1 0 < sin θ ≤ 1 . For ε / 2 ≤ φ ≤ π − ε / 2 \varepsilon/2\leq\varphi\leq\pi-\varepsilon/2 ε /2 ≤ φ ≤ π − ε /2 , the angles are separated and L s ( θ − φ ) ≤ C ε ( 1 − s ) L_s(\theta-\varphi)\leq C_\varepsilon(1-s) L s ( θ − φ ) ≤ C ε ( 1 − s ) . For π − ε / 2 < φ < π \pi-\varepsilon/2<\varphi<\pi π − ε /2 < φ < π , they are also separated, and
∫ π − ε / 2 π L s ( θ , φ ) p ∗ ( φ ) p ∗ ( θ ) d φ ≤ C ε ( 1 − s ) ( sin θ ) s − 1 / 2 ∫ π − ε / 2 π ( sin φ ) s − 3 / 2 d φ ≤ C ε . \int_{\pi-\varepsilon/2}^{\pi}L_s(\theta,\varphi)\frac{p_*(\varphi)}{p_*(\theta)}\,d\varphi
\leq C_\varepsilon(1-s)(\sin\theta)^{s-1/2}
\int_{\pi-\varepsilon/2}^{\pi}(\sin\varphi)^{s-3/2}\,d\varphi
\leq C_\varepsilon. ∫ π − ε /2 π L s ( θ , φ ) p ∗ ( θ ) p ∗ ( φ ) d φ ≤ C ε ( 1 − s ) ( sin θ ) s − 1/2 ∫ π − ε /2 π ( sin φ ) s − 3/2 d φ ≤ C ε .
The range between 2 θ 2\theta 2 θ and ε / 2 \varepsilon/2 ε /2 is covered by the preceding near-zero estimate. Reflection treats θ \theta θ near π \pi π . For θ \theta θ in a fixed middle interval, the near-diagonal term is bounded by ∥ L s ∥ 1 \lVert L_s\rVert_1 ∥ L s ∥ 1 . On its complement the kernel is at most C ε ( 1 − s ) C_\varepsilon(1-s) C ε ( 1 − s ) ; the endpoint integral is finite because s 0 > 1 / 2 s_0 > 1/2 s 0 > 1/2 . Thus
sup θ 1 p ∗ ( θ ) ∫ 0 π L s ( θ , φ ) p ∗ ( φ ) d φ ≤ C . \sup_{\theta}\frac{1}{p_*(\theta)}\int_0^\pi\mathcal{L}_s(\theta,\varphi)p_*(\varphi)\,d\varphi\le C. θ sup p ∗ ( θ ) 1 ∫ 0 π L s ( θ , φ ) p ∗ ( φ ) d φ ≤ C .
The kernel is symmetric. Thus the transposed Schur bound also holds. Schur’s test proves
∣ t s [ y , z ] ∣ ≤ C ∥ y ∥ L 2 ( sin θ d θ ) ∥ z ∥ L 2 ( sin θ d θ ) . |t_s[y,z]| \le C\lVert y\rVert_{L^2(\sin\theta\,d\theta)}\lVert z\rVert_{L^2(\sin\theta\,d\theta)}. ∣ t s [ y , z ] ∣ ≤ C ∥ y ∥ L 2 ( s i n θ d θ ) ∥ z ∥ L 2 ( s i n θ d θ ) .
The endpoint estimate (104) , used with σ = 2 \sigma=2 σ = 2 , gives
∫ ( 0 , δ ) ∪ ( π − δ , π ) sin θ ∣ y n ∣ 2 d θ ≤ C δ . \int_{(0,\delta)\cup(\pi-\delta,\pi)} \sin\theta\,|y_n|^2\,d\theta\le C\delta. ∫ ( 0 , δ ) ∪ ( π − δ , π ) sin θ ∣ y n ∣ 2 d θ ≤ C δ .
Together with local convergence, this proves
y n ⟶ y in L 2 ( ( 0 , π ) , sin θ d θ ) . y_n \longrightarrow y \quad\text{in } L^2((0,\pi),\sin\theta\,d\theta). y n ⟶ y in L 2 (( 0 , π ) , sin θ d θ ) .
For a fixed v ∈ C c ∞ ( 0 , π ) v\in C^\infty_c(0,\pi) v ∈ C c ∞ ( 0 , π ) , the moment computation in the core part of Lemma 8.6 gives
t s [ v , v ] ⟶ ∫ 0 π sin θ ∣ v ∣ 2 d θ . t_s[v,v]\longrightarrow\int_0^\pi\sin\theta\,|v|^2\,d\theta. t s [ v , v ] ⟶ ∫ 0 π sin θ ∣ v ∣ 2 d θ .
Write ∥ ⋅ ∥ ∗ = ∥ ⋅ ∥ L 2 ( sin θ d θ ) \lVert\cdot\rVert_*=\lVert\cdot\rVert_{L^2(\sin\theta\,d\theta)} ∥ ⋅ ∥ ∗ = ∥ ⋅ ∥ L 2 ( s i n θ d θ ) . For a fixed core function v v v , the uniform Schur bound gives
∣ t s n [ y n , y n ] − t s n [ v , v ] ∣ ≤ C ∥ y n − v ∥ ∗ ( ∥ y n ∥ ∗ + ∥ v ∥ ∗ ) , |t_{s_n}[y_n,y_n]-t_{s_n}[v,v]|\le C\lVert y_n-v\rVert_*(\lVert y_n\rVert_*+\lVert v\rVert_*), ∣ t s n [ y n , y n ] − t s n [ v , v ] ∣ ≤ C ∥ y n − v ∥ ∗ (∥ y n ∥ ∗ + ∥ v ∥ ∗ ) ,
∣ ∥ y ∥ ∗ 2 − ∥ v ∥ ∗ 2 ∣ ≤ ∥ y − v ∥ ∗ ( ∥ y ∥ ∗ + ∥ v ∥ ∗ ) . \bigl|\lVert y\rVert_*^2-\lVert v\rVert_*^2\bigr|\le\lVert y-v\rVert_*(\lVert y\rVert_*+\lVert v\rVert_*). ∥ y ∥ ∗ 2 − ∥ v ∥ ∗ 2 ≤ ∥ y − v ∥ ∗ (∥ y ∥ ∗ + ∥ v ∥ ∗ ) .
First let n → ∞ n\to\infty n → ∞ , using y n → y y_n\to y y n → y in this norm and the core moment limit for v v v . Then choose v v v with ∥ y − v ∥ ∗ \lVert y-v\rVert_* ∥ y − v ∥ ∗ arbitrarily small. The two bounds prove the asserted limit. □
Proof of Theorem 1.6. After the ground-state conjugation, the denominator of the zero-fiber Rayleigh quotient is the fixed H 1 , τ H_{1,\tau} H 1 , τ norm. Hence
2 ( 1 − s ) π Λ s , τ = inf y ≠ 0 a s , 0 [ y ] ∥ y ∥ H 1 , τ 2 . \frac{2(1-s)}{\pi}\Lambda_{s,\tau}=\inf_{y\ne0}\frac{a_{s,0}[y]}{\lVert y\rVert_{H_{1,\tau}}^2}. π 2 ( 1 − s ) Λ s , τ = y = 0 inf ∥ y ∥ H 1 , τ 2 a s , 0 [ y ] .
A recovery sequence for the local minimizer gives
lim sup s ↑ 1 2 ( 1 − s ) π Λ s , τ ≤ Λ 1 , τ . \limsup_{s\uparrow1}\frac{2(1-s)}{\pi}\Lambda_{s,\tau}\le\Lambda_{1,\tau}. s ↑ 1 lim sup π 2 ( 1 − s ) Λ s , τ ≤ Λ 1 , τ .
For the opposite inequality, choose a sequence s n ↑ 1 s_n\uparrow1 s n ↑ 1 such that
2 ( 1 − s n ) π Λ s n , τ ⟶ lim inf s ↑ 1 2 ( 1 − s ) π Λ s , τ , \frac{2(1-s_n)}{\pi}\Lambda_{s_n,\tau}\longrightarrow\liminf_{s\uparrow1}\frac{2(1-s)}{\pi}\Lambda_{s,\tau}, π 2 ( 1 − s n ) Λ s n , τ ⟶ s ↑ 1 lim inf π 2 ( 1 − s ) Λ s , τ ,
and put y n = ( sin θ ) 1 / 2 − s n ϕ s n , τ y_n=(\sin\theta)^{1/2-s_n}\phi_{s_n,\tau} y n = ( sin θ ) 1/2 − s n ϕ s n , τ . The functions y n y_n y n are normalized in H 1 , τ H_{1,\tau} H 1 , τ . The preceding upper bound makes a s n , 0 [ y n ] a_{s_n,0}[y_n] a s n , 0 [ y n ] bounded. Compact Mosco convergence therefore gives, after passage to a subsequence, y n → y y_n\to y y n → y strongly in H 1 , τ H_{1,\tau} H 1 , τ . Thus ∥ y ∥ H 1 , τ = 1 \lVert y\rVert_{H_{1,\tau}}=1 ∥ y ∥ H 1 , τ = 1 , and the Mosco lower bound gives
lim inf s ↑ 1 2 ( 1 − s ) π Λ s , τ = lim n → ∞ a s n , 0 [ y n ] ≥ l 1 , 0 [ y ] ≥ Λ 1 , τ . \liminf_{s\uparrow1}\frac{2(1-s)}{\pi}\Lambda_{s,\tau}
=\lim_{n\to\infty}a_{s_n,0}[y_n]\ge l_{1,0}[y]\ge\Lambda_{1,\tau}. s ↑ 1 lim inf π 2 ( 1 − s ) Λ s , τ = n → ∞ lim a s n , 0 [ y n ] ≥ l 1 , 0 [ y ] ≥ Λ 1 , τ .
This proves convergence of the first eigenvalue. Finally, take an arbitrary sequence s n ↑ 1 s_n\uparrow1 s n ↑ 1 . Every subsequence of the corresponding positive normalized y n y_n y n has a further subsequence converging strongly to a nonnegative normalized local minimizer. The simplicity of the local first eigenvalue identifies that limit with y 1 , τ y_{1,\tau} y 1 , τ . Hence the whole family converges strongly to y 1 , τ y_{1,\tau} y 1 , τ .
For the higher eigenvalues, we use the min–max principle. For the upper bound, fix ε > 0 \varepsilon>0 ε > 0 . Approximate an orthonormal basis of the first j j j local eigenspaces in the local graph norm by angular core functions. Their span E ε E_\varepsilon E ε has dimension j j j and largest local Rayleigh quotient at most λ j ( 1 ) ( ξ ) + ε \lambda_j^{(1)}(\xi)+\varepsilon λ j ( 1 ) ( ξ ) + ε . Core convergence is uniform on its unit sphere. The min–max principle, followed by ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 , gives the upper bound.
For the lower bound, write λ j ( s ) ( ξ ) \lambda_j^{(s)}(\xi) λ j ( s ) ( ξ ) for the nonlocal eigenvalue in this paragraph. Choose s n ↑ 1 s_n \uparrow1 s n ↑ 1 so that
L : = lim n → ∞ 2 ( 1 − s n ) π λ j ( s n ) ( ξ ) = lim inf s ↑ 1 2 ( 1 − s ) π λ j ( s ) ( ξ ) . L := \lim_{n\to\infty} \frac{2(1-s_n)}{\pi}\lambda_j^{(s_n)}(\xi) = \liminf_{s\uparrow1} \frac{2(1-s)}{\pi}\lambda_j^{(s)}(\xi). L := n → ∞ lim π 2 ( 1 − s n ) λ j ( s n ) ( ξ ) = s ↑ 1 lim inf π 2 ( 1 − s ) λ j ( s ) ( ξ ) .
For 1 ≤ k ≤ j 1 \leq k \leq j 1 ≤ k ≤ j , let e k , n e_{k,n} e k , n be orthonormal eigenfunctions of a s n , ξ \mathbf{a}_{s_n,\xi} a s n , ξ associated with its first j j j eigenvalues. The upper bound just proved makes their form values bounded. By compactness and a diagonal extraction, e k , n → e k e_{k,n} \to e_k e k , n → e k strongly in H 1 , τ H_{1,\tau} H 1 , τ for every k ≤ j k \leq j k ≤ j ; the limits are orthonormal. For every c = ( c 1 , … , c j ) \mathbf{c}=(c_1,\ldots,c_j) c = ( c 1 , … , c j ) with rational real and imaginary parts, the Mosco lower bound and the diagonal spectral representation give
ℓ 1 , ξ [ ∑ k = 1 j c k e k ] ≤ lim inf n → ∞ ∑ k = 1 j 2 ( 1 − s n ) π λ k ( s n ) ( ξ ) ∣ c k ∣ 2 ≤ L ∑ k = 1 j ∣ c k ∣ 2 . \ell_{1,\xi}\left[\sum_{k=1}^{j}c_ke_k\right]\leq\liminf_{n\to\infty}\sum_{k=1}^{j}\frac{2(1-s_n)}{\pi}\lambda_k^{(s_n)}(\xi)|c_k|^2\leq L\sum_{k=1}^{j}|c_k|^2. ℓ 1 , ξ [ k = 1 ∑ j c k e k ] ≤ n → ∞ lim inf k = 1 ∑ j π 2 ( 1 − s n ) λ k ( s n ) ( ξ ) ∣ c k ∣ 2 ≤ L k = 1 ∑ j ∣ c k ∣ 2 .
By finite-dimensional density the inequality holds for all c ∈ C j \mathbf{c}\in\mathbb{C}^j c ∈ C j . The min–max principle gives λ j ( 1 ) ( ξ ) ≤ L \lambda_j^{(1)}(\xi)\leq L λ j ( 1 ) ( ξ ) ≤ L .
For the last assertion, put
y s = ( sin θ ) 1 / 2 − s ϕ s , τ . y_s=(\sin\theta)^{1/2-s}\phi_{s,\tau}. y s = ( sin θ ) 1/2 − s ϕ s , τ .
The preceding part of the proof gives y s → y 1 , τ y_s\to y_{1,\tau} y s → y 1 , τ locally in L 2 ( 0 , π ) L^2(0,\pi) L 2 ( 0 , π ) and
sup s n e a r 1 { a s , 0 [ y s ] + ∥ y s ∥ H 1 , τ 2 } < ∞ . \sup_{s\ \mathrm{near}\ 1}\{\mathbf{a}_{s,0}[y_s]+\|y_s\|_{H_{1,\tau}}^2\}<\infty. s near 1 sup { a s , 0 [ y s ] + ∥ y s ∥ H 1 , τ 2 } < ∞.
Moreover, the definition of κ s , τ \kappa_{s,\tau} κ s , τ gives
2 ( 1 − s ) π κ s , τ = t s [ y s , y s ] . \frac{2(1-s)}{\pi}\kappa_{s,\tau}=t_s[y_s,y_s]. π 2 ( 1 − s ) κ s , τ = t s [ y s , y s ] .
Lemma 8.7 now yields
lim s ↑ 1 2 ( 1 − s ) π κ s , τ = ∫ 0 π sin θ y 1 , τ 2 d θ . \lim_{s\uparrow1}\frac{2(1-s)}{\pi}\kappa_{s,\tau}=\int_0^\pi\sin\theta\,y_{1,\tau}^2\,d\theta. s ↑ 1 lim π 2 ( 1 − s ) κ s , τ = ∫ 0 π sin θ y 1 , τ 2 d θ .
Corollary 8.8 (Limit of the sharp deficit). For every τ > 0 \tau>0 τ > 0 and u ∈ C ∞ ( R 2 ) u\in C^\infty(\mathbb{R}^2) u ∈ C ∞ ( R 2 ) ,
lim s ↑ 1 2 ( 1 − s ) π { H s [ u ] − Λ s , τ R s , τ [ u ] } = H 1 [ u ] − Λ 1 , τ R 1 , τ [ u ] . \lim_{s\uparrow1}\frac{2(1-s)}{\pi}\{\mathcal{H}_s[u]-\Lambda_{s,\tau}\mathcal{R}_{s,\tau}[u]\}=\mathcal{H}_1[u]-\Lambda_{1,\tau}\mathcal{R}_{1,\tau}[u]. s ↑ 1 lim π 2 ( 1 − s ) { H s [ u ] − Λ s , τ R s , τ [ u ]} = H 1 [ u ] − Λ 1 , τ R 1 , τ [ u ] . For every d ≥ 2 d\geq2 d ≥ 2 ,
lim s ↑ 1 ( 1 − s ) Λ d , s , τ K d , 2 = Λ 1 , τ . \lim_{s\uparrow1}\frac{(1-s)\Lambda_{d,s,\tau}}{K_{d,2}}=\Lambda_{1,\tau}. s ↑ 1 lim K d , 2 ( 1 − s ) Λ d , s , τ = Λ 1 , τ . Proof. The first assertion follows from Proposition 8.1 and Theorem 1.6. For the second one, use Λ d , s , τ = A d , s Λ s , τ \Lambda_{d,s,\tau}=A_{d,s}\Lambda_{s,\tau} Λ d , s , τ = A d , s Λ s , τ , and
A d , s ⟶ π ( d − 2 ) / 2 Γ ( ( d + 2 ) / 2 ) = K d , 2 K 2 , 2 . A_{d,s}\longrightarrow\frac{\pi^{(d-2)/2}}{\Gamma((d+2)/2)}=\frac{K_{d,2}}{K_{2,2}}. A d , s ⟶ Γ (( d + 2 ) /2 ) π ( d − 2 ) /2 = K 2 , 2 K d , 2 . Since K 2 , 2 = π / 2 K_{2,2}=\pi/2 K 2 , 2 = π /2 , the conclusion follows.
Remark 8.9. For τ = 1 \tau=1 τ = 1 , Lemma 8.2 gives the Maz’ya–Shaposhnikova eigenvalue
Λ 1 , 1 = min y ≠ 0 ∫ 0 π sin θ ( ∣ y ′ ∣ 2 + ∣ y ∣ 2 / 4 ) d θ ∫ 0 π ∣ y ∣ 2 d θ . \Lambda_{1,1}=\min_{y\ne0}\frac{\int_0^\pi\sin\theta(|y'|^2+|y|^2/4)\,d\theta}{\int_0^\pi|y|^2\,d\theta}. Λ 1 , 1 = y = 0 min ∫ 0 π ∣ y ∣ 2 d θ ∫ 0 π sin θ ( ∣ y ′ ∣ 2 + ∣ y ∣ 2 /4 ) d θ .
For τ = 2 \tau= 2 τ = 2 , Proposition 8.3 gives the full local spectrum and an attained local stability coefficient. In contrast, Theorem 1.4 proves non-attainment for every s < 1 s < 1 s < 1 ; both statements use the same fixed multiplier h h h . Convergence of Γ s , τ h p \Gamma^{\mathrm{hp}}_{s,\tau} Γ s , τ hp would require both large-frequency coercivity uniform in s s s and uniform control of the generalized bands at zero. The fixed-s s s estimates above do not give these bounds.
Appendix A. The spectrally periodized cylinder
Fix 0 < s < 1 0 < s < 1 0 < s < 1 and τ > 0 \tau> 0 τ > 0 . For L > 0 L > 0 L > 0 , let C L = ( R / L Z ) × ( 0 , π ) C_L = (\mathbb{R}/L\mathbb{Z}) \times(0,\pi) C L = ( R / L Z ) × ( 0 , π ) . Its form is defined by the discrete fibers b s , 2 π k / L b_{s,2\pi k/L} b s , 2 π k / L . Let M L \mathcal{M}_L M L be the one-dimensional space of functions c ϕ s , τ ( θ ) c\phi_{s,\tau}(\theta) c ϕ s , τ ( θ ) . The optimal stability constant for the spectrally periodized cylinder is
G s , τ ( L ) = inf w ∉ M L q s , L [ w ] − Λ s , τ ∥ w ∥ L 2 ( C L , m s , τ ) 2 dist L 2 ( C L , m s , τ ) ( w , M L ) 2 (108) G_{s,\tau}(L) = \inf_{w \notin\mathcal{M}_L} \frac{q_{s,L}[w] - \Lambda_{s,\tau}\lVert w\rVert_{L^2(C_L,m_{s,\tau})}^2}{\operatorname{dist}_{L^2(C_L,m_{s,\tau})}(w,\mathcal{M}_L)^2} \tag*{(108)} G s , τ ( L ) = w ∈ / M L inf dist L 2 ( C L , m s , τ ) ( w , M L ) 2 q s , L [ w ] − Λ s , τ ∥ w ∥ L 2 ( C L , m s , τ ) 2 (108)
Let L > 0 L > 0 L > 0 . A function on C L C_L C L has the expansion
w ( t , θ ) = L − 1 / 2 ∑ k ∈ Z e 2 π i k t / L ψ k ( θ ) . w(t,\theta) = L^{-1/2}\sum_{k\in\mathbb{Z}} e^{2\pi ikt/L}\psi_k(\theta). w ( t , θ ) = L − 1/2 k ∈ Z ∑ e 2 π ik t / L ψ k ( θ ) .
Define
q s , L [ w ] = ∑ k ∈ Z b s , 2 π k / L [ ψ k ] (109) q_{s,L}[w] = \sum_{k\in\mathbb{Z}} b_{s,2\pi k/L}[\psi_k] \tag*{(109)} q s , L [ w ] = k ∈ Z ∑ b s , 2 π k / L [ ψ k ] (109)
and
∥ w ∥ L 2 ( C L , m s , τ ) 2 = ∑ k ∈ Z ∥ ψ k ∥ L 2 ( m s , τ ) 2 . \lVert w\rVert_{L^2(C_L,m_{s,\tau})}^2 = \sum_{k\in\mathbb{Z}}\lVert\psi_k\rVert_{L^2(m_{s,\tau})}^2. ∥ w ∥ L 2 ( C L , m s , τ ) 2 = k ∈ Z ∑ ∥ ψ k ∥ L 2 ( m s , τ ) 2 .
The periodic form domain is
X s , L = { w : ψ k ∈ Y s , ∑ k ∈ Z ( b s , 2 π k / L [ ψ k ] + ∥ ψ k ∥ H s , τ 2 ) < ∞ } . \mathcal{X}_{s,L} = \left\{w : \psi_k \in\mathcal{Y}_s,\ \sum_{k\in\mathbb{Z}}\left(b_{s,2\pi k/L}[\psi_k] + \lVert\psi_k\rVert_{H_{s,\tau}}^2\right) < \infty\right\}. X s , L = { w : ψ k ∈ Y s , k ∈ Z ∑ ( b s , 2 π k / L [ ψ k ] + ∥ ψ k ∥ H s , τ 2 ) < ∞ } .
The graph norm makes this a Hilbert space. Finite Fourier sums with coefficients in C c ∞ ( 0 , π ) C^\infty_c(0,\pi) C c ∞ ( 0 , π ) form a core.
Proposition A.1 (Periodic extremal). The sharp spectrally periodized quotient is
inf w ≠ 0 q s , L [ w ] ∥ w ∥ L 2 ( C L , m s , τ ) 2 = Λ s , τ . \inf_{w\ne0}\frac{q_{s,L}[w]}{\lVert w\rVert_{L^2(C_L,m_{s,\tau})}^2} = \Lambda_{s,\tau}. w = 0 inf ∥ w ∥ L 2 ( C L , m s , τ ) 2 q s , L [ w ] = Λ s , τ . It is attained precisely by w ( t , θ ) = c ϕ s , τ ( θ ) w(t,\theta) = c\phi_{s,\tau}(\theta) w ( t , θ ) = c ϕ s , τ ( θ ) .
Proof. Each summand in (A.2) is bounded below by λ 1 ( 2 π k / L ) ∥ ψ k ∥ 2 \lambda_1(2\pi k/L)\lVert\psi_k\rVert^2 λ 1 ( 2 π k / L ) ∥ ψ k ∥ 2 . Lemma 4.7 gives λ 1 ( 2 π k / L ) ≥ λ 1 ( 0 ) \lambda_1(2\pi k/L) \ge\lambda_1(0) λ 1 ( 2 π k / L ) ≥ λ 1 ( 0 ) . Equality is possible only when k = 0 k = 0 k = 0 . In that fiber, Lemma 4.6 shows that equality is possible only for a multiple of ϕ s , τ \phi_{s,\tau} ϕ s , τ . □
Theorem A.2 (Attainment on the spectrally periodized cylinder). Let 0 < s < 1 0 < s < 1 0 < s < 1 , τ > 0 \tau> 0 τ > 0 , and L > 0 L > 0 L > 0 . The constant G s , τ ( L ) G_{s,\tau}(L) G s , τ ( L ) is positive and is attained. If λ 2 ( 0 ) \lambda_2(0) λ 2 ( 0 ) denotes the second eigenvalue of the zero fiber, then
G s , τ ( L ) = min { λ 2 ( 0 ) − λ 1 ( 0 ) , inf k ∈ Z ∖ { 0 } [ λ 1 ( 2 π k / L ) − λ 1 ( 0 ) ] } . (110) G_{s,\tau}(L) = \min\left\{\lambda_2(0) - \lambda_1(0),\ \inf_{k\in\mathbb{Z}\setminus\{0\}}\left[\lambda_1(2\pi k/L) - \lambda_1(0)\right]\right\}. \tag*{(110)} G s , τ ( L ) = min { λ 2 ( 0 ) − λ 1 ( 0 ) , k ∈ Z ∖ { 0 } inf [ λ 1 ( 2 π k / L ) − λ 1 ( 0 ) ] } . (110) Moreover,
G s , τ ( L ) = 4 π 2 κ s , τ L 2 + O ( L − 4 ) as L → ∞ . (111) G_{s,\tau}(L) = \frac{4\pi^2\kappa_{s,\tau}}{L^2} + O(L^{-4}) \quad\text{as } L \to\infty. \tag*{(111)} G s , τ ( L ) = L 2 4 π 2 κ s , τ + O ( L − 4 ) as L → ∞. (111) For all sufficiently large L L L , the orthogonal component w − P M L w w - P_{\mathcal{M}_L}w w − P M L w of every extremizer is supported in the modes k = ± 1 k = \pm1 k = ± 1 . A complex representative orthogonal to M L \mathcal{M}_L M L may use one mode. A real representative has conjugate coefficients.
Proof of Theorem A.2. Write w = P M L w + w ⊥ w=P_{\mathcal{M}_L}w+w^\perp w = P M L w + w ⊥ . Since the zero-mode ground state solves (75) , the numerator in (A.1) is unchanged when P M L w P_{\mathcal{M}_L}w P M L w is removed. The denominator is ∥ w ⊥ ∥ 2 \lVert w^\perp\rVert^2 ∥ w ⊥ ∥ 2 . Hence it is enough to take w ⊥ M L w\perp\mathcal{M}_L w ⊥ M L . Then ψ 0 ⊥ φ s , τ \psi_0\perp\varphi_{s,\tau} ψ 0 ⊥ φ s , τ . The spectral theorem gives
b s , 0 [ ψ 0 ] − Λ s , τ ∥ ψ 0 ∥ 2 ≥ ( λ 2 ( 0 ) − λ 1 ( 0 ) ) ∥ ψ 0 ∥ 2 . b_{s,0}[\psi_0]-\Lambda_{s,\tau}\lVert\psi_0\rVert^2\geq\left(\lambda_2(0)-\lambda_1(0)\right)\lVert\psi_0\rVert^2. b s , 0 [ ψ 0 ] − Λ s , τ ∥ ψ 0 ∥ 2 ≥ ( λ 2 ( 0 ) − λ 1 ( 0 ) ) ∥ ψ 0 ∥ 2 .
For k ≠ 0 k\neq0 k = 0 ,
b s , 2 π k / L [ ψ k ] − Λ s , τ ∥ ψ k ∥ 2 ≥ ( λ 1 ( 2 π k / L ) − λ 1 ( 0 ) ) ∥ ψ k ∥ 2 . b_{s,2\pi k/L}[\psi_k]-\Lambda_{s,\tau}\lVert\psi_k\rVert^2\geq\left(\lambda_1(2\pi k/L)-\lambda_1(0)\right)\lVert\psi_k\rVert^2. b s , 2 π k / L [ ψ k ] − Λ s , τ ∥ ψ k ∥ 2 ≥ ( λ 1 ( 2 π k / L ) − λ 1 ( 0 ) ) ∥ ψ k ∥ 2 .
Summation proves the lower bound in (A.3). Equality is obtained by a second eigenfunction of the zero fiber or by a first eigenfunction of a minimizing nonzero fiber. Such a fiber exists because λ 1 ( 2 π k / L ) → ∞ \lambda_1(2\pi k/L)\to\infty λ 1 ( 2 π k / L ) → ∞ as ∣ k ∣ → ∞ |k|\to\infty ∣ k ∣ → ∞ . Thus the constant is positive and attained. Proposition 6.1 gives
λ 1 ( 2 π / L ) − λ 1 ( 0 ) = 4 π 2 κ s , τ L 2 + O ( L − 4 ) . (112) \lambda_1(2\pi/L)-\lambda_1(0)=\frac{4\pi^2\kappa_{s,\tau}}{L^2}+O(L^{-4}). \tag*{(112)} λ 1 ( 2 π / L ) − λ 1 ( 0 ) = L 2 4 π 2 κ s , τ + O ( L − 4 ) . (112)
Differentiating the even analytic expansion gives
λ 1 ′ ( ξ ) = 2 κ s , τ ξ + O ( ξ 3 ) . \lambda_1'(\xi)=2\kappa_{s,\tau}\xi+O(\xi^3). λ 1 ′ ( ξ ) = 2 κ s , τ ξ + O ( ξ 3 ) .
Choose δ > 0 \delta>0 δ > 0 so that λ 1 ′ ( ξ ) > 0 \lambda_1'(\xi)>0 λ 1 ′ ( ξ ) > 0 for 0 < ξ < δ 0<\xi<\delta 0 < ξ < δ . By Lemma 6.4, there is c δ > 0 c_\delta>0 c δ > 0 with λ 1 ( ξ ) − λ 1 ( 0 ) ≥ c δ \lambda_1(\xi)-\lambda_1(0)\geq c_\delta λ 1 ( ξ ) − λ 1 ( 0 ) ≥ c δ for ∣ ξ ∣ ≥ δ |\xi|\geq\delta ∣ ξ ∣ ≥ δ . For sufficiently large L L L ,
2 π L < δ , λ 1 ( 2 π / L ) − λ 1 ( 0 ) < min { c δ , λ 2 ( 0 ) − λ 1 ( 0 ) } . \frac{2\pi}{L}<\delta,\qquad\lambda_1(2\pi/L)-\lambda_1(0)<\min\{c_\delta,\lambda_2(0)-\lambda_1(0)\}. L 2 π < δ , λ 1 ( 2 π / L ) − λ 1 ( 0 ) < min { c δ , λ 2 ( 0 ) − λ 1 ( 0 )} .
If ∣ k ∣ ≥ 2 |k|\geq2 ∣ k ∣ ≥ 2 and 2 π ∣ k ∣ / L < δ 2\pi|k|/L<\delta 2 π ∣ k ∣/ L < δ , strict monotonicity gives a larger gap than at ∣ k ∣ = 1 |k|=1 ∣ k ∣ = 1 . If 2 π ∣ k ∣ / L ≥ δ 2\pi|k|/L\geq\delta 2 π ∣ k ∣/ L ≥ δ , the gap is at least c δ c_\delta c δ . Thus only k = ± 1 k=\pm1 k = ± 1 minimize the nonzero modal gap. Equations (A.3) and (A.5) prove (A.4). For the equality cases, subtract G s , τ ( L ) ∥ w ⊥ ∥ 2 G_{s,\tau}(L)\lVert w^\perp\rVert^2 G s , τ ( L ) ∥ w ⊥ ∥ 2 from the numerator. Each modal term is nonnegative. For large L L L , equality forces all modes except k = ± 1 k=\pm1 k = ± 1 to vanish, and each remaining coefficient belongs to the corresponding first eigenspace. The two complex coefficients are independent; for real functions they are complex conjugates. An arbitrary element of M L \mathcal{M}_L M L may be added without changing the quotient. □ \square □
Angular Ritz approximation
Dima and Ignat [4 ] obtain sharp finite element error rates for the full-space fractional Hardy constant. The compact angular problem here has the following Ritz convergence; no mesh error rate is claimed.
Proposition B.1 (Angular Ritz approximation). Let Y N ⊂ Y s Y_N\subset\mathcal{Y}_s Y N ⊂ Y s be nested finite-dimensional spaces of dimension at least two whose union is form-dense. Let Λ N \Lambda_N Λ N and λ 2 , N \lambda_{2,N} λ 2 , N be their first two Ritz eigenvalues, and let φ N \varphi_N φ N be a normalized first Ritz eigenvector with ⟨ φ N , φ s , τ ⟩ H s , τ ≥ 0 \langle\varphi_N,\varphi_{s,\tau}\rangle_{\mathcal{H}_{s,\tau}}\geq0 ⟨ φ N , φ s , τ ⟩ H s , τ ≥ 0 . Then
Λ N ↓ Λ s , τ , λ 2 , N ↓ λ 2 ( 0 ) , φ N → φ s , τ in Y s . \Lambda_N\downarrow\Lambda_{s,\tau},\qquad\lambda_{2,N}\downarrow\lambda_2(0),\qquad\varphi_N\to\varphi_{s,\tau}\ \text{in }\mathcal{Y}_s. Λ N ↓ Λ s , τ , λ 2 , N ↓ λ 2 ( 0 ) , φ N → φ s , τ in Y s . Consequently, λ 2 , N − Λ N → Γ s , τ a n g \lambda_{2,N}-\Lambda_N\to\Gamma_{s,\tau}^{\mathrm{ang}} λ 2 , N − Λ N → Γ s , τ ang . If κ N \kappa_N κ N is obtained from (11) by replacing φ s , τ ( θ ) φ s , τ ( φ ) \varphi_{s,\tau}(\theta)\varphi_{s,\tau}(\varphi) φ s , τ ( θ ) φ s , τ ( φ ) with φ N ( θ ) φ N ( φ ) \varphi_N(\theta)\varphi_N(\varphi) φ N ( θ ) φ N ( φ ) , then κ N → κ s , τ \kappa_N\to\kappa_{s,\tau} κ N → κ s , τ .
Proof. Nesting and the min–max principle give monotone eigenvalue bounds. Approximate the first two angular eigenfunctions in the form norm by vectors in Y N Y_N Y N . Their Gram and form matrices tend to I 2 I_2 I 2 and diag ( λ 1 ( 0 ) , λ 2 ( 0 ) ) \operatorname{diag}(\lambda_1(0),\lambda_2(0)) diag ( λ 1 ( 0 ) , λ 2 ( 0 )) , proving the eigenvalue limits. Compactness and simplicity identify every weak form limit of φ N \varphi_N φ N with φ s , τ \varphi_{s,\tau} φ s , τ . Moreover,
b s , 0 [ φ N ] + 1 = Λ N + 1 ⟶ b s , 0 [ φ s , τ ] + 1. b_{s,0}[\varphi_N]+1=\Lambda_N+1\longrightarrow b_{s,0}[\varphi_{s,\tau}]+1. b s , 0 [ φ N ] + 1 = Λ N + 1 ⟶ b s , 0 [ φ s , τ ] + 1.
The form norm is induced by b s , 0 ( ⋅ , ⋅ ) + ⟨ ⋅ , ⋅ ⟩ H s , τ b_{s,0}(\cdot,\cdot)+\langle\cdot,\cdot\rangle_{H^{s,\tau}} b s , 0 ( ⋅ , ⋅ ) + ⟨ ⋅ , ⋅ ⟩ H s , τ . Weak convergence and convergence of these squared norms give
∥ ϕ N − ϕ s , τ ∥ Y s 2 = ∥ ϕ N ∥ Y s 2 + ∥ ϕ s , τ ∥ Y s 2 − 2 Re ⟨ ϕ N , ϕ s , τ ⟩ Y s ⟶ 0. \|\phi_N-\phi_{s,\tau}\|_{Y_s}^2=\|\phi_N\|_{Y_s}^2+\|\phi_{s,\tau}\|_{Y_s}^2-2\operatorname{Re}\langle\phi_N,\phi_{s,\tau}\rangle_{Y_s}\longrightarrow0. ∥ ϕ N − ϕ s , τ ∥ Y s 2 = ∥ ϕ N ∥ Y s 2 + ∥ ϕ s , τ ∥ Y s 2 − 2 Re ⟨ ϕ N , ϕ s , τ ⟩ Y s ⟶ 0.
Every subsequence has the same limit by the phase normalization, so this holds for the whole sequence. Subtraction of the eigenvalue limits gives the gap limit. By (53) and the integrable second-moment kernel in (72) , Young’s inequality gives
∣ κ N − κ s , τ ∣ ≤ C ( ∥ ϕ N ∥ 2 + ∥ ϕ s , τ ∥ 2 ) ∥ ϕ N − ϕ s , τ ∥ 2 ⟶ 0. |\kappa_N-\kappa_{s,\tau}|\le C(\|\phi_N\|_2+\|\phi_{s,\tau}\|_2)\|\phi_N-\phi_{s,\tau}\|_2\longrightarrow0. ∣ κ N − κ s , τ ∣ ≤ C ( ∥ ϕ N ∥ 2 + ∥ ϕ s , τ ∥ 2 ) ∥ ϕ N − ϕ s , τ ∥ 2 ⟶ 0.
Subcritical nonlinear coercivity
Throughout this appendix, p ≥ 2 p\ge2 p ≥ 2 and all functions are real-valued. Put
ω s , p ( x ) = x d ( s p − 1 ) / p . \omega_{s,p}(x)=x_d^{(sp-1)/p}. ω s , p ( x ) = x d ( s p − 1 ) / p .
For u ∈ C c ∞ ( R + d ) u\in C_c^\infty(\mathbb{R}_+^d) u ∈ C c ∞ ( R + d ) , define
G d , s , p [ u ] = ∫ R + d ∫ R + d ∣ u ( x ) ω s , p ( x ) − u ( y ) ω s , p ( y ) ∣ p ω s , p ( x ) p / 2 ω s , p ( y ) p / 2 ∣ x − y ∣ d + s p d x d y . \mathcal{G}_{d,s,p}[u]=\int_{\mathbb{R}_+^d}\int_{\mathbb{R}_+^d}\left|\frac{u(x)}{\omega_{s,p}(x)}-\frac{u(y)}{\omega_{s,p}(y)}\right|^p\frac{\omega_{s,p}(x)^{p/2}\omega_{s,p}(y)^{p/2}}{|x-y|^{d+sp}}\,dx\,dy. G d , s , p [ u ] = ∫ R + d ∫ R + d ω s , p ( x ) u ( x ) − ω s , p ( y ) u ( y ) p ∣ x − y ∣ d + s p ω s , p ( x ) p /2 ω s , p ( y ) p /2 d x d y .
The functional is p p p -homogeneous. If G d , s , p [ u ] = 0 \mathcal{G}_{d,s,p}[u]=0 G d , s , p [ u ] = 0 , Fubini’s theorem gives u / ω s , p = c u/\omega_{s,p}=c u / ω s , p = c almost everywhere. Since u u u has compact support, c = 0 c=0 c = 0 . Thus its kernel on the test core is trivial.
Lemma C.1 (Nonlinear Hardy ground-state remainder). Let d ≥ 2 d\ge2 d ≥ 2 , 0 < s < 1 0<s<1 0 < s < 1 , and p ≥ 2 p\ge2 p ≥ 2 . Then
H d , s , p [ u ] ≥ c p G d , s , p [ u ] for u ∈ C c ∞ ( R + d ) , \mathcal{H}_{d,s,p}[u]\ge c_p\mathcal{G}_{d,s,p}[u]\qquad\text{for }u\in C_c^\infty(\mathbb{R}_+^d), H d , s , p [ u ] ≥ c p G d , s , p [ u ] for u ∈ C c ∞ ( R + d ) , where
c p = min 0 < t < 1 / 2 { ( 1 − t ) p − t p + p t p − 1 } > 0. c_p=\min_{0<t<1/2}\left\{(1-t)^p-t^p+pt^{p-1}\right\}>0. c p = 0 < t < 1/2 min { ( 1 − t ) p − t p + p t p − 1 } > 0. For p = 2 p=2 p = 2 , one has c 2 = 1 c_2=1 c 2 = 1 , and the inequality is an identity.
Proof. For s p ≠ 1 sp\ne1 s p = 1 , this is [11 Theorem 1.2] , based on the ground-state representation [10 Proposition 2.3 and Lemma 2.6] . The half-space ground state there is x d ( s p − 1 ) / p x_d^{(sp-1)/p} x d ( s p − 1 ) / p , and its Hardy constant agrees with (20) in our ordered-double-integral normalization. Every test-core function belongs to the energy space of that theorem. Its remainder is finite: near the diagonal, smoothness of u / ω s , p u/\omega_{s,p} u / ω s , p gives an integrable power because p − s p > 0 p-sp>0 p − s p > 0 ; away from the support, tangential integration reduces the boundary and infinity bounds to the integrable powers y d ( s p − 1 ) / 2 y_d^{(sp-1)/2} y d ( s p − 1 ) /2 and y d − ( s p + 3 ) / 2 y_d^{-(sp+3)/2} y d − ( s p + 3 ) /2 , respectively. For s p = 1 sp=1 s p = 1 , one has ω s , p = 1 \omega_{s,p}=1 ω s , p = 1 and D d , s , p = 0 \mathcal{D}_{d,s,p}=0 D d , s , p = 0 , so H d , s , p [ u ] = G d , s , p [ u ] \mathcal{H}_{d,s,p}[u]=\mathcal{G}_{d,s,p}[u] H d , s , p [ u ] = G d , s , p [ u ] . Since 0 < c p ≤ 1 0<c_p\le1 0 < c p ≤ 1 by [10 Lemma 2.6] , the asserted inequality also follows in this case. The cited representation is an identity for p = 2 p=2 p = 2 , with c 2 = 1 c_2=1 c 2 = 1 . □
Corollary C.2 (Subcritical nonlinear coercivity). Let d ≥ 2 d\ge2 d ≥ 2 , 0 < s < 1 0<s<1 0 < s < 1 , p ≥ 2 p\ge2 p ≥ 2 , τ > 0 \tau>0 τ > 0 , and 0 ≤ λ < Λ d , s , p , τ 0\le\lambda<\Lambda_{d,s,p,\tau} 0 ≤ λ < Λ d , s , p , τ . Then
H d , s , p [ u ] − λ R d , s , p , τ [ u ] ≥ c p ( 1 − λ Λ d , s , p , τ ) G d , s , p [ u ] \mathcal{H}_{d,s,p}[u]-\lambda\mathcal{R}_{d,s,p,\tau}[u]\ge c_p\left(1-\frac{\lambda}{\Lambda_{d,s,p,\tau}}\right)\mathcal{G}_{d,s,p}[u] H d , s , p [ u ] − λ R d , s , p , τ [ u ] ≥ c p ( 1 − Λ d , s , p , τ λ ) G d , s , p [ u ] for u ∈ C c ∞ ( R + d ) u\in C_c^\infty(\mathbb{R}_+^d) u ∈ C c ∞ ( R + d ) . Together with the sharp inequality, this gives
H d , s , p [ u ] − λ R d , s , p , τ [ u ] ≥ Λ d , s , p , τ − λ 2 { R d , s , p , τ [ u ] + c p Λ d , s , p , τ G d , s , p [ u ] } . \mathcal{H}_{d,s,p}[u]-\lambda\mathcal{R}_{d,s,p,\tau}[u]\ge\frac{\Lambda_{d,s,p,\tau}-\lambda}{2}\left\{\mathcal{R}_{d,s,p,\tau}[u]+\frac{c_p}{\Lambda_{d,s,p,\tau}}\mathcal{G}_{d,s,p}[u]\right\}. H d , s , p [ u ] − λ R d , s , p , τ [ u ] ≥ 2 Λ d , s , p , τ − λ { R d , s , p , τ [ u ] + Λ d , s , p , τ c p G d , s , p [ u ] } . The first bound is equivalently
H d , s , p [ u ] ≥ ( 1 − δ ) Λ d , s , p , τ R d , s , p , τ [ u ] + δ c p G d , s , p [ u ] , 0 < δ ≤ 1 , \mathcal{H}_{d,s,p}[u]\ge(1-\delta)\Lambda_{d,s,p,\tau}\mathcal{R}_{d,s,p,\tau}[u]+\delta c_p\mathcal{G}_{d,s,p}[u],\qquad0<\delta\le1, H d , s , p [ u ] ≥ ( 1 − δ ) Λ d , s , p , τ R d , s , p , τ [ u ] + δ c p G d , s , p [ u ] , 0 < δ ≤ 1 ,
with λ = ( 1 − δ ) Λ d , s , p , τ \lambda= (1-\delta)\Lambda_{d,s,p,\tau} λ = ( 1 − δ ) Λ d , s , p , τ . Taking λ = Λ d , s , p , τ − ε \lambda= \Lambda_{d,s,p,\tau}-\varepsilon λ = Λ d , s , p , τ − ε in the two-remainder bound gives the corresponding estimate for the regularized sharp deficit, for every 0 < ε ≤ Λ d , s , p , τ 0 < \varepsilon\le\Lambda_{d,s,p,\tau} 0 < ε ≤ Λ d , s , p , τ .
Proof. Write Λ = Λ d , s , p , τ \Lambda= \Lambda_{d,s,p,\tau} Λ = Λ d , s , p , τ and δ = 1 − λ / Λ ∈ ( 0 , 1 ] \delta= 1-\lambda/\Lambda\in(0,1] δ = 1 − λ /Λ ∈ ( 0 , 1 ] . The exact decomposition is
H d , s , p [ u ] − λ R d , s , p , τ [ u ] = ( 1 − δ ) ( H d , s , p [ u ] − Λ R d , s , p , τ [ u ] ) + δ H d , s , p [ u ] . \mathcal{H}_{d,s,p}[u]-\lambda\mathcal{R}_{d,s,p,\tau}[u] = (1-\delta)(\mathcal{H}_{d,s,p}[u]-\Lambda\mathcal{R}_{d,s,p,\tau}[u])+\delta\mathcal{H}_{d,s,p}[u]. H d , s , p [ u ] − λ R d , s , p , τ [ u ] = ( 1 − δ ) ( H d , s , p [ u ] − Λ R d , s , p , τ [ u ]) + δ H d , s , p [ u ] .
The first summand is nonnegative. Applying Lemma C.1 to the second gives δ c p G d , s , p [ u ] \delta c_p\mathcal{G}_{d,s,p}[u] δ c p G d , s , p [ u ] . Average this estimate with H d , s , p [ u ] − λ R d , s , p , τ [ u ] ≥ ( Λ d , s , p , τ − λ ) R d , s , p , τ [ u ] \mathcal{H}_{d,s,p}[u]-\lambda\mathcal{R}_{d,s,p,\tau}[u] \ge(\Lambda_{d,s,p,\tau}-\lambda)\mathcal{R}_{d,s,p,\tau}[u] H d , s , p [ u ] − λ R d , s , p , τ [ u ] ≥ ( Λ d , s , p , τ − λ ) R d , s , p , τ [ u ] to obtain the two-remainder estimate. □
Remark C.3 (Scope of the nonlinear estimate). For p > 2 p > 2 p > 2 , the coefficient c p ( 1 − λ / Λ d , s , p , τ ) c_p(1-\lambda/\Lambda_{d,s,p,\tau}) c p ( 1 − λ / Λ d , s , p , τ ) vanishes at the sharp level. The estimate therefore gives subcritical stability . Fischer’s two-sided representation [9 ] is proved for weighted graphs. To apply a corresponding ground-state argument here, one would need a positive ground state for the full Hardy–Maz’ya potential.
Data availability statement
No datasets were generated or analysed during the present study.
Conflict of interest
The authors declare that they have no conflict of interest.
References
Email address: shbdeng@swu.edu.cn