This paper was converted on www.awesomepapers.org from LaTeX by an anonymous user.
Want to know more? Visit the Converter page.

Spectral asymptotics for the semiclassical Bochner Laplacian of a line bundle with constant rank curvature

Léo Morin
Abstract.

The goal of this paper is manyfold. Firstly, we want to give a short introduction to the Bochner Laplacian and explain why it acts locally as a magnetic Laplacian. Secondly, given a confining magnetic field, we use Agmon-like estimates to reduce its spectral study to magnetic Laplacians, in the semiclassical limit. Finally, we use this to translate already-known spectral asymptotics for the magnetic Laplacian to the Bochner Laplacian.

1 –   Introduction

1.1 – Motivations and context

The spectral theory of the magnetic Laplacian, and the Bochner Laplacian, has given rise to many interesting questions. First motivated by the Ginzburg-Landau theory, bound states of the magnetic Laplacian (ihd+A)(ihd+A)(ih\mathrm{d}+A)^{*}(ih\mathrm{d}+A) on a Riemannian manifold in the semiclassical limit h0h\rightarrow 0 were studied in many works (see the books [2, 12]), and appears to have very various behaviours according to the variations of the magnetic field B=dAB=\mathrm{d}A and the boundary conditions. The first main technique consisted in the construction of approximated eigenfunctions (see for instance the works of Helffer-Mohamed [6] and Helffer-Kordyukov [3, 4, 5]). More recently, an other approach was developped, which consists in an approximation of the operator itself, using semiclassical tools such as microlocalisation estimates and Birkhoff normal forms (As in Raymond-Vu Ngoc [13] and Helffer-Kordyukov-Raymond-Vu Ngoc [7]). In the semiclassical limit h0h\rightarrow 0, we recover the classical behaviour of a particle exposed to the magnetic field BB, since the magnetic Laplacian is the quantification of the classical energy.

If we are given a magnetic field BB which is not exact, there is no potential AA and we cannot define the magnetic Laplacian. However, the Bochner Laplacian 1p2ΔLp\frac{1}{p^{2}}\Delta^{L^{p}} appears to be the suitable generalization in this case, since it acts locally as a magnetic Laplacian. In this context the semiclassical parameter is p=h1p=h^{-1}. Its spectral theory appears to be deeply related to holomorphic structures and to the Kodaira Laplacian (or the renormalized Bochner Laplacian more generaly). For instance, this is exploited in the works of Marinescu-Savale [9], and Kordyukov [8]. In this last paper, the case of non-degenerate magnetic wells with full-rank magnetic field is studied, and expansions of the ground states energies are given, using quasimodes. In [11], similar results were obtained in the special case of the magnetic Laplacian, using a Birkhoff normal form, but also giving a description of semi-excited states, and a Weyl law. In [10], these result are generalized to constant-rank magnetic fields.

The spectral theory of the Bochner Laplacian is also deeply related to the global geometry of complex manifolds. For instance, in [1], a Weyl law was proven for ΔLp\Delta^{L^{p}}, and used to get Morse inequalities and Riemann-Roch formulas.

In this paper, we use Agmon-like estimates to reduce the spectral theory of the Bochner Laplacian with magnetic wells to magnetic Laplacians. Then, we deduce spectral asymptotics for the Bochner Laplacian using the results of [11, 10].

1.2 – The Bochner Laplacian on a line bundle

Let (M,g)(M,g) be a compact oriented manifold of dimension d>1d>1. We consider a complex line bundle LML\rightarrow M over MM, endowed with a Hermitian metric hh. In other words, we associate to each xMx\in M a 11-dimensional complex vector space LxL_{x}, and a Hermitian product hxh_{x} on LxL_{x}. LL is a d+1d+1-dimensional manifold such that L=xMLxL=\bigcup_{x\in M}L_{x}. A smooth section of LL (or LL-valued function) is a smooth function s:MLs:M\rightarrow L such that s(x)Lxs(x)\in L_{x}. It is the generalisation of the notion of function f:M𝐂f:M\rightarrow\mathbf{C}, but here the target space can vary with xMx\in M. Similarily, LL-valued kk-forms are sections of kTML\wedge^{k}T^{*}M\otimes L. We denote by 𝒞(M,L)\mathcal{C}^{\infty}(M,L) the set of smooth sections of LL, and Ωk(M,L)\Omega^{k}(M,L) the set of smooth LL-valued kk-forms.

We take L\nabla^{L} a Hermitian connexion on (L,h)(L,h). It is the generalisation of the exterior derivative d\mathrm{d}. The underlying idea is that the "derivative" of a LL-valued function should be LL-valued too. L:Ωk(M,L)Ωk+1(M,L)\nabla^{L}:\Omega^{k}(M,L)\rightarrow\Omega^{k+1}(M,L) satisfies:

(1.1) L(sα)=Lsα+sdα,s𝒞(M,L),αΩk(M,𝐂),\nabla^{L}(s\alpha)=\nabla^{L}s\wedge\alpha+s\mathrm{d}\alpha,\quad\forall s\in\mathcal{C}^{\infty}(M,L),\quad\alpha\in\Omega^{k}(M,\mathbf{C}),
(1.2) dh(s1,s2)=h(Ls1,s2)+h(s1,Ls2),s1,s2𝒞(M,L).\mathrm{d}h(s_{1},s_{2})=h(\nabla^{L}s_{1},s_{2})+h(s_{1},\nabla^{L}s_{2}),\quad\forall s_{1},s_{2}\in\mathcal{C}^{\infty}(M,L).

One can prove that (L)2:Ω0(M,L)Ω2(M,L)(\nabla^{L})^{2}:\Omega^{0}(M,L)\rightarrow\Omega^{2}(M,L) acts as a multiplication. There exists a real closed 22-form BB on MM such that:

(1.3) (L)2s=iBs,s𝒞(M,L).(\nabla^{L})^{2}s=iBs,\quad\forall s\in\mathcal{C}^{\infty}(M,L).

Example : The trivial line bundle. The line bundle L=M×𝐂L=M\times\mathbf{C}, such that Lx={x}×𝐂L_{x}=\{x\}\times\mathbf{C} is called the trivial line bundle. We identify sections s𝒞(M,L)s\in\mathcal{C}^{\infty}(M,L) with functions f𝒞(M)f\in\mathcal{C}^{\infty}(M) by s(x)=(x,f(x))s(x)=(x,f(x)). Similarily, LL-valued kk-forms are identified with 𝐂\mathbf{C}-valued kk-forms, and we recover the usual differential objects on MM. If LL is endowed with the Hermitian product hx((x,z1),(x,z2))=z1z¯2h_{x}((x,z_{1}),(x,z_{2}))=z_{1}\overline{z}_{2}, we call (L,h)(L,h) the trivial Hermitian line bundle. We write h(z1,z2)h(z_{1},z_{2}) for short. Hermitian connexions on the trivial line bundle are given by α=d+iα\nabla_{\alpha}=\mathrm{d}+i\alpha where αΩ1(M,𝐑)\alpha\in\Omega^{1}(M,\mathbf{R}) and d\mathrm{d} is the exterior derivative. The curvature of α\nabla_{\alpha} is α2=idα\nabla_{\alpha}^{2}=i\mathrm{d}\alpha, as shown by the easy but enlightening calculation:

(1.4) α2f=(d+iα)(df+ifα)=d2f+iαdf+id(fα)+ifαα=iαdf+idfα+ifdα=ifdα.\begin{matrix}[l]\nabla_{\alpha}^{2}f&=(\mathrm{d}+i\alpha)(\mathrm{d}f+if\alpha)=\mathrm{d}^{2}f+i\alpha\wedge\mathrm{d}f+i\mathrm{d}(f\alpha)+if\alpha\wedge\alpha\\ &=i\alpha\wedge\mathrm{d}f+i\mathrm{d}f\wedge\alpha+if\mathrm{d}\alpha=if\mathrm{d}\alpha.\end{matrix}

Let us describe now the Bochner Laplacian ΔL\Delta^{L} associated to a Hermitian connexion L\nabla^{L} on a Hermitian complex line bundle (L,h)(L,h). First note that the spaces 𝒞(M,L)=Ω0(M,L)\mathcal{C}^{\infty}(M,L)=\Omega^{0}(M,L) and Ω1(M,L)\Omega^{1}(M,L) are endowed with 𝖫2\mathsf{L}^{2}-norms. The norm of a section s𝒞(M,L)s\in\mathcal{C}^{\infty}(M,L) is:

(1.5) s2=Mhx(s(x),s(x))dνg(x),\|s\|^{2}=\int_{M}h_{x}(s(x),s(x))\mathrm{d}\nu_{g}(x),

where dνg\mathrm{d}\nu_{g} denotes the volume form of the oriented Riemannian manifold (M,g)(M,g). We denote by 𝖫2(M,L)\mathsf{L}^{2}(M,L) the completion of 𝒞(M,L)\mathcal{C}^{\infty}(M,L) for this norm. The definition of the norm of a LL-valued 11-form α\alpha is a little more involved. First, using a partition of unity, it is enough to define it for αΩ1(U,L)\alpha\in\Omega^{1}(U,L) where UU is a small open subset of MM. If UU is small enough, there exists a section e𝒞(U,L)e\in\mathcal{C}^{\infty}(U,L) such that hx(e(x),e(x))=1h_{x}(e(x),e(x))=1. Then for any αΩ1(U,L)\alpha\in\Omega^{1}(U,L), there exists a unique XTMX\in TM such that αx()=gx(Xx,)ex\alpha_{x}(\bullet)=g_{x}(X_{x},\bullet)e_{x} (we identify 11-forms with tangent vectors using the metric gg). We define:

(1.6) α2=Mgx(Xx,Xx)dνg(x).\|\alpha\|^{2}=\int_{M}g_{x}(X_{x},X_{x})\mathrm{d}\nu_{g}(x).

The completion of Ω1(M,L)\Omega^{1}(M,L) for this norm is denoted by 𝖫2Ω1(M,L)\mathsf{L}^{2}\Omega^{1}(M,L): it is the space of square-integrable LL-valued 11-forms. These norms are associated with scalar products, denoted by brackets .,.\langle.,.\rangle.

The formal adjoint of L:Ω0(M,L)Ω1(M,L)\nabla^{L}:\Omega^{0}(M,L)\rightarrow\Omega^{1}(M,L) for these scalar products is denoted by (L):Ω1(M,L)Ω0(M,L)(\nabla^{L})^{*}:\Omega^{1}(M,L)\rightarrow\Omega^{0}(M,L). The Bochner Laplacian ΔL\Delta^{L} is the self-adjoint extension of (L)L(\nabla^{L})^{*}\nabla^{L}. It is the operator associated with the quadratic form:

(1.7) Q(s1,s2)=Ls1,Ls2.Q(s_{1},s_{2})=\langle\nabla^{L}s_{1},\nabla^{L}s_{2}\rangle.

We denote by 𝖣𝗈𝗆(ΔL)\mathsf{Dom}(\Delta^{L}) its domain. 𝒞(M,L)\mathcal{C}^{\infty}(M,L) is a dense subspace of 𝖣𝗈𝗆(ΔL)\mathsf{Dom}(\Delta^{L}) and:

(1.8) ΔLs1,s2=Ls1,Ls2,s1,s2𝖣𝗈𝗆(ΔL).\langle\Delta^{L}s_{1},s_{2}\rangle=\langle\nabla^{L}s_{1},\nabla^{L}s_{2}\rangle,\quad\forall s_{1},s_{2}\in\mathsf{Dom}(\Delta^{L}).

Since MM is compact, one can prove that ΔL\Delta^{L} has compact resolvent, and we denote by

(1.9) λ1(ΔL)λ2(ΔL)\lambda_{1}(\Delta^{L})\leq\lambda_{2}(\Delta^{L})\leq...

the non-decreasing sequence of its eigenvalues. We will use the following notation for the Weil counting function:

N(ΔL,λ):={j;λj(ΔL)λ}.N(\Delta^{L},\lambda):=\sharp\{j;\lambda_{j}(\Delta^{L})\leq\lambda\}.

In this paper, we are interested in the semiclassical limit, i.e. the high curvature limit "B+B\rightarrow+\infty". We can increase the curvature BB using tensor products of LL. For any p𝐍p\in\mathbf{N}, we denote by Lp=LLL^{p}=L\otimes...\otimes L the pp-th tensor power of LL. LpL^{p} is still a complex line bundle of MM, with Lxp=LxLxL^{p}_{x}=L_{x}\otimes...\otimes L_{x}. It is endowed with the Hermitian product hxp(s1sp,s1sp)=Πi=1phx(si,si)h^{p}_{x}(s_{1}\otimes...\otimes s_{p},s_{1}\otimes...\otimes s_{p})=\Pi_{i=1}^{p}h_{x}(s_{i},s_{i}). The connexion L\nabla^{L} induces a Hermitian connexion Lp\nabla^{L^{p}} on LpL^{p} by the formula:

Lp(s1sp)=(Ls1)sp++s1(Lsp).\nabla^{L^{p}}(s_{1}\otimes...\otimes s_{p})=(\nabla^{L}s_{1})\otimes...\otimes s_{p}+...+s_{1}\otimes...\otimes(\nabla^{L}s_{p}).

The curvature of Lp\nabla^{L^{p}} is

(1.10) (Lp)2=ipB.(\nabla^{L^{p}})^{2}=ipB.

Hence, the high curvature limit is p+p\rightarrow+\infty. We want to invastigate the behaviour of λj(ΔLp)\lambda_{j}(\Delta^{L^{p}}) and the corresponding eigensections in the limit p+p\rightarrow+\infty.

1.3 – Main results

One can measure the "intensity" of the curvature BB (the magnetic field) in the following way. We denote by 𝐁x:TxMTxM\mathbf{B}_{x}:T_{x}M\rightarrow T_{x}M the linear operator defined by

(1.11) gx(𝐁xU,V)=Bx(U,V),U,VTxM.g_{x}(\mathbf{B}_{x}U,V)=B_{x}(U,V),\quad\forall U,V\in T_{x}M.

𝐁x\mathbf{B}_{x} is real and skew-symmetric with respect to gxg_{x}, and thus its eigenvalues lie on the imaginary axis and are symmetric with respect to the origin. We denote its non-zero eigenvalues by:

(1.12) ±iβ1(x),,±iβs(x),\pm i\beta_{1}(x),\cdots,\pm i\beta_{s}(x),

with βj(x)>0\beta_{j}(x)>0. Hence, the rank of 𝐁x\mathbf{B}_{x} is 2s2s and might depend on xx, but we will soon assume that ss is constant, at least locally. The magnetic intensity is the function b:M𝐑+b:M\rightarrow\mathbf{R}_{+} defined by:

(1.13) b(x)=j=1s(x)βj(x).b(x)=\sum_{j=1}^{s(x)}\beta_{j}(x).

This function is continuous on MM, but not smooth in general. However, note that it is smooth on a neighborhood of any point x0x_{0} where the (βj(x0))1js(\beta_{j}(x_{0}))_{1\leq j\leq s} are simple (if ss is locally constant near x0x_{0}).

One of the purposes of this article is to show that the eigensections of ΔLp\Delta^{L^{p}} are localized near the minimum points of bb, and to deduce that the low-lying spectrum of ΔLp\Delta^{L^{p}} is given by magnetic Laplacians on neighborhoods of the minimal points of bb.

We will do the following assumptions.

Assumptions.
  1. (A1)

    The minimal value of bb is only reached at non degenerate points x1,,xNMx_{1},\cdots,x_{N}\in M. We denote by b0=b(xj)=minxMbb_{0}=b(x_{j})=\min_{x\in M}b.

  2. (A2)

    The rank of 𝐁\mathbf{B} is constant on small neighborhoods U1,,UNU_{1},\cdots,U_{N} of x1,,xNx_{1},\cdots,x_{N}. We denote by 2sj2s_{j} the rank of 𝐁xj\mathbf{B}_{x_{j}}.

  3. (A3)

    We assume b0>0b_{0}>0, which is equivalent to say that sj>0s_{j}>0 for any jj.

As noticed in several papers, one can prove using Agmon-like estimates that the eigensections of ΔLp\Delta^{L^{p}} associated to low-lying eigenvalues are exponentially localized near {xM,b(x)=b0},\{x\in M,\quad b(x)=b_{0}\}, in the limit p+p\rightarrow+\infty. Now let us present the local model operators on UjU_{j}.

Recall that the 22-form BB is closed: dB=0\mathrm{d}B=0. Hence, if the open sets UjU_{j} are small enough, BB is exact on UjU_{j}: there exists AjΩ1(Uj)A_{j}\in\Omega^{1}(U_{j}) such that B=dAjB=\mathrm{d}A_{j} on UjU_{j}. We denote by p(j)\mathcal{L}_{p}^{(j)} the Dirichlet realization of (d+ipAj)(d+ipAj)(d+ipA_{j})^{*}(d+ipA_{j}) on 𝖫2(Uj)\mathsf{L}^{2}(U_{j}). It is the self-adjoint operator associated to the following sesquilinear form on 𝒞0(Uj)\mathcal{C}^{\infty}_{0}(U_{j}):

(1.14) Qj(u,v)=M(du+ipAju)(dv+ipAjv)¯dνg.Q_{j}(u,v)=\int_{M}(\mathrm{d}u+ipA_{j}u)\overline{(\mathrm{d}v+ipA_{j}v)}\mathrm{d}\nu_{g}.

We prove the following Theorem.

Theorem 1.

Let α(0,1/2)\alpha\in(0,1/2). Under assumptions (A1)(A1) and (A3)(A3), if η,ε>0\eta,\varepsilon>0 are small enough, then:

(1.15) λk(ΔLp)=λk(p(1)p(N))+𝒪(exp(εpα)),\lambda_{k}(\Delta^{L^{p}})=\lambda_{k}\left(\mathcal{L}_{p}^{(1)}\oplus...\oplus\mathcal{L}_{p}^{(N)}\right)+\mathcal{O}(exp(-\varepsilon p^{\alpha})),

uniformly with respect to k[1,Kp]k\in[1,K_{p}], where

Kp=min(N(ΔLp,(b0+η)p),N(p(1)p(N),(b0+η)p)),K_{p}=\min\left(N(\Delta^{L^{p}},(b_{0}+\eta)p),N(\mathcal{L}_{p}^{(1)}\oplus...\oplus\mathcal{L}_{p}^{(N)},(b_{0}+\eta)p)\right),

and N(𝒜,λ)N(\mathcal{A},\lambda) denotes the number of eigenvalues of an operator 𝒜\mathcal{A} below λ\lambda, counted with multiplicities.

As a corollary, we can deduce spectral asymptotics for ΔLp\Delta^{L^{p}} from already-known results for p(j)\mathcal{L}_{p}^{(j)}. Let us recall some of these results here.

1.3.1 –  The full-rank case

Under the assumptions (A1)(A2)(A3)(A1)-(A2)-(A3), we fix a j{1,,N}j\in\{1,\cdots,N\}, and we denote by Bj=dAjB_{j}=\mathrm{d}A_{j}. Hence, BjB_{j} is just the restriction of BB to the small open set UjU_{j}, where it admits a primitive AjA_{j}. h(j)\mathcal{L}_{h}^{(j)} is the magnetic Laplacian with Dirichlet boundary conditions on UjU_{j}, with magnetic field BjB_{j}. We first focus on the full-rank case, when the rank of BjB_{j} is maximal: 2sj=d2s_{j}=d. We define rj𝐍r_{j}\in\mathbf{N} by the condition

(1.16) n𝐙sj,0<=1sj|n|<rj=1sjnβ(xj)0.\forall n\in\mathbf{Z}^{s_{j}},\quad 0<\sum_{\ell=1}^{s_{j}}|n_{\ell}|<r_{j}\Rightarrow\sum_{\ell=1}^{s_{j}}n_{\ell}\beta_{\ell}(x_{j})\neq 0.

Note that, if the β(xj)\beta_{\ell}(x_{j}) are pairwise distinct, we can choose rj3r_{j}\geq 3. Moreover, if the open set UjU_{j} is small enough we have, for all xUjx\in U_{j} and n𝐙sjn\in\mathbf{Z}^{s_{j}},

(1.17) 0<=1sj|n|<rj=1sjnβ(x)0.0<\sum_{\ell=1}^{s_{j}}|n_{\ell}|<r_{j}\implies\sum_{\ell=1}^{s_{j}}n_{\ell}\beta_{\ell}(x)\neq 0.

The following Theorem was proved in [11].

Theorem 2.

We assume (A1)(A2)(A3)(A1)-(A2)-(A3) with 2sj=d2s_{j}=d and r3r\geq 3 in (1.16). Let η,ε>0\eta,\varepsilon>0 small enough. Then there exists a symplectomorphism ψ:UjT𝐑d/2\psi:U_{j}\rightarrow T^{*}\mathbf{R}^{d/2} such that:

(1.18) 1p2λk(p(j))=λk(n𝐍d𝒩p[j,n])+𝒪(prj/2+ε),\frac{1}{p^{2}}\lambda_{k}(\mathcal{L}_{p}^{(j)})=\lambda_{k}\left(\bigoplus_{n\in\mathbf{N}^{d}}\mathcal{N}_{p}^{[j,n]}\right)+\mathcal{O}(p^{-r_{j}/2+\varepsilon}),

uniformly with respect to k[1,K~p]k\in[1,\tilde{K}_{p}], where 𝒩p[j,n]\mathcal{N}_{p}^{[j,n]} is a pseudo-differential operator with principal symbol:

σ(𝒩p[j,n])=1p=1sj(2n+1)βψ1(x,ξ),\sigma(\mathcal{N}_{p}^{[j,n]})=\frac{1}{p}\sum_{\ell=1}^{s_{j}}(2n_{\ell}+1)\beta_{\ell}\circ\psi^{-1}(x,\xi),

and

K~p=min(N(p(j),(b0+η)p),N(n𝒩p[j,n],(b0+η)p1)).\tilde{K}_{p}=\min\left(N(\mathcal{L}_{p}^{(j)},(b_{0}+\eta)p),N(\oplus_{n}\mathcal{N}_{p}^{[j,n]},(b_{0}+\eta)p^{-1})\right).

Hence, we have a description of the semi-excited states of p(j)\mathcal{L}_{p}^{(j)}. In the same paper, Weyl estimates are proven for p(j)\mathcal{L}_{p}^{(j)}. We directly deduce the following Weyl estimates for ΔLp\Delta^{L^{p}}. A Similar formula was proven in [1] using a local approximation of the magnetic field by a constant field.

Corollary 3.

Assume (A1)(A2)(A3)(A_{1})-(A_{2})-(A_{3}), and for any j{1,,N}j\in\{1,\cdots,N\} that sj=d/2s_{j}=d/2, and that (β(xj))1N(\beta_{\ell}(x_{j}))_{1\leq\ell\leq N} are pairwise distinct. Then, for η>0\eta>0 small enough,

(1.19) N(ΔLp,(b0+η)p)(p2π)d/2n𝐍d/2b[n](x)b0+ηBd/2(d/2)!,N(\Delta^{L^{p}},(b_{0}+\eta)p)\sim\left(\frac{p}{2\pi}\right)^{d/2}\sum_{n\in\mathbf{N}^{d/2}}\int_{b^{[n]}(x)\leq b_{0}+\eta}\frac{B^{d/2}}{(d/2)!},

in the limit p+p\rightarrow+\infty, where b[n](x)==1d/2nβ(x)b^{[n]}(x)=\sum_{\ell=1}^{d/2}n_{\ell}\beta_{\ell}(x).

Finally, we also deduce asymptotic expansions of the first eigenvalues.

Corollary 4.

Assume (A1)(A2)(A3)(A1)-(A2)-(A3), and for any j{1,N}j\in\{1,\cdots N\} that sj=d/2s_{j}=d/2 and that (β(xj))1N(\beta_{\ell}(x_{j}))_{1\leq\ell\leq N} are pairwise distinct, and r:=minjrj5r:=\min_{j}r_{j}\geq 5. Then, for any k𝐍k\in\mathbf{N} and ε>0\varepsilon>0,

λk(ΔLp)=b0p+i=0r5αi,kpi/2+𝒪(p2r/2+ε),\lambda_{k}(\Delta^{L^{p}})=b_{0}p+\sum_{i=0}^{r-5}\alpha_{i,k}p^{-i/2}+\mathcal{O}(p^{2-r/2+\varepsilon}),

for some coefficients αi,k𝐑\alpha_{i,k}\in\mathbf{R}.

Remark. We also have geometric interpretations of the coefficients. First, the full expansion comes from the effective operator 𝒩p[j,0]\mathcal{N}_{p}^{[j,0]}, which is the reduction of p(j)\mathcal{L}_{p}^{(j)} to the lowest energy of the Harmonic oscillator describing the classical cyclotron motion. Moreover, α0,k\alpha_{0,k} is given by an eigenvalue of an other Harmonic oscillator whose symbol is the Hessian of bb at xjx_{j} (for some 1jN1\leq j\leq N): it describes a slow drift of the classical particle arround xjx_{j}. If the eigenvalues of this oscillator are simple, then a Birkhoff normal form can be used to show that αi,k=0\alpha_{i,k}=0 if ii is odd.

1.3.2 –  The constant-rank case

In the non-full-rank case, the kernel of BB (which corresponds to the directions of the field lines), has a great influence on the spectrum of ΔLp\Delta^{L^{p}}. Fix 1jN1\leq j\leq N. If the rank of BjB_{j} is constant, equal to 2sj2s_{j}, then its kernel as dimension kj=d2sjk_{j}=d-2s_{j}. The partial Hessian of bb at xjx_{j}, in the directions of the Kernel of BjB_{j}, is non-degenerate. we denote by

(1.20) νj,12,,νj,kj2\nu_{j,1}^{2},\cdots,\nu_{j,k_{j}}^{2}

its eigenvalues. For simplicity, we will make the following non-resonance assumptions (however, we can deal with resonances using a resonance order rr as in the full-rank case).

Assumptions.

(A4) For every jj, (β(xj))1sj(\beta_{\ell}(x_{j}))_{1\leq\ell\leq s_{j}} are non-resonnant in the following sense:

n𝐙sj,n0=1sjnβ(xj)0.\forall n\in\mathbf{Z}^{s_{j}},\quad n\neq 0\implies\sum_{\ell=1}^{s_{j}}n_{\ell}\beta_{\ell}(x_{j})\neq 0.

(A5) For every jj such that kj>0k_{j}>0, (νj,)1kj(\nu_{j,\ell})_{1\leq\ell\leq k_{j}} are non resonnant in the following sense:

n𝐙kj,n0=1kjnνj,0.\forall n\in\mathbf{Z}^{k_{j}},\quad n\neq 0\implies\sum_{\ell=1}^{k_{j}}n_{\ell}\nu_{j,\ell}\neq 0.

Applying the results of [10] to get spectral asymptotics for h(j)\mathcal{L}_{h}^{(j)}, we deduce from Theorem 1 the following corollary.

Corollary 5.

Assume (A1)(A2)(A3)(A4)(A5)(A1)-(A2)-(A3)-(A4)-(A5), and let n𝐍n\in\mathbf{N}. Then λn(ΔLp)\lambda_{n}(\Delta^{L^{p}}) admits a full asymptotic expansion in powers of p1/2p^{-1/2}:

λn(ΔLp)=b0p+κp1/2+i0αi,npi/2+𝒪(p).\lambda_{n}(\Delta^{L^{p}})=b_{0}p+\kappa p^{1/2}+\sum_{i\geq 0}\alpha_{i,n}p^{-i/2}+\mathcal{O}(p^{-\infty}).

Moreover:

  1. If there is at least one jj such that kj=0k_{j}=0, then κ=0\kappa=0.

  2. If j{1,,N},kj>0\forall j\in\{1,\cdots,N\},k_{j}>0, then κ=minj=1,,N=1kjνj,.\kappa=\min_{j=1,\cdots,N}\sum_{\ell=1}^{k_{j}}\nu_{j,\ell}.

2 –   Some Remarks

2.1 – The Bochner Laplacian and the magnetic Laplacian are locally the same

If UU is any open subset of MM such that there exists a non-vanishing section e𝒞(U,L)e\in\mathcal{C}^{\infty}(U,L), then any s𝒞(U,L)s\in\mathcal{C}^{\infty}(U,L) can be written s=ues=ue for some u𝒞(M)u\in\mathcal{C}^{\infty}(M). Hence,

s=(e)u+e(du)=e[(d+iA)u],\nabla s=(\nabla e)u+e(\mathrm{d}u)=e[(\mathrm{d}+iA)u],

with e=eiA\nabla e=eiA. Moreover,

2s=e[(d+iA)u]+ed[(d+iA)u]=e(iAdu)+e(iAiA)u+ed2u+ieudA+eduiA=ieudA=(idA)s,\begin{matrix}[l]\nabla^{2}s&=\nabla e\wedge[(\mathrm{d}+iA)u]+e\mathrm{d}[(\mathrm{d}+iA)u]\\ &=e(iA\wedge\mathrm{d}u)+e(iA\wedge iA)u+e\mathrm{d}^{2}u+ieu\mathrm{d}A+e\mathrm{d}u\wedge iA\\ &=ieu\mathrm{d}A=(i\mathrm{d}A)s,\end{matrix}

and thus B=dAB=\mathrm{d}A. Hence \nabla acts locally as d+iA\mathrm{d}+iA, and ΔL\Delta^{L} as the magnetic Laplacian (d+iA)(d+iA)(\mathrm{d}+iA)^{*}(\mathrm{d}+iA). This is the core of Theorem 1.

2.2 – On the quantization of a magnetic field

If we are given a closed 22-form BB (the magnetic field), the quantization question constist in finding a quantum operator associated to BB. If BB is exact, this question is answered by the semiclassical magnetic Laplacian (d+iA)(d+iA)(\hbar\mathrm{d}+iA)^{*}(\hbar\mathrm{d}+iA), with B=dAB=\mathrm{d}A. Here, >0\hbar>0 is the semiclassical parameter (Planck’s constant) and the semiclassical limit is 0\hbar\rightarrow 0.

If BB is not exact, but if there exists an Hermitian line bundle with Hermitian connexion such that 2=iB\nabla^{2}=iB, then the Bochner Laplacian \nabla^{*}\nabla acts locally as the magnetic Laplacian and hence it is a good candidate. Moreover, we have locally

ΔLp=(d+ipA)(d+ipA)=p2(1pd+iA)(1pd+iA),\Delta^{L^{p}}=(\mathrm{d}+ipA)^{*}(\mathrm{d}+ipA)=p^{2}(\frac{1}{p}\mathrm{d}+iA)^{*}(\frac{1}{p}\mathrm{d}+iA),

so that the semiclassical parameter is now =1p\hbar=\frac{1}{p} (Also notice the p2p^{2} factor which is important for the eigenvalue asymptotics). The limit 0\hbar\rightarrow 0 is equivalent to p+p\rightarrow+\infty exept that the semiclassical parameter becomes discrete (p𝐍p\in\mathbf{N}).

A new question arises : When does such an Hermitian line bundle exists ? Weil’s Theorem states that it exists if and only if BB satisfies the prequantization condition:

(2.1) [B]2π𝐙,[B]\in 2\pi\mathbf{Z},

where [B][B] denotes the cohomology class of BB. This condition also enlightens the discreteness of the semiclassical parameter. Indeed, if one wants to quantize the magnetic field 1B\frac{1}{\hbar}B, then one must have [1B]2π𝐙\left[\frac{1}{\hbar}B\right]\in 2\pi\mathbf{Z}, and thus 1p𝐙\frac{1}{p}\in\mathbf{Z}, unless [B]=0[B]=0 which means that BB is exact (and thus we can use the magnetic Laplacian !).

3 –   Proof of Theorem 1

3.1 – Agmon-like estimates

In this section, we prove exponential decay on the eigensections of ΔLp\Delta^{L^{p}}, away from the set {x1,xN}\{x_{1},\cdots x_{N}\}. We need the following Lemma.

Lemma 6.

There exist p0>0p_{0}>0 and C0>0C_{0}>0 such that, for pp0p\geq p_{0} and s𝒞(M,L)s\in\mathcal{C}^{\infty}(M,L),

(1+C0p1/4)Lps2pM(b(x)C0p1/4)|s(x)|2dx.(1+\frac{C_{0}}{p^{1/4}})\|\nabla^{L^{p}}s\|^{2}\geq p\int_{M}(b(x)-\frac{C_{0}}{p^{1/4}})|s(x)|^{2}\mathrm{d}x.
Proof.

Take a partition of unity (χα)α(\chi_{\alpha})_{\alpha} on MM, such that 𝗌𝗎𝗉𝗉χαUα\mathsf{supp}\chi_{\alpha}\subset U_{\alpha} with UαU_{\alpha} a small open subset of MM, and

(3.1) 1=αχα2,α|dχα|2C.1=\sum_{\alpha}\chi_{\alpha}^{2},\quad\sum_{\alpha}|\mathrm{d}\chi_{\alpha}|^{2}\leq C.

Then, for any s𝒞(M,L)s\in\mathcal{C}^{\infty}(M,L) we have

(3.2) Lps2=αLp(χαs)2α|dχα|s2,\|\nabla^{L^{p}}s\|^{2}=\sum_{\alpha}\|\nabla^{L^{p}}(\chi_{\alpha}s)\|^{2}-\sum_{\alpha}\||\mathrm{d}\chi_{\alpha}|s\|^{2},

and χαs𝒞0(Uα,L)\chi_{\alpha}s\in\mathcal{C}^{\infty}_{0}(U_{\alpha},L). If every UαU_{\alpha} are small enough, we can find a non-vanishing section eαe_{\alpha} on UαU_{\alpha} which we can use to trivialize the line bundle. Writting χαs=uαeα\chi_{\alpha}s=u_{\alpha}e_{\alpha} for some uα𝒞0(Uα)u_{\alpha}\in\mathcal{C}^{\infty}_{0}(U_{\alpha}), we have

(3.3) Lp(χαs)=[(d+ipAα)uα]eα,\nabla^{L^{p}}(\chi_{\alpha}s)=\left[(\mathrm{d}+ipA_{\alpha})u_{\alpha}\right]e_{\alpha},

where AαΩ1(Uα)A_{\alpha}\in\Omega^{1}(U_{\alpha}) is such that B=dAαB=\mathrm{d}A_{\alpha}. For the magnetic Laplacian (d+ipAα)(d+ipAα)(\mathrm{d}+ipA_{\alpha})^{*}(\mathrm{d}+ipA_{\alpha}), the desired inequality is well-known: There exist pα𝐍p_{\alpha}\in\mathbf{N}, Cα>0C_{\alpha}>0 such that, for every uα𝒞(Uα)u_{\alpha}\in\mathcal{C}^{\infty}(U_{\alpha}), and ppαp\geq p_{\alpha}:

(3.4) (1+Cαp1/4)(d+ipAα)uα2pUα(b(x)Cαp1/4)|uα(x)|2dx.(1+\frac{C_{\alpha}}{p^{1/4}})\|(\mathrm{d}+ipA_{\alpha})u_{\alpha}\|^{2}\geq p\int_{U_{\alpha}}(b(x)-\frac{C_{\alpha}}{p^{1/4}})|u_{\alpha}(x)|^{2}\mathrm{d}x.

Using (3.2), (3.3), and (3.4), we deduce that

(3.5) (1+C0p1/4)Lps2pM(b(x)C0p1/4)|s(x)|2dx(1+C0p1/4)α|dχα|s2,(1+\frac{C_{0}}{p^{1/4}})\|\nabla^{L^{p}}s\|^{2}\geq p\int_{M}(b(x)-\frac{C_{0}}{p^{1/4}})|s(x)|^{2}\mathrm{d}x-(1+\frac{C_{0}}{p^{1/4}})\sum_{\alpha}\||\mathrm{d}\chi_{\alpha}|s\|^{2},

with C0=maxαCαC_{0}=\max_{\alpha}C_{\alpha}, and for pmaxαpαp\geq\max_{\alpha}p_{\alpha}. Finally, (3.1) yields

(3.6) (1+C0p1/4)α|dχα|s2C(1+C0p1/4)s2C~p3/4s2.(1+\frac{C_{0}}{p^{1/4}})\sum_{\alpha}\||\mathrm{d}\chi_{\alpha}|s\|^{2}\leq C(1+\frac{C_{0}}{p^{1/4}})\|s\|^{2}\leq\tilde{C}p^{3/4}\|s\|^{2}.

Hence, up to changing C0C_{0} into C0+C~C_{0}+\tilde{C}, Lemma 6 is proved. ∎

Now we can use Lemma 6 to prove Agmon-like decay estimates.

Theorem 7.

Let α(0,1/2)\alpha\in(0,1/2), η>0\eta>0, and Kη={b(x)b0+2η}K_{\eta}=\{b(x)\leq b_{0}+2\eta\}. There exist C>0C>0 and p0>0p_{0}>0 such that, for all pp0p\geq p_{0} and all eigenpair (λ,ψ)(\lambda,\psi) of ΔLp\Delta^{L^{p}} with λ(b0+η)p\lambda\leq(b_{0}+\eta)p,

M|ed(x,Kη)pαψ|2dqCψ2.\int_{M}|e^{\mathrm{d}(x,K_{\eta})p^{\alpha}}\psi|^{2}\mathrm{d}q\leq C\|\psi\|^{2}.
Proof.

Let Φ:M𝐑\Phi:M\rightarrow\mathbf{R} be a Lipschitz function. The Agmon formula is:

(3.7) ΔLpeΦψ,eΦψ=λeΦψ2+dΦeΦψ2.\langle\Delta^{L^{p}}e^{\Phi}\psi,e^{\Phi}\psi\rangle=\lambda\|e^{\Phi}\psi\|^{2}+\|\mathrm{d}\Phi e^{\Phi}\psi\|^{2}.

Using Lemma 6, we deduce that:

(pb(x)C0p3/4(1+C0p1/4)(λ+|dΦ|2))|eΦψ|2dx0.\displaystyle\int(pb(x)-C_{0}p^{3/4}-(1+C_{0}p^{-1/4})(\lambda+|\mathrm{d}\Phi|^{2}))|e^{\Phi}\psi|^{2}\mathrm{d}x\leq 0.

We split this integral into two parts.

Kηc\displaystyle\int_{K^{c}_{\eta}} (pb(x)C0p3/4(1+C0p1/4)(λ+|dΦ|2))|eΦψ|2dx\displaystyle(pb(x)-C_{0}p^{3/4}-(1+C_{0}p^{-1/4})(\lambda+|\mathrm{d}\Phi|^{2}))|e^{\Phi}\psi|^{2}\mathrm{d}x
Kη(pb(x)+C0p3/4+(1+C0p1/4)(λ+|dΦ|2))|eΦψ|2dx\displaystyle\leq\int_{K_{\eta}}(-pb(x)+C_{0}p^{3/4}+(1+C_{0}p^{-1/4})(\lambda+|\mathrm{d}\Phi|^{2}))|e^{\Phi}\psi|^{2}\mathrm{d}x

We choose Φ\Phi:

Φm(x)=χm(d(x,Kη))pα,for m>0,\Phi_{m}(x)=\chi_{m}(d(x,K_{\eta}))p^{\alpha},\quad\text{for }m>0,

where χm(t)=t\chi_{m}(t)=t for t<mt<m, χm(t)=0\chi_{m}(t)=0 for t>2mt>2m, and χm\chi_{m}^{\prime} uniformly bounded with respect to mm. Since Φm=0\Phi_{m}=0 on KηK_{\eta} and pb(x)C0p3/4>0pb(x)-C_{0}p^{3/4}>0 for pp large enough, we have:

Kηc\displaystyle\int_{K^{c}_{\eta}} (pb(x)C0p3/4(1+C0p1/4)(λ+|dΦm|2))|eΦmψ|2dx\displaystyle(pb(x)-C_{0}p^{3/4}-(1+C_{0}p^{-1/4})(\lambda+|\mathrm{d}\Phi_{m}|^{2}))|e^{\Phi_{m}}\psi|^{2}\mathrm{d}x
(b0+η)pKη(1+C0p1/4)|ψ|2dxCpψ2.\displaystyle\leq(b_{0}+\eta)p\int_{K_{\eta}}(1+C_{0}p^{-1/4})|\psi|^{2}\mathrm{d}x\leq Cp\|\psi\|^{2}.

Moreover, since λ(b0+η)p\lambda\leq(b_{0}+\eta)p and |dΦm|2Cp2α|\mathrm{d}\Phi_{m}|^{2}\leq Cp^{2\alpha},

Kηc(pb(x)C0p3/4(1+C0p1/4)(b0p+ηp+Cp2α)|eΦmψ|2dx\displaystyle\int_{K^{c}_{\eta}}(pb(x)-C_{0}p^{3/4}-(1+C_{0}p^{-1/4})(b_{0}p+\eta p+Cp^{2\alpha})|e^{\Phi_{m}}\psi|^{2}\mathrm{d}x Cpψ2\displaystyle\leq Cp\|\psi\|^{2}
pKηc(b(x)(b0+η)C0p1/4C~p2α1)|eΦmψ|2dx\displaystyle p\int_{K^{c}_{\eta}}(b(x)-(b_{0}+\eta)-C_{0}p^{-1/4}-\tilde{C}p^{2\alpha-1})|e^{\Phi_{m}}\psi|^{2}\mathrm{d}x Cpψ2,\displaystyle\leq Cp\|\psi\|^{2},

for pp large enough. But b(x)>b0+2ηb(x)>b_{0}+2\eta on KηcK_{\eta}^{c}, so there is a δ>0\delta>0 and p0>0p_{0}>0 such that, for pp0p\geq p_{0}:

δKc|eΦmψ|2dqCψ2.\displaystyle\delta\int_{K^{c}}|e^{\Phi_{m}}\psi|^{2}\mathrm{d}q\leq C\|\psi\|^{2}.

Since Φm=0\Phi_{m}=0 on KK, we get a new C>0C>0 such that:

eΦmψ2Cψ2,\|e^{\Phi_{m}}\psi\|^{2}\leq C\|\psi\|^{2},

and we can use Fatou’s lemma in the limit m+m\rightarrow+\infty to get the desired inequality. ∎

Corollary 8.

Let ε>0\varepsilon>0 and χ:M[0,1]\chi:M\rightarrow[0,1] be a smooth cutoff function, being 1 on a small neighborhood of

Kη+ε={x;d(x,Kη)<ε}.K_{\eta}+\varepsilon=\{x;\quad\mathrm{d}(x,K_{\eta})<\varepsilon\}.

Then, for any eigenpair (λ,ψ)(\lambda,\psi) of ΔLp\Delta^{L^{p}}, with λ(b0+η)p\lambda\leq(b_{0}+\eta)p we have:

ψ=χψ+𝒪(eεpα)ψ,\psi=\chi\psi+\mathcal{O}(e^{-\varepsilon p^{\alpha}})\|\psi\|,

and

Lp(χψ)=Lpψ+𝒪(p1/2eεpα)ψ,\nabla^{L^{p}}(\chi\psi)=\nabla^{L^{p}}\psi+\mathcal{O}(p^{1/2}e^{-\varepsilon p^{\alpha}})\|\psi\|,

uniformly with respect to (λ,ψ)(\lambda,\psi).

Proof.

By Theorem 7, we have:

(3.8) (1χ)ψ2(Kη+ε)c|ψ|2dqMe2εpα|ed(x,Kη)pαψ|2dxCe2εpαψ2,\|(1-\chi)\psi\|^{2}\leq\int_{(K_{\eta}+\varepsilon)^{c}}|\psi|^{2}\mathrm{d}q\leq\int_{M}e^{-2\varepsilon p^{\alpha}}|e^{\mathrm{d}(x,K_{\eta})p^{\alpha}}\psi|^{2}\mathrm{d}x\leq Ce^{-2\varepsilon p^{\alpha}}\|\psi\|^{2},

which gives the first estimates. Moreover, we have with Φ(x)=d(x,Kη)\Phi(x)=\mathrm{d}(x,K_{\eta}),

eΦpαLpψLp(eΦpαψ)+pαdΦeΦpαψ,\|e^{\Phi p^{\alpha}}\nabla^{L^{p}}\psi\|\leq\|\nabla^{L^{p}}(e^{\Phi p^{\alpha}}\psi)\|+p^{\alpha}\|\mathrm{d}\Phi e^{\Phi p^{\alpha}}\psi\|,

and using Agmon’s formula 3.7 and Theorem 7:

Lp(eΦpαψ)2=λeΦpαψ2+p2αdΦeΦpαψ2C2pψ2.\|\nabla^{L^{p}}(e^{\Phi p^{\alpha}}\psi)\|^{2}=\lambda\|e^{\Phi p^{\alpha}}\psi\|^{2}+p^{2\alpha}\|\mathrm{d}\Phi e^{\Phi p^{\alpha}}\psi\|^{2}\leq C^{2}p\|\psi\|^{2}.

Thus,

(3.9) eΦpαLpψCp1/2ψ2.\|e^{\Phi p^{\alpha}}\nabla^{L^{p}}\psi\|\leq Cp^{1/2}\|\psi\|^{2}.

We can use these Agmon estimates on Lpψ\nabla^{L^{p}}\psi to get our second result.

(3.10) Lp((1χ)ψ)(Lpχ)ψ+(1χ)Lpψ\|\nabla^{L^{p}}((1-\chi)\psi)\|\leq\|(\nabla^{L^{p}}\chi)\psi\|+\|(1-\chi)\nabla^{L^{p}}\psi\|

The first term is dominated by

(3.11) (Lpχ)ψC(1χ¯)ψ\|(\nabla^{L^{p}}\chi)\psi\|\leq C\|(1-\overline{\chi})\psi\|

where χ¯\overline{\chi} is a cutoff function such that χ¯=1\overline{\chi}=1 on Kη+εK_{\eta}+\varepsilon and χ¯=0\overline{\chi}=0 on 𝗌𝗎𝗉𝗉(1χ)\mathsf{supp}(1-\chi). We can apply (3.8) to χ¯\overline{\chi} to get:

(3.12) (Lpχ)ψCeεpαψ.\|(\nabla^{L^{p}}\chi)\psi\leq Ce^{-\varepsilon p^{\alpha}}\|\psi\|.

The second term of (3.10) is dominated as in (3.8), using (3.9):

(3.13) (1χ)LpψCp1/2eεpαψ.\|(1-\chi)\nabla^{L^{p}}\psi\|\leq Cp^{1/2}e^{-\varepsilon p^{\alpha}}\|\psi\|.

Finally, (3.10) with (3.12) and (3.13) yields

Lp((1χ)ψ)Cp1/2eεpαψ.\|\nabla^{L^{p}}((1-\chi)\psi)\|\leq Cp^{1/2}e^{-\varepsilon p^{\alpha}}\|\psi\|.

3.2 – Comparison of the spectrum of ΔLp\Delta^{L^{p}} and p(j)\mathcal{L}_{p}^{(j)}

We recall that the minimum b0b_{0} of bb is reached at x1,,xNx_{1},\cdots,x_{N} in a non-degenerate way. For η>0\eta>0 small enough, the compact set Kη={b(x)b0+η}K_{\eta}=\{b(x)\leq b_{0}+\eta\} has NN disjoint connected components Kη(j)K_{\eta}^{(j)} such that xjKη(j)x_{j}\in K_{\eta}^{(j)}. We fix the value of η\eta, and we take UjU_{j} a neighborhood of Kη(j)K_{\eta}^{(j)}. For ε>0\varepsilon>0 sufficiently small, Kη(j)+2εUjK_{\eta}^{(j)}+2\varepsilon\subset U_{j}.

We denote by BjB_{j} the restriction of BB to UjU_{j}. p(j)\mathcal{L}_{p}^{(j)} is the Dirichlet realisation of (d+ipAj)(d+ipAj)(\mathrm{d}+ipA_{j})^{*}(\mathrm{d}+ipA_{j}), with AjΩ1(Uj,L)A_{j}\in\Omega^{1}(U_{j},L) such that Bj=dAjB_{j}=\mathrm{d}A_{j}. It is the self adjoint operator associated to the quadratic form:

(3.14) Qj(u,v)=Uj(d+ipAj)u(d+ipAj)v¯dx,u,v𝖧01(Uj).Q_{j}(u,v)=\int_{U_{j}}(\mathrm{d}+ipA_{j})u\overline{(\mathrm{d}+ipA_{j})v}\mathrm{d}x,\quad\forall u,v\in\mathsf{H}^{1}_{0}(U_{j}).

Let us denote by

(3.15) Kp=min[N(ΔLp,(b0+η)p);N(j=1Np(j),(0,b0+η)p)].K_{p}=\min\left[N(\Delta^{L^{p}},(b_{0}+\eta)p);N\left(\bigoplus_{j=1}^{N}\mathcal{L}_{p}^{(j)},(0,b_{0}+\eta)p\right)\right].

We split the proof of Theorem 1 into two Lemmas.

Lemma 9.

Let α(0,1/2)\alpha\in(0,1/2). We have:

λk(j=1Np(j))λk(ΔLp)+𝒪(exp(εpα)),\lambda_{k}(\bigoplus_{j=1}^{N}\mathcal{L}_{p}^{(j)})\leq\lambda_{k}(\Delta^{L^{p}})+\mathcal{O}(\exp(-\varepsilon p^{\alpha})),

uniformily with respect to k[1,Kp]k\in[1,K_{p}].

Proof.

We prove this using the min-max principle. For k[1,Jp]k\in[1,J_{p}], let ψk\psi_{k} be the normalized eigenfunction associated to λk(ΔLp)\lambda_{k}(\Delta^{L^{p}}). We will define the quasimode uj,k𝒞0(Uj)u_{j,k}\in\mathcal{C}^{\infty}_{0}(U_{j}) using a local trivialisation of LpL^{p} on UjU_{j}. Let ej𝒞(Uj,L)e_{j}\in\mathcal{C}^{\infty}(U_{j},L) be the non-vanishing local section of LL such that, for any u𝒞(Uj)u\in\mathcal{C}^{\infty}(U_{j}),

(3.16) Lp(uej)=[(d+ipAj)u]ej.\nabla^{L^{p}}(ue_{j})=\left[(\mathrm{d}+ipA_{j})u\right]e_{j}.

Let χj𝒞0(Uj)\chi_{j}\in\mathcal{C}^{\infty}_{0}(U_{j}) be a smooth cutoff function, such that χj=1\chi_{j}=1 on Kη(j)+εK_{\eta}^{(j)}+\varepsilon. We define uj,k𝒞0(Uj)u_{j,k}\in\mathcal{C}^{\infty}_{0}(U_{j}) by χjψk=uj,kej\chi_{j}\psi_{k}=u_{j,k}e_{j}, and

uk=u1,kuN,k.u_{k}=u_{1,k}\oplus...\oplus u_{N,k}.

Then

jp(j)uk,uk=j=1Np(j)uj,k,uj,k=j=1N(d+ipAj)uj,k2.\displaystyle\langle\bigoplus_{j}\mathcal{L}_{p}^{(j)}u_{k},u_{k}\rangle=\sum_{j=1}^{N}\langle\mathcal{L}_{p}^{(j)}u_{j,k},u_{j,k}\rangle=\sum_{j=1}^{N}\|(\mathrm{d}+ipA_{j})u_{j,k}\|^{2}.

Moreover, by (3.16),

(d+ipAj)uj,k2=Uj|(d+ipAj)uj,k|2dx=Uj|Lp(χjψk)|2dx.\|(\mathrm{d}+ipA_{j})u_{j,k}\|^{2}=\int_{U_{j}}|(\mathrm{d}+ipA_{j})u_{j,k}|^{2}\mathrm{d}x=\int_{U_{j}}|\nabla^{L^{p}}(\chi_{j}\psi_{k})|^{2}\mathrm{d}x.

Now, χ=j=1Nχj\chi=\sum_{j=1}^{N}\chi_{j} satisfies the assumptions of Corollary 8 (with 2ε2\varepsilon instead of ε\varepsilon). Thus,

jp(j)uk,uk=M|Lp(χψk)|2dx=Lpψk2+𝒪(p1/2e2εpα)ψk,\langle\bigoplus_{j}\mathcal{L}_{p}^{(j)}u_{k},u_{k}\rangle=\int_{M}|\nabla^{L^{p}}(\chi\psi_{k})|^{2}\mathrm{d}x=\|\nabla^{L^{p}}\psi_{k}\|^{2}+\mathcal{O}(p^{1/2}e^{-2\varepsilon p^{\alpha}})\|\psi_{k}\|,

uniformly with respect to kk. ψk\psi_{k} being the eigensection associated to λk(ΔLp)\lambda_{k}(\Delta^{L^{p}}), it remains:

jp(j)uk,uk=(λk(ΔLp)+𝒪(p1/2e2εpα))ψk.\langle\bigoplus_{j}\mathcal{L}_{p}^{(j)}u_{k},u_{k}\rangle=\left(\lambda_{k}(\Delta^{L^{p}})+\mathcal{O}(p{1/2}e^{-2\varepsilon p^{\alpha}})\right)\|\psi_{k}\|.

This is true for every k[1,Kp]k\in[1,K_{p}]. Hence, for 1ikKp1\leq i\leq k\leq K_{p} we have

jp(j)ui,ui(λk(ΔLp)+𝒪(p1/2e2εpα))ψk,\langle\bigoplus_{j}\mathcal{L}_{p}^{(j)}u_{i},u_{i}\rangle\leq\left(\lambda_{k}(\Delta^{L^{p}})+\mathcal{O}(p^{1/2}e^{-2\varepsilon p^{\alpha}})\right)\|\psi_{k}\|,

and the Lemma follows from the min-max principle, because the vector space ranged by (ui)1ik(u_{i})_{1\leq i\leq k} is kk-dimensional (and p1/2e2εpα=𝒪(eεpα)p^{1/2}e^{-2\varepsilon p^{\alpha}}=\mathcal{O}(e^{-\varepsilon p^{\alpha}})). ∎

The reverse inequality is proven similarily.

Lemma 10.

Let α(0,1/2)\alpha\in(0,1/2). We have:

λk(ΔLp)λk(j=1Np(j))+𝒪(exp(εpα)),\lambda_{k}(\Delta^{L^{p}})\leq\lambda_{k}(\bigoplus_{j=1}^{N}\mathcal{L}_{p}^{(j)})+\mathcal{O}(\exp(-\varepsilon p^{\alpha})),

uniformily with respect to k[1,Kp]k\in[1,K_{p}].

Proof.

The kk-th eigenvalue of j=1Np(j)\bigoplus_{j=1}^{N}\mathcal{L}_{p}^{(j)} is given by an eigenpair (μk,uk)(\mu_{k},u_{k}) of p(jk)\mathcal{L}_{p}^{(j_{k})} for some jk{1,,N}j_{k}\in\{1,\cdots,N\}. Let χk𝒞0(Ujk)\chi_{k}\in\mathcal{C}^{\infty}_{0}(U_{j_{k}}) be a cutoff function equal to 11 on Kη(jk)+2εK_{\eta}^{(j_{k})}+2\varepsilon. Then, Agmon estimates (Theorem 7) for p(j)\mathcal{L}_{p}^{(j)} imply that

(d+ipA)uk=(d+ipA)(χkuk)+𝒪(eεpα)uk\displaystyle(\mathrm{d}+ipA)u_{k}=(\mathrm{d}+ipA)(\chi_{k}u_{k})+\mathcal{O}(e^{-\varepsilon p^{\alpha}})\|u_{k}\|

uniformly with respect to kk. We define sk=χkukejks_{k}=\chi_{k}u_{k}e_{j_{k}}, where ejke_{j_{k}} satisfies (3.16), and we extend sks_{k} by 0 outside UjkU_{j_{k}}. Then,

ΔLpsk,sk=Ujk|(d+ipA)χkuk|2dx=Ujk|(d+ipA)uk|2dx+𝒪(eεpα)=μkuk2+𝒪(eεpα).\begin{matrix}[l]\langle\Delta^{L^{p}}s_{k},s_{k}\rangle&=\int_{U_{j_{k}}}|(\mathrm{d}+ipA)\chi_{k}u_{k}|^{2}\mathrm{d}x\\ &=\int_{U_{j_{k}}}|(\mathrm{d}+ipA)u_{k}|^{2}\mathrm{d}x+\mathcal{O}(e^{-\varepsilon p^{\alpha}})\\ &=\mu_{k}\|u_{k}\|^{2}+\mathcal{O}(e^{-\varepsilon p^{\alpha}}).\end{matrix}

Hence the min-max principle implies

λk(ΔLp)μk+𝒪(eεpα),\lambda_{k}(\Delta^{L^{p}})\leq\mu_{k}+\mathcal{O}(e^{-\varepsilon p^{\alpha}}),

which is the desired inequality. ∎

References

  • [1] J.P. Demailly “Champs magnétiques et inégalités de Morse pour la d"-cohomologie” In Annales de l’Institut Fourier 35-4, 1985
  • [2] S. Fournais and B. Helffer “Spectral methods in surface superconductivity” Progress in Nonlinear Differential Equationstheir Applications. Birkhäuser Boston Inc., Boston, MA, 2010
  • [3] B. Helffer and Y. Kordyukov “Semiclassical spectral asymptotics for a magnetic Schrödinger operator with non-vanishing magnetic field” In Geometric Methods in Physics: XXXII Workshop, Bialowieza, Trends in Mathematics. Birkhäuser, Basel, 2014
  • [4] B. Helffer and Y. Kordyukov “Semiclassical spectral asymptotics for a two-dimensional magnetic Schrödinger operator. II The case of degenerate wells” In Communications in Partial Differential Equations 37, 2012
  • [5] B. Helffer and Y. Kordyukov “Semiclassical spectral asymptotics for a two-dimensional magnetic Schrödinger operator: The case of discrete wells” In Spectral Theory and Geometric Analysis 535, 2011
  • [6] B. Helffer and A. Mohamed “Semiclassical Analysis for the Ground State Energy of a Schrödinger Operator with Magnetic Wells” In Journal of Functional Analysis 138, 1996
  • [7] B. Helffer, Y. Kordyukov, N. Raymond and S. Vũ Ngoc “Magnetic Wells in Dimension Three” In Anal. and PDE 9, 2016
  • [8] Y.A. Kordyukov “Semiclassical eigenvalue asymptotics for the Bochner Laplacian of a positive line bundle on a symplectic manifold” In arXiv: 1908.01756v1, 2019
  • [9] G. Marinescu and N. Savale “Bochner Laplacian and Bergman kernel expansion of semi-positive line bundles on a Riemann surface” In arXiv: 1811.00992, 2018
  • [10] L. Morin “A semiclassical Birkhoff normal form for constant-rank magnetic fields” In arXiv:2005.09386, 2020
  • [11] L. Morin “A Semiclassical Birkhoff Normal Form for Symplectic Magnetic Wells” In arXiv:1907.03493, 2019
  • [12] N. Raymond “Bound States of the Magnetic Schrödinger Operator” EMS Tracts in Mathematics, 2017
  • [13] N. Raymond and S. Vũ Ngoc “Geometry and Spectrum in 2D Magnetic wells” In Annales de l’Institut Fourier 65 (1), 2015