Introduction
Multivalued harmonic functions, 1-forms, and spinors arise naturally in geometric measure theory [4 ] , gauge theory [Tau13], calibrated geometry [10 ] , and special holonomy [6 ] . In this paper we study the regularity of the branching locus of a Z 2 \mathbb{Z}_2 Z 2 -harmonic spinor.
Let M M M be a Riemannian manifold of dimension n ≥ 2 n \ge2 n ≥ 2 , equipped with a real or complex spinor bundle E E E carrying a fibre metric h h h , Clifford multiplication c c c , and a compatible connection ∇ E \nabla^E ∇ E , and write D = ∑ i c ( e i ) ∇ e i E D = \sum_i c(e_i)\nabla^E_{e_i} D = ∑ i c ( e i ) ∇ e i E for the associated Dirac operator in some local orthonormal frame { e i } \{e_i\} { e i } . We call a symmetric two-valued section u u u a Z 2 \mathbb{Z}_2 Z 2 -harmonic spinor if u ∈ W l o c 1 , 2 ( M , A 2 s y m ( E ) ) u \in W^{1,2}_{\mathrm{loc}}(M,\mathcal{A}^{\mathrm{sym}}_2(E)) u ∈ W loc 1 , 2 ( M , A 2 sym ( E )) and
D u = 0 almost everywhere. D u = 0 \quad\text{almost everywhere.} D u = 0 almost everywhere.
Here A 2 s y m ( E x ) : = { ⟦ v ⟧ + ⟦ − v ⟧ : v ∈ E x } \mathcal{A}^{\mathrm{sym}}_2(E_x) := \{\llbracket v\rrbracket+ \llbracket-v\rrbracket: v \in E_x\} A 2 sym ( E x ) := { [ [ v ] ] + [ [ − v ] ] : v ∈ E x } denotes the space of symmetric unordered pairs in the fibre E x E_x E x for x ∈ M x \in M x ∈ M , and D u D u D u is well-defined almost everywhere for W 1 , 2 W^{1,2} W 1 , 2 two-valued sections [4 ] .
It turns out that u u u admits a Hölder continuous representative (see § 2.1.4 ). The branching set B u B_u B u and the zero locus Z u Z_u Z u are defined by
B u : = { x : u has no continuous local lift near x } ⊂ Z u : = { x : u ( x ) = 2 ⟦ 0 ⟧ } . B_u := \{x : u \text{ has no continuous local lift near } x\} \subset Z_u := \{x : u(x) = 2\llbracket0\rrbracket\}. B u := { x : u has no continuous local lift near x } ⊂ Z u := { x : u ( x ) = 2 [ [ 0 ] ] } .
The branching set B u B_u B u is closed.
Our main result concerns branch points of frequency 1 2 \frac{1}{2} 2 1 .
Theorem 1. Let u u u be a nonzero Z 2 \mathbb{Z}_2 Z 2 -harmonic spinor and let p 0 ∈ B u p_0 \in B_u p 0 ∈ B u have frequency N u ( p 0 , 0 ) = 1 2 N_u(p_0,0) = \frac{1}{2} N u ( p 0 , 0 ) = 2 1 . Then B u B_u B u is a smooth (resp. real-analytic) embedded submanifold of codimension two near p 0 p_0 p 0 for smooth (resp. real-analytic) ambient defining data ( M , E , h , c , ∇ E ) (M,E,h,c,\nabla^E) ( M , E , h , c , ∇ E ) .
(3) (Frequency-1 2 \frac{1}{2} 2 1 and geometric applications ) Certain gluing constructions in gauge theory and calibrated geometry [12 , 24 ] involve non-degenerate Z 2 \mathbb{Z}_2 Z 2 -harmonic spinors with a smooth branching locus and ∣ u ( x ) ∣ ∼ dist ( x , B u ) 1 / 2 |u(x)| \sim\operatorname{dist}(x,\mathcal{B}_u)^{1/2} ∣ u ( x ) ∣ ∼ dist ( x , B u ) 1/2 near the branching locus. To address the loss of derivatives in the deformation theory, one uses Nash–Moser methods, which require sufficiently high regularity of B u \mathcal{B}_u B u [7 , 23 , 12 ] . Frequency measures the local vanishing order, so the frequency-1 2 \frac{1}{2} 2 1 assumption gives a weak formulation of non-degeneracy. Theorem 1 then supplies the regularity of the branching locus needed for this deformation theory.
(4) (Regularity of singular sets ) The branching set is itself part of the unknown, and no equation is prescribed for it a priori . Its regularity is therefore a codimension-two free boundary problem. Related regularity questions arise in the classical obstacle problem [1 , 8 , 9 ] and the Signorini (thin obstacle) problem [15 ] . They also arise in the study of cylindrical tangent cones in minimal submanifolds [Sim93], multivalued harmonic functions [17 , 18 , 16 ] , singular sets of area-minimizing integral currents [3 , 19 ] , and cylindrical singular sets in mean curvature flow [SWX25].
Proof strategy
The proof has two parts. We first show that the branching locus is a C 1 , α C^{1,\alpha} C 1 , α graph and obtain quantitative estimates for approximation by its tangent models. We then use certain components of the spinor as normal coordinates and prove higher regularity in the tangential directions.
Part I: C 1 , α C^{1,\alpha} C 1 , α -regularity of the branching locus and quantitative estimates. We first show that a Z 2 \mathbb{Z}_2 Z 2 -harmonic spinor locally minimises a modified Dirichlet energy J \mathcal{J} J . The key input is an integrated Weitzenböck formula, obtained from a transgression form on the total space E E E . This observation was motivated by work on branched deformations of special Lagrangians [10 ] and by the compactness problem for Fueter sections with values in moduli spaces of monopoles of charge ≥ 2 \ge2 ≥ 2 [11 , 20 ] .
Local minimality yields stationarity and the variational estimates needed to establish almost monotonicity of frequency, Hölder continuity, and strong W 1 , 2 W^{1,2} W 1 , 2 -compactness for blow-ups [4 Stu19] . With these inputs, we adapt the excess-decay method of Leon Simon [Sim93], as developed for two-valued harmonic functions by Krummel–Wickramasekera [17 ] .
The main output of Part I is the following.
Theorem 2 (C 1 , α C^{1,\alpha} C 1 , α -regularity with quantitative estimates). For smooth ambient data, there is a ball B ( p 0 , r 0 ) B(p_0,r_0) B ( p 0 , r 0 ) such that the set B u ∩ B ( p 0 , r 0 ) \mathcal{B}_u\cap B(p_0,r_0) B u ∩ B ( p 0 , r 0 ) is a single C 1 , α C^{1,\alpha} C 1 , α -graph of codimension two over the spine at p 0 p_0 p 0 , for some α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) . For each p ∈ B u ∩ B ( p 0 , r 0 ) p\in\mathcal{B}_u\cap B(p_0,r_0) p ∈ B u ∩ B ( p 0 , r 0 ) , there is a unique non-degenerate flat Dirac tangent model Φ p \Phi_p Φ p with spine T p B u T_p\mathcal{B}_u T p B u . Moreover, the following quantitative estimates hold:
sup B r / 2 ( p ) G ( u , Φ p ) ≤ C r 1 / 2 + α , r 1 − n ∫ B r / 2 ( p ) G ( ∇ u , ∇ Φ p ) 2 ≤ C r 2 α , 0 < r ≤ r 0 , (1) \sup_{B_{r/2}(p)}\mathcal{G}(u,\Phi_p)\le Cr^{1/2+\alpha},\qquad r^{1-n}\int_{B_{r/2}(p)}\mathcal{G}(\nabla u,\nabla\Phi_p)^2\le Cr^{2\alpha},\qquad0<r\le r_0, \tag*{(1)} B r /2 ( p ) sup G ( u , Φ p ) ≤ C r 1/2 + α , r 1 − n ∫ B r /2 ( p ) G ( ∇ u , ∇ Φ p ) 2 ≤ C r 2 α , 0 < r ≤ r 0 , (1) where G \mathcal{G} G denotes the distance between unordered two-valued sections.
Part II: The partial Legendre transform and Campanato iterations. Choose local coordinates ( z , y ) (z,y) ( z , y ) in which the spine at p 0 p_0 p 0 is { 0 } × R n − 2 \{0\}\times\mathbb{R}^{n-2} { 0 } × R n − 2 . By Part I, the branching locus is the graph { ( g ( y ) , y ) } \{(g(y),y)\} {( g ( y ) , y )} of a C 1 , α C^{1,\alpha} C 1 , α function, and the tangent model at p 0 p_0 p 0 has the form Φ p 0 = a 0 z 1 / 2 + b 0 z ˉ 1 / 2 \Phi_{p_0}=a_0z^{1/2}+b_0\bar z^{1/2} Φ p 0 = a 0 z 1/2 + b 0 z ˉ 1/2 , with ∣ a 0 ∣ 2 + ∣ b 0 ∣ 2 > 0 |a_0|^2+|b_0|^2>0 ∣ a 0 ∣ 2 + ∣ b 0 ∣ 2 > 0 . The quantitative approximation by these models allows us to use two real components of the spinor as new normal coordinates ξ \xi ξ . On the branched double cover, the resulting partial Legendre transform is described by
X ( ξ , y ) = ( φ ( ξ , y ) , y ) , u ^ ( ξ , y ) = ( ξ , V ( ξ , y ) ) , U = ( φ , V ) , X(\xi,y)=(\varphi(\xi,y),y),\qquad\widehat{u}(\xi,y)=(\xi,V(\xi,y)),\qquad U=(\varphi,V), X ( ξ , y ) = ( φ ( ξ , y ) , y ) , u ( ξ , y ) = ( ξ , V ( ξ , y )) , U = ( φ , V ) ,
where φ \varphi φ is the inverse normal coordinate map and V V V contains the remaining spinor components. The map φ \varphi φ is even in ξ \xi ξ , while V V V is odd. At p 0 = p ( y 0 ) = ( g ( y 0 ) , y 0 ) p_0=p(y_0)=(g(y_0),y_0) p 0 = p ( y 0 ) = ( g ( y 0 ) , y 0 ) , the transformed tangent model has the form
P = ( a + B ( y − y 0 ) + q ( ξ ) , M ξ ) , (2) P=(a+B(y-y_0)+q(\xi),M\xi), \tag*{(2)} P = ( a + B ( y − y 0 ) + q ( ξ ) , M ξ ) , (2)
where a ∈ C a \in\mathbb{C} a ∈ C , B B B and M M M are linear maps acting on y − y 0 y-y_0 y − y 0 and ξ \xi ξ , respectively, and q q q is a homogeneous quadratic polynomial. The simplest example is the square model U 0 = ( ξ 2 , 0 ) U_0=(\xi^2,0) U 0 = ( ξ 2 , 0 ) .
In these coordinates, the branching set is the fixed plane { ξ = 0 } \{\xi=0\} { ξ = 0 } , and its geometry is encoded in the equations for U U U . These form a quasilinear degenerate elliptic system with the inherited variational structure. The identity g ( y ) = φ ( 0 , y ) g(y)=\varphi(0,y) g ( y ) = φ ( 0 , y ) reduces regularity of the original branching locus to tangential regularity of φ \varphi φ . We write D y β = D y 1 β 1 ⋯ D y n − 2 β n − 2 D_y^\beta=D_{y_1}^{\beta_1}\cdots D_{y_{n-2}}^{\beta_{n-2}} D y β = D y 1 β 1 ⋯ D y n − 2 β n − 2 for a tangential derivative of order ∣ β ∣ = ∑ k = 1 n − 2 β k |\beta|=\sum_{k=1}^{n-2}\beta_k ∣ β ∣ = ∑ k = 1 n − 2 β k .
The main result of Part II is the following.
Theorem 3. For smooth (resp. real-analytic) ambient data, on every sufficiently small cylinder Q r ( y 0 ) = B r 2 ( 0 ) × B r n − 2 ( y 0 ) Q_r(y_0)=B_r^2(0)\times B_r^{n-2}(y_0) Q r ( y 0 ) = B r 2 ( 0 ) × B r n − 2 ( y 0 ) centred at a frequency-1 2 \frac{1}{2} 2 1 branch point ( 0 , y 0 ) (0,y_0) ( 0 , y 0 ) in the transformed coordinates, the fields D y β U D_y^\beta U D y β U have a unique tangent polynomial of the form (2) and satisfy
∥ D y β U ∥ L ∞ ( Q r ( y 0 ) ) ≤ C β , (resp. ∥ D y β U ∥ L ∞ ( Q r ( y 0 ) ) ≤ C ∣ β ∣ + 1 ∣ β ∣ ! ), (3) \|D_y^\beta U\|_{L^\infty(Q_r(y_0))}\le C_\beta,\qquad\text{(resp. }\|D_y^\beta U\|_{L^\infty(Q_r(y_0))}\le C^{|\beta|+1}|\beta|!\text{),} \tag*{(3)} ∥ D y β U ∥ L ∞ ( Q r ( y 0 )) ≤ C β , (resp. ∥ D y β U ∥ L ∞ ( Q r ( y 0 )) ≤ C ∣ β ∣ + 1 ∣ β ∣ ! ), (3) where C C C is independent of the multi-index β \beta β in the real-analytic case. In particular, the branching set near p 0 = p ( y 0 ) p_0=p(y_0) p 0 = p ( y 0 ) is smooth (resp. real-analytic). In Legendre coordinates, the Z 2 \mathbb{Z}_2 Z 2 -harmonic spinor and the model coefficients a = g ( y ) a=g(y) a = g ( y ) , B = D y g B=D_y g B = D y g , q q q , and M M M depend smoothly (resp. real-analytically) on y y y .
The proof proceeds in three steps.
We first study the square model U 0 = ( ξ 2 , 0 ) U_0=(\xi^2,0) U 0 = ( ξ 2 , 0 ) in flat ambient space. Its linearised Euler–Lagrange equation B 0 ( Z , Y ) = 0 \mathcal{B}_0(Z,Y)=0 B 0 ( Z , Y ) = 0 is the weak formulation of the system
{ Δ ξ Z φ − 4 ξ ∂ ξ Z φ + 4 ∣ ξ ∣ 2 Δ y Z φ = 0 , Δ ξ Z V + 4 ∣ ξ ∣ 2 Δ y Z V = 0 , (4) \begin{cases}
\Delta_\xi Z_\varphi-\dfrac{4}{\xi}\partial_\xi Z_\varphi+4|\xi|^2\Delta_y Z_\varphi=0,\\
\Delta_\xi Z_V+4|\xi|^2\Delta_y Z_V=0, \tag*{(4)}
\end{cases} ⎩ ⎨ ⎧ Δ ξ Z φ − ξ 4 ∂ ξ Z φ + 4∣ ξ ∣ 2 Δ y Z φ = 0 , Δ ξ Z V + 4∣ ξ ∣ 2 Δ y Z V = 0 , (4)
where Z φ Z_\varphi Z φ is even and Z V Z_V Z V is odd in ξ \xi ξ . In weighted Sobolev and Hölder spaces adapted to Grushin Laplacian Δ ξ + 4 ∣ ξ ∣ 2 Δ y \Delta_\xi+4|\xi|^2\Delta_y Δ ξ + 4∣ ξ ∣ 2 Δ y , we prove a weighted Caccioppoli inequality and excess decay for weak solutions of (4) .
The Grushin Laplacian commutes with tangential derivatives. Differentiating the nonlinear system in the y y y -directions gives an equation of the schematic form
B U ( D y β U , Y ) = F β ( Y ) , for every admissible test field Y , \mathcal{B}_U(D_y^\beta U,Y)=\mathcal{F}_\beta(Y),\qquad\text{for every admissible test field }Y, B U ( D y β U , Y ) = F β ( Y ) , for every admissible test field Y ,
where B U \mathcal{B}_U B U is the second variation of the transformed J \mathcal{J} J -energy and F β \mathcal{F}_\beta F β contains only field derivatives with lower tangential order. The estimates from Part I control the difference between B U \mathcal{B}_U B U and the form obtained by freezing at a transformed tangent model. A Campanato iteration then transfers the frozen excess decay to D y β U D_y^\beta U D y β U .
To continue the induction, we need derivative estimates for the fields appearing in the source F β \mathcal{F}_\beta F β . We therefore upgrade the L 2 L^2 L 2 -height estimate to a weighted C 1 , α C^{1,\alpha} C 1 , α -estimate. In particular, we obtain
∣ D ξ D y β φ ( ξ , y ) ∣ ≲ β ∣ ξ ∣ , |D_\xi D_y^\beta\varphi(\xi,y)|\lesssim_\beta|\xi|, ∣ D ξ D y β φ ( ξ , y ) ∣ ≲ β ∣ ξ ∣ ,
in accordance with the quadratic normal term and the evenness of φ \varphi φ . These estimates close the induction for smooth ambient data. For real-analytic data, we track their dependence on the derivative order to obtain the factorial bounds in (3) , and hence real analyticity of the branching locus.
Remark 1.2. Krummel [16 ] studied real analyticity of the branch set for two-valued Dirichlet minimisers and C 1 , α C^{1,\alpha} C 1 , α two-valued harmonic functions under the assumption that the frequency is locally constant along the branch set. His approach uses quantitative asymptotics at branch points to construct a partial Legendre transform, following Kinderlehrer–Nirenberg–Spruck [14 ] . The intended route to analyticity is then through Schauder estimates for the resulting singular elliptic system and a majorant argument. Partial Legendre transforms have also been used for the Signorini problem [15 ] . Our treatment of the transformed problem instead uses excess decay and Campanato iteration. This approach is of independent interest.
Remark 1.3 (Frequency jumping set). We expect that B u \mathcal{B}_u B u is smooth (resp. real-analytic) near every branch point at which N u ( ⋅ , 0 ) N_u(\cdot,0) N u ( ⋅ , 0 ) is locally constant on B u \mathcal{B}_u B u , with value 1 2 + k \frac{1}{2}+k 2 1 + k for some k ∈ Z ≥ 0 k \in\mathbb{Z}_{\geq0} k ∈ Z ≥ 0 (cf. [16 ] ). How large is the set of branch points near which N u ( ⋅ , 0 ) N_u(\cdot,0) N u ( ⋅ , 0 ) is not constant on B u \mathcal{B}_u B u ? For real-analytic ambient data, does it have vanishing H n − 2 \mathcal{H}^{n-2} H n − 2 -measure, or even Hausdorff dimension at most n − 3 n-3 n − 3 ? By contrast, for smooth data, it may have positive H n − 2 \mathcal{H}^{n-2} H n − 2 -measure (cf. the fat Cantor set example by extending the construction of [2 ] ).
Organisation
Part I begins in Section 2 with J \mathcal{J} J -minimality, Hölder regularity, strong blow-up compactness and the frequency-1 2 \frac{1}{2} 2 1 models. § 2.2 establishes sheet matching, the no-hole property and the non-concentration estimates. Section 4 establishes linear excess decay and carries out the Campanato iteration, completing the proof of Theorem 2 . In Part II, § 3.1 introduces the partial Legendre transform and the weighted spaces. § 3.4 and § 3.5 prove the frozen and variable-coefficient linear estimates, corresponding to the first two steps above. § 3.6 carries out the tangential regularity induction and proves real analyticity, completing the proof of Theorem 1 .
Acknowledgements.
Part I develops and refines material from the second author’s Ph.D. thesis [22 ] . Y. L. is supported by the Royal Society URF. D. P. was partially supported by the Cecilia Tanner Research Impulse Fund from Imperial College London. Y. Z. is supported by the London School of Geometry and Number Theory Centre for Doctoral Training (CDT), a joint venture between University College London, Imperial College London and King’s College London.
We thank Simon Donaldson, Siqi He, Gregory Parker, Luca Spolaor, Thomas Walpuski, and Neshan Wickramasekera for their interest in this work and for useful discussions. Parts of this work were discussed during the 2026 Clay workshop Multi-valued Harmonicity at Oxford. We are also grateful to the organisers and all participants for an inspiring meeting.
AI disclosure.
The main ideas of this paper were developed by the authors. Large language models (LLMs) were used to check routine calculations and assist with manuscript preparation. The authors take full responsibility for the content and correctness of the paper.
Part 1. C1 , α ^{1,\alpha} 1 , α -regularity of the branching locus
This part establishes C 1 , α C^{1,\alpha} C 1 , α -regularity of the branching locus together with the quantitative estimates, which are later used in Part II (cf. Proposition 5.4 ). The principal new ingredient is the integrated Weitzenböck formula in § 2.1.2 , proved without using the Laplacian. It leads to the local minimising property in § 2.1.3 . Hölder regularity is then essentially a consequence of Stuvard’s work [Stu19]; we briefly discuss the argument for the benefit of the reader. The remainder of Part I is adapted from the second author’s thesis [22 ] , which follows the strategy of Simon [Sim93] and Krummel–Wickramasekera [17 ] , and uses the Dirac equation to simplify several steps.
J-energy minimality, Holder regularity and frequency-one-half models
Geometric setting
Recall that E → M E \to M E → M is a real or complex spinor bundle over a Riemannian manifold ( M , g ) (M,g) ( M , g ) , equipped with a fibre metric h h h , Clifford multiplication c c c , and a compatible connection ∇ E \nabla^E ∇ E , which we denote by ∇ \nabla ∇ below. The associated Dirac operator is D = ∑ i c ( e i ) ∇ e i \mathcal{D}=\sum_i c(e_i)\nabla_{e_i} D = ∑ i c ( e i ) ∇ e i in a local orthonormal frame { e i } \{e_i\} { e i } . We use the Clifford conventions
c ( v ) c ( w ) + c ( w ) c ( v ) = − 2 g ( v , w ) , c ( v ) ∗ = − c ( v ) . c(v)c(w)+c(w)c(v)=-2g(v,w), \qquad c(v)^*=-c(v). c ( v ) c ( w ) + c ( w ) c ( v ) = − 2 g ( v , w ) , c ( v ) ∗ = − c ( v ) .
The Weitzenböck formula on C 2 C^2 C 2 sections of E E E is
D ∗ D = ∇ ∗ ∇ + R , R = ∑ i < j c ( e i ) c ( e j ) R e i , e j E . \mathcal{D}^*\mathcal{D}=\nabla^*\nabla+\mathcal{R}, \qquad\mathcal{R}=\sum_{i<j}c(e_i)c(e_j)R^E_{e_i,e_j}. D ∗ D = ∇ ∗ ∇ + R , R = i < j ∑ c ( e i ) c ( e j ) R e i , e j E .
where R E R^{E} R E is the curvature of ∇ \nabla ∇ . We write ⟨ v , w ⟩ = Re h ( v , w ) \langle v,w\rangle=\operatorname{Re}h(v,w) ⟨ v , w ⟩ = Re h ( v , w ) . For a two-valued W 1 , 2 W^{1,2} W 1 , 2 -section W = [ [ w 1 ] ] + [ [ w 2 ] ] W=[\![w_{1}]\!]+[\![w_{2}]\!] W = [ [ w 1 ] ] + [ [ w 2 ] ] , set
∣ W ∣ 2 = 1 2 ∑ ℓ = 1 2 ∣ w ℓ ∣ 2 , ∣ ∇ W ∣ 2 = 1 2 ∑ ℓ = 1 2 ∣ ∇ w ℓ ∣ 2 , ∣ D W ∣ 2 = 1 2 ∑ ℓ = 1 2 ∣ D w ℓ ∣ 2 . |W|^{2}=\frac{1}{2}\sum_{\ell=1}^{2}|w_{\ell}|^{2},\qquad|\nabla W|^{2}=\frac{1}{2}\sum_{\ell=1}^{2}|\nabla w_{\ell}|^{2},\qquad|DW|^{2}=\frac{1}{2}\sum_{\ell=1}^{2}|Dw_{\ell}|^{2}. ∣ W ∣ 2 = 2 1 ℓ = 1 ∑ 2 ∣ w ℓ ∣ 2 , ∣∇ W ∣ 2 = 2 1 ℓ = 1 ∑ 2 ∣∇ w ℓ ∣ 2 , ∣ D W ∣ 2 = 2 1 ℓ = 1 ∑ 2 ∣ D w ℓ ∣ 2 .
Here the covariant gradient is defined almost everywhere using the approximate derivative of W W W in local trivialisations. We measure the pointwise distance between two sections of A 2 ( E ) A_{2}(E) A 2 ( E ) using the Almgren metric [4 ] , normalised by
G ( ∑ ℓ = 1 2 [ [ v ℓ ] ] , ∑ ℓ = 1 2 [ [ w ℓ ] ] ) 2 = 1 2 min π ∈ S 2 ∑ ℓ = 1 2 ∣ v ℓ − w π ( ℓ ) ∣ 2 . \mathcal{G}\left(\sum_{\ell=1}^{2}[\![v_{\ell}]\!],\sum_{\ell=1}^{2}[\![w_{\ell}]\!]\right)^{2}=\frac{1}{2}\min_{\pi\in S_{2}}\sum_{\ell=1}^{2}|v_{\ell}-w_{\pi(\ell)}|^{2}. G ( ℓ = 1 ∑ 2 [ [ v ℓ ] ] , ℓ = 1 ∑ 2 [ [ w ℓ ] ] ) 2 = 2 1 π ∈ S 2 min ℓ = 1 ∑ 2 ∣ v ℓ − w π ( ℓ ) ∣ 2 .
so ∣ W ∣ 2 = G ( W , 2 [ [ 0 ] ] ) 2 |W|^{2}=\mathcal{G}(W,2[\![0]\!])^{2} ∣ W ∣ 2 = G ( W , 2 [ [ 0 ] ] ) 2 . The same convention applies to covariant derivatives in T ∗ M ⊗ E T^{*}M\otimes E T ∗ M ⊗ E and higher covariant tensors, with one permutation matching all components of each sheet.
On a geodesic ball B r ( p ) ⋐ M B_{r}(p)\Subset M B r ( p ) ⋐ M , define the energy of W ∈ W 1 , 2 ( B r ( p ) ; A 2 ( E ) ) W\in W^{1,2}(B_{r}(p);A_{2}(E)) W ∈ W 1 , 2 ( B r ( p ) ; A 2 ( E )) by
J B r ( p ) ( W ) = 1 2 ∑ ℓ = 1 2 ∫ B r ( p ) ( ∣ ∇ w ℓ ∣ 2 + ⟨ R w ℓ , w ℓ ⟩ ) d V g . \mathcal{J}_{B_{r}(p)}(W)=\frac{1}{2}\sum_{\ell=1}^{2}\int_{B_{r}(p)}\left(|\nabla w_{\ell}|^{2}+\langle\mathcal{R}w_{\ell},w_{\ell}\rangle\right)\,dV_{g}. J B r ( p ) ( W ) = 2 1 ℓ = 1 ∑ 2 ∫ B r ( p ) ( ∣∇ w ℓ ∣ 2 + ⟨ R w ℓ , w ℓ ⟩ ) d V g .
This is independent of the ordering of the sheets. For p ∈ Z u p\in Z_{u} p ∈ Z u , set
H u ( p , r ) = ∫ ∂ B r ( p ) ∣ u ∣ 2 d Area g , N u ( p , r ) = r J B r ( p ) ( u ) H u ( p , r ) , N u ( p , 0 ) = lim r ↓ 0 N u ( p , r ) . H_{u}(p,r)=\int_{\partial B_{r}(p)}|u|^{2}\,d\operatorname{Area}_{g},\qquad N_{u}(p,r)=\frac{r\mathcal{J}_{B_{r}(p)}(u)}{H_{u}(p,r)},\qquad N_{u}(p,0)=\lim_{r\downarrow0}N_{u}(p,r). H u ( p , r ) = ∫ ∂ B r ( p ) ∣ u ∣ 2 d Area g , N u ( p , r ) = H u ( p , r ) r J B r ( p ) ( u ) , N u ( p , 0 ) = r ↓ 0 lim N u ( p , r ) .
The existence of this limit is proved in § 2.1.5 .
The tangent models are described in Subsection 2.7. We compare them in geodesic coordinates and the exponential gauge centred at p p p : an orthonormal basis of E p E_{p} E p is extended by parallel transport along radial geodesics. This identifies E x E_{x} E x with E p E_{p} E p . We write X = exp p − 1 x X=\exp_{p}^{-1}x X = exp p − 1 x and suppress the parallel transport maps below.
The integrated Weitzenböck formula
Although our applications concern two-valued sections, the following integrated Weitzenböck formula holds for arbitrary Q Q Q -valued W 1 , 2 W^{1,2} W 1 , 2 -sections. The conventions in the previous Section 2.1 extend to A Q ( E ) A_{Q}(E) A Q ( E ) , with quantities averaged over the Q Q Q sheets.
Let π : E → M \pi:E\to M π : E → M be the bundle projection. Then the pullback bundle π ∗ E → E \pi^{*}E\to E π ∗ E → E admits a tautological section τ ( e ) = e \tau(e)=e τ ( e ) = e ; we write
ϑ = ∇ π ∗ E τ ∈ Ω 1 ( E ; π ∗ E ) . \vartheta=\nabla^{\pi^{*}E}\tau\in\Omega^{1}(E;\pi^{*}E). ϑ = ∇ π ∗ E τ ∈ Ω 1 ( E ; π ∗ E ) .
The form ϑ \vartheta ϑ vanishes on horizontal vectors and identifies each vertical tangent space T e ( E π ( e ) ) T_{e}(E_{\pi(e)}) T e ( E π ( e ) ) with E π ( e ) E_{\pi(e)} E π ( e ) . For a smooth section v : Ω → E v:\Omega\to E v : Ω → E over an open set Ω ⊂ M \Omega\subset M Ω ⊂ M , the natural identification v ∗ ( π ∗ E ) = E ∣ Ω v^{*}(\pi^{*}E)=E|_{\Omega} v ∗ ( π ∗ E ) = E ∣ Ω implies
v ∗ ϑ = ∇ v ∈ Ω 1 ( Ω ; E ) . v^{*}\vartheta=\nabla v\in\Omega^{1}(\Omega;E). v ∗ ϑ = ∇ v ∈ Ω 1 ( Ω ; E ) .
In a local oriented orthonormal frame { e i } i = 1 n \{e_{i}\}_{i=1}^{n} { e i } i = 1 n of T M TM T M , set
Θ = ∑ i < j ⟨ τ , c ( e i ) c ( e j ) ϑ ⟩ ∧ π ∗ ( ι e j ι e i d V g ) ∈ Ω n − 1 ( E ; R ) . \Theta=\sum_{i<j}\langle\tau,c(e_{i})c(e_{j})\vartheta\rangle\wedge\pi^{*}(\iota_{e_{j}}\iota_{e_{i}}dV_{g})\in\Omega^{n-1}(E;\mathbb{R}). Θ = i < j ∑ ⟨ τ , c ( e i ) c ( e j ) ϑ ⟩ ∧ π ∗ ( ι e j ι e i d V g ) ∈ Ω n − 1 ( E ; R ) .
Here the Clifford endomorphisms act on the pullback bundle. Thus Θ \Theta Θ is a real differential form on the total space E E E , with horizontal degree n − 2 n-2 n − 2 and vertical degree one relative to the connection.
Proposition 2.1 (Integrated Weitzenböck formula). For every W = ∑ ℓ = 1 Q [ [ w ℓ ] ] ∈ W 1 , 2 ( Ω ; A Q ( E ) ) W=\sum_{\ell=1}^{Q}[\![w_{\ell}]\!]\in W^{1,2}(\Omega;A_{Q}(E)) W = ∑ ℓ = 1 Q [ [ w ℓ ] ] ∈ W 1 , 2 ( Ω ; A Q ( E )) and every χ ∈ C ∞ ( Ω ) \chi\in C^{\infty}(\Omega) χ ∈ C ∞ ( Ω ) ,
∑ ℓ = 1 Q ∫ Ω χ ( ∣ ∇ w ℓ ∣ 2 + ⟨ R w ℓ , w ℓ ⟩ − ∣ D w ℓ ∣ 2 ) d V g = ∑ ℓ = 1 Q ∫ Ω ( ⟨ D w ℓ , c ( d χ ) w ℓ ⟩ − ⟨ ∇ ∇ χ w ℓ , w ℓ ⟩ ) d V g . (5) \sum_{\ell=1}^{Q}\int_{\Omega}\chi\left(|\nabla w_{\ell}|^{2}+\langle\mathcal{R}w_{\ell},w_{\ell}\rangle-|Dw_{\ell}|^{2}\right)\,dV_{g}
=
\sum_{\ell=1}^{Q}\int_{\Omega}\left(\langle Dw_{\ell},c(d\chi)w_{\ell}\rangle-\langle\nabla\nabla\chi w_{\ell},w_{\ell}\rangle\right)\,dV_{g}. \tag*{(5)} ℓ = 1 ∑ Q ∫ Ω χ ( ∣∇ w ℓ ∣ 2 + ⟨ R w ℓ , w ℓ ⟩ − ∣ D w ℓ ∣ 2 ) d V g = ℓ = 1 ∑ Q ∫ Ω ( ⟨ D w ℓ , c ( d χ ) w ℓ ⟩ − ⟨ ∇∇ χ w ℓ , w ℓ ⟩ ) d V g . (5) :::`
Proof. We establish (5) successively for smooth, Lipschitz, and W 1 , 2 W^{1,2} W 1 , 2 -sections. Note that the definition of Θ \Theta Θ is independent of the orthonormal frame. For a smooth section v v v , the Clifford relation implies the pointwise identity
∣ D v ∣ 2 = ∣ ∇ v ∣ 2 − 2 ∑ i < j ⟨ ∇ e i v , c ( e i ) c ( e j ) ∇ e j v ⟩ . \lvert Dv\rvert^2=\lvert\nabla v\rvert^2-2\sum_{i<j}\langle\nabla_{e_i}v,c(e_i)c(e_j)\nabla_{e_j}v\rangle. ∣ D v ∣ 2 = ∣ ∇ v ∣ 2 − 2 i < j ∑ ⟨ ∇ e i v , c ( e i ) c ( e j ) ∇ e j v ⟩ .
Computing in local normal coordinates,
d ⟨ v , c ( e i ) c ( e j ) ∇ v ⟩ ( e i , e j ) = 2 ⟨ ∇ e i v , c ( e i ) c ( e j ) ∇ e j v ⟩ + ⟨ v , c ( e i ) c ( e j ) R e i , e j E v ⟩ . d\langle v,c(e_i)c(e_j)\nabla v\rangle(e_i,e_j)=2\langle\nabla_{e_i}v,c(e_i)c(e_j)\nabla_{e_j}v\rangle+\langle v,c(e_i)c(e_j)R^E_{e_i,e_j}v\rangle. d ⟨ v , c ( e i ) c ( e j ) ∇ v ⟩ ( e i , e j ) = 2 ⟨ ∇ e i v , c ( e i ) c ( e j ) ∇ e j v ⟩ + ⟨ v , c ( e i ) c ( e j ) R e i , e j E v ⟩ .
Summing over i < j i<j i < j yields the smooth transgression formula
v ∗ d Θ = ( ∣ ∇ v ∣ 2 + ⟨ R v , v ⟩ − ∣ D v ∣ 2 ) d V g . v^*d\Theta=(\lvert\nabla v\rvert^2+\langle\mathcal{R}v,v\rangle-\lvert Dv\rvert^2)dV_g. v ∗ d Θ = (∣ ∇ v ∣ 2 + ⟨ R v , v ⟩ − ∣ D v ∣ 2 ) d V g .
Thus v ∗ d Θ v^*d\Theta v ∗ d Θ is the divergence term in the Weitzenböck formula.
Now, let W W W be a Lipschitz Q Q Q -valued section, and let Γ W \Gamma_W Γ W be the current associated with its graph in E ∣ Ω E|_\Omega E ∣ Ω , counting each sheet with multiplicity one. The boundary theorem [5 Theorem 2.1] gives ∂ Γ W = 0 \partial\Gamma_W=0 ∂ Γ W = 0 in E ∣ Ω E|_\Omega E ∣ Ω . Since W W W is bounded on spt χ \operatorname{spt}\chi spt χ , the support of π ∗ χ Θ \pi^*\chi\,\Theta π ∗ χ Θ meets the graph in a compact set. Stokes’ theorem therefore yields
0 = ∂ Γ W ( π ∗ χ Θ ) = Γ W ( d ( π ∗ χ Θ ) ) = ∑ ℓ = 1 Q ∫ Ω ( χ w ℓ ∗ d Θ + d χ ∧ w ℓ ∗ Θ ) . 0=\partial\Gamma_W(\pi^*\chi\,\Theta)=\Gamma_W(d(\pi^*\chi\,\Theta))=\sum_{\ell=1}^{Q}\int_\Omega(\chi\,w_\ell^*d\Theta+d\chi\wedge w_\ell^*\Theta). 0 = ∂ Γ W ( π ∗ χ Θ ) = Γ W ( d ( π ∗ χ Θ )) = ℓ = 1 ∑ Q ∫ Ω ( χ w ℓ ∗ d Θ + d χ ∧ w ℓ ∗ Θ ) .
The pullbacks w ℓ ∗ Θ w_\ell^*\Theta w ℓ ∗ Θ and w ℓ ∗ d Θ w_\ell^*d\Theta w ℓ ∗ d Θ are interpreted using the approximate differentials of W W W and do not require globally defined Lipschitz sheets. It follows from the Clifford relation that
− d χ ∧ w ℓ ∗ Θ = ( ⟨ D w ℓ , c ( d χ ) w ℓ ⟩ − ⟨ ∇ ∇ χ w ℓ , w ℓ ⟩ ) d V g , -d\chi\wedge w_\ell^*\Theta=(\langle Dw_\ell,c(d\chi)w_\ell\rangle-\langle\nabla_{\nabla\chi}w_\ell,w_\ell\rangle)dV_g, − d χ ∧ w ℓ ∗ Θ = (⟨ D w ℓ , c ( d χ ) w ℓ ⟩ − ⟨ ∇ ∇ χ w ℓ , w ℓ ⟩) d V g ,
for almost every point in Ω \Omega Ω . Dividing the sum over the sheets by Q Q Q proves (5) for Lipschitz W W W .
Finally, for a general W 1 , 2 W^{1,2} W 1 , 2 -section, choose Lipschitz approximations W k W_k W k converging strongly to W W W in W 1 , 2 W^{1,2} W 1 , 2 , as in [4 Proposition 4.4] . The approximations may be chosen with λ k → ∞ \lambda_k\to\infty λ k → ∞ , Lip W k ≤ C λ k \operatorname{Lip}W_k\le C\lambda_k Lip W k ≤ C λ k , and
A k = { W k ≠ W } , ∣ A k ∣ = o ( λ k − 2 ) , ∥ G ( W k , W ) ∥ L 2 ( Ω ) → 0. A_k=\{W_k\ne W\},\qquad\lvert A_k\rvert=o(\lambda_k^{-2}),\qquad\lVert\mathcal{G}(W_k,W)\rVert_{L^2(\Omega)}\to0. A k = { W k = W } , ∣ A k ∣ = o ( λ k − 2 ) , ∥ G ( W k , W ) ∥ L 2 ( Ω ) → 0.
Absolute continuity of the integral and the Lipschitz bound imply
ε k : = ∫ A k ( ∣ ∇ W k ∣ 2 + ∣ ∇ W ∣ 2 + ∣ W k ∣ 2 + ∣ W ∣ 2 ) d V g → 0. \varepsilon_k:=\int_{A_k}(\lvert\nabla W_k\rvert^2+\lvert\nabla W\rvert^2+\lvert W_k\rvert^2+\lvert W\rvert^2)dV_g\to0. ε k := ∫ A k (∣ ∇ W k ∣ 2 + ∣ ∇ W ∣ 2 + ∣ W k ∣ 2 + ∣ W ∣ 2 ) d V g → 0.
Outside A k A_k A k , the sections and their covariant derivatives agree almost everywhere. With the L 1 L^1 L 1 -norms taken over spt χ \operatorname{spt}\chi spt χ , Clifford contraction and the boundedness of the curvature give
∥ ∣ ∇ W k ∣ 2 − ∣ ∇ W ∣ 2 ∥ L 1 + ∥ ∣ D W k ∣ 2 − ∣ D W ∣ 2 ∥ L 1 + ∥ ⟨ R W k , W k ⟩ − ⟨ R W , W ⟩ ∥ L 1 ≲ ε k → 0. \big\lVert\lvert\nabla W_k\rvert^2-\lvert\nabla W\rvert^2\big\rVert_{L^1}+\big\lVert\lvert DW_k\rvert^2-\lvert DW\rvert^2\big\rVert_{L^1}+\big\lVert\langle\mathcal{R}W_k,W_k\rangle-\langle\mathcal{R}W,W\rangle\big\rVert_{L^1}\lesssim\varepsilon_k\to0. ∣ ∇ W k ∣ 2 − ∣ ∇ W ∣ 2 L 1 + ∣ D W k ∣ 2 − ∣ D W ∣ 2 L 1 + ⟨ R W k , W k ⟩ − ⟨ R W , W ⟩ L 1 ≲ ε k → 0.
Denote the integrand on the right-hand side of (5) by b χ b_\chi b χ . Over A k A_k A k , the Cauchy–Schwarz inequality implies that
∥ b χ ( W k ) − b χ ( W ) ∥ L 1 ≲ ∥ d χ ∥ ∞ ( ∫ A k ( ∣ ∇ W k ∣ 2 + ∣ ∇ W ∣ 2 ) d V g ) 1 / 2 ( ∫ A k ( ∣ W k ∣ 2 + ∣ W ∣ 2 ) d V g ) 1 / 2 ≲ χ ε k → 0 , k → ∞ . \begin{aligned}
\lVert b_\chi(W_k)-b_\chi(W)\rVert_{L^1}
&\lesssim\lVert d\chi\rVert_\infty\left(\int_{A_k}(\lvert\nabla W_k\rvert^2+\lvert\nabla W\rvert^2)dV_g\right)^{1/2}
\left(\int_{A_k}(\lvert W_k\rvert^2+\lvert W\rvert^2)dV_g\right)^{1/2}\\
&\lesssim_\chi\varepsilon_k\to0,\qquad k\to\infty.
\end{aligned} ∥ b χ ( W k ) − b χ ( W ) ∥ L 1 ≲ ∥ d χ ∥ ∞ ( ∫ A k (∣ ∇ W k ∣ 2 + ∣ ∇ W ∣ 2 ) d V g ) 1/2 ( ∫ A k (∣ W k ∣ 2 + ∣ W ∣ 2 ) d V g ) 1/2 ≲ χ ε k → 0 , k → ∞.
Both sides of (5) therefore converge in L 1 L^1 L 1 . The integrated Weitzenböck formula follows. □
Local minimality of J \mathcal{J} J -energy for Z 2 \mathbb{Z}_2 Z 2 -harmonic spinors
Corollary 2.2 (Dirac solutions are J \mathcal{J} J -energy local minimisers). Let u u u be a Z 2 \mathbb{Z}_2 Z 2 -harmonic spinor. For every ball B ⋐ M B\Subset M B ⋐ M and every V ∈ W 1 , 2 ( B ; A 2 ( E ) ) V\in W^{1,2}(B;\mathcal{A}_2(E)) V ∈ W 1 , 2 ( B ; A 2 ( E )) with the same trace as u u u on ∂ B \partial B ∂ B , we have
J B ( V ) − J B ( u ) = ∥ D V ∥ L 2 ( B ) 2 ≥ 0. \mathcal{J}_B(V)-\mathcal{J}_B(u)=\lVert DV\rVert_{L^2(B)}^2\ge0. J B ( V ) − J B ( u ) = ∥ D V ∥ L 2 ( B ) 2 ≥ 0.
Proof. Choose a slightly larger ball B ~ \widetilde{B} B with B ⋐ B ~ ⋐ M B \Subset\widetilde{B} \Subset M B ⋐ B ⋐ M , and set
V ~ = { V in B , u in B ~ ∖ B . \widetilde{V} =
\begin{cases}
V & \text{in } B,\\
u & \text{in } \widetilde{B} \setminus B.
\end{cases} V = { V u in B , in B ∖ B .
The equality of the traces implies V ~ ∈ W 1 , 2 ( B ~ ; A 2 ( E ) ) \widetilde{V} \in W^{1,2}(\widetilde{B};\mathcal{A}_2(E)) V ∈ W 1 , 2 ( B ; A 2 ( E )) : in local trivialisations, the traces agree after the Almgren embedding [4 Proposition 2.10] , so the ordinary Sobolev gluing theorem applies; using the Lipschitz inverse on the embedded target, we then obtain the asserted two-valued Sobolev section. Choose a cutoff
χ ∈ C ∞ ( B ~ ) , 0 ≤ χ ≤ 1 , χ = 1 on a neighbourhood of B ‾ . \chi\in C^\infty(\widetilde{B}), \qquad0 \le\chi\le1, \qquad\chi= 1 \text{ on a neighbourhood of } \overline{B}. χ ∈ C ∞ ( B ) , 0 ≤ χ ≤ 1 , χ = 1 on a neighbourhood of B .
In particular, spt d χ ⊂ B ~ ∖ B ‾ \operatorname{spt} d\chi\subset\widetilde{B} \setminus\overline{B} spt d χ ⊂ B ∖ B , where V ~ = u \widetilde{V}=u V = u and their approximate differentials agree almost everywhere. Applying (5) to V ~ \widetilde{V} V and u u u and taking differences, we find
0 = J B ( V ~ ) − J B ( u ) − ∥ D V ~ ∥ L 2 ( B ) 2 + ∥ D u ∥ L 2 ( B ) 2 ≤ J B ( V ~ ) − J B ( u ) , 0=\mathcal{J}_{B}(\widetilde{V})-\mathcal{J}_{B}(u)-\lVert D\widetilde{V}\rVert_{L^2(B)}^2+\lVert Du\rVert_{L^2(B)}^2\leq\mathcal{J}_{B}(\widetilde{V})-\mathcal{J}_{B}(u), 0 = J B ( V ) − J B ( u ) − ∥ D V ∥ L 2 ( B ) 2 + ∥ D u ∥ L 2 ( B ) 2 ≤ J B ( V ) − J B ( u ) ,
as D u = 0 Du=0 D u = 0 .
Minimality implies the outer and inner variation identities recorded below.
Corollary 2.4 (Caccioppoli estimate). Let W W W be a local minimiser of the J \mathcal{J} J -energy in B t ( x 0 ) B_t(x_0) B t ( x 0 ) . For concentric balls B s ( x 0 ) ⊂ B t ( x 0 ) B_s(x_0)\subset B_t(x_0) B s ( x 0 ) ⊂ B t ( x 0 ) , where 0 < s < t < 1 0<s<t<1 0 < s < t < 1 , we have
∫ B s ( x 0 ) ∣ ∇ W ∣ 2 ≤ C ( t − s ) 2 ∫ B t ( x 0 ) ∣ W ∣ 2 . (6) \int_{B_s(x_0)}|\nabla W|^2\leq\frac{C}{(t-s)^2}\int_{B_t(x_0)}|W|^2. \tag*{(6)} ∫ B s ( x 0 ) ∣∇ W ∣ 2 ≤ ( t − s ) 2 C ∫ B t ( x 0 ) ∣ W ∣ 2 . (6) Proof. We simply sketch the proof as it is standard. Choose ζ ∈ C c ∞ ( B t ( x 0 ) ) \zeta\in C^\infty_c(B_t(x_0)) ζ ∈ C c ∞ ( B t ( x 0 )) with ζ = 1 \zeta=1 ζ = 1 on B s ( x 0 ) B_s(x_0) B s ( x 0 ) , 0 ≤ ζ ≤ 1 0\leq\zeta\leq1 0 ≤ ζ ≤ 1 , and ∣ d ζ ∣ ≲ ( t − s ) − 1 |d\zeta|\lesssim(t-s)^{-1} ∣ d ζ ∣ ≲ ( t − s ) − 1 . The outer variation W τ = ( 1 + τ ζ 2 ) W W_\tau=(1+\tau\zeta^2)W W τ = ( 1 + τ ζ 2 ) W and Young’s inequality give
∫ B t ζ 2 ∣ ∇ W ∣ 2 ≲ ∫ B t ( ∣ d ζ ∣ 2 + ζ 2 ) ∣ W ∣ 2 ≲ 1 ( t − s ) 2 ∫ B t ∣ W ∣ 2 . \int_{B_t}\zeta^2|\nabla W|^2\lesssim\int_{B_t}(|d\zeta|^2+\zeta^2)|W|^2\lesssim\frac{1}{(t-s)^2}\int_{B_t}|W|^2. ∫ B t ζ 2 ∣∇ W ∣ 2 ≲ ∫ B t ( ∣ d ζ ∣ 2 + ζ 2 ) ∣ W ∣ 2 ≲ ( t − s ) 2 1 ∫ B t ∣ W ∣ 2 .
Following [4 Theorem 3.9 and Proposition 3.10] and [Stu19, Theorem 5.1 and Proposition 5.2], we first obtain a uniform Hölder estimate and frequency control for Z 2 \mathbb{Z}_2 Z 2 -harmonic spinors, then prove strong W 1 , 2 W^{1,2} W 1 , 2 -convergence of blow-ups. The former is a standard consequence of J \mathcal{J} J -minimality and Campanato–Morrey iteration, while the latter follows by applying the integrated Weitzenböck formula to the rescaled spinors.
A uniform Hölder estimate
In a smooth metric trivialisation over B s ( x 0 ) B_s(x_0) B s ( x 0 ) , the J \mathcal{J} J -energy has the form
J B s ( x 0 ) ( W ) = 1 2 ∑ ℓ = 1 2 ∫ B s ( x 0 ) ( a α β i j ∂ i w ℓ α ∂ j w ℓ β + 2 b α β i w ℓ α ∂ i w ℓ β + c α β w ℓ α w ℓ β ) d x . (7) \mathcal{J}_{B_s(x_0)}(W)=\frac{1}{2}\sum_{\ell=1}^{2}\int_{B_s(x_0)}\left(a_{\alpha\beta}^{ij}\partial_iw_\ell^\alpha\partial_jw_\ell^\beta+2b_{\alpha\beta}^i w_\ell^\alpha\partial_iw_\ell^\beta+c_{\alpha\beta}w_\ell^\alpha w_\ell^\beta\right)\,dx. \tag*{(7)} J B s ( x 0 ) ( W ) = 2 1 ℓ = 1 ∑ 2 ∫ B s ( x 0 ) ( a α β ij ∂ i w ℓ α ∂ j w ℓ β + 2 b α β i w ℓ α ∂ i w ℓ β + c α β w ℓ α w ℓ β ) d x . (7)
where a α β i j = det g g i j h α β a_{\alpha\beta}^{ij}=\sqrt{\det g}\,g^{ij}h_{\alpha\beta} a α β ij = det g g ij h α β is uniformly positive, and a , b , c a,b,c a , b , c are smooth and bounded. The Euclidean Dirichlet energy is Dir ( W , B s ( x 0 ) ) = ∫ B s ( x 0 ) ∣ d W ∣ 2 = 1 2 ∑ ℓ = 1 2 ∫ B s ( x 0 ) ∣ d w ℓ ∣ 2 \operatorname{Dir}(W,B_s(x_0))=\int_{B_s(x_0)}|dW|^2=\frac{1}{2}\sum_{\ell=1}^{2}\int_{B_s(x_0)}|dw_\ell|^2 Dir ( W , B s ( x 0 )) = ∫ B s ( x 0 ) ∣ d W ∣ 2 = 2 1 ∑ ℓ = 1 2 ∫ B s ( x 0 ) ∣ d w ℓ ∣ 2 , with Dir ( W , ∂ B s ( x 0 ) ) \operatorname{Dir}(W,\partial B_s(x_0)) Dir ( W , ∂ B s ( x 0 )) denoting the tangential energy on the boundary.
Proposition 2.5 (Hölder regularity for local minimisers). Let W W W be a local minimiser of the functional J \mathcal{J} J . There exist γ > 0 \gamma>0 γ > 0 , Λ , C < ∞ \Lambda,C<\infty Λ , C < ∞ , and s 0 > 0 s_0>0 s 0 > 0 , depending only on the ellipticity and coefficient bounds, such that, if W W W is defined in B s 0 ( x 0 ) B_{s_0}(x_0) B s 0 ( x 0 ) , then for 0 < s < t < s 0 0<s<t<s_0 0 < s < t < s 0 ,
Dir ( W , B s ) ≤ C ( s t ) n − 2 + γ ( Dir ( W , B t ) + Λ ∫ B t ∣ W ∣ 2 ) . (8) \operatorname{Dir}(W,B_s)\le C\left(\frac{s}{t}\right)^{n-2+\gamma}\left(\operatorname{Dir}(W,B_t)+\Lambda\int_{B_t}|W|^2\right). \tag*{(8)} Dir ( W , B s ) ≤ C ( t s ) n − 2 + γ ( Dir ( W , B t ) + Λ ∫ B t ∣ W ∣ 2 ) . (8)
In particular, W ∈ C l o c γ / 2 W\in C_{\mathrm{loc}}^{\gamma/2} W ∈ C loc γ /2 .
Proof. We use geodesic coordinates centred at x 0 x_0 x 0 with radius s 0 ≪ 1 s_0\ll1 s 0 ≪ 1 and an orthonormal bundle frame. The following estimates are uniform for centres in a fixed compact set of M M M . As a competitor, take a Euclidean Dirichlet minimiser f f f with the same boundary data as W W W on ∂ B s \partial B_s ∂ B s . Since ∣ a − I ∣ ≲ s |a-I|\lesssim s ∣ a − I ∣ ≲ s and the lower-order coefficients are bounded, minimality and Young’s inequality give, for any ε > 0 \varepsilon>0 ε > 0 and sufficiently small s 0 s_0 s 0 ,
Dir ( f , B s ) ≤ Dir ( W , B s ) ≤ ( 1 + ε ) Dir ( f , B s ) + C ε ∫ B s ( ∣ W ∣ 2 + ∣ f ∣ 2 ) , 0 < s < s 0 . (9) \operatorname{Dir}(f,B_s)\le\operatorname{Dir}(W,B_s)\le(1+\varepsilon)\operatorname{Dir}(f,B_s)+C_\varepsilon\int_{B_s}(|W|^2+|f|^2),\qquad0<s<s_0. \tag*{(9)} Dir ( f , B s ) ≤ Dir ( W , B s ) ≤ ( 1 + ε ) Dir ( f , B s ) + C ε ∫ B s ( ∣ W ∣ 2 + ∣ f ∣ 2 ) , 0 < s < s 0 . (9)
By [4 Proposition 3.10] and [Stu19, Corollary 5.4], we may fix γ ∈ ( 0 , 1 ) \gamma\in(0,1) γ ∈ ( 0 , 1 ) sufficiently small so that
Dir ( f , B s ) ≤ s n − 2 + 2 γ Dir ( W , ∂ B s ) , (10) \operatorname{Dir}(f,B_s)\le\frac{s}{n-2+2\gamma}\operatorname{Dir}(W,\partial B_s), \tag*{(10)} Dir ( f , B s ) ≤ n − 2 + 2 γ s Dir ( W , ∂ B s ) , (10)
∫ B s ∣ v ∣ 2 ≤ s n − 2 + 2 γ ∫ ∂ B s ∣ v ∣ 2 + C s 2 Dir ( v , B s ) , v = W , f . (11) \int_{B_s}|v|^2\le\frac{s}{n-2+2\gamma}\int_{\partial B_s}|v|^2+Cs^2\operatorname{Dir}(v,B_s),\qquad v=W,f. \tag*{(11)} ∫ B s ∣ v ∣ 2 ≤ n − 2 + 2 γ s ∫ ∂ B s ∣ v ∣ 2 + C s 2 Dir ( v , B s ) , v = W , f . (11)
To absorb the lower-order terms, we follow [Stu19, Proposition 5.2] and add Λ ∫ B s ∣ W ∣ 2 \Lambda\int_{B_s}|W|^2 Λ ∫ B s ∣ W ∣ 2 to the Dirichlet energy. Choose ε \varepsilon ε small, then Λ \Lambda Λ large, and finally decrease s 0 s_0 s 0 . Combining these estimates and absorbing the terms proportional to s 2 Dir ( W , B s ) s^2\operatorname{Dir}(W,B_s) s 2 Dir ( W , B s ) gives, for almost every 0 < s < s 0 0<s<s_0 0 < s < s 0 ,
Dir ( W , B s ) + Λ ∫ B s ∣ W ∣ 2 ≤ s n − 2 + γ ( Dir ( W , ∂ B s ) + Λ ∫ ∂ B s ∣ W ∣ 2 ) . (12) \operatorname{Dir}(W,B_s)+\Lambda\int_{B_s}|W|^2\le\frac{s}{n-2+\gamma}\left(\operatorname{Dir}(W,\partial B_s)+\Lambda\int_{\partial B_s}|W|^2\right). \tag*{(12)} Dir ( W , B s ) + Λ ∫ B s ∣ W ∣ 2 ≤ n − 2 + γ s ( Dir ( W , ∂ B s ) + Λ ∫ ∂ B s ∣ W ∣ 2 ) . (12)
With φ ( r ) = Dir ( W , B r ) + Λ ∫ B r ∣ W ∣ 2 \varphi(r)=\operatorname{Dir}(W,B_r)+\Lambda\int_{B_r}|W|^2 φ ( r ) = Dir ( W , B r ) + Λ ∫ B r ∣ W ∣ 2 , the coarea formula and (12) imply
r φ ′ ( r ) ≥ ( n − 2 + γ ) φ ( r ) r\varphi'(r)\ge(n-2+\gamma)\varphi(r) r φ ′ ( r ) ≥ ( n − 2 + γ ) φ ( r )
for almost every r ∈ ( 0 , s 0 ) r\in(0,s_0) r ∈ ( 0 , s 0 ) . Integrating from s s s to t t t gives (8) . The local C γ / 2 C^{\gamma/2} C γ /2 -regularity follows from the Campanato–Morrey estimate.
Frequency control
We collect the frequency estimates needed to construct nonzero homogeneous blow-ups. For Z 2 \mathbb{Z}_2 Z 2 -harmonic spinors, the underlying identities and curvature estimates appear in [Tau14, Lemmas 3.1–3.3] and [22 Section 3.2] ; the corresponding variational argument for multiple-valued Jacobi fields is developed in [Stu19, Lemma 6.13 and Propositions 6.9–6.10].
Let u u u be a symmetric local J \mathcal{J} J -minimiser and p ∈ B u p\in B_u p ∈ B u . Suppressing the centre p p p , write
H ( r ) = H u ( p , r ) , J ( r ) = J B r ( p ) ( u ) , N ( r ) = r J ( r ) H ( r ) , h ( r ) = r 1 − n H ( r ) . H(r)=H_u(p,r),\qquad\mathcal{J}(r)=\mathcal{J}_{B_r(p)}(u),\qquad N(r)=\frac{r\mathcal{J}(r)}{H(r)},\qquad h(r)=r^{1-n}H(r). H ( r ) = H u ( p , r ) , J ( r ) = J B r ( p ) ( u ) , N ( r ) = H ( r ) r J ( r ) , h ( r ) = r 1 − n H ( r ) .
For sufficiently small r 0 > 0 r_0 > 0 r 0 > 0 , we have H ( r ) > 0 H(r) > 0 H ( r ) > 0 and N ( r ) ≥ − C r 2 N(r) \ge-Cr^2 N ( r ) ≥ − C r 2 for 0 < r ≤ r 0 0 < r \le r_0 0 < r ≤ r 0 . The key ingredients are the inner and outer variation identities in Remark 2.3 . Applying the outer variation to radial cutoffs, we get
J ( r ) = ∫ ∂ B r ( p ) ⟨ u , ∇ ν u ⟩ . (13) \mathcal{J}(r)=\int_{\partial B_r(p)}\langle u,\nabla_\nu u\rangle. \tag*{(13)} J ( r ) = ∫ ∂ B r ( p ) ⟨ u , ∇ ν u ⟩ . (13)
Differentiating the height and using radial inner variations give, respectively,
H ′ ( r ) = n − 1 r H ( r ) + 2 J ( r ) + O ( r H ( r ) ) , (14) H'(r)=\frac{n-1}{r}H(r)+2\mathcal{J}(r)+O(rH(r)), \tag*{(14)} H ′ ( r ) = r n − 1 H ( r ) + 2 J ( r ) + O ( r H ( r )) , (14)
J ′ ( r ) = n − 2 r J ( r ) + 2 ∫ ∂ B r ( p ) ∣ ∇ ν u ∣ 2 + O ( r ∫ B r ( p ) ∣ ∇ u ∣ 2 + H ( r ) ) . (15) \mathcal{J}'(r)=\frac{n-2}{r}\mathcal{J}(r)+2\int_{\partial B_r(p)}|\nabla_\nu u|^2+O\left(r\int_{B_r(p)}|\nabla u|^2+H(r)\right). \tag*{(15)} J ′ ( r ) = r n − 2 J ( r ) + 2 ∫ ∂ B r ( p ) ∣ ∇ ν u ∣ 2 + O ( r ∫ B r ( p ) ∣∇ u ∣ 2 + H ( r ) ) . (15)
These identities hold for almost every 0 < r < r 0 0<r<r_0 0 < r < r 0 and imply
N ′ ( r ) ≥ 2 r H ( r ) ∫ ∂ B r ( p ) ∣ ∇ ν u − N ( r ) r u ∣ 2 − C r ( 1 + N ( r ) ) . (16) N'(r)\ge\frac{2r}{H(r)}\int_{\partial B_r(p)}\left|\nabla_\nu u-\frac{N(r)}{r}u\right|^2-Cr(1+N(r)). \tag*{(16)} N ′ ( r ) ≥ H ( r ) 2 r ∫ ∂ B r ( p ) ∇ ν u − r N ( r ) u 2 − C r ( 1 + N ( r )) . (16)
Thus e A r 2 ( 1 + N ( r ) ) e^{Ar^2}(1+N(r)) e A r 2 ( 1 + N ( r )) is nondecreasing for sufficiently large A A A , and the limit ν p : = N u ( p , 0 ) = lim r ↓ 0 N ( r ) ∈ [ 0 , ∞ ) \nu_p:=N_u(p,0)=\lim_{r\downarrow0}N(r)\in[0,\infty) ν p := N u ( p , 0 ) = lim r ↓ 0 N ( r ) ∈ [ 0 , ∞ ) exists. Integrating (14) gives, for 0 < s < t ≤ r 0 0<s<t\le r_0 0 < s < t ≤ r 0 ,
∣ log h ( t ) h ( s ) − 2 ∫ s t N ( τ ) τ d τ ∣ ≤ C ( t 2 − s 2 ) . (17) \left|\log\frac{h(t)}{h(s)}-2\int_s^t\frac{N(\tau)}{\tau}\,d\tau\right|\le C(t^2-s^2). \tag*{(17)} log h ( s ) h ( t ) − 2 ∫ s t τ N ( τ ) d τ ≤ C ( t 2 − s 2 ) . (17)
Here r 0 r_0 r 0 , A A A and C C C depend only on the local geometry and the coefficients of J \mathcal{J} J .
Strong W 1 , 2 W^{1,2} W 1 , 2 -convergence of blow-ups
In the exponential gauge centred at a branch point p p p , define the normalised blow-ups by u r ( X ) = h ( r ) − 1 / 2 u ( exp p ( r X ) ) u_r(X)=h(r)^{-1/2}u(\exp_p(rX)) u r ( X ) = h ( r ) − 1/2 u ( exp p ( r X )) .
Theorem 2.6 (Blow-ups of Z 2 \mathbb{Z}_2 Z 2 -harmonic spinors). Let u u u be a Z 2 \mathbb{Z}_2 Z 2 -harmonic spinor, and let p ∈ B u p\in B_u p ∈ B u . For any sequence r j ↓ 0 r_j\downarrow0 r j ↓ 0 , the normalised blow-ups u r j u_{r_j} u r j admit a subsequence converging strongly in W l o c 1 , 2 W^{1,2}_{\mathrm{loc}} W loc 1 , 2 to a nonzero limit Φ \Phi Φ on T p M T_pM T p M .
The limit is locally Dirichlet-minimising, is homogeneous of degree N u ( p , 0 ) N_u(p,0) N u ( p , 0 ) , and satisfies D 0 Φ = 0 D_0\Phi=0 D 0 Φ = 0 , where D 0 D_0 D 0 is the constant-coefficient Dirac operator obtained by freezing the principal symbol of D D D at p p p .
Proof. Write u j = u r j u_j=u_{r_j} u j = u r j and g j = r j − 2 ( exp p ( r j ⋅ ) ) ∗ g g_j=r_j^{-2}(\exp_p(r_j\cdot))^*g g j = r j − 2 ( exp p ( r j ⋅ ) ) ∗ g . We first establish the W 1 , 2 W^{1,2} W 1 , 2 -weak convergence using the frequency function and then apply the integrated Weitzenböck formula (5) to derive minimality of the limit and energy convergence.
For each fixed R ≥ 1 R\ge1 R ≥ 1 , the frequency bounds and (17) give h ( ρ r j ) ≤ C R h ( r j ) h(\rho r_j)\le C_Rh(r_j) h ( ρ r j ) ≤ C R h ( r j ) for 0 < ρ ≤ 2 R 0<\rho\le2R 0 < ρ ≤ 2 R and r j < r 0 / ( 2 R ) r_j<r_0/(2R) r j < r 0 / ( 2 R ) . Here C R C_R C R depends also on N u ( p , r 0 ) N_u(p,r_0) N u ( p , r 0 ) , but not on j j j . By the coarea formula,
∫ B 2 R ∣ u j ∣ 2 d V g j = ∫ 0 2 R ρ n − 1 h ( ρ r j ) h ( r j ) d ρ ≤ C R . \int_{B_{2R}}|u_j|^2\,dV_{g_j}=\int_0^{2R}\rho^{n-1}\frac{h(\rho r_j)}{h(r_j)}\,d\rho\le C_R. ∫ B 2 R ∣ u j ∣ 2 d V g j = ∫ 0 2 R ρ n − 1 h ( r j ) h ( ρ r j ) d ρ ≤ C R .
The Caccioppoli estimate (6) and uniform equivalence of the rescaled metrics then yield
∫ B R ( ∣ u j ∣ 2 + ∣ d u j ∣ 2 ) d X ≤ C R . (18) \int_{B_R}\left(|u_j|^2+|du_j|^2\right)\,dX\le C_R. \tag*{(18)} ∫ B R ( ∣ u j ∣ 2 + ∣ d u j ∣ 2 ) d X ≤ C R . (18)
Proposition 2.5 also gives uniform Hölder bounds on each fixed ball. Passing to a diagonal subsequence, we can assume
u j → Φ strongly in L l o c 2 ∩ C l o c γ ′ / 2 , u j ⇀ Φ in W l o c 1 , 2 , γ ′ < γ , (19) u_j\to\Phi\quad\text{strongly in }L^2_{\mathrm{loc}}\cap C^{\gamma'/2}_{\mathrm{loc}},\qquad u_j\rightharpoonup\Phi\quad\text{in }W^{1,2}_{\mathrm{loc}},\qquad\gamma'<\gamma, \tag*{(19)} u j → Φ strongly in L loc 2 ∩ C loc γ ′ /2 , u j ⇀ Φ in W loc 1 , 2 , γ ′ < γ , (19)
with Φ ( 0 ) = 2 ∥ ⟦ ∅ ⟧ ∥ \Phi(0)=2\lVert\mathord{\llbracket}\mathord{\varnothing}\mathord{\rrbracket}\rVert Φ ( 0 ) = 2 ∥ [ [ ∅ ] ] ∥ . Passing to the limit in the weak formulation of the rescaled Dirac equations, using (19) and convergence of the coefficients, gives D 0 Φ = 0 D_0\Phi=0 D 0 Φ = 0 . The integrated Weitzenböck formula, as in Corollary 2.2, then shows that Φ \Phi Φ is locally Dirichlet-minimising. The usual homogeneity argument from (16) , using strong L l o c 2 L^2_{\mathrm{loc}} L loc 2 -convergence, shows that Φ \Phi Φ is homogeneous of degree N u ( p , 0 ) N_u(p,0) N u ( p , 0 ) . The normalisation then gives
∫ ∂ B ρ ∣ Φ ∣ 2 d Area = ρ n − 1 + 2 N u ( p , 0 ) > 0 , ρ > 0. (20) \int_{\partial B_\rho}|\Phi|^2\,d\operatorname{Area}=\rho^{n-1+2N_u(p,0)}>0,\qquad\rho>0. \tag*{(20)} ∫ ∂ B ρ ∣Φ ∣ 2 d Area = ρ n − 1 + 2 N u ( p , 0 ) > 0 , ρ > 0. (20)
Now, for χ ∈ C ∞ ( R n ) \chi\in C^\infty(\mathbb{R}^n) χ ∈ C ∞ ( R n ) , apply (5) to u j u_j u j in the rescaled geometry and use its Dirac equation. Smooth convergence of the rescaled coefficients and the bound (18) give
∫ χ ∣ d u j ∣ 2 d X = 1 2 ∫ ( Δ χ ) ∣ u j ∣ 2 d X + o ( 1 ) ⟶ 1 2 ∫ ( Δ χ ) ∣ Φ ∣ 2 d X = ∫ χ ∣ d Φ ∣ 2 d X . \int\chi|d u_j|^2\,dX = \frac{1}{2}\int(\Delta\chi)|u_j|^2\,dX + o(1)
\longrightarrow\frac{1}{2}\int(\Delta\chi)|\Phi|^2\,dX = \int\chi|d\Phi|^2\,dX. ∫ χ ∣ d u j ∣ 2 d X = 2 1 ∫ ( Δ χ ) ∣ u j ∣ 2 d X + o ( 1 ) ⟶ 2 1 ∫ ( Δ χ ) ∣Φ ∣ 2 d X = ∫ χ ∣ d Φ ∣ 2 d X .
Here Δ = ∑ i ∂ i 2 \Delta=\sum_i\partial_i^2 Δ = ∑ i ∂ i 2 , the limit follows from strong L l o c 2 L^2_{\mathrm{loc}} L loc 2 -convergence, and the last equality follows from the same formula and D 0 Φ = 0 D_0\Phi=0 D 0 Φ = 0 . Approximating ball indicators by smooth cutoffs yields
lim j → ∞ Dir ( u j , B R ) = Dir ( Φ , B R ) , R > 0. (21) \lim_{j\to\infty}\operatorname{Dir}(u_j,B_R)=\operatorname{Dir}(\Phi,B_R),\qquad R>0. \tag*{(21)} j → ∞ lim Dir ( u j , B R ) = Dir ( Φ , B R ) , R > 0. (21)
and hence strong W l o c 1 , 2 W^{1,2}_{\mathrm{loc}} W loc 1 , 2 -convergence. □
Frequency-one-half models
For any codimension-two plane P ⊂ R n P\subset\mathbb{R}^n P ⊂ R n , we decompose R n = P ⊥ ⊕ P ≅ C z ⊕ P \mathbb{R}^n=P^\perp\oplus P\cong\mathbb{C}_z\oplus P R n = P ⊥ ⊕ P ≅ C z ⊕ P . If E E E is real, we complexify E E E and extend Clifford multiplication complex linearly.
Definition 2.7 (Frequency-1 2 \frac{1}{2} 2 1 models). Let C D \mathcal{C}_D C D be the class of nonzero maps
Φ P , a , b ( z ) : = ⟦ a z 1 / 2 + b z − 1 / 2 ⟧ + ⟦ − a z 1 / 2 − b z − 1 / 2 ⟧ , z ∈ C , \Phi_{P,a,b}(z):=\llbracket az^{1/2}+bz^{-1/2}\rrbracket+\llbracket-az^{1/2}-bz^{-1/2}\rrbracket,\qquad z\in\mathbb{C}, Φ P , a , b ( z ) := [ [ a z 1/2 + b z − 1/2 ] ] + [ [ − a z 1/2 − b z − 1/2 ] ] , z ∈ C , where
J a = − i a , J b = i b , J = c ( e 1 ) c ( e 2 ) , Ja=-ia,\qquad Jb=ib,\qquad J=c(e_1)c(e_2), J a = − ia , J b = ib , J = c ( e 1 ) c ( e 2 ) , for an oriented orthonormal basis e 1 , e 2 e_1,e_2 e 1 , e 2 of P ⊥ P^\perp P ⊥ . The model is translation invariant along P P P . The coefficients lie in E p E_p E p in the complex case and in its complexification in the real case. For a real spinor bundle E E E , we impose in addition b = a ‾ b=\overline{a} b = a , where the bar denotes the natural conjugation on the complexification of E p E_p E p .
Proposition 2.8 (Minimum frequency and tangent classification). Let u u u be a nonzero Z 2 \mathbb{Z}_2 Z 2 -harmonic spinor and p ∈ B u p\in\mathcal{B}_u p ∈ B u . Then
N u ( p , 0 ) ≥ 1 2 . N_u(p,0)\geq\frac{1}{2}. N u ( p , 0 ) ≥ 2 1 . If N u ( p , 0 ) = 1 2 N_u(p,0)=\frac{1}{2} N u ( p , 0 ) = 2 1 , every tangent map at p p p belongs to C D \mathcal{C}_D C D .
Proof. The dimension-reduction argument in Taubes [Tau14, Lemmas 5.1, 5.4, 5.5 and 6.4] applies in arbitrary dimension. We recall the main points. By Theorem 2.6, every tangent map Φ \Phi Φ is a nonzero homogeneous Dirichlet minimiser satisfying D 0 Φ = 0 D_0\Phi=0 D 0 Φ = 0 , of positive degree α = N u ( p , 0 ) \alpha=N_u(p,0) α = N u ( p , 0 ) . In dimension two, the angular Dirac equation gives α ∈ 1 2 Z ≥ 1 \alpha\in\frac{1}{2}\mathbb{Z}_{\geq1} α ∈ 2 1 Z ≥ 1 . For n ≥ 3 n\geq3 n ≥ 3 , if Z Φ = { 0 } Z_\Phi=\{0\} Z Φ = { 0 } , a sheet extends smoothly across the origin, so α \alpha α is a positive integer. Otherwise, at a zero q ≠ 0 q\neq0 q = 0 , the cylindrical blow-up lemma [4 Lemma 3.24] gives a tangent map invariant in the direction q q q . Induction on the dimension, together with frequency monotonicity and N Φ ( q , R ) → α N_\Phi(q,R)\to\alpha N Φ ( q , R ) → α as R → ∞ R\to\infty R → ∞ , yields
1 2 ≤ N Φ ( q , 0 ) ≤ N Φ ( q , R ) ≤ α . \frac{1}{2}\leq N_\Phi(q,0)\leq N_\Phi(q,R)\leq\alpha. 2 1 ≤ N Φ ( q , 0 ) ≤ N Φ ( q , R ) ≤ α .
If α = 1 2 \alpha=\frac{1}{2} α = 2 1 , equality holds throughout. Thus Φ \Phi Φ is homogeneous about both 0 0 0 and q q q , hence invariant in the direction q q q . Iterating reduces to the two-dimensional angular equation and gives a sheet
ϕ ( z ) = a z 1 / 2 + b z ‾ 1 / 2 , J a = − i a , J b = i b , ∣ Φ ( z ) ∣ 2 = ( ∣ a ∣ 2 + ∣ b ∣ 2 ) ∣ z ∣ . (22) \phi(z)=az^{1/2}+b\overline{z}^{1/2},\qquad Ja=-ia,\qquad Jb=ib,\qquad|\Phi(z)|^2=(|a|^2+|b|^2)|z|. \tag*{(22)} ϕ ( z ) = a z 1/2 + b z 1/2 , J a = − ia , J b = ib , ∣Φ ( z ) ∣ 2 = ( ∣ a ∣ 2 + ∣ b ∣ 2 ) ∣ z ∣. (22)
translation invariant along an ( n − 2 ) (n-2) ( n − 2 ) -plane. In the real case b = a ‾ b=\overline{a} b = a . Hence Φ ∈ C D \Phi\in\mathcal{C}_D Φ ∈ C D . □
determinant. By continuity, so is π T \pi T π T for every model Φ \Phi Φ near Φ 0 \Phi_0 Φ 0 with the same centre, and by the polar decomposition there is a unique ϕ ∈ [ 0 , 2 π ) \phi\in[0,2\pi) ϕ ∈ [ 0 , 2 π ) such that π T ∘ e − i ϕ \pi T \circ e^{-i\phi} π T ∘ e − i ϕ is symmetric positive definite. We call the corresponding normal frame and square root the model gauge . In this gauge the parametrisation is unique, and we can compare nearby models by
d C ( Φ , Φ ′ ) 2 = ∑ j = 1 2 ∣ e j − e j ′ ∣ 2 + ∣ a − a ′ ∣ 2 + ∣ b − b ′ ∣ 2 , d_{\mathcal C}(\Phi,\Phi')^2=\sum_{j=1}^{2}|e_j-e'_j|^2+|a-a'|^2+|b-b'|^2, d C ( Φ , Φ ′ ) 2 = j = 1 ∑ 2 ∣ e j − e j ′ ∣ 2 + ∣ a − a ′ ∣ 2 + ∣ b − b ′ ∣ 2 ,
where ( e 1 , e 2 ) (e_1,e_2) ( e 1 , e 2 ) and ( e 1 ′ , e 2 ′ ) (e'_1,e'_2) ( e 1 ′ , e 2 ′ ) are the normal frames of Φ \Phi Φ and Φ ′ \Phi' Φ ′ in the model gauge; the first term also measures the distance between their spines. A compact family of models in such a neighbourhood is called a non-degenerate model chart .
We record the infinitesimal variations needed in the excess-decay argument.
Lemma 2.10 (Variations of Dirac models). Fix Φ = { ± φ } ∈ C D \Phi=\{\pm\varphi\}\in\mathcal C_D Φ = { ± φ } ∈ C D with φ = a z 1 / 2 + b z ˉ 1 / 2 \varphi=az^{1/2}+b\bar{z}^{1/2} φ = a z 1/2 + b z ˉ 1/2 and spine P P P . Write z = x 1 + i x 2 z=x_1+ix_2 z = x 1 + i x 2 and choose orthonormal coordinates ( y 1 , … , y n − 2 ) (y_1,\ldots,y_{n-2}) ( y 1 , … , y n − 2 ) on P P P . The tangent space T Φ C D T_\Phi\mathcal C_D T Φ C D consists of the fields
ℓ = A z 1 / 2 + B z ˉ 1 / 2 + ∑ α = 1 n − 2 ∑ i = 1 2 c α i y α ∂ i φ , (23) \ell=Az^{1/2}+B\bar{z}^{1/2}+\sum_{\alpha=1}^{n-2}\sum_{i=1}^{2}c_{\alpha i}y_\alpha\partial_i\varphi, \tag*{(23)} ℓ = A z 1/2 + B z ˉ 1/2 + α = 1 ∑ n − 2 i = 1 ∑ 2 c α i y α ∂ i φ , (23) where A , B A,B A , B are fibre constants and c α i ∈ R c_{\alpha i}\in\mathbb R c α i ∈ R , satisfying D 0 ℓ = 0 D_0\ell=0 D 0 ℓ = 0 on R n ∖ P \mathbb R^n\setminus P R n ∖ P ; in the real case, B = A ˉ B=\bar{A} B = A ˉ . Every such field is realised by a smooth one-parameter family Φ s ∈ C D \Phi_s\in\mathcal C_D Φ s ∈ C D with Φ 0 = Φ \Phi_0=\Phi Φ 0 = Φ and
∥ G ( Φ s , { ± ( φ + s ℓ ) } ) ∥ L 2 ( B 1 ) = O ( ∣ s ∣ 3 / 2 ) . (24) \left\|G\left(\Phi_s,\{\pm(\varphi+s\ell)\}\right)\right\|_{L^2(B_1)}=O\left(|s|^{3/2}\right). \tag*{(24)} ∥ G ( Φ s , { ± ( φ + s ℓ )} ) ∥ L 2 ( B 1 ) = O ( ∣ s ∣ 3/2 ) . (24)
Sketch of the proof. Varying the coefficients gives the first two terms in (23) ; tilting the normal plane gives the terms c α i y α ∂ i φ c_{\alpha i}y_\alpha\partial_i\varphi c α i y α ∂ i φ . The Dirac constraints are ( c ( e 1 ) + i c ( e 2 ) ) a = 0 (c(e_1)+ic(e_2))a=0 ( c ( e 1 ) + i c ( e 2 )) a = 0 and ( c ( e 1 ) − i c ( e 2 ) ) b = 0 (c(e_1)-ic(e_2))b=0 ( c ( e 1 ) − i c ( e 2 )) b = 0 . Differentiating these constraints restricts the fields in (23) to those satisfying D 0 ℓ = 0 D_0\ell=0 D 0 ℓ = 0 . Since the corresponding eigenspaces of J J J vary smoothly with the normal plane, every such infinitesimal variation is realised by a smooth one-parameter family in C D \mathcal C_D C D . The error estimate (24) follows from Taylor expansion. □
Estimates near a cylindrical spine
This section gives estimates for spinors close to cylindrical models of frequency 1 / 2 1/2 1/2 , beginning with sheet matching and the topological exclusion of holes. We adapt Simon’s cylindrical cone technique, as developed further by Krummel–Wickramasekera [Sim93, KW13], taking advantage of some simplifications from the Dirac equation. See also the special case of triple junctions [21 ] .
Given a Z 2 \mathbb Z_2 Z 2 -harmonic spinor on a geodesic ball B r ( 0 ) B_r(0) B r ( 0 ) , dilate the ball to B 1 B_1 B 1 and multiply the spinor values by r − 1 / 2 r^{-1/2} r − 1/2 . Throughout this section, write u u u for the resulting spinor on B 1 B_1 B 1 and g g g , ∇ \nabla ∇ , R \mathcal R R for the rescaled ambient data, with δ = C r 2 \delta=Cr^2 δ = C r 2 as in Remark 2.3 . We use geodesic coordinates centred at 0 0 0 , identifying fibres by parallel transport along radial geodesics.
Write
X = ( z , y ) ∈ C × R n − 2 , P = { z = 0 } , ρ = ∣ z ∣ , R = ∣ X ∣ . X=(z,y)\in\mathbb C\times\mathbb R^{n-2},\qquad P=\{z=0\},\qquad\rho=|z|,\qquad R=|X|. X = ( z , y ) ∈ C × R n − 2 , P = { z = 0 } , ρ = ∣ z ∣ , R = ∣ X ∣.
Unless indicated otherwise, balls and volume integrals use g g g . We write D D D for coordinate differentiation and d X dX d X for coordinate measure, and set z = x 1 + i x 2 z=x_1+ix_2 z = x 1 + i x 2 . Throughout this section, Φ = { ± φ } \Phi=\{\pm\varphi\} Φ = { ± φ } belongs to a fixed compact nondegenerate Dirac model chart C 0 ∈ C D \mathcal C_0\in\mathcal C_D C 0 ∈ C D , with spine P P P . Its 1 / 2 1/2 1/2 -homogeneity makes Φ \Phi Φ invariant under the dilation. Small L 2 L^2 L 2 -excess relative to this family gives a uniform L 2 L^2 L 2 -bound on u u u .
Sheet matching away from the branching locus
Away from the model spine, the Hölder estimate and small L 2 L^2 L 2 -excess allow us to match the local sheets of u u u and Φ \Phi Φ and apply elliptic estimates to their difference.
Lemma 3.1 (Sheet matching). Let Φ ∈ C 0 \Phi\in C_0 Φ ∈ C 0 have spine P P P , and let K K K be a compact set in B 3 / 4 B_{3/4} B 3/4 at distance at least c 0 > 0 c_0 > 0 c 0 > 0 from both P P P and ∂ B 3 / 4 \partial B_{3/4} ∂ B 3/4 . There exist ε , C > 0 \varepsilon, C > 0 ε , C > 0 and κ 0 ∈ ( 0 , 1 ) \kappa_0 \in(0,1) κ 0 ∈ ( 0 , 1 ) such that, if
∫ B 1 G ( u , Φ ) 2 + δ < ε 2 , \int_{B_1} \mathcal{G}(u,\Phi)^2 + \delta< \varepsilon^2, ∫ B 1 G ( u , Φ ) 2 + δ < ε 2 , then u u u has no zero near K K K and has a unique matching with Φ \Phi Φ such that G ( u , Φ ) 2 = ∣ u − φ ∣ 2 \mathcal{G}(u,\Phi)^2 = \lvert u-\varphi\rvert^2 G ( u , Φ ) 2 = ∣ u − φ ∣ 2 and
∥ u − φ ∥ C 1 ( K ) ≤ C ( ∥ G ( u , Φ ) ∥ L 2 ( B 1 ) κ 0 + δ ) . (25) \lVert u-\varphi\rVert_{C^1(K)} \le C\left(\lVert\mathcal{G}(u,\Phi)\rVert_{L^2(B_1)}^{\kappa_0}+\delta\right). \tag*{(25)} ∥ u − φ ∥ C 1 ( K ) ≤ C ( ∥ G ( u , Φ ) ∥ L 2 ( B 1 ) κ 0 + δ ) . (25) The constants are uniform for Φ ∈ C 0 \Phi\in C_0 Φ ∈ C 0 .
Proof. Corollary 2.4 and Proposition 2.5 give a uniform Hölder bound for G ( u , Φ ) \mathcal{G}(u,\Phi) G ( u , Φ ) . Interpolation with its L ∞ L^\infty L ∞ -norm yields
G ( u , Φ ) ≤ C ∥ G ( u , Φ ) ∥ L 2 ( B 1 ) κ 0 \mathcal{G}(u,\Phi) \le C\lVert\mathcal{G}(u,\Phi)\rVert_{L^2(B_1)}^{\kappa_0} G ( u , Φ ) ≤ C ∥ G ( u , Φ ) ∥ L 2 ( B 1 ) κ 0
on a fixed neighbourhood of K K K , for some κ 0 ∈ ( 0 , 1 ) \kappa_0 \in(0,1) κ 0 ∈ ( 0 , 1 ) .
Choose a local model branch φ \varphi φ . Its norm has a uniform positive lower bound; choose ε \varepsilon ε so that C ε κ 0 < 1 2 ∣ φ ∣ C\varepsilon^{\kappa_0}<\frac{1}{2}\lvert\varphi\rvert C ε κ 0 < 2 1 ∣ φ ∣ on this neighbourhood. The hypothesis then gives G ( u , Φ ) < ∣ φ ∣ / 2 \mathcal{G}(u,\Phi)<\lvert\varphi\rvert/2 G ( u , Φ ) < ∣ φ ∣ /2 . The disjoint fibre balls of radius ∣ φ ∣ / 2 \lvert\varphi\rvert/2 ∣ φ ∣ /2 around ± φ \pm\varphi ± φ therefore contain one value of u u u each. This uniquely selects continuous W 1 , 2 W^{1,2} W 1 , 2 -branches, which satisfy the Dirac equation weakly.
Writing u u u for this branch, we have G ( u , Φ ) = ∣ u − φ ∣ \mathcal{G}(u,\Phi)=\lvert u-\varphi\rvert G ( u , Φ ) = ∣ u − φ ∣ and D ( u − φ ) = ( D 0 − D ) φ D(u-\varphi)=(D_0-D)\varphi D ( u − φ ) = ( D 0 − D ) φ . The derivatives of φ \varphi φ are uniformly bounded away from P P P , while the coefficients of D − D 0 D-D_0 D − D 0 and their derivatives are O ( δ ) O(\delta) O ( δ ) . Interior elliptic estimates therefore give (25) . □
Topological exclusion of holes
The matching in Lemma 3.1 identifies the Z 2 \mathbb{Z}_2 Z 2 -monodromy of u u u and Φ \Phi Φ : their sheets have the same sign change around closed curves. In particular, u u u has nontrivial monodromy around each small normal circle where the sheets are matched.
Lemma 3.2 (No-hole principle). For each 0 < τ < 1 / 4 0<\tau<1/4 0 < τ < 1/4 , there exists ε ( τ ) > 0 \varepsilon(\tau)>0 ε ( τ ) > 0 , uniform for Φ ∈ C 0 \Phi\in C_0 Φ ∈ C 0 , such that
∫ B 1 G ( u , Φ ) 2 d X + δ < ε ( τ ) \int_{B_1} \mathcal{G}(u,\Phi)^2\,dX+\delta<\varepsilon(\tau) ∫ B 1 G ( u , Φ ) 2 d X + δ < ε ( τ ) implies that every normal disk of radius τ / 4 \tau/4 τ /4 centred at a point of P ∩ B 1 / 2 P\cap B_{1/2} P ∩ B 1/2 contains a branch point.
Proof. The boundary circles of these disks stay a fixed positive distance from P P P and ∂ B 3 / 4 \partial B_{3/4} ∂ B 3/4 , depending only on τ \tau τ . For sufficiently small ε ( τ ) \varepsilon(\tau) ε ( τ ) , Lemma 3.1 gives nonzero matched sheets near all these circles. Thus u u u has the same nontrivial Z 2 \mathbb{Z}_2 Z 2 -monodromy as Φ \Phi Φ around each circle.
If a disk contained no branch point, choose a contractible neighbourhood of the disk disjoint from B u \mathcal{B}_u B u , using that B u \mathcal{B}_u B u is closed. Thus we can choose local continuous sheets of u u u . They satisfy the Dirac equation weakly and hence are smooth. By unique continuation and contractibility, u u u admits a single-valued smooth lift on the disk. The monodromy around the boundary circle is therefore trivial, a contradiction. □
The estimates at branch points
Assume now that 0 ∈ B u 0\in\mathcal{B}_u 0 ∈ B u , and define
E ( u , Φ ; p , s ) = s − n − 1 ∫ B s ( p ) G ( u , Φ ) 2 d V g , E : = E ( u , Φ ; 0 , 1 ) . \mathcal{E}(u,\Phi;p,s)=s^{-n-1}\int_{B_s(p)}\mathcal{G}(u,\Phi)^2\,dV_g,\qquad E:=\mathcal{E}(u,\Phi;0,1). E ( u , Φ ; p , s ) = s − n − 1 ∫ B s ( p ) G ( u , Φ ) 2 d V g , E := E ( u , Φ ; 0 , 1 ) .
For p = ( ξ , ζ ) p=(\xi,\zeta) p = ( ξ , ζ ) , write R p = d g ( p , ⋅ ) R_p=d_g(p,\cdot) R p = d g ( p , ⋅ ) and let ∂ R p \partial_{R_p} ∂ R p denote radial differentiation in geodesic coordinates centred at p p p .
Proposition 3.3 (L 2 L^2 L 2 -estimates at branch points). There exists ε > 0 \varepsilon>0 ε > 0 , depending only on the fixed model family and ambient bounds, such that if E + δ < ε E+\delta<\varepsilon E + δ < ε , then for every p = ( ξ , ζ ) ∈ B u ∩ B 9 / 16 p=(\xi,\zeta)\in\mathcal{B}_u\cap B_{9/16} p = ( ξ , ζ ) ∈ B u ∩ B 9/16 ,
∣ ξ ∣ 2 + ∫ B 1 / 2 R 2 − n ∣ ∂ R ( u R 1 / 2 ) ∣ 2 + ∫ B 1 / 2 ∣ D g u ∣ 2 ≤ C ( E + δ ) . (26) |\xi|^2+\int_{B_{1/2}}R^{2-n}\left|\partial_R\left(\frac{u}{R^{1/2}}\right)\right|^2+\int_{B_{1/2}}|D_g u|^2\leq C(E+\delta). \tag*{(26)} ∣ ξ ∣ 2 + ∫ B 1/2 R 2 − n ∂ R ( R 1/2 u ) 2 + ∫ B 1/2 ∣ D g u ∣ 2 ≤ C ( E + δ ) . (26) For every 0 < σ < 1 0<\sigma<1 0 < σ < 1 , we have the non-concentration estimate, with translation taken in the original coordinates,
∫ B 1 / 4 ( p ) G ( u ( X ) , Φ ( X − p ) ) 2 R p ( X ) n + 1 − σ d V g ( X ) + ∫ B 1 / 2 G ( u ( X ) , Φ ( X ) ) 2 R p ( X ) n − 1 − σ d V g ( X ) ≤ C σ ( E + δ ) . (27) \int_{B_{1/4}(p)}\frac{\mathcal{G}(u(X),\Phi(X-p))^2}{R_p(X)^{n+1-\sigma}}\,dV_g(X)+\int_{B_{1/2}}\frac{\mathcal{G}(u(X),\Phi(X))^2}{R_p(X)^{n-1-\sigma}}\,dV_g(X)\leq C_\sigma(E+\delta). \tag*{(27)} ∫ B 1/4 ( p ) R p ( X ) n + 1 − σ G ( u ( X ) , Φ ( X − p ) ) 2 d V g ( X ) + ∫ B 1/2 R p ( X ) n − 1 − σ G ( u ( X ) , Φ ( X ) ) 2 d V g ( X ) ≤ C σ ( E + δ ) . (27)
Remark 3.4. Estimate (26) controls the normal displacement, homogeneity defect and tangential energy. The bound for the second integral in (27) , combined with the no-hole principle, yields L 2 L^2 L 2 non-concentration of the normalised errors near the spine.
Proof. Step 1: the Weiss energy and variational inequalities. Let H H H and F \mathcal{F} F be as in Subsection 2.5, with centre 0. Define the Weiss energy
W ( s ) = s 1 − n G ( s ) − 1 2 s − n H ( s ) , 0 < s < 3 / 4. W(s)=s^{1-n}\mathcal{G}(s)-\frac{1}{2}s^{-n}H(s),\qquad0<s<3/4. W ( s ) = s 1 − n G ( s ) − 2 1 s − n H ( s ) , 0 < s < 3/4.
By Proposition 2.8 and the frequency estimates, N ( s ) ≥ 1 2 − C δ s 2 N(s)\geq\frac{1}{2}-C\delta s^2 N ( s ) ≥ 2 1 − C δ s 2 . The uniform L 2 L^2 L 2 -bound and (6) then imply
H ( s ) ≤ C s n , ∫ B s ∣ ∇ u ∣ g 2 d V g ≤ C s n − 1 . H(s)\leq Cs^n,\qquad\int_{B_s}|\nabla u|_g^2\,dV_g\leq Cs^{n-1}. H ( s ) ≤ C s n , ∫ B s ∣∇ u ∣ g 2 d V g ≤ C s n − 1 .
Consequently W ( s ) = s − n H ( s ) ( N ( s ) − 1 2 ) ≥ − C δ s 2 W(s)=s^{-n}H(s)(N(s)-\frac{1}{2})\geq-C\delta s^2 W ( s ) = s − n H ( s ) ( N ( s ) − 2 1 ) ≥ − C δ s 2 , so lim inf s ↓ 0 W ( s ) ≥ 0 \liminf_{s\downarrow0}W(s)\geq0 lim inf s ↓ 0 W ( s ) ≥ 0 . Using (13) –(15) for the rescaled data, we obtain
W ′ ( s ) = 2 s 1 − n ∫ ∂ B s ∣ ∇ ∂ R u − u 2 s ∣ 2 d A r e a g + O ( δ s 2 − n ∫ B s ∣ ∇ u ∣ g 2 d V g + δ s 1 − n H ( s ) ) . W'(s)=2s^{1-n}\int_{\partial B_s}\left|\nabla_{\partial_R}u-\frac{u}{2s}\right|^2\,dA_{\mathrm{reag}}+O\left(\delta s^{2-n}\int_{B_s}|\nabla u|_g^2\,dV_g+\delta s^{1-n}H(s)\right). W ′ ( s ) = 2 s 1 − n ∫ ∂ B s ∇ ∂ R u − 2 s u 2 d A reag + O ( δ s 2 − n ∫ B s ∣∇ u ∣ g 2 d V g + δ s 1 − n H ( s ) ) .
The preceding bounds give an O ( δ s ) O(\delta s) O ( δ s ) error. Integrating yields
W ( s ) − W ( t ) = 2 ∫ B s ∖ B t R 2 − n ∣ ∂ R ( u R 1 / 2 ) ∣ 2 d V g + O ( δ s 2 ) , 0 < t < s < 3 / 4. (28) W(s)-W(t)=2\int_{B_s\setminus B_t}R^{2-n}\left|\partial_R\left(\frac{u}{R^{1/2}}\right)\right|^2\,dV_g+O(\delta s^2),\qquad0<t<s<3/4. \tag*{(28)} W ( s ) − W ( t ) = 2 ∫ B s ∖ B t R 2 − n ∂ R ( R 1/2 u ) 2 d V g + O ( δ s 2 ) , 0 < t < s < 3/4. (28)
Letting t ↓ 0 t\downarrow0 t ↓ 0 along a sequence attaining the lower limit, we obtain
W ( s ) ≥ 2 ∫ B s R 2 − n ∣ ∂ R ( u R 1 / 2 ) ∣ 2 d V g − C δ s 2 . W(s)\geq2\int_{B_s}R^{2-n}\left|\partial_R\left(\frac{u}{R^{1/2}}\right)\right|^2\,dV_g-C\delta s^2. W ( s ) ≥ 2 ∫ B s R 2 − n ∂ R ( R 1/2 u ) 2 d V g − C δ s 2 .
Choose a smooth nonincreasing function ψ : [ 0 , ∞ ) → [ 0 , 1 ] \psi:[0,\infty)\to[0,1] ψ : [ 0 , ∞ ) → [ 0 , 1 ] with ψ = 1 \psi=1 ψ = 1 on [ 0 , 1 / 2 ] [0,1/2] [ 0 , 1/2 ] and ψ = 0 \psi=0 ψ = 0 on [ 3 / 4 , ∞ ) [3/4,\infty) [ 3/4 , ∞ ) . Throughout the following calculations, we write ψ \psi ψ and ψ ′ \psi' ψ ′ for ψ ( R ) \psi(R) ψ ( R ) and ψ ′ ( R ) \psi'(R) ψ ′ ( R ) . By Remark 2.3 and the Caccioppoli estimate (6) , replacing ∣ ∇ u ∣ g 2 + ⟨ R u , u ⟩ |\nabla u|_g^2+\langle\mathcal{R}u,u\rangle ∣∇ u ∣ g 2 + ⟨ R u , u ⟩ and d V g dV_g d V g by ∣ D u ∣ 2 |Du|^2 ∣ D u ∣ 2 and d X dX d X , respectively, introduces an O ( δ ) O(\delta) O ( δ ) error. Combined with the coarea formula and integration by parts, we find
∫ B 3 / 4 ( ∣ D u ∣ 2 ψ 2 + R − 1 ∣ u ∣ 2 ψ ψ ′ ) d X = ∫ B 3 / 4 [ ( ∣ ∇ u ∣ g 2 + ⟨ R u , u ⟩ ) ψ 2 + R − 1 ∣ u ∣ 2 ψ ψ ′ ] d V g + O ( δ ) = ∫ 1 / 2 3 / 4 ( − 2 ψ ( s ) ψ ′ ( s ) ) s n − 1 W ( s ) d s + O ( δ ) ≥ 2 ∫ 1 / 2 3 / 4 ( − 2 ψ ψ ′ ) s n − 1 ( ∫ B s R 2 − n ∣ ∂ R ( u R 1 / 2 ) ∣ 2 d V g ) d s − C δ ≥ c ∫ B 1 / 2 R 2 − n ∣ ∂ R ( u R 1 / 2 ) ∣ 2 d X − C δ . (29) \begin{aligned}
\int_{B_{3/4}} \left(|Du|^2\psi^2+R^{-1}|u|^2\psi\psi'\right)\,dX
={}&\int_{B_{3/4}}\left[(|\nabla u|_g^2+\langle\mathcal{R}u,u\rangle)\psi^2+R^{-1}|u|^2\psi\psi'\right]\,dV_g+O(\delta)\\
={}&\int_{1/2}^{3/4}(-2\psi(s)\psi'(s))s^{n-1}W(s)\,ds+O(\delta)\\
\geq{}&2\int_{1/2}^{3/4}(-2\psi\psi')s^{n-1}\left(\int_{B_s}R^{2-n}\left|\partial_R\left(\frac{u}{R^{1/2}}\right)\right|^2\,dV_g\right)\,ds-C\delta\\
\geq{}&c\int_{B_{1/2}}R^{2-n}\left|\partial_R\left(\frac{u}{R^{1/2}}\right)\right|^2\,dX-C\delta. \tag*{(29)}
\end{aligned} ∫ B 3/4 ( ∣ D u ∣ 2 ψ 2 + R − 1 ∣ u ∣ 2 ψ ψ ′ ) d X = = ≥ ≥ ∫ B 3/4 [ ( ∣∇ u ∣ g 2 + ⟨ R u , u ⟩) ψ 2 + R − 1 ∣ u ∣ 2 ψ ψ ′ ] d V g + O ( δ ) ∫ 1/2 3/4 ( − 2 ψ ( s ) ψ ′ ( s )) s n − 1 W ( s ) d s + O ( δ ) 2 ∫ 1/2 3/4 ( − 2 ψ ψ ′ ) s n − 1 ( ∫ B s R 2 − n ∂ R ( R 1/2 u ) 2 d V g ) d s − C δ c ∫ B 1/2 R 2 − n ∂ R ( R 1/2 u ) 2 d X − C δ . (29)
where c > 0 c>0 c > 0 depends only on n n n . To control the tangential energy as well, apply the outer variation ψ 2 u \psi^2u ψ 2 u and the inner variation ψ 2 ( z , 0 ) \psi^2(z,0) ψ 2 ( z , 0 ) to the rescaled J \mathcal{J} J -energy, with the same O ( δ ) O(\delta) O ( δ ) control of the coordinate errors. Thus
1 2 ∫ ∣ D u ∣ 2 ψ 2 d X = − ∫ ⟨ u , ρ ∂ ρ u + y ⋅ D y u ⟩ R − 1 ψ ψ ′ d X + O ( δ ) , \frac{1}{2}\int|Du|^2\psi^2\,dX=-\int\langle u,\rho\partial_\rho u+y\cdot D_yu\rangle R^{-1}\psi\psi'\,dX+O(\delta), 2 1 ∫ ∣ D u ∣ 2 ψ 2 d X = − ∫ ⟨ u , ρ ∂ ρ u + y ⋅ D y u ⟩ R − 1 ψ ψ ′ d X + O ( δ ) ,
∫ ∣ D y u ∣ 2 ψ 2 d X = − 2 ∫ ( 1 2 ρ 2 ∣ D u ∣ 2 − ⟨ ρ ∂ ρ u , ρ ∂ ρ u + y ⋅ D y u ⟩ ) R − 1 ψ ψ ′ d X + O ( δ ) . \int|D_yu|^2\psi^2\,dX=-2\int\left(\frac{1}{2}\rho^2|Du|^2-\langle\rho\partial_\rho u,\rho\partial_\rho u+y\cdot D_yu\rangle\right)R^{-1}\psi\psi'\,dX+O(\delta). ∫ ∣ D y u ∣ 2 ψ 2 d X = − 2 ∫ ( 2 1 ρ 2 ∣ D u ∣ 2 − ⟨ ρ ∂ ρ u , ρ ∂ ρ u + y ⋅ D y u ⟩ ) R − 1 ψ ψ ′ d X + O ( δ ) .
To express the error relative to Φ \Phi Φ , set
Err ( u , Φ ) = 1 2 ρ 2 ( ∣ D u ∣ 2 − ∣ D Φ ∣ 2 ) − ⟨ ρ ∂ ρ u , ρ ∂ ρ u − u / 2 ⟩ − 1 4 ( ∣ u ∣ 2 − ∣ Φ ∣ 2 ) . \operatorname{Err}(u,\Phi)=\frac{1}{2}\rho^2\left(|Du|^2-|D\Phi|^2\right)-\langle\rho\partial_\rho u,\rho\partial_\rho u-u/2\rangle-\frac{1}{4}\left(|u|^2-|\Phi|^2\right). Err ( u , Φ ) = 2 1 ρ 2 ( ∣ D u ∣ 2 − ∣ D Φ ∣ 2 ) − ⟨ ρ ∂ ρ u , ρ ∂ ρ u − u /2 ⟩ − 4 1 ( ∣ u ∣ 2 − ∣Φ ∣ 2 ) .
The two terms involving only Φ \Phi Φ cancel pointwise, since ρ 2 ∣ D Φ ∣ 2 = 1 2 ∣ Φ ∣ 2 \rho^2|D\Phi|^2=\frac{1}{2}|\Phi|^2 ρ 2 ∣ D Φ ∣ 2 = 2 1 ∣Φ ∣ 2 for the Dirac models in Definition 2.7. Adding the variation identities and applying Young’s inequality to ⟨ y ⋅ D y u , ρ ∂ ρ u − u / 2 ⟩ \langle y\cdot D_yu,\rho\partial_\rho u-u/2\rangle ⟨ y ⋅ D y u , ρ ∂ ρ u − u /2 ⟩ gives
1 2 ∫ ( ∣ D u ∣ 2 + ∣ D y u ∣ 2 ) ψ 2 d X + 1 2 ∫ R − 1 ∣ u ∣ 2 ψ ψ ′ d X ≤ − 2 ∫ Err ( u , Φ ) R − 1 ψ ψ ′ d X (30) \frac{1}{2}\int\left(|Du|^2+|D_yu|^2\right)\psi^2\,dX+\frac{1}{2}\int R^{-1}|u|^2\psi\psi'\,dX\leq-2\int\operatorname{Err}(u,\Phi)R^{-1}\psi\psi'\,dX \tag*{(30)} 2 1 ∫ ( ∣ D u ∣ 2 + ∣ D y u ∣ 2 ) ψ 2 d X + 2 1 ∫ R − 1 ∣ u ∣ 2 ψ ψ ′ d X ≤ − 2 ∫ Err ( u , Φ ) R − 1 ψ ψ ′ d X (30)
+ C ∫ ∣ ρ ∂ ρ u − u / 2 ∣ 2 ( ψ ′ ) 2 d X + C δ . {}+C\int|\rho\partial_\rho u-u/2|^2(\psi')^2\,dX+C\delta. + C ∫ ∣ ρ ∂ ρ u − u /2 ∣ 2 ( ψ ′ ) 2 d X + C δ .
Step 2: Tilt-excess along the spine. To estimate the right-hand side of (30) , we localise at the distance to the spine, as in the graphical description of triple junctions in [21 ] . For s > 0 s>0 s > 0 , y 0 ∈ R n − 2 y_0\in\mathbb{R}^{n-2} y 0 ∈ R n − 2 and 0 < κ ≤ 1 / 8 0<\kappa\leq1/8 0 < κ ≤ 1/8 , set
A s , κ ( y 0 ) = { ( z , y ) : ( ρ − s ) 2 + ∣ y − y 0 ∣ 2 < ( κ s ) 2 } , A_{s,\kappa}(y_0)=\{(z,y):( \rho-s)^2+|y-y_0|^2<(\kappa s)^2\}, A s , κ ( y 0 ) = {( z , y ) : ( ρ − s ) 2 + ∣ y − y 0 ∣ 2 < ( κ s ) 2 } ,
on which ρ \rho ρ is comparable to s s s . We sometimes suppress y 0 y_0 y 0 from the notation. Choose a cover of { 1 / 2 < R < 3 / 4 } ∖ P \{1/2<R<3/4\}\setminus P { 1/2 < R < 3/4 } ∖ P by A s i , 1 / 16 ( y i ) A_{s_i,1/16}(y_i) A s i , 1/16 ( y i ) , with the enlarged regions A s i , 1 / 8 ( y i ) A_{s_i,1/8}(y_i) A s i , 1/8 ( y i ) contained in B 1 B_1 B 1 and having bounded overlap. The main strategy is a covering argument: regions with small rescaled L 2 L^2 L 2 -excess admit matched sheets by Lemma 3.1 , while on the remaining regions the Caccioppoli estimate suffices.
Choose ε 0 > 0 \varepsilon_0>0 ε 0 > 0 small enough for the sheet-matching argument in Lemma 3.1 to apply after rescaling. If s − n + 1 ∫ A s , 1 / 8 G ( u , Φ ) 2 d X < ε 0 s^{-n+1}\int_{A_{s,1/8}}\mathcal{G}(u,\Phi)^2\,dX<\varepsilon_0 s − n + 1 ∫ A s , 1/8 G ( u , Φ ) 2 d X < ε 0 , write v = u − φ v=u-\varphi v = u − φ for the difference of the locally matched sheets. It satisfies a δ \delta δ -perturbed harmonic equation. Here ∣ φ ∣ 2 + s 2 ∣ D φ ∣ 2 ≤ C s |\varphi|^2+s^2|D\varphi|^2\leq Cs ∣ φ ∣ 2 + s 2 ∣ D φ ∣ 2 ≤ C s on A s , 1 / 8 A_{s,1/8} A s , 1/8 . Test with η 2 v \eta^2v η 2 v , where η \eta η depends only on ρ \rho ρ and y y y , equals one on A s , 1 / 16 A_{s,1/16} A s , 1/16 , is supported in A s , 3 / 32 A_{s,3/32} A s , 3/32 and satisfies ∣ D η ∣ ≤ C / s |D\eta|\leq C/s ∣ D η ∣ ≤ C / s . This gives
∫ A s , 1 / 16 ( ∣ v ∣ 2 + s 2 ∣ D v ∣ 2 ) d X ≤ C ∫ A s , 1 / 8 G ( u , Φ ) 2 d X + C δ 2 ∫ A s , 1 / 8 ( ∣ φ ∣ 2 + s 2 ∣ D φ ∣ 2 ) d X ≤ C ∫ A s , 1 / 8 G ( u , Φ ) 2 d X + C δ 2 s n + 1 . (31) \begin{aligned}
\int_{A_{s,1/16}}\left(|v|^2+s^2|Dv|^2\right)\,dX\leq C\int_{A_{s,1/8}}\mathcal{G}(u,\Phi)^2\,dX+C\delta^2\int_{A_{s,1/8}}\left(|\varphi|^2+s^2|D\varphi|^2\right)\,dX \\
&\leq C\int_{A_{s,1/8}}\mathcal{G}(u,\Phi)^2\,dX+C\delta^2s^{n+1}. \tag*{(31)}
\end{aligned} ∫ A s , 1/16 ( ∣ v ∣ 2 + s 2 ∣ D v ∣ 2 ) d X ≤ C ∫ A s , 1/8 G ( u , Φ ) 2 d X + C δ 2 ∫ A s , 1/8 ( ∣ φ ∣ 2 + s 2 ∣ D φ ∣ 2 ) d X ≤ C ∫ A s , 1/8 G ( u , Φ ) 2 d X + C δ 2 s n + 1 . (31)
Writing z = ρ e i θ z=\rho e^{i\theta} z = ρ e i θ and using ρ ∂ ρ φ = φ / 2 \rho\partial_\rho\varphi=\varphi/2 ρ ∂ ρ φ = φ /2 and D y φ = 0 D_y\varphi=0 D y φ = 0 , we have
Err ( u , Φ ) = ⟨ ∂ θ φ , ∂ θ v ⟩ − 1 4 ⟨ φ , v ⟩ + O ( s 2 ∣ D v ∣ 2 + ∣ v ∣ 2 ) , \operatorname{Err}(u,\Phi)=\langle\partial_\theta\varphi,\partial_\theta v\rangle-\frac{1}{4}\langle\varphi,v\rangle+O(s^2|Dv|^2+|v|^2), Err ( u , Φ ) = ⟨ ∂ θ φ , ∂ θ v ⟩ − 4 1 ⟨ φ , v ⟩ + O ( s 2 ∣ D v ∣ 2 + ∣ v ∣ 2 ) ,
ρ ∂ ρ u − u / 2 = ρ ∂ ρ v − v / 2. \rho\partial_\rho u-u/2=\rho\partial_\rho v-v/2. ρ ∂ ρ u − u /2 = ρ ∂ ρ v − v /2.
The two linear terms have zero angular integral:
∫ 0 2 π ( ⟨ ∂ θ φ , ∂ θ v ⟩ − 1 4 ⟨ φ , v ⟩ ) d θ = [ ⟨ ∂ θ φ , v ⟩ ] 0 2 π − ∫ 0 2 π ⟨ ∂ θ 2 φ + 1 4 φ , v ⟩ d θ = 0 , (32) \int_0^{2\pi}\left(\langle\partial_\theta\varphi,\partial_\theta v\rangle-\frac{1}{4}\langle\varphi,v\rangle\right)\,d\theta
=\left[\langle\partial_\theta\varphi,v\rangle\right]_0^{2\pi}-\int_0^{2\pi}\left\langle\partial_\theta^2\varphi+\frac{1}{4}\varphi,v\right\rangle\,d\theta=0, \tag*{(32)} ∫ 0 2 π ( ⟨ ∂ θ φ , ∂ θ v ⟩ − 4 1 ⟨ φ , v ⟩ ) d θ = [ ⟨ ∂ θ φ , v ⟩ ] 0 2 π − ∫ 0 2 π ⟨ ∂ θ 2 φ + 4 1 φ , v ⟩ d θ = 0 , (32)
since φ \varphi φ and v v v are 2 π 2\pi 2 π -antiperiodic and ∂ θ 2 φ = − φ / 4 \partial_\theta^2\varphi=-\varphi/4 ∂ θ 2 φ = − φ /4 . The cutoff weights are independent of θ \theta θ , so (31) controls both error terms after first integrating in θ \theta θ and then taking absolute values.
On the region not covered by such annuli, s n + 1 ≤ ε 0 − 1 ∫ A s , 1 / 8 G ( u , Φ ) 2 d X s^{n+1}\leq\varepsilon_0^{-1}\int_{A_{s,1/8}}\mathcal{G}(u,\Phi)^2\,dX s n + 1 ≤ ε 0 − 1 ∫ A s , 1/8 G ( u , Φ ) 2 d X . The Caccioppoli estimate (6) , with a cutoff in A s , 1 / 8 A_{s,1/8} A s , 1/8 , and ∣ u ∣ 2 ≤ C ( G ( u , Φ ) 2 + ∣ Φ ∣ 2 ) |u|^2\leq C(\mathcal{G}(u,\Phi)^2+|\Phi|^2) ∣ u ∣ 2 ≤ C ( G ( u , Φ ) 2 + ∣Φ ∣ 2 ) give
∫ A s , 1 / 16 ( ∣ u ∣ 2 + ∣ Φ ∣ 2 + s 2 ∣ D u ∣ 2 + s 2 ∣ D Φ ∣ 2 ) d X ≤ C ∫ A s , 1 / 8 G ( u , Φ ) 2 d X . (33) \int_{A_{s,1/16}}\left(|u|^2+|\Phi|^2+s^2|Du|^2+s^2|D\Phi|^2\right)\,dX\leq C\int_{A_{s,1/8}}\mathcal{G}(u,\Phi)^2\,dX. \tag*{(33)} ∫ A s , 1/16 ( ∣ u ∣ 2 + ∣Φ ∣ 2 + s 2 ∣ D u ∣ 2 + s 2 ∣ D Φ ∣ 2 ) d X ≤ C ∫ A s , 1/8 G ( u , Φ ) 2 d X . (33)
This controls the same error terms directly from the definition of $\operatorname{Err}(u,\Phi).
The weights R − 1 ψ ψ ′ R^{-1}\psi\psi' R − 1 ψ ψ ′ and ( ψ ′ ) 2 (\psi')^2 ( ψ ′ ) 2 are bounded. Since the enlarged regions have bounded overlap, the sum of their local excesses is at most C E CE C E . Their volumes are comparable to s i n s_i^n s i n , so
δ 2 ∑ i s i n + 1 ≤ C δ 2 ∑ i ∫ A s i , 1 / 8 ( y i ) ρ d X ≤ C δ 2 ∫ B 1 ρ d X ≤ C δ 2 . \delta^2\sum_i s_i^{n+1}\le C\delta^2\sum_i\int_{A_{s_i,1/8}(y_i)}\rho\,dX\le C\delta^2\int_{B_1}\rho\,dX\le C\delta^2. δ 2 i ∑ s i n + 1 ≤ C δ 2 i ∑ ∫ A s i , 1/8 ( y i ) ρ d X ≤ C δ 2 ∫ B 1 ρ d X ≤ C δ 2 .
Summing the nonnegative integrals obtained above therefore bounds the right-hand side of (30) by C ( E + δ ) C(E+\delta) C ( E + δ ) . Combining this with (29) and comparability of d X dX d X and d V g dV_g d V g , we obtain
∫ B 1 / 2 R 2 − n ∣ ∂ R ( u R 1 / 2 ) ∣ 2 d V g + ∫ B 1 / 2 ∣ D y u ∣ 2 d V g ≤ C ( E + δ ) . (34) \int_{B_{1/2}}R^{2-n}\left|\partial_R\left(\frac{u}{R^{1/2}}\right)\right|^2dV_g+\int_{B_{1/2}}|D_yu|^2dV_g\le C(E+\delta). \tag*{(34)} ∫ B 1/2 R 2 − n ∂ R ( R 1/2 u ) 2 d V g + ∫ B 1/2 ∣ D y u ∣ 2 d V g ≤ C ( E + δ ) . (34)
Step 3: Changing the centre. Fix p = ( ξ , ζ ) ∈ B u ∩ B 9 / 16 p=(\xi,\zeta)\in B_u\cap B_{9/16} p = ( ξ , ζ ) ∈ B u ∩ B 9/16 . Translation invariance along P P P and ∣ D Φ ∣ 2 ≤ C / ∣ z ∣ |D\Phi|^2\le C/|z| ∣ D Φ ∣ 2 ≤ C /∣ z ∣ give
G ( Φ ( X − p ) , Φ ( X ) ) 2 ≤ C ∣ ξ ∣ 2 ∫ 0 1 ∣ z − t ξ ∣ − 1 d t . (35) \mathcal{G}(\Phi(X-p),\Phi(X))^2\le C|\xi|^2\int_0^1|z-t\xi|^{-1}\,dt. \tag*{(35)} G ( Φ ( X − p ) , Φ ( X ) ) 2 ≤ C ∣ ξ ∣ 2 ∫ 0 1 ∣ z − t ξ ∣ − 1 d t . (35)
Thus ∫ B 1 G ( u ( X ) , Φ ( X − p ) ) 2 d V g ( X ) ≤ C ( E + ∣ ξ ∣ 2 ) \int_{B_1}\mathcal{G}(u(X),\Phi(X-p))^2dV_g(X)\le C(E+|\xi|^2) ∫ B 1 G ( u ( X ) , Φ ( X − p ) ) 2 d V g ( X ) ≤ C ( E + ∣ ξ ∣ 2 ) , with ∣ ξ ∣ → 0 |\xi|\to0 ∣ ξ ∣ → 0 as E + δ → 0 E+\delta\to0 E + δ → 0 by Lemma 3.1.
Choose geodesic coordinates Y Y Y centred at p p p whose coordinate basis at p p p is obtained by orthonormalising the original coordinate basis. Extend the original bundle frame at p p p by parallel transport along radial geodesics. The coordinate change has the form
X = p + ( I + O ( δ ) ) Y + O ( δ R p 2 ) , X=p+(I+O(\delta))Y+O(\delta R_p^2), X = p + ( I + O ( δ )) Y + O ( δ R p 2 ) ,
and the bundle frame change is I + O ( δ R p ) I+O(\delta R_p) I + O ( δ R p ) . Write Φ = { ± ( a z 1 / 2 + b z ˉ 1 / 2 ) } \Phi=\{\pm(az^{1/2}+b\bar z^{1/2})\} Φ = { ± ( a z 1/2 + b z ˉ 1/2 )} . Parallel transport the oriented normal frame and the coefficients a , b a,b a , b from 0 0 0 to p p p along the radial geodesic. This defines a cylindrical Dirac model Ψ \Psi Ψ at p p p , since parallel transport preserves the Clifford relations. Comparing in the same bundle frame, for corresponding coordinates X , Y X,Y X , Y with X ∉ p + P X\notin p+P X ∈ / p + P , we have
G ( Φ ( X − p ) , Ψ ( Y ) ) 2 ≤ C δ 2 ( R p 2 dist ( X , p + P ) + R p 3 ) ≤ C δ 2 R p 2 dist ( X , p + P ) . \mathcal{G}(\Phi(X-p),\Psi(Y))^2\le C\delta^2\left(\frac{R_p^2}{\operatorname{dist}(X,p+P)}+R_p^3\right)\le C\delta^2\frac{R_p^2}{\operatorname{dist}(X,p+P)}. G ( Φ ( X − p ) , Ψ ( Y ) ) 2 ≤ C δ 2 ( dist ( X , p + P ) R p 2 + R p 3 ) ≤ C δ 2 dist ( X , p + P ) R p 2 .
The O ( δ 2 R p 2 / dist ( X , p + P ) ) O(\delta^2R_p^2/\operatorname{dist}(X,p+P)) O ( δ 2 R p 2 / dist ( X , p + P )) error comes from the O ( δ R p ) O(\delta R_p) O ( δ R p ) change in the variable z z z in the model formula (22) ; the O ( δ 2 R p 3 ) O(\delta^2R_p^3) O ( δ 2 R p 3 ) error comes from the bundle frame change. Integrating in polar coordinates for X − p X-p X − p gives
∫ B 3 / 8 G ( Φ ( X − p ) , Ψ ( Y ) ) 2 R p ( X ) n + 1 − σ d Y ≤ C δ 2 ∫ 0 1 t σ − 1 d t ≤ C σ δ 2 . (36) \int_{B_{3/8}}\frac{\mathcal{G}(\Phi(X-p),\Psi(Y))^2}{R_p(X)^{n+1-\sigma}}\,dY\le C\delta^2\int_0^1t^{\sigma-1}\,dt\le C_\sigma\delta^2. \tag*{(36)} ∫ B 3/8 R p ( X ) n + 1 − σ G ( Φ ( X − p ) , Ψ ( Y ) ) 2 d Y ≤ C δ 2 ∫ 0 1 t σ − 1 d t ≤ C σ δ 2 . (36)
It follows that ∫ B 3 / 8 G ( u , Ψ ) 2 d Y ≤ C ( E + ∣ ξ ∣ 2 + δ 2 ) \int_{B_{3/8}}\mathcal{G}(u,\Psi)^2\,dY\le C(E+|\xi|^2+\delta^2) ∫ B 3/8 G ( u , Ψ ) 2 d Y ≤ C ( E + ∣ ξ ∣ 2 + δ 2 ) . Applying Steps 1 and 2 with a cutoff supported in B 3 / 8 B_{3/8} B 3/8 and equal to one on B 5 / 16 B_{5/16} B 5/16 gives
∫ B 5 / 16 R p 2 − n ∣ ∂ R p ( u R p 1 / 2 ) ∣ 2 d Y ≤ C ( E + ∣ ξ ∣ 2 + δ ) . (37) \int_{B_{5/16}}R_p^{2-n}\left|\partial_{R_p}\left(\frac{u}{R_p^{1/2}}\right)\right|^2\,dY\le C(E+|\xi|^2+\delta). \tag*{(37)} ∫ B 5/16 R p 2 − n ∂ R p ( R p 1/2 u ) 2 d Y ≤ C ( E + ∣ ξ ∣ 2 + δ ) . (37)
Choose a radial η ∈ C ∞ ( B 5 / 16 ) \eta\in C^\infty(B_{5/16}) η ∈ C ∞ ( B 5/16 ) with 0 ≤ η ≤ 1 0\le\eta\le1 0 ≤ η ≤ 1 , η = 1 \eta=1 η = 1 on B 9 / 32 B_{9/32} B 9/32 , and ∣ D η ∣ ≤ C |D\eta|\le C ∣ D η ∣ ≤ C . Integration by parts, with nonpositive inner boundary term, Young’s inequality and the homogeneity of Ψ \Psi Ψ give
∫ B 9 / 32 G ( u , Ψ ) 2 R p n + 1 − σ d Y ≤ C σ ∫ B 5 / 16 R p 2 − n + σ ∣ ∂ R p ( η G ( u , Ψ ) R p 1 / 2 ) ∣ 2 d Y ≤ C σ ∫ B 5 / 16 ( R p 2 − n ∣ ∂ R p ( u R p 1 / 2 ) ∣ 2 + G ( u , Ψ ) 2 ) d Y ≤ C σ ( E + ∣ ξ ∣ 2 + δ ) . \begin{aligned}
\int_{B_{9/32}}\frac{\mathcal{G}(u,\Psi)^2}{R_p^{n+1-\sigma}}\,dY\le C_\sigma\int_{B_{5/16}}R_p^{2-n+\sigma}\left|\partial_{R_p}\left(\frac{\eta\mathcal{G}(u,\Psi)}{R_p^{1/2}}\right)\right|^2\,dY \\
&\le C_\sigma\int_{B_{5/16}}\left(R_p^{2-n}\left|\partial_{R_p}\left(\frac{u}{R_p^{1/2}}\right)\right|^2+\mathcal{G}(u,\Psi)^2\right)\,dY \\
&\le C_\sigma(E+|\xi|^2+\delta).
\end{aligned} ∫ B 9/32 R p n + 1 − σ G ( u , Ψ ) 2 d Y ≤ C σ ∫ B 5/16 R p 2 − n + σ ∂ R p ( R p 1/2 η G ( u , Ψ ) ) 2 d Y ≤ C σ ∫ B 5/16 R p 2 − n ∂ R p ( R p 1/2 u ) 2 + G ( u , Ψ ) 2 d Y ≤ C σ ( E + ∣ ξ ∣ 2 + δ ) .
Since B 1 / 4 ( p ) B_{1/4}(p) B 1/4 ( p ) lies inside B 9 / 32 B_{9/32} B 9/32 in the new coordinates, (3.12) and volume comparability give, in the original chart,
∫ B 1 / 4 ( p ) G ( u ( X ) , Φ ( X − p ) ) 2 R p ( X ) n + 1 − σ d V g ( X ) ≤ C σ ( E + ∣ ξ ∣ 2 + δ ) . (38) \int_{B_{1/4}(p)} \frac{\mathcal{G}(u(X),\Phi(X-p))^2}{R_p(X)^{n+1-\sigma}}\,dV_g(X) \leq C_\sigma(E+|\xi|^2+\delta). \tag*{(38)} ∫ B 1/4 ( p ) R p ( X ) n + 1 − σ G ( u ( X ) , Φ ( X − p ) ) 2 d V g ( X ) ≤ C σ ( E + ∣ ξ ∣ 2 + δ ) . (38)
Step 4: Estimating the displacement ∣ ξ ∣ = dist ( p , R n − 2 ) |\xi|=\operatorname{dist}(p,\mathbb{R}^{n-2}) ∣ ξ ∣ = dist ( p , R n − 2 ) . By the explicit formula (2.18) for Φ \Phi Φ and uniform comparability of geodesic and coordinate balls, if ∣ ξ ∣ ≤ s ≪ 1 8 |\xi|\leq s\ll\frac{1}{8} ∣ ξ ∣ ≤ s ≪ 8 1 ,
∫ B s ( p ) G ( Φ ( X ) , Φ ( X − p ) ) 2 d V g ( X ) ≳ ∫ B s ( p ) ∣ ξ ∣ 2 s d V g ( X ) ≳ s n − 1 ∣ ξ ∣ 2 . (39) \int_{B_s(p)} \mathcal{G}(\Phi(X),\Phi(X-p))^2\,dV_g(X) \gtrsim\int_{B_s(p)} \frac{|\xi|^2}{s}\,dV_g(X) \gtrsim s^{n-1}|\xi|^2. \tag*{(39)} ∫ B s ( p ) G ( Φ ( X ) , Φ ( X − p ) ) 2 d V g ( X ) ≳ ∫ B s ( p ) s ∣ ξ ∣ 2 d V g ( X ) ≳ s n − 1 ∣ ξ ∣ 2 . (39)
Taking σ = 1 / 2 \sigma=1/2 σ = 1/2 in (3.14), we have
c s n − 1 ∣ ξ ∣ 2 ≲ ∫ B s ( p ) ( G ( u ( X ) , Φ ( X ) ) 2 + G ( u ( X ) , Φ ( X − p ) ) 2 ) d V g ( X ) cs^{n-1}|\xi|^2 \lesssim\int_{B_s(p)} \left(\mathcal{G}(u(X),\Phi(X))^2+\mathcal{G}(u(X),\Phi(X-p))^2\right)\,dV_g(X) c s n − 1 ∣ ξ ∣ 2 ≲ ∫ B s ( p ) ( G ( u ( X ) , Φ ( X ) ) 2 + G ( u ( X ) , Φ ( X − p ) ) 2 ) d V g ( X )
≲ E + s n + 1 / 2 ( E + ∣ ξ ∣ 2 + δ ) . \lesssim E+s^{n+1/2}(E+|\xi|^2+\delta). ≲ E + s n + 1/2 ( E + ∣ ξ ∣ 2 + δ ) .
Thus ∣ ξ ∣ 2 ≤ C ( E + δ ) |\xi|^2\leq C(E+\delta) ∣ ξ ∣ 2 ≤ C ( E + δ ) . Together with (3.10) and (3.14), this proves (3.2) and bounds the first integral in (3.3).
For the second integral in (3.3), the triangle inequality reduces the remaining estimate to
∫ B 1 / 2 G ( Φ ( X − p ) , Φ ( X ) ) 2 R p ( X ) n − 1 − σ d V g ( X ) . \int_{B_{1/2}} \frac{\mathcal{G}(\Phi(X-p),\Phi(X))^2}{R_p(X)^{n-1-\sigma}}\,dV_g(X). ∫ B 1/2 R p ( X ) n − 1 − σ G ( Φ ( X − p ) , Φ ( X ) ) 2 d V g ( X ) .
By (3.11) and comparability of d V g dV_g d V g and d X dX d X , it suffices to show
∫ B 1 / 2 d X ∣ z − t ξ ∣ R p n − 1 − σ ≤ C σ , ∀ t ∈ [ 0 , 1 ] . \int_{B_{1/2}} \frac{dX}{|z-t\xi|R_p^{n-1-\sigma}}\leq C_\sigma,\qquad\forall t\in[0,1]. ∫ B 1/2 ∣ z − t ξ ∣ R p n − 1 − σ d X ≤ C σ , ∀ t ∈ [ 0 , 1 ] .
In fact, when n ≥ 3 n\geq3 n ≥ 3 , using R p ≳ ∣ z − ξ ∣ , ∣ y − ζ ∣ R_p\gtrsim|z-\xi|,|y-\zeta| R p ≳ ∣ z − ξ ∣ , ∣ y − ζ ∣ ,
∫ B 1 / 2 d X ∣ z − t ξ ∣ R p n − 1 − σ ≲ ( ∫ ∣ z ∣ ≲ 1 d z ∣ z − t ξ ∣ ∣ z − ξ ∣ 1 − σ / 2 ) ( ∫ ∣ y − ζ ∣ ≲ 1 d y ∣ y − ζ ∣ n − 2 − σ / 2 ) < ∞ . \int_{B_{1/2}} \frac{dX}{|z-t\xi|R_p^{n-1-\sigma}}\lesssim\left(\int_{|z|\lesssim1}\frac{dz}{|z-t\xi||z-\xi|^{1-\sigma/2}}\right)\left(\int_{|y-\zeta|\lesssim1}\frac{dy}{|y-\zeta|^{n-2-\sigma/2}}\right)<\infty. ∫ B 1/2 ∣ z − t ξ ∣ R p n − 1 − σ d X ≲ ( ∫ ∣ z ∣ ≲ 1 ∣ z − t ξ ∣∣ z − ξ ∣ 1 − σ /2 d z ) ( ∫ ∣ y − ζ ∣ ≲ 1 ∣ y − ζ ∣ n − 2 − σ /2 d y ) < ∞.
When n = 2 n=2 n = 2 , it directly follows from ∣ z ∣ − 2 + σ ∈ L l o c 1 |z|^{-2+\sigma}\in L^1_{\mathrm{loc}} ∣ z ∣ − 2 + σ ∈ L loc 1 .
Corollary 3.5 (L 2 L^2 L 2 non-concentration near the spine). For each sufficiently small τ > 0 \tau>0 τ > 0 , if the height excess satisfies E + δ < ε ( τ ) E+\delta<\varepsilon(\tau) E + δ < ε ( τ ) , we have
∫ B 1 / 2 G ( u , Φ ) 2 max { ρ , τ } 1 − σ ≤ C σ ( E + δ ) . (40) \int_{B_{1/2}} \frac{\mathcal{G}(u,\Phi)^2}{\max\{\rho,\tau\}^{1-\sigma}} \leq C_\sigma(E+\delta). \tag*{(40)} ∫ B 1/2 max { ρ , τ } 1 − σ G ( u , Φ ) 2 ≤ C σ ( E + δ ) . (40) The constant C σ C_\sigma C σ is independent of τ \tau τ .
Proof. At the prescribed scale τ \tau τ , by Lemma 3.2 we know that for every y ∈ P ∩ B 1 / 2 y\in P\cap B_{1/2} y ∈ P ∩ B 1/2 , the normal slice contains a branch point p y ∈ B u p_y\in\mathcal{B}_u p y ∈ B u with ∣ p y − y ∣ < τ / 4 |p_y-y|<\tau/4 ∣ p y − y ∣ < τ /4 . For τ ≤ s ≤ 1 4 \tau\leq s\leq\frac{1}{4} τ ≤ s ≤ 4 1 , applying (3.3) with exponent σ / 2 \sigma/2 σ /2 , we find
∫ B s ( y ) ∩ B 1 / 2 G ( u , Φ ) 2 ≤ C σ s n − 1 − σ / 2 ( E + δ ) . \int_{B_s(y)\cap B_{1/2}}\mathcal{G}(u,\Phi)^2\leq C_\sigma s^{n-1-\sigma/2}(E+\delta). ∫ B s ( y ) ∩ B 1/2 G ( u , Φ ) 2 ≤ C σ s n − 1 − σ /2 ( E + δ ) .
Covering the s / 2 s/2 s /2 -tube of the spine by at most C s − ( n − 2 ) Cs^{-(n-2)} C s − ( n − 2 ) such balls, we get M ( s ) ≤ C σ s 1 − σ / 2 ( E + δ ) M(s)\leq C_\sigma s^{1-\sigma/2}(E+\delta) M ( s ) ≤ C σ s 1 − σ /2 ( E + δ ) for all s ≥ τ s\geq\tau s ≥ τ , where M ( t ) : = ∫ B 1 / 2 ∩ { ρ < t } G ( u , Φ ) 2 M(t):=\int_{B_{1/2}\cap\{\rho<t\}}\mathcal{G}(u,\Phi)^2 M ( t ) := ∫ B 1/2 ∩ { ρ < t } G ( u , Φ ) 2 .
Thus,
∫ B 1 / 2 G ( u , Φ ) 2 max { ρ , τ } 1 − σ ≲ C σ τ − 1 + σ M ( τ ) + C σ ∫ τ 1 / 4 s − 2 + σ M ( s ) d s + C E ≲ C σ ( E + δ ) . \int_{B_{1/2}}\frac{\mathcal{G}(u,\Phi)^2}{\max\{\rho,\tau\}^{1-\sigma}}\lesssim C_\sigma\tau^{-1+\sigma}M(\tau)+C_\sigma\int_\tau^{1/4}s^{-2+\sigma}M(s)\,ds+CE\lesssim C_\sigma(E+\delta). ∫ B 1/2 max { ρ , τ } 1 − σ G ( u , Φ ) 2 ≲ C σ τ − 1 + σ M ( τ ) + C σ ∫ τ 1/4 s − 2 + σ M ( s ) d s + C E ≲ C σ ( E + δ ) .
This completes the proof.
Harmonic blow-ups
Let ( u j , Φ j ) (u_j,\Phi_j) ( u j , Φ j ) be as above, with their model spines identified with P P P . Assume Φ j → Φ 0 = { ± φ } \Phi_j \to\Phi_0=\{\pm\varphi\} Φ j → Φ 0 = { ± φ } , E j : = E ( u j , Φ j ; 0 , 1 ) → 0 E_j:=\mathcal{E}(u_j,\Phi_j;0,1)\to0 E j := E ( u j , Φ j ; 0 , 1 ) → 0 and δ j / E j → 0 \delta_j/E_j\to0 δ j / E j → 0 . We study the blow-up limit of the renormalised differences
v j = u j − φ j E j on B 1 / 2 ∩ { ρ > τ j } . v_j=\frac{u_j-\varphi_j}{\sqrt{E_j}}\quad\text{on }B_{1/2}\cap\{\rho>\tau_j\}. v j = E j u j − φ j on B 1/2 ∩ { ρ > τ j } .
Here τ j ↓ 0 \tau_j\downarrow0 τ j ↓ 0 is chosen so that the sheets of u j u_j u j and Φ j \Phi_j Φ j are uniquely matched on this region by Lemma 3.1 . For λ = λ 1 + i λ 2 \lambda=\lambda_1+i\lambda_2 λ = λ 1 + i λ 2 , write λ ⋅ D z = λ 1 ∂ x 1 + λ 2 ∂ x 2 \lambda\cdot D_z=\lambda_1\partial_{x_1}+\lambda_2\partial_{x_2} λ ⋅ D z = λ 1 ∂ x 1 + λ 2 ∂ x 2 .
Proposition 3.6. After taking a subsequence, v j → w v_j\to w v j → w strongly in L 2 ( B 1 / 2 ) L^2(B_{1/2}) L 2 ( B 1/2 ) . It is smooth and satisfies D 0 w = 0 D_0w=0 D 0 w = 0 on B 1 / 2 ∖ P B_{1/2}\setminus P B 1/2 ∖ P . Moreover,
∫ B 1 / 4 R 2 − n ∣ ∂ R ( w / R 1 / 2 ) ∣ 2 ≤ C . (41) \int_{B_{1/4}} R^{2-n}\left|\partial_R\left(w/R^{1/2}\right)\right|^2\le C. \tag*{(41)} ∫ B 1/4 R 2 − n ∂ R ( w / R 1/2 ) 2 ≤ C . (41) For each ∣ y 0 ∣ < 1 / 8 |y_0|<1/8 ∣ y 0 ∣ < 1/8 , there is λ ( y 0 ) ∈ C \lambda(y_0)\in\mathbb{C} λ ( y 0 ) ∈ C with ∣ λ ( y 0 ) ∣ ≤ C |\lambda(y_0)|\le C ∣ λ ( y 0 ) ∣ ≤ C such that for every 0 < σ < 1 0<\sigma<1 0 < σ < 1 ,
∫ B 1 / 8 ( 0 , y 0 ) ∣ w − λ ( y 0 ) ⋅ D z φ ∣ 2 ∣ X − ( 0 , y 0 ) ∣ n + 1 − σ ≤ C σ . (42) \int_{B_{1/8}(0,y_0)}\frac{\left|w-\lambda(y_0)\cdot D_z\varphi\right|^2}{|X-(0,y_0)|^{n+1-\sigma}}\le C_\sigma. \tag*{(42)} ∫ B 1/8 ( 0 , y 0 ) ∣ X − ( 0 , y 0 ) ∣ n + 1 − σ ∣ w − λ ( y 0 ) ⋅ D z φ ∣ 2 ≤ C σ . (42)
Proof. We first prove the strong L 2 L^2 L 2 -convergence and then derive the limiting equation for w w w . Fix 0 < τ < 1 / 8 0<\tau<1/8 0 < τ < 1/8 and 0 < σ < 1 0<\sigma<1 0 < σ < 1 . Summing (31) away from P P P and applying Corollary 3.5 give, for all sufficiently large j j j ,
∫ B 1 / 2 ∩ { ρ > τ } ∣ D v j ∣ 2 ≤ C τ − 2 ( 1 + δ j / E j ) , \int_{B_{1/2}\cap\{\rho>\tau\}}|Dv_j|^2\le C\tau^{-2}(1+\delta_j/E_j), ∫ B 1/2 ∩ { ρ > τ } ∣ D v j ∣ 2 ≤ C τ − 2 ( 1 + δ j / E j ) ,
1 E j ∫ B 1 / 2 ∩ { ρ < τ } G ( u j , Φ j ) 2 ≤ C σ τ 1 − σ ( 1 + δ j / E j ) . \frac{1}{E_j}\int_{B_{1/2}\cap\{\rho<\tau\}}\mathcal{G}(u_j,\Phi_j)^2\le C_\sigma\tau^{1-\sigma}(1+\delta_j/E_j). E j 1 ∫ B 1/2 ∩ { ρ < τ } G ( u j , Φ j ) 2 ≤ C σ τ 1 − σ ( 1 + δ j / E j ) .
Rellich compactness away from P P P and the second estimate give the claimed convergence by first letting j → ∞ j\to\infty j → ∞ and then τ ↓ 0 \tau\downarrow0 τ ↓ 0 , with ∫ B 1 / 2 ∣ w ∣ 2 ≤ 1 \int_{B_{1/2}}|w|^2\le1 ∫ B 1/2 ∣ w ∣ 2 ≤ 1 . The source terms in the Dirac equations for v j v_j v j are bounded on compact subsets away from P P P by C δ j / E j → 0 C\delta_j/\sqrt{E_j}\to0 C δ j / E j → 0 . Passing to the limit and applying interior elliptic regularity gives a smooth solution of D 0 w = 0 D_0w=0 D 0 w = 0 on B 1 / 2 ∖ P B_{1/2}\setminus P B 1/2 ∖ P .
Now we pass the L 2 L^2 L 2 -estimates (cf. Proposition 3.3) to establish the one for w w w . The 1 / 2 1/2 1/2 -homogeneity of φ j \varphi_j φ j gives ∂ R ( v j / R 1 / 2 ) = E j − 1 / 2 ∂ R ( u j / R 1 / 2 ) \partial_R(v_j/R^{1/2})=E_j^{-1/2}\partial_R(u_j/R^{1/2}) ∂ R ( v j / R 1/2 ) = E j − 1/2 ∂ R ( u j / R 1/2 ) . Thus (41) follows from (26) by weak lower semicontinuity away from P P P , followed by an exhaustion argument.
Fix ∣ y 0 ∣ < 1 / 8 |y_0|<1/8 ∣ y 0 ∣ < 1/8 . The no-hole principle gives branch points p j = ( ξ j , y 0 ) p_j=(\xi_j,y_0) p j = ( ξ j , y 0 ) approaching ( 0 , y 0 ) (0,y_0) ( 0 , y 0 ) , and (26) gives ∣ ξ j ∣ 2 ≤ C ( E j + δ j ) |\xi_j|^2\le C(E_j+\delta_j) ∣ ξ j ∣ 2 ≤ C ( E j + δ j ) . For this fixed y 0 y_0 y 0 , pass to a further subsequence so that
− ξ j E j → λ ( y 0 ) , u j − φ j ( ⋅ − p j ) E j → w − λ ( y 0 ) ⋅ D z φ -\frac{\xi_j}{\sqrt{E_j}}\to\lambda(y_0),\qquad\frac{u_j-\varphi_j(\cdot-p_j)}{\sqrt{E_j}}\to w-\lambda(y_0)\cdot D_z\varphi − E j ξ j → λ ( y 0 ) , E j u j − φ j ( ⋅ − p j ) → w − λ ( y 0 ) ⋅ D z φ
locally away from P P P , with ∣ λ ( y 0 ) ∣ ≤ C |\lambda(y_0)|\le C ∣ λ ( y 0 ) ∣ ≤ C . Divide the bound for the first integral in (27) by E j E_j E j and apply Fatou’s lemma on B 1 / 8 ( 0 , y 0 ) B_{1/8}(0,y_0) B 1/8 ( 0 , y 0 ) to obtain (42) . □
Excess decay and C1-alpha-regularity
In this section, we prove the excess decay via Campanato iteration and finally achieve the C 1 , α C^{1,\alpha} C 1 , α regularity of the frequency 1 2 \frac{1}{2} 2 1 branching set for the given Z 2 \mathbb{Z}_2 Z 2 -harmonic spinor.
The leading term of the linear error
Lemma 4.1 (Excess decay of the blow-up). Let w w w be a limit in Proposition 3.6 , with comparison models converging to Φ 0 \Phi_0 Φ 0 . Then there exists ℓ ∈ T Φ 0 C D \ell\in T_{\Phi_0}\mathcal{C}_D ℓ ∈ T Φ 0 C D such that
∥ ℓ ∥ L 2 ( B 1 / 2 ) 2 + s − n − 3 ∫ B s ∣ w − ℓ ∣ 2 ≤ C ∫ B 1 / 2 ∣ w ∣ 2 , ∀ 0 < s < 1 32 . (43) \|\ell\|_{L^2(B_{1/2})}^2+s^{-n-3}\int_{B_s}|w-\ell|^2\le C\int_{B_{1/2}}|w|^2,\qquad\forall0<s<\frac{1}{32}. \tag*{(43)} ∥ ℓ ∥ L 2 ( B 1/2 ) 2 + s − n − 3 ∫ B s ∣ w − ℓ ∣ 2 ≤ C ∫ B 1/2 ∣ w ∣ 2 , ∀0 < s < 32 1 . (43)
Proof. We fix σ = 1 / 2 \sigma= 1/2 σ = 1/2 in Proposition 3.6 . In polar coordinates ( ρ , θ ) (\rho,\theta) ( ρ , θ ) in the normal direction, w w w is 4 π 4\pi 4 π -periodic in θ \theta θ , and we decompose it as
w = ∑ ν w ν = ∑ ν ( a ν ( ρ , y ) cos ( ν θ ) + b ν ( ρ , y ) sin ( ν θ ) ) , ν = 1 2 , 3 2 , ⋯ . w=\sum_{\nu}w_\nu=\sum_{\nu}\left(a_\nu(\rho,y)\cos(\nu\theta)+b_\nu(\rho,y)\sin(\nu\theta)\right),\qquad\nu=\frac{1}{2},\frac{3}{2},\cdots. w = ν ∑ w ν = ν ∑ ( a ν ( ρ , y ) cos ( ν θ ) + b ν ( ρ , y ) sin ( ν θ ) ) , ν = 2 1 , 2 3 , ⋯ .
Each coefficient f ν = a ν f_\nu=a_\nu f ν = a ν or b ν b_\nu b ν solves
∂ ρ 2 f ν + ρ − 1 ∂ ρ f ν − ν 2 ρ − 2 f ν + Δ y f ν = 0. (44) \partial_\rho^2f_\nu+\rho^{-1}\partial_\rho f_\nu-\nu^2\rho^{-2}f_\nu+\Delta_yf_\nu=0. \tag*{(44)} ∂ ρ 2 f ν + ρ − 1 ∂ ρ f ν − ν 2 ρ − 2 f ν + Δ y f ν = 0. (44)
Case 1: The frequency ν = 1 / 2 \nu=1/2 ν = 1/2 . Write g ( ρ , y ) : = ρ 1 / 2 w 1 / 2 g(\rho,y):=\rho^{1/2}w_{1/2} g ( ρ , y ) := ρ 1/2 w 1/2 , which takes values in the real vector space V = { a cos ( θ / 2 ) + b sin ( θ / 2 ) : a , b ∈ E 0 } \mathcal{V}=\{a\cos(\theta/2)+b\sin(\theta/2):a,b\in E_0\} V = { a cos ( θ /2 ) + b sin ( θ /2 ) : a , b ∈ E 0 } with the real L 2 ( 0 , 2 π ) L^2(0,2\pi) L 2 ( 0 , 2 π ) -inner product. Equation (44) gives ( ∂ ρ 2 + Δ y ) g = 0 (\partial_\rho^2+\Delta_y)g=0 ( ∂ ρ 2 + Δ y ) g = 0 . Set t i ( θ ) = ρ 1 / 2 ∂ i φ t_i(\theta)=\rho^{1/2}\partial_i\varphi t i ( θ ) = ρ 1/2 ∂ i φ for i = 1 , 2 i=1,2 i = 1 , 2 ; these are independent of ρ \rho ρ . Decompose g = g ∥ + g ⊥ g=g_\parallel+g_\perp g = g ∥ + g ⊥ relative to T = span R { t 1 , t 2 } \mathcal{T}=\operatorname{span}_{\mathbb{R}}\{t_1,t_2\} T = span R { t 1 , t 2 } and its orthogonal complement in V \mathcal{V} V . We write B + B^+ B + for the half-ball in the variables ( ρ , y ) (\rho,y) ( ρ , y ) .
Since ρ 1 / 2 λ ( y 0 ) ⋅ D z φ ∈ T \rho^{1/2}\lambda(y_0)\cdot D_z\varphi\in\mathcal{T} ρ 1/2 λ ( y 0 ) ⋅ D z φ ∈ T , the weighted estimate (42) gives, for ∣ y 0 ∣ < 1 / 8 |y_0|<1/8 ∣ y 0 ∣ < 1/8 and 0 < s < 1 / 16 0<s<1/16 0 < s < 1/16 ,
∫ B s + ( 0 , y 0 ) ∥ g ⊥ ∥ V 2 ρ d ρ d y ≤ C s n + 1 − σ . \int_{B_s^+(0,y_0)}\lVert g_\perp\rVert_{\mathcal{V}}^2\,\rho\,d\rho\,dy\leq Cs^{n+1-\sigma}. ∫ B s + ( 0 , y 0 ) ∥ g ⊥ ∥ V 2 ρ d ρ d y ≤ C s n + 1 − σ .
The mean-value inequality in a ball of radius ρ / 2 \rho/2 ρ /2 , contained in B ρ / 2 + ( 0 , y ) B_{\rho/2}^+(0,y) B ρ /2 + ( 0 , y ) , now implies ∣ g ⊥ ( ρ , y ) ∣ ≤ C ρ 1 − σ / 2 |g_\perp(\rho,y)|\leq C\rho^{1-\sigma/2} ∣ g ⊥ ( ρ , y ) ∣ ≤ C ρ 1 − σ /2 for 0 < ρ < 1 / 32 0<\rho<1/32 0 < ρ < 1/32 and ∣ y ∣ < 1 / 8 |y|<1/8 ∣ y ∣ < 1/8 . Thus g ⊥ g_\perp g ⊥ has zero continuous boundary values and its odd reflection is harmonic. Interior estimates for this reflection give
∥ g ⊥ ∥ C k ( B 1 / 16 + ) ≤ C k ∥ g ⊥ ∥ L 2 ( B 1 / 8 + ) ≤ C k ∥ w ∥ L 2 ( B 1 / 2 ) , k ≥ 0. \lVert g_\perp\rVert_{C^k(B_{1/16}^+)}\leq C_k\lVert g_\perp\rVert_{L^2(B_{1/8}^+)}\leq C_k\lVert w\rVert_{L^2(B_{1/2})},\qquad k\geq0. ∥ g ⊥ ∥ C k ( B 1/16 + ) ≤ C k ∥ g ⊥ ∥ L 2 ( B 1/8 + ) ≤ C k ∥ w ∥ L 2 ( B 1/2 ) , k ≥ 0.
To control g ∥ g_\parallel g ∥ , use the Dirac equation. With J = c ( e 1 ) c ( e 2 ) J=c(e_1)c(e_2) J = c ( e 1 ) c ( e 2 ) , its polar form is
∂ ρ w = ρ − 1 J ∂ θ w + ∑ α = 1 n − 2 c ( ∂ ρ ) c ( e α + 2 ) ∂ y α w . \partial_\rho w=\rho^{-1}J\partial_\theta w+\sum_{\alpha=1}^{n-2}c(\partial_\rho)c(e_{\alpha+2})\partial_{y_\alpha}w. ∂ ρ w = ρ − 1 J ∂ θ w + α = 1 ∑ n − 2 c ( ∂ ρ ) c ( e α + 2 ) ∂ y α w .
The model Dirac equation gives t 2 = − J t 1 t_2=-Jt_1 t 2 = − J t 1 and J ∂ θ t i = − t i / 2 J\partial_\theta t_i=-t_i/2 J ∂ θ t i = − t i /2 . The operator c ( ∂ ρ ) c ( e α + 2 ) c(\partial_\rho)c(e_{\alpha+2}) c ( ∂ ρ ) c ( e α + 2 ) is skew-adjoint and anticommutes with J J J , so ⟨ c ( ∂ ρ ) c ( e α + 2 ) t j , t i ⟩ = 0 \langle c(\partial_\rho)c(e_{\alpha+2})t_j,t_i\rangle=0 ⟨ c ( ∂ ρ ) c ( e α + 2 ) t j , t i ⟩ = 0 for i , j = 1 , 2 i,j=1,2 i , j = 1 , 2 . Moreover, c ( ∂ ρ ) c ( e α + 2 ) t i c(\partial_\rho)c(e_{\alpha+2})t_i c ( ∂ ρ ) c ( e α + 2 ) t i has only angular frequencies ± 1 / 2 \pm1/2 ± 1/2 , by the explicit form (22) of Φ \Phi Φ . Since J ∗ = − J J^*=-J J ∗ = − J and both w w w and t i t_i t i change sign after one turn, integration by parts gives
ρ − 1 / 2 ∫ 0 2 π ⟨ J ∂ θ w , t i ⟩ d θ = ρ − 1 / 2 ∫ 0 2 π ⟨ w , J ∂ θ t i ⟩ d θ = − 1 2 ρ ⟨ g , t i ⟩ V . \rho^{-1/2}\int_0^{2\pi}\langle J\partial_\theta w,t_i\rangle\,d\theta=\rho^{-1/2}\int_0^{2\pi}\langle w,J\partial_\theta t_i\rangle\,d\theta=-\frac{1}{2\rho}\langle g,t_i\rangle_{\mathcal{V}}. ρ − 1/2 ∫ 0 2 π ⟨ J ∂ θ w , t i ⟩ d θ = ρ − 1/2 ∫ 0 2 π ⟨ w , J ∂ θ t i ⟩ d θ = − 2 ρ 1 ⟨ g , t i ⟩ V .
This cancels the term 1 2 ρ ⟨ g , t i ⟩ V \frac{1}{2\rho}\langle g,t_i\rangle_{\mathcal{V}} 2 ρ 1 ⟨ g , t i ⟩ V arising when we differentiate the factor ρ 1 / 2 \rho^{1/2} ρ 1/2 in g g g . Hence
∂ ρ ⟨ g , t i ⟩ V = ∑ α = 1 n − 2 ∂ y α ∫ 0 2 π ⟨ c ( ∂ ρ ) c ( e α + 2 ) g ⊥ , t i ⟩ d θ . \partial_\rho\langle g,t_i\rangle_{\mathcal{V}}=\sum_{\alpha=1}^{n-2}\partial_{y_\alpha}\int_0^{2\pi}\langle c(\partial_\rho)c(e_{\alpha+2})g_\perp,t_i\rangle\,d\theta. ∂ ρ ⟨ g , t i ⟩ V = α = 1 ∑ n − 2 ∂ y α ∫ 0 2 π ⟨ c ( ∂ ρ ) c ( e α + 2 ) g ⊥ , t i ⟩ d θ .
The right-hand side extends smoothly to ρ = 0 \rho=0 ρ = 0 and vanishes there, since g ⊥ g_\perp g ⊥ has zero boundary values. As t 1 , t 2 t_1,t_2 t 1 , t 2 form a basis of T \mathcal{T} T , integration from a fixed positive value of ρ \rho ρ shows that g ∥ g_\parallel g ∥ extends smoothly with ∂ ρ g ∥ ( 0 , y ) = 0 \partial_\rho g_\parallel(0,y)=0 ∂ ρ g ∥ ( 0 , y ) = 0 . Its even reflection is therefore harmonic. Interior estimates give
∥ g ∥ C 2 ( B 1 / 32 + ) ≤ C ∥ w ∥ L 2 ( B 1 / 2 ) , g ( 0 , y ) = g ∥ ( 0 , y ) ∈ T . \lVert g\rVert_{C^2(B_{1/32}^+)}\leq C\lVert w\rVert_{L^2(B_{1/2})},\qquad g(0,y)=g_\parallel(0,y)\in\mathcal{T}. ∥ g ∥ C 2 ( B 1/32 + ) ≤ C ∥ w ∥ L 2 ( B 1/2 ) , g ( 0 , y ) = g ∥ ( 0 , y ) ∈ T .
The radial estimate (41) forces g ( 0 , 0 ) = 0 g(0,0)=0 g ( 0 , 0 ) = 0 , since a nonzero constant term in g g g would make the integral diverge. Thus, near ( 0 , 0 ) (0,0) ( 0 , 0 ) , the Taylor expansion of g g g takes the form
g ( ρ , y ) = ρ B ( θ ) + ∑ α , i c α i y α ( ρ 1 / 2 ∂ i φ ) + O ( ( ρ 2 + ∣ y ∣ 2 ) ∥ w ∥ L 2 ) , B ( θ ) ∈ V . g(\rho,y)=\rho B(\theta)+\sum_{\alpha,i}c_{\alpha i}y_\alpha(\rho^{1/2}\partial_i\varphi)+O\left((\rho^2+|y|^2)\lVert w\rVert_{L^2}\right),\qquad B(\theta)\in\mathcal{V}. g ( ρ , y ) = ρB ( θ ) + α , i ∑ c α i y α ( ρ 1/2 ∂ i φ ) + O ( ( ρ 2 + ∣ y ∣ 2 ) ∥ w ∥ L 2 ) , B ( θ ) ∈ V .
Then we define ℓ : = ρ 1 / 2 B ( θ ) + ∑ α , i c α i y α ∂ i φ \ell:=\rho^{1/2}B(\theta)+\sum_{\alpha,i}c_{\alpha i}y_\alpha\partial_i\varphi ℓ := ρ 1/2 B ( θ ) + ∑ α , i c α i y α ∂ i φ . For 0 < s < 1 / 32 0<s<1/32 0 < s < 1/32 , the remainder estimate is given by
∫ B s ∣ w 1 / 2 − ℓ ∣ 2 ≲ ∥ w ∥ L 2 ( B 1 / 2 ) 2 ∫ ρ , ∣ y ∣ ≲ s ( ρ 2 + ∣ y ∣ 2 ) 2 ρ − 1 d ρ d y ≲ s n + 3 ∥ w ∥ L 2 ( B 1 / 2 ) 2 . \int_{B_s}|w_{1/2}-\ell|^2\lesssim\lVert w\rVert_{L^2(B_{1/2})}^2\int_{\rho,|y|\lesssim s}(\rho^2+|y|^2)^2\rho^{-1}\,d\rho\,dy\lesssim s^{n+3}\lVert w\rVert_{L^2(B_{1/2})}^2. ∫ B s ∣ w 1/2 − ℓ ∣ 2 ≲ ∥ w ∥ L 2 ( B 1/2 ) 2 ∫ ρ , ∣ y ∣ ≲ s ( ρ 2 + ∣ y ∣ 2 ) 2 ρ − 1 d ρ d y ≲ s n + 3 ∥ w ∥ L 2 ( B 1/2 ) 2 .
Case 2: The higher angular frequencies ν ≥ 3 / 2 \nu\ge3/2 ν ≥ 3/2 . For ν ≥ 3 / 2 \nu\ge3/2 ν ≥ 3/2 set g ν = ρ − ν f ν g_\nu=\rho^{-\nu}f_\nu g ν = ρ − ν f ν . Equation (44) becomes
∂ ρ 2 g ν + ( 2 ν + 1 ) ρ − 1 ∂ ρ g ν + Δ y g ν = 0. \partial_\rho^2g_\nu+(2\nu+1)\rho^{-1}\partial_\rho g_\nu+\Delta_y g_\nu=0. ∂ ρ 2 g ν + ( 2 ν + 1 ) ρ − 1 ∂ ρ g ν + Δ y g ν = 0.
Regarding g ν g_\nu g ν as the function g ν ( ∣ x ∣ , y ) g_\nu(|x|,y) g ν ( ∣ x ∣ , y ) of x ∈ R 2 ν + 2 x\in\mathbb{R}^{2\nu+2} x ∈ R 2 ν + 2 , we see that g ν g_\nu g ν is harmonic away from the codimension 2 ν + 2 ≥ 5 2\nu+2\ge5 2 ν + 2 ≥ 5 axis { x = 0 } \{x=0\} { x = 0 } with L 2 L^2 L 2 integral ∫ ∣ f ν ∣ 2 ρ d ρ d y \int|f_\nu|^2\rho\,d\rho\,dy ∫ ∣ f ν ∣ 2 ρ d ρ d y , up to constants. A standard approximation argument implies that g ν g_\nu g ν is weakly harmonic across the axis, and hence harmonic by Weyl’s lemma. The monotonicity formula for harmonic functions implies that
∫ B s ∣ w ν ∣ 2 ≤ ( 4 s ) n + 2 ν ∫ B 1 / 4 ∣ w ν ∣ 2 , 0 < s < 1 / 4. \int_{B_s}|w_\nu|^2\le(4s)^{n+2\nu}\int_{B_{1/4}}|w_\nu|^2,\qquad0<s<1/4. ∫ B s ∣ w ν ∣ 2 ≤ ( 4 s ) n + 2 ν ∫ B 1/4 ∣ w ν ∣ 2 , 0 < s < 1/4.
Since 4 s < 1 4s<1 4 s < 1 and 2 ν ≥ 3 2\nu\ge3 2 ν ≥ 3 , Parseval’s identity implies
∑ ν ≥ 3 / 2 ∫ B s ∣ w ν ∣ 2 ≤ ( 4 s ) n + 3 ∑ ν ≥ 3 / 2 ∫ B 1 / 4 ∣ w ν ∣ 2 ≤ C s n + 3 ∥ w ∥ L 2 ( B 1 / 2 ) 2 . \sum_{\nu\ge3/2}\int_{B_s}|w_\nu|^2\le(4s)^{n+3}\sum_{\nu\ge3/2}\int_{B_{1/4}}|w_\nu|^2\le Cs^{n+3}\|w\|_{L^2(B_{1/2})}^2. ν ≥ 3/2 ∑ ∫ B s ∣ w ν ∣ 2 ≤ ( 4 s ) n + 3 ν ≥ 3/2 ∑ ∫ B 1/4 ∣ w ν ∣ 2 ≤ C s n + 3 ∥ w ∥ L 2 ( B 1/2 ) 2 .
The coefficients of ℓ \ell ℓ are bounded by the preceding C 1 C^1 C 1 estimate for g g g , so ∥ ℓ ∥ L 2 ( B 1 / 2 ) ≤ C ∥ w ∥ L 2 ( B 1 / 2 ) \|\ell\|_{L^2(B_{1/2})}\le C\|w\|_{L^2(B_{1/2})} ∥ ℓ ∥ L 2 ( B 1/2 ) ≤ C ∥ w ∥ L 2 ( B 1/2 ) as well.
Finally s − 1 / 2 w ( s ⋅ ) → ℓ s^{-1/2}w(s\cdot)\to\ell s − 1/2 w ( s ⋅ ) → ℓ in L 2 L^2 L 2 on fixed balls as s ↓ 0 s\downarrow0 s ↓ 0 . Passing to the limit in the Dirac equation away from P P P gives D 0 ℓ = 0 D_0\ell=0 D 0 ℓ = 0 and hence ℓ ∈ T Φ 0 C D \ell\in T_{\Phi_0}\mathcal{C}_D ℓ ∈ T Φ 0 C D by Lemma 2.10 . This completes the proof. □
Campanato iteration
Given Proposition 3.6 and Lemma 4.1 , we are ready to run the Campanato iteration scheme. Fix any Hölder exponent β ∈ ( 0 , 1 ) \beta\in(0,1) β ∈ ( 0 , 1 ) .
Proposition 4.2. There are θ ∈ ( 0 , 1 / 8 ) \theta\in(0,1/8) θ ∈ ( 0 , 1/8 ) , ε > 0 \varepsilon>0 ε > 0 and C < ∞ C<\infty C < ∞ with the following property. Suppose that p ∈ B u p\in B_u p ∈ B u with N u ( p , 0 ) ≥ 1 2 N_u(p,0)\ge\frac12 N u ( p , 0 ) ≥ 2 1 , and a frequency-1 2 \frac12 2 1 comparison Dirac model Φ \Phi Φ in a compact non-degenerate model chart, satisfy the smallness of L 2 L^2 L 2 height excess, i.e.
E ( u , Φ ; p , r ) + r 2 ≤ ε . \mathcal{E}(u,\Phi;p,r)+r^2\le\varepsilon. E ( u , Φ ; p , r ) + r 2 ≤ ε . Then there exists a nearby Dirac model Φ ′ \Phi' Φ ′ such that
E ( u , Φ ′ ; p , θ r ) ≤ 1 2 θ 2 β E ( u , Φ ; p , r ) + C r 2 . (45) \mathcal{E}(u,\Phi';p,\theta r)\le\frac12\theta^{2\beta}\mathcal{E}(u,\Phi;p,r)+Cr^2. \tag*{(45)} E ( u , Φ ′ ; p , θ r ) ≤ 2 1 θ 2 β E ( u , Φ ; p , r ) + C r 2 . (45) d C ( Φ ′ , Φ ) ≤ C ( E ( u , Φ ; p , r ) 1 / 2 + r ) . (46) d_{\mathcal C}(\Phi',\Phi)\le C\left(\mathcal{E}(u,\Phi;p,r)^{1/2}+r\right). \tag*{(46)} d C ( Φ ′ , Φ ) ≤ C ( E ( u , Φ ; p , r ) 1/2 + r ) . (46)
Proof. Choose θ ∈ ( 0 , 1 / 32 ) \theta\in(0,1/32) θ ∈ ( 0 , 1/32 ) such that C θ 2 ≤ θ 2 β / 8 C\theta^2\le\theta^{2\beta}/8 C θ 2 ≤ θ 2 β /8 for the constant C C C in Lemma 4.1 . The estimates in Proposition 3.3 require only N u ( p , 0 ) ≥ 1 2 N_u(p,0)\ge\frac12 N u ( p , 0 ) ≥ 2 1 ; the comparison models remain homogeneous of degree 1 2 \frac12 2 1 .
Suppose the conclusion fails. For each j ≥ 1 j\ge1 j ≥ 1 , choose ( u j , Φ j , p j , r j ) (u_j,\Phi_j,p_j,r_j) ( u j , Φ j , p j , r j ) with E j + r j 2 ≤ 1 / j E_j+r_j^2\le1/j E j + r j 2 ≤ 1/ j , where E j : = E ( u j , Φ j ; p j , r j ) E_j:=\mathcal{E}(u_j,\Phi_j;p_j,r_j) E j := E ( u j , Φ j ; p j , r j ) , such that for every nearby Dirac model Φ j ′ \Phi'_j Φ j ′ , at least one of (45) and (46) fails with C = j C=j C = j . Taking Φ j ′ = Φ j \Phi'_j=\Phi_j Φ j ′ = Φ j , we know
E ( u j , Φ j ; p j , θ r j ) > 1 2 θ 2 β E j + j r j 2 . \mathcal{E}(u_j,\Phi_j;p_j,\theta r_j)>\frac12\theta^{2\beta}E_j+jr_j^2. E ( u j , Φ j ; p j , θ r j ) > 2 1 θ 2 β E j + j r j 2 .
On the other hand, comparing the volume, we know E ( u j , Φ j ; p j , θ r j ) ≤ θ − n − 1 E j \mathcal{E}(u_j,\Phi_j;p_j,\theta r_j)\le\theta^{-n-1}E_j E ( u j , Φ j ; p j , θ r j ) ≤ θ − n − 1 E j . It follows that r j 2 E j → 0 \frac{r_j^2}{E_j}\to0 E j r j 2 → 0 . Rescale to the unit ball and pass to a subsequence with Φ j → Φ 0 \Phi_j\to\Phi_0 Φ j → Φ 0 . Proposition 3.6 and Lemma 4.1 give ℓ ∈ T Φ 0 C D \ell\in T_{\Phi_0}\mathcal{C}_D ℓ ∈ T Φ 0 C D with excess decay estimates. Choose ℓ j ∈ T Φ j C D \ell_j\in T_{\Phi_j}\mathcal{C}_D ℓ j ∈ T Φ j C D converging to ℓ \ell ℓ . Lemma 2.10 , with parameter E j \sqrt{E_j} E j , gives nearby Dirac models Φ j ′ \Phi'_j Φ j ′ with d C ( Φ j ′ , Φ j ) ≤ C E j d_{\mathcal C}(\Phi'_j,\Phi_j)\le C\sqrt{E_j} d C ( Φ j ′ , Φ j ) ≤ C E j . The expansion (24) and strong L 2 L^2 L 2 -convergence, together with Corollary 3.5 , imply
lim sup j → ∞ E ( u j , Φ j ′ ; 0 , θ ) E j ≤ θ − n − 1 ∫ B θ ∣ w − ℓ ∣ 2 ≤ C θ 2 . \limsup_{j\to\infty}\frac{\mathcal{E}(u_j,\Phi'_j;0,\theta)}{E_j}\le\theta^{-n-1}\int_{B_\theta}|w-\ell|^2\le C\theta^2. j → ∞ lim sup E j E ( u j , Φ j ′ ; 0 , θ ) ≤ θ − n − 1 ∫ B θ ∣ w − ℓ ∣ 2 ≤ C θ 2 .
By the choice of θ \theta θ , both required estimates hold for large j j j , a contradiction. □
At a fixed centre p ∈ B u p\in B_u p ∈ B u with N ( p , 0 ) = 1 2 N(p,0)=\frac12 N ( p , 0 ) = 2 1 , Theorem 2.6 gives a nonzero tangent model Φ \Phi Φ and arbitrarily small scales r 0 r_0 r 0 at which the height-normalised spinor is close to Φ \Phi Φ . To start the iteration with the rescaling used here, multiply u u u once by ( r 0 / h ( r 0 ) ) 1 / 2 (r_0/h(r_0))^{1/2} ( r 0 / h ( r 0 ) ) 1/2 , where h h h is defined in Subsection 2.5, and retain the name u u u . Its r 0 − 1 / 2 r_0^{-1/2} r 0 − 1/2 rescaling is precisely the height-normalised spinor, so E ( u , Φ ; p , r 0 ) + r 0 2 \mathcal{E}(u,\Phi;p,r_0)+r_0^2 E ( u , Φ ; p , r 0 ) + r 0 2 is sufficiently small. This multiplicative constant remains fixed throughout the iteration; it can be undone at the end. Set r k = θ k r 0 r_k=\theta^k r_0 r k = θ k r 0 . Applying Proposition 4.2 iteratively, we obtain nearby flat Dirac models Φ k \Phi_k Φ k with excesses E k : = E ( u , Φ k ; p , r k ) E_k:=\mathcal{E}(u,\Phi_k;p,r_k) E k := E ( u , Φ k ; p , r k ) satisfying
E k ≤ C ( E 0 + r 0 2 ) θ 2 β k , ∑ j ≥ k d C ( Φ j + 1 , Φ j ) ≤ C ( E 0 1 / 2 + r 0 ) θ β k , ∀ k ≥ 1. (47) E_k\leq C(E_0+r_0^2)\theta^{2\beta k},\qquad\sum_{j\geq k}d_{\mathcal{C}}(\Phi_{j+1},\Phi_j)\leq C(E_0^{1/2}+r_0)\theta^{\beta k},\qquad\forall k\geq1. \tag*{(47)} E k ≤ C ( E 0 + r 0 2 ) θ 2 β k , j ≥ k ∑ d C ( Φ j + 1 , Φ j ) ≤ C ( E 0 1/2 + r 0 ) θ β k , ∀ k ≥ 1. (47)
Thus { Φ j } \{\Phi_j\} { Φ j } is Cauchy, converging to a nondegenerate Dirac model Φ p \Phi_p Φ p with L 2 L^2 L 2 -excess estimate at all small scales,
E ( u , Φ p ; p , s ) ≤ C ( E 0 + r 0 2 ) ( s / r 0 ) 2 β , 0 < s ≤ r 0 . (48) \mathcal{E}(u,\Phi_p;p,s)\leq C(E_0+r_0^2)(s/r_0)^{2\beta},\qquad0<s\leq r_0. \tag*{(48)} E ( u , Φ p ; p , s ) ≤ C ( E 0 + r 0 2 ) ( s / r 0 ) 2 β , 0 < s ≤ r 0 . (48)
In particular, the tangent map Φ p \Phi_p Φ p is unique.
The following lemma compares the parameters of nearby models with their L 2 L^2 L 2 -distance.
Lemma 4.3. Let p , p ′ ∈ B u p,p'\in\mathcal{B}_u p , p ′ ∈ B u with d g ( p , p ′ ) < r / 2 d_g(p,p')<r/2 d g ( p , p ′ ) < r /2 , r ≪ 1 r\ll1 r ≪ 1 , have frequency at least 1 / 2 1/2 1/2 , and let Φ , Φ ′ \Phi,\Phi' Φ , Φ ′ be flat Dirac comparison models centred at these points respectively, in a fixed nondegenerate model chart. Then
d C ( Φ , Φ ′ ) 2 + r − 2 ∣ Π P ⊥ exp p − 1 ( p ′ ) ∣ 2 ≤ C ( E ( u , Φ ; p , r ) + E ( u , Φ ′ ; p ′ , r ) + r 2 ) , (49) d_{\mathcal{C}}(\Phi,\Phi')^2+r^{-2}\lvert\Pi_{P^\perp}\exp_p^{-1}(p')\rvert^2\leq C\bigl(\mathcal{E}(u,\Phi;p,r)+\mathcal{E}(u,\Phi';p',r)+r^2\bigr), \tag*{(49)} d C ( Φ , Φ ′ ) 2 + r − 2 ∣ Π P ⊥ exp p − 1 ( p ′ ) ∣ 2 ≤ C ( E ( u , Φ ; p , r ) + E ( u , Φ ′ ; p ′ , r ) + r 2 ) , (49)
provided that the right-hand side is sufficiently small, where we compare Φ \Phi Φ and Φ ′ \Phi' Φ ′ by parallel transport, and P , P ′ P,P' P , P ′ stand for their spines respectively.
Proof. Let Ψ \Psi Ψ be the model at p ′ p' p ′ obtained by parallel transport of the normal frame and coefficients of Φ \Phi Φ . Proposition 3.3 bounds the normal displacement ∣ Π P ⊥ exp p − 1 ( p ′ ) ∣ \lvert\Pi_{P^\perp}\exp_p^{-1}(p')\rvert ∣ Π P ⊥ exp p − 1 ( p ′ )∣ . Together with the translation and model comparisons in Step 3 of its proof, this gives
r − 2 ∣ Π P ⊥ exp p − 1 ( p ′ ) ∣ 2 + r − n − 1 ∫ B r / 4 ( p ′ ) G ( Ψ , Φ ′ ) 2 d V g ≤ C ( E ( u , Φ ; p , r ) + r 2 ) . r^{-2}\lvert\Pi_{P^\perp}\exp_p^{-1}(p')\rvert^2+r^{-n-1}\int_{B_{r/4}(p')} \mathcal{G}(\Psi,\Phi')^2\,dV_g\leq C\bigl(\mathcal{E}(u,\Phi;p,r)+r^2\bigr). r − 2 ∣ Π P ⊥ exp p − 1 ( p ′ ) ∣ 2 + r − n − 1 ∫ B r /4 ( p ′ ) G ( Ψ , Φ ′ ) 2 d V g ≤ C ( E ( u , Φ ; p , r ) + r 2 ) .
Since B r / 4 ( p ′ ) ⊂ B r ( p ) B_{r/4}(p')\subset B_r(p) B r /4 ( p ′ ) ⊂ B r ( p ) , the triangle inequality gives E ( u , Ψ ; p ′ , r / 4 ) ≤ C ( E ( u , Φ ; p , r ) + r 2 ) \mathcal{E}(u,\Psi;p',r/4)\leq C\bigl(\mathcal{E}(u,\Phi;p,r)+r^2\bigr) E ( u , Ψ ; p ′ , r /4 ) ≤ C ( E ( u , Φ ; p , r ) + r 2 ) .
In d C ( Φ , Φ ′ ) d_{\mathcal{C}}(\Phi,\Phi') d C ( Φ , Φ ′ ) , the parameters of Φ \Phi Φ are compared at p ′ p' p ′ , where they are represented by Ψ \Psi Ψ . For models with the same centre, the explicit formula (22) and the fixed model gauge control parameter distance by normalised L 2 L^2 L 2 -distance. Applying the triangle inequality through u u u gives
d C ( Φ , Φ ′ ) 2 ≤ C ( E ( u , Ψ ; p ′ , r / 4 ) + E ( u , Φ ′ ; p ′ , r / 4 ) ) ≤ C ( E ( u , Φ ; p , r ) + E ( u , Φ ′ ; p ′ , r ) + r 2 ) . \begin{aligned}
d_{\mathcal{C}}(\Phi,\Phi')^2\leq C\bigl(\mathcal{E}(u,\Psi;p',r/4)+\mathcal{E}(u,\Phi';p',r/4)\bigr) \\
&\leq C\bigl(\mathcal{E}(u,\Phi;p,r)+\mathcal{E}(u,\Phi';p',r)+r^2\bigr).
\end{aligned} d C ( Φ , Φ ′ ) 2 ≤ C ( E ( u , Ψ ; p ′ , r /4 ) + E ( u , Φ ′ ; p ′ , r /4 ) ) ≤ C ( E ( u , Φ ; p , r ) + E ( u , Φ ′ ; p ′ , r ) + r 2 ) .
Together these prove (49) .
Nearby centres and the branch graph
Fix a branch point p 0 p_0 p 0 with frequency 1 2 \frac12 2 1 ; we aim to transfer the estimates E ( u , Φ p 0 ; p 0 , s ) ≤ C s 2 β \mathcal{E}(u,\Phi_{p_0};p_0,s)\leq Cs^{2\beta} E ( u , Φ p 0 ; p 0 , s ) ≤ C s 2 β , s ≪ 1 s\ll1 s ≪ 1 to nearby branch points.
Proposition 4.4. There are a small neighbourhood U U U of p 0 p_0 p 0 , radius r 0 ≪ 1 r_0\ll1 r 0 ≪ 1 , and C < ∞ C<\infty C < ∞ such that every p ∈ B u ∩ U p\in\mathcal{B}_u\cap U p ∈ B u ∩ U has frequency 1 / 2 1/2 1/2 and a unique nondegenerate Dirac model Φ p \Phi_p Φ p , with
E ( u , Φ p ; p , s ) ≤ C s 2 β , ∀ 0 < s < r 0 . (50) \mathcal{E}(u,\Phi_p;p,s)\leq Cs^{2\beta},\qquad\forall0<s<r_0. \tag*{(50)} E ( u , Φ p ; p , s ) ≤ C s 2 β , ∀0 < s < r 0 . (50)
Moreover, for p ≠ p ′ ∈ B u ∩ U p\neq p'\in\mathcal{B}_u\cap U p = p ′ ∈ B u ∩ U ,
d C ( Φ p , Φ p ′ ) + ∣ Π P p ⊥ exp p − 1 ( p ′ ) ∣ d g ( p , p ′ ) ≤ C d g ( p , p ′ ) β . (51) d_{\mathcal{C}}(\Phi_p,\Phi_{p'})+\frac{\lvert\Pi_{P_p^\perp}\exp_p^{-1}(p')\rvert}{d_g(p,p')}\leq Cd_g(p,p')^\beta. \tag*{(51)} d C ( Φ p , Φ p ′ ) + d g ( p , p ′ ) ∣ Π P p ⊥ exp p − 1 ( p ′ )∣ ≤ C d g ( p , p ′ ) β . (51)
In particular, B u ∩ U \mathcal{B}_u\cap U B u ∩ U is a single C 1 , β C^{1,\beta} C 1 , β -graph over the spine P p 0 P_{p_0} P p 0 with T p B u = P p T_p\mathcal{B}_u=P_p T p B u = P p .
Proof. Choosing r 0 r_0 r 0 sufficiently small, we can make E 0 + r 0 2 E_0+r_0^2 E 0 + r 0 2 arbitrarily small by the fixed-centre Campanato iteration. Now for p ∈ B u ∩ B r 0 / 2 ( p 0 ) p\in\mathcal{B}_u\cap B_{r_0/2}(p_0) p ∈ B u ∩ B r 0 /2 ( p 0 ) , applying Proposition 3.3, we have
r 0 − 2 ∣ ξ ∣ 2 ≲ ( E 0 + r 0 2 ) . r_0^{-2}|\xi|^2\lesssim(E_0+r_0^2). r 0 − 2 ∣ ξ ∣ 2 ≲ ( E 0 + r 0 2 ) .
Let Ψ \Psi Ψ be obtained by parallel transport of the oriented normal frame and coefficients of Φ p 0 \Phi_{p_0} Φ p 0 from p 0 p_0 p 0 to p p p , as in Step 3 of Proposition 3.3. Then
E ( u , Ψ ; p , r 0 / 4 ) + ( r 0 / 4 ) 2 ≲ ( E 0 + r 0 − 2 ∣ ξ ∣ 2 + r 0 2 ) ≲ C ( E 0 + r 0 2 ) . \mathcal{E}(u,\Psi;p,r_0/4)+(r_0/4)^2\lesssim(E_0+r_0^{-2}|\xi|^2+r_0^2)\lesssim C(E_0+r_0^2). E ( u , Ψ ; p , r 0 /4 ) + ( r 0 /4 ) 2 ≲ ( E 0 + r 0 − 2 ∣ ξ ∣ 2 + r 0 2 ) ≲ C ( E 0 + r 0 2 ) .
Since N u ( p , 0 ) ≥ 1 / 2 N_u(p,0) \ge1/2 N u ( p , 0 ) ≥ 1/2 , Proposition 4.2 applies. Then Campanato iteration implies a unique tangent Dirac model Φ p \Phi_p Φ p with excess decay estimates
E ( u , Φ p ; p , s ) ≤ C ( E 0 + r 0 2 ) ( s / r 0 ) 2 β , 0 < s < r 0 / 4. \mathcal{E}(u,\Phi_p;p,s) \le C(E_0+r_0^2)(s/r_0)^{2\beta}, \qquad0<s<r_0/4. E ( u , Φ p ; p , s ) ≤ C ( E 0 + r 0 2 ) ( s / r 0 ) 2 β , 0 < s < r 0 /4.
where the constant C C C is independent of p p p . Therefore, N ( p , 0 ) = 1 2 N(p,0)=\frac{1}{2} N ( p , 0 ) = 2 1 by Theorem 2.6 . To compare nearby centres p , p ′ p,p' p , p ′ with distance d = d g ( p , p ′ ) d=d_g(p,p') d = d g ( p , p ′ ) , Lemma 4.3 (with r = 10 d r=10d r = 10 d ) implies (4.9):
d C ( Φ p , Φ p ′ ) 2 + d − 2 ∣ Π P p ⊥ exp p − 1 ( p ′ ) ∣ 2 ≲ ( d 2 β + d 2 β + d 2 ) ≲ d 2 β . (52) d_C(\Phi_p,\Phi_{p'})^2+d^{-2}\lvert\Pi_{P_p^\perp}\exp_{p}^{-1}(p')\rvert^2\lesssim(d^{2\beta}+d^{2\beta}+d^2)\lesssim d^{2\beta}. \tag*{(52)} d C ( Φ p , Φ p ′ ) 2 + d − 2 ∣ Π P p ⊥ exp p − 1 ( p ′ ) ∣ 2 ≲ ( d 2 β + d 2 β + d 2 ) ≲ d 2 β . (52)
The C 1 , β C^{1,\beta} C 1 , β -graphical representation then follows from the no-hole principle and (4.9). □
Tangent estimates and conclusion
For later use in Part II, we record in the following corollary certain quantitative bounds of the height and tilt excess.
Corollary 4.5. For a sufficiently small neighbourhood U U U , there exists α ∈ ( 0 , β ) \alpha\in(0,\beta) α ∈ ( 0 , β ) such that for all small scales r ≪ 1 r\ll1 r ≪ 1 ,
r − n − 1 ∫ B r ( p ) G ( u , Φ p ) 2 ≤ C r 2 α , sup B r / 2 ( p ) G ( u , Φ p ) ≤ C r 1 / 2 + α , (53) r^{-n-1}\int_{B_r(p)}\mathcal{G}(u,\Phi_p)^2\le Cr^{2\alpha},\qquad\sup_{B_{r/2}(p)}\mathcal{G}(u,\Phi_p)\le Cr^{1/2+\alpha}, \tag*{(53)} r − n − 1 ∫ B r ( p ) G ( u , Φ p ) 2 ≤ C r 2 α , B r /2 ( p ) sup G ( u , Φ p ) ≤ C r 1/2 + α , (53)
r 1 − n ∫ B r / 2 ( p ) G ( ∇ u , ∇ Φ p ) 2 ≤ C r 2 α . (54) r^{1-n}\int_{B_{r/2}(p)}\mathcal{G}(\nabla u,\nabla\Phi_p)^2\le Cr^{2\alpha}. \tag*{(54)} r 1 − n ∫ B r /2 ( p ) G ( ∇ u , ∇ Φ p ) 2 ≤ C r 2 α . (54)
For every fixed λ ∈ ( 0 , 1 / 4 ) \lambda\in(0,1/4) λ ∈ ( 0 , 1/4 ) and integer k ≥ 0 k\ge0 k ≥ 0 , locally matched sheets satisfy
sup x ∈ B r / 2 ( p ) dist ( x , B u ) > λ r r k − 1 / 2 ∣ ∇ k ( u − φ p ) ( x ) ∣ ≤ C λ , k r α , ∀ r ≪ 1. (55) \sup_{\substack{x\in B_{r/2}(p)\\ \operatorname{dist}(x,\mathcal{B}_u)>\lambda r}}r^{k-1/2}\lvert\nabla^k(u-\varphi_p)(x)\rvert\le C_{\lambda,k}r^\alpha,\qquad\forall r\ll1. \tag*{(55)} x ∈ B r /2 ( p ) dist ( x , B u ) > λ r sup r k − 1/2 ∣ ∇ k ( u − φ p ) ( x )∣ ≤ C λ , k r α , ∀ r ≪ 1. (55)
Sketch of the Proof. We choose
0 < α < min { β 3 , β γ n + γ } . 0<\alpha<\min\left\{\frac{\beta}{3},\frac{\beta\gamma}{n+\gamma}\right\}. 0 < α < min { 3 β , n + γ β γ } .
where γ \gamma γ is as in Proposition 2.5, which gives local C γ / 2 C^{\gamma/2} C γ /2 estimates for u u u and Φ p \Phi_p Φ p . By interpolation with the L 2 L^2 L 2 excess decay estimates, (4.10) follows. For the gradient estimate (4.11), the branch locus is (locally) contained in a C r 1 + β Cr^{1+\beta} C r 1 + β -tube over the spine P p P_p P p in B r ( p ) B_r(p) B r ( p ) . Away from this tube at scale s r sr sr (with s ≫ r β s\gg r^\beta s ≫ r β and s 1 / 2 ≳ r β γ / ( n + γ ) s^{1/2}\gtrsim r^{\beta\gamma/(n+\gamma)} s 1/2 ≳ r β γ / ( n + γ ) ), we can select locally matched sheets by Lemma 3.1 . Since G ( D u , D Φ p ) ≤ ∣ D ( u − φ p ) ∣ \mathcal{G}(Du,D\Phi_p)\le\lvert D(u-\varphi_p)\rvert G ( D u , D Φ p ) ≤ ∣ D ( u − φ p )∣ for these choices of signs, the interior estimates give
r 1 − n ∫ ρ > s r G ( D u , D Φ p ) 2 ≲ s − 2 r − n − 1 ∫ ρ > s r G ( u , Φ p ) 2 + r 2 ≲ s − 2 r 2 β + r 2 . r^{1-n}\int_{\rho>sr}\mathcal{G}(Du,D\Phi_p)^2\lesssim s^{-2}r^{-n-1}\int_{\rho>sr}\mathcal{G}(u,\Phi_p)^2+r^2\lesssim s^{-2}r^{2\beta}+r^2. r 1 − n ∫ ρ > sr G ( D u , D Φ p ) 2 ≲ s − 2 r − n − 1 ∫ ρ > sr G ( u , Φ p ) 2 + r 2 ≲ s − 2 r 2 β + r 2 .
We cover the tube by at most C s 2 − n Cs^{2-n} C s 2 − n balls of radius C s r Csr C sr centred on B u \mathcal{B}_u B u . The frequency estimates for u u u and the explicit formula for Φ p \Phi_p Φ p give
r 1 − n ∫ ρ < s r G ( D u , D Φ p ) 2 ≲ r 1 − n s 2 − n ( s r ) n − 1 ≲ s . r^{1-n}\int_{\rho<sr}\mathcal{G}(Du,D\Phi_p)^2\lesssim r^{1-n}s^{2-n}(sr)^{n-1}\lesssim s. r 1 − n ∫ ρ < sr G ( D u , D Φ p ) 2 ≲ r 1 − n s 2 − n ( sr ) n − 1 ≲ s .
Taking s = r 2 α s=r^{2\alpha} s = r 2 α , we obtain (4.11). At points whose distance from the spine is at least λ r \lambda r λ r , the higher order estimate (4.12) follows from the interior elliptic estimates. □
Theorem 2 follows directly from Proposition 4.4 and Corollary 4.5.
Part 2. Higher regularity of the branching locus
In Part II, we prove that the branching locus is smooth (resp. real-analytic) for smooth (resp. real-analytic) ambient data. The input from Part I is Theorem 2 : near a frequency-1 2 \frac{1}{2} 2 1 branch point, the branching locus is a C 1 , α C^{1,\alpha} C 1 , α -graph, and the spinor is approximated by non-degenerate tangent models at rate r 1 / 2 + α r^{1/2+\alpha} r 1/2 + α .
The partial Legendre transform (cf. § 3.1 ) sends the branching locus to { ξ = 0 } \{\xi=0\} { ξ = 0 } and reduces its regularity to the tangential regularity of the transformed spinor U U U . Then the field U U U satisfies a quasilinear elliptic system with inherited variational structure, degenerating along { ξ = 0 } \{\xi= 0\} { ξ = 0 } . This degeneracy is captured by Grushin-type weighted spaces. Moreover, tangential derivatives D y β U D_y^\beta U D y β U satisfy a linearised equation of the schematic form
B U ( Z , Y ) = F ( Y ) , \mathcal{B}_U(Z,Y)=\mathcal{F}(Y), B U ( Z , Y ) = F ( Y ) ,
with sources determined by lower-order derivatives. Higher regularity therefore reduces to a linear theory for B U \mathcal{B}_U B U near { ξ = 0 } \{\xi= 0\} { ξ = 0 } , which we develop in three steps.
The frozen operator (§ 3.4 ). By reduction to the square model U 0 = ( ξ 2 , 0 ) U_0=(\xi^2,0) U 0 = ( ξ 2 , 0 ) , we prove coercivity of the second variation at a transformed tangent model and establish the excess decay for weak solutions of the corresponding linear equation.
Excess decay (§ 3.5 ). Treating B U \mathcal{B}_U B U as a perturbation of the frozen operator yields a Campanato-type decay to a comparison polynomial, where an extra polynomial correction (using Lemma 6.5 ) is required to absorb the principal part of the source term.
Induction on the order of derivatives (§ 3.6 ). We upgrade the L 2 L^2 L 2 -excess decay of § 3.5 to weighted C 1 , α C^{1,\alpha} C 1 , α -estimates, which control the sources of the tangentially differentiated equations. Induction on the order of differentiation then yields smoothness and real-analyticity.
Partial Legendre transform and weighted spaces
We set up the partial Legendre transform in § 3.1.1 and the weighted spaces adapted to the degeneracy at { ξ = 0 } \{\xi= 0\} { ξ = 0 } in § 3.1.2 . § 3.2 translates the estimates of Part I into these coordinates, and § 3.3 derives the transformed Euler–Lagrange equation.
Partial Legendre transform
5.1. Partial Legendre transform. Throughout Part II, we fix a frequency-1 2 \frac{1}{2} 2 1 branch point ( 0 , y 0 ) (0,y_0) ( 0 , y 0 ) and work in the ball B r 0 ( 0 , y 0 ) B_{r_0}(0,y_0) B r 0 ( 0 , y 0 ) such that Theorem 2 applies. For sufficiently small r 0 ≪ 1 r_0\ll1 r 0 ≪ 1 , all constants in the estimate of Part II will be uniform over this ball.
We use coordinates X = ( z , y ) ∈ C × R n − 2 X=(z,y)\in\mathbb{C}\times\mathbb{R}^{n-2} X = ( z , y ) ∈ C × R n − 2 in which P 0 = { z = 0 } P_0=\{z=0\} P 0 = { z = 0 } is the spine at ( 0 , y 0 ) (0,y_0) ( 0 , y 0 ) , and the exponential gauge centred at ( 0 , y 0 ) (0,y_0) ( 0 , y 0 ) , which identifies the fibres of E E E with the fibre E 0 E_0 E 0 at ( 0 , y 0 ) (0,y_0) ( 0 , y 0 ) . By Theorem 2 ,
B u ∩ B r 0 ( 0 , y 0 ) = { p ( y ) = ( g ( y ) , y ) } , g ∈ C 1 , α , g ( y 0 ) = 0 , D g ( y 0 ) = 0. \mathcal{B}_u\cap B_{r_0}(0,y_0)=\{p(y)=(g(y),y)\},\qquad g\in C^{1,\alpha},\qquad g(y_0)=0,\qquad Dg(y_0)=0. B u ∩ B r 0 ( 0 , y 0 ) = { p ( y ) = ( g ( y ) , y )} , g ∈ C 1 , α , g ( y 0 ) = 0 , D g ( y 0 ) = 0.
For p ∈ B u ∩ B r 0 ( 0 , y 0 ) p\in\mathcal{B}_u\cap B_{r_0}(0,y_0) p ∈ B u ∩ B r 0 ( 0 , y 0 ) , let Φ p \Phi_p Φ p be its tangent model with spine P p = T p B u P_p=T_p\mathcal{B}_u P p = T p B u . As in Remark 2.9 , we write Φ p = ± T p ( Π P p ⊥ ( X − p ) ) 1 / 2 \Phi_p=\pm T_p(\Pi_{P_p^\perp}(X-p))^{1/2} Φ p = ± T p ( Π P p ⊥ ( X − p ) ) 1/2 with T p ( w ) = a p w + b p w ‾ T_p(w)=a_pw+b_p\overline{w} T p ( w ) = a p w + b p w , where Π P p ⊥ \Pi_{P_p^\perp} Π P p ⊥ is the orthogonal projection onto P p ⊥ ≅ C P_p^\perp\cong\mathbb{C} P p ⊥ ≅ C . Fix a real-linear projection π : E 0 → C \pi:E_0\to\mathbb{C} π : E 0 → C such that π T p ( y 0 ) \pi T_{p(y_0)} π T p ( y 0 ) is invertible with positive determinant; by continuity, so is π T p \pi T_p π T p . Let π ⊥ \pi^\perp π ⊥ be the complementary orthogonal projection, and write ξ \xi ξ and V V V for the components of a spinor in E 0 = im π ⊕ im π ⊥ E_0=\operatorname{im}\pi\oplus\operatorname{im}\pi^\perp E 0 = im π ⊕ im π ⊥ .
By Theorem 2 , the map ( z , y ) ↦ ( ξ , y ) : = ( π u ( z , y ) , y ) (z,y)\mapsto(\xi,y):=(\pi u(z,y),y) ( z , y ) ↦ ( ξ , y ) := ( π u ( z , y ) , y ) is invertible, and we write its inverse and the spinor in the new coordinates as
X ( ξ , y ) = ( φ ( ξ , y ) , y ) , u ^ ( ξ , y ) : = u ( X ( ξ , y ) ) = ( ξ , V ( ξ , y ) ) . (56) X(\xi,y)=(\varphi(\xi,y),y),\qquad\widehat{u}(\xi,y):=u(X(\xi,y))=(\xi,V(\xi,y)). \tag*{(56)} X ( ξ , y ) = ( φ ( ξ , y ) , y ) , u ( ξ , y ) := u ( X ( ξ , y )) = ( ξ , V ( ξ , y )) . (56)
We call U : = ( φ , V ) U:=(\varphi,V) U := ( φ , V ) the transformed spinor . Since ξ ↦ − ξ \xi\mapsto-\xi ξ ↦ − ξ exchanges the two sheets of the original Z 2 \mathbb{Z}_2 Z 2 -harmonic spinor u u u , φ \varphi φ is even and V V V is odd in ξ \xi ξ . Moreover, φ ( 0 , y ) = g ( y ) \varphi(0,y)=g(y) φ ( 0 , y ) = g ( y ) and V ( 0 , y ) = 0 V(0,y)=0 V ( 0 , y ) = 0 , so the branching locus becomes { ξ = 0 } \{\xi=0\} { ξ = 0 } , and its regularity is encoded in the tangential regularity of φ \varphi φ along { ξ = 0 } \{\xi=0\} { ξ = 0 } .
Near each spine point ( 0 , y ) (0,y) ( 0 , y ) , we compare U U U with its transformed tangent model at p ( y ) p(y) p ( y ) , which is a polynomial with coefficients frozen at y y y .
Example 5.1 (The transformed tangent model). Fix p = p ( y ) p=p(y) p = p ( y ) , write X = ( z , y ′ ) X=(z,y') X = ( z , y ′ ) , and apply the same construction to Φ p \Phi_p Φ p , i.e. set ξ = π Φ p \xi=\pi\Phi_p ξ = π Φ p , so that Π P p ⊥ ( X − p ) = ( ( π T p ) − 1 ξ ) 2 \Pi_{P_p^\perp}(X-p)=((\pi T_p)^{-1}\xi)^2 Π P p ⊥ ( X − p ) = (( π T p ) − 1 ξ ) 2 . Since B u \mathcal{B}_u B u is C 1 , α C^{1,\alpha} C 1 , α , the restriction Π P p ⊥ ∣ P 0 ⊥ : P 0 ⊥ → P p ⊥ \Pi_{P_p^\perp}|_{P_0^\perp}:P_0^\perp\to P_p^\perp Π P p ⊥ ∣ P 0 ⊥ : P 0 ⊥ → P p ⊥ is invertible. Decomposing
X − p = ( D g ( y ) ( y ′ − y ) , y ′ − y ) + ( z − g ( y ) − D g ( y ) ( y ′ − y ) , 0 ) X-p=\bigl(Dg(y)(y'-y),y'-y\bigr)+\bigl(z-g(y)-Dg(y)(y'-y),0\bigr) X − p = ( D g ( y ) ( y ′ − y ) , y ′ − y ) + ( z − g ( y ) − D g ( y ) ( y ′ − y ) , 0 )
∈ P p \in P_p ∈ P p and applying Π P p ⊥ \Pi_{P_p^\perp} Π P p ⊥ , we express the original normal coordinate z z z in terms of ( ξ , y ′ ) (\xi,y') ( ξ , y ′ ) :
z = φ p ( ξ , y ′ ) = g ( y ) + D g ( y ) ( y ′ − y ) + ( Π P p ⊥ ∣ P 0 ⊥ ) − 1 ( ( π T p ) − 1 ξ ) 2 . (57) z=\varphi_p(\xi,y')=g(y)+Dg(y)(y'-y)+(\Pi_{P_p^\perp}|_{P_0^\perp})^{-1}((\pi T_p)^{-1}\xi)^2. \tag*{(57)} z = φ p ( ξ , y ′ ) = g ( y ) + D g ( y ) ( y ′ − y ) + ( Π P p ⊥ ∣ P 0 ⊥ ) − 1 (( π T p ) − 1 ξ ) 2 . (57)
The remaining V V V -component of the tangent model is
V p ( ξ , y ′ ) = π ⊥ T p ( π T p ) − 1 ξ = : M ( y ) ξ . (58) V_p(\xi,y')=\pi^\perp T_p(\pi T_p)^{-1}\xi=:M(y)\xi. \tag*{(58)} V p ( ξ , y ′ ) = π ⊥ T p ( π T p ) − 1 ξ =: M ( y ) ξ . (58)
By non-degeneracy, the homogeneous quadratic polynomial q ( y , ⋅ ) q(y,\cdot) q ( y , ⋅ ) satisfies c ∣ ξ ∣ ≤ ∣ D ξ q ( y , ξ ) ∣ ≤ C ∣ ξ ∣ c|\xi|\leq|D_\xi q(y,\xi)|\leq C|\xi| c ∣ ξ ∣ ≤ ∣ D ξ q ( y , ξ ) ∣ ≤ C ∣ ξ ∣ , uniformly in y y y . We call U p = ( φ p , V p ) U_p=(\varphi_p,V_p) U p = ( φ p , V p ) the transformed model at p p p . Its components φ p \varphi_p φ p and V p V_p V p are even and odd in ξ \xi ξ , respectively, with φ p ( 0 , y ′ ) = g ( y ) + D g ( y ) ( y ′ − y ) \varphi_p(0,y')=g(y)+Dg(y)(y'-y) φ p ( 0 , y ′ ) = g ( y ) + D g ( y ) ( y ′ − y ) and V p ( 0 , y ′ ) = 0 V_p(0,y')=0 V p ( 0 , y ′ ) = 0 .
In the variables ( ξ , y ′ ) (\xi,y') ( ξ , y ′ ) , it is an affine function of y ′ y' y ′ plus a homogeneous polynomial in ξ \xi ξ , whose coefficients g ( y ) g(y) g ( y ) , D g ( y ) Dg(y) D g ( y ) , q ( y , ⋅ ) q(y,\cdot) q ( y , ⋅ ) , M ( y ) M(y) M ( y ) are frozen at y y y . We shall use U p U_p U p only for fixed y y y , as a comparison model near ( 0 , y ) (0,y) ( 0 , y ) , and postpone the study of the dependence of the frozen coefficients on y y y to § 3.6 .
Since ∣ Π P p ⊥ ( X − p ) ∣ ∼ ∣ ξ ∣ 2 |\Pi_{P_p^\perp}(X-p)|\sim|\xi|^2 ∣ Π P p ⊥ ( X − p ) ∣ ∼ ∣ ξ ∣ 2 , the rescaling X ↦ r 2 X X\mapsto r^2X X ↦ r 2 X in the original coordinates corresponds to the anisotropic rescaling ( ξ , y ) ↦ ( r ξ , r 2 y ) (\xi,y)\mapsto(r\xi,r^2y) ( ξ , y ) ↦ ( r ξ , r 2 y ) in the new coordinates.
Let f ( X , u , D u ) f(X,u,Du) f ( X , u , D u ) be the J \mathcal{J} J -energy density of Section 2. For a field U = ( φ , V ) U=(\varphi,V) U = ( φ , V ) with det D ξ φ > 0 \det D_\xi\varphi>0 det D ξ φ > 0 away from { ξ = 0 } \{\xi=0\} { ξ = 0 } , define the partial Legendre transform by (56) . Changing variables in J \mathcal{J} J implies the transformed density
f ^ ( ξ , y , U , D U ) = ( det D ξ φ ) f ( X , u ^ , D u ^ ( D X ) − 1 ) , ξ ≠ 0. (59) \widehat{f}(\xi,y,U,DU)=(\det D_\xi\varphi)f(X,\widehat{u},D\widehat{u}(DX)^{-1}),\qquad\xi\ne0. \tag*{(59)} f ( ξ , y , U , D U ) = ( det D ξ φ ) f ( X , u , D u ( D X ) − 1 ) , ξ = 0. (59)
Example 5.2 (The Euclidean case). For the Dirichlet density 1 2 ∣ D u ∣ 2 \frac{1}{2}|Du|^2 2 1 ∣ D u ∣ 2 ,
f ^ 0 = det D ξ φ 2 ∣ D u ^ ( D X ) − 1 ∣ 2 , D u ^ ( D X ) − 1 = ( ( D ξ φ ) − 1 − ( D ξ φ ) − 1 D y φ D ξ V ( D ξ φ ) − 1 D y V − D ξ V ( D ξ φ ) − 1 D y φ ) . (60) \widehat{f}_0=\frac{\det D_\xi\varphi}{2}|D\widehat{u}(DX)^{-1}|^2,\qquad D\widehat{u}(DX)^{-1}=
\begin{pmatrix}
(D_\xi\varphi)^{-1} & -(D_\xi\varphi)^{-1}D_y\varphi\\
D_\xi V(D_\xi\varphi)^{-1} & D_yV-D_\xi V(D_\xi\varphi)^{-1}D_y\varphi
\end{pmatrix}. \tag*{(60)} f 0 = 2 det D ξ φ ∣ D u ( D X ) − 1 ∣ 2 , D u ( D X ) − 1 = ( ( D ξ φ ) − 1 D ξ V ( D ξ φ ) − 1 − ( D ξ φ ) − 1 D y φ D y V − D ξ V ( D ξ φ ) − 1 D y φ ) . (60)
Weighted spaces
We use the Euclidean structure in ( ξ , y ) (\xi,y) ( ξ , y ) for all integrals and weak equations. The degeneracy at { ξ = 0 } \{\xi=0\} { ξ = 0 } is encoded in the weighted derivatives and norms below. We write ρ = ∣ ξ ∣ \rho=|\xi| ρ = ∣ ξ ∣ and work on the anisotropic cylinders
Q r ( y 0 ) = B r 2 × B r 2 n − 2 ( y 0 ) , ∣ Q r ( y 0 ) ∣ = r N G ∣ Q 1 ∣ , N G = 2 n − 2. Q_r(y_0)=B_r^2\times B_{r^2}^{n-2}(y_0),\qquad|Q_r(y_0)|=r^{N_G}|Q_1|,\qquad N_G=2n-2. Q r ( y 0 ) = B r 2 × B r 2 n − 2 ( y 0 ) , ∣ Q r ( y 0 ) ∣ = r N G ∣ Q 1 ∣ , N G = 2 n − 2.
A field Z = ( Z φ , Z V ) Z=(Z_\varphi,Z_V) Z = ( Z φ , Z V ) is admissible if Z φ Z_\varphi Z φ is even and Z V Z_V Z V is odd in ξ \xi ξ . Its Grushin (weighted) derivatives are defined by
∇ G Z = ( ρ − 1 D ξ Z φ , D y Z φ , D ξ Z V , ρ D y Z V ) , (61) \nabla^G Z=(\rho^{-1}D_\xi Z_\varphi,D_yZ_\varphi,D_\xi Z_V,\rho D_yZ_V), \tag*{(61)} ∇ G Z = ( ρ − 1 D ξ Z φ , D y Z φ , D ξ Z V , ρ D y Z V ) , (61)
which are of degree zero under the rescaling
( S r Z ) ( ξ , y ) = ( r − 2 Z φ , r − 1 Z V ) ( r ξ , y 0 + r 2 y ) . (S_rZ)(\xi,y)=(r^{-2}Z_\varphi,r^{-1}Z_V)(r\xi,y_0+r^2y). ( S r Z ) ( ξ , y ) = ( r − 2 Z φ , r − 1 Z V ) ( r ξ , y 0 + r 2 y ) .
The L 2 L^2 L 2 -height excess on Q r ( y 0 ) Q_r(y_0) Q r ( y 0 ) is defined by
E r ( Z ) = r − N G − 4 ∫ Q r ( y 0 ) ∣ Z φ ∣ 2 + r − N G − 2 ∫ Q r ( y 0 ) ∣ Z V ∣ 2 , (62) \mathcal{E}_r(Z)=r^{-N_G-4}\int_{Q_r(y_0)}|Z_\varphi|^2+r^{-N_G-2}\int_{Q_r(y_0)}|Z_V|^2, \tag*{(62)} E r ( Z ) = r − N G − 4 ∫ Q r ( y 0 ) ∣ Z φ ∣ 2 + r − N G − 2 ∫ Q r ( y 0 ) ∣ Z V ∣ 2 , (62)
so that E r ( Z ) = E 1 ( S r Z ) \mathcal{E}_r(Z)=\mathcal{E}_1(S_rZ) E r ( Z ) = E 1 ( S r Z ) and ∫ Q 1 ∣ ∇ G S r Z ∣ 2 = r − N G ∫ Q r ( y 0 ) ∣ ∇ G Z ∣ 2 \int_{Q_1}|\nabla^G S_rZ|^2=r^{-N_G}\int_{Q_r(y_0)}|\nabla^G Z|^2 ∫ Q 1 ∣ ∇ G S r Z ∣ 2 = r − N G ∫ Q r ( y 0 ) ∣ ∇ G Z ∣ 2 . On Ω = Q r ( y 0 ) \Omega=Q_r(y_0) Ω = Q r ( y 0 ) , we define the adapted Sobolev norms as
∥ Z ∥ L G 2 ( Ω ) 2 = ∫ Ω ( ∣ Z φ ∣ 2 + ρ 2 ∣ Z V ∣ 2 ) , ∥ Z ∥ W G 1 , 2 ( Ω ) 2 = ∫ Ω ( ∣ Z φ ∣ 2 + ∣ Z V ∣ 2 + ∣ ∇ G Z ∣ 2 ) . (63) \|Z\|_{L_G^2(\Omega)}^2=\int_\Omega(|Z_\varphi|^2+\rho^2|Z_V|^2),\qquad\|Z\|_{W_G^{1,2}(\Omega)}^2=\int_\Omega(|Z_\varphi|^2+|Z_V|^2+|\nabla^GZ|^2). \tag*{(63)} ∥ Z ∥ L G 2 ( Ω ) 2 = ∫ Ω ( ∣ Z φ ∣ 2 + ρ 2 ∣ Z V ∣ 2 ) , ∥ Z ∥ W G 1 , 2 ( Ω ) 2 = ∫ Ω ( ∣ Z φ ∣ 2 + ∣ Z V ∣ 2 + ∣ ∇ G Z ∣ 2 ) . (63)
Write W G 1 , 2 ( Ω ) W_G^{1,2}(\Omega) W G 1 , 2 ( Ω ) for the completion of the smooth admissible fields under the norm above, and W G , 0 1 , 2 ( Ω ) W_{G,0}^{1,2}(\Omega) W G , 0 1 , 2 ( Ω ) for the closure of those compactly supported ones. Here, the trace Z φ ( 0 , y ) Z_\varphi(0,y) Z φ ( 0 , y ) need not vanish.
We shall study the second variation of the transformed J \mathcal{J} J -energy ∫ f ^ ( ξ , y , U , D U ) \int\widehat{f}(\xi,y,U,DU) ∫ f ( ξ , y , U , D U ) . For this, we introduce the following
A = ( A 1 , A 0 ) = ( D D U f ^ , D U f ^ ) , A U Z = d d t ∣ t = 0 A ( ξ , y , U + t Z , D U + t D Z ) , (64) \mathbf{A}=(A^1,A^0)=(D_{DU}\widehat{f},D_U\widehat{f}),\qquad\mathbf{A}_U Z=\left.\frac{d}{dt}\right|_{t=0}\mathbf{A}(\xi,y,U+tZ,DU+tDZ), \tag*{(64)} A = ( A 1 , A 0 ) = ( D D U f , D U f ) , A U Z = d t d t = 0 A ( ξ , y , U + tZ , D U + t D Z ) , (64)
with the associated bilinear form
B U ( Z , Y ) = ∫ Ω ( ( A U Z ) 1 : D Y + ( A U Z ) 0 ⋅ Y ) . (65) \mathcal{B}_{U}(Z,Y)=\int_{\Omega}\left((\mathbf{A}_{U}Z)^{1}:DY+(\mathbf{A}_{U}Z)^{0}\cdot Y\right). \tag*{(65)} B U ( Z , Y ) = ∫ Ω ( ( A U Z ) 1 : D Y + ( A U Z ) 0 ⋅ Y ) . (65)
for admissible Z ∈ W G 1 , 2 ( Ω ) Z\in W_{G}^{1,2}(\Omega) Z ∈ W G 1 , 2 ( Ω ) and smooth admissible Y Y Y compactly supported in Ω \Omega Ω . Here A 1 \mathbf{A}^{1} A 1 and A 0 \mathbf{A}^{0} A 0 have the same structure as D U DU D U and U U U , respectively, with the block decomposition A 1 : D Y = A φ , ξ 1 : D ξ Y φ + A φ , y 1 : D y Y φ + A V , ξ 1 : D ξ Y V + A V , y 1 : D y Y V \mathbf{A}^{1}:DY=\mathbf{A}_{\varphi,\xi}^{1}:D_{\xi}Y_{\varphi}+\mathbf{A}_{\varphi,y}^{1}:D_{y}Y_{\varphi}+\mathbf{A}_{V,\xi}^{1}:D_{\xi}Y_{V}+\mathbf{A}_{V,y}^{1}:D_{y}Y_{V} A 1 : D Y = A φ , ξ 1 : D ξ Y φ + A φ , y 1 : D y Y φ + A V , ξ 1 : D ξ Y V + A V , y 1 : D y Y V and A 0 ⋅ Y = A φ 0 ⋅ Y φ + A V 0 ⋅ Y V \mathbf{A}^{0}\cdot Y=\mathbf{A}_{\varphi}^{0}\cdot Y_{\varphi}+\mathbf{A}_{V}^{0}\cdot Y_{V} A 0 ⋅ Y = A φ 0 ⋅ Y φ + A V 0 ⋅ Y V . The weights dual to (61) are
A G 1 = ( ρ A φ , ξ 1 , A φ , y 1 , A V , ξ 1 , ρ − 1 A V , y 1 ) , so that A 1 : D Y = A G 1 ⋅ ∇ G Y . \mathbf{A}_{G}^{1}=\left(\rho\mathbf{A}_{\varphi,\xi}^{1},\mathbf{A}_{\varphi,y}^{1},\mathbf{A}_{V,\xi}^{1},\rho^{-1}\mathbf{A}_{V,y}^{1}\right),
\qquad\text{so that}\qquad
\mathbf{A}^{1}:DY=\mathbf{A}_{G}^{1}\cdot\nabla^{G}Y. A G 1 = ( ρ A φ , ξ 1 , A φ , y 1 , A V , ξ 1 , ρ − 1 A V , y 1 ) , so that A 1 : D Y = A G 1 ⋅ ∇ G Y .
For such coefficient fields A = ( A 1 , A 0 ) \mathbf{A}=(\mathbf{A}^{1},\mathbf{A}^{0}) A = ( A 1 , A 0 ) , we use script letters for the corresponding functionals and introduce the dual weighted (semi-)norm
A ( Y ) = ∫ Ω ( A 1 : D Y + A 0 ⋅ Y ) , ∥ A ∥ ( W G 1 , 2 ( Ω ) ) ∗ = sup Y ∈ C c ∞ ( Ω ) a d m i s s i b l e ∥ ∇ G Y ∥ L 2 ( Ω ) ≤ 1 ∣ A ( Y ) ∣ . (66) \mathcal{A}(Y)=\int_{\Omega}\left(\mathbf{A}^{1}:DY+\mathbf{A}^{0}\cdot Y\right),
\qquad
\lVert\mathbf{A}\rVert_{(W_{G}^{1,2}(\Omega))^{*}}
=
\sup_{\substack{Y\in C_{c}^{\infty}(\Omega)\ \mathrm{admissible}\\ \lVert\nabla^{G}Y\rVert_{L^{2}(\Omega)}\leq1}}
\left|\mathcal{A}(Y)\right|. \tag*{(66)} A ( Y ) = ∫ Ω ( A 1 : D Y + A 0 ⋅ Y ) , ∥ A ∥ ( W G 1 , 2 ( Ω ) ) ∗ = Y ∈ C c ∞ ( Ω ) admissible ∥ ∇ G Y ∥ L 2 ( Ω ) ≤ 1 sup ∣ A ( Y ) ∣ . (66)
The following elementary inequalities are used throughout Part II.
Lemma 5.3 (Weighted inequalities). There hold:
(i) Every admissible field Z Z Z on Q r ( y 0 ) Q_{r}(y_{0}) Q r ( y 0 ) satisfies
∫ ρ − 2 ∣ Z V ∣ 2 ≤ ∫ ∣ D ξ Z V ∣ 2 . (67) \int\rho^{-2}|Z_{V}|^{2}\leq\int|D_{\xi}Z_{V}|^{2}. \tag*{(67)} ∫ ρ − 2 ∣ Z V ∣ 2 ≤ ∫ ∣ D ξ Z V ∣ 2 . (67)
If further Z Z Z is compactly supported in Q r ( y 0 ) Q_{r}(y_{0}) Q r ( y 0 ) , then
∥ Z φ ∥ L 2 ≤ C r 2 ∥ ∇ G Z ∥ L 2 , ∥ Z φ ( 0 , ⋅ ) ∥ L y 2 ≤ C r ∥ ∇ G Z ∥ L 2 . (68) \lVert Z_{\varphi}\rVert_{L^{2}}\leq Cr^{2}\lVert\nabla^{G}Z\rVert_{L^{2}},
\qquad
\lVert Z_{\varphi}(0,\cdot)\rVert_{L_{y}^{2}}\leq Cr\lVert\nabla^{G}Z\rVert_{L^{2}}. \tag*{(68)} ∥ Z φ ∥ L 2 ≤ C r 2 ∥ ∇ G Z ∥ L 2 , ∥ Z φ ( 0 , ⋅ ) ∥ L y 2 ≤ C r ∥ ∇ G Z ∥ L 2 . (68)
(ii) If ∣ A G 1 ∣ ≤ C 1 |\mathbf{A}_{G}^{1}|\leq C_{1} ∣ A G 1 ∣ ≤ C 1 , ∣ A φ 0 ∣ ≤ C 0 |\mathbf{A}_{\varphi}^{0}|\leq C_{0} ∣ A φ 0 ∣ ≤ C 0 and ∣ A V 0 ∣ ≤ C 0 ρ |\mathbf{A}_{V}^{0}|\leq C_{0}\rho ∣ A V 0 ∣ ≤ C 0 ρ on Q r ( y 0 ) Q_{r}(y_{0}) Q r ( y 0 ) , then
∥ A ∥ ( W G 1 , 2 ( Q r ( y 0 ) ) ) ∗ ≤ C r N G / 2 ( C 1 + r 2 C 0 ) . (69) \lVert\mathbf{A}\rVert_{(W_{G}^{1,2}(Q_{r}(y_{0})))^{*}}\leq Cr^{N_{G}/2}(C_{1}+r^{2}C_{0}). \tag*{(69)} ∥ A ∥ ( W G 1 , 2 ( Q r ( y 0 )) ) ∗ ≤ C r N G /2 ( C 1 + r 2 C 0 ) . (69)
In particular, ∥ Z ∥ W G 1 , 2 2 \lVert Z\rVert_{W_{G}^{1,2}}^{2} ∥ Z ∥ W G 1 , 2 2 is uniformly equivalent to ∫ ( ∣ Z φ ∣ 2 + ∣ ∇ G Z ∣ 2 ) \int(|Z_{\varphi}|^{2}+|\nabla^{G}Z|^{2}) ∫ ( ∣ Z φ ∣ 2 + ∣ ∇ G Z ∣ 2 ) , and to ∫ ∣ ∇ G Z ∣ 2 \int|\nabla^{G}Z|^{2} ∫ ∣ ∇ G Z ∣ 2 on W G , 0 1 , 2 W_{G,0}^{1,2} W G , 0 1 , 2 ; the trace Z ↦ Z φ ( 0 , ⋅ ) Z\mapsto Z_{\varphi}(0,\cdot) Z ↦ Z φ ( 0 , ⋅ ) is bounded from W G 1 , 2 ( Q r ( y 0 ) ) W_{G}^{1,2}(Q_{r}(y_{0})) W G 1 , 2 ( Q r ( y 0 )) to L 2 ( B r n − 2 n − 2 ( y 0 ) ) L^{2}(B_{r^{n-2}}^{n-2}(y_{0})) L 2 ( B r n − 2 n − 2 ( y 0 )) .
Proof. (i) On each normal circle, an odd function has only odd Fourier modes, so Wirtinger’s inequality implies
∫ 0 2 π ∣ Z V ∣ 2 d θ ≤ ∫ 0 2 π ∣ ∂ θ Z V ∣ 2 d θ ≤ ρ 2 ∫ 0 2 π ∣ D ξ Z V ∣ 2 d θ . \int_{0}^{2\pi}|Z_{V}|^{2}\,d\theta
\leq\int_{0}^{2\pi}|\partial_{\theta}Z_{V}|^{2}\,d\theta
\leq\rho^{2}\int_{0}^{2\pi}|D_{\xi}Z_{V}|^{2}\,d\theta. ∫ 0 2 π ∣ Z V ∣ 2 d θ ≤ ∫ 0 2 π ∣ ∂ θ Z V ∣ 2 d θ ≤ ρ 2 ∫ 0 2 π ∣ D ξ Z V ∣ 2 d θ .
Integrating in ρ d ρ d y \rho\,d\rho\,dy ρ d ρ d y proves (67) . The first bound in (68) is the standard Poincaré inequality in y y y on B r n − 2 n − 2 ( y 0 ) B_{r^{n-2}}^{n-2}(y_{0}) B r n − 2 n − 2 ( y 0 ) . To prove Z φ ( 0 , ⋅ ) ∈ L y 2 Z_{\varphi}(0,\cdot)\in L_{y}^{2} Z φ ( 0 , ⋅ ) ∈ L y 2 , we integrate along normal rays from ρ = r \rho=r ρ = r , where Z Z Z vanishes:
∣ Z φ ( 0 , y ) ∣ 2 = ∣ ∫ 0 r ∂ ρ Z φ d ρ ∣ 2 ≤ r 2 2 ∫ 0 r ρ − 1 ∣ ∂ ρ Z φ ∣ 2 d ρ , |Z_{\varphi}(0,y)|^{2}
=\left|\int_{0}^{r}\partial_{\rho}Z_{\varphi}\,d\rho\right|^{2}
\leq\frac{r^{2}}{2}\int_{0}^{r}\rho^{-1}|\partial_{\rho}Z_{\varphi}|^{2}\,d\rho, ∣ Z φ ( 0 , y ) ∣ 2 = ∫ 0 r ∂ ρ Z φ d ρ 2 ≤ 2 r 2 ∫ 0 r ρ − 1 ∣ ∂ ρ Z φ ∣ 2 d ρ ,
and average in θ \theta θ and integrate in y y y . Then (ii) follows from (i) and Cauchy–Schwarz. □
Weighted Hölder norms. Although the equations degenerate along { ξ = 0 } \{\xi=0\} { ξ = 0 } , they are uniformly elliptic on balls whose radii are proportional to their distance from this set, with ellipticity constants independent of that distance. More precisely, fix c 0 ∈ ( 0 , 1 / 100 ) c_{0}\in(0,1/100) c 0 ∈ ( 0 , 1/100 ) . For each ( ξ , y ) ∈ Ω (\xi,y)\in\Omega ( ξ , y ) ∈ Ω with ρ = ∣ ξ ∣ > 0 \rho=|\xi|>0 ρ = ∣ ξ ∣ > 0 , we define Q ξ , y = ( B c 0 ρ 2 ( ξ ) × B c 0 2 ρ n − 2 ( y ) ) ∩ Ω Q_{\xi,y}=(B_{c_{0}\rho}^{2}(\xi)\times B_{c_{0}^{2}\rho}^{n-2}(y))\cap\Omega Q ξ , y = ( B c 0 ρ 2 ( ξ ) × B c 0 2 ρ n − 2 ( y )) ∩ Ω , on which ∣ ξ ∣ |\xi| ∣ ξ ∣ is comparable to ρ \rho ρ .
We measure fields Z Z Z , degree-zero quantities F F F (such as the blocks of ∇ G Z \nabla^{G}Z ∇ G Z ), and coefficient fields A \mathbf{A} A by adapted Hölder norms
∥ Z ∥ C G k , α ( Ω ) = sup ∥ ( ρ − 2 Z φ , ρ − 1 Z V ) ∥ C k , α ( Q ξ , y , g ρ ) , ∥ F ∥ C G k , α ( Ω ) = sup ∥ F ∥ C k , α ( Q ξ , y , g ρ ) , ∥ A ∥ C G k , α ( Ω ) = sup ∥ ( A G 1 , ρ 2 A φ 0 , ρ A V 0 ) ∥ C k , α ( Q ξ , y , g ρ ) . (70) \begin{aligned}
\lVert Z\rVert_{C_{G}^{k,\alpha}(\Omega)}
&=\sup\left\lVert(\rho^{-2}Z_{\varphi},\rho^{-1}Z_{V})\right\rVert_{C^{k,\alpha}(Q_{\xi,y},g_{\rho})},\\
\lVert F\rVert_{C_{G}^{k,\alpha}(\Omega)}
&=\sup\lVert F\rVert_{C^{k,\alpha}(Q_{\xi,y},g_{\rho})},\\
\lVert\mathbf{A}\rVert_{C_{G}^{k,\alpha}(\Omega)}
&=\sup\left\lVert(\mathbf{A}_{G}^{1},\rho^{2}\mathbf{A}_{\varphi}^{0},\rho\mathbf{A}_{V}^{0})\right\rVert_{C^{k,\alpha}(Q_{\xi,y},g_{\rho})}. \tag*{(70)}
\end{aligned} ∥ Z ∥ C G k , α ( Ω ) ∥ F ∥ C G k , α ( Ω ) ∥ A ∥ C G k , α ( Ω ) = sup ( ρ − 2 Z φ , ρ − 1 Z V ) C k , α ( Q ξ , y , g ρ ) , = sup ∥ F ∥ C k , α ( Q ξ , y , g ρ ) , = sup ( A G 1 , ρ 2 A φ 0 , ρ A V 0 ) C k , α ( Q ξ , y , g ρ ) . (70)
Here each supremum is over ( ξ , y ) ∈ Ω (\xi,y) \in\Omega ( ξ , y ) ∈ Ω with ρ = ∣ ξ ∣ > 0 \rho= |\xi| > 0 ρ = ∣ ξ ∣ > 0 , and the C k , α C^{k,\alpha} C k , α -norms are taken with respect to the rescaled product metric g ρ = ρ − 2 g ξ + ρ − 2 g y g_\rho= \rho^{-2}g_\xi+ \rho^{-2}g_y g ρ = ρ − 2 g ξ + ρ − 2 g y .
Comparison polynomials. The transformed models (57) –(58) lie in the finite-dimensional space with norms
P = { P = ( a + B ( y − y 0 ) + q ( ξ ) , M ξ ) } , ∥ ∇ P ∥ = ∣ B ∣ + ∣ q ∣ + ∣ M ∣ , ∥ P ∥ r = r − 2 ∣ a ∣ + ∥ ∇ P ∥ , (71) \mathcal{P} = \{P = (a + B(y-y_0) + q(\xi),M\xi)\},
\qquad
\lVert\nabla P\rVert= |B| + |q| + |M|,
\qquad
\lVert P\rVert_r = r^{-2}|a| + \lVert\nabla P\rVert, \tag*{(71)} P = { P = ( a + B ( y − y 0 ) + q ( ξ ) , M ξ )} , ∥ ∇ P ∥ = ∣ B ∣ + ∣ q ∣ + ∣ M ∣ , ∥ P ∥ r = r − 2 ∣ a ∣ + ∥ ∇ P ∥ , (71)
where a ∈ C a \in\mathbb{C} a ∈ C , B : R n − 2 → C B : \mathbb{R}^{n-2} \to\mathbb{C} B : R n − 2 → C and M : C → im π ⊥ M : \mathbb{C} \to\operatorname{im}\pi^\perp M : C → im π ⊥ are real-linear, and q q q is a homogeneous quadratic polynomial. Denote the norm of the coefficients of q q q by ∣ q ∣ |q| ∣ q ∣ . Since ∇ G P = ( ρ − 1 D ξ q , B , M , 0 ) \nabla^G P = (\rho^{-1}D_\xi q,B,M,0) ∇ G P = ( ρ − 1 D ξ q , B , M , 0 ) and S r P = ( r − 2 a + B y + q ( ξ ) , M ξ ) S_rP = (r^{-2}a+By+q(\xi),M\xi) S r P = ( r − 2 a + B y + q ( ξ ) , M ξ ) ,
∥ ∇ P ∥ ∼ ∥ ∇ G P ∥ L 2 ( Q 1 ( y 0 ) ) , ∥ P ∥ r ∼ E r ( P ) 1 / 2 , ∀ P ∈ P . (72) \lVert\nabla P\rVert\sim\lVert\nabla^G P\rVert_{L^2(Q_1(y_0))},
\qquad
\lVert P\rVert_r \sim\mathcal{E}_r(P)^{1/2},
\qquad
\forall P \in\mathcal{P}. \tag*{(72)} ∥ ∇ P ∥ ∼ ∥ ∇ G P ∥ L 2 ( Q 1 ( y 0 )) , ∥ P ∥ r ∼ E r ( P ) 1/2 , ∀ P ∈ P . (72)
For a field Z Z Z defined near ( 0 , y 0 ) (0,y_0) ( 0 , y 0 ) , we call P ∈ P P \in\mathcal{P} P ∈ P the comparison polynomial of Z Z Z at ( 0 , y 0 ) (0,y_0) ( 0 , y 0 ) if E r ( Z − P ) → 0 \mathcal{E}_r(Z-P) \to0 E r ( Z − P ) → 0 as r → 0 r \to0 r → 0 . It is unique if it exists: by (72) , two such polynomials satisfy ∥ P − P ′ ∥ r → 0 \lVert P-P'\rVert_r \to0 ∥ P − P ′ ∥ r → 0 , so all coefficients of P − P ′ P-P' P − P ′ vanish.
Initial estimates
The quantitative estimates of Part I translate into the following estimates in Legendre coordinates:
Proposition 5.4 (Initial estimates in Legendre coordinates). For r 0 ≪ 1 r_0 \ll1 r 0 ≪ 1 ,
The partial Legendre transform ( z , y ) ↦ ( π u ( z , y ) , y ) (z,y) \mapsto(\pi u(z,y),y) ( z , y ) ↦ ( π u ( z , y ) , y ) is well-defined.
For every ( ξ , y ) (\xi,y) ( ξ , y ) with 0 < ρ = ∣ ξ ∣ ≤ r 0 0 < \rho= |\xi| \le r_0 0 < ρ = ∣ ξ ∣ ≤ r 0 and ∣ y − y 0 ∣ ≤ r 0 |y-y_0| \le r_0 ∣ y − y 0 ∣ ≤ r 0 ,
∥ ( ρ − 2 ( φ − φ p ( y ) ) , ρ − 1 ( V − V p ( y ) ) ) ∥ C 1 , α ( Q ξ , y , g ρ ) ≤ C ρ α . (73) \left\lVert\left(\rho^{-2}(\varphi-\varphi_{p(y)}),\rho^{-1}(V-V_{p(y)})\right)\right\rVert_{C^{1,\alpha}(Q_{\xi,y},g_\rho)} \le C\rho^\alpha. \tag*{(73)} ( ρ − 2 ( φ − φ p ( y ) ) , ρ − 1 ( V − V p ( y ) ) ) C 1 , α ( Q ξ , y , g ρ ) ≤ C ρ α . (73) For every branch point p 0 = p ( y 0 ) p_0=p(y_0) p 0 = p ( y 0 ) and 0 < r ≤ r 0 0<r\le r_0 0 < r ≤ r 0 ,
sup Q r ( y 0 ) ∖ { ξ = 0 } ( r − 2 ∣ φ − φ p 0 ∣ + r − 1 ∣ V − V p 0 ∣ + ∣ ∇ G ( U − U p 0 ) ∣ ) ≤ C r α . (74) \sup_{Q_r(y_0)\setminus\{\xi=0\}}
\left(r^{-2}|\varphi-\varphi_{p_0}|+r^{-1}|V-V_{p_0}|+|\nabla^G(U-U_{p_0})|\right)
\le Cr^\alpha. \tag*{(74)} Q r ( y 0 ) ∖ { ξ = 0 } sup ( r − 2 ∣ φ − φ p 0 ∣ + r − 1 ∣ V − V p 0 ∣ + ∣ ∇ G ( U − U p 0 ) ∣ ) ≤ C r α . (74) In particular, the comparison polynomial of U U U at ( 0 , y ) (0,y) ( 0 , y ) is the transformed model U p ( y ) U_{p(y)} U p ( y ) , with estimates
E r ( U − U p 0 ) ≤ C r 2 α . (75) \mathcal{E}_r(U-U_{p_0}) \le Cr^{2\alpha}. \tag*{(75)} E r ( U − U p 0 ) ≤ C r 2 α . (75)
The transformed equation
We now determine the structure of the transformed density (59) in the weighted variables, and derive the weak equation.
Lemma 5.5 (The transformed equation). Consider fields U = ( φ , V ) U=(\varphi,V) U = ( φ , V ) with uniformly bounded normalised data, i.e. ( X , u ^ / ρ , ∇ G U ) (X,\widehat{u}/\rho,\nabla^G U) ( X , u / ρ , ∇ G U ) is uniformly bounded on { ξ ≠ 0 } \{\xi\ne0\} { ξ = 0 } and ρ − 1 D ξ φ \rho^{-1}D_\xi\varphi ρ − 1 D ξ φ is uniformly invertible.
The transformed density has the form
f ^ = f ^ p r ( X , ∇ G U ) + ρ 2 Err ( ρ , X , u ^ / ρ , ∇ G U ) , (76) \widehat{f}=\widehat{f}_{\mathrm{pr}}(X,\nabla^G U)+\rho^2\operatorname{Err}(\rho,X,\widehat{u}/\rho,\nabla^G U), \tag*{(76)} f = f pr ( X , ∇ G U ) + ρ 2 Err ( ρ , X , u / ρ , ∇ G U ) , (76) where f ^ p r \widehat{f}_{\mathrm{pr}} f pr is independent of u ^ / ρ \widehat{u}/\rho u / ρ . Both f ^ p r \widehat{f}_{\mathrm{pr}} f pr and Err \operatorname{Err} Err are smooth (resp. real-analytic) functions of ρ \rho ρ and the normalised data for smooth (resp. real-analytic) ambient data.
The coefficient field A \mathbf{A} A in (64) satisfies
∣ A G 1 ∣ + ∣ A φ 0 ∣ + ρ − 1 ∣ A V 0 ∣ ≤ C . (77) |A^1_G|+|A^0_\varphi|+\rho^{-1}|A^0_V|\le C. \tag*{(77)} ∣ A G 1 ∣ + ∣ A φ 0 ∣ + ρ − 1 ∣ A V 0 ∣ ≤ C . (77) The transformed spinor U U U and its transformed models U p U_p U p have uniformly bounded normalised data, with bounds independent of p p p , and U U U satisfies the weak equation
∫ ( A 1 : D Y + A 0 ⋅ Y ) = 0 (78) \int(A^1:DY+A^0\cdot Y)=0 \tag*{(78)} ∫ ( A 1 : D Y + A 0 ⋅ Y ) = 0 (78) for every smooth admissible Y Y Y compactly supported in Q r 0 ( y 0 ) Q_{r_0}(y_0) Q r 0 ( y 0 ) .
Proof. (i) In a smooth trivialisation, f ( X , u , D u ) f(X,u,Du) f ( X , u , D u ) is the sum of a quadratic form in D u Du D u , a bilinear term in ( u , D u ) (u,Du) ( u , D u ) , and a quadratic form in u u u , with smooth coefficients in X X X (cf. the J \mathcal{J} J -energy (7) ). Writing D ξ φ = ρ ⋅ ρ − 1 D ξ φ D_{\xi}\varphi=\rho\cdot\rho^{-1}D_{\xi}\varphi D ξ φ = ρ ⋅ ρ − 1 D ξ φ in (60) , each block of D u = D u ^ ( D X ) − 1 Du=D\widehat{u}(DX)^{-1} D u = D u ( D X ) − 1 is ρ − 1 \rho^{-1} ρ − 1 times a smooth function of ∇ G U \nabla_G U ∇ G U , and det D ξ φ = ρ 2 det ( ρ − 1 D ξ φ ) \det D_{\xi}\varphi=\rho^2\det(\rho^{-1}D_{\xi}\varphi) det D ξ φ = ρ 2 det ( ρ − 1 D ξ φ ) . Then writing u ^ = ρ ⋅ u / ρ \widehat{u}=\rho\cdot u/\rho u = ρ ⋅ u / ρ , the first term contributes f ^ p r ( X , ∇ G U ) \widehat{f}_{\mathrm{pr}}(X,\nabla_G U) f pr ( X , ∇ G U ) , the transform of the quadratic form in D u Du D u frozen at X X X (for the Euclidean metric, f ^ p r \widehat{f}_{\mathrm{pr}} f pr is the density f ^ 0 \widehat{f}_0 f 0 in (60) , a smooth function of ∇ G U \nabla_G U ∇ G U ), with O ( ρ 2 ) O(\rho^2) O ( ρ 2 ) -remainders from the other parts.
(ii) Since f ^ \widehat{f} f depends on D U DU D U only through ∇ G U \nabla_G U ∇ G U , we have A G 1 = D ∇ G U f ^ A^1_G=D_{\nabla_G U}\widehat{f} A G 1 = D ∇ G U f , which is bounded. The field φ \varphi φ enters only through X X X , so A φ 0 A^0_{\varphi} A φ 0 is bounded; V V V enters only through u ^ / ρ \widehat{u}/\rho u / ρ in the O ( ρ 2 ) O(\rho^2) O ( ρ 2 ) -remainder of (76) , so A V 0 = O ( ρ ) A^0_V=O(\rho) A V 0 = O ( ρ ) .
(iii) The first assertion follows from the initial bounds (74) , since ρ − 1 D ξ φ p = ρ − 1 D ξ q \rho^{-1}D_{\xi}\varphi_p=\rho^{-1}D_{\xi}q ρ − 1 D ξ φ p = ρ − 1 D ξ q is uniformly invertible. To derive the weak equation, consider in the original coordinates the combined domain and target variation
u t ( Ψ t ( z , y ) ) = u ( z , y ) + t ( 0 , Y V ( π u , y ) ) , Ψ t ( z , y ) = ( z + t Y φ ( π u , y ) , y ) . u_t(\Psi_t(z,y))=u(z,y)+t(0,Y_V(\pi_u,y)),\qquad\Psi_t(z,y)=(z+tY_{\varphi}(\pi_u,y),y). u t ( Ψ t ( z , y )) = u ( z , y ) + t ( 0 , Y V ( π u , y )) , Ψ t ( z , y ) = ( z + t Y φ ( π u , y ) , y ) .
We check that this is an admissible variation in the sense of Remark 2.3:
(inner variation) Since Y φ Y_{\varphi} Y φ is even in ξ \xi ξ , Ψ t \Psi_t Ψ t is single-valued. Moreover, D ξ Y φ = O ( ρ ) D_{\xi}Y_{\varphi}=O(\rho) D ξ Y φ = O ( ρ ) and D X ξ = O ( ρ − 1 ) D_X\xi=O(\rho^{-1}) D X ξ = O ( ρ − 1 ) guarantees that for ∣ t ∣ ≪ 1 |t|\ll1 ∣ t ∣ ≪ 1 small, Ψ t \Psi_t Ψ t is a bi-Lipschitz homeomorphism, equal to the identity outside a compact set.
(outer variation) Since Y V Y_V Y V is odd, u t u_t u t is a two-valued section.
The variation preserves π u \pi_u π u , so ( ξ , y ) = ( π u t ∘ Ψ t , y ) = ( π u , y ) (\xi,y)=(\pi_{u_t}\circ\Psi_t,y)=(\pi_u,y) ( ξ , y ) = ( π u t ∘ Ψ t , y ) = ( π u , y ) are Legendre coordinates for u t u_t u t , in which the transformed field is U t = U + t Y U_t=U+tY U t = U + t Y . Therefore, the inner and outer variation identities of Remark 2.3 apply to u t u_t u t , and by (59) , the local J \mathcal{J} J -minimality of u u u yields
0 = d d t ∣ t = 0 ∫ f ^ ( ξ , y , U + t Y , D U + t D Y ) = ∫ ( A 1 : D Y + A 0 ⋅ Y ) . 0=\left.\frac{d}{dt}\right|_{t=0}\int\widehat{f}(\xi,y,U+tY,DU+tDY)=\int(A^1:DY+A^0\cdot Y). 0 = d t d t = 0 ∫ f ( ξ , y , U + t Y , D U + t D Y ) = ∫ ( A 1 : D Y + A 0 ⋅ Y ) .
Let A p r ( X , D U ) = ( D D U f ^ p r ( X , ∇ G U ) , 0 ) A_{\mathrm{pr}}(X,DU)=(D_{DU}\widehat{f}_{\mathrm{pr}}(X,\nabla_G U),0) A pr ( X , D U ) = ( D D U f pr ( X , ∇ G U ) , 0 ) be the coefficient field of the principal part of (76) .
The frozen operator
In Section 7, we shall treat the second variation B U \mathcal{B}_U B U as a perturbation of the second variation at a transformed model, with coefficients frozen at its centre. In this section we study this frozen second variation: we prove its coercivity and an excess decay for its weak solutions (cf. Proposition 6.1), and construct the polynomial corrections (cf. Lemma 6.5) used in the Campanato estimate (cf. Theorem 7.4) to absorb the O ( 1 ) O(1) O ( 1 ) principal parts of the source term.
Fix a branch point p 0 = p ( y 0 ) p_0=p(y_0) p 0 = p ( y 0 ) . Freezing the principal part of J \mathcal{J} J at p 0 p_0 p 0 yields the density
f p 0 ( D u ) = 1 2 det g i j ( p 0 ) ∣ D u ∣ g i j ( p 0 ) , h ( p 0 ) 2 , f_{p_0}(Du)=\frac{1}{2}\sqrt{\det g_{ij}(p_0)}\lvert Du\rvert^2_{g_{ij}(p_0),h(p_0)}, f p 0 ( D u ) = 2 1 det g ij ( p 0 ) ∣ D u ∣ g ij ( p 0 ) , h ( p 0 ) 2 ,
where g i j ( p 0 ) , h ( p 0 ) g_{ij}(p_0),h(p_0) g ij ( p 0 ) , h ( p 0 ) are the Riemannian metric and the spinor bundle metric at p 0 p_0 p 0 . We write f ^ p 0 ( D U ) \widehat{f}_{p_0}(DU) f p 0 ( D U ) for its transform (59) ; it depends only on D U DU D U , and f ^ p 0 ( D U ) = f ^ p r ( p 0 , ∇ G U ) \widehat{f}_{p_0}(DU)=\widehat{f}_{\mathrm{pr}}(p_0,\nabla_G U) f p 0 ( D U ) = f pr ( p 0 , ∇ G U ) . By Theorem 2.6, Φ p 0 \Phi_{p_0} Φ p 0 minimises the energy of f p 0 f_{p_0} f p 0 , so U p 0 U_{p_0} U p 0 is a critical point of ∫ f ^ p 0 \int\widehat{f}_{p_0} ∫ f p 0 . The frozen linearised coefficient field and second variation are
A p r , p 0 Z = ( D 2 f ^ p 0 ( D U p 0 ) ( D Z ) , 0 ) = D D U A p r ( p 0 , D U p 0 ) ( D Z ) , A_{\mathrm{pr},p_0}Z=(D^2\widehat{f}_{p_0}(DU_{p_0})(DZ),0)=D_{DU}A_{\mathrm{pr}}(p_0,DU_{p_0})(DZ), A pr , p 0 Z = ( D 2 f p 0 ( D U p 0 ) ( D Z ) , 0 ) = D D U A pr ( p 0 , D U p 0 ) ( D Z ) ,
B p 0 ( Z , Y ) = ( A p 0 Z ) ( Y ) = ∫ D 2 f ^ p 0 ( D U p 0 ) ( D Z , D Y ) . (79) \mathcal{B}_{p_0}(Z,Y)=(\mathcal{A}_{p_0}Z)(Y)=\int D^2\widehat{f}_{p_0}(DU_{p_0})(DZ,DY). \tag*{(79)} B p 0 ( Z , Y ) = ( A p 0 Z ) ( Y ) = ∫ D 2 f p 0 ( D U p 0 ) ( D Z , D Y ) . (79)
Proposition 6.1 (The frozen operator). There are constants c , C > 0 c,C>0 c , C > 0 and θ 0 ∈ ( 0 , 1 ) \theta_0\in(0,1) θ 0 ∈ ( 0 , 1 ) , uniform in p 0 p_0 p 0 , with the following properties.
(i) (Coercivity) For every Z ∈ W G , 0 1 , 2 Z\in W^{1,2}_{G,0} Z ∈ W G , 0 1 , 2 ,
c ∥ ∇ G Z ∥ L 2 2 ≤ B p 0 ( Z , Z ) ≤ C ∥ ∇ G Z ∥ L 2 2 . (80) c\lVert\nabla_G Z\rVert_{L^2}^2\leq\mathcal{B}_{p_0}(Z,Z)\leq C\lVert\nabla_G Z\rVert_{L^2}^2. \tag*{(80)} c ∥ ∇ G Z ∥ L 2 2 ≤ B p 0 ( Z , Z ) ≤ C ∥ ∇ G Z ∥ L 2 2 . (80)
(ii) (Excess decay ) Let Z ∈ W G , l o c 1 , 2 ( Q 1 ( y 0 ) ) Z \in W^{1,2}_{G,\mathrm{loc}}(Q_1(y_0)) Z ∈ W G , loc 1 , 2 ( Q 1 ( y 0 )) satisfy E 1 1 ( Z ) < ∞ \mathcal{E}_1^1(Z)<\infty E 1 1 ( Z ) < ∞ and B p 0 ( Z , Y ) = 0 \mathcal{B}_{p_0}(Z,Y)=0 B p 0 ( Z , Y ) = 0 for all compactly supported admissible Y Y Y . Then there is a comparison polynomial H ∈ P H\in\mathcal{P} H ∈ P (see (71) ) with A p 0 H = 0 \mathcal{A}_{p_0}H=0 A p 0 H = 0 such that
∥ H ∥ 1 ≤ C E 1 1 ( Z ) 1 / 2 , E θ 1 ( Z − H ) ≤ C θ 4 E 1 1 ( Z ) , 0 < θ < θ 0 . (81) \|H\|_1\le C\mathcal{E}_1^1(Z)^{1/2},\qquad\mathcal{E}_\theta^1(Z-H)\le C\theta^4\mathcal{E}_1^1(Z),\qquad0<\theta<\theta_0. \tag*{(81)} ∥ H ∥ 1 ≤ C E 1 1 ( Z ) 1/2 , E θ 1 ( Z − H ) ≤ C θ 4 E 1 1 ( Z ) , 0 < θ < θ 0 . (81)
The proof goes through the rest of this section: we reduce to the square model in § 3.4.1 , and prove (i) in § 3.4.2 and (ii) in § 3.4.3 . The polynomial corrections are treated in § 3.4.4 .
Reduction to the square model
Example 6.2 (The square model). Let f ^ 0 \hat f_0 f ^ 0 be the Euclidean density (60) and B 0 \mathcal{B}_0 B 0 the second variation at U 0 U_0 U 0 . Then
B 0 ( Z , Z ) = ∫ ( ρ − 2 ∣ ∂ ξ ˉ Z φ ∣ 2 + ∣ D y Z φ ∣ 2 + ∣ D ξ Z V ∣ 2 + 4 ρ 2 ∣ D y Z V ∣ 2 ) . (82) \mathcal{B}_0(Z,Z)=\int\left(\rho^{-2}|\partial_{\bar{\xi}}Z_\varphi|^2+|D_yZ_\varphi|^2+|D_\xi Z_V|^2+4\rho^2|D_yZ_V|^2\right). \tag*{(82)} B 0 ( Z , Z ) = ∫ ( ρ − 2 ∣ ∂ ξ ˉ Z φ ∣ 2 + ∣ D y Z φ ∣ 2 + ∣ D ξ Z V ∣ 2 + 4 ρ 2 ∣ D y Z V ∣ 2 ) . (82) where ∂ ξ = 1 2 ( ∂ ξ 1 − − 1 ∂ ξ 2 ) \partial_\xi=\frac{1}{2}(\partial_{\xi_1}-\sqrt{-1}\partial_{\xi_2}) ∂ ξ = 2 1 ( ∂ ξ 1 − − 1 ∂ ξ 2 ) and ∂ ξ ˉ = 1 2 ( ∂ ξ 1 + − 1 ∂ ξ 2 ) \partial_{\bar{\xi}}=\frac{1}{2}(\partial_{\xi_1}+\sqrt{-1}\partial_{\xi_2}) ∂ ξ ˉ = 2 1 ( ∂ ξ 1 + − 1 ∂ ξ 2 ) for ξ = ξ 1 + − 1 ξ 2 ∈ C \xi=\xi_1+\sqrt{-1}\xi_2\in\mathbb{C} ξ = ξ 1 + − 1 ξ 2 ∈ C . Indeed, since ∣ D ξ φ ∣ 2 = 2 ( ∣ ∂ ξ φ ∣ 2 + ∣ ∂ ξ ˉ φ ∣ 2 ) |D_\xi\varphi|^2=2(|\partial_\xi\varphi|^2+|\partial_{\bar{\xi}}\varphi|^2) ∣ D ξ φ ∣ 2 = 2 ( ∣ ∂ ξ φ ∣ 2 + ∣ ∂ ξ ˉ φ ∣ 2 ) and det D ξ φ = ∣ ∂ ξ φ ∣ 2 − ∣ ∂ ξ ˉ φ ∣ 2 \det D_\xi\varphi=|\partial_\xi\varphi|^2-|\partial_{\bar{\xi}}\varphi|^2 det D ξ φ = ∣ ∂ ξ φ ∣ 2 − ∣ ∂ ξ ˉ φ ∣ 2 , we find (cf. the top-left block in (60) )
det D ξ φ 2 ∣ ( D ξ φ ) − 1 ∣ 2 = ∣ D ξ φ ∣ 2 2 det D ξ φ = 1 + 2 ∣ ∂ ξ ˉ φ ∣ 2 ∣ ∂ ξ φ ∣ 2 − ∣ ∂ ξ ˉ φ ∣ 2 . \frac{\det D_\xi\varphi}{2}|(D_\xi\varphi)^{-1}|^2=\frac{|D_\xi\varphi|^2}{2\det D_\xi\varphi}=1+\frac{2|\partial_{\bar{\xi}}\varphi|^2}{|\partial_\xi\varphi|^2-|\partial_{\bar{\xi}}\varphi|^2}. 2 det D ξ φ ∣ ( D ξ φ ) − 1 ∣ 2 = 2 det D ξ φ ∣ D ξ φ ∣ 2 = 1 + ∣ ∂ ξ φ ∣ 2 − ∣ ∂ ξ ˉ φ ∣ 2 2∣ ∂ ξ ˉ φ ∣ 2 . Along the variation U 0 + t Z U_0+tZ U 0 + tZ , we have ∂ ξ φ = 2 ξ + t ∂ ξ Z φ \partial_\xi\varphi=2\xi+t\partial_\xi Z_\varphi ∂ ξ φ = 2 ξ + t ∂ ξ Z φ and ∂ ξ ˉ φ = t ∂ ξ ˉ Z φ \partial_{\bar{\xi}}\varphi=t\partial_{\bar{\xi}}Z_\varphi ∂ ξ ˉ φ = t ∂ ξ ˉ Z φ , and substituting into (60) yields
f ^ 0 ( D U 0 + t D Z ) = 1 + t 2 2 ( ρ − 2 ∣ ∂ ξ ˉ Z φ ∣ 2 + ∣ D y Z φ ∣ 2 + ∣ D ξ Z V ∣ 2 + 4 ρ 2 ∣ D y Z V ∣ 2 ) + O ( t 3 ) . \hat f_0(DU_0+tDZ)=1+\frac{t^2}{2}\left(\rho^{-2}|\partial_{\bar{\xi}}Z_\varphi|^2+|D_yZ_\varphi|^2+|D_\xi Z_V|^2+4\rho^2|D_yZ_V|^2\right)+O(t^3). f ^ 0 ( D U 0 + t D Z ) = 1 + 2 t 2 ( ρ − 2 ∣ ∂ ξ ˉ Z φ ∣ 2 + ∣ D y Z φ ∣ 2 + ∣ D ξ Z V ∣ 2 + 4 ρ 2 ∣ D y Z V ∣ 2 ) + O ( t 3 ) .
Lemma 6.3 (Reduction to the square model). There exist an isometry ( z , y ) ↦ ( z ′ , y ′ ) (z,y)\mapsto(z',y') ( z , y ) ↦ ( z ′ , y ′ ) of the base, sending p 0 p_0 p 0 to ( 0 , y 0 ) (0,y_0) ( 0 , y 0 ) and P p 0 P_{p_0} P p 0 to { z ′ = 0 } \{z'=0\} { z ′ = 0 } , and a homothety of the fibre E 0 E_0 E 0 with factor κ − 1 / 2 = ∣ T p 0 ( 1 ) ∣ − 1 \kappa^{-1/2}=|T_{p_0}^{(1)}|^{-1} κ − 1/2 = ∣ T p 0 ( 1 ) ∣ − 1 , which together transform U p 0 U_{p_0} U p 0 into the square model U 0 U_0 U 0 . Let ( ξ ′ , y ′ ) = Ψ ( ξ , y ) (\xi',y')=\Psi(\xi,y) ( ξ ′ , y ′ ) = Ψ ( ξ , y ) be the induced change of Legendre coordinates, and Z ↦ Z ′ Z\mapsto Z' Z ↦ Z ′ be the linearisation at U p 0 U_{p_0} U p 0 of the induced map on fields. Then:
(i) B p 0 ( Z , Y ) = κ B 0 ( Z ′ , Y ′ ) \mathcal{B}_{p_0}(Z,Y)=\kappa\mathcal{B}_0(Z',Y') B p 0 ( Z , Y ) = κ B 0 ( Z ′ , Y ′ ) for admissible Z Z Z and compactly supported admissible Y Y Y ;
(ii) ∥ ∇ G Z ∥ L 2 \|\nabla^G Z\|_{L^2} ∥ ∇ G Z ∥ L 2 , ∥ Z ∥ L G 2 \|Z\|_{L^2_G} ∥ Z ∥ L G 2 and E r ( Z ) \mathcal{E}_r(Z) E r ( Z ) are uniformly comparable to the corresponding quantities of Z ′ Z' Z ′ on the image cylinders;
(iii) Z ↦ Z ′ Z\mapsto Z' Z ↦ Z ′ preserves admissibility, compact support, P \mathcal{P} P (see (71) ) and fields of the form ( q ( ξ ) , 0 ) (q(\xi),0) ( q ( ξ ) , 0 ) ; Ψ \Psi Ψ commutes with anisotropic dilations, and Q c r ( y 0 ) ⊂ Ψ ( Q r ( y 0 ) ) ⊂ Q C r ( y 0 ) Q_{\mathrm{cr}}(y_0)\subset\Psi(Q_r(y_0))\subset Q_{Cr}(y_0) Q cr ( y 0 ) ⊂ Ψ ( Q r ( y 0 )) ⊂ Q C r ( y 0 ) with uniform c , C > 0 c,C>0 c , C > 0 .
Proof. The tangent model is determined by the plane P p 0 P_{p_0} P p 0 and the real-linear map T = T p 0 : C → E 0 T=T_{p_0}:\mathbb{C}\to E_0 T = T p 0 : C → E 0 , T ( w ) = a p 0 w + b p 0 w ‾ T(w)=a_{p_0}w+b_{p_0}\overline{w} T ( w ) = a p 0 w + b p 0 w . We normalise the former by an isometry of the base and the latter by a homothety of the fibre, and read off the induced non-linear change of Legendre coordinates.
Since a p 0 w a_{p_0}w a p 0 w and b p 0 w ‾ b_{p_0}\overline{w} b p 0 w lie in orthogonal eigenspaces of J J J (cf. Definition 2.7 ), we have ∣ T ( w ) ∣ 2 = κ ∣ w ∣ 2 |T(w)|^2=\kappa|w|^2 ∣ T ( w ) ∣ 2 = κ ∣ w ∣ 2 . Let G G G be a linear isometry from ( R n , g ( p 0 ) ) (\mathbb{R}^n,g(p_0)) ( R n , g ( p 0 )) to Euclidean space sending P p 0 P_{p_0} P p 0 to { z ′ = 0 } \{z'=0\} { z ′ = 0 } , so that z ′ = Π P p 0 ⊥ ( X − p 0 ) z'=\Pi_{P_{p_0}^{\perp}}(X-p_0) z ′ = Π P p 0 ⊥ ( X − p 0 ) under the identification P p 0 ⊥ ≅ C P_{p_0}^{\perp}\cong\mathbb{C} P p 0 ⊥ ≅ C of § 3.1.1 , and set X ′ = G ( X − p 0 ) + ( 0 , y 0 ) X'=G(X-p_0)+(0,y_0) X ′ = G ( X − p 0 ) + ( 0 , y 0 ) . Let Q Q Q be an orthogonal map of E 0 E_0 E 0 with Q T ( w ) = ( κ 1 / 2 w , 0 ) QT(w)=(\kappa^{1/2}w,0) QT ( w ) = ( κ 1/2 w , 0 ) in the decomposition ( π , π ⊥ ) (\pi,\pi^\perp) ( π , π ⊥ ) , and set ( ξ ′ , V ′ ) = H ( ξ , V ) (\xi',V')=H(\xi,V) ( ξ ′ , V ′ ) = H ( ξ , V ) with H = κ − 1 / 2 Q H=\kappa^{-1/2}Q H = κ − 1/2 Q . Then the model becomes ± ( ( z ′ ) 1 / 2 , 0 ) \pm((z')^{1/2},0) ± (( z ′ ) 1/2 , 0 ) , i.e. z ′ = ( ξ ′ ) 2 z'=(\xi')^2 z ′ = ( ξ ′ ) 2 and V ′ = 0 V'=0 V ′ = 0 in the new Legendre coordinates.
In block form with respect to ( z , y ) (z,y) ( z , y ) and ( π , π ⊥ ) (\pi,\pi^\perp) ( π , π ⊥ ) , this reads
G 11 D g ( y 0 ) + G 12 = 0 , H 11 + H 12 M = ( π T ) − 1 , G_{11}Dg(y_0)+G_{12}=0,\qquad H_{11}+H_{12}M=(\pi T)^{-1}, G 11 D g ( y 0 ) + G 12 = 0 , H 11 + H 12 M = ( π T ) − 1 ,
H 21 + H 22 M = 0 , and G 11 q ( y 0 , ξ ) = ( ( π T ) − 1 ξ ) 2 , H_{21}+H_{22}M=0,\qquad\text{and}\qquad G_{11}q(y_0,\xi)=((\pi T)^{-1}\xi)^2, H 21 + H 22 M = 0 , and G 11 q ( y 0 , ξ ) = (( π T ) − 1 ξ ) 2 ,
where M = M ( y 0 ) M=M(y_0) M = M ( y 0 ) . For a field U U U , the new Legendre coordinates are Ψ U = ( H 11 ξ + H 12 V , y 0 + G 21 ( φ − g ( y 0 ) ) + G 22 ( y − y 0 ) ) \Psi_U=(H_{11}\xi+H_{12}V,y_0+G_{21}(\varphi-g(y_0))+G_{22}(y-y_0)) Ψ U = ( H 11 ξ + H 12 V , y 0 + G 21 ( φ − g ( y 0 )) + G 22 ( y − y 0 )) , and at U = U p 0 U=U_{p_0} U = U p 0 ,
Ψ ( ξ , y ) = ( ( π T ) − 1 ξ , y 0 + ( G 21 D g ( y 0 ) + G 22 ) ( y − y 0 ) + G 21 q ( y 0 , ξ ) ) . (83) \Psi(\xi,y)=\left((\pi T)^{-1}\xi,\ y_0+(G_{21}Dg(y_0)+G_{22})(y-y_0)+G_{21}q(y_0,\xi)\right). \tag*{(83)} Ψ ( ξ , y ) = ( ( π T ) − 1 ξ , y 0 + ( G 21 D g ( y 0 ) + G 22 ) ( y − y 0 ) + G 21 q ( y 0 , ξ ) ) . (83)
Linearising the induced map U ↦ U ′ U\mapsto U' U ↦ U ′ at U p 0 U_{p_0} U p 0 yields
Z ′ ∘ Ψ = ( G 11 − 2 ξ ′ H 12 0 H 22 ) Z . Z'\circ\Psi=\begin{pmatrix}G_{11}&-2\xi'H_{12}\\0&H_{22}\end{pmatrix}Z. Z ′ ∘ Ψ = ( G 11 0 − 2 ξ ′ H 12 H 22 ) Z .
We can now prove (i)–(iii). Since G G G and κ 1 / 2 H \kappa^{1/2}H κ 1/2 H are isometries, f p 0 ( D u ) = κ ∣ det g ( p 0 ) ∣ 1 / 2 f 0 ( D u ′ ) f_{p_0}(Du)=\kappa\lvert\det g(p_0)\rvert^{1/2}f_0(Du') f p 0 ( D u ) = κ ∣ det g ( p 0 ) ∣ 1/2 f 0 ( D u ′ ) , and changing variables gives ∫ Ω f ^ p 0 ( D U ) = κ ∫ Ψ U ( Ω ) f ^ 0 ( D U ′ ) \int_\Omega\widehat f_{p_0}(DU)=\kappa\int_{\Psi_U(\Omega)}\widehat f_0(DU') ∫ Ω f p 0 ( D U ) = κ ∫ Ψ U ( Ω ) f 0 ( D U ′ ) for fields U U U near U p 0 U_{p_0} U p 0 . Taking the second variation at U p 0 U_{p_0} U p 0 proves (i).
By (83) , Ψ \Psi Ψ is triangular with constant Jacobian, preserves parity and commutes with the anisotropic dilations. Since ξ ′ \xi' ξ ′ and Z V Z_V Z V are odd, the matrix above preserves admissibility, P \mathcal P P and fields of the form ( q ( ξ ) , 0 ) (q(\xi),0) ( q ( ξ ) , 0 ) . This proves (iii).
For (ii), we have ρ ′ ∼ ρ \rho'\sim\rho ρ ′ ∼ ρ , and the only new terms are 2 ξ ′ H 12 Z V = O ( ρ ∣ Z V ∣ ) 2\xi'H_{12}Z_V=O(\rho\lvert Z_V\rvert) 2 ξ ′ H 12 Z V = O ( ρ ∣ Z V ∣) in Z φ ′ Z'_\varphi Z φ ′ by the chain rule, which is compatible with the weights in ∥ ⋅ ∥ L G 2 \lVert\cdot\rVert_{L^2_G} ∥ ⋅ ∥ L G 2 and E r \mathcal E_r E r , and ρ − 1 D ξ ( 2 ξ ′ H 12 Z V ) = O ( ρ − 1 ∣ Z V ∣ + ∣ D ξ Z V ∣ ) \rho^{-1}D_\xi(2\xi'H_{12}Z_V)=O(\rho^{-1}\lvert Z_V\rvert+\lvert D_\xi Z_V\rvert) ρ − 1 D ξ ( 2 ξ ′ H 12 Z V ) = O ( ρ − 1 ∣ Z V ∣ + ∣ D ξ Z V ∣) in ∇ G Z ′ \nabla_G Z' ∇ G Z ′ , whose L 2 L^2 L 2 -norm is bounded by ∥ D ξ Z V ∥ L 2 \lVert D_\xi Z_V\rVert_{L^2} ∥ D ξ Z V ∥ L 2 by (67) .
Coercivity
To prove Proposition 6.1(i), it remains to bound ∫ ρ − 2 ∣ ∂ ξ ˉ Z φ ∣ 2 \int\rho^{-2}\lvert\partial_{\bar\xi}Z_\varphi\rvert^2 ∫ ρ − 2 ∣ ∂ ξ ˉ Z φ ∣ 2 from below. We start with a Hardy-type estimate for Z φ Z_\varphi Z φ . For such an even Z φ Z_\varphi Z φ , regarded as C \mathbb C C -valued, we write
f ( ξ , y ) = Z φ ( 0 , y ) − Z φ ( ξ , y ) 2 ξ . (84) f(\xi,y)=\frac{Z_\varphi(0,y)-Z_\varphi(\xi,y)}{2\xi}. \tag*{(84)} f ( ξ , y ) = 2 ξ Z φ ( 0 , y ) − Z φ ( ξ , y ) . (84)
Then f f f is odd in ξ \xi ξ , with f = O ( ρ ) f=O(\rho) f = O ( ρ ) and D ξ f = O ( 1 ) D_\xi f=O(1) D ξ f = O ( 1 ) , and ∂ ξ ˉ f = − ∂ ξ ˉ Z φ / ( 2 ξ ) \partial_{\bar\xi}f=-\partial_{\bar\xi}Z_\varphi/(2\xi) ∂ ξ ˉ f = − ∂ ξ ˉ Z φ / ( 2 ξ ) , so
ρ − 2 ∣ ∂ ξ ˉ Z φ ∣ 2 = 4 ∣ ∂ ξ ˉ f ∣ 2 = ∣ D ξ f ∣ 2 − 2 det D ξ f . (85) \rho^{-2}\lvert\partial_{\bar\xi}Z_\varphi\rvert^2=4\lvert\partial_{\bar\xi}f\rvert^2=\lvert D_\xi f\rvert^2-2\det D_\xi f. \tag*{(85)} ρ − 2 ∣ ∂ ξ ˉ Z φ ∣ 2 = 4 ∣ ∂ ξ ˉ f ∣ 2 = ∣ D ξ f ∣ 2 − 2 det D ξ f . (85)
We then prove the a Hardy-type inequality for the even component Z φ Z_\varphi Z φ .
Lemma 6.4. Every smooth Z φ Z_\varphi Z φ on Q r ( y 0 ) Q_r(y_0) Q r ( y 0 ) , even in ξ \xi ξ , satisfies
∫ Q r ( y 0 ) ∣ Z φ ( ξ , y ) − Z φ ( 0 , y ) ∣ 2 ρ 4 ≤ ∫ Q r ( y 0 ) ∣ ∂ ρ Z φ ∣ 2 ρ 2 , ∥ D ξ f ∥ L 2 + r − 1 ∥ f ∥ L 2 ≤ C ∥ ρ − 1 D ξ Z φ ∥ L 2 . (86) \int_{Q_r(y_0)}\frac{\lvert Z_\varphi(\xi,y)-Z_\varphi(0,y)\rvert^2}{\rho^4}\leq\int_{Q_r(y_0)}\frac{\lvert\partial_\rho Z_\varphi\rvert^2}{\rho^2},\qquad\lVert D_\xi f\rVert_{L^2}+r^{-1}\lVert f\rVert_{L^2}\leq C\lVert\rho^{-1}D_\xi Z_\varphi\rVert_{L^2}. \tag*{(86)} ∫ Q r ( y 0 ) ρ 4 ∣ Z φ ( ξ , y ) − Z φ ( 0 , y ) ∣ 2 ≤ ∫ Q r ( y 0 ) ρ 2 ∣ ∂ ρ Z φ ∣ 2 , ∥ D ξ f ∥ L 2 + r − 1 ∥ f ∥ L 2 ≤ C ∥ ρ − 1 D ξ Z φ ∥ L 2 . (86) Proof. Note that Z φ − Z φ ( 0 , y ) = O ( ρ 2 ) Z_\varphi-Z_\varphi(0,y)=O(\rho^2) Z φ − Z φ ( 0 , y ) = O ( ρ 2 ) by evenness. For fixed ( θ , y ) (\theta,y) ( θ , y ) and 0 < s < r 0<s<r 0 < s < r , integration by parts and Young’s inequality give
∫ 0 s ∣ Z φ − Z φ ( 0 , y ) ∣ 2 ρ 3 d ρ ≤ ∫ 0 s Re ⟨ Z φ − Z φ ( 0 , y ) , ∂ ρ Z φ ⟩ ρ 2 d ρ \int_0^s\frac{\lvert Z_\varphi-Z_\varphi(0,y)\rvert^2}{\rho^3}\,d\rho\leq\int_0^s\frac{\operatorname{Re}\langle Z_\varphi-Z_\varphi(0,y),\partial_\rho Z_\varphi\rangle}{\rho^2}\,d\rho ∫ 0 s ρ 3 ∣ Z φ − Z φ ( 0 , y ) ∣ 2 d ρ ≤ ∫ 0 s ρ 2 Re ⟨ Z φ − Z φ ( 0 , y ) , ∂ ρ Z φ ⟩ d ρ ≤ 1 2 ∫ 0 s ∣ Z φ − Z φ ( 0 , y ) ∣ 2 ρ 3 d ρ + 1 2 ∫ 0 s ∣ ∂ ρ Z φ ∣ 2 ρ d ρ . \leq\frac12\int_0^s\frac{\lvert Z_\varphi-Z_\varphi(0,y)\rvert^2}{\rho^3}\,d\rho+\frac12\int_0^s\frac{\lvert\partial_\rho Z_\varphi\rvert^2}{\rho}\,d\rho. ≤ 2 1 ∫ 0 s ρ 3 ∣ Z φ − Z φ ( 0 , y ) ∣ 2 d ρ + 2 1 ∫ 0 s ρ ∣ ∂ ρ Z φ ∣ 2 d ρ . Integrating in ( θ , y ) (\theta,y) ( θ , y ) and letting s → r s\to r s → r proves the first inequality. The second follows from it and ∣ D ξ f ∣ + r − 1 ∣ f ∣ ≲ ρ − 1 ∣ D ξ Z φ ∣ + ρ − 2 ∣ Z φ − Z φ ( 0 , y ) ∣ \lvert D_\xi f\rvert+r^{-1}\lvert f\rvert\lesssim\rho^{-1}\lvert D_\xi Z_\varphi\rvert+\rho^{-2}\lvert Z_\varphi-Z_\varphi(0,y)\rvert ∣ D ξ f ∣ + r − 1 ∣ f ∣ ≲ ρ − 1 ∣ D ξ Z φ ∣ + ρ − 2 ∣ Z φ − Z φ ( 0 , y )∣ .
Proof of Proposition 6.1(i). By Lemma 6.3 , it suffices to consider the square model. The upper bound is clear from (82) . For the lower bound, we aim to bound ∫ ρ − 2 ∣ D ξ Z φ ∣ 2 \int\rho^{-2}\lvert D_\xi Z_\varphi\rvert^2 ∫ ρ − 2 ∣ D ξ Z φ ∣ 2 by ∫ ρ − 2 ∣ ∂ ξ ˉ Z φ ∣ 2 \int\rho^{-2}\lvert\partial_{\bar\xi}Z_\varphi\rvert^2 ∫ ρ − 2 ∣ ∂ ξ ˉ Z φ ∣ 2 for smooth compactly supported admissible Z Z Z . Fix y y y and let f f f be as in (84) ; outside the support, f = Z φ ( 0 , y ) / ( 2 ξ ) = O ( ρ − 1 ) f=Z_\varphi(0,y)/(2\xi)=O(\rho^{-1}) f = Z φ ( 0 , y ) / ( 2 ξ ) = O ( ρ − 1 ) . The determinant in (85) integrates to zero: on an annulus { ε < ρ < s } \{\varepsilon<\rho<s\} { ε < ρ < s } ,
∫ det D ξ f d ξ = 1 2 ∫ 0 2 π Im ( f ‾ ∂ θ f ) d θ ∣ ρ = ε ρ = s = O ( ε 2 ) + O ( s − 2 ) . \int\det D_\xi f\,d\xi=\frac12\int_0^{2\pi}\operatorname{Im}(\overline f\,\partial_\theta f)\,d\theta\bigg|_{\rho=\varepsilon}^{\rho=s}=O(\varepsilon^2)+O(s^{-2}). ∫ det D ξ f d ξ = 2 1 ∫ 0 2 π Im ( f ∂ θ f ) d θ ρ = ε ρ = s = O ( ε 2 ) + O ( s − 2 ) .
Hence ∫ R 2 ∣ D ξ f ∣ 2 = ∫ R 2 ρ − 2 ∣ ∂ ξ ˉ Z φ ∣ 2 \int_{\mathbb R^2}\lvert D_\xi f\rvert^2=\int_{\mathbb R^2}\rho^{-2}\lvert\partial_{\bar\xi}Z_\varphi\rvert^2 ∫ R 2 ∣ D ξ f ∣ 2 = ∫ R 2 ρ − 2 ∣ ∂ ξ ˉ Z φ ∣ 2 . Since D ξ Z φ = − 2 ( f d ξ + ξ D ξ f ) D_\xi Z_\varphi=-2(f\,d\xi+\xi D_\xi f) D ξ Z φ = − 2 ( f d ξ + ξ D ξ f ) and f f f is odd, the weighted inequality in (67) yield
∫ R 2 ρ − 2 ∣ D ξ Z φ ∣ 2 ≲ ∫ R 2 ( ∣ D ξ f ∣ 2 + ρ − 2 ∣ f ∣ 2 ) ≲ ∫ R 2 ∣ D ξ f ∣ 2 = ∫ R 2 ρ − 2 ∣ ∂ ξ ˉ Z φ ∣ 2 . \int_{\mathbb R^2}\rho^{-2}\lvert D_\xi Z_\varphi\rvert^2\lesssim\int_{\mathbb R^2}\left(\lvert D_\xi f\rvert^2+\rho^{-2}\lvert f\rvert^2\right)\lesssim\int_{\mathbb R^2}\lvert D_\xi f\rvert^2=\int_{\mathbb R^2}\rho^{-2}\lvert\partial_{\bar\xi}Z_\varphi\rvert^2. ∫ R 2 ρ − 2 ∣ D ξ Z φ ∣ 2 ≲ ∫ R 2 ( ∣ D ξ f ∣ 2 + ρ − 2 ∣ f ∣ 2 ) ≲ ∫ R 2 ∣ D ξ f ∣ 2 = ∫ R 2 ρ − 2 ∣ ∂ ξ ˉ Z φ ∣ 2 .
Integrating in y y y completes the proof. □
Excess decay
Proof of Proposition 6.1(ii). By Lemma 6.3, we may assume U p 0 = U 0 U_{p_0}=U_0 U p 0 = U 0 , and write Q a = Q a ( y 0 ) Q_a=Q_a(y_0) Q a = Q a ( y 0 ) . The strategy (essentially the same as in Part I) is to prove that Z Z Z is smooth up to { ξ = 0 } \{\xi=0\} { ξ = 0 } , with bounded C m , α C^{m,\alpha} C m , α -norm, and take H H H to be its Taylor polynomial in P \mathcal{P} P (see (71) ). We divide this proof in three steps.
Step 1: Tangential regularity. For 0 < a < b < 1 0<a<b<1 0 < a < b < 1 , let η = η 1 ( ρ ) η 2 ( y ) \eta=\eta_1(\rho)\eta_2(y) η = η 1 ( ρ ) η 2 ( y ) be a cutoff of Q a ⊂ Q b Q_a\subset Q_b Q a ⊂ Q b with η 1 = 1 \eta_1=1 η 1 = 1 near ρ = 0 \rho=0 ρ = 0 . Since the coefficients of B 0 \mathcal{B}_0 B 0 do not depend on y y y , we have the exact identity
B 0 ( η Z , η Z ) = B 0 ( Z , η 2 Z ) + ∫ D 2 f ^ 0 ( D U 0 ) ( Z ⊗ D η , Z ⊗ D η ) . \mathcal{B}_0(\eta Z,\eta Z)=\mathcal{B}_0(Z,\eta^2Z)+\int D^2\widehat f_0(DU_0)(Z\otimes D\eta,Z\otimes D\eta). B 0 ( η Z , η Z ) = B 0 ( Z , η 2 Z ) + ∫ D 2 f 0 ( D U 0 ) ( Z ⊗ D η , Z ⊗ D η ) .
The first term vanishes. Coercivity (cf. Proposition 6.1 (i)), the square-model form (82) and D ξ η = 0 D_\xi\eta=0 D ξ η = 0 near ρ = 0 \rho=0 ρ = 0 imply the weighted Caccioppoli inequality
∫ Q a ∣ ∇ G Z ∣ 2 ≤ C a , b ∥ Z ∥ L G 2 ( Q b ) 2 ≤ C a , b E 1 ( Z ) . (87) \int_{Q_a}|\nabla^G Z|^2\le C_{a,b}\lVert Z\rVert_{L_G^2(Q_b)}^2\le C_{a,b}\mathcal{E}_1(Z). \tag*{(87)} ∫ Q a ∣ ∇ G Z ∣ 2 ≤ C a , b ∥ Z ∥ L G 2 ( Q b ) 2 ≤ C a , b E 1 ( Z ) . (87)
The equation is invariant under translations in y y y , so tangential difference quotients δ t , α Z ( ξ , y ) = Z ( ξ , y + t ∂ y α ) − Z ( ξ , y ) t \delta_{t,\alpha}Z(\xi,y)=\frac{Z(\xi,y+t\partial_{y_\alpha})-Z(\xi,y)}{t} δ t , α Z ( ξ , y ) = t Z ( ξ , y + t ∂ y α ) − Z ( ξ , y ) of Z Z Z are also weak solutions with L G 2 L_G^2 L G 2 -norms bounded by ∥ ∇ G Z ∥ L 2 \lVert\nabla^G Z\rVert_{L^2} ∥ ∇ G Z ∥ L 2 . Applying (87) to δ t , α Z \delta_{t,\alpha}Z δ t , α Z , we get
∑ ∣ γ ∣ ≤ m ( ∥ D y γ Z ∥ L 2 ( Q a ) + ∥ ∇ G D y γ Z ∥ L 2 ( Q a ) ) ≤ C m , a E 1 ( Z ) 1 / 2 . (88) \sum_{|\gamma|\le m}\left(\lVert D_y^\gamma Z\rVert_{L^2(Q_a)}+\lVert\nabla^G D_y^\gamma Z\rVert_{L^2(Q_a)}\right)\le C_{m,a}\mathcal{E}_1(Z)^{1/2}. \tag*{(88)} ∣ γ ∣ ≤ m ∑ ( ∥ D y γ Z ∥ L 2 ( Q a ) + ∥ ∇ G D y γ Z ∥ L 2 ( Q a ) ) ≤ C m , a E 1 ( Z ) 1/2 . (88)
Step 2: Regularity across { ξ = 0 } \{\xi=0\} { ξ = 0 } . Let f f f be the normal difference quotient (84) of Z φ Z_\varphi Z φ . Lemma 6.4 and the trace bound of Lemma 5.3, applied to cutoffs of D y γ Z D_y^\gamma Z D y γ Z , give
∑ ∣ γ ∣ ≤ m ( ∥ D y γ Z φ ( 0 , ⋅ ) ∥ L 2 ( B a n − 2 ( y 0 ) ) + ∥ D ξ γ f ∥ L 2 ( Q a ) + ∥ D ξ D y γ f ∥ L 2 ( Q a ) ) ≤ C m , a E 1 ( Z ) 1 / 2 . (89) \sum_{|\gamma|\le m}\left(\lVert D_y^\gamma Z_\varphi(0,\cdot)\rVert_{L^2(B_a^{n-2}(y_0))}+\lVert D_\xi^\gamma f\rVert_{L^2(Q_a)}+\lVert D_\xi D_y^\gamma f\rVert_{L^2(Q_a)}\right)\le C_{m,a}\mathcal{E}_1(Z)^{1/2}. \tag*{(89)} ∣ γ ∣ ≤ m ∑ ( ∥ D y γ Z φ ( 0 , ⋅ ) ∥ L 2 ( B a n − 2 ( y 0 )) + ∥ D ξ γ f ∥ L 2 ( Q a ) + ∥ D ξ D y γ f ∥ L 2 ( Q a ) ) ≤ C m , a E 1 ( Z ) 1/2 . (89)
Testing the equation with Y = ( − 2 ξ ψ , 0 ) Y=(-2\xi\psi,0) Y = ( − 2 ξ ψ , 0 ) , with ψ \psi ψ odd, we find that f f f and Z V Z_V Z V are weak solutions of
Δ ξ f + 4 ρ 2 Δ y f = 2 ξ ‾ Δ y Z φ ( 0 , y ) , Δ ξ Z V + 4 ρ 2 Δ y Z V = 0 , (90) \Delta_\xi f+4\rho^2\Delta_y f=2\overline{\xi}\Delta_y Z_\varphi(0,y),\qquad\Delta_\xi Z_V+4\rho^2\Delta_y Z_V=0, \tag*{(90)} Δ ξ f + 4 ρ 2 Δ y f = 2 ξ Δ y Z φ ( 0 , y ) , Δ ξ Z V + 4 ρ 2 Δ y Z V = 0 , (90)
where odd test functions suffice by oddness of f f f and Z V Z_V Z V . On each normal slice, the standard W 2 , 2 W^{2,2} W 2 , 2 -estimates imply
∥ D ξ 2 f ∥ L 2 ( B a 2 ) ≤ C a , b ( ∥ Δ ξ f ∥ L 2 ( B b 2 ) + ∥ f ∥ L 2 ( B b 2 ) ) . \lVert D_\xi^2f\rVert_{L^2(B_a^2)}\le C_{a,b}\left(\lVert\Delta_\xi f\rVert_{L^2(B_b^2)}+\lVert f\rVert_{L^2(B_b^2)}\right). ∥ D ξ 2 f ∥ L 2 ( B a 2 ) ≤ C a , b ( ∥ Δ ξ f ∥ L 2 ( B b 2 ) + ∥ f ∥ L 2 ( B b 2 ) ) .
Using the equation and integrating in y y y , we obtain
∥ D ξ 2 f ∥ L 2 ( Q a ) ≤ C a , b ( ∥ Δ y f ∥ L 2 ( Q b ) + ∥ Δ y Z φ ( 0 , ⋅ ) ∥ L 2 ( B b n − 2 ( y 0 ) ) + ∥ f ∥ L 2 ( Q b ) ) ≤ C a , b E 1 ( Z ) 1 / 2 , \lVert D_\xi^2f\rVert_{L^2(Q_a)}\le C_{a,b}\left(\lVert\Delta_yf\rVert_{L^2(Q_b)}+\lVert\Delta_yZ_\varphi(0,\cdot)\rVert_{L^2(B_b^{n-2}(y_0))}+\lVert f\rVert_{L^2(Q_b)}\right)\le C_{a,b}\mathcal{E}_1(Z)^{1/2}, ∥ D ξ 2 f ∥ L 2 ( Q a ) ≤ C a , b ( ∥ Δ y f ∥ L 2 ( Q b ) + ∥ Δ y Z φ ( 0 , ⋅ ) ∥ L 2 ( B b n − 2 ( y 0 )) + ∥ f ∥ L 2 ( Q b ) ) ≤ C a , b E 1 ( Z ) 1/2 ,
by (89) . The same bound holds for D ξ 2 D y γ f D_\xi^2D_y^\gamma f D ξ 2 D y γ f and their higher order weighted derivatives. The same estimates apply to Z V Z_V Z V . Thus, by Sobolev embedding, we find
∥ Z φ ∥ C m ( Q a ) + ∥ Z V ∥ C m ( Q a ) ≤ C m , a E 1 ( Z ) 1 / 2 . \lVert Z_\varphi\rVert_{C^m(Q_a)}+\lVert Z_V\rVert_{C^m(Q_a)}\le C_{m,a}\mathcal{E}_1(Z)^{1/2}. ∥ Z φ ∥ C m ( Q a ) + ∥ Z V ∥ C m ( Q a ) ≤ C m , a E 1 ( Z ) 1/2 .
Step 3: Approximation by a kernel element. By parity, the Taylor polynomial of Z Z Z at ( 0 , y 0 ) (0,y_0) ( 0 , y 0 ) in P \mathcal{P} P (see (71) ) is
H = ( Z φ ( 0 , y 0 ) + D y Z φ ( 0 , y 0 ) ( y − y 0 ) + 1 2 D ξ 2 Z φ ( 0 , y 0 ) ( ξ , ξ ) , D ξ Z V ( 0 , y 0 ) ξ ) , H=\left(Z_\varphi(0,y_0)+D_yZ_\varphi(0,y_0)(y-y_0)+\frac{1}{2}D_\xi^2Z_\varphi(0,y_0)(\xi,\xi),D_\xi Z_V(0,y_0)\xi\right), H = ( Z φ ( 0 , y 0 ) + D y Z φ ( 0 , y 0 ) ( y − y 0 ) + 2 1 D ξ 2 Z φ ( 0 , y 0 ) ( ξ , ξ ) , D ξ Z V ( 0 , y 0 ) ξ ) ,
and Step 2 gives, on Q θ 0 Q_{\theta_0} Q θ 0 with θ 0 ≪ 1 \theta_0\ll1 θ 0 ≪ 1 ,
∣ Z φ − H φ ∣ ≲ ( ρ 4 + ρ 2 ∣ y − y 0 ∣ + ∣ y − y 0 ∣ 2 ) E 1 ( Z ) 1 / 2 , ∣ Z V − H V ∣ ≲ ( ρ 3 + ρ ∣ y − y 0 ∣ ) E 1 ( Z ) 1 / 2 . |Z_\varphi-H_\varphi|\lesssim\left(\rho^4+\rho^2|y-y_0|+|y-y_0|^2\right)\mathcal{E}_1(Z)^{1/2},\qquad|Z_V-H_V|\lesssim\left(\rho^3+\rho|y-y_0|\right)\mathcal{E}_1(Z)^{1/2}. ∣ Z φ − H φ ∣ ≲ ( ρ 4 + ρ 2 ∣ y − y 0 ∣ + ∣ y − y 0 ∣ 2 ) E 1 ( Z ) 1/2 , ∣ Z V − H V ∣ ≲ ( ρ 3 + ρ ∣ y − y 0 ∣ ) E 1 ( Z ) 1/2 .
Integrating over Q θ Q_\theta Q θ with θ ≤ θ 0 \theta\le\theta_0 θ ≤ θ 0 proves (81) . It remains to show A 0 H = 0 \mathcal{A}_0H=0 A 0 H = 0 . The same expansion gives ∥ ∇ G S r ( Z − H ) ∥ L 2 ( Q 1 ) ≲ r 2 E 1 ( Z ) 1 / 2 \lVert\nabla^G S_r(Z-H)\rVert_{L^2(Q_1)}\lesssim r^2\mathcal{E}_1(Z)^{1/2} ∥ ∇ G S r ( Z − H ) ∥ L 2 ( Q 1 ) ≲ r 2 E 1 ( Z ) 1/2 . Since B 0 \mathcal{B}_0 B 0 depends only on derivatives, D ( S r H ) = D H D(\mathbf{S}_rH)=DH D ( S r H ) = D H for the rescaling S r \mathbf{S}_r S r centred at y 0 y_0 y 0 , and the equation is invariant under S r \mathbf{S}_r S r , we find
∣ B 0 ( H , Y ) ∣ = ∣ B 0 ( S r ( H − Z ) , Y ) ∣ ≲ r 2 E 1 ( Z ) 1 / 2 ∥ ∇ G Y ∥ L 2 |\mathcal{B}_0(H,Y)|=|\mathcal{B}_0(\mathbf{S}_r(H-Z),Y)|\lesssim r^2\mathcal{E}_1(Z)^{1/2}\lVert\nabla^G Y\rVert_{L^2} ∣ B 0 ( H , Y ) ∣ = ∣ B 0 ( S r ( H − Z ) , Y ) ∣ ≲ r 2 E 1 ( Z ) 1/2 ∥ ∇ G Y ∥ L 2
for all small r r r and all compactly supported admissible Y Y Y . Taking r → 0 r\to0 r → 0 completes the proof. □
Polynomial corrections
For P ∈ P P\in\mathcal{P} P ∈ P (see (71) ), the functional A p 0 P \mathcal{A}_{p_0}P A p 0 P acts on compactly supported admissible test fields. Let A p 0 ( P ) \mathcal{A}_{p_0}(\mathcal{P}) A p 0 ( P ) be the image of P \mathcal{P} P under A p 0 \mathcal{A}_{p_0} A p 0 and ker A p 0 = { P ∈ P : A p 0 P = 0 } \ker\mathcal{A}_{p_0}=\{P\in\mathcal{P}:\mathcal{A}_{p_0}P=0\} ker A p 0 = { P ∈ P : A p 0 P = 0 } . In § 3.5 , the principal parts of the sources lie in the finite-dimensional space A p 0 ( P ) \mathcal{A}_{p_0}(\mathcal{P}) A p 0 ( P ) .
Lemma 6.5 (Polynomial corrections). There is a linear right inverse R p 0 : A p 0 ( P ) → P \mathbf{R}_{p_0}:\mathcal{A}_{p_0}(\mathcal{P})\to\mathcal{P} R p 0 : A p 0 ( P ) → P such that each R p 0 F \mathbf{R}_{p_0}\mathcal{F} R p 0 F is of the form ( q ( ξ ) , 0 ) (q(\xi),0) ( q ( ξ ) , 0 ) and
A p 0 R p 0 F = F , ∥ R p 0 F ∥ 1 ≤ C ∥ F ∥ ( W G 1 , 2 ( Q 1 ( y 0 ) ) ) ∗ , F ∈ A p 0 ( P ) , (91) \mathcal{A}_{p_0}\mathbf{R}_{p_0}\mathcal{F}=\mathcal{F},\qquad\lVert\mathbf{R}_{p_0}\mathcal{F}\rVert_1\le C\lVert\mathcal{F}\rVert_{(W_G^{1,2}(Q_1(y_0)))^*},\qquad\mathcal{F}\in\mathcal{A}_{p_0}(\mathcal{P}), \tag*{(91)} A p 0 R p 0 F = F , ∥ R p 0 F ∥ 1 ≤ C ∥ F ∥ ( W G 1 , 2 ( Q 1 ( y 0 )) ) ∗ , F ∈ A p 0 ( P ) , (91) with C C C uniform in p 0 p_0 p 0 .
Proof. By Lemma 6.3, it suffices to consider the square model. For P = ( a + B ( y − y 0 ) + q ( ξ ) , M ξ ) P=(a+B(y-y_0)+q(\xi),M\xi) P = ( a + B ( y − y 0 ) + q ( ξ ) , M ξ ) , the affine part and M ξ M\xi M ξ contribute zero to B 0 ( P , Y ) \mathcal{B}_0(P,Y) B 0 ( P , Y ) by (82) and integration by parts, so A 0 P = A 0 ( q , 0 ) \mathcal{A}_0P=\mathcal{A}_0(q,0) A 0 P = A 0 ( q , 0 ) . Thus A 0 \mathcal{A}_0 A 0 maps the finite-dimensional space of fields ( q ( ξ ) , 0 ) (q(\xi),0) ( q ( ξ ) , 0 ) onto A 0 ( P ) \mathcal{A}_0(\mathcal{P}) A 0 ( P ) , and any linear right inverse of this map satisfies (91) , since all norms on a finite-dimensional space are equivalent. □
Excess decay for the second variation
In this section we prove a general Campanato-type excess decay result for weak solutions Z Z Z of the linearised equation
B U ( Z , Y ) = F ( Y ) , \mathcal{B}_U(Z,Y)=\mathcal{F}(Y), B U ( Z , Y ) = F ( Y ) ,
whose source term F \mathcal{F} F with F ∈ ( W G 1 , 2 ( Q r ( y 0 ) ) ) ∗ \mathcal{F}\in(W_G^{1,2}(Q_r(y_0)))^* F ∈ ( W G 1 , 2 ( Q r ( y 0 )) ) ∗ (see (66) for definition) has a principal part F p r ∈ A p 0 ( P ) \mathcal{F}_{\mathrm{pr}}\in\mathcal{A}_{p_0}(\mathcal{P}) F pr ∈ A p 0 ( P ) with size O ( 1 ) O(1) O ( 1 ) and a remainder of order r N G / 2 + α r^{N_G/2+\alpha} r N G /2 + α at scale r r r (cf. Theorem 7.4 ). In § 3.6 , Z Z Z will be taken as tangential derivatives D y β U D^\beta_yU D y β U , and F \mathcal{F} F will be specified by (120) , consisting of lower-order tangential derivatives of the transformed spinor U U U , and F p r \mathcal{F}_{\mathrm{pr}} F pr will be lower-order derivatives frozen at p 0 p_0 p 0 .
The idea is to treat B U \mathcal{B}_U B U as a perturbation of the frozen form B p 0 \mathcal{B}_{p_0} B p 0 of § 3.4 . By Proposition 5.4, the coefficients of B U \mathcal{B}_U B U differ from the frozen ones by O ( r α ) O(r^\alpha) O ( r α ) on Q r ( y 0 ) Q_r(y_0) Q r ( y 0 ) . If only the difference B U − B p 0 \mathcal{B}_U-\mathcal{B}_{p_0} B U − B p 0 were present, the usual Campanato iteration would apply: at each scale, we approximate Z Z Z by a polynomial P r P_r P r , compare Z − P r Z-P_r Z − P r with a solution of the frozen equation, and update P r P_r P r by the kernel element given by Proposition 6.1(ii). However, the density (76) depends on the original position X = ( φ , y ) X=(\varphi,y) X = ( φ , y ) . The constant term a a a of a kernel element does move the centre p 0 p_0 p 0 in the normal z z z -direction, and thus contributes a principal part N p 0 a \mathcal{N}_{p_0}a N p 0 a to the source (see (93) below), of order ∣ a ∣ = O ( 1 ) |a|=O(1) ∣ a ∣ = O ( 1 ) , which is not small at scale r r r . To remove this systematic error caused by freezing, together with the given principal part F p r \mathcal{F}_{\mathrm{pr}} F pr , we further correct the polynomials by Lemma 6.5, so that they lie in the affine class (106) . The resulting argument is a variant of the Campanato iteration.
Throughout this section, we fix a branch point p 0 = p ( y 0 ) p_0=p(y_0) p 0 = p ( y 0 ) . By Proposition 5.4(iii) and φ p 0 − g ( y 0 ) = O ( r 2 ) \varphi_{p_0}-g(y_0)=O(r^2) φ p 0 − g ( y 0 ) = O ( r 2 ) on Q r ( y 0 ) Q_r(y_0) Q r ( y 0 ) , for 0 < r ≤ r 0 0<r\le r_0 0 < r ≤ r 0 , we have
∣ ∇ G ( U − U p 0 ) ∣ ≤ C r α , ∣ X − p 0 ∣ ≤ C r 2 on Q r ( y 0 ) ∖ { ξ = 0 } . (92) |\nabla^G(U-U_{p_0})|\le Cr^\alpha,\qquad|X-p_0|\le Cr^2\quad\text{on }Q_r(y_0)\setminus\{\xi=0\}. \tag*{(92)} ∣ ∇ G ( U − U p 0 ) ∣ ≤ C r α , ∣ X − p 0 ∣ ≤ C r 2 on Q r ( y 0 ) ∖ { ξ = 0 } . (92)
We write
N p 0 a = d d t ∣ t = 0 A p r ( p 0 + t ( a , 0 ) , D U p 0 ) , a ∈ C , (93) \mathcal{N}_{p_0}a=\left.\frac{d}{dt}\right|_{t=0}\mathcal{A}_{\mathrm{pr}}(p_0+t(a,0),DU_{p_0}),\qquad a\in\mathbb{C}, \tag*{(93)} N p 0 a = d t d t = 0 A pr ( p 0 + t ( a , 0 ) , D U p 0 ) , a ∈ C , (93)
that is, the change of the frozen coefficients when the centre moves in the normal direction by a a a as discussed above, with D U p 0 DU_{p_0} D U p 0 fixed.
Comparison with the frozen operator
The following lemma collects the estimates comparing B U \mathcal{B}_U B U with B p 0 \mathcal{B}_{p_0} B p 0 that are used below.
Lemma 7.1 (Comparison with the frozen operator). For 0 < r ≤ r 0 ≪ 1 0<r\le r_0\ll1 0 < r ≤ r 0 ≪ 1 , the following hold on Q r = Q r ( y 0 ) Q_r=Q_r(y_0) Q r = Q r ( y 0 ) .
(i) For Z ∈ W G 1 , 2 ( Q r ) Z\in W^{1,2}_G(Q_r) Z ∈ W G 1 , 2 ( Q r ) and Y ∈ W G , 0 1 , 2 ( Q r ) Y\in W^{1,2}_{G,0}(Q_r) Y ∈ W G , 0 1 , 2 ( Q r ) ,
∣ B U ( Z , Y ) − B p 0 ( Z , Y ) ∣ ≤ C r α ( ∥ ∇ G Z ∥ L 2 + r − 2 ∥ Z φ ∥ L 2 ) ∥ ∇ G Y ∥ L 2 . (94) \left|\mathcal{B}_U(Z,Y)-\mathcal{B}_{p_0}(Z,Y)\right|\le Cr^\alpha\left(\|\nabla^G Z\|_{L^2}+r^{-2}\|Z_\varphi\|_{L^2}\right)\|\nabla^G Y\|_{L^2}. \tag*{(94)} ∣ B U ( Z , Y ) − B p 0 ( Z , Y ) ∣ ≤ C r α ( ∥ ∇ G Z ∥ L 2 + r − 2 ∥ Z φ ∥ L 2 ) ∥ ∇ G Y ∥ L 2 . (94) In particular, B U \mathcal{B}_U B U is uniformly coercive on W G , 0 1 , 2 ( Q r ) W^{1,2}_{G,0}(Q_r) W G , 0 1 , 2 ( Q r ) .
(ii) For P ∈ P P\in\mathcal{P} P ∈ P with constant term a a a , the error field E = A U P − A p r , p 0 P − N p 0 a \mathbf{E}=\mathcal{A}_U P-\mathcal{A}_{\mathrm{pr},p_0}P-\mathcal{N}_{p_0}a E = A U P − A pr , p 0 P − N p 0 a satisfies
r − N G / 2 ∥ E ∥ ( W G 1 , 2 ( Q r ) ) ∗ ≤ C r α ( ∣ a ∣ + ∥ ∇ P ∥ ) , (95) r^{-N_G/2}\|\mathbf{E}\|_{(W^{1,2}_G(Q_r))^*}\le Cr^\alpha(|a|+\|\nabla P\|), \tag*{(95)} r − N G /2 ∥ E ∥ ( W G 1 , 2 ( Q r ) ) ∗ ≤ C r α ( ∣ a ∣ + ∥∇ P ∥ ) , (95) ∥ E ∥ C G 0 , α ( Q r ) ≤ C r α ( ∣ a ∣ + ∥ ∇ P ∥ ) . (96) \|\mathbf{E}\|_{C^{0,\alpha}_G(Q_r)}\le Cr^\alpha(|a|+\|\nabla P\|). \tag*{(96)} ∥ E ∥ C G 0 , α ( Q r ) ≤ C r α ( ∣ a ∣ + ∥∇ P ∥ ) . (96) (iii) For every a ∈ C a\in\mathbb{C} a ∈ C there is P a ∈ P P_a\in\mathcal{P} P a ∈ P with constant term a a a such that N p 0 a = − A p 0 P a \mathcal{N}_{p_0}a=-\mathcal{A}_{p_0}P_a N p 0 a = − A p 0 P a . In particular,
N p 0 a ∈ A p 0 ( P ) , ∥ R p 0 N p 0 a ∥ 1 ≤ C ∣ a ∣ . (97) \mathcal{N}_{p_0}a\in\mathcal{A}_{p_0}(\mathcal{P}),\qquad\|\mathcal{R}_{p_0}\mathcal{N}_{p_0}a\|_1\le C|a|. \tag*{(97)} N p 0 a ∈ A p 0 ( P ) , ∥ R p 0 N p 0 a ∥ 1 ≤ C ∣ a ∣. (97)
Proof. The main ingredients are the structure (5.21) of the transformed density, the deviation estimate (7.1), and the weighted inequalities in Lemma 5.3.
(i) Write the second variation as B U ( Z , Y ) = ∫ ( a ( D Z , D Y ) + b ( Z , D Y ) + b ( Y , D Z ) + c ( Z , Y ) ) \mathcal{B}_U(Z,Y)=\int(a(DZ,DY)+b(Z,DY)+b(Y,DZ)+c(Z,Y)) B U ( Z , Y ) = ∫ ( a ( D Z , D Y ) + b ( Z , D Y ) + b ( Y , D Z ) + c ( Z , Y )) , with coefficients given by second derivatives of f ^ \hat f f ^ at U U U . By (5.21), the principal part is D D U 2 f ^ p r ( X , ∇ G U ) D^2_{DU}\hat f_{\mathrm{pr}}(X,\nabla^G U) D D U 2 f ^ pr ( X , ∇ G U ) up to O ( ρ 2 ) O(\rho^2) O ( ρ 2 ) , and B p 0 \mathcal{B}_{p_0} B p 0 is obtained by evaluating it at ( p 0 , ∇ G U p 0 ) (p_0,\nabla^G U_{p_0}) ( p 0 , ∇ G U p 0 ) . By (7.1), the difference of the principal parts is bounded by C r α ∣ ∇ G Z ∣ ∣ ∇ G Y ∣ Cr^\alpha|\nabla^G Z||\nabla^G Y| C r α ∣ ∇ G Z ∣∣ ∇ G Y ∣ . The remaining terms are bounded by
C ( ∣ Z φ ∣ + ρ ∣ Z V ∣ + ρ 2 ∣ ∇ G Z ∣ ) ∣ ∇ G Y ∣ + C ( ∣ Y φ ∣ + ρ ∣ Y V ∣ ) ∣ ∇ G Z ∣ + C ( ∣ Z φ ∣ + ∣ Z V ∣ ) ( ∣ Y φ ∣ + ∣ Y V ∣ ) . C(|Z_\varphi|+\rho|Z_V|+\rho^2|\nabla^G Z|)|\nabla^G Y|+C(|Y_\varphi|+\rho|Y_V|)|\nabla^G Z|+C(|Z_\varphi|+|Z_V|)(|Y_\varphi|+|Y_V|). C ( ∣ Z φ ∣ + ρ ∣ Z V ∣ + ρ 2 ∣ ∇ G Z ∣ ) ∣ ∇ G Y ∣ + C ( ∣ Y φ ∣ + ρ ∣ Y V ∣ ) ∣ ∇ G Z ∣ + C ( ∣ Z φ ∣ + ∣ Z V ∣ ) ( ∣ Y φ ∣ + ∣ Y V ∣ ) .
By Lemma 5.3,
∥ Z V ∥ L 2 ≤ r ∥ ∇ G Z ∥ L 2 , ∥ Y φ ∥ L 2 + r ∥ Y V ∥ L 2 ≤ C r 2 ∥ ∇ G Y ∥ L 2 . \|Z_V\|_{L^2}\le r\|\nabla^G Z\|_{L^2},\qquad\|Y_\varphi\|_{L^2}+r\|Y_V\|_{L^2}\le Cr^2\|\nabla^G Y\|_{L^2}. ∥ Z V ∥ L 2 ≤ r ∥ ∇ G Z ∥ L 2 , ∥ Y φ ∥ L 2 + r ∥ Y V ∥ L 2 ≤ C r 2 ∥ ∇ G Y ∥ L 2 .
The Cauchy–Schwarz inequality then bounds the remaining terms by
C ( ∥ Z φ ∥ L 2 + r 2 ∥ ∇ G Z ∥ L 2 ) ∥ ∇ G Y ∥ L 2 = C r 2 ( ∥ ∇ G Z ∥ L 2 + r − 2 ∥ Z φ ∥ L 2 ) ∥ ∇ G Y ∥ L 2 , C(\|Z_\varphi\|_{L^2}+r^2\|\nabla^G Z\|_{L^2})\|\nabla^G Y\|_{L^2}=Cr^2(\|\nabla^G Z\|_{L^2}+r^{-2}\|Z_\varphi\|_{L^2})\|\nabla^G Y\|_{L^2}, C ( ∥ Z φ ∥ L 2 + r 2 ∥ ∇ G Z ∥ L 2 ) ∥ ∇ G Y ∥ L 2 = C r 2 ( ∥ ∇ G Z ∥ L 2 + r − 2 ∥ Z φ ∥ L 2 ) ∥ ∇ G Y ∥ L 2 ,
which proves (7.3). For compactly supported Z Z Z , we have r − 2 ∥ Z φ ∥ L 2 ≤ C ∥ ∇ G Z ∥ L 2 r^{-2}\|Z_\varphi\|_{L^2}\le C\|\nabla^G Z\|_{L^2} r − 2 ∥ Z φ ∥ L 2 ≤ C ∥ ∇ G Z ∥ L 2 , so coercivity of B U \mathcal{B}_U B U follows from the coercivity (6.2) of the frozen form.
(ii) Write P = ( a , 0 ) + ( P − ( a , 0 ) ) P=(a,0)+(P-(a,0)) P = ( a , 0 ) + ( P − ( a , 0 )) . Since D U DU D U is unchanged by the constant variation ( a , 0 ) (a,0) ( a , 0 ) ,
A U ( a , 0 ) = d d t ∣ t = 0 A p r ( X + t ( a , 0 ) , D U ) + O ( ρ 2 ∣ a ∣ ) , \mathcal{A}_U(a,0)=\left.\frac{d}{dt}\right|_{t=0}\mathcal{A}_{\mathrm{pr}}(X+t(a,0),DU)+O(\rho^2|a|), A U ( a , 0 ) = d t d t = 0 A pr ( X + t ( a , 0 ) , D U ) + O ( ρ 2 ∣ a ∣ ) ,
where the O ( ρ 2 ∣ a ∣ ) O(\rho^2|a|) O ( ρ 2 ∣ a ∣ ) -remainder is measured by the weighted C 0 , α C^{0,\alpha} C 0 , α -norm (5.15). Compared with N p 0 a \mathcal{N}_{p_0}a N p 0 a , the coefficients are evaluated at ( X , D U ) (X,DU) ( X , D U ) instead of ( p 0 , D U p 0 ) (p_0,DU_{p_0}) ( p 0 , D U p 0 ) , so (7.1) gives an error O ( r α ∣ a ∣ ) O(r^\alpha|a|) O ( r α ∣ a ∣ ) . The non-constant part satisfies
∣ ∇ G P ∣ ≤ C ∥ ∇ P ∥ , ∣ P φ − a ∣ ≤ C r 2 ∥ ∇ P ∥ , ∣ P V ∣ ≤ ρ ∥ ∇ P ∥ , on Q r , |\nabla^G P|\le C\|\nabla P\|,\qquad|P_\varphi-a|\le Cr^2\|\nabla P\|,\qquad|P_V|\le\rho\|\nabla P\|,\qquad\text{on }Q_r, ∣ ∇ G P ∣ ≤ C ∥∇ P ∥ , ∣ P φ − a ∣ ≤ C r 2 ∥∇ P ∥ , ∣ P V ∣ ≤ ρ ∥∇ P ∥ , on Q r ,
so the proof of (i), applied pointwise, implies
∣ E G 1 ∣ ≤ C r α ( ∣ a ∣ + ∥ ∇ P ∥ ) , ∣ E φ 0 ∣ + ρ − 1 ∣ E V 0 ∣ ≤ C ( ∣ a ∣ + ∥ ∇ P ∥ ) . |\mathbf{E}^1_G|\le Cr^\alpha(|a|+\|\nabla P\|),\qquad|\mathbf{E}^0_\varphi|+\rho^{-1}|\mathbf{E}^0_V|\le C(|a|+\|\nabla P\|). ∣ E G 1 ∣ ≤ C r α ( ∣ a ∣ + ∥∇ P ∥ ) , ∣ E φ 0 ∣ + ρ − 1 ∣ E V 0 ∣ ≤ C ( ∣ a ∣ + ∥∇ P ∥ ) .
Then the dual error estimate (7.4) follows from Lemma 5.3(ii). Together with the transformed C 1 , α C^{1,\alpha} C 1 , α -estimates (5.18) of Proposition 5.4(ii), the same computation shows (7.5).
(iii) This follows essentially from the Leibniz rule. Let p t = p 0 + t ( a , 0 ) p_t=p_0+t(a,0) p t = p 0 + t ( a , 0 ) . Using the exponential gauge and parallel transport, we transport Φ p 0 \Phi_{p_0} Φ p 0 to a smooth family of flat Dirac models Φ t \Phi_t Φ t centred at p t p_t p t , each minimising the energy frozen at p t p_t p t . Their transformed models have the form
U t = ( g ( y 0 ) + t a + B t ( y − y 0 ) + q t ( ξ ) , M t ξ ) , U 0 = U p 0 , U_t=(g(y_0)+ta+B_t(y-y_0)+q_t(\xi),M_t\xi),\qquad U_0=U_{p_0}, U t = ( g ( y 0 ) + t a + B t ( y − y 0 ) + q t ( ξ ) , M t ξ ) , U 0 = U p 0 ,
and satisfy
∫ A p r 1 ( p t , D U t ) : D Y = 0 , \int\mathcal{A}_{\mathrm{pr}}^1(p_t,DU_t):DY=0, ∫ A pr 1 ( p t , D U t ) : D Y = 0 ,
@page 31/42 starts_mid=0 ends_mid=1 The following lemma collects the estimates comparing B U \mathcal{B}_U B U with B p 0 \mathcal{B}_{p_0} B p 0 that are used below.
Comparison with the frozen operator
Lemma 7.1 (Comparison with the frozen operator). For 0 < r ≤ r 0 ≪ 1 0<r\le r_0\ll1 0 < r ≤ r 0 ≪ 1 , the following hold on Q r = Q r ( y 0 ) Q_r=Q_r(y_0) Q r = Q r ( y 0 ) .
(i) For Z ∈ W G 1 , 2 ( Q r ) Z\in W^{1,2}_G(Q_r) Z ∈ W G 1 , 2 ( Q r ) and Y ∈ W G , 0 1 , 2 ( Q r ) Y\in W^{1,2}_{G,0}(Q_r) Y ∈ W G , 0 1 , 2 ( Q r ) ,
∣ B U ( Z , Y ) − B p 0 ( Z , Y ) ∣ ≤ C r α ( ∥ ∇ G Z ∥ L 2 + r − 2 ∥ Z φ ∥ L 2 ) ∥ ∇ G Y ∥ L 2 . (98) \left|\mathcal{B}_U(Z,Y)-\mathcal{B}_{p_0}(Z,Y)\right|
\le Cr^\alpha\left(\|\nabla^G Z\|_{L^2}+r^{-2}\|Z_\varphi\|_{L^2}\right)\|\nabla^G Y\|_{L^2}. \tag*{(98)} ∣ B U ( Z , Y ) − B p 0 ( Z , Y ) ∣ ≤ C r α ( ∥ ∇ G Z ∥ L 2 + r − 2 ∥ Z φ ∥ L 2 ) ∥ ∇ G Y ∥ L 2 . (98) In particular, B U \mathcal{B}_U B U is uniformly coercive on W G , 0 1 , 2 ( Q r ) W^{1,2}_{G,0}(Q_r) W G , 0 1 , 2 ( Q r ) .
(ii) For P ∈ P P\in\mathcal{P} P ∈ P with constant term a a a , the error field E = A U P − A p r , p 0 P − N p 0 a \mathbf{E}=\mathcal{A}_U P-\mathcal{A}_{\mathrm{pr},p_0}P-\mathcal{N}_{p_0}a E = A U P − A pr , p 0 P − N p 0 a satisfies
r − N G / 2 ∥ E ∥ ( W G 1 , 2 ( Q r ) ) ∗ ≤ C r α ( ∣ a ∣ + ∥ ∇ P ∥ ) , (99) r^{-N_G/2}\|\mathbf{E}\|_{(W^{1,2}_G(Q_r))^*}\le Cr^\alpha(|a|+\|\nabla P\|), \tag*{(99)} r − N G /2 ∥ E ∥ ( W G 1 , 2 ( Q r ) ) ∗ ≤ C r α ( ∣ a ∣ + ∥∇ P ∥ ) , (99) ∥ E ∥ C G 0 , α ( Q r ) ≤ C r α ( ∣ a ∣ + ∥ ∇ P ∥ ) . (100) \|\mathbf{E}\|_{C^{0,\alpha}_G(Q_r)}\le Cr^\alpha(|a|+\|\nabla P\|). \tag*{(100)} ∥ E ∥ C G 0 , α ( Q r ) ≤ C r α ( ∣ a ∣ + ∥∇ P ∥ ) . (100) (iii) For every a ∈ C a\in\mathbb{C} a ∈ C there is P a ∈ P P_a\in\mathcal{P} P a ∈ P with constant term a a a such that N p 0 a = − A p 0 P a \mathcal{N}_{p_0}a=-\mathcal{A}_{p_0}P_a N p 0 a = − A p 0 P a . In particular,
N p 0 a ∈ A p 0 ( P ) , ∥ R p 0 N p 0 a ∥ 1 ≤ C ∣ a ∣ . (101) \mathcal{N}_{p_0}a\in\mathcal{A}_{p_0}(\mathcal{P}),\qquad\|\mathcal{R}_{p_0}\mathcal{N}_{p_0}a\|_1\le C|a|. \tag*{(101)} N p 0 a ∈ A p 0 ( P ) , ∥ R p 0 N p 0 a ∥ 1 ≤ C ∣ a ∣. (101)
Proof. The main ingredients are the structure (5.21) of the transformed density, the deviation estimate (7.1), and the weighted inequalities in Lemma 5.3.
(i) Write the second variation as B U ( Z , Y ) = ∫ ( a ( D Z , D Y ) + b ( Z , D Y ) + b ( Y , D Z ) + c ( Z , Y ) ) \mathcal{B}_U(Z,Y)=\int(a(DZ,DY)+b(Z,DY)+b(Y,DZ)+c(Z,Y)) B U ( Z , Y ) = ∫ ( a ( D Z , D Y ) + b ( Z , D Y ) + b ( Y , D Z ) + c ( Z , Y )) , with coefficients given by second derivatives of f ^ \hat f f ^ at U U U . By (5.21), the principal part is D D U 2 f ^ p r ( X , ∇ G U ) D^2_{DU}\hat f_{\mathrm{pr}}(X,\nabla^G U) D D U 2 f ^ pr ( X , ∇ G U ) up to O ( ρ 2 ) O(\rho^2) O ( ρ 2 ) , and B p 0 \mathcal{B}_{p_0} B p 0 is obtained by evaluating it at ( p 0 , ∇ G U p 0 ) (p_0,\nabla^G U_{p_0}) ( p 0 , ∇ G U p 0 ) . By (7.1), the difference of the principal parts is bounded by C r α ∣ ∇ G Z ∣ ∣ ∇ G Y ∣ Cr^\alpha|\nabla^G Z||\nabla^G Y| C r α ∣ ∇ G Z ∣∣ ∇ G Y ∣ . The remaining terms are bounded by
C ( ∣ Z φ ∣ + ρ ∣ Z V ∣ + ρ 2 ∣ ∇ G Z ∣ ) ∣ ∇ G Y ∣ + C ( ∣ Y φ ∣ + ρ ∣ Y V ∣ ) ∣ ∇ G Z ∣ + C ( ∣ Z φ ∣ + ∣ Z V ∣ ) ( ∣ Y φ ∣ + ∣ Y V ∣ ) . C(|Z_\varphi|+\rho|Z_V|+\rho^2|\nabla^G Z|)|\nabla^G Y|
+C(|Y_\varphi|+\rho|Y_V|)|\nabla^G Z|
+C(|Z_\varphi|+|Z_V|)(|Y_\varphi|+|Y_V|). C ( ∣ Z φ ∣ + ρ ∣ Z V ∣ + ρ 2 ∣ ∇ G Z ∣ ) ∣ ∇ G Y ∣ + C ( ∣ Y φ ∣ + ρ ∣ Y V ∣ ) ∣ ∇ G Z ∣ + C ( ∣ Z φ ∣ + ∣ Z V ∣ ) ( ∣ Y φ ∣ + ∣ Y V ∣ ) .
By Lemma 5.3,
∥ Z V ∥ L 2 ≤ r ∥ ∇ G Z ∥ L 2 , ∥ Y φ ∥ L 2 + r ∥ Y V ∥ L 2 ≤ C r 2 ∥ ∇ G Y ∥ L 2 . \|Z_V\|_{L^2}\le r\|\nabla^G Z\|_{L^2},\qquad
\|Y_\varphi\|_{L^2}+r\|Y_V\|_{L^2}\le Cr^2\|\nabla^G Y\|_{L^2}. ∥ Z V ∥ L 2 ≤ r ∥ ∇ G Z ∥ L 2 , ∥ Y φ ∥ L 2 + r ∥ Y V ∥ L 2 ≤ C r 2 ∥ ∇ G Y ∥ L 2 .
The Cauchy–Schwarz inequality then bounds the remaining terms by
C ( ∥ Z φ ∥ L 2 + r 2 ∥ ∇ G Z ∥ L 2 ) ∥ ∇ G Y ∥ L 2 = C r 2 ( ∥ ∇ G Z ∥ L 2 + r − 2 ∥ Z φ ∥ L 2 ) ∥ ∇ G Y ∥ L 2 , C(\|Z_\varphi\|_{L^2}+r^2\|\nabla^G Z\|_{L^2})\|\nabla^G Y\|_{L^2}
=Cr^2(\|\nabla^G Z\|_{L^2}+r^{-2}\|Z_\varphi\|_{L^2})\|\nabla^G Y\|_{L^2}, C ( ∥ Z φ ∥ L 2 + r 2 ∥ ∇ G Z ∥ L 2 ) ∥ ∇ G Y ∥ L 2 = C r 2 ( ∥ ∇ G Z ∥ L 2 + r − 2 ∥ Z φ ∥ L 2 ) ∥ ∇ G Y ∥ L 2 ,
which proves (7.3). For compactly supported Z Z Z , we have r − 2 ∥ Z φ ∥ L 2 ≤ C ∥ ∇ G Z ∥ L 2 r^{-2}\|Z_\varphi\|_{L^2}\le C\|\nabla^G Z\|_{L^2} r − 2 ∥ Z φ ∥ L 2 ≤ C ∥ ∇ G Z ∥ L 2 , so coercivity of B U \mathcal{B}_U B U follows from the coercivity (6.2) of the frozen form.
(ii) Write P = ( a , 0 ) + ( P − ( a , 0 ) ) P=(a,0)+(P-(a,0)) P = ( a , 0 ) + ( P − ( a , 0 )) . Since D U DU D U is unchanged by the constant variation ( a , 0 ) (a,0) ( a , 0 ) ,
A U ( a , 0 ) = d d t ∣ t = 0 A p r ( X + t ( a , 0 ) , D U ) + O ( ρ 2 ∣ a ∣ ) , \mathcal{A}_U(a,0)
=\left.\frac{d}{dt}\right|_{t=0}\mathcal{A}_{\mathrm{pr}}(X+t(a,0),DU)+O(\rho^2|a|), A U ( a , 0 ) = d t d t = 0 A pr ( X + t ( a , 0 ) , D U ) + O ( ρ 2 ∣ a ∣ ) ,
where the O ( ρ 2 ∣ a ∣ ) O(\rho^2|a|) O ( ρ 2 ∣ a ∣ ) -remainder is measured by the weighted C 0 , α C^{0,\alpha} C 0 , α -norm (5.15). Compared with N p 0 a \mathcal{N}_{p_0}a N p 0 a , the coefficients are evaluated at ( X , D U ) (X,DU) ( X , D U ) instead of ( p 0 , D U p 0 ) (p_0,DU_{p_0}) ( p 0 , D U p 0 ) , so (7.1) gives an error O ( r α ∣ a ∣ ) O(r^\alpha|a|) O ( r α ∣ a ∣ ) . The non-constant part satisfies
∣ ∇ G P ∣ ≤ C ∥ ∇ P ∥ , ∣ P φ − a ∣ ≤ C r 2 ∥ ∇ P ∥ , ∣ P V ∣ ≤ ρ ∥ ∇ P ∥ , on Q r , |\nabla^G P|\le C\|\nabla P\|,\qquad
|P_\varphi-a|\le Cr^2\|\nabla P\|,\qquad
|P_V|\le\rho\|\nabla P\|,\qquad\text{on }Q_r, ∣ ∇ G P ∣ ≤ C ∥∇ P ∥ , ∣ P φ − a ∣ ≤ C r 2 ∥∇ P ∥ , ∣ P V ∣ ≤ ρ ∥∇ P ∥ , on Q r ,
so the proof of (i), applied pointwise, implies
∣ E G 1 ∣ ≤ C r α ( ∣ a ∣ + ∥ ∇ P ∥ ) , ∣ E φ 0 ∣ + ρ − 1 ∣ E V 0 ∣ ≤ C ( ∣ a ∣ + ∥ ∇ P ∥ ) . |\mathbf{E}^1_G|\le Cr^\alpha(|a|+\|\nabla P\|),\qquad
|\mathbf{E}^0_\varphi|+\rho^{-1}|\mathbf{E}^0_V|
\le C(|a|+\|\nabla P\|). ∣ E G 1 ∣ ≤ C r α ( ∣ a ∣ + ∥∇ P ∥ ) , ∣ E φ 0 ∣ + ρ − 1 ∣ E V 0 ∣ ≤ C ( ∣ a ∣ + ∥∇ P ∥ ) .
Then the dual error estimate (7.4) follows from Lemma 5.3(ii). Together with the transformed C 1 , α C^{1,\alpha} C 1 , α -estimates (5.18) of Proposition 5.4(ii), the same computation shows (7.5).
(iii) This follows essentially from the Leibniz rule. Let p t = p 0 + t ( a , 0 ) p_t=p_0+t(a,0) p t = p 0 + t ( a , 0 ) . Using the exponential gauge and parallel transport, we transport Φ p 0 \Phi_{p_0} Φ p 0 to a smooth family of flat Dirac models Φ t \Phi_t Φ t centred at p t p_t p t , each minimising the energy frozen at p t p_t p t . Their transformed models have the form
U t = ( g ( y 0 ) + t a + B t ( y − y 0 ) + q t ( ξ ) , M t ξ ) , U 0 = U p 0 , U_t=(g(y_0)+ta+B_t(y-y_0)+q_t(\xi),M_t\xi),\qquad U_0=U_{p_0}, U t = ( g ( y 0 ) + t a + B t ( y − y 0 ) + q t ( ξ ) , M t ξ ) , U 0 = U p 0 ,
and satisfy
∫ A p r 1 ( p t , D U t ) : D Y = 0 , (102) \int\mathcal{A}_{\mathrm{pr}}^1(p_t,DU_t):DY=0, \tag*{(102)} ∫ A pr 1 ( p t , D U t ) : D Y = 0 , (102)
for every compactly supported admissible Y Y Y . Differentiating at t = 0 t=0 t = 0 implies
N p 0 a = − A p 0 P a , where P a : = d d t ∣ t = 0 U t ∈ P , \mathcal{N}_{p_0}a=-\mathcal{A}_{p_0}P_a,\qquad\text{where }P_a:=\left.\frac{d}{dt}\right|_{t=0}U_t\in\mathcal{P}, N p 0 a = − A p 0 P a , where P a := d t d t = 0 U t ∈ P ,
whose constant term is a a a . The bound (7.6) follows from the bound (91) of the right inverse in Lemma 6.5 . □
Remark 7.2. For h ∈ R n − 2 h\in\mathbb{R}^{n-2} h ∈ R n − 2 with ∣ h ∣ ≤ r 2 / 100 |h|\le r^2/100 ∣ h ∣ ≤ r 2 /100 , write U h ( ξ , y ) = U ( ξ , y + h ) U_h(\xi,y)=U(\xi,y+h) U h ( ξ , y ) = U ( ξ , y + h ) and U h , t = ( 1 − t ) U + t U h U_{h,t}=(1-t)U+tU_h U h , t = ( 1 − t ) U + t U h for t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] . The proof of Lemma 7.1(i) applies verbatim to the following bilinear forms:
B h ( Z , Y ) = d d s ∣ s = 0 ∫ ( A 1 ( ξ , y + h , U h + s Z , D U h + s D Z ) : D Y + A 0 ( ξ , y + h , U h + s Z , D U h + s D Z ) ⋅ Y ) . (103) \mathcal{B}_h(Z,Y)=\left.\frac{d}{ds}\right|_{s=0}\int\left(\mathbf{A}^{1}(\xi,y+h,U_h+sZ,DU_h+sDZ):DY+\mathbf{A}^{0}(\xi,y+h,U_h+sZ,DU_h+sDZ)\cdot Y\right). \tag*{(103)} B h ( Z , Y ) = d s d s = 0 ∫ ( A 1 ( ξ , y + h , U h + s Z , D U h + sD Z ) : D Y + A 0 ( ξ , y + h , U h + s Z , D U h + sD Z ) ⋅ Y ) . (103)
B ‾ h ( Z , Y ) = d d s ∣ s = 0 ∫ 0 1 ∫ ( A 1 ( ξ , y + t h , U h , t + s Z , D U h , t + s D Z ) : D Y + A 0 ( ξ , y + t h , U h , t + s Z , D U h , t + s D Z ) ⋅ Y ) d t . (104) \overline{\mathcal{B}}_h(Z,Y)=\left.\frac{d}{ds}\right|_{s=0}\int_{0}^{1}\int\left(\mathbf{A}^{1}(\xi,y+th,U_{h,t}+sZ,DU_{h,t}+sDZ):DY+\mathbf{A}^{0}(\xi,y+th,U_{h,t}+sZ,DU_{h,t}+sDZ)\cdot Y\right)\,dt. \tag*{(104)} B h ( Z , Y ) = d s d s = 0 ∫ 0 1 ∫ ( A 1 ( ξ , y + t h , U h , t + s Z , D U h , t + sD Z ) : D Y + A 0 ( ξ , y + t h , U h , t + s Z , D U h , t + sD Z ) ⋅ Y ) d t . (104)
As a consequence, we obtain the following weighted Caccioppoli inequality with source F \mathcal{F} F .
Corollary 7.3. Let B \mathcal{B} B be B U \mathcal{B}_U B U or one of the forms in Remark 7.2. If Z ∈ W G 1 , 2 ( Q r ( y 0 ) ) Z\in W^{1,2}_{G}(Q_r(y_0)) Z ∈ W G 1 , 2 ( Q r ( y 0 )) satisfies B ( Z , Y ) = F ( Y ) \mathcal{B}(Z,Y)=\mathcal{F}(Y) B ( Z , Y ) = F ( Y ) for every compactly supported admissible Y Y Y , then
∥ ∇ G Z ∥ L 2 ( Q r / 2 ( y 0 ) ) ≤ C ( r − 2 ∥ Z ∥ L G 2 ( Q r ( y 0 ) ) + ∥ F ∥ ( W G 1 , 2 ( Q r ( y 0 ) ) ) ∗ ) . (105) \|\nabla_G Z\|_{L^2(Q_{r/2}(y_0))}\le C\left(r^{-2}\|Z\|_{L^2_G(Q_r(y_0))}+\|\mathcal{F}\|_{(W^{1,2}_{G}(Q_r(y_0)))^*}\right). \tag*{(105)} ∥ ∇ G Z ∥ L 2 ( Q r /2 ( y 0 )) ≤ C ( r − 2 ∥ Z ∥ L G 2 ( Q r ( y 0 )) + ∥ F ∥ ( W G 1 , 2 ( Q r ( y 0 )) ) ∗ ) . (105)
Proof. Let η = η 1 ( ρ ) η 2 ( y ) \eta=\eta_1(\rho)\eta_2(y) η = η 1 ( ρ ) η 2 ( y ) be a cutoff of Q r / 2 ⊂ Q r Q_{r/2}\subset Q_r Q r /2 ⊂ Q r with η 1 = 1 \eta_1=1 η 1 = 1 for ρ ≤ r / 2 \rho\le r/2 ρ ≤ r /2 . Since B \mathcal{B} B is symmetric, B ( η Z , η Z ) = B ( Z , η 2 Z ) + ∫ a ( Z ⊗ D η , Z ⊗ D η ) \mathcal{B}(\eta Z,\eta Z)=\mathcal{B}(Z,\eta^2Z)+\int a(Z\otimes D\eta,Z\otimes D\eta) B ( η Z , η Z ) = B ( Z , η 2 Z ) + ∫ a ( Z ⊗ D η , Z ⊗ D η ) , as in Step 1 of the proof of Proposition 6.1(ii) without source term. Since D ξ η D_\xi\eta D ξ η is supported where ρ ≥ r / 2 \rho\ge r/2 ρ ≥ r /2 , the weighted bounds on a a a give ∣ a ( Z ⊗ D η , Z ⊗ D η ) ∣ ≤ C r − 4 ( ∣ Z ϕ ∣ 2 + ρ 2 ∣ Z V ∣ 2 ) |a(Z\otimes D\eta,Z\otimes D\eta)|\le Cr^{-4}(|Z_\phi|^2+\rho^2|Z_V|^2) ∣ a ( Z ⊗ D η , Z ⊗ D η ) ∣ ≤ C r − 4 ( ∣ Z ϕ ∣ 2 + ρ 2 ∣ Z V ∣ 2 ) . The conclusion follows from the coercivity in Lemma 7.1(i) and ∣ F ( η 2 Z ) ∣ ≤ ∥ F ∥ ( W G 1 , 2 ) ∗ ∥ ∇ G ( η 2 Z ) ∥ L 2 |\mathcal{F}(\eta^2Z)|\le\|\mathcal{F}\|_{(W^{1,2}_{G})^*}\|\nabla_G(\eta^2Z)\|_{L^2} ∣ F ( η 2 Z ) ∣ ≤ ∥ F ∥ ( W G 1 , 2 ) ∗ ∥ ∇ G ( η 2 Z ) ∥ L 2 . □
Campanato iteration
By Lemma 6.5 and Lemma 7.1(iii), for F ∈ A p 0 ( P ) \mathcal{F}\in\mathcal{A}_{p_0}(\mathcal{P}) F ∈ A p 0 ( P ) , the polynomials solving the frozen equation with source F \mathcal{F} F , corrected for the motion of the centre, form an affine space
P ( F ) = { P ∈ P : A p 0 P + N p 0 a P = F } = R p 0 F + { H − R p 0 N p 0 a H : H ∈ ker A p 0 } . (106) \mathcal{P}(\mathcal{F})=\{P\in\mathcal{P}:\mathcal{A}_{p_0}P+\mathcal{N}_{p_0}a_P=\mathcal{F}\}=\mathcal{R}_{p_0}\mathcal{F}+\{H-\mathcal{R}_{p_0}\mathcal{N}_{p_0}a_H:H\in\ker\mathcal{A}_{p_0}\}. \tag*{(106)} P ( F ) = { P ∈ P : A p 0 P + N p 0 a P = F } = R p 0 F + { H − R p 0 N p 0 a H : H ∈ ker A p 0 } . (106)
Theorem 7.4 (Campanato estimate). Let F p r ∈ A p 0 ( P ) \mathcal{F}_{\mathrm{pr}}\in\mathcal{A}_{p_0}(\mathcal{P}) F pr ∈ A p 0 ( P ) , and suppose that Z ∈ W G 1 , 2 ( Q r 0 ( y 0 ) ) Z\in W^{1,2}_{G}(Q_{r_0}(y_0)) Z ∈ W G 1 , 2 ( Q r 0 ( y 0 )) satisfies B U ( Z , Y ) = F ( Y ) \mathcal{B}_U(Z,Y)=\mathcal{F}(Y) B U ( Z , Y ) = F ( Y ) for every compactly supported admissible Y Y Y , where
∥ F − F p r ∥ ( W G 1 , 2 ( Q r ( y 0 ) ) ) ∗ ≤ C r N G / 2 + α , 0 < r ≤ r 0 . (107) \|\mathcal{F}-\mathcal{F}_{\mathrm{pr}}\|_{(W^{1,2}_{G}(Q_r(y_0)))^*}\le Cr^{N_G/2+\alpha},\qquad0<r\le r_0. \tag*{(107)} ∥ F − F pr ∥ ( W G 1 , 2 ( Q r ( y 0 )) ) ∗ ≤ C r N G /2 + α , 0 < r ≤ r 0 . (107) Then there exists a unique comparison polynomial P Z , y 0 ∈ P ( F p r ) P_{Z,y_0}\in\mathcal{P}(\mathcal{F}_{\mathrm{pr}}) P Z , y 0 ∈ P ( F pr ) of Z Z Z at ( 0 , y 0 ) (0,y_0) ( 0 , y 0 ) with estimates,
E r ( Z − P Z , y 0 ) ≤ C r 2 α , 0 < r ≤ r 0 , (108) \mathcal{E}_r(Z-P_{Z,y_0})\le Cr^{2\alpha},\qquad0<r\le r_0, \tag*{(108)} E r ( Z − P Z , y 0 ) ≤ C r 2 α , 0 < r ≤ r 0 , (108) where the constant depends linearly on the data at scale r 0 r_0 r 0 .
Proof. For 0 < r ≤ r 0 0<r\le r_0 0 < r ≤ r 0 , let P r P_r P r minimise E r ( Z − P ) \mathcal{E}_r(Z-P) E r ( Z − P ) over P ( F p r ) \mathcal{P}(\mathcal{F}_{\mathrm{pr}}) P ( F pr ) , and set e ( r ) = E r ( Z − P r ) 1 / 2 e(r)=\mathcal{E}_r(Z-P_r)^{1/2} e ( r ) = E r ( Z − P r ) 1/2 : this minimiser exists and is unique, since E r ( Z − ⋅ ) \mathcal{E}_r(Z-\cdot) E r ( Z − ⋅ ) is a coercive strictly convex quadratic functional (cf. (72) ) on the finite-dimensional affine space P ( F p r ) \mathcal{P}(\mathcal{F}_{\mathrm{pr}}) P ( F pr ) . As A p 0 P r + N p 0 a P r = F p r \mathcal{A}_{p_0}P_r+\mathcal{N}_{p_0}a_{P_r}=\mathcal{F}_{\mathrm{pr}} A p 0 P r + N p 0 a P r = F pr , the field W = Z − P r W=Z-P_r W = Z − P r then satisfies
B U ( W , Y ) = ( F − F p r ) ( Y ) − ∫ ( E 1 : D Y + E 0 ⋅ Y ) . (109) \mathcal{B}_U(W,Y)=(\mathcal{F}-\mathcal{F}_{\mathrm{pr}})(Y)-\int\left(\mathbf{E}^{1}:DY+\mathbf{E}^{0}\cdot Y\right). \tag*{(109)} B U ( W , Y ) = ( F − F pr ) ( Y ) − ∫ ( E 1 : D Y + E 0 ⋅ Y ) . (109)
where E \mathbf{E} E is the error field of P r P_r P r . By (107) and the error estimate (7.4) of Lemma 7.1(ii), the right-hand side has ( W G 1 , 2 ( Q r ( y 0 ) ) ) ∗ (W^{1,2}_{G}(Q_r(y_0)))^* ( W G 1 , 2 ( Q r ( y 0 )) ) ∗ -norm O ( r N G / 2 + α ( 1 + ∥ P r ∥ r 0 ) ) O(r^{N_G/2+\alpha}(1+\|P_r\|_{r_0})) O ( r N G /2 + α ( 1 + ∥ P r ∥ r 0 )) . Let now W 0 W_0 W 0 solve B p 0 ( W 0 , Y ) = 0 \mathcal{B}_{p_0}(W_0,Y)=0 B p 0 ( W 0 , Y ) = 0 on Q r / 2 ( y 0 ) Q_{r/2}(y_0) Q r /2 ( y 0 ) with W − W 0 ∈ W G , 0 1 , 2 ( Q r / 2 ( y 0 ) ) W-W_0\in W^{1,2}_{G,0}(Q_{r/2}(y_0)) W − W 0 ∈ W G , 0 1 , 2 ( Q r /2 ( y 0 )) , whose existence is guaranteed by the coercivity estimate (80) of the frozen form and the Lax–Milgram theorem. Testing (109) and the frozen equation with W − W 0 W-W_0 W − W 0 and applying weighted Caccioppoli estimates (cf. Corollary 7.3), we find
∥ ∇ G ( W − W 0 ) ∥ L 2 2 ≲ ( B p 0 − B U ) ( W , W − W 0 ) + ( F − F p r ) ( W − W 0 ) − ∫ ( E 1 : D ( W − W 0 ) + E 0 ⋅ ( W − W 0 ) ) ≲ ( r α ( ∥ ∇ G W ∥ L 2 ( Q r / 2 ) + r − 2 ∥ W φ ∥ L 2 ( Q r / 2 ) ) + r N G / 2 + α ( 1 + ∥ P r ∥ r 0 ) ) ∥ ∇ G ( W − W 0 ) ∥ L 2 ≲ r N G / 2 + α ( e ( r ) + 1 + ∥ P r ∥ r 0 ) ∥ ∇ G ( W − W 0 ) ∥ L 2 . \begin{aligned}
\lVert\nabla^G(W-W_0)\rVert_{L^2}^2 &\lesssim(\mathcal{B}_{p_0}-\mathcal{B}_U)(W,W-W_0)\\
&\quad+(\mathcal{F}-\mathcal{F}_{\mathrm{pr}})(W-W_0)-\int\left(\mathbf{E}^1:D(W-W_0)+\mathbf{E}^0\cdot(W-W_0)\right)\\
&\lesssim\left(r^\alpha\left(\lVert\nabla^G W\rVert_{L^2(Q_{r/2})}+r^{-2}\lVert W\varphi\rVert_{L^2(Q_{r/2})}\right)\right.\\
&\qquad\left.+r^{N_G/2+\alpha}(1+\lVert P_r\rVert_{r_0})\right)\lVert\nabla^G(W-W_0)\rVert_{L^2}\\
&\lesssim r^{N_G/2+\alpha}\left(e(r)+1+\lVert P_r\rVert_{r_0}\right)\lVert\nabla^G(W-W_0)\rVert_{L^2}.
\end{aligned} ∥ ∇ G ( W − W 0 ) ∥ L 2 2 ≲ ( B p 0 − B U ) ( W , W − W 0 ) + ( F − F pr ) ( W − W 0 ) − ∫ ( E 1 : D ( W − W 0 ) + E 0 ⋅ ( W − W 0 ) ) ≲ ( r α ( ∥ ∇ G W ∥ L 2 ( Q r /2 ) + r − 2 ∥ W φ ∥ L 2 ( Q r /2 ) ) + r N G /2 + α ( 1 + ∥ P r ∥ r 0 ) ) ∥ ∇ G ( W − W 0 ) ∥ L 2 ≲ r N G /2 + α ( e ( r ) + 1 + ∥ P r ∥ r 0 ) ∥ ∇ G ( W − W 0 ) ∥ L 2 .
By weighted Poincaré inequalities, this implies
E r / 2 ( W − W 0 ) 1 / 2 ≤ C r − N G / 2 ∥ ∇ G ( W − W 0 ) ∥ L 2 ( Q r / 2 ) ≤ C r α ( e ( r ) + 1 + ∥ P r ∥ r 0 ) . (110) \mathcal{E}_{r/2}(W-W_0)^{1/2}\leq Cr^{-N_G/2}\lVert\nabla^G(W-W_0)\rVert_{L^2(Q_{r/2})}\leq Cr^\alpha(e(r)+1+\lVert P_r\rVert_{r_0}). \tag*{(110)} E r /2 ( W − W 0 ) 1/2 ≤ C r − N G /2 ∥ ∇ G ( W − W 0 ) ∥ L 2 ( Q r /2 ) ≤ C r α ( e ( r ) + 1 + ∥ P r ∥ r 0 ) . (110)
By Proposition 6.1(ii), applied to scale r / 2 r/2 r /2 , there is H ∈ ker A p 0 H\in\ker\mathcal{A}_{p_0} H ∈ ker A p 0 with ∥ H ∥ r / 2 ≤ C E r / 2 ( W 0 ) 1 / 2 \lVert H\rVert_{r/2}\leq C\mathcal{E}_{r/2}(W_0)^{1/2} ∥ H ∥ r /2 ≤ C E r /2 ( W 0 ) 1/2 and E θ r ( W 0 − H ) ≤ C θ 4 E r / 2 ( W 0 ) \mathcal{E}_{\theta r}(W_0-H)\leq C\theta^4\mathcal{E}_{r/2}(W_0) E θ r ( W 0 − H ) ≤ C θ 4 E r /2 ( W 0 ) for θ < θ 0 / 2 ≪ 1 \theta<\theta_0/2\ll1 θ < θ 0 /2 ≪ 1 . As H H H may have a non-zero constant term, we consider the competitor P r + H − R p 0 N p 0 a H ∈ P ( F p r ) P_r+H-R_{p_0}N_{p_0}a_H\in\mathcal{P}(\mathcal{F}_{\mathrm{pr}}) P r + H − R p 0 N p 0 a H ∈ P ( F pr ) . Since the correction R p 0 N p 0 a H R_{p_0}N_{p_0}a_H R p 0 N p 0 a H is purely quadratic in ξ \xi ξ , we know E θ r ( R p 0 N p 0 a H ) 1 / 2 ≲ ∥ R p 0 N p 0 a H ∥ 1 ≲ ∣ a H ∣ ≲ r 2 ∥ H ∥ r / 2 \mathcal{E}_{\theta r}(R_{p_0}N_{p_0}a_H)^{1/2}\lesssim\lVert R_{p_0}N_{p_0}a_H\rVert_1\lesssim\lvert a_H\rvert\lesssim r^2\lVert H\rVert_{r/2} E θ r ( R p 0 N p 0 a H ) 1/2 ≲ ∥ R p 0 N p 0 a H ∥ 1 ≲ ∣ a H ∣ ≲ r 2 ∥ H ∥ r /2 (cf. Lemma 7.1 (iii)). Together with (110) , this gives
e ( θ r ) ≤ C 0 θ 2 e ( r ) + C θ r α ( e ( r ) + 1 + ∥ P r ∥ r 0 ) , e(\theta r)\leq C_0\theta^2e(r)+C\theta r^\alpha(e(r)+1+\lVert P_r\rVert_{r_0}), e ( θ r ) ≤ C 0 θ 2 e ( r ) + C θ r α ( e ( r ) + 1 + ∥ P r ∥ r 0 ) ,
with C 0 C_0 C 0 independent of θ \theta θ . Fixing θ \theta θ with C 0 θ 2 ≤ θ α / 4 C_0\theta^2\leq\theta^\alpha/4 C 0 θ 2 ≤ θ α /4 , and then r 0 r_0 r 0 with C θ r 0 α ≤ θ α / 4 C\theta r_0^\alpha\leq\theta^\alpha/4 C θ r 0 α ≤ θ α /4 , we obtain
e ( θ r ) ≤ 1 2 θ α e ( r ) + C r α ( 1 + ∥ P r ∥ r 0 ) . (111) e(\theta r)\leq\tfrac{1}{2}\theta^\alpha e(r)+Cr^\alpha(1+\lVert P_r\rVert_{r_0}). \tag*{(111)} e ( θ r ) ≤ 2 1 θ α e ( r ) + C r α ( 1 + ∥ P r ∥ r 0 ) . (111)
and
∥ P θ r − P r ∥ r ≲ E θ r ( P θ r − P r ) 1 / 2 ≤ e ( θ r ) + E θ r ( Z − P r ) 1 / 2 ≤ C ( e ( r ) + r α ( 1 + ∥ P r ∥ r 0 ) ) . (112) \lVert P_{\theta r}-P_r\rVert_r\lesssim\mathcal{E}_{\theta r}(P_{\theta r}-P_r)^{1/2}\leq e(\theta r)+\mathcal{E}_{\theta r}(Z-P_r)^{1/2}\leq C(e(r)+r^\alpha(1+\lVert P_r\rVert_{r_0})). \tag*{(112)} ∥ P θ r − P r ∥ r ≲ E θ r ( P θ r − P r ) 1/2 ≤ e ( θ r ) + E θ r ( Z − P r ) 1/2 ≤ C ( e ( r ) + r α ( 1 + ∥ P r ∥ r 0 )) . (112)
Theorem 7.4 follows from iterating (111) and (112) along r k = θ k r 0 r_k=\theta^k r_0 r k = θ k r 0 in the standard way. □ \square □
Tangential smoothness and real analyticity
We now prove Theorem 3 by induction on the order of tangential derivatives. Suppose that D y γ U D_y^\gamma U D y γ U has been controlled for all ∣ γ ∣ < k \lvert\gamma\rvert<k ∣ γ ∣ < k . Then at order k k k , differentiating the transformed weak equation (78) shows that Z β = D y β U Z_\beta=D_y^\beta U Z β = D y β U with ∣ β ∣ = k \lvert\beta\rvert=k ∣ β ∣ = k solves the linearised equation B U ( Z β , Y ) = F β ( Y ) \mathcal{B}_U(Z_\beta,Y)=\mathcal{F}_\beta(Y) B U ( Z β , Y ) = F β ( Y ) , whose source depends only on lower-order derivatives . To apply Theorem 7.4, we need two properties of F β \mathcal{F}_\beta F β :
Principal part. F β \mathcal{F}_\beta F β is close to its principal part F β , p r , y \mathcal{F}_{\beta,\mathrm{pr},y} F β , pr , y at every centre ( 0 , y ) (0,y) ( 0 , y ) . This requires pointwise and L ∞ L^\infty L ∞ /Hölder bounds on the lower-order derivatives, not only L 2 L^2 L 2 -bounds. We therefore upgrade the L 2 L^2 L 2 -excess decay of Theorem 7.4 to weighted C 1 , α C^{1,\alpha} C 1 , α -estimates (Proposition 8.1), using interior Schauder estimates at the scale of the distance to { ξ = 0 } \{\xi=0\} { ξ = 0 } .
Compatibility and polynomial corrections. The principal part lies in A p ( y ) ( P ) \mathcal{A}_{p(y)}(\mathcal{P}) A p ( y ) ( P ) . This follows by taking β \beta β -th derivative of the identity expressing that transformed models are critical points of the frozen energy, along a family of models following the branching locus (Proposition 8.2(ii)).
The smoothness of the tangential derivatives of U U U (and hence of the branching locus) shall follow from above discussion. To obtain real-analyticity, we use Kato’s trick [13 ] together with the diffeomorphism invariance of the transformed system and a power-series majorant to obtain factorial bounds on the tangential derivatives (cf. Theorem 8.3).
From L 2 L^2 L 2 excess decay to weighted C 1 , α C^{1,\alpha} C 1 , α -estimates
Let Ω = Q r ( y 0 ) \Omega= Q_r(y_0) Ω = Q r ( y 0 ) with 0 < r ≤ r 0 0 < r \le r_0 0 < r ≤ r 0 , and let Z Z Z be a field on Ω \Omega Ω which has a comparison polynomial
P Z , y = ( a Z ( y ) + B Z ( y ) ( y ′ − y ) + q Z ( y , ξ ) , M Z ( y ) ξ ) (113) P_{Z,y} = (a_Z(y) + B_Z(y)(y' - y) + q_Z(y,\xi), M_Z(y)\xi) \tag*{(113)} P Z , y = ( a Z ( y ) + B Z ( y ) ( y ′ − y ) + q Z ( y , ξ ) , M Z ( y ) ξ ) (113)
at every centre ( 0 , y ) ∈ Ω (0,y) \in\Omega ( 0 , y ) ∈ Ω . We define the following scaling-invariant X α X_\alpha X α -norm for fields Z Z Z ,
∥ Z ∥ X α ( Ω ) : = r − N G / 2 − 2 ∥ Z ∥ L G 2 ( Ω ) + r − N G / 2 ∥ ∇ G Z ∥ L 2 ( Ω ) + r − 2 ∥ a Z ∥ C 0 + ∥ D y a Z ∥ C 0 + ∥ q Z ∥ C 0 + ∥ M Z ∥ C 0 + r α ( [ D y a Z ] C 0 , α / 2 + [ q Z ] C 0 , α / 2 + [ M Z ] C 0 , α / 2 ) + sup y , s ( r s ) α E s ( Z − P Z , y ) 1 / 2 + sup ξ , y ( r ρ ) α ∥ ( ρ − 2 ( Z − P Z , y ) φ , ρ − 1 ( Z − P Z , y ) V ) ∥ C 1 , α ( Q ξ , y , g ρ ) . (114) \begin{aligned}
\|Z\|_{X_\alpha(\Omega)} := {}& r^{-N_G/2-2}\|Z\|_{L^2_G(\Omega)} + r^{-N_G/2}\|\nabla^G Z\|_{L^2(\Omega)} \\
&+ r^{-2}\|a_Z\|_{C^0} + \|D_y a_Z\|_{C^0} + \|q_Z\|_{C^0} + \|M_Z\|_{C^0} \\
&+ r^\alpha\left([D_y a_Z]_{C^{0,\alpha/2}} + [q_Z]_{C^{0,\alpha/2}} + [M_Z]_{C^{0,\alpha/2}}\right) \\
&+ \sup_{y,s}\left(\frac{r}{s}\right)^\alpha\mathcal{E}_s(Z-P_{Z,y})^{1/2} \\
&+ \sup_{\xi,y}\left(\frac{r}{\rho}\right)^\alpha
\left\|(\rho^{-2}(Z-P_{Z,y})_\varphi,\rho^{-1}(Z-P_{Z,y})_V)\right\|_{C^{1,\alpha}(Q_{\xi,y},g_\rho)}. \tag*{(114)}
\end{aligned} ∥ Z ∥ X α ( Ω ) := r − N G /2 − 2 ∥ Z ∥ L G 2 ( Ω ) + r − N G /2 ∥ ∇ G Z ∥ L 2 ( Ω ) + r − 2 ∥ a Z ∥ C 0 + ∥ D y a Z ∥ C 0 + ∥ q Z ∥ C 0 + ∥ M Z ∥ C 0 + r α ( [ D y a Z ] C 0 , α /2 + [ q Z ] C 0 , α /2 + [ M Z ] C 0 , α /2 ) + y , s sup ( s r ) α E s ( Z − P Z , y ) 1/2 + ξ , y sup ( ρ r ) α ( ρ − 2 ( Z − P Z , y ) φ , ρ − 1 ( Z − P Z , y ) V ) C 1 , α ( Q ξ , y , g ρ ) . (114)
Here the norms of coefficients are taken over B r n − 2 ( y 0 ) B_{r}^{n-2}(y_0) B r n − 2 ( y 0 ) ; the first supremum is over centres ( 0 , y ) ∈ Ω (0,y) \in\Omega ( 0 , y ) ∈ Ω and radii 0 < s ≤ r 0 < s \le r 0 < s ≤ r with Q s ( y ) ⊂ Ω Q_s(y) \subset\Omega Q s ( y ) ⊂ Ω ; the second supremum is over Q ξ , y ⊂ Ω Q_{\xi,y} \subset\Omega Q ξ , y ⊂ Ω with ρ = ∣ ξ ∣ > 0 \rho= |\xi| > 0 ρ = ∣ ξ ∣ > 0 . As B Z = D y a Z B_Z = D_y a_Z B Z = D y a Z (established below), the coefficient bounds and the last term in (114) imply
sup Ω ∖ { ξ = 0 } ( r − 2 ∣ Z φ ∣ + ρ − 1 ∣ Z V ∣ + ∣ ∇ G Z ∣ ) ≤ C ∥ Z ∥ X α ( Ω ) . \sup_{\Omega\setminus\{\xi=0\}}\left(r^{-2}|Z_\varphi|+\rho^{-1}|Z_V|+|\nabla^G Z|\right) \le C\|Z\|_{X_\alpha(\Omega)}. Ω ∖ { ξ = 0 } sup ( r − 2 ∣ Z φ ∣ + ρ − 1 ∣ Z V ∣ + ∣ ∇ G Z ∣ ) ≤ C ∥ Z ∥ X α ( Ω ) .
Correspondingly, for a source F \mathbf{F} F with prescribed principal coefficient fields F p r , y \mathbf{F}_{\mathrm{pr},y} F pr , y at the centres ( 0 , y ) (0,y) ( 0 , y ) , we define the scaling-invariant Y α Y_\alpha Y α -norm
∥ F ∥ Y α ( Ω ) = r − N G / 2 ∥ F ∥ ( W G 1 , 2 ( Ω ) ) ∗ + sup y ∥ F p r , y ∥ C G 0 , α ( Q 1 ( y ) ) + sup y , s ( r s ) α s − N G / 2 ∥ F − F p r , y ∥ ( W G 1 , 2 ( Q s ( y ) ) ) ∗ + sup ξ , y ( r ρ ) α ∥ F − F p r , y ∥ C G 0 , α ( Q ξ , y ) . (115) \begin{aligned}
\|\mathbf{F}\|_{Y_\alpha(\Omega)} ={}& r^{-N_G/2}\|\mathbf{F}\|_{(W_G^{1,2}(\Omega))^*} + \sup_y\|\mathbf{F}_{\mathrm{pr},y}\|_{C_G^{0,\alpha}(Q_1(y))} \\
&+ \sup_{y,s}\left(\frac{r}{s}\right)^\alpha s^{-N_G/2}
\|\mathbf{F}-\mathbf{F}_{\mathrm{pr},y}\|_{(W_G^{1,2}(Q_s(y)))^*} \\
&+ \sup_{\xi,y}\left(\frac{r}{\rho}\right)^\alpha
\|\mathbf{F}-\mathbf{F}_{\mathrm{pr},y}\|_{C_G^{0,\alpha}(Q_{\xi,y})}. \tag*{(115)}
\end{aligned} ∥ F ∥ Y α ( Ω ) = r − N G /2 ∥ F ∥ ( W G 1 , 2 ( Ω ) ) ∗ + y sup ∥ F pr , y ∥ C G 0 , α ( Q 1 ( y )) + y , s sup ( s r ) α s − N G /2 ∥ F − F pr , y ∥ ( W G 1 , 2 ( Q s ( y )) ) ∗ + ξ , y sup ( ρ r ) α ∥ F − F pr , y ∥ C G 0 , α ( Q ξ , y ) . (115)
with the suprema as above, Q s ( y ) , Q ξ , y ⊂ Ω Q_s(y), Q_{\xi,y} \subset\Omega Q s ( y ) , Q ξ , y ⊂ Ω .
We remark that the above norms just collect all the norms involved (adapted weighted W 1 , 2 W^{1,2} W 1 , 2 and C 1 , α C^{1,\alpha} C 1 , α norms) in the Campanato iteration procedure.
Proposition 8.1 (Weighted C 1 , α C^{1,\alpha} C 1 , α -estimates). Let 0 < 2 r ≤ r 0 0 < 2r \le r_0 0 < 2 r ≤ r 0 . Suppose that Z ∈ W G 1 , 2 ( Q 2 r ( y 0 ) ) Z \in W_G^{1,2}(Q_{2r}(y_0)) Z ∈ W G 1 , 2 ( Q 2 r ( y 0 )) satisfies B U ( Z , Y ) = F ( Y ) \mathcal{B}_U(Z,Y)=\mathcal{F}(Y) B U ( Z , Y ) = F ( Y ) for every compactly supported admissible Y Y Y , where ∥ F ∥ Y α ( Q 2 r ( y 0 ) ) < ∞ \|\mathcal{F}\|_{Y_\alpha(Q_{2r}(y_0))}<\infty ∥ F ∥ Y α ( Q 2 r ( y 0 )) < ∞ and F p r , y ∈ A p ( y ) ( P ) \mathbf{F}_{\mathrm{pr},y}\in\mathcal{A}_{p(y)}(\mathcal{P}) F pr , y ∈ A p ( y ) ( P ) at every centre. Then Z Z Z has a comparison polynomial P Z , y ∈ P ( F p r , y ) P_{Z,y}\in\mathcal{P}(\mathbf{F}_{\mathrm{pr},y}) P Z , y ∈ P ( F pr , y ) at every centre ( 0 , y ) ∈ Q r ( y 0 ) (0,y)\in Q_r(y_0) ( 0 , y ) ∈ Q r ( y 0 ) , with B Z = D y a Z B_Z=D_y a_Z B Z = D y a Z , and
∥ Z ∥ X α ( Q r ( y 0 ) ) ≤ C ( r − N G / 2 − 2 ∥ Z ∥ L G 2 ( Q 2 r ( y 0 ) ) + ∥ F ∥ Y α ( Q 2 r ( y 0 ) ) ) . (116) \|Z\|_{X_\alpha(Q_r(y_0))} \le C\left(r^{-N_G/2-2}\|Z\|_{L_G^2(Q_{2r}(y_0))}+\|\mathcal{F}\|_{Y_\alpha(Q_{2r}(y_0))}\right). \tag*{(116)} ∥ Z ∥ X α ( Q r ( y 0 )) ≤ C ( r − N G /2 − 2 ∥ Z ∥ L G 2 ( Q 2 r ( y 0 )) + ∥ F ∥ Y α ( Q 2 r ( y 0 )) ) . (116) In particular,
∣ D ξ Z φ ( ξ , y ) ∣ ≤ C ∣ ξ ∣ ( r − N G / 2 − 2 ∥ Z ∥ L G 2 ( Q 2 r ( y 0 ) ) + ∥ F ∥ Y α ( Q 2 r ( y 0 ) ) ) |D_\xi Z_\varphi(\xi,y)| \le C|\xi|\left(r^{-N_G/2-2}\|Z\|_{L_G^2(Q_{2r}(y_0))}+\|\mathcal{F}\|_{Y_\alpha(Q_{2r}(y_0))}\right) ∣ D ξ Z φ ( ξ , y ) ∣ ≤ C ∣ ξ ∣ ( r − N G /2 − 2 ∥ Z ∥ L G 2 ( Q 2 r ( y 0 )) + ∥ F ∥ Y α ( Q 2 r ( y 0 )) ) on Q r ( y 0 ) ∖ { ξ = 0 } Q_r(y_0)\setminus\{\xi=0\} Q r ( y 0 ) ∖ { ξ = 0 } .
If further Z ∈ W G , 0 1 , 2 ( Q 2 r ( y 0 ) ) Z\in W_{G,0}^{1,2}(Q_{2r}(y_0)) Z ∈ W G , 0 1 , 2 ( Q 2 r ( y 0 )) is supported in K ⋐ Q 2 r ( y 0 ) K\Subset Q_{2r}(y_0) K ⋐ Q 2 r ( y 0 ) , then
∥ Z ∥ X α ( Q 2 r ( y 0 ) ) ≤ C K ∥ F ∥ Y α ( Q 2 r ( y 0 ) ) . (117) \|Z\|_{X_\alpha(Q_{2r}(y_0))} \le C_K\|\mathcal{F}\|_{Y_\alpha(Q_{2r}(y_0))}. \tag*{(117)} ∥ Z ∥ X α ( Q 2 r ( y 0 )) ≤ C K ∥ F ∥ Y α ( Q 2 r ( y 0 )) . (117)
Proof. Since the equation is linear, it suffices to prove (116) under the normalisation r − N G / 2 − 2 ∥ Z ∥ L G 2 ( Q 2 r ( y 0 ) ) + ∥ F ∥ Y α ( Q 2 r ( y 0 ) ) ≤ 1 r^{-N_G/2-2}\|Z\|_{L_G^2(Q_{2r}(y_0))}+\|\mathcal{F}\|_{Y_\alpha(Q_{2r}(y_0))}\le1 r − N G /2 − 2 ∥ Z ∥ L G 2 ( Q 2 r ( y 0 )) + ∥ F ∥ Y α ( Q 2 r ( y 0 )) ≤ 1 . Throughout the proof, C C C is independent of r r r .
Step 1: Excess decay at each centre. Applying the weighted Caccioppoli estimate (Corollary 7.3, with the fixed interior radius ratio 3 / 4 3/4 3/4 ) to Q 2 r ( y 0 ) Q_{2r}(y_0) Q 2 r ( y 0 ) , we find
r − N G / 2 ∥ ∇ G Z ∥ L 2 ( Q 3 r / 2 ( y 0 ) ) ≤ C ( r − N G / 2 − 2 ∥ Z ∥ L G 2 ( Q 2 r ( y 0 ) ) + r − N G / 2 ∥ F ∥ ( W G 1 , 2 ( Q 2 r ( y 0 ) ) ) ∗ ) ≤ C . r^{-N_G/2}\|\nabla^G Z\|_{L^2(Q_{3r/2}(y_0))}
\le C\left(r^{-N_G/2-2}\|Z\|_{L_G^2(Q_{2r}(y_0))}+r^{-N_G/2}\|\mathcal{F}\|_{(W_G^{1,2}(Q_{2r}(y_0)))^*}\right)\le C. r − N G /2 ∥ ∇ G Z ∥ L 2 ( Q 3 r /2 ( y 0 )) ≤ C ( r − N G /2 − 2 ∥ Z ∥ L G 2 ( Q 2 r ( y 0 )) + r − N G /2 ∥ F ∥ ( W G 1 , 2 ( Q 2 r ( y 0 )) ) ∗ ) ≤ C .
Together with the normalisation, this bounds the first two terms of (114) . For ( 0 , y ) ∈ Q r ( y 0 ) (0,y)\in Q_r(y_0) ( 0 , y ) ∈ Q r ( y 0 ) , we have Q r ( y ) ⊂ Q 3 r / 2 ( y 0 ) Q_r(y)\subset Q_{3r/2}(y_0) Q r ( y ) ⊂ Q 3 r /2 ( y 0 ) , so (67) implies
r − N G / 2 − 2 ∥ Z φ ∥ L 2 ( Q r ( y ) ) + r − N G / 2 − 1 ∥ Z V ∥ L 2 ( Q r ( y ) ) ≤ C . r^{-N_G/2-2}\|Z_\varphi\|_{L^2(Q_r(y))}+r^{-N_G/2-1}\|Z_V\|_{L^2(Q_r(y))}\le C. r − N G /2 − 2 ∥ Z φ ∥ L 2 ( Q r ( y )) + r − N G /2 − 1 ∥ Z V ∥ L 2 ( Q r ( y )) ≤ C .
Thus Theorem 7.4 applies at scale r r r centred at ( 0 , y ) (0,y) ( 0 , y ) , and we obtain a unique comparison polynomial P Z , y ∈ P ( F p r , y ) P_{Z,y} \in\mathcal{P}(\mathcal{F}_{\mathrm{pr},y}) P Z , y ∈ P ( F pr , y ) with estimates
( r s ) α E s 1 / 2 ( Z − P Z , y ) + r − 2 ∣ a Z ( y ) ∣ + ∥ ∇ P Z , y ∥ ≤ C , 0 < s ≤ r . \left(\frac{r}{s}\right)^{\alpha}\mathcal{E}_{s}^{1/2}(Z-P_{Z,y})+r^{-2}|a_Z(y)|+\|\nabla P_{Z,y}\|\leq C,\qquad0<s\leq r. ( s r ) α E s 1/2 ( Z − P Z , y ) + r − 2 ∣ a Z ( y ) ∣ + ∥∇ P Z , y ∥ ≤ C , 0 < s ≤ r .
Step 2: Estimates away from { ξ = 0 } \{\xi=0\} { ξ = 0 } . Since P Z , y ∈ P ( F p r , y ) P_{Z,y}\in\mathcal{P}(\mathcal{F}_{\mathrm{pr},y}) P Z , y ∈ P ( F pr , y ) , the difference Z − P Z , y Z-P_{Z,y} Z − P Z , y satisfies
B U ( Z − P Z , y , Y ) = ( F − F p r , y ) ( Y ) − ∫ ( E 1 : D Y + E 0 ⋅ Y ) , \mathcal{B}_{U}(Z-P_{Z,y},Y)=(\mathcal{F}-\mathcal{F}_{\mathrm{pr},y})(Y)-\int(\mathbf{E}^{1}:DY+\mathbf{E}^{0}\cdot Y), B U ( Z − P Z , y , Y ) = ( F − F pr , y ) ( Y ) − ∫ ( E 1 : D Y + E 0 ⋅ Y ) ,
where E \mathbf{E} E is the error field of P Z , y P_{Z,y} P Z , y at p ( y ) p(y) p ( y ) . Let ( ξ , y ) ∈ Q r ( y 0 ) (\xi,y)\in Q_r(y_0) ( ξ , y ) ∈ Q r ( y 0 ) with ρ = ∣ ξ ∣ > 0 \rho=|\xi|>0 ρ = ∣ ξ ∣ > 0 . With respect to the rescaled metric g ρ g_\rho g ρ of (70) , this is a uniformly elliptic system in divergence form on the annulus Q ξ , y Q_{\xi,y} Q ξ , y with coefficients uniformly bounded in C 0 , α C^{0,\alpha} C 0 , α , by Proposition 5.4 (ii) and Lemma 5.5 . The source has C 0 , α C^{0,\alpha} C 0 , α -norm at most C ( ρ / r ) α C(\rho/r)^\alpha C ( ρ / r ) α , by (115) , the polynomial coefficient bounds from Step 1, and (7.5). Moreover, Step 1 yields an L 2 L^2 L 2 -bound C ( ρ / r ) α C(\rho/r)^\alpha C ( ρ / r ) α for the normalised field ( ρ − 2 ( Z − P Z , y ) φ , ρ − 1 ( Z − P Z , y ) V ) (\rho^{-2}(Z-P_{Z,y})_\varphi,\rho^{-1}(Z-P_{Z,y})_V) ( ρ − 2 ( Z − P Z , y ) φ , ρ − 1 ( Z − P Z , y ) V ) with respect to the rescaled metric g ρ = ρ − 2 g ξ + ρ − 2 g y g_\rho=\rho^{-2}g_\xi+\rho^{-2}g_y g ρ = ρ − 2 g ξ + ρ − 2 g y . The interior elliptic estimate therefore implies
∥ ( ρ − 2 ( Z − P Z , y ) φ , ρ − 1 ( Z − P Z , y ) V ) ∥ C 1 , α ( Q ξ , y , g ρ ) ≤ C ( ρ r ) α , \left\|(\rho^{-2}(Z-P_{Z,y})_\varphi,\rho^{-1}(Z-P_{Z,y})_V)\right\|_{C^{1,\alpha}(Q_{\xi,y},g_\rho)}\leq C\left(\frac{\rho}{r}\right)^\alpha, ( ρ − 2 ( Z − P Z , y ) φ , ρ − 1 ( Z − P Z , y ) V ) C 1 , α ( Q ξ , y , g ρ ) ≤ C ( r ρ ) α ,
which bounds the last term of the X α \mathcal{X}_\alpha X α -norm (114) .
Step 3: Estimates of the comparison polynomials. Let ( 0 , y 1 ) , ( 0 , y 2 ) ∈ Q r ( y 0 ) (0,y_1),(0,y_2)\in Q_r(y_0) ( 0 , y 1 ) , ( 0 , y 2 ) ∈ Q r ( y 0 ) and set s = 4 ∣ y 1 − y 2 ∣ 1 / 2 ≤ r s=4|y_1-y_2|^{1/2}\leq r s = 4∣ y 1 − y 2 ∣ 1/2 ≤ r . Then Q s / 2 ( y 1 ) ⊂ Q s ( y 1 ) ∩ Q s ( y 2 ) Q_{s/2}(y_1)\subset Q_s(y_1)\cap Q_s(y_2) Q s /2 ( y 1 ) ⊂ Q s ( y 1 ) ∩ Q s ( y 2 ) . Step 1 and the triangle inequality imply
E s / 2 1 / 2 ( P Z , y 1 − P Z , y 2 ) ≤ C ( s r ) α , \mathcal{E}_{s/2}^{1/2}(P_{Z,y_1}-P_{Z,y_2})\leq C\left(\frac{s}{r}\right)^\alpha, E s /2 1/2 ( P Z , y 1 − P Z , y 2 ) ≤ C ( r s ) α ,
with E s / 2 \mathcal{E}_{s/2} E s /2 centred at y 1 y_1 y 1 . Expanding P Z , y 2 P_{Z,y_2} P Z , y 2 around y 1 y_1 y 1 and using (72) , we obtain
∣ B Z ( y 2 ) − B Z ( y 1 ) ∣ + ∣ q Z ( y 2 ) − q Z ( y 1 ) ∣ + ∣ M Z ( y 2 ) − M Z ( y 1 ) ∣ ≤ C r − α ∣ y 2 − y 1 ∣ α / 2 , |B_Z(y_2)-B_Z(y_1)|+|q_Z(y_2)-q_Z(y_1)|+|M_Z(y_2)-M_Z(y_1)|\leq Cr^{-\alpha}|y_2-y_1|^{\alpha/2}, ∣ B Z ( y 2 ) − B Z ( y 1 ) ∣ + ∣ q Z ( y 2 ) − q Z ( y 1 ) ∣ + ∣ M Z ( y 2 ) − M Z ( y 1 ) ∣ ≤ C r − α ∣ y 2 − y 1 ∣ α /2 ,
∣ a Z ( y 2 ) − a Z ( y 1 ) − B Z ( y 1 ) ( y 2 − y 1 ) ∣ ≤ C r − α ∣ y 2 − y 1 ∣ 1 + α / 2 . (118) |a_Z(y_2)-a_Z(y_1)-B_Z(y_1)(y_2-y_1)|\leq Cr^{-\alpha}|y_2-y_1|^{1+\alpha/2}. \tag*{(118)} ∣ a Z ( y 2 ) − a Z ( y 1 ) − B Z ( y 1 ) ( y 2 − y 1 ) ∣ ≤ C r − α ∣ y 2 − y 1 ∣ 1 + α /2 . (118)
Hence D y a Z = B Z D_ya_Z=B_Z D y a Z = B Z . For more distant pairs of points, the coefficient supremum bounds in Step 1 imply the required Hölder bounds. Consequently,
r − 2 ∥ a Z ∥ C 0 + ∥ D y a Z ∥ C 0 + ∥ q Z ∥ C 0 + ∥ M Z ∥ C 0 + r α ( [ D y a Z ] C 0 , α / 2 + [ q Z ] C 0 , α / 2 + [ M Z ] C 0 , α / 2 ) ≤ C . r^{-2}\|a_Z\|_{C^0}+\|D_ya_Z\|_{C^0}+\|q_Z\|_{C^0}+\|M_Z\|_{C^0}+r^\alpha\left([D_ya_Z]_{C^{0,\alpha/2}}+[q_Z]_{C^{0,\alpha/2}}+[M_Z]_{C^{0,\alpha/2}}\right)\leq C. r − 2 ∥ a Z ∥ C 0 + ∥ D y a Z ∥ C 0 + ∥ q Z ∥ C 0 + ∥ M Z ∥ C 0 + r α ( [ D y a Z ] C 0 , α /2 + [ q Z ] C 0 , α /2 + [ M Z ] C 0 , α /2 ) ≤ C .
Steps 1–3 prove (116) .
Step 4: Compactly supported fields. Suppose that Z ∈ W G , 0 1 , 2 ( Q 2 r ( y 0 ) ) Z\in W^{1,2}_{G,0}(Q_{2r}(y_0)) Z ∈ W G , 0 1 , 2 ( Q 2 r ( y 0 )) is supported in K ⋐ Q 2 r ( y 0 ) K\Subset Q_{2r}(y_0) K ⋐ Q 2 r ( y 0 ) . Testing the equation with Z Z Z , and using coercivity and the weighted Poincaré inequalities, we find
r − N G / 2 − 2 ∥ Z ∥ L G 2 ( Q 2 r ( y 0 ) ) + r − N G / 2 ∥ ∇ G Z ∥ L 2 ( Q 2 r ( y 0 ) ) ≤ C r − N G / 2 ∥ F ∥ ( W G 1 , 2 ( Q 2 r ( y 0 ) ) ) ∗ ≤ C ∥ F ∥ Y α ( Q 2 r ( y 0 ) ) . r^{-N_G/2-2}\|Z\|_{L^2_G(Q_{2r}(y_0))}+r^{-N_G/2}\|\nabla_GZ\|_{L^2(Q_{2r}(y_0))}
\leq Cr^{-N_G/2}\|F\|_{(W^{1,2}_G(Q_{2r}(y_0)))^*}
\leq C\|F\|_{\mathcal{Y}_\alpha(Q_{2r}(y_0))}. r − N G /2 − 2 ∥ Z ∥ L G 2 ( Q 2 r ( y 0 )) + r − N G /2 ∥ ∇ G Z ∥ L 2 ( Q 2 r ( y 0 )) ≤ C r − N G /2 ∥ F ∥ ( W G 1 , 2 ( Q 2 r ( y 0 )) ) ∗ ≤ C ∥ F ∥ Y α ( Q 2 r ( y 0 )) .
Choose a fixed interior neighbourhood of K K K and cover it by finitely many spine-centred cylinders and annuli Q ξ , y Q_{\xi,y} Q ξ , y . Apply (116) on the spine-centred cylinders and interior Schauder estimates on the annuli. Outside this neighbourhood, the comparison polynomials and the excesses at sufficiently small scales vanish. Patching these estimates together yields (117) . □
The differentiated equations
For a multi-index β = ( β 1 , … , β n − 2 ) \beta=(\beta_1,\ldots,\beta_{n-2}) β = ( β 1 , … , β n − 2 ) , write ∣ β ∣ = ∑ i β i |\beta|=\sum_i\beta_i ∣ β ∣ = ∑ i β i , β ! = ∏ i β i ! \beta!=\prod_i\beta_i! β ! = ∏ i β i ! and Z β = D y β U Z_\beta=D_y^\beta U Z β = D y β U , and let e j e_j e j be the j j j -th unit multi-index. Let D q A D^q\mathbf{A} D q A be the q q q -th derivative of A \mathbf{A} A with respect to ( y , U , D U ) (y,U,DU) ( y , U , D U ) , with ξ \xi ξ fixed, and set
E γ = D y γ ( y , U , D U ) = { ( e j , Z e j , D Z e j ) , γ = e j , ( 0 , Z γ , D Z γ ) , ∣ γ ∣ ≥ 2. (119) \mathbf{E}_\gamma=D_y^\gamma(y,U,DU)=
\begin{cases}
(e_j,Z_{e_j},DZ_{e_j}), & \gamma=e_j,\\
(0,Z_\gamma,DZ_\gamma), & |\gamma|\geq2.
\end{cases} \tag*{(119)} E γ = D y γ ( y , U , D U ) = { ( e j , Z e j , D Z e j ) , ( 0 , Z γ , D Z γ ) , γ = e j , ∣ γ ∣ ≥ 2. (119)
Formally differentiating the transformed weak equation (78) k = ∣ β ∣ k=|\beta| k = ∣ β ∣ times by the chain rule, we find that Z β Z_\beta Z β satisfies B U ( Z β , Y ) = F β ( Y ) \mathcal{B}_U(Z_\beta,Y)=\mathcal{F}_\beta(Y) B U ( Z β , Y ) = F β ( Y ) , with F e j = − ∂ y j A \mathbf{F}_{e_j}=-\partial_{y_j}\mathbf{A} F e j = − ∂ y j A and
F β = − ∑ q = 2 k 1 q ! ∑ γ 1 + ⋯ + γ q = β γ i ≠ 0 β ! γ 1 ! ⋯ γ q ! D q A ( Ξ γ 1 , … , Ξ γ q ) , k ≥ 2. (120) \mathbf{F}_\beta=-\sum_{q=2}^{k}\frac{1}{q!}\sum_{\substack{\gamma_1+\cdots+\gamma_q=\beta\\ \gamma_i\ne0}}\frac{\beta!}{\gamma_1!\cdots\gamma_q!}D^q\mathbf{A}(\mathbf{\Xi}_{\gamma_1},\ldots,\mathbf{\Xi}_{\gamma_q}),\qquad k\ge2. \tag*{(120)} F β = − q = 2 ∑ k q ! 1 γ 1 + ⋯ + γ q = β γ i = 0 ∑ γ 1 ! ⋯ γ q ! β ! D q A ( Ξ γ 1 , … , Ξ γ q ) , k ≥ 2. (120)
The source involves only Z γ Z_\gamma Z γ with ∣ γ ∣ ≤ k − 1 |\gamma|\le k-1 ∣ γ ∣ ≤ k − 1 . At a centre ( 0 , y ) (0,y) ( 0 , y ) , the principal part of the source is
F β , p r , y = − ∑ q = 2 k 1 q ! ∑ γ 1 + ⋯ + γ q = β γ i ≠ 0 β ! γ 1 ! ⋯ γ q ! D q A p r ( p ( y ) , D U p ( y ) ) ( Ξ γ 1 p r , … , Ξ γ q p r ) , k ≥ 2 , (121) \mathbf{F}_{\beta,\mathrm{pr},y}=-\sum_{q=2}^{k}\frac{1}{q!}\sum_{\substack{\gamma_1+\cdots+\gamma_q=\beta\\ \gamma_i\ne0}}\frac{\beta!}{\gamma_1!\cdots\gamma_q!}D^q\mathbf{A}_{\mathrm{pr}}(p(y),D U_{p(y)})(\mathbf{\Xi}^{\mathrm{pr}}_{\gamma_1},\ldots,\mathbf{\Xi}^{\mathrm{pr}}_{\gamma_q}),\qquad k\ge2, \tag*{(121)} F β , pr , y = − q = 2 ∑ k q ! 1 γ 1 + ⋯ + γ q = β γ i = 0 ∑ γ 1 ! ⋯ γ q ! β ! D q A pr ( p ( y ) , D U p ( y ) ) ( Ξ γ 1 pr , … , Ξ γ q pr ) , k ≥ 2 , (121)
where Ξ γ p r = ( D y γ y , a Z γ ( y ) , D P Z γ , y ) \mathbf{\Xi}^{\mathrm{pr}}_\gamma=(D^\gamma_y y,a_{Z_\gamma}(y),D P_{Z_\gamma,y}) Ξ γ pr = ( D y γ y , a Z γ ( y ) , D P Z γ , y ) , in the order of ( y , U , D U ) (y,U,DU) ( y , U , D U ) as in (119) (the U U U -component is a Z γ ( y ) a_{Z_\gamma}(y) a Z γ ( y ) , since A p r ( X , D U ) \mathbf{A}_{\mathrm{pr}}(X,DU) A pr ( X , D U ) depends on U U U only through X = ( φ , y ) X=(\varphi,y) X = ( φ , y ) ), and F e j , p r , y = − ∂ y j A p r ( p ( y ) , D U p ( y ) ) \mathbf{F}_{e_j,\mathrm{pr},y}=-\partial_{y_j}\mathbf{A}_{\mathrm{pr}}(p(y),D U_{p(y)}) F e j , pr , y = − ∂ y j A pr ( p ( y ) , D U p ( y ) ) for k = 1 k=1 k = 1 . On Q ξ , y Q_{\xi,y} Q ξ , y , in the normalised quantities of the weighted Hölder norms (70) , the independence of A p r \mathbf{A}_{\mathrm{pr}} A pr from the value of V V V gives
F β − F β , p r , y = O ( ρ α ) + O ( ∣ X − p ( y ) ∣ + ∣ ∇ G ( U − U p ( y ) ) ∣ ) + O ( ∑ 1 ≤ ∣ γ ∣ < k ( ∣ ( Z γ ) φ − a Z γ ( y ) ∣ + ∣ ∇ G ( Z γ − P Z γ , y ) ∣ ) ) = O ( ρ α ) , \mathbf{F}_\beta-\mathbf{F}_{\beta,\mathrm{pr},y}=O(\rho^\alpha)+O\left(|X-p(y)|+|\nabla^G(U-U_{p(y)})|\right)
+O\left(\sum_{1\le|\gamma|<k}\left(|(Z_\gamma)_\varphi-a_{Z_\gamma}(y)|+|\nabla^G(Z_\gamma-P_{Z_\gamma,y})|\right)\right)=O(\rho^\alpha), F β − F β , pr , y = O ( ρ α ) + O ( ∣ X − p ( y ) ∣ + ∣ ∇ G ( U − U p ( y ) ) ∣ ) + O 1 ≤ ∣ γ ∣ < k ∑ ( ∣ ( Z γ ) φ − a Z γ ( y ) ∣ + ∣ ∇ G ( Z γ − P Z γ , y ) ∣ ) = O ( ρ α ) ,
by the density decomposition (76) , the initial estimates of Proposition 5.4(ii), and the lower-order X α X_\alpha X α -estimates.
The following proposition collects the properties of the sources used in the induction.
Proposition 8.2 (Source estimates). Let Ω = Q r ( y 0 ) \Omega=Q_r(y_0) Ω = Q r ( y 0 ) , k ≥ 1 k\ge1 k ≥ 1 , and suppose that ∥ Z γ ∥ X α ( Ω ) < ∞ \|Z_\gamma\|_{X_\alpha(\Omega)}<\infty ∥ Z γ ∥ X α ( Ω ) < ∞ for ∣ γ ∣ ≤ k − 1 |\gamma|\le k-1 ∣ γ ∣ ≤ k − 1 , with
a Z γ = D y γ g , q Z γ = D y γ q , M Z γ = D y γ M . (122) a_{Z_\gamma}=D^\gamma_y g,\qquad q_{Z_\gamma}=D^\gamma_y q,\qquad M_{Z_\gamma}=D^\gamma_y M. \tag*{(122)} a Z γ = D y γ g , q Z γ = D y γ q , M Z γ = D y γ M . (122) (i) (Source estimate) For k = 1 k=1 k = 1 , the source F e j = − ∂ y j A \mathbf{F}_{e_j}=-\partial_{y_j}\mathbf{A} F e j = − ∂ y j A satisfies ∥ F e j ∥ Y α ( Ω ) ≤ C 1 \|\mathbf{F}_{e_j}\|_{Y_\alpha(\Omega)}\le C_1 ∥ F e j ∥ Y α ( Ω ) ≤ C 1 by the background estimates . For k ≥ 2 k\ge2 k ≥ 2 , q ≥ 1 q\ge1 q ≥ 1 and nonzero γ 1 , … , γ q \gamma_1,\ldots,\gamma_q γ 1 , … , γ q with ∣ γ i ∣ ≤ k − 1 |\gamma_i|\le k-1 ∣ γ i ∣ ≤ k − 1 , the term D q A ( Ξ γ 1 , … , Ξ γ q ) D^q\mathbf{A}(\mathbf{\Xi}_{\gamma_1},\ldots,\mathbf{\Xi}_{\gamma_q}) D q A ( Ξ γ 1 , … , Ξ γ q ) with the principal parts defined above satisfies
∥ D q A ( Ξ γ 1 , … , Ξ γ q ) ∥ Y α ( Ω ) ≤ C q ∏ i = 1 q ( ∥ Z γ i ∥ X α ( Ω ) + 1 { ∣ γ i ∣ = 1 } ) , (123) \left\|D^q\mathbf{A}(\mathbf{\Xi}_{\gamma_1},\ldots,\mathbf{\Xi}_{\gamma_q})\right\|_{Y_\alpha(\Omega)}
\le C_q\prod_{i=1}^{q}\left(\|Z_{\gamma_i}\|_{X_\alpha(\Omega)}+\mathbf{1}_{\{|\gamma_i|=1\}}\right), \tag*{(123)} D q A ( Ξ γ 1 , … , Ξ γ q ) Y α ( Ω ) ≤ C q i = 1 ∏ q ( ∥ Z γ i ∥ X α ( Ω ) + 1 { ∣ γ i ∣ = 1 } ) , (123) where C q ≤ Λ 0 Λ 1 q q ! C_q\le\Lambda_0\Lambda_1^q q! C q ≤ Λ 0 Λ 1 q q ! for real-analytic ambient data, with Λ 0 , Λ 1 \Lambda_0,\Lambda_1 Λ 0 , Λ 1 independent of q q q and the multi-indices. In particular, ∥ F β ∥ Y α ( Ω ) < ∞ \|\mathbf{F}_\beta\|_{Y_\alpha(\Omega)}<\infty ∥ F β ∥ Y α ( Ω ) < ∞ for ∣ β ∣ = k |\beta|=k ∣ β ∣ = k .
(ii) (Validity of polynomial corrections) For ∣ β ∣ = k |\beta|=k ∣ β ∣ = k and every centre ( 0 , y ) (0,y) ( 0 , y ) , F β , p r , y ∈ A p ( y ) ( P ) \mathbf{F}_{\beta,\mathrm{pr},y}\in\mathcal{A}_{p(y)}(\mathcal{P}) F β , pr , y ∈ A p ( y ) ( P ) .
We remark that the indicator function in (123) accounts for the entry e j e_j e j of Ξ e j \mathbf{\Xi}_{e_j} Ξ e j .
Proof. Step 1: Pointwise structure and the constants. By (76) , D q A ( Ξ γ 1 , … , Ξ γ q ) D^q\mathbf{A}(\mathbf{\Xi}_{\gamma_1},\ldots,\mathbf{\Xi}_{\gamma_q}) D q A ( Ξ γ 1 , … , Ξ γ q ) is a multilinear expression in the normalised quantities ( Z γ i ) φ (Z_{\gamma_i})_\varphi ( Z γ i ) φ , ∇ G Z γ i \nabla^G Z_{\gamma_i} ∇ G Z γ i , ρ − 1 ( Z γ i ) V \rho^{-1}(Z_{\gamma_i})_V ρ − 1 ( Z γ i ) V (and e j e_j e j ), whose coefficients are derivatives of order at most q + 1 q+1 q + 1 of the energy density f ^ \hat f f ^ , regarded as a function of ρ \rho ρ and the normalised data ( X , u ^ / ρ , ∇ G U ) (X,\hat u/\rho,\nabla^G U) ( X , u ^ / ρ , ∇ G U ) (see Lemma 5.5). As in the proof of Lemma 5.5(ii), the resulting coefficient field satisfies the weighted bounds (77) with constant C q ∏ i ( ∥ Z γ i ∥ X α + 1 { ∣ γ i ∣ = 1 } ) C_q\prod_i(\|Z_{\gamma_i}\|_{X_\alpha}+\mathbf{1}_{\{|\gamma_i|=1\}}) C q ∏ i ( ∥ Z γ i ∥ X α + 1 { ∣ γ i ∣ = 1 } ) , where
C q = C q + 1 ( 1 + max m ≤ q + 3 sup ∥ D m f ^ ∥ ) (124) C_q=C^{q+1}\left(1+\max_{m\le q+3}\sup\|D^m\hat f\|\right) \tag*{(124)} C q = C q + 1 ( 1 + m ≤ q + 3 max sup ∥ D m f ^ ∥ ) (124)
with C C C independent of q q q , where D m f ^ D^m\hat f D m f ^ denotes the m m m -th derivative in ( X , u ^ / ρ , ∇ G U ) (X,\hat u/\rho,\nabla^G U) ( X , u ^ / ρ , ∇ G U ) , and the supremum is over Ω ∖ { ξ = 0 } \Omega\setminus\{\xi=0\} Ω ∖ { ξ = 0 } along U U U . For real-analytic data, f ^ \hat f f ^ is real-analytic, so sup ∥ D m f ^ ∥ ≤ Λ Λ 1 m m ! \sup\|D^m\hat f\|\le\Lambda\Lambda_1^m m! sup ∥ D m f ^ ∥ ≤ Λ Λ 1 m m ! for some constants Λ , Λ 1 > 0 \Lambda,\Lambda_1>0 Λ , Λ 1 > 0 . Since ( q + 3 ) ! ≤ 4 q + 3 q ! (q+3)!\le4^{q+3}q! ( q + 3 )! ≤ 4 q + 3 q ! , we obtain C q ≤ Λ 0 Λ 1 q q ! C_q\le\Lambda_0\Lambda_1^q q! C q ≤ Λ 0 Λ 1 q q ! for sufficiently large Λ 0 , Λ 1 ≫ 1 \Lambda_0,\Lambda_1\gg1 Λ 0 , Λ 1 ≫ 1 .
Step 2: Closeness to the principal part. Fix a centre ( 0 , y ) (0,y) ( 0 , y ) , and let D q A p r D^q A_{\mathrm{pr}} D q A pr denote the derivative of A p r A_{\mathrm{pr}} A pr evaluated at ( p ( y ) , D U p ( y ) ) (p(y),DU_{p(y)}) ( p ( y ) , D U p ( y ) ) . The difference of the two multilinear expressions telescopes:
D q A ( Ξ ‾ γ 1 , … , Ξ ‾ γ q ) − D q A p r ( Ξ ‾ γ 1 p r , … , Ξ ‾ γ q p r ) = ( D q A − D q A p r ) ( Ξ ‾ γ 1 , … , Ξ ‾ γ q ) + ∑ i = 1 q D q A p r ( Ξ ‾ γ 1 p r , … , Ξ ‾ γ i − 1 p r , Ξ ‾ γ i − Ξ ‾ γ i p r , Ξ ‾ γ i + 1 , … , Ξ ‾ γ q ) . (125) \begin{aligned}
D^q A(\overline{\Xi}_{\gamma_1},\ldots,\overline{\Xi}_{\gamma_q})-D^q A_{\mathrm{pr}}(\overline{\Xi}^{\mathrm{pr}}_{\gamma_1},\ldots,\overline{\Xi}^{\mathrm{pr}}_{\gamma_q})
={}&(D^q A-D^q A_{\mathrm{pr}})(\overline{\Xi}_{\gamma_1},\ldots,\overline{\Xi}_{\gamma_q})\\
&+\sum_{i=1}^{q}D^q A_{\mathrm{pr}}(\overline{\Xi}^{\mathrm{pr}}_{\gamma_1},\ldots,\overline{\Xi}^{\mathrm{pr}}_{\gamma_{i-1}},\overline{\Xi}_{\gamma_i}-\overline{\Xi}^{\mathrm{pr}}_{\gamma_i},\overline{\Xi}_{\gamma_{i+1}},\ldots,\overline{\Xi}_{\gamma_q}). \tag*{(125)}
\end{aligned} D q A ( Ξ γ 1 , … , Ξ γ q ) − D q A pr ( Ξ γ 1 pr , … , Ξ γ q pr ) = ( D q A − D q A pr ) ( Ξ γ 1 , … , Ξ γ q ) + i = 1 ∑ q D q A pr ( Ξ γ 1 pr , … , Ξ γ i − 1 pr , Ξ γ i − Ξ γ i pr , Ξ γ i + 1 , … , Ξ γ q ) . (125)
On Q ξ , y Q_{\xi,y} Q ξ , y , with the normalised quantities (70) and the C 0 , α C^{0,\alpha} C 0 , α -norms with respect to g ρ g_\rho g ρ , we have
∥ Ξ ‾ γ i ∥ + ∥ Ξ ‾ γ i p r ∥ ≤ C ( ∥ Z γ i ∥ X α + 1 { ∣ γ i ∣ = 1 } ) \|\overline{\Xi}_{\gamma_i}\|+\|\overline{\Xi}^{\mathrm{pr}}_{\gamma_i}\|\leq C(\|Z_{\gamma_i}\|_{X_\alpha}+\mathbf{1}_{\{|\gamma_i|=1\}}) ∥ Ξ γ i ∥ + ∥ Ξ γ i pr ∥ ≤ C ( ∥ Z γ i ∥ X α + 1 { ∣ γ i ∣ = 1 } )
and
∥ ( Z γ i ) φ − a Z γ i ( y ) ∥ C 0 , α + ∥ ∇ G ( Z γ i − P Z γ i , y ) ∥ C 0 , α ≤ C ( ρ r ) α ∥ Z γ i ∥ X α , \|(Z_{\gamma_i})_\varphi-a_{Z_{\gamma_i}}(y)\|_{C^{0,\alpha}}+\|\nabla^G(Z_{\gamma_i}-P_{Z_{\gamma_i},y})\|_{C^{0,\alpha}}\leq C\left(\frac{\rho}{r}\right)^\alpha\|Z_{\gamma_i}\|_{X_\alpha}, ∥ ( Z γ i ) φ − a Z γ i ( y ) ∥ C 0 , α + ∥ ∇ G ( Z γ i − P Z γ i , y ) ∥ C 0 , α ≤ C ( r ρ ) α ∥ Z γ i ∥ X α ,
by the last term in (114) and ∣ ( P Z γ i , y ) φ − a Z γ i ( y ) ∣ ≤ C ρ 2 ∥ ∇ P Z γ i , y ∥ \lvert(P_{Z_{\gamma_i},y})_\varphi-a_{Z_{\gamma_i}}(y)\rvert\leq C\rho^2\|\nabla P_{Z_{\gamma_i},y}\| ∣( P Z γ i , y ) φ − a Z γ i ( y )∣ ≤ C ρ 2 ∥∇ P Z γ i , y ∥ . Since A p r A_{\mathrm{pr}} A pr is independent of the value of V V V , the V V V -value component of Ξ ‾ γ i − Ξ ‾ γ i p r \overline{\Xi}_{\gamma_i}-\overline{\Xi}^{\mathrm{pr}}_{\gamma_i} Ξ γ i − Ξ γ i pr contributes zero in each summand of (125) . Moreover, in the corresponding normalised coefficient blocks,
∥ D q A ∥ + ∥ D q A p r ∥ + ρ − α ∥ D q A − D q A p r ∥ ≤ C q , \|D^q A\|+\|D^q A_{\mathrm{pr}}\|+\rho^{-\alpha}\|D^q A-D^q A_{\mathrm{pr}}\|\leq C_q, ∥ D q A ∥ + ∥ D q A pr ∥ + ρ − α ∥ D q A − D q A pr ∥ ≤ C q ,
by Proposition 5.4(ii) and (76) . Substituting these bounds into (125) implies
∥ F − F p r , y ∥ C G 0 , α ( Q ξ , y ) ≤ C q ( ρ r ) α ∏ i = 1 q ( ∥ Z γ i ∥ X α + 1 { ∣ γ i ∣ = 1 } ) , \|\mathbf{F}-\mathbf{F}_{\mathrm{pr},y}\|_{C_G^{0,\alpha}(Q_{\xi,y})}\leq C_q\left(\frac{\rho}{r}\right)^\alpha\prod_{i=1}^{q}(\|Z_{\gamma_i}\|_{X_\alpha}+\mathbf{1}_{\{|\gamma_i|=1\}}), ∥ F − F pr , y ∥ C G 0 , α ( Q ξ , y ) ≤ C q ( r ρ ) α i = 1 ∏ q ( ∥ Z γ i ∥ X α + 1 { ∣ γ i ∣ = 1 } ) ,
after enlarging C C C in the definition (124) of C q C_q C q to absorb the factor q q q . This bounds the last term in the Y α \mathcal{Y}_\alpha Y α -norm (115) . On Q s ( y ) Q_s(y) Q s ( y ) , we decompose
F − F p r , y = ( F − F p r , y ′ ) + ( F p r , y ′ − F p r , y ) \mathbf{F}-\mathbf{F}_{\mathrm{pr},y}=(\mathbf{F}-\mathbf{F}_{\mathrm{pr},y'})+(\mathbf{F}_{\mathrm{pr},y'}-\mathbf{F}_{\mathrm{pr},y}) F − F pr , y = ( F − F pr , y ′ ) + ( F pr , y ′ − F pr , y )
at each point ( ξ ′ , y ′ ) (\xi',y') ( ξ ′ , y ′ ) . The first part is O ( ( s / r ) α ) O((s/r)^\alpha) O (( s / r ) α ) by the previous bound, and the second is O ( ρ − α ∣ y ′ − y ∣ α / 2 ) = O ( ( s / r ) α ) O(\rho^{-\alpha}|y'-y|^{\alpha/2})=O((s/r)^\alpha) O ( ρ − α ∣ y ′ − y ∣ α /2 ) = O (( s / r ) α ) by the Hölder continuity of the coefficients of the comparison polynomials. Then Cauchy–Schwarz, in the form of the dual estimate (69) of Lemma 5.3(ii), bounds the deviation F − F p r , y \mathbf{F}-\mathbf{F}_{\mathrm{pr},y} F − F pr , y , completing (i).
Step 3: Validity of polynomial corrections F β , p r , y ∈ A p ( y ) ( P ) \mathfrak{F}_{\beta,\mathrm{pr},y}\in\mathcal{A}_{p(y)}(\mathcal{P}) F β , pr , y ∈ A p ( y ) ( P ) . This is essentially the Leibniz rule again, now in the form of the (generalised) chain rule. Fix ∣ β ∣ = k |\beta|=k ∣ β ∣ = k and a centre ( 0 , y ) (0,y) ( 0 , y ) , and write p = p ( y ) p=p(y) p = p ( y ) . Let
p t = ( g t , y + t ) , g t = g ( y ) + ∑ 1 ≤ ∣ γ ∣ ≤ k − 1 t γ γ ! D γ g ( y ) , t ∈ R n − 2 , ∣ t ∣ ≪ 1 , p_t=(g_t,y+t),\qquad g_t=g(y)+\sum_{1\leq|\gamma|\leq k-1}\frac{t^\gamma}{\gamma!}D^\gamma g(y),\qquad t\in\mathbb{R}^{n-2},\quad|t|\ll1, p t = ( g t , y + t ) , g t = g ( y ) + 1 ≤ ∣ γ ∣ ≤ k − 1 ∑ γ ! t γ D γ g ( y ) , t ∈ R n − 2 , ∣ t ∣ ≪ 1 ,
which agrees with p ( y + t ) p(y+t) p ( y + t ) to order k − 1 k-1 k − 1 at t = 0 t=0 t = 0 , and the k k k -th order derivatives of g t g_t g t vanish. Using the exponential gauge and parallel transport as in the proof of Lemma 7.1(iii), we find a smooth family of flat Dirac models Φ t \Phi_t Φ t centred at p t p_t p t , each minimising the energy frozen at p t p_t p t , whose transformed models U t U_t U t agree with U p ( y + t ) U_{p(y+t)} U p ( y + t ) to order k − 1 k-1 k − 1 at t = 0 t=0 t = 0 . They satisfy
∫ A p r 1 ( p t , D U t ) : D Y = 0 , (126) \int A_{\mathrm{pr}}^1(p_t,DU_t):DY=0, \tag*{(126)} ∫ A pr 1 ( p t , D U t ) : D Y = 0 , (126)
compactly supported admissible Y Y Y and, by (122) ,
∂ t γ ( y + t , g t , D U t ) ∣ t = 0 = ( D γ y , a Z γ ( y ) , D P Z γ , y ) = Ξ ‾ γ p r , 1 ≤ ∣ γ ∣ ≤ k − 1. \left.\partial_t^\gamma(y+t,g_t,DU_t)\right|_{t=0}=(D^\gamma y,a_{Z_\gamma}(y),DP_{Z_\gamma,y})=\overline{\Xi}^{\mathrm{pr}}_\gamma,\qquad1\leq|\gamma|\leq k-1. ∂ t γ ( y + t , g t , D U t ) ∣ t = 0 = ( D γ y , a Z γ ( y ) , D P Z γ , y ) = Ξ γ pr , 1 ≤ ∣ γ ∣ ≤ k − 1.
At order k k k , since differentiation in t t t preserves parity and the polynomial form of the models, we have
a : = ∂ t β g t ∣ t = 0 = 0 , P : = ∂ t β U t ∣ t = 0 ∈ P . a:=\left.\partial_t^\beta g_t\right|_{t=0}=0,\qquad P:=\left.\partial_t^\beta U_t\right|_{t=0}\in\mathcal{P}. a := ∂ t β g t t = 0 = 0 , P := ∂ t β U t t = 0 ∈ P .
In the source expansion (120) , the term with q = 1 q=1 q = 1 is the first derivative of A p r 1 A_{\mathrm{pr}}^{1} A pr 1 applied to ∂ t β ( y + t , g t , D U t ) ∣ t = 0 = ( D y β , a , D P ) \partial_t^\beta(y+t,g_t,DU_t)|_{t=0}=(D_y^\beta,a,DP) ∂ t β ( y + t , g t , D U t ) ∣ t = 0 = ( D y β , a , D P ) , and the terms with q ≥ 2 q\ge2 q ≥ 2 only involve the jets Ξ γ i p r \Xi_{\gamma_i}^{\mathrm{pr}} Ξ γ i pr with ∣ γ i ∣ ≤ k − 1 |\gamma_i|\le k-1 ∣ γ i ∣ ≤ k − 1 . Hence, by (121) ,
∂ t β A p r 1 ( p t , D U t ) ∣ t = 0 = ( N p a ) 1 + D D U A p r 1 ( p , D U p ) ( D P ) + { ∂ y j A p r 1 ( p , D U p ) , β = e j , ∑ q = 2 k 1 q ! ∑ γ 1 + ⋯ + γ q = β γ i ≠ 0 β ! γ 1 ! ⋯ γ q ! D q A p r 1 ( p , D U p ) ( Ξ γ 1 p r , … , Ξ γ q p r ) , k ≥ 2 , = ( N p a ) 1 + ( A p r , p P ) 1 − F β , p r , y 1 = ( A p r , p P ) 1 − F β , p r , y 1 . \begin{aligned}
\partial_t^\beta A_{\mathrm{pr}}^{1}(p_t,DU_t)\big|_{t=0}=(N_pa)^1+D_{DU}A_{\mathrm{pr}}^{1}(p,DU_p)(DP)
+\begin{cases}
\partial_{y_j}A_{\mathrm{pr}}^{1}(p,DU_p), & \beta=e_j,\\
\sum_{q=2}^{k}\frac{1}{q!}\sum_{\substack{\gamma_1+\cdots+\gamma_q=\beta\\\gamma_i\ne0}}\frac{\beta!}{\gamma_1!\cdots\gamma_q!}D^qA_{\mathrm{pr}}^{1}(p,DU_p)(\Xi_{\gamma_1}^{\mathrm{pr}},\ldots,\Xi_{\gamma_q}^{\mathrm{pr}}), & k\ge2,
\end{cases} \\
&=(N_pa)^1+(A_{\mathrm{pr},p}P)^1-F_{\beta,\mathrm{pr},y}^{1}=(A_{\mathrm{pr},p}P)^1-F_{\beta,\mathrm{pr},y}^{1}.
\end{aligned} ∂ t β A pr 1 ( p t , D U t ) t = 0 = ( N p a ) 1 + D D U A pr 1 ( p , D U p ) ( D P ) + ⎩ ⎨ ⎧ ∂ y j A pr 1 ( p , D U p ) , ∑ q = 2 k q ! 1 ∑ γ 1 + ⋯ + γ q = β γ i = 0 γ 1 ! ⋯ γ q ! β ! D q A pr 1 ( p , D U p ) ( Ξ γ 1 pr , … , Ξ γ q pr ) , β = e j , k ≥ 2 , = ( N p a ) 1 + ( A pr , p P ) 1 − F β , pr , y 1 = ( A pr , p P ) 1 − F β , pr , y 1 .
For ∣ t ∣ ≪ 1 |t|\ll1 ∣ t ∣ ≪ 1 , the t t t -derivatives of A p r 1 ( p t , D U t ) A_{\mathrm{pr}}^{1}(p_t,DU_t) A pr 1 ( p t , D U t ) are bounded on supp Y \operatorname{supp}Y supp Y uniformly in t t t , so by the Lebesgue’s dominated convergence theorem we differentiate (126) β \beta β times and obtain (ii):
F β , p r , y = A p r P ∈ A p r ( P ) . \mathcal{F}_{\beta,\mathrm{pr},y}=\mathcal{A}_{\mathrm{pr}}P\in\mathcal{A}_{\mathrm{pr}}(P). F β , pr , y = A pr P ∈ A pr ( P ) .
Tangential regularity
We now combine Propositions 8.1 and 8.2 in an induction on ∣ β ∣ |\beta| ∣ β ∣ . For real-analytic data, we make the estimates quantitative in ∣ β ∣ |\beta| ∣ β ∣ .
Theorem 8.3 (Tangential regularity). *Suppose that the ambient data are smooth (resp. real-analytic), and let 0 < r ≤ r 0 / 2 0<r\le r_0/2 0 < r ≤ r 0 /2 . For every multi-index β ≠ 0 \beta\ne0 β = 0 , the field Z β = D y β U Z_\beta=D_y^\beta U Z β = D y β U belongs to W G 1 , 2 ( Q r ( y 0 ) ) W_G^{1,2}(Q_r(y_0)) W G 1 , 2 ( Q r ( y 0 )) and solves B U ( Z β , Y ) = F β ( Y ) \mathcal{B}_U(Z_\beta,Y)=\mathcal{F}_\beta(Y) B U ( Z β , Y ) = F β ( Y ) for every compactly supported admissible Y Y Y , with the source (120) . Moreover, the comparison polynomials of Z β Z_\beta Z β satisfy (122) , and
∥ D y β U ∥ X α ( Q r ( y 0 ) ) ≤ C β ( resp. ≤ C ∣ β ∣ + 1 ∣ β ∣ ! ) , (127) \|D_y^\beta U\|_{X_\alpha(Q_r(y_0))}\le C_\beta\qquad(\text{resp. }\le C^{|\beta|+1}|\beta|!), \tag*{(127)} ∥ D y β U ∥ X α ( Q r ( y 0 )) ≤ C β ( resp. ≤ C ∣ β ∣ + 1 ∣ β ∣ !) , (127) where C C C is independent of β \beta β . In particular,
∣ D ξ D y β ϕ ( ξ , y ) ∣ ≤ C β ∣ ξ ∣ , ( ξ , y ) ∈ Q r ( y 0 ) ∖ { ξ = 0 } , (128) |D_\xi D_y^\beta\phi(\xi,y)|\le C_\beta|\xi|,\qquad(\xi,y)\in Q_r(y_0)\setminus\{\xi=0\}, \tag*{(128)} ∣ D ξ D y β ϕ ( ξ , y ) ∣ ≤ C β ∣ ξ ∣ , ( ξ , y ) ∈ Q r ( y 0 ) ∖ { ξ = 0 } , (128) the branching locus is smooth (resp. real-analytic), and U U U and the coefficients g , q , M g,q,M g , q , M of the transformed models U p ( y ) U_p(y) U p ( y ) depend smoothly (resp. real-analytically) on $y.
Proof for smooth data. We argue by induction on k = ∣ β ∣ k=|\beta| k = ∣ β ∣ . For β = 0 \beta=0 β = 0 , these are the transformed estimates from Part I (Proposition 5.4). Each step below shrinks the domain by a fixed factor, from Q 4 r ( y 0 ) Q_{4r}(y_0) Q 4 r ( y 0 ) to Q r ( y 0 ) Q_r(y_0) Q r ( y 0 ) . This is harmless since the assertions are local: we prove them at each order on every compact subset of Q r 0 ( y 0 ) Q_{r_0}(y_0) Q r 0 ( y 0 ) , by covering it with finitely many such cylinders. So we assume that the assertions hold for all orders below k k k on Q 4 r ( y 0 ) Q_{4r}(y_0) Q 4 r ( y 0 ) .
Now, let ∣ β ∣ = k |\beta|=k ∣ β ∣ = k , choose j j j with β j > 0 \beta_j>0 β j > 0 , and set γ = ( β 1 , … , β j − 1 , … , β n − 2 ) \gamma=(\beta_1,\ldots,\beta_j-1,\ldots,\beta_{n-2}) γ = ( β 1 , … , β j − 1 , … , β n − 2 ) . For 0 < ∣ h ∣ ≤ r 2 / 100 0<|h|\le r^2/100 0 < ∣ h ∣ ≤ r 2 /100 , consider the tangential difference quotient
δ h , j Z γ ( ξ , y ) = Z γ ( ξ , y + h e j ) − Z γ ( ξ , y ) h . \delta_{h,j}Z_\gamma(\xi,y)=\frac{Z_\gamma(\xi,y+he_j)-Z_\gamma(\xi,y)}{h}. δ h , j Z γ ( ξ , y ) = h Z γ ( ξ , y + h e j ) − Z γ ( ξ , y ) .
Step 1: Z β ∈ W G 1 , 2 ( Q 2 r ( y 0 ) ) Z_\beta\in W_G^{1,2}(Q_{2r}(y_0)) Z β ∈ W G 1 , 2 ( Q 2 r ( y 0 )) solves the weak equation. For k = 1 k=1 k = 1 , subtracting the transformed weak equation (78) from its translate by h e j he_j h e j and applying the Newton–Leibniz formula along the segment from U U U to U h e j U_{he_j} U h e j , we have
B ‾ h e j ( δ h , j U , Y ) = − ∫ 0 1 ( ∂ y j A 1 ( ξ , y + t h e j , U h e j , t , D U h e j , t ) : D Y + ∂ y j A 0 ( ξ , y + t h e j , U h e j , t , D U h e j , t ) ⋅ Y ) d t , \overline{\mathcal{B}}_{he_j}(\delta_{h,j}U,Y)
=-\int_0^1\left(\partial_{y_j}A^1(\xi,y+the_j,U_{he_j,t},DU_{he_j,t}):DY+\partial_{y_j}A^0(\xi,y+the_j,U_{he_j,t},DU_{he_j,t})\cdot Y\right)\,dt, B h e j ( δ h , j U , Y ) = − ∫ 0 1 ( ∂ y j A 1 ( ξ , y + t h e j , U h e j , t , D U h e j , t ) : D Y + ∂ y j A 0 ( ξ , y + t h e j , U h e j , t , D U h e j , t ) ⋅ Y ) d t ,
where B ‾ h e j \overline{\mathcal{B}}_{he_j} B h e j is the averaged bilinear form considered in Remark 7.2.
For k ≥ 2 k\ge2 k ≥ 2 , subtracting the equation of Z γ Z_\gamma Z γ from its translate by h e j he_j h e j implies
B h e j ( δ h , j Z γ , Y ) = ( δ h , j F γ ) ( Y ) − ( B h e j − B U ) ( Z γ , Y ) h , \mathcal{B}_{he_j}(\delta_{h,j}Z_\gamma,Y)=(\delta_{h,j}\mathcal{F}_\gamma)(Y)-\frac{(\mathcal{B}_{he_j}-\mathcal{B}_U)(Z_\gamma,Y)}{h}, B h e j ( δ h , j Z γ , Y ) = ( δ h , j F γ ) ( Y ) − h ( B h e j − B U ) ( Z γ , Y ) ,
where B h e j \mathcal{B}_{he_j} B h e j is the spine-translated bilinear form considered in Remark 7.2. Since F γ \mathcal{F}_\gamma F γ and the coefficients of B U \mathcal{B}_U B U only involve Z γ ′ Z_{\gamma'} Z γ ′ with ∣ γ ′ ∣ ≤ k − 2 |\gamma'|\le k-2 ∣ γ ′ ∣ ≤ k − 2 , their difference quotients only involve Z γ ′ Z_{\gamma'} Z γ ′ with ∣ γ ′ ∣ ≤ k − 1 |\gamma'| \le k-1 ∣ γ ′ ∣ ≤ k − 1 . These are bounded pointwise in the normalised blocks, by the induction hypothesis and the last term in the X α \mathcal{X}_{\alpha} X α -norm ((114) ). Together with the weighted bounds ((77) ) on the coefficients and the dual estimate ((69) ) for weighted sources, the right-hand sides above are bounded in ( W G 1 , 2 ( Q 3 r ( y 0 ) ) ) ∗ (W^{1,2}_{G}(Q_{3r}(y_{0})))^{*} ( W G 1 , 2 ( Q 3 r ( y 0 )) ) ∗ uniformly in h h h . Since ∥ δ h , j Z γ ∥ L G 2 ( Q 3 r ( y 0 ) ) ≤ C ∥ ∇ G Z γ ∥ L 2 ( Q 4 r ( y 0 ) ) \|\delta_{h,j}Z_{\gamma}\|_{L^{2}_{G}(Q_{3r}(y_{0}))} \le C\|\nabla_{G}Z_{\gamma}\|_{L^{2}(Q_{4r}(y_{0}))} ∥ δ h , j Z γ ∥ L G 2 ( Q 3 r ( y 0 )) ≤ C ∥ ∇ G Z γ ∥ L 2 ( Q 4 r ( y 0 )) , the weighted Caccioppoli estimates (Corollary 7.3) imply that
∥ ∇ G δ h , j Z γ ∥ L 2 ( Q 2 r ( y 0 ) ) ≤ C k , 0 < ∣ h ∣ ≤ r 2 / 100. \|\nabla_{G}\delta_{h,j}Z_{\gamma}\|_{L^{2}(Q_{2r}(y_{0}))} \le C_{k}, \qquad0<|h|\le r^{2}/100. ∥ ∇ G δ h , j Z γ ∥ L 2 ( Q 2 r ( y 0 )) ≤ C k , 0 < ∣ h ∣ ≤ r 2 /100.
Letting h → 0 h\to0 h → 0 , we obtain δ h , j Z γ ⇀ Z β \delta_{h,j}Z_{\gamma}\rightharpoonup Z_{\beta} δ h , j Z γ ⇀ Z β in W G 1 , 2 ( Q 2 r ( y 0 ) ) W^{1,2}_{G}(Q_{2r}(y_{0})) W G 1 , 2 ( Q 2 r ( y 0 )) , with B U ( Z β , Y ) = F β ( Y ) \mathcal{B}_{U}(Z_{\beta},Y)=\mathcal{F}_{\beta}(Y) B U ( Z β , Y ) = F β ( Y ) on Q 2 r ( y 0 ) Q_{2r}(y_{0}) Q 2 r ( y 0 ) for all compactly supported admissible test fields.
Step 2: Z β Z_{\beta} Z β satisfies the hypotheses of Proposition 8.1. By the induction hypothesis, the fields Z γ ′ Z_{\gamma'} Z γ ′ , ∣ γ ′ ∣ ≤ k − 1 |\gamma'|\le k-1 ∣ γ ′ ∣ ≤ k − 1 , have finite X α ( Q 2 r ( y 0 ) ) \mathcal{X}_{\alpha}(Q_{2r}(y_{0})) X α ( Q 2 r ( y 0 )) -norms and satisfy ((122) ). These are the conditions of Proposition 8.2 on Q 2 r ( y 0 ) Q_{2r}(y_{0}) Q 2 r ( y 0 ) , so
∥ F β ∥ Y α ( Q 2 r ( y 0 ) ) < ∞ , F β , p r , y ∈ A p ( y ) ( P ) at every centre ( 0 , y ) . \|\mathcal{F}_{\beta}\|_{\mathcal{Y}_{\alpha}(Q_{2r}(y_{0}))}<\infty,\qquad\mathcal{F}_{\beta,\mathrm{pr},y}\in\mathcal{A}_{p(y)}(\mathcal{P})\quad\text{at every centre }(0,y). ∥ F β ∥ Y α ( Q 2 r ( y 0 )) < ∞ , F β , pr , y ∈ A p ( y ) ( P ) at every centre ( 0 , y ) .
Proposition 8.1 therefore implies that Z β Z_{\beta} Z β has the unique comparison polynomial P Z β , y ∈ P ( F β , p r , y ) P_{Z_{\beta},y}\in\mathcal{P}(\mathcal{F}_{\beta,\mathrm{pr},y}) P Z β , y ∈ P ( F β , pr , y ) at every centre ( 0 , y ) ∈ Q r ( y 0 ) (0,y)\in Q_{r}(y_{0}) ( 0 , y ) ∈ Q r ( y 0 ) , and
∥ Z β ∥ X α ( Q r ( y 0 ) ) ≤ C ( r − N G / 2 − 2 ∥ Z β ∥ L G 2 ( Q 2 r ( y 0 ) ) + ∥ F β ∥ Y α ( Q 2 r ( y 0 ) ) ) < ∞ . \|Z_{\beta}\|_{\mathcal{X}_{\alpha}(Q_{r}(y_{0}))}\le C\left(r^{-N_{G}/2-2}\|Z_{\beta}\|_{L^{2}_{G}(Q_{2r}(y_{0}))}+\|\mathcal{F}_{\beta}\|_{\mathcal{Y}_{\alpha}(Q_{2r}(y_{0}))}\right)<\infty. ∥ Z β ∥ X α ( Q r ( y 0 )) ≤ C ( r − N G /2 − 2 ∥ Z β ∥ L G 2 ( Q 2 r ( y 0 )) + ∥ F β ∥ Y α ( Q 2 r ( y 0 )) ) < ∞.
Since D ξ D y β φ = D ξ ( Z β ) φ D_{\xi}D_{y}^{\beta}\varphi=D_{\xi}(Z_{\beta})\varphi D ξ D y β φ = D ξ ( Z β ) φ , ((128) ) follows.
Step 3: The coefficients of P Z β , y P_{Z_{\beta},y} P Z β , y satisfy ((122) ). By the last term in the X α \mathcal{X}_{\alpha} X α -norm ((114) ), every field Z Z Z of finite X α \mathcal{X}_{\alpha} X α -norm satisfies, uniformly in y y y and ∣ ω ∣ = 1 |\omega|=1 ∣ ω ∣ = 1 ,
Z φ ( ρ ω , y ) = a Z ( y ) + ρ 2 q Z ( y , ω ) + O ( ρ 2 + α ) , Z V ( ρ ω , y ) = ρ M Z ( y ) ω + O ( ρ 1 + α ) . Z_{\varphi}(\rho\omega,y)=a_{Z}(y)+\rho^{2}q_{Z}(y,\omega)+O(\rho^{2+\alpha}),\qquad Z_{V}(\rho\omega,y)=\rho M_{Z}(y)\omega+O(\rho^{1+\alpha}). Z φ ( ρ ω , y ) = a Z ( y ) + ρ 2 q Z ( y , ω ) + O ( ρ 2 + α ) , Z V ( ρ ω , y ) = ρ M Z ( y ) ω + O ( ρ 1 + α ) .
Applying this to Z γ Z_{\gamma} Z γ and Z β = ∂ y j Z γ Z_{\beta}=\partial_{y_{j}}Z_{\gamma} Z β = ∂ y j Z γ , we find ( a , q , M ) Z β = ∂ y j ( a , q , M ) Z γ = D y β ( g , q , M ) (a,q,M)_{Z_{\beta}}=\partial_{y_{j}}(a,q,M)_{Z_{\gamma}}=D_{y}^{\beta}(g,q,M) ( a , q , M ) Z β = ∂ y j ( a , q , M ) Z γ = D y β ( g , q , M ) by ((122) ) for γ \gamma γ .
Steps 1–3 complete the induction. In particular, g g g , q q q and M M M are smooth, and so is the branching locus, the graph of g g g . □
For real-analytic data, the difficulty is that each application of the local estimate ((116) ) shrinks the domain. Instead, we consider χ k D y β U \chi^{k}D_{y}^{\beta}U χ k D y β U for a smooth cut-off function χ \chi χ following Kato [13 ] and use the diffeomorphism invariance of the transformed system. The advantage of this method is that only compactly supported fields are involved, so the estimate ((117) ) applies at every order. The constants are then tracked through Proposition 8.2(i), and a majorant series implies factorial bounds.
Proof for real-analytic data. Step 1: Kato’s trick and diffeomorphism invariance. Let Ω = Q 2 r ( y 0 ) \Omega=Q_{2r}(y_{0}) Ω = Q 2 r ( y 0 ) and K = Q 3 r / 2 ( y 0 ) K=Q_{3r/2}(y_{0}) K = Q 3 r /2 ( y 0 ) . Choose a smooth cutoff χ = χ 1 ( y ) χ 2 ( ξ ) \chi=\chi_{1}(y)\chi_{2}(\xi) χ = χ 1 ( y ) χ 2 ( ξ ) supported in K K K , even in ξ \xi ξ , equal to one near Q r ( y 0 ) Q_{r}(y_{0}) Q r ( y 0 ) , and with χ ( ξ , y ) = χ ( 0 , y ) \chi(\xi,y)=\chi(0,y) χ ( ξ , y ) = χ ( 0 , y ) for ∣ ξ ∣ ≤ r |\xi|\le r ∣ ξ ∣ ≤ r . For t ∈ R n − 2 t\in\mathbb{R}^{n-2} t ∈ R n − 2 with ∣ t ∣ ≪ 1 |t|\ll1 ∣ t ∣ ≪ 1 , the map
Ψ t ( ξ , y ) = ( ξ , y + χ ( ξ , y ) t ) , U t = U ∘ Ψ t , (129) \Psi_{t}(\xi,y)=(\xi,y+\chi(\xi,y)t),\qquad U_{t}=U\circ\Psi_{t}, \tag*{(129)} Ψ t ( ξ , y ) = ( ξ , y + χ ( ξ , y ) t ) , U t = U ∘ Ψ t , (129)
is a diffeomorphism of Ω \Omega Ω which equals the identity outside K K K and preserves parity and { ξ = 0 } \{\xi=0\} { ξ = 0 } . We have
D Ψ t = ( I 2 0 t D ξ χ I n − 2 + t D y χ ) , ( D Ψ t ) − 1 = ( I 2 0 − t D ξ χ 1 + D y χ t I n − 2 − t D y χ 1 + D y χ t ) , (130) D\Psi_{t}=
\begin{pmatrix}
I_{2} & 0\\
tD_{\xi}\chi& I_{n-2}+tD_{y}\chi
\end{pmatrix},
\qquad
(D\Psi_{t})^{-1}=
\begin{pmatrix}
I_{2} & 0\\
-\dfrac{tD_{\xi}\chi}{1+D_{y}\chi\,t} & I_{n-2}-\dfrac{tD_{y}\chi}{1+D_{y}\chi\,t}
\end{pmatrix}, \tag*{(130)} D Ψ t = ( I 2 t D ξ χ 0 I n − 2 + t D y χ ) , ( D Ψ t ) − 1 = I 2 − 1 + D y χ t t D ξ χ 0 I n − 2 − 1 + D y χ t t D y χ , (130)
and det D Ψ t = 1 + D y χ t \det D\Psi_{t}=1+D_{y}\chi\,t det D Ψ t = 1 + D y χ t in the Legendre coordinates ( ξ , y ) (\xi,y) ( ξ , y ) . Moreover, U t U_{t} U t solves the weak equation ∫ Ω A t 1 ( ξ , y , U t , D U t ) ⋅ D Y + A t 0 ( ξ , y , U t , D U t ) ⋅ Y = 0 \int_{\Omega}A_{t}^{1}(\xi,y,U_{t},DU_{t)\cdot DY+A_{t}^{0}(\xi,y,U_{t},DU_{t})\cdot Y=0} ∫ Ω A t 1 ( ξ , y , U t , D U t ) ⋅ D Y + A t 0 ( ξ , y , U t , D U t ) ⋅ Y = 0 for every compactly supported admissible Y Y Y , with coefficients
A t 1 ( ξ , y , U , D U ) = ( det D Ψ t ) A 1 ( ξ , y + χ t , U , D U ( D Ψ t ) − 1 ( D Ψ t ) − T ) , A_{t}^{1}(\xi,y,U,DU)=(\det D\Psi_{t})A^{1}(\xi,y+\chi t,U,DU(D\Psi_{t})^{-1}(D\Psi_{t})^{-T}), A t 1 ( ξ , y , U , D U ) = ( det D Ψ t ) A 1 ( ξ , y + χ t , U , D U ( D Ψ t ) − 1 ( D Ψ t ) − T ) ,
A t 0 ( ξ , y , U , D U ) = ( det D Ψ t ) A 0 ( ξ , y + χ t , U , D U ( D Ψ t ) − 1 ) . (131) A_{t}^{0}(\xi,y,U,DU)=(\det D\Psi_{t})A^{0}(\xi,y+\chi t,U,DU(D\Psi_{t})^{-1}). \tag*{(131)} A t 0 ( ξ , y , U , D U ) = ( det D Ψ t ) A 0 ( ξ , y + χ t , U , D U ( D Ψ t ) − 1 ) . (131)
For ∣ β ∣ = k ≥ 1 |\beta|=k\ge1 ∣ β ∣ = k ≥ 1 , consider the localised field
Z β : = ∂ t β U t ∣ t = 0 = χ k D y β U . Z_{\beta}:=\left.\partial_t^\beta U_t\right|_{t=0}=\chi^kD_y^\beta U. Z β := ∂ t β U t t = 0 = χ k D y β U .
By the smooth case, Z β ∈ W G , 0 1 , 2 ( Ω ) Z_{\beta}\in W^{1,2}_{G,0}(\Omega) Z β ∈ W G , 0 1 , 2 ( Ω ) is supported in K K K and has finite X α ( Ω ) \mathcal{X}_{\alpha}(\Omega) X α ( Ω ) -norm. Since χ = χ ( 0 , y ) \chi=\chi(0,y) χ = χ ( 0 , y ) for ∣ ξ ∣ ≤ r |\xi|\le r ∣ ξ ∣ ≤ r , its comparison polynomials have coefficients
a Z β = χ ( 0 , y ) k D y β g , q Z β = χ ( 0 , y ) k D y β q , M Z β = χ ( 0 , y ) k D y β M . a_{Z_\beta}=\chi(0,y)^kD_y^\beta g,\qquad q_{Z_\beta}=\chi(0,y)^kD_y^\beta q,\qquad M_{Z_\beta}=\chi(0,y)^kD_y^\beta M. a Z β = χ ( 0 , y ) k D y β g , q Z β = χ ( 0 , y ) k D y β q , M Z β = χ ( 0 , y ) k D y β M .
Step 2: The recursive estimate for A k A_k A k : : = max ∣ β ∣ = k ∥ Z β ∥ X α ( Ω ) :=\max_{|\beta|=k}\|Z_\beta\|_{\mathcal{X}_{\alpha}(\Omega)} := max ∣ β ∣ = k ∥ Z β ∥ X α ( Ω ) . Differentiating the equation of U t U_t U t at t = 0 t=0 t = 0 as in Step 1 of the proof for smooth data, we see that Z β Z_\beta Z β solves B U ( Z β , Y ) = F β ( Y ) \mathcal{B}_U(Z_\beta,Y)=\mathcal{F}_\beta(Y) B U ( Z β , Y ) = F β ( Y ) , where F β \mathcal{F}_\beta F β is given by (120) with ( y , A ) (y,A) ( y , A ) replaced by ( t , A t ) (t,A_t) ( t , A t ) and Ξ γ = ∂ t γ ( t , U t , D U t ) ∣ t = 0 \Xi_\gamma=\left.\partial_t^\gamma(t,U_t,DU_t)\right|_{t=0} Ξ γ = ∂ t γ ( t , U t , D U t ) ∣ t = 0 . For ∣ ξ ∣ ≤ r |\xi|\le r ∣ ξ ∣ ≤ r , we have D ξ χ = 0 D_\xi\chi=0 D ξ χ = 0 , so Ψ t \Psi_t Ψ t is induced by ( z , y ) ↦ ( z , y + χ ( 0 , y ) t ) (z,y)\mapsto(z,y+\chi(0,y)t) ( z , y ) ↦ ( z , y + χ ( 0 , y ) t ) in the original coordinate chart. Hence right multiplication by ( D Ψ t ) − 1 (D\Psi_t)^{-1} ( D Ψ t ) − 1 only changes the tangential columns of D U DU D U , and A t A_t A t keeps the structure (76) of the transformed density. For ∣ ξ ∣ ≥ r |\xi|\ge r ∣ ξ ∣ ≥ r , the weighted and ordinary derivatives are uniformly comparable. Moreover, (130) and (131) are real-analytic in ( t , U , D U ) (t,U,DU) ( t , U , D U ) on a fixed neighbourhood of the normalised data of U U U .
Applying Proposition 8.2 to A t A_t A t , we get
∥ D ( t , U , D U ) q A t ∣ t = 0 ( Ξ γ 1 , … , Ξ γ q ) ∥ Y α ( Ω ) ≤ Λ 0 Λ 1 q q ! ∏ i = 1 q ( ∥ Ξ γ i ∥ X α ( Ω ) + 1 { ∣ γ i ∣ = 1 } ) . (132) \left\|\left.D^q_{(t,U,DU)}A_t\right|_{t=0}(\Xi_{\gamma_1},\ldots,\Xi_{\gamma_q})\right\|_{\mathcal{Y}_{\alpha}(\Omega)}
\le\Lambda_0\Lambda_1^q q!\prod_{i=1}^q\left(\|\Xi_{\gamma_i}\|_{\mathcal{X}_{\alpha}(\Omega)}+\mathbf{1}_{\{|\gamma_i|=1\}}\right). \tag*{(132)} D ( t , U , D U ) q A t t = 0 ( Ξ γ 1 , … , Ξ γ q ) Y α ( Ω ) ≤ Λ 0 Λ 1 q q ! i = 1 ∏ q ( ∥ Ξ γ i ∥ X α ( Ω ) + 1 { ∣ γ i ∣ = 1 } ) . (132)
where Λ 0 , Λ 1 \Lambda_0,\Lambda_1 Λ 0 , Λ 1 depend on χ \chi χ , but not on q q q or the multi-indices;
the same proof gives F β , p r , y ∈ A p ( y ) ( P ) \mathcal{F}_{\beta,\mathrm{pr},y}\in\mathcal{A}_{p(y)}(\mathcal{P}) F β , pr , y ∈ A p ( y ) ( P ) .
Since Z β Z_\beta Z β is supported in K K K , the estimate (117) of Proposition 8.1 for compactly supported fields gives
A k ≤ C K max ∣ β ∣ = k ∥ F β ∥ Y α ( Ω ) . (133) A_k\le C_K\max_{|\beta|=k}\|\mathcal{F}_\beta\|_{\mathcal{Y}_{\alpha}(\Omega)}. \tag*{(133)} A k ≤ C K ∣ β ∣ = k max ∥ F β ∥ Y α ( Ω ) . (133)
with C K C_K C K independent of k k k . Summing (132) over the terms of (120) , we obtain A 1 ≤ C K Λ 0 Λ 1 A_1\le C_K\Lambda_0\Lambda_1 A 1 ≤ C K Λ 0 Λ 1 and
A k ≤ C K Λ 0 ∑ q = 2 k Λ 1 q ∑ γ 1 + ⋯ + γ q = β γ i ≠ 0 β ! γ 1 ! ⋯ γ q ! ∏ i = 1 q ( A ∣ γ i ∣ + 1 { ∣ γ i ∣ = 1 } ) , k ≥ 2. (134) A_k\le C_K\Lambda_0\sum_{q=2}^{k}\Lambda_1^q\sum_{\substack{\gamma_1+\cdots+\gamma_q=\beta\\\gamma_i\ne0}}\frac{\beta!}{\gamma_1!\cdots\gamma_q!}\prod_{i=1}^q\left(A_{|\gamma_i|}+\mathbf{1}_{\{|\gamma_i|=1\}}\right),\qquad k\ge2. \tag*{(134)} A k ≤ C K Λ 0 q = 2 ∑ k Λ 1 q γ 1 + ⋯ + γ q = β γ i = 0 ∑ γ 1 ! ⋯ γ q ! β ! i = 1 ∏ q ( A ∣ γ i ∣ + 1 { ∣ γ i ∣ = 1 } ) , k ≥ 2. (134)
Step 3: Factorial bounds via a majorant series. For fixed j 1 + ⋯ + j q = k j_1+\cdots+j_q=k j 1 + ⋯ + j q = k , comparing coefficients in ∏ i ( x 1 + ⋯ + x n − 2 ) j i = ( x 1 + ⋯ + x n − 2 ) k \prod_i(x_1+\cdots+x_{n-2})^{j_i}=(x_1+\cdots+x_{n-2})^k ∏ i ( x 1 + ⋯ + x n − 2 ) j i = ( x 1 + ⋯ + x n − 2 ) k yields the multinomial identity
∑ γ 1 + ⋯ + γ q = β ∣ γ i ∣ = j i β ! γ 1 ! ⋯ γ q ! = k ! j 1 ! ⋯ j q ! . \sum_{\substack{\gamma_1+\cdots+\gamma_q=\beta\\|\gamma_i|=j_i}}\frac{\beta!}{\gamma_1!\cdots\gamma_q!}=\frac{k!}{j_1!\cdots j_q!}. γ 1 + ⋯ + γ q = β ∣ γ i ∣ = j i ∑ γ 1 ! ⋯ γ q ! β ! = j 1 ! ⋯ j q ! k ! .
Thus (134) becomes
A k k ! ≤ C K Λ 0 ∑ q = 2 k Λ 1 q ∑ j 1 + ⋯ + j q = k j i ≥ 1 ∏ i = 1 q ( A j i j i ! + 1 { j i = 1 } ) , k ≥ 2. (135) \frac{A_k}{k!}\le C_K\Lambda_0\sum_{q=2}^{k}\Lambda_1^q\sum_{\substack{j_1+\cdots+j_q=k\\j_i\ge1}}\prod_{i=1}^q\left(\frac{A_{j_i}}{j_i!}+\mathbf{1}_{\{j_i=1\}}\right),\qquad k\ge2. \tag*{(135)} k ! A k ≤ C K Λ 0 q = 2 ∑ k Λ 1 q j 1 + ⋯ + j q = k j i ≥ 1 ∑ i = 1 ∏ q ( j i ! A j i + 1 { j i = 1 } ) , k ≥ 2. (135)
Let b = max { 1 , C K Λ 0 Λ 1 } ≥ A 1 b=\max\{1,C_K\Lambda_0\Lambda_1\}\ge A_1 b = max { 1 , C K Λ 0 Λ 1 } ≥ A 1 , and let w ( s ) = ∑ k ≥ 1 w k s k w(s)=\sum_{k\ge1}w_ks^k w ( s ) = ∑ k ≥ 1 w k s k be the solution with w ( 0 ) = 0 w(0)=0 w ( 0 ) = 0 of
w = b s + C K Λ 0 Λ 1 2 ( s + w ) 2 1 − Λ 1 ( s + w ) . (136) w=bs+C_K\Lambda_0\frac{\Lambda_1^2(s+w)^2}{1-\Lambda_1(s+w)}. \tag*{(136)} w = b s + C K Λ 0 1 − Λ 1 ( s + w ) Λ 1 2 ( s + w ) 2 . (136)
Comparing coefficients of s k s^k s k , we know the recurrence for w k w_k w k is exactly the equality case of (135) , with nonnegative coefficients and w 1 = b ≥ A 1 w_1=b\ge A_1 w 1 = b ≥ A 1 . By induction, A k / k ! ≤ w k A_k/k!\le w_k A k / k ! ≤ w k . With c = Λ 1 + C K Λ 0 Λ 1 2 c=\Lambda_1+C_K\Lambda_0\Lambda_1^2 c = Λ 1 + C K Λ 0 Λ 1 2 , we have the explicit formula for w ( s ) w(s) w ( s )
w = − s + 1 + ( b + 1 ) Λ 1 s − ( 1 + ( b + 1 ) Λ 1 s ) 2 − 4 ( b + 1 ) c s 2 c , w=-s+\frac{1+(b+1)\Lambda_1s-\sqrt{(1+(b+1)\Lambda_1s)^2-4(b+1)cs}}{2c}, w = − s + 2 c 1 + ( b + 1 ) Λ 1 s − ( 1 + ( b + 1 ) Λ 1 s ) 2 − 4 ( b + 1 ) cs ,
showing that w ( s ) w(s) w ( s ) is real-analytic near s = 0 s=0 s = 0 ; hence w k ≤ C k + 1 w_k \le C^{k+1} w k ≤ C k + 1 for some C > 0 C>0 C > 0 , so A k ≤ C k + 1 k ! A_k \le C^{k+1}k! A k ≤ C k + 1 k ! . Since χ = 1 \chi=1 χ = 1 near Q r ( y 0 ) Q_r(y_0) Q r ( y 0 ) , we have Z β = D y β U Z_\beta=D_y^\beta U Z β = D y β U there, which proves (127) . Finally, ∣ D y β g ∣ = ∣ a Z β ∣ ≤ C ∣ β ∣ + 1 ∣ β ∣ ! |D_y^\beta g|=|a_{Z_\beta}|\le C^{|\beta|+1}|\beta|! ∣ D y β g ∣ = ∣ a Z β ∣ ≤ C ∣ β ∣ + 1 ∣ β ∣ ! by (122) and the definition (114) of the X α \mathcal{X}_\alpha X α -norm, and similarly for q q q , M M M and U U U . This proves real analyticity. □
References
Department of Pure Mathematics and Mathematical Statistics, Cambridge University, Wilberforce Road, Cambridge CB3 0WB, UK
Email address: yl454@cam.ac.uk
Department of Mathematics, Zeeman Building, University of Warwick, Gibbet Hill Road, Coventry CV4 7AL, UK
Email address: Davide.Parise@warwick.ac.uk
Department of Mathematics, University College London, Gower Street, London WC1E 6BT, UK
Email address: yichen.zhang.24@ucl.ac.uk