Eigenvalue asymptotics for weighted Laplace equations on rough Riemannian manifolds with boundary
EEIGENVALUE ASYMPTOTICS FOR WEIGHTED LAPLACEEQUATIONS ON ROUGH RIEMANNIAN MANIFOLDS WITHBOUNDARY
LASHI BANDARA, MEDET NURSULTANOV, AND JULIE ROWLETT
This work is dedicated to the memory of Alan McIntosh.
Abstract.
Our topological setting is a smooth compact manifold of dimensiontwo or higher with smooth boundary. Although this underlying topological struc-ture is smooth, the Riemannian metric tensor is only assumed to be bounded andmeasurable. This is known as a rough Riemannian manifold.
For a large classof boundary conditions we demonstrate a Weyl law for the asymptotics of theeigenvalues of the Laplacian associated to a rough metric. Moreover, we obtaineigenvalue asymptotics for weighted Laplace equations associated to a rough met-ric. Of particular novelty is that the weight function is not assumed to be of fixedsign, and thus the eigenvalues may be both positive and negative. Key ingredi-ents in the proofs were demonstrated by Birman and Solomjak nearly fifty yearsago in their seminal work on eigenvalue asymptotics. In addition to determiningthe eigenvalue asymptotics in the rough Riemannian manifold setting for weightedLaplace equations, we also wish to promote their achievements which may havefurther applications to modern problems.
Contents
1. Introduction 1Acknowledgements 72. The rough Riemannian manifold setting 73. Analytic preliminaries 124. Proof of the main results 165. Concluding Remarks 29References 301.
Introduction
Let M be a smooth, n -dimensional topological manifold with smooth boundary, ∂M ,such that the closure, M = M ∪ ∂M , is compact. If M is equipped with a smoothRiemannian metric, g , then there is a naturally associated Laplace operator, whichin local coordinates is∆ g = − (cid:112) det( g ) n (cid:88) i,j =1 ∂∂x i (cid:18) g ij (cid:112) det( g ) ∂∂x j · (cid:19) . (1)This is a second order elliptic operator with smooth coefficients, inherited from thesmoothness of the Riemannian metric. It is well known in this setting that theLaplacian, ∆ g , has a discrete, non-negative set of eigenvalues which accumulate Date : December 4, 2018.2010
Mathematics Subject Classification.
Primary 58J50, Secondary 58B20.
Key words and phrases.
Weyl asymptotics, rough metrics. a r X i v : . [ m a t h . DG ] D ec LASHI BANDARA, MEDET NURSULTANOV, AND JULIE ROWLETT only at ∞ . The set of all eigenvalues is known as the spectrum. Connectionsbetween the spectrum and the geometry of the underlying manifold have captivatedmathematicians and physicists alike for many years; see for example [53], [36], [43],[31],[32], [46], [40], and [41]. Whereas it is impossible in general to analyticallycompute the individual eigenvalues, in order to discover relationships between theeigenvalues and the geometry, one may study quantities determined by the spectrum.Any such quantity is known as a spectral invariant. The most fundamental spectralinvariants are determined by the rate at which the eigenvalues tend to infinity andwere discovered by Hermann Weyl in 1911 [53].Weyl proved that in the special case in which M = Ω is a smoothly bounded domainin R n , and the Dirichlet boundary condition is taken for the Euclidean Laplacian,then lim Λ →∞ N (Λ)Λ n/ = ω n Vol(Ω)(2 π ) n . (2)Above, N (Λ) is the number of eigenvalues of the Laplacian, counted with multiplic-ity, which do not exceed Λ, ω n is the volume of the unit ball in R n , and Vol(Ω) is thevolume of the domain, Ω. Hence, the rate at which the eigenvalues tend to infinitydetermines both the dimension, n , as well as the volume of Ω. These quantitiesare therefore spectral invariants. The asymptotic formula (2) is known as Weyl’sLaw.
Weyl’s law has both geometric generalisations, in which the underlying domainor manifold is no longer smooth; as well as analytic generalisations, in which theLaplace operator is replaced by a different, but typically Laplace-like operator.Here we simultaneously consider both a geometric generalisation as well as an an-alytic generalisation. We consider compact manifolds with a smooth differentiablestructure and allow the possibility that such manifolds also carry a smooth bound-ary. However, the Riemannian-like metric in our setting, known as a rough metric, is only assumed to be measurable, which is the primary novelty and a great sourceof difficulty in the analysis. Such a rough metric is only required to be boundedabove in an L ∞ sense, and essentially bounded below. A smooth topological mani-fold, M , equipped with a rough Riemannian metric, g , is henceforth dubbed a roughRiemannian manifold, with abbreviation RRM.Given that the coefficients of the metric tensor are merely measurable for a roughRiemannian manifold, the length functional over a curve is not well defined. There-fore, unlike for smooth or even continuous metrics, it is not possible to obtain alength structure via minimisation over curves. An alternative may be to considersupremums over the difference of certain classes of functions evaluated at two pointsin an attempt to obtain a distance between these points. In the smooth context,locally Lipschitz functions with gradient almost-everywhere bounded above recoversthe usual distance metric. In our context, it is not clear that this yields a reasonablenotion of distance. Rough metrics may have a dense set of singularities. Moreover,given that we allow for boundary further complicates matters. Although there isno canonical distance metric, we show that a rough metric does induce a canonicalRadon measure which allows for an L p theory of tensors. It is this fact, along withthe fact that the exterior derivative is purely determined by the differential topology, EYL’S LAW FOR WEIGHTED LAPLACE EQUATION ON RRM WITH BOUNDARY 3 that we will employ to see the rough metric as a measure space with a Dirichlet formthat we will use to define a Laplacian.The explicit study of these rough metrics arose with connections to the Kato squareroot problem (c.f. [6, 9, 14]), where these metrics can be seen as geometric invariancesof this problem (c.f. [12]). These objects have appeared implicitly in the past,particularly in the setting of real-variable harmonic analysis where the L ∞ topologyis a natural one. They are also a useful device when singular information can betransferred purely into the Riemannian metric. This happens when the singularobject is actually a differentiable manifold, and the singular information can bepurely seen as a lack of regularity of the Riemannian metric. A typical situation forthis is when the manifold is obtained as a limit of a Riemannian manifolds in theGromov-Hausdorff sense.In the manifold context, rough metrics were treated in [49] by Saloff-Coste in connec-tion with Harnack estimates. Norris studied Lipschitz manifolds [45], where roughmetrics are the natural replacement of smooth Riemannian metrics due to the reg-ularity of the differentiable structure. Higher regularity versions of rough metrics,namely C metrics, were used by Simon in [50] to study the Ricci flow with initialdata given by a C metric. Burtscher [20] also used these higher regularity versionsof rough metrics to study length structures, since for these metrics length structuresexist as they do in the smooth context.One of the main reasons to study general rough metrics with only bounded, mea-surable coefficients is that a pullback of a smooth metric by a lipeomorphism is onlyguaranteed to have such regularity. Such a transformation allows for objects withsingularities to be studied more simply. For example, a Euclidean box can be writtenas the Lipschitz graph over a sphere, and hence, a Euclidean box can be analysedas a rough metric on the sphere. In [13], rough metrics played a central role in theanalysis of the regularity properties of a geometric flow that is related to the Ricciflow in the context of optimal transport.Although rough metrics arise in a variety of contexts and have been studied byseveral authors, Weyl’s law has remained unknown in this context. Indeed, due tothe highly singular nature and lack of a distance metric, one could expect results inthe spirit of those for sub-Riemannian manifolds. Although one may define a sub-Riemannian Laplacian which has discrete spectrum, spectral asymptotics are stilla largely open question [15], [39]. Under certain assumptions one does, however,have a Weyl asymptotic [26]. Yet, in the same work, it is shown that in generalthere may be only a local Weyl asymptotic which varies from point to point; thereis no single asymptotic rate at which the eigenvalues tend to infinity, thus no Weyllaw for the eigenvalues of the Laplacian. Given the lack of smoothness and thelack of a distance-metric structure in the rough Riemannian manifold context, it isnot immediately clear whether or not one would expect a Weyl asymptotic for theeigenvalues of the Laplacian. However, a very crude indication that this may be thecase can be seen from the fact that the Kato square root problem can be solved fora rough metric on any closed manifold (c.f. [11]). LASHI BANDARA, MEDET NURSULTANOV, AND JULIE ROWLETT
The Laplace operator for a rough metric in our analysis is applicable to a wide classof boundary conditions, which we call admissible boundary conditions.
The precisedefinition of admissible boundary condition and examples are given in §
3. In essence,we begin with a closed subspace W of the Sobolev space H ( M ) containing H ( M ).Then, we define a Dirichlet form on the subspace W , which in turn gives rise to anassociated Laplace operator, denoted ∆ g, W . We not only demonstrate Weyl’s lawfor such a Laplace operator, but we also demonstrate a Weyl law for a weightedLaplace equation.To state our main result in full generality, let M be a smooth compact manifold ofdimension n ≥ g be a rough metric on M (seeDefinition 2.1). Let β > n , and ρ ∈ L β ( M, dµ g ) be a real-valued function such that ˆ M ρ dµ g (cid:54) = 0 . For an admissible boundary condition, W , we consider the form E g, W [ u, v ] = ( ∇ u, ∇ v ) L ( M, dµ g ) defined on the subspace, Z ( ρ ) = W if E g, W generates the norm in W , which is equivalent to the standard H norm , (cid:8) u ∈ W : ´ M ρu dµ g = 0 (cid:9) otherwise . We consider the eigenvalue problem ˆ M ρuv dµ g = λ E g, W [ u, v ] , u, v ∈ Z ( ρ ) , (3)Let us see that this problem is natural. If we assume for a moment that ρ = 1, and W = H ( M ), then u ∈ Z ( ρ ) if and only if u is orthogonal to the constant functions inthe L sense. On the other hand, a constant function is an eigenfunction of Laplaceoperator with the Neumann boundary condition corresponding to the eigenvaluezero. Since the eigenfunctions of a self-adjoint operator are orthogonal, it followsthat the eigenvalues of the Laplace operator on Z ( ρ ) are the non-zero eigenvaluesof Laplace operator with Neumann boundary condition. In case W = H ( M ), Z ( ρ ) = H ( M ). Moreover, the non-zero eigenvalues, λ , of (3) are in bijection withthe eigenvalues Λ of the weighted Laplace equation∆ g, W u = Λ ρu (4)In this way, we refer to the eigenvalues of (3) as eigenvalues of a weighted Laplaceequation.This type of equation arises in the study of hydrodynamics and elasticity, specificallyin the linearisation of certain nonlinear problems; see [2] and references therein.Further motivation comes from quantum mechanics and the study of the behaviour of EYL’S LAW FOR WEIGHTED LAPLACE EQUATION ON RRM WITH BOUNDARY 5 eigenvalues for Schr¨odinger operators with a large parameter; see [48]. The equation(∆ − qV ( x )) u = λu, q → ∞ reduces to the study of the spectral problem∆ u = λV ( x ) u. Above, V is the electric potential, for which there are no reasons to assume that thesign is constant. Consequently, it is quite interesting to study (4) in the generalityconsidered here, where we only assume the weight function ρ ∈ L β for some β > n ,but we do not assume that ρ is of constant sign. Investigation of this type of problemincludes, but is not limited to [18], [19], [34]. This equation is not only interesting andrelevant to physics but also has applications in biology such as modelling populationgenetics; see [28].The main result of this paper is Theorem 1.1 ( Weyl asymptotics for weighted Laplace equation with ad-missible boundary conditions ) . Let M be a smooth compact manifold of dimen-sion ≥ with smooth boundary, and let g be a rough metric on M . Then, theeigenvalues of (3) are discrete with finite dimensional eigenspaces with positive andnegative eigenvalues, {− λ − j ( W ); λ + j ( W ) } ∞ j =1 , such that − λ − ( W ) ≤ − λ − ( W ) ≤ . . . < < . . . ≤ λ +2 ( W ) ≤ λ +1 ( W ) . Moreover, they satisfy the Weyl asymptotic formula lim k →∞ λ ± k ( W ) k n = (cid:18) ω n (2 π ) n (cid:19) n (cid:18) ˆ M ± | ρ ( x ) | n dµ g (cid:19) n = (cid:18) ω n (2 π ) n (cid:19) n (cid:107) ρ (cid:107) L n ( M + , dµ g ) . Above, M ± := { x ∈ M : ± ρ ( x ) > } . As a corollary, we obtain classical Weyl asymptotics for the unweighted Laplaceeigenvalue problem.
Corollary 1.2 (Classical Weyl asymptotics) . Let M be a smooth compact manifoldof dimension ≥ with smooth boundary, and let g be a rough metric on M . Then,the Laplacian ∆ g, W associated to an admissible boundary condition W has discretespectrum with finite dimensional eigenspaces, and lim λ →∞ N ( λ, ∆ g, W ) λ n = ω n (2 π ) n Vol(
M, g ) . Above N ( λ, ∆ g, W ) is the number of eigenvalues of ∆ g, W less than λ . Weyl’s law in singular geometric settings.
Weyl’s law has been previouslydemonstrated in many singular geometric settings. Perhaps the most robust methodfor obtaining Weyl’s law in these settings is the so-called heat kernel or semi-groupmethod. This method can be used to obtain Weyl’s law on the manifolds with conicalsingularities studied by Cheeger [22]. In that case, the Laplacian has terms withcoefficients r − with r tending to 0 at the conical singularity. The heat kernel methodcan also be used to obtain Weyl’s law for non-smooth spaces which arise as thelimits of smooth, compact Riemannian manifolds. Any sequence of smooth, compact LASHI BANDARA, MEDET NURSULTANOV, AND JULIE ROWLETT
Riemannian manifolds with Ricci curvature bounded below has a subsequence whichconverges in the pointed Gromov-Hausdorff sense to a limit space. These limit spaceswere studied in 1997–2000 by Cheeger and Colding [23–25]. In the non-collapsedcase, they were able to define a Laplace operator on the limit space and obtaindiscreteness of its spectrum.In 2002, Ding [27] used heat kernel techniques to obtain Weyl’s law for the non-collapsed limit spaces studied by Cheeger and Colding. Ding showed that the sin-gular limit space has a well-defined heat kernel. Relating this heat kernel to thosefor the smooth spaces, he could extract the Weyl law for the singular limit space.More recently, Weyl’s law has been studied in the context of metric spaces satisfyingthe Riemannian Curvature Dimension (RCD) condition [54]. Since it is impracticalto provide an exhaustive list of references, we point the reader to the survey article[35] by Ivrii and references therein, which provides an overview of RCD spaces andtheir development in a time-linear narrative. The method used to prove Weyl’s lawin [54] is also through the short time asymptotic behaviour of the trace of the heatkernel. There is also a probabilistic approach via heat kernels; see [3] by Ambrosio,Honda and Tewodrose where this method is described in the setting of RCD spaces.Whereas the semi-group method can be used to demonstrate Weyl’s law in bothsmooth as well as many singular settings, it is inaccessible in the rough metricsetting. To see this, recall that a key step in this method is to compare the heatkernel H ( t, x, y ) to the function (4 πt ) − n/ e − d ( x,y )24 t , for small times, t , for points x and y which are sufficiently close. Above d ( x, y ) isthe distance between the points x and y on the underlying space. No such functioncan be defined without a well-defined notion of distance between points; hence thismethod is not available in the rough Riemannian manifold setting. Our space is notobtained as a limit of smooth objects, so the approach of Ding [27] in the contextof the limit spaces studied by Cheeger & Colding [23–25] is also not available.1.2. Strategy and structure of the paper.
Since heat kernel methods are un-available, as is any method which requires a well-defined notion of distance betweenpoints, we focus on abstract approaches rooted in functional analysis. We are in-spired by the work of the Soviet mathematicians, Birman and Solomjak [17], whomade a fundamental contribution to the study of the eigenvalue asymptotics forelliptic operators with non-smooth coefficients nearly fifty years ago. In the presentpaper, which can be considered as a tribute to their achievements and a popularisa-tion of their results, we demonstrate, for the particular case of second-order operatorsin divergence form, how their results can be carried over from the original settingfor a domain in a Euclidean space to rather general compact manifolds with roughmetrics. Birman and Solomjak [17] obtained the principal terms of the asymptoticbehaviour of the Dirichlet and Neumann problems for the equation B u = λ A u ,where A is a self-adjoint elliptic operator, and B is a self-adjoint operator of lowerorder. In particular, they considered a generalised Dirichlet eigenvalue problem of EYL’S LAW FOR WEIGHTED LAPLACE EQUATION ON RRM WITH BOUNDARY 7 the form − div A ∇ u = λBu inside bounded Euclidean domains Ω ⊂ R n . Above, A ( x ) is a positive matrix foralmost every x ∈ Ω such that A − ∈ L α (Ω), A ∈ L κ (Ω), and B ∈ L β (Ω) is a functionwhich is in general not of constant sign. Under the conditions α − + β − < n − , α > n, α − + κ − < n − , they obtain the asymptotic behaviour of both positive and negative eigenvalues interms of the L β norm of the function B . They mentioned, in [17], that this resultstill hold for the case α − + β − = 2 n − and n >
2, for the complete proof see [47].Although their work is an invaluable technical tool, due to the different geometricsetting and boundary conditions we consider, several additional results must bedemonstrated to be able to apply [17].This work is organised as follows. In § §
3, we show how a Laplace operator associated to a rough metric may be defined, andwe introduce the admissible boundary conditions together with examples thereof.Moreover, we demonstrate variational principles in the spirit of Courant, Rayleigh,and Poincar´e for general eigenvalue problems like those considered here. Our mainresults are proven in §
4. Concluding remarks are offered in § Acknowledgements
The first author was supported by the Knut and Alice Wallenberg foundation, KAW2013.0322 postdoctoral program in Mathematics for researchers from outside Swe-den, and by SPP2026 from the German Research Foundation (DFG). The secondauthor was partially supported by the Ministry of Education Science of the Repub-lic of Kazakhstan under the grant AP05132071. These authors also acknowledgethe gracious support of the organisers of the event “Harmonic Analysis of Ellipticand Parabolic Partial Differential Equations” at CIRM Luminy as well as the latterorganisation. 2.
The rough Riemannian manifold setting
Throughout, we fix M to denote a compact manifold of dimension equal to or exceed-ing 2 with a smooth differentiable structure. If the manifold has nonempty boundary ∂M , then we assume that it is smooth. We let T x M and T ∗ x M be the tangent andcotangent spaces at x , respectively, and T M and T ∗ M be the corresponding associ-ated bundles. The tensor bundles of covariant rank q and contravariant rank p arethen denoted by T ( p,q ) M = ( ⊕ pj =0 T ∗ M ) ⊕ ( ⊕ qk =0 T M ).In addition to a differentiable structure, such a space affords us with a notion ofmeasurability independent of a Riemannian metric: we say that a set A is measur-able if for every chart ( U, ψ ) with U ∩ A (cid:54) = ∅ , we have that ψ ( A ∩ U ) is Lebesgue LASHI BANDARA, MEDET NURSULTANOV, AND JULIE ROWLETT measurable in R n . We shall use L to denote the Lebesgue measure in R n . Propo-sition 1 in [12] shows that this notion of measurability is equivalent to asking for A to be µ h -measurable, where h is any smooth Riemannian metric on M , and µ h is itsinduced volume measure. With this, we obtain a notion of a measurable section ofa ( p, q ) tensor. The set in which these objects live will be denoted by Γ( T ( p,q ) M ).Similarly, we can define a measurable set Z ⊂ M to be of zero measure if for everychart ( U, ψ ), when U ∩ Z (cid:54) = ∅ , we have that ψ ( U ∩ Z ) has zero Lebesgue measure.This yields a notion of almost-everywhere in M without alluding to a measure. Itis straightforward to verify that if Z is of zero measure, then µ h ( Z ) = 0 for anysmooth metric h . Similarly, a property P holds almost-everywhere precisely when P holds µ h almost-everywhere for any smooth metric h .We can now present the precise notion of a rough metric . Definition 2.1 (Rough metric) . We say that a symmetric (2 ,
0) measurable tensor-field g is a rough metric if it satisfies the following local comparability condition : foreach x ∈ M , there exists a chart ( U x , ψ x ) containing x and a constant C ( U x ) ≥ C ( U x ) − | u | ψ ∗ x δ ( y ) ≤ | u | g ( y ) ≤ C ( U x ) | u | ψ ∗ x δ ( y ) for almost-every y ∈ U x , for all u ∈ T y M . Above, ψ ∗ x δ is the pullback to U x of the R n scalar product inside ψ ( U x ). Remark 2.2.
As a consequence of the compactness of M , we note that the compat-ibility condition is equivalent to demanding that there exists a smooth Riemannianmetric, h , on M such that C ( U x ) − | u | h ≤ | u | g ≤ C ( U x ) | u | h for almost-every y ∈ U x , where U x , u , and C ( U x ) are as in Definition 2.1.Due to the regularity of the coefficients of a general rough metric g , it is unclearhow to associate a canonical distance structure to g . However, the expression (cid:112) det g ( x ) dψ ∗ x L , for almost-every x ∈ U x inside a compatible a chart ( U x , ψ x ), can readily be checkedto transform consistently under a change of coordinates. This yields a Radon mea-sure that is independent of coordinates, which we denote by µ g .2.1. Rough metrics in harmonic analysis.
These rough metrics which are afocus in this paper were observed by Bandara in [12] to be geometric invariances ofthe
Kato square root problem on manifolds without boundary. In a nutshell, thisproblem is to prove that D ( √− div B ∇ ) = H ( M ) for bounded, measurable, complex,non-symmetric, elliptic coefficient matrices, x (cid:55)→ B ( x ). In the case of M = R n , thisproblem resisted resolution for over forty years. It was finally settled by Auscher,Hofmann, Lacey, McIntosh, and Tchamitchian in [6]. The first-order formulationof the problem by Axelsson , Keith and McIntosh in [10] allowed the problem tobe considered in geometric settings. Their approach was to obtain a solution to Andreas Axelsson is the former name of Andreas Ros´en.
EYL’S LAW FOR WEIGHTED LAPLACE EQUATION ON RRM WITH BOUNDARY 9 this problem by showing that an associated operator Π B , which in part consists ofthe original operator in question, has an H ∞ functional calculus. In this paper, theauthors obtained a solution of this problem on compact manifolds without boundary.In the non-compact setting, Morris in [44] first obtained results in this direction forEuclidean submanifolds under second fundamental form bounds. Later, Bandaraand McIntosh considered the intrinsic picture in [14] and demonstrated that thisproblem can be solved under appropriate lower bounds on injectivity radius as wellas Ricci curvature bounds.A fundamental question of McIntosh was to understand the limitations of the meth-ods used in this geometric version of the problem. Exploiting the stability of theproblem under L ∞ perturbations, Bandara in [12] showed that the problem couldbe solved far more widely than the previously used tools appeared to allow. Heshowed that if the problem can be solved for some Riemannian metric h , then italso admits a solution for any rough metric g which is L ∞ -close to h . In this sense,rough metrics naturally emerged as geometric invariances of the Kato square rootproblem. Indeed, the rough metrics as we have defined them here were introducedand investigated in [14] and [12] as geometric invariances of the Kato square rootproblem seen through the functional calculus of its first-order characterisation.It was shown in both [10] by Axelsson (Ros´en), Keith and McIntosh and later in[11] by Bandara, that on smooth boundaryless compact manifolds, the Kato squareroot problem has a positive solution. Counterexamples were first demonstrated byMcIntosh in [42] and later adapted by Auscher in [7]. These counterexamples reliedon having an operator whose spectrum grows exponentially. Since the Kato squareroot problem can be solved in the boundaryless case for rough metrics, one mayconjecture that the Laplacian associated to a rough metric ought to satisfy Weylasymptotics. Indeed, this was a key observation that prompted our investigation ofthe spectrum of the Laplacian on rough Riemannian manifolds with boundary.2.2. Geometric examples.
There are many natural examples of rough metrics,and here, we present here a small motivating collection. It is readily checked thatevery smooth or even continuous metric is rough. In particular, a metric of theform g = ψ ∗ h is rough whenever ψ : ( M, h ) → ( M, h ) is a lipeomorphism, and h is smooth. Recall that a lipeomorphism is a homeomorphism that is also locallyLipschitz with locally Lipschitz inverse. Example 2.3 (Rough metrics arising from Lipschitz graphs) . Let M be a smoothcompact manifold with smooth boundary, and let h be a smooth metric on M . Let N be some other smooth compact manifold with smooth boundary with metric h (cid:48) .Fix a Lipschitz function f : M → N . Note that Φ f : M → graph( f ) ⊂ M × N givenby Φ f ( x ) = ( x, f ( x )) is a lipeomorphism to its image. Moreover, g f ( u, v ) = ( d Φ f ( x ) u, d Φ f ( x ) v ) h ⊗ h (cid:48) , defines a metric tensor on M . Given the regularity of f , we have that Φ f is alipeomorphism. To see that g f is a rough metric, fix x ∈ M , and let ψ x : U x → B (2 ,
0) be a chart.Letting σ be a curve inside U x , we obtain that | σ (cid:48) ( t ) | g f = | σ (cid:48) ( t ) | h + | df ( σ (cid:48) ( t )) | h (cid:48) . Now, | df ( σ (cid:48) ( t )) | h (cid:48) ≤ sup t ∈ [0 , | Lip f ( σ ( t )) || σ (cid:48) ( t ) | h ≤ sup y ∈ U x | Lip f ( y ) || σ (cid:48) ( t ) | h where | Lip f ( y ) | = lim sup z → y d h (cid:48) ( f ( y ) , f ( z )) d h ( y, z ) . Since f is a Lipeomorphism, this supremum is finite. Thus df is defined for almostevery t , and it is a linear map between T x M → T f ( x ) N . Therefore, we have that for µ g almost every y ∈ U x and u ∈ T y M , setting C ( U x ) = max (cid:8) sup U x | Lip f | , (cid:9) C ( U x ) − | u | h ≤ | u | g f ≤ C ( U x ) | u | h . Consequently, g f is indeed a rough metric on M .As a concrete example, let M = B (1 ,
0) be the unit ball with the Euclidean metricin R n , and N = R also with the standard Euclidean distance. Then, M × N = B (1 , × R ⊂ R n +1 , and the the metric tensor h ⊗ dt in this case is the usualEuclidean metric on R n +1 . If f : B (1 , → R a Lipschitz map, then g f definedas above is a rough metric on B (1 , M = B (1 , ⊂ R , given any set E which has Lebesguemeasure zero, there exists a Lipschitz function for which the singular set of thisLipschitz function, i.e. where it fails to be differentiable, contains the set E . Thisset, E , can be a dense subset of M = B (1 , B (1 , ⊂ R , there exist non-trivial and highly singular rough metrics.We remark that in general, we do not treat the case of Lipschitz boundary, but ourmethods apply to those manifolds with Lipschitz boundary which are lipeomorphicto a smooth manifold. There are many Lipschitz manifolds that are not lipeomorphicor even homeomorphic to a smooth manifold. In dimensions exceeding 4, there areLipschitz manifolds that do not admit a smooth structure. This is seen by combin-ing [38] by Kervaire, where he demonstrates the existence of topological manifoldswithout a smooth structure in dimensions exceeding 3, and [52] by Sullivan whoshows that every topological manifold can be made into a Lipschitz manifold fordimensions exceeding 4. Although we do not treat these cases, our methods may behelpful for understanding these settings in the future. Example 2.4 (Manifolds with geometric cones) . Let M be a smooth compact man-ifold with smooth boundary. Suppose that there are points x , . . . , x k ⊂ M andneighbourhoods ( ψ i , U i ) of x i with ψ i ( U i ) = B (1 , i : U i → R n +1 byΦ i ( x ) = ( ψ i ( x ) , cot( α i / − | ψ i ( x ) | R n )) , α i ∈ (0 , π ] . A metric g ∈ C ∞ ( M \ { x , . . . , x i } ) has geometric cones at x i if g ( x ) = Φ ∗ i ( x )( · , · ) R n +1 inside ψ − i ( B (1 , α i . That is, there existsa chart near x i such that the metric takes the form g = dr + sin ( α i ) r dy . To see EYL’S LAW FOR WEIGHTED LAPLACE EQUATION ON RRM WITH BOUNDARY 11
Figure 1.
The witch’s hat sphere metric.this, fix such a point x i with associated α i ∈ (0 , π ]. Let ( r, y ) ∈ (0 , × S n − bepolar coordinates in B (1 , γ : I → B (1 ,
0) with γ (0) (cid:54) = 0, and let ˜ γ = ψ − i ( γ ). Define σ ( t ) := (Φ i ◦ ˜ γ )( t ) = ( γ ( t ) , cot( α i / − γ r ( t ))) , where γ r ( t ) = | γ ( t ) | R n . Therefore, σ (cid:48) ( t ) = ( γ (cid:48) ( t ) , − cot( α i / γ (cid:48) r ( t )) . In polar coordinates, γ ( t ) = γ r ( t ) γ y ( t ), with | γ y ( t ) | = 1. We therefore compute that | ˜ γ (cid:48) (0) | g = ( σ (cid:48) (0) , σ (cid:48) (0)) R n +1 = γ r (0) γ (cid:48) y (0) + γ (cid:48) r (0) + cot ( α i / γ (cid:48) r (0) = γ r (0) γ (cid:48) y (0) + (1 + cot ( α i / γ (cid:48) r (0) = γ r (0) γ (cid:48) y (0) + csc ( α i / γ (cid:48) r (0) . This shows that inside this chart, g = csc ( α i / dr + r dy . A simple change tothe coordinate system ( r, y ) (cid:55)→ (˜ r, y ) given by ˜ r = csc( α i / r , shows that in thesecoordinates, g = d ˜ r + sin ( α i / r dy . The quintessential example in the situation with boundary is the standard cone ofangle π/
2, given by M = graph( f ) where f : B (1 , → R is given by f ( x ) = 1 − | x | .In the absence of boundary, the “witch’s hat sphere metric” on the n -sphere S n is a particular example of a manifold with a geometric cone which was consideredby Bandara, Lakzian and Munn in [13]. They studied a geometric flow tangentialto the Ricci flow in a suitable sense that was defined by Gigli and Mantegazza in[29]. An appealing feature of this flow is that it can be defined in many singulargeometric settings such as metric spaces satisfying the RCD criterion. The regularityproperties of this flow were not considered in [29], which motivated the study of theflow in [13] on the “witch’s hat sphere metric.” This metric is the standard n -spheremetric away from a neighbourhood of the north pole, at the north pole, there is ageometric cone singularity of angle π/ The term “witch’s hat” arises from the Australian vernacular for a traffic cone. Analytic preliminaries
Notation.
Throughout this paper, we assume the Einstein summation con-vention. That is, whenever a raised index appears against a lowered index, unlessspecified otherwise, we sum over that index. By S , we denote the cardinality ofa given set S . In our analysis, we often write a (cid:46) b to mean that a ≤ Cb , where C > C will either be explicitly specified orotherwise, clear from context. By a ≈ b we mean that a (cid:46) b and b (cid:46) a .3.2. Dirichlet Forms and operators.
Here we introduce some facts regardingclosed symmetric densely defined forms and the self-adjoint operators they generate.We let D ( · ) denote the domain of either an operator or a form.Let H be a separable Hilbert space with scalar product ( · , · ), and E be a closedsymmetric densely defined form in H such that E [ x, x ] (cid:38) , x ∈ D ( E ) . Then, E generates a unique self-adjoint, non-negative operator T in H with domain D ( T ) ⊂ D ( E ) such that E [ x, y ] = ( T x, y )for all x ∈ D ( T ) and y ∈ D ( E ). Moreover, D (cid:0) T / (cid:1) = D ( E ) and (cid:0) T / x, T / y (cid:1) = E [ x, y ]for all x , y ∈ D ( E ). If, additionally, the form E is strictly positive, i.e. E [ x, x ] (cid:38) (cid:107) x (cid:107) ,then T is a strictly positive operator. For a more detailed exposition of these results,see Theorem 2.1 and Theorem 2.32 in § a is a completely continuous symmetric form in H . Then itgenerates a unique completely continuous self-adjoint operator A in H such that a [ u, v ] = ( Au, v ) for u , v ∈ H ; see Section 2.2 in [51]. We note that a completelycontinuous operator in a Hilbert space is compact. Remark 3.1.
In this work we often investigate the eigenvalues of the problem a [ u, v ] = λ ( u, v ) in H , by which we mean the eigenvalues of the unique operator A such that a [ u, v ] =( Au, v ). Recall that every completely continuous self-adjoint operator is a compactself-adjoint operator and therefore has discrete spectrum accumulating at 0; seeSection 2.5 in [51].In the subsequent analysis, the following variational principles are indispensabletools. The first variational principle is the so-called min-max characterisation ofCourant; see [30, Theorem 6.1]. The second and third ones are called Poincar´e’sand Rayleigh’s variational principles, respectively. We expect they are known butwere unable to locate a proof in the generality required here, so we include theproofs.
EYL’S LAW FOR WEIGHTED LAPLACE EQUATION ON RRM WITH BOUNDARY 13
Theorem 3.1.
Let H be a Hilbert space with scalar product ( · , · ) . Assume that A is a completely continuous self-adjoint operator in H , with the positive eigenvalues { λ + j } ∞ j =1 and negative eigenvalues {− λ − j } ∞ j =1 such that λ ± k +1 ≤ λ ± k . Let { u ± j } ∞ j =1 bethe corresponding eigenfunctions. Then(i) Courant’s variational principle λ ± k = min L ⊂H , dim L ⊥ = k − max u ∈ L \{ } ± ( Au, u )( u, u ) . (ii) Poincar´e’s variational principles λ ± k = max V ⊂H , dim V = k min u ∈ V \{ } ± ( Au, u )( u, u ) . (iii) Rayleigh’s variational principles λ ± k = max (cid:26) ± ( Au, u )( u, u ) : u ∈ { u ± , ..., u ± k − } ⊥ (cid:27) . Proof.
As we mentioned above, we prove (ii) and (iii). We first derive (ii). Since { u ± j } ∞ j =1 are the eigenfunctions corresponding to eigenvalues {± λ ± j } ∞ j =1 , it followsmax V ⊂H , dim V = k min u ∈ V \{ } ± ( Au, u )( u, u ) ≥ min u ∈ span { u ± ,...,u ± k } ± ( Au, u )( u, u ) = λ ± k . (5)Let L be the space for which the right side of (i) is achieved. Then for any k − dimensional V ⊂ H , there exists f ∈ L ∩ V . Therefore λ ± k = max u ∈ L \{ } ± ( Au, u )( u, u ) ≥ ± ( Af, f )( f, f ) ≥ min u ∈ V \{ } ± ( Au, u )( u, u ) . Since this is true for all k − dimensional V ⊂ H , we conclude λ ± k ≥ max V ⊂H , dim V = k min u ∈ V \{ } ± ( Au, u )( u, u ) . This together with (5) proves (ii).Next, let us prove (iii). Let H − and H + be the spectral spaces corresponding tothe negative and positive spectrum of A . Fix u ∈ H ± ∩ { u ± , ..., u ± k − } ⊥ , and assume u = (cid:80) ∞ j =1 a j u ± j . Then a j = 0 for j < k and ± ( Au, u ) = ∞ (cid:88) j = k λ ± j a j (cid:107) u ± j (cid:107) ≤ λ ± k ∞ (cid:88) j = k a j (cid:107) u ± j (cid:107) = λ ± k (cid:107) u (cid:107) . Note that for u = u ± k the inequality above becomes an equality. Therefore λ ± k = max (cid:26) ± ( Au, u )( u, u ) : u ∈ H ± ∩ { u ± , ..., u ± k − } ⊥ (cid:27) . (6)Let f ∈ { u ± , ..., u ± k − } ⊥ and f ± be projections of f into H ± . Then we write( Af, f ) = ( Af − , f − ) + 2( Af + , f − ) + ( Af + , f + ) . Hence, since ± ( Af ± , f ± ) ≥ Af + , f − ) = 0, we derive ± ( Af, f ) ≤ ± ( Af ± , f ± ).Therefore the right hand side of (iii) does not exceed the right hand side of (6). Onthe other hand, the right side of (6) by its very definition does not exceed the rightside of (iii). (cid:3) Laplacian associated to admissible boundary conditions.
As aforemen-tioned, a rough metric has a canonically associated Radon measure, µ g , and so wemay define L k ( T ( p,q ) M, dµ g ) spaces in the usual way. The Sobolev spaces H k ( M )and H k ( M ) on a compact Riemannian manifold with boundary are independent ofthe metric. The central issue for us is to ensure that H ( M ) and H ( M ) agree withthe domains of the defining operators for us in our Dirichlet forms. For this, weneed the fact that ∇ = ∇ = d : C ∞ ∩ L ( M, dµ g ) → C ∞ ∩ L ( T ∗ M, dµ g ) is aclosable operator. This uses the fact that the exterior derivative d depends only onthe differential structure of M as well as the properties of the rough metric. See[12] for details. Armed with this fact, we assert that H ( M ) = D ( ∇ ). Moreover,we also consider ∇ c = ∇ : C ∞ c ( M ) → C ∞ c ( M ). Since D ( ∇ c ) ⊂ D ( ∇ ) , we ob-tain H ( M ) = D ( ∇ c ) = C ∞ c (cid:107)·(cid:107) H . In the situation that ∂M = ∅ , we obtain that H ( M ) = H ( M ).Recall that in the case of a smooth metric, the Laplacian obtained by the Dirichletforms E N [ u, v ] = ( ∇ u, ∇ v ) L ( T ∗ M, dµ g ) with D ( E N ) = H ( M ) and E D [ u, v ] = E N [ u, v ]but with domain D ( E D ) = H ( M ) respectively yield the Neumann and DirichletLaplacians (c.f. § § Definition 3.2 (Admissible boundary condition) . Let
W ⊂ H ( M ) be a closedsubspace of H ( M ) such that H ⊂ W . Then we call W an admissible boundarycondition.Define the Dirichlet form, E g, W : W × W → C associated to W by E g, W [ u, v ] = ( ∇ u, ∇ v ) L ( T ∗ M, dµ g ) . From the representation theorems, namely, Theorem 2.1 in § § g, W . It is a non-negative self-adjoint operator with domain D (∆ g, W ) ⊂ W and with D ( (cid:112) ∆ g, W ) = W .Defining ∇ W as the operator ∇ with domain W , we see that it is a closed operatorand hence obtain a densely-defined and closed adjoint ∇ ∗W ,g by Theorem 5.29 in § EYL’S LAW FOR WEIGHTED LAPLACE EQUATION ON RRM WITH BOUNDARY 15 of Chapter 3 in [37]. A routine operator theory argument yields∆ g, W = ∇ ∗ g, W ∇ W . Proposition 3.2.
The spectrum of ∆ g, W is discrete with no finite accumulationpoints and with each eigenspace being of finite dimension. Proof.
Since we have assumed that the boundary of M is smooth, we have by theRellich-Kondrachov theorem (c.f. Theorem 2.34 in [5]) that H ( M ) (cid:44) → L ( M, dµ g )compactly. Therefore, we can factor the resolvent ( i + ∆ g, W ) : L ( M, dµ g ) → L ( M, dµ g ) as( i + ∆ g, W ) : L ( M, dµ g ) → D (∆ g, W ) → D ( (cid:112) ∆ g, W ) = W → H ( M ) (cid:44) → L ( M, dµ g ) . Hence, we obtain that this is a compact map.A sufficient condition for a self-adjoint operator T to have discrete spectrum is for( ζ − T ) − to be a compact operator, for some ζ in the resolvent set. In this situation,one also has that the operator T has no finite accumulation points, and that eacheigenspace is finite dimensional; this follows from Theorem 6.29 in § § g, W is self-adjoint, and i is anelement of the resolvent set. (cid:3) Examples of boundary conditions.
We shall see that Dirichlet, Neumann,and mixed boundary conditions are all admissible.
Example 3.3 (Dirichlet and Neumann conditions) . As mentioned in § M with smooth boundary and a smooth metric g , W = H ( M ) corresponds to Dirichlet boundary condition, and W = H ( M ) is the Neu-mann counterpart. This is easily verified by Stokes’s theorem, and by using theexistence of a unit outer normal to the boundary. See § § Example 3.4 (Mixed boundary conditions) . Let (
M, g ) be a smooth compact man-ifold and assume also that g is smooth. Fix Σ ⊂ ∂M a closed subset of the boundary ∂M with nonempty interior. Then, define W = (cid:8) u ∈ H (Ω) : spt u | ∂M ⊂ Σ a.e. in ∂M (cid:9) . This is a closed subspace of H ( M, g ). To see this, let u n ∈ W converge to u ∈ H ( M ). Then, letting Σ c = ∂M \ Σ, (cid:107) u | ∂M (cid:107) L (Σ c ) ≤ (cid:107) u | ∂M − ( u n ) | ∂M (cid:107) L (Σ c ) + (cid:107) ( u n ) | ∂M (cid:107) L (Σ c ) ≤ (cid:107) u | ∂M − ( u n ) | ∂M (cid:107) L ( ∂M ) ≤ (cid:107) u − u n (cid:107) H ( M ) where the penultimate inequality follows from the fact that ( u n ) | ∂M = 0 almost-everywhere in Σ c , whereas the ultimate from the boundedness of the trace map u (cid:55)→ u | ∂M : H ( M ) → H ( ∂M ) (cid:44) → L ( ∂M ) . By letting n → ∞ , this shows that u | ∂M = 0 almost-everywhere in Σ c and hence u ∈ W . The Laplacian ∆ g, W then has mixed-boundary conditions, with Dirichletboundary conditions on ∂M \ Σ and Neumann on Σ.
When M = Ω ⊂ R n , a bounded domain, with g as the standard Euclidean metricwere considered by Axelsson (Ros´en), Keith and McIntosh in [8] to study the Katosquare root problem under mixed boundary conditions. As mentioned in § B has a H ∞ functional calculus. In thepresence of boundary, the operator Π B does not have a canonical domain. Thedomains considered are built from closed subspaces V ⊂ H (Ω) satisfying H (Ω) ⊂V . The two extremes, V = H (Ω) and V = H (Ω) correspond to the Dirichlet andNeumann conditions for the relevant part of the operator Π B respectively, wherein the functional calculus, this operator is accessed by simply taking a relevantfunction f and considering a new function z (cid:55)→ f ( z ). The conditions W which wehave defined above, in this context, are precisely the “mixed-boundary conditions”of [8]. 4. Proof of the main results
The statement of the problem.
For the convenience of the reader, we recallthe key notions from the introduction. Let M be a smooth compact manifold ofdimension n ≥ g be a rough metric on M . Let β > n , and ρ ∈ L β ( M, dµ g ) be a real valued function such that ˆ M ρ dµ g (cid:54) = 0 . We consider the Dirichlet form E g, W [ u, v ] = ( ∇ u, ∇ v ) L ( M, C n , dµ g ) . Let W be an admissible boundary condition. Associated to the Dirichlet form is asubspace of W , Z ( ρ ) = W if E g, W generates the norm in W , which is equivalent to the standard H ( M ) norm , (cid:8) u ∈ W : ´ M ρu dµ g = 0 (cid:9) otherwise . Note, as we mentioned in the introduction, for ρ = 1, Z ( ρ ) is the intersection of W and the closure of the operator ∆ g, W . Proposition 4.1.
The subspace Z ( ρ ) ⊂ W is closed in H ( M ) norm. Moreover,dim Z ( ρ ) ⊥ = τ ≤
1, where orthogonality is in the W sense. Proof.
Let us choose βnnβ − n +2 β < q < ≤ n , then nqn − q < ββ − . Therefore The SobolevEmbedding Theorem, see [33, Theorem 10.1], gives the continuous embeddings H ( M ) = W , ( M ) ⊂ W ,q ( M ) ⊂ L ββ − ( M ) ⊂ L ββ − ( M ) , (7) EYL’S LAW FOR WEIGHTED LAPLACE EQUATION ON RRM WITH BOUNDARY 17 where the second embedding is compact. Therefore the H¨older inequality gives (cid:12)(cid:12)(cid:12)(cid:12) ˆ M ρu dµ g (cid:12)(cid:12)(cid:12)(cid:12) (cid:46) (cid:107) ρ (cid:107) L β (cid:107) u (cid:107) L ββ − (cid:46) (cid:107) ρ (cid:107) L β (cid:107) u (cid:107) H for u ∈ H ( M ). The implicit constants depend only on the volume of M withrespect to dµ g . In the case Z ( ρ ) = W , obviously Z ( ρ ) is closed in W . So, let usassume we are in the second case. Let { f j } ∞ j =1 ⊂ Z ( ρ ) and f ∈ W such that f n → f in the H norm. Since ´ M ρf j dµ g = 0, we obtain (cid:12)(cid:12)(cid:12)(cid:12) ˆ M ρf dµ g (cid:12)(cid:12)(cid:12)(cid:12) = (cid:12)(cid:12)(cid:12)(cid:12) ˆ M ( ρf − ρf j ) dµ g (cid:12)(cid:12)(cid:12)(cid:12) (cid:46) (cid:107) ρ (cid:107) L β (cid:107) f − f j (cid:107) H . By letting j → ∞ , we conclude that ´ M ρf dµ g = 0, and hence f ∈ Z ( ρ ), so Z ( ρ ) isa closed subspace of W with respect to the H ( M ) norm.To prove the statement regarding the dimension of Z ( ρ ), we assume for the sake ofcontradiction that dim Z ( ρ ) ⊥ = τ >
1. Then there exists linearly independent v , v ∈ Z ( ρ ) ⊥ . Since Z ( ρ ) ⊥ is a subspace, any linear combination of v and v shouldalso be in Z ( ρ ) ⊥ . Let c j = ´ M ρv j dµ g for j = 1 ,
2. Note that c j (cid:54) = 0, so there exists a ∈ R such that ac − c = 0, which is equivalent to a ˆ M ρv dµ g − ˆ M ρv dµ g = 0 = ˆ M ρ ( av − v ) dµ g . Therefore av − v ∈ Z ( ρ ). However, since av − v is a linear combination of v , v ∈ Z ( ρ ) ⊥ , we have that av − v ∈ Z ( ρ ) ∩ Z ( ρ ) ⊥ . This means that av − v = 0which contradicts the linear independence of v and v . (cid:3) Remark 4.1 (Notational simplifications) . We may, for the sake of simplicity use L k to denote L k ( M, dµ g ). We shall do this when we are only working with respectto the measure, dµ g . In case we are working with different measures, we shall usethe more cumbersome notation to indicate the measure of integration. Lemma 4.2 (Poincar´e inequality) . There exists a constant
C > such that (cid:13)(cid:13)(cid:13)(cid:13) u − ´ M ρ dµ g ˆ M ρu dµ g (cid:13)(cid:13)(cid:13)(cid:13) L ≤ C (cid:107)∇ u (cid:107) L (8) holds for any u ∈ H ( M ) .Proof. Without lost of generality, assume that ´ M ρ dµ g = 1. Assume that (8) falsefor every C >
0. Then there exists a sequence of functions { u j } ∞ j =1 such that theleft side of (8) equals 1 for all j while the right hand side tends to zero as j → ∞ .Let h j = u j − ˆ M ρu j dµ g . Since (cid:107) h j (cid:107) L = 1, and (cid:107)∇ h j (cid:107) L = (cid:107)∇ u j (cid:107) L , the sequence { h j } ∞ j =1 is bounded in H ( M ). Therefore, since every bounded sequence in a Hilbert space has a weaklyconvergent subsequence, we may assume that there exists h ∈ H ( M ) such that h j → h in weakly in H ( M ). Since (cid:107)∇ h j (cid:107) L → j → ∞ , we obtain that ∇ h = 0in the distributional sense. Since M is connected, it follows that h is constant function. Since ´ M ρh j dµ g = 0 and the second embedding in (7) is compact, wederive (cid:12)(cid:12)(cid:12)(cid:12) ˆ M ρh dµ g (cid:12)(cid:12)(cid:12)(cid:12) = (cid:12)(cid:12)(cid:12)(cid:12) ˆ M ρh dµ g − ˆ M ρh j dµ g (cid:12)(cid:12)(cid:12)(cid:12) ≤ (cid:107) ρ (cid:107) L β (cid:107) h − h j (cid:107) L ββ − → j → ∞ . Since ´ M ρ dµ g (cid:54) = 0, and h is constant, h = 0. This contradicts || h || L = 1. (cid:3) Corollary 4.2.
The form E g, W [ · , · ] generates the norm in Z ( ρ ), which is equivalentto the standard H ( M ) norm. Proof.
The proof follows from Lemma 4.2 and Proposition 4.1. (cid:3)
By Proposition 4.2, Z ( ρ ), equipped with the norm E g, W [ · , · ], is a Hilbert space, whichwe denote by ( Z ( ρ ) , E g, W ).Let us consider the form ρ [ u, v ] = ˆ M ρuv dµ g in ( Z ( ρ ) , E g, W ) . In order to see that ρ [ · , · ] is well defined, recall( Z ( ρ ) , E g, W ) ⊂ H ( M ) ⊂ L ββ − ( M ) , (9)where the first embedding is continuous by Corollary 4.2 and the second embeddingis compact by (7). Hence, by (7), the H¨older inequality implies | ρ [ u, v ] | ≤ (cid:107) ρ (cid:107) L β (cid:107) u (cid:107) L ββ − (cid:107) v (cid:107) L ββ − , (10)for u , v ∈ Z ( ρ ), so that ρ [ · , · ] is well defined. Moreover, the next proposition holds. Proposition 4.3.
The form ρ [ · , · ] is a completely continuous form in the Hilbertspace ( Z ( ρ ) , E g, W ). Proof.
Let u j → u and v j → v weakly in ( Z ( ρ ) , E g, W ). Then, by (10), we estimate | ρ [ u, v ] − ρ [ u j , v j ] | = | ρ [ u, v − v j ] − ρ [ u j − u, v j ] | (cid:46) (cid:107) ρ (cid:107) L β (cid:107) u (cid:107) L ββ − (cid:107) v − v j (cid:107) L ββ − + (cid:107) ρ (cid:107) L β (cid:107) u j − u (cid:107) L ββ − (cid:107) v j (cid:107) L ββ − . Since (9), u j → u and v j → v strongly in L ββ − ( M ). Hence the last estimate impliesthat ρ [ · , · ] is a completely continuous form in ( Z ( ρ ) , E g, W ). (cid:3) Proposition 4.3 allows us to consider eigenvalues of the problem ρ [ u, v ] = λ E g, W [ u, v ] , u, v ∈ Z ( p ) , (11)in the sense of Remark 3.1. We henceforth denote the non-zero eigenvalues of eigen-value problem (11) by {− λ − j ( W ); λ + j ( W ) } ∞ j =1 , EYL’S LAW FOR WEIGHTED LAPLACE EQUATION ON RRM WITH BOUNDARY 19 such that − λ − ( W ) ≤ − λ − ( W ) ≤ ... < < ... ≤ λ +2 ( W ) ≤ λ +1 ( W ) . Our main task is to investigate the asymptotic behaviour of these eigenvalues. Webegin by recalling results obtained by Birman and Solomjak [17] for domains in R n .4.2. Eigenvalue asymptotics for the weighted Laplace equation on Eu-clidean domains.
Here we state the simplified version of Theorems 3.2 and 3.5 in[17], and modify them for our propose.Let Ω ⊂ R n be a domain with Lipschitz boundary. Assume that, for almost every x ∈ Ω, B ( x ) is a positive number such that B − , B ∈ L ∞ (Ω), and A ( x ) is a positive n × n matrix such that A − , A ∈ L ∞ (Ω). Let P ∈ L β (Ω), where β > n/ W is one of the spaces H (Ω) or H (Ω). Considerthe following forms, for t > a t [ u, v ] := ( A ∇ u, ∇ v ) L (Ω ,d L ) + t ( u, v ) L (Ω ,d L ) , D ( a t ) = W ,a B [ u, v ] := ( A ∇ u, ∇ v ) L (Ω ,d L ) + ( Bu, v ) L (Ω ,d L ) , D ( a B ) = W . Let ( W , a t ) and ( W , a B ) be the spaces of functions u ∈ W equipped with the norms a t [ · , · ] and a B [ · , · ] respectively. Since t >
0, and A ( x ) is positive for almost every x ∈ Ω, the norms a t [ · , · ] and a B [ · , · ] are equivalent to the standard norm in H .Therefore ( W , a t ), ( W , a B ) are the Hilbert spaces, and they are equal to W as a set.Consider the form p [ u, v ] = ˆ Ω P uv d L in W . By the same arguments we do in Proposition 4.3, this form, p [ · , · ], is a completelycontinuous symmetric form in both Hilbert spaces ( W , a t ) and ( W , a B ). Therefore,in the sense of Remark 3.1, the eigenvalue problems p [ u, v ] = λa t [ u, v ] , in W , (12) p [ u, v ] = λa B [ u, v ] , in W , (13)have the discrete spectrum, eigenvalues with finite multiplicity, and accumulating at0. Let us denote the non-zero eigenvalues of (12) and (13) by {− λ − j ( a t ) , λ + j ( a t ) } ∞ j =1 and {− λ − j ( a B ) , λ + j ( a B ) } ∞ j =1 , respectively, ordered such that − λ − ( a t ) ≤ − λ − ( a t ) ≤ ... < < ... ≤ λ +2 ( a t ) ≤ λ +1 ( a t ) , − λ − ( a B ) ≤ − λ − ( a B ) ≤ ... < < ... ≤ λ +2 ( a B ) ≤ λ +1 ( a B ) . Let N ± ( λ, a t ) and N ± ( λ, a B ) be the distribution functions of eigenvalues of problems(12) and (13) respectively, N ± ( λ, a t ) = (cid:8) λ ± j ( a t ) > λ (cid:9) , N ± ( λ, a B ) = (cid:8) λ ± j ( a B ) > λ (cid:9) . The eigenvalues above are counted according to multiplicity.The following theorem, in a more general form, was proved in [17]. We state herethe simple version which shall be an essential ingredient in the proof of our mainresult.
Theorem 4.4 (Theorem 3.2 and 3.5 in [17]) . We have the asymptotic formulaslim λ → λ (cid:0) N ± ( λ, a t ) (cid:1) n = (cid:32) ω n (2 π ) n ˆ Ω ± | P ( x ) | n (cid:112) det A ( x ) dx (cid:33) n , where Ω ± := { x ∈ Ω : ± P ( x ) ≥ } , and ω n is the volume of the unit ball in R n .We shall use the preceding result to obtain an asymptotic formula for N ± ( λ, a B ). Theorem 4.3.
We have the asymptotic formulas lim λ → λ (cid:0) N ± ( λ, a B ) (cid:1) n = (cid:32) ω n (2 π ) n ˆ Ω ± | P ( x ) | n (cid:112) det A ( x ) dx (cid:33) n , where Ω ± := { x ∈ Ω : ± P ( x ) ≥ } , and ω n is the volume of the unit ball in R n .Proof. Since Ω ⊂ R n is a bounded domain with piecewise smooth boundary, theSobolev Embedding Theorem gives that id : W (cid:44) → L (Ω , d L ) is compact. Therefore,the multiplication operator ( B − t ) : W → L (Ω , d L )can be factored as an operator W id (cid:44) → L (Ω , d L ) → L (Ω , d L ) which shows that itis compact. In particular, this guarantees that it is a completely continuous map.Letting E [ u, v ] = a B [ u, v ] − a t [ u, v ], and taking u j → u and v j → v weakly, E [ u, v ] − E [ u j , v j ] = (( B − t ) u, v ) L (Ω ,d L ) − (( B − t ) u j , v j ) L (Ω ,d L ) = (( B − t ) u, v − v j ) L (Ω ,d L ) − (( B − t )( u j − u ) , v j ) L (Ω ,d L ) . By an application of the Banach-Steinhaus theorem, we can deduce that (cid:107) v j (cid:107) (cid:46) | (( B − t )( u j − u ) , v j ) L (Ω ,d L ) | (cid:46) (cid:107) ( B − t )( u j − u ) (cid:107) , and this tends to zero by the complete continuity of ( B − t ). The remaining termtends to zero by the fact that v j → v weakly, and therefore, E is a completelycontinuous Dirichlet form on W . Finally, [17, Lemma 1.3] implies thatlim λ → λ (cid:0) N ± ( λ, a B ) (cid:1) n = lim λ → λ (cid:0) N ± ( λ, a t ) (cid:1) n and hence Theorem 4.4 implies the statement. (cid:3) An auxiliary problem.
We shall demonstrate a Weyl asymptotic formula fora Dirichlet form in the spirit of a t . This will then be used to obtain our main result.Let W be an admissible boundary condition. Consider the following Dirichlet form,for t > E g, W ,t [ u, v ] = E g, W [ u, v ] + t ( u, v ) L ( M, dµ g ) , in W . We are interested in the eigenvalues, ν , of the following problem ρ [ u, v ] = ν E g, W ,t [ u, v ] , in ( W , E g, W ,t ) , (14)where ρ [ · , · ] is the form defined in Section 4.1. Note that the norm, obtained by E g, W ,t [ · , · ] is equivalent to the standard norm in H ( M ). Therefore we can equip W EYL’S LAW FOR WEIGHTED LAPLACE EQUATION ON RRM WITH BOUNDARY 21 with the norm E g, W ,t [ · , · ], and derive the new Hilbert space ( W , E g, W ,t ). By the samearguments we do in Proposition 4.3, one can see that ρ [ · , · ] is a completely continuousform in the Hilbert space ( W , E g, W ,t ). Therefore, in the sense of Remark 3.1, theeigenvalue problem (14) has discrete spectrum. We denote its non-zero eigenvaluesby {− λ − j ( W , t ); λ + j ( W , t ) } ∞ j =1 , such that − λ − ( W , t ) ≤ − λ − ( W , t ) ≤ ... < < ... ≤ λ +2 ( W , t ) ≤ λ +1 ( W , t ) . The following lemma allows us to localise the problem.
Lemma 4.4.
There exists a finite collection of open sets { M j } and functions { Φ j } such that(i) ( M j , Φ j ) is a coordinate patch,(ii) ∂M j is piecewise smooth and Lipschitz,(iii) M = ∪ Kj =1 M j and µ g ( M \ ∪ Kj =1 M j ) = 0 .Proof. Every smooth manifold with boundary is smoothly triangulable (c.f. [21]),and so we take { M j } as the interior of the simplices in the triangulation. Thefiniteness of the { M j } simply follows from compactness. It is easy to see that ∂M j is piecewise smooth and Lipschitz. The measure condition follows simply from thefact that each { M j } is a simplex. (cid:3) For each k = 1 , ..., K , we define the forms E Dk [ u, v ] := ( ∇ u, ∇ v ) L ( M k , dµ g ) + t ( u, v ) L ( M k , dµ g ) , D ( E Dk ) = H ( M k ) , E Nk [ u, v ] := ( ∇ u, ∇ v ) L ( M k , dµ g ) + t ( u, v ) L ( M k , dµ g ) , D ( E Nk ) = H ( M k ) , and ρ k [ u, v ] := ˆ M k ρuv dµ g D ( ρ k ) = H ( M k ) . Note that the form ρ k [ · , · ] is a completely continuous symmetric form on H ( M k ),and its restriction on H ( M k ) is also a completely continuous symmetric form on H ( M k ). Therefore the eigenvalue problems ρ k [ u, v ] = λ E Dk [ u, v ] in H ( M k ) , (15) ρ k [ u, v ] = λ E Nk [ u, v ] in H ( M k ) , (16)are well defined; see Remark 3.1. We will investigate eigenvalues of the problemsabove by reducing them into Euclidean space. In order to do this let us introducethe following notions.Given T = T kj dx k ⊗ dx j with the matrix ( T kj ) being invertible, we define T ∗ = T kj ∂∂x k ⊗ ∂∂x j with T kj T kj = δ jk , where δ jk is the Kronecker delta.Let Ω k := Φ k ( M k ) and Φ k ( · , · ) be the pullback of the usual Euclidean inner productin Ω k . Fix a smooth metric, h , on M , as in Remark 2.2. Then, there exist G and H k such that g ( u, v ) = h ( Gu, v ) h ( u, v ) = Φ k ( H k u, v )for u , v ∈ L ( T ∗ M k , dµ g ) (c.f. Proposition 10 in [12]). Let θ k := √ det H k and γ := √ det G . We also set (cid:101) G k := G ◦ Φ − k , (cid:101) H k := H k ◦ Φ − k ˜ γ k := γ ◦ Φ − k , ˜ θ k := θ k ◦ Φ − k , ˜ ρ k := ρ ◦ Φ − k Finally, we define A k := (cid:101) H k ∗ (cid:101) G k ∗ ˜ θ k ˜ γ k , B k := t ˜ θ k ˜ γ k , P k := ˜ ρ k ˜ θ k ˜ γ k . Next we consider the following reduced forms˜ E Dk [ u, v ] := ( A k ∇ u, ∇ v ) L (Ω k ,d L ) + ( B k u, v ) L (Ω k ,d L ) , D ( E Dk ) = H (Ω k ) , ˜ E Nk [ u, v ] := ( A k ∇ u, ∇ v ) L (Ω k ,d L )) + ( B k u, v ) L (Ω k ,d L ) , D ( E Nk ) = H (Ω k ) , ˜ p k [ u, v ] := ˆ Ω k P k uvd L , and the corresponding eigenvalue problems˜ p k [ u, v ] = λ ˜ E Dk [ u, v ] in H (Ω k ) , (17)˜ p k [ u, v ] = λ ˜ E Nk [ u, v ] in H (Ω k ) . (18)According to Section 4.2, these problems have discrete spectrum. We denote theirnon-zero eigenvalues by (cid:8) − ν − k,j ; ν + k,j (cid:9) ∞ j =1 and (cid:8) − η − k,j ; η + k,j (cid:9) ∞ j =1 , respectively. Next weprove that these are also eigenvalues of the problems (15) and (16) respectively. Lemma 4.5.
The eigenvalues of problems (15) and (16) coincide with eigenvaluesof problems (17) and (18) respectively.Proof.
Let both u and v belong to H (Ω k ) or H (Ω k ). Then, by the definitions of g ∗ , h ∗ , G ∗ , Φ k ∗ , H k ∗ , γ , ˜ H k ∗ , ˜ θ k , ˜ G ∗ , ˜ γ k , ˜ u , and ˜ v ,( ∇ u, ∇ v ) L ( M k , dµ g ) = ˆ M k g ∗ ( ∇ u, ∇ v ) dµ g = ˆ M k h ∗ ( G ∗ ∇ u, ∇ v ) dµ g = ˆ M k h ∗ ( G ∗ γ ∇ u, ∇ v ) dµ h = ˆ M k Φ k ∗ ( H k ∗ G ∗ γ ∇ u, ∇ v ) dµ h = ˆ Ω k ( (cid:101) H k ∗ ˜ θ k (cid:101) G ∗ ˜ γ k ∇ ˜ u, ∇ ˜ v ) d L = ( A k ∇ ˜ u, ∇ ˜ v ) L (Ω k ,d L ) . We refer to these forms as reduced because the geometric setting has been reduced to a domainin R n . EYL’S LAW FOR WEIGHTED LAPLACE EQUATION ON RRM WITH BOUNDARY 23
Similarly, one can check that t ( u, v ) L ( M k , dµ g ) = ( t ˜ θ k ˜ γ k ˜ u, ˜ v ) L (Ω k ,d L ) = ( B k ˜ u, ˜ v ) L (Ω k ,d L ) ,ρ k [ u, v ] = ˆ M k ρuv dµ g = ˆ Ω k ˜ ρ k ˜ θ k ˜ γ k ˜ u ˜ vd L = ˜ p k [˜ u, ˜ v ] . Therefore, for λ ∈ C and I = D or I = N , we obtain ρ k [ u, v ] − λ E Ik [ u, v ] = ˜ p k [˜ u, ˜ v ] − λ ˜ E Ik [˜ u, ˜ v ] . (cid:3) By Lemma 4.5, problems (15), (16) have the same eigenvalues as problems (17),(18) respectively. From now on, by (cid:8) − ν − k,j ; ν + k,j (cid:9) ∞ j =1 and (cid:8) − η − k,j ; η + k,j (cid:9) ∞ j =1 , we referto the eigenvalues of the problems (15) and (16) respectively. The correspondingeigenfunctions we denote by (cid:8) φ − k,j ; φ + k,j (cid:9) ∞ j =1 and (cid:8) ψ − k,j ; ψ + k,j (cid:9) ∞ j =1 respectively. Proposition 4.5.
The eigenvalues of the problems (15) and (16) satisfy the followingasymptotic formulaslim j →∞ ν ± k,j j n = lim j →∞ η ± k,j j n = (cid:18) ω n (2 π ) n (cid:19) n (cid:32) ˆ M ± k | ρ | n dµ g (cid:33) n , where M ± k := { x ∈ M k : ± ρ ( x ) ≥ } . Proof.
Since (cid:8) − ν − k,j ; ν + k,j (cid:9) ∞ j =1 and (cid:8) − η − k,j ; η + k,j (cid:9) ∞ j =1 are also eigenvalues of problems(17), (18) respectively, Theorem 4.3 implieslim j →∞ ν ± k,j j n = lim j →∞ η ± k,j j n = (cid:18) ω n (2 π ) n (cid:19) n (cid:32) ˆ Ω ± k | P k ( x ) | n (cid:112) det A ( x ) dx (cid:33) n . (19)where Ω ± k := { x ∈ Ω : ± P k ( x ) ≥ } and ω n is the volume of a unit ball in R n .Since, for x ∈ Ω k , ˜ θ k ( x ) > γ k ( x ) >
0, we obtain Ω ± k = { x ∈ Ω : ± ˜ ρ k ( x ) ≥ } .Therefore ˆ Ω ± k | P k ( x ) | n (cid:112) det A ( x ) dx = ˆ Ω ± k | ˜ ρ k ( x )˜ θ k ( x )˜ γ k ( x ) | n (cid:113) det ˜ G k ∗ ( x ) ˜ H k ∗ ( x ) (cid:16) ˜ θ k ( x )˜ γ k ( x ) (cid:17) n dx = ˆ Ω ± k | ˜ ρ k ( x ) | n (cid:113) det ˜ G k ∗ ( x ) ˜ H k ∗ ( x ) dx = ˆ Ω ± k | ˜ ρ k ( x ) | n ˜ θ k ( x )˜ γ k ( x ) dx. Since ˜ ρ k = ρ ◦ Φ − k , ˜ ρ k ( x ) > ρ (cid:0) Φ − k ( x ) (cid:1) >
0, and hence Φ − k (Ω ± k ) = M ± k .Therefore the above equation gives ˆ Ω ± k | P k ( x ) | n (cid:112) det A ( x ) dx = ˆ Ω ± k | ˜ ρ k ( x ) | n ˜ θ k ( x )˜ γ k ( x ) dx = ˆ M ± k | ρ | n dµ g . Hence (19) implieslim j →∞ ν ± k,j j n = lim j →∞ η ± k,j j n = (cid:18) ω n (2 π ) n (cid:19) n (cid:32) ˆ M ± k | ρ | n dµ g (cid:33) n . (cid:3) Next we introduce the following notations. We set (cid:8) − ν − j ; ν + j (cid:9) ∞ j =1 := (cid:8) − ν − k,j ; ν + k,j (cid:9) ∞ k,j =1 (cid:8) − η − j ; η + j (cid:9) ∞ j =1 := (cid:8) − η − k,j ; η + k,j (cid:9) ∞ k,j =1 such that − ν − ≤ − ν − ≤ ... < < ...ν +2 ≤ ν +1 , − η − ≤ − η − ≤ ... < < ...η +2 ≤ η +1 . Therefore, every ± ν ± k is an eigenvalue of one of the problems (15), correspondingto some form E ± D,k ∈ {E Dj } Kj =1 and domain M ± D,k ∈ { M j } Kj =1 . Similarly, every ± η ± k is an eigenvalue of one of the problems (16) corresponding to some form E ± N,k ∈{E Nj } Kj =1 and domain M ± N,k ∈ { M j } Kj =1 . Let { φ ± k } ∞ k =1 and { ψ ± k } ∞ k =1 be eigenfunctionscorresponding to eigenvalues {± ν ± k } ∞ k =1 and {± η ± k } ∞ k =1 .eig.val eig.funct domain D. form k = 1 , ..., K j ∈ N ± ν ± k,j φ ± k,j M k E Dk k = 1 , ..., K j ∈ N ± η ± k,j ψ ± k,j M k E Nk j ∈ N ± ν ± j φ ± j M ± D,j E ± D,j j ∈ N ± η ± j ψ ± j M ± N,j E ± N,j j ∈ N ± λ ± j ( W , t ) f ± j M E g, W ,t j ∈ N ± λ ± j ( W ) M E g, W Figure 2.
The eigenvalues and functions, together with their respec-tive geometric domains and Dirichlet forms are organised in the abovetable.
Proposition 4.6.
We have the following estimate λ ± k ( W , t ) ≥ ν ± k , k ∈ N . (20) Proof.
For 1 ≤ j ≤ k , we extend the eigenfunctions φ + j to M \ M + D,j by zero. Notethat this extension is in H ( M ) since it is an eigenfunction for the Dirichlet problem.Since a system of ( k −
1) linear equations with k unknowns has a solution, we canfind α , ..., α k such that f := (cid:80) kj =1 α j φ + j ∈ H ( M ) with f (cid:54) = 0 satisfies E g, W ,t [ f, f + j ] = 0 . Moreover, f ∈ H ( M ) and so we also have that f ∈ W by our assumption on W .Recall that ( W , E g, W ,t ) is a Hilbert space with scalar product E g, W ,t [ · , · ]. Also recallthat { λ ± j ( W , t ) } ∞ j =1 are non-zero eigenvalues of the completely continuous operatorgenerated by the form ρ [ · , · ]. Let us denote this operator by B , so that D ( B ) = EYL’S LAW FOR WEIGHTED LAPLACE EQUATION ON RRM WITH BOUNDARY 25 ( W , E g, W ,t ), and E g, W ,t [ B u, v ] = ρ [ u, v ]. Then Theorem 3.1(iii), with A = B and H = ( W , E g, W ,t ), implies λ + k ( W , t ) E g, W ,t [ f, f ] ≥ E g, W ,t [ B f, f ] = ρ [ f, f ] = k (cid:88) j,l =1 ρ [ α j φ + j , α l φ + l ] . Next let us note that ρ [ φ + j , φ + l ] = 0 for j (cid:54) = l . Indeed, if their supports are disjointthen the claim is obviously true. If their supports intersect, then M + D,j = M + D,l = M i for some i = 1 ...K . This means that φ + j and φ + l are distinct eigenfunctions of the i -th problem of (15). Therefore ν + l E Di [ ρφ + j , φ + l ] = ρ i [ φ + j , φ + l ] = ν + j E Di [ φ + j , φ + l ] . This is possible only if ρ i [ φ + j , φ + l ] = 0, and therefore ρ [ φ + j , φ + l ] = 0. Hence the lastestimate implies λ + k ( W , t ) E g, W ,t [ f, f ] ≥ k (cid:88) j ρ [ α j φ + j , α j φ + j ] . (21)On the other hand k (cid:88) j =1 ρ [ α j φ + j , α j φ + j ] = k (cid:88) j =1 ν + j (cid:0) ( ∇ α j φ + j , ∇ α j φ + j ) L ( M, dµ g ) + t ( α j φ + j , α j φ + j ) L ( M, dµ g ) (cid:1) ≥ ν + k k (cid:88) j =1 (cid:0) ( ∇ α j φ + j , ∇ α j φ + j ) L ( M, dµ g ) + t ( α j φ + j , α j φ + j ) L ( M, dµ g ) (cid:1) = ν + k E g, W ,t [ f, f ]since ρ [ φ + j , φ + l ] = 0 for l (cid:54) = j . This, together with (21), implies (20). An analogousargument gives the result for the negative eigenvalues. (cid:3) Proposition 4.7.
We have the following estimate η ± k ≥ λ ± k ( W , t ) , (22)for sufficiently large k ∈ N . Proof.
As in the previous proposition, we can find f = (cid:80) kj =1 β j f + j such that E + N,j (cid:104) f | M + N,j , ψ + j (cid:105) = 0 , j = 1 , ..., k − . Let us fix 1 ≤ l ≤ K to fix a chart (Φ l , M l ). Next, we will prove the estimate η + k E Nl [ f | M l , f | M l ] ≥ ρ (cid:104) f | M l , f | M l (cid:105) . (23)Assume, for now, that there exists m ( l ) ∈ N such that η + l,m ( l )+1 ≤ η + k ≤ η + l,m ( l ) . (24)Recall that { η + l,j } ∞ j =1 and { ψ + l,j } ∞ j =1 are eigenvalues and eigenfunctions correspondingto the form E Nl on the domain M l . The last estimates imply { η + l,j } m ( l ) j =1 ⊂ { η + j } kj =1 .Therefore { ψ + l,j } m ( l ) j =1 ⊂ { ψ + j } kj =1 , and consequently, by from the construction of f , itfollows E Nl (cid:2) f | M l , ψ + l,j (cid:3) = 0 , j = 1 , ..., m ( l ) . (25) Since f + j ∈ W , we see that f ∈ W ⊂ H ( M ), and therefore f | M l ∈ H ( M ) = D ( E Nl ).Moreover, by (25), we see that f | M l ⊥ { ψ + l,j } m ( l ) j =1 in (cid:0) H ( M l ) , E Nl (cid:1) . Therefore, byTheorem 3.1(iii), we obtain η + k E Nl [ f | M l , f | M l ] ≥ η + l,m ( l )+1 E Nl [ f | M l , f | M l ] ≥ ρ (cid:104) f | M l , f | M l (cid:105) . Next, assume that there is no such m ( l ) as in (24). This is possible only if theeigenvalue problem (16), with number l , does not have positive eigenvalues. Thismeans that the right hand side of (23) is negative, so that (23) still holds.Summing (23) over 1 ≤ l ≤ K gives η + k E g, W ,t [ f, f ] ≥ ρ [ f, f ] . (26)Since ρ [ f + j , f + l ] = λ + j ( W , t ) E g, W ,t [ f + j , f + l ] , we conclude that ρ [ f + j , f + l ] = 0 for j (cid:54) = l . Therefore ρ [ f, f ] = k (cid:88) j =1 ρ [ β j f + j , β j f + j ] = k (cid:88) j =1 λ + j ( W , t ) E g, W ,t [ β j f + j , β j f + j ] ≥ λ + k ( W , t ) k (cid:88) j =1 E g, W ,t [ β j f + j , β j f + j ]= λ + k ( W , t ) E g, W ,t [ f, f ] . Comparing this with (26) we derive the statement. A similar argument proves theanalogous result for the negative eigenvalues. (cid:3)
Now we are ready to prove the main theorem of this subsection.
Theorem 4.6.
The eigenvalues of problem (14) satisfy the following asymptoticformula lim k →∞ λ ± k ( W , t ) k n = (cid:18) ω n (2 π ) n (cid:19) n (cid:18) ˆ M ± | ρ | n dµ g (cid:19) n , where M ± := { x ∈ M : ± ρ ( x ) > } .Proof. We note that for each k , by Proposition 4.5, the counting functions satisfylim λ → λ n/ { η ± k,j ≥ λ } = ω n (2 π ) n ˆ M ± k | ρ | n dµ g , lim λ → λ n/ { ν ± k,j ≥ λ } = ω n (2 π ) n ˆ M ± k | ρ | n dµ g . Consequently,lim λ → K (cid:88) k =1 λ n/ { η ± k,j ≥ λ } = ω n (2 π ) n K (cid:88) k =1 ˆ M ± k | ρ | n dµ g = ω n (2 π ) n ˆ M ± | ρ | n dµ g , EYL’S LAW FOR WEIGHTED LAPLACE EQUATION ON RRM WITH BOUNDARY 27 and similarly lim λ → K (cid:88) k =1 λ n/ { ν ± k,j ≥ λ } = ω n (2 π ) n ˆ M ± | ρ | n dµ g . Let N ± ( λ, E g, W ,t ), N ± ( λ, E Dk ), and N ± ( λ, E Nk ) be the counting functions of the eigen-values of problems (14), (15), and (16) respectively. By their very definitions, K (cid:88) k =1 { η ± k,j ≥ λ } = K (cid:88) k =1 N ± ( λ, E Nk ) , and similarly, K (cid:88) k =1 { ν ± k,j ≥ λ } = K (cid:88) k =1 N ± ( λ, E Dk ) . We therefore havelim λ → λ n K (cid:88) k =1 N ± ( λ, E Nk ) = lim λ → λ n K (cid:88) k =1 N ± ( λ, E Dk ) = ω n (2 π ) n ˆ M ± | ρ | n dµ g . By Propositions (4.6) and (4.7) K (cid:88) k =1 λ n N ± ( λ, E Dk ) ≤ λ n N ± ( λ, E g, W ,t ) ≤ K (cid:88) k =1 λ n N ± ( λ, E Nk ) . Thus, we obtain lim λ → λ n N ± ( λ, E g, W ,t ) = ω n (2 π ) n ˆ M ± | ρ | n dµ g . The statement of the theorem is an immediate consequence. (cid:3)
Eigenvalue asymptotics for the weighted Laplacian on a rough Rie-mannian manifold.
In this subsection we will prove our main result. We startwith the following lemma which allows us to derive the asymptotics of λ k ( W ) fromthose of λ k ( W , t ). We note that this lemma is an adaptation of [16, Lemma 2.1]. Lemma 4.7.
We have the following estimates λ ± k + τ ( W , t ) ≤ λ ± k ( W ) ≤ (1 − t ) − λ ± k ( W , Ct ) , < t < for some C > independent of t > . (Recall that τ = dim Z ( ρ ) ⊥ , and τ ≤ ).Proof. By Proposition 4.2, there exists
C > E g, W [ u, u ] ≥ C ( u, u ) L ( M, dµ g ) for all u ∈ Z ( ρ ). Therefore E g, W [ u, u ] ≥ (1 − t ) (cid:0) E g, W [ u, u ] + Ct ( u, u ) L ( M, dµ g ) (cid:1) , < t < . Therefore, by applying Theorem 3.1(ii) with H = ( Z ( ρ ) , E g, W ) and A being theoperator generated by the form ρ [ · , · ], we conclude λ + k ( W ) = max L ⊂ Z ( p ) , dim L = k min u ∈ L \{ } ρ [ u, u ] E g, W [ u, u ] ≤ max L ⊂ Z ( p ) , dim L = k min u ∈ L \{ } ρ [ u, u ](1 − t ) (cid:0) E g, W [ u, u ] + Ct ( u, u ) L ( M, dµ g ) (cid:1) ≤ max L ⊂W , dim L = k min u ∈ L \{ } ρ [ u, u ](1 − t ) (cid:0) E g, W [ u, u ] + Ct ( u, u ) L ( M, dµ g ) (cid:1) = (1 − t ) − λ + k ( W , Ct ) . In the last equation we again applied Theorem 3.1(ii), but with H = ( W , E g, W ,Ct )and A being the operator generated by the form ρ [ · , · ]. This proves the secondinequality.In the same way, but using Theorem 3.1(i), we derive λ + k + τ ( W , t ) = min L ⊂W , dim L ⊥ = k + τ − max u ∈ L \{ } ρ [ u, u ] (cid:0) E g, W [ u, u ] + t ( u, u ) L ( M, dµ g ) (cid:1) ≤ min L ⊂W , dim L ⊥ = k + τ − , ( Z ( p )) ⊥ ⊂ L ⊥ max u ∈ L \{ } ρ [ u, u ] (cid:0) E g, W [ u, u ] + t ( u, u ) L ( M, dµ g ) (cid:1) ≤ min L ⊂ Z ( ρ ) , dim L ⊥ = k − max u ∈ L \{ } ρ [ u, u ] (cid:0) E g, W [ u, u ] + t ( u, u ) L ( M, dµ g ) (cid:1) = λ + k ( W ) . This proves the first estimate. An analogous argument shows the same result forthe negative eigenvalues. (cid:3)
We are now poised to prove the main theorem. The statements concerning thediscreteness of the spectrum have already been proven, so it only remains to demon-strate
Theorem 4.8 ( Weyl’s law for a weighted Laplacian with an admissibleboundary condition ) . We have the following asymptotic formula lim k →∞ λ ± k ( W ) k n = (cid:18) ω n (2 π ) n (cid:19) n (cid:18) ˆ M ± | ρ ( x ) | n dµ g (cid:19) n = (cid:18) ω n (2 π ) n (cid:19) n (cid:107) ρ (cid:107) L n ( M ± , dµ g ) . Proof.
Let us multiply (27) by k n and take the limit as k → ∞ ,lim k →∞ λ ± k + τ ( W , t ) k n ≤ lim k →∞ λ ± k ( W ) k n ≤ (1 − t ) − lim k →∞ λ ± k ( W , Ct ) k n . We have already demonstrated thatlim k →∞ λ ± k ( W , Ct ) k n = (cid:18) ω n (2 π ) n (cid:19) n (cid:18) ˆ M ± | ρ ( x ) | n dµ g (cid:19) n . EYL’S LAW FOR WEIGHTED LAPLACE EQUATION ON RRM WITH BOUNDARY 29
Similarly, lim k →∞ λ ± k + τ ( W , t )( k + τ ) n = (cid:18) ω n (2 π ) n (cid:19) n (cid:18) ˆ M ± | ρ ( x ) | n dµ g (cid:19) n . Since τ ∈ { , } , we clearly have lim k →∞ k n ( k + τ ) n = 1 . Therefore, lim k →∞ λ ± k + τ ( W , t )( k ) n = (cid:18) ω n (2 π ) n (cid:19) n (cid:18) ˆ M ± | ρ ( x ) | n dµ g (cid:19) n . Thus, we derive (cid:18) ω n (2 π ) n (cid:19) n (cid:18) ˆ M ± | ρ ( x ) | n dµ g (cid:19) n ≤ lim k →∞ λ ± k ( W ) k k ≤ (1 − t ) − (cid:18) ω n (2 π ) n (cid:19) n (cid:18) ˆ M ± | ρ ( x ) | n dµ g (cid:19) n . Finally, by taking t →
0, we obtain the statement. (cid:3) Concluding Remarks
In this paper, we considered Laplacians induced by rough metrics g and weightedeigenvalue equations involving a weight function, ρ , which need not have constantsign. The manifold in question was also permitted to have boundary, and we wereable to consider a large class of admissible boundary conditions including mixedboundary conditions. There are a number of directions that further research forsuch problems could take.An immediate and interesting question is to determine estimates for the remainderterm in Weyl’s law. Since this contains curvature information in the smooth case, itwould be interesting to understand what this reveals about the structure of roughmetrics, and perhaps this would allow us to extract a weak notion of curvature orcurvature bounds for such objects.Beyond this question, we can consider this problem in more general settings. Onedirection would be to consider ( V, h ) → M , a Hermitian vector bundle with metric h over M equipped with a measure µ . Fixing a connection ∇ and a closed subspace W ⊂ H ( V, h ) with H ( V, g ) ⊂ W , we could consider the eigenvalue problem forthe divergence form equation b ∇ ∗W B ∇ W , where D ( ∇ W ) = W , b is a measurablefunction bounded below, and B is an elliptic, bounded, measurable endomorphismover V . The complication of this analysis is the fact that we can no longer localisethe problem by pulling it into R n , and we would have to devise a method by whichwe only use the trivialisations available to us from the bundle structure. If the measure µ were to not be induced by a rough metric, then we would also need tounderstand which classes of measures would be appropriate.A bundle in which we have the commutativity of the pullbacks with a differentialoperator are the differential forms Ω M , where we would be forced to consider theexterior derivative d instead of a connection ∇ . Fixing a rough metric g , we wouldobtain adjoints d ∗ g , and the Hodge-Dirac operator, D H = d + d ∗ g , would be an oper-ator of interest. This analysis would be plausible on a manifold without boundaryto obtain spectral asymptotics for the Hodge-Laplacian ∆ H = D , using similarmethods to those that we used here. However, the presence of boundary wouldcomplicate matters, since it is known that even in the setting when the metric issmooth, the operator ∆ H = D ∗ max D max where D max = ( D c ) ∗ with domain D ( D c ) = C ∞ c (Ω M ), admits an infinite dimen-sional kernel (c.f. Proposition 3.5 in [1]). The analysis would therefore require anunderstanding of the boundary conditions we impose on the boundary. We intendto investigate these problems in forthcoming work and regard the present paper asa solid foundation upon which to initiate a more general study.
References [1] Pierre Albin,
Analysis on non-compact manifolds notes for 18.158 . Unpublished, url: https://faculty.math.illinois.edu/~palbin/18158/18158May26.pdf .[2] Walter Allegretto,
Principal eigenvalues for indefinite-weight elliptic problems in R n , Proc.Amer. Math. Soc. (1992), no. 3, 701–706. MR1098396[3] Luigi Ambrosio, Shouhei Honda, and David Tewodrose, Short-time behavior of the heat kerneland Weyl’s law on
RCD ∗ ( K, N ) spaces. , Ann. Global Anal. Geom. (2018), no. 1, 97–119(English).[4] Wolfgang Arendt, Ralph Chill, Christian Seifert, Hendrik Vogt, and J¨urgen Voigt, Form meth-ods for evolution equations, and applications , 18th Internet Seminar on Evolution Equations,2015.[5] Thierry Aubin,
Nonlinear analysis on manifolds. Monge-Amp`ere equations , Grundlehren derMathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences], vol. 252,Springer-Verlag, New York, 1982. MR681859[6] Pascal Auscher, Steve Hofmann, Michael Lacey, Alan McIntosh, and Philippe. Tchamitchian,
The solution of the Kato square root problem for second order elliptic operators on R n , Ann.of Math. (2) (2002), no. 2, 633–654.[7] Pascal Auscher and Philippe Tchamitchian, Square root problem for divergence operators andrelated topics , Ast´erisque (1998), viii+172. MR1651262[8] Andreas Axelsson, Stephen Keith, and Alan McIntosh,
The Kato square root problem for mixedboundary value problems , J. London Math. Soc. (2) (2006), no. 1, 113–130.[9] , Quadratic estimates and functional calculi of perturbed Dirac operators , Invent. Math. (2006), no. 3, 455–497.[10] ,
Quadratic estimates and functional calculi of perturbed Dirac operators , Invent. Math. (2006), no. 3, 455–497. MR2207232 (2007k:58029)[11] Lashi Bandara,
Continuity of solutions to space-varying pointwise linear elliptic equations ,Publ. Mat. , 239–258.[12] , Rough metrics on manifolds and quadratic estimates , Math. Z. (2016), no. 3-4,1245–1281. MR3520003 This result was communicated privately to us by Matthias Ludewig, Batu G¨uneysu andFrancesco Bei.
EYL’S LAW FOR WEIGHTED LAPLACE EQUATION ON RRM WITH BOUNDARY 31 [13] ,
Geometric singularities and a flow tangent to the Ricci flow , Ann. Sc. Norm. Super.Pisa Cl. Sci. (5) (2017), no. 2, 763–804. MR3700383[14] Lashi Bandara and Alan McIntosh, The Kato Square Root Problem on Vector Bundles withGeneralised Bounded Geometry , J. Geom. Anal. (2016), no. 1, 428–462. MR3441522[15] G´erard Ben Arous, D´eveloppement asymptotique du noyau de la chaleur hypoelliptique sur ladiagonale , Ann. Inst. Fourier (Grenoble) (1989), no. 1, 73–99. MR1011978[16] Mikhail Sh. Birman and Mikhail Z. Solomyak, The principal term of the spectral asymp-totics for “non-smooth” elliptic problems , Funkcional. Anal. i Priloˇzen. (1970), no. 4, 1–13.MR0278126[17] , Spectral asymptotics of nonsmooth elliptic operators. I, II , Trudy Moskov. Mat. Obˇsˇc. (1972), 3–52; ibid. 28 (1973), 3–34. MR0364898[18] Kenneth J. Brown, Chris Cosner, and Jacqueline Fleckinger, Principal eigenvalues for problemswith indefinite weight function on R n , Proc. Amer. Math. Soc. (1990), no. 1, 147–155.MR1007489[19] Kenneth J. Brown and Achilles Tertikas, The existence of principal eigenvalues for problemswith indefinite weight function on R k , Proc. Roy. Soc. Edinburgh Sect. A (1993), no. 3,561–569. MR1226617[20] Annegret Y. Burtscher, Length structures on manifolds with continuous Riemannian metrics ,New York J. Math. (2015), 273–296. MR3358543[21] Stewart S. Cairns, A simple triangulation method for smooth manifolds , Bull. Amer. Math.Soc. (1961), 389–390. MR0149491[22] Jeff Cheeger, Spectral geometry of singular Riemannian spaces , J. Differential Geom. (1983),no. 4, 575–657 (1984). MR730920[23] Jeff Cheeger and Tobias H. Colding, On the structure of spaces with Ricci curvature boundedbelow. I , J. Differential Geom. (1997), no. 3, 406–480. MR1484888[24] , On the structure of spaces with Ricci curvature bounded below. II , J. DifferentialGeom. (2000), no. 1, 13–35. MR1815410[25] , On the structure of spaces with Ricci curvature bounded below. III , J. DifferentialGeom. (2000), no. 1, 37–74. MR1815411[26] Yves Colin de Verdi`ere, Luc Hillairet, and Emmanuel Tr´elat, Spectral asymptotics for sub-Riemannian Laplacians, I: Quantum ergodicity and quantum limits in the 3-dimensional con-tact case , Duke Math. J. (2018), no. 1, 109–174. MR3743700[27] Yu Ding,
Heat kernels and Green’s functions on limit spaces , Comm. Anal. Geom. (2002),no. 3, 475–514. MR1912256[28] Wendell H. Fleming, A selection-migration model in population genetics , J. Math. Biol. (1975), no. 3, 219–233. MR0403720[29] Nicola Gigli and Carlo Mantegazza, A flow tangent to the Ricci flow via heat kernels and masstransport , Adv. Math. (2014), 74–104. MR3122163[30] Valentyna I. Gorbachuk and Myroslav L. Gorbachuk,
Boundary value problems for operatordifferential equations , Mathematics and Its Applications( Soviet Series), vol. 48, Springer,Dordrecht, 1991.[31] Carolyn S. Gordon, David L. Webb, and Scott A. Wolpert,
Isospectral plane domains andsurfaces via Riemannian orbifolds , Invent. Math. (1992), no. 1, 1–22. MR1181812[32] Stephen W. Hawking,
Zeta function regularization of path integrals in curved spacetime ,Comm. Math. Phys. (1977), no. 2, 133–148. MR0524257[33] Emmanuel Hebey, Nonlinear analysis on manifolds: Sobolev spaces and inequalities , CourantLecture Notes in Mathematics, vol. 5, New York University, Courant Institute of MathematicalSciences, New York; American Mathematical Society, Providence, RI, 1999. MR1688256[34] Peter Hess and Tosio Kato,
On some linear and nonlinear eigenvalue problems with an in-definite weight function , Comm. Partial Differential Equations (1980), no. 10, 999–1030.MR588690[35] Victor Ivrii,
100 years of Weyl’s law , Bull. Math. Sci. (2016), no. 3, 379–452. MR3556544[36] Mark Kac, Can one hear the shape of a drum? , Amer. Math. Monthly (1966), no. 4, partII, 1–23. MR0201237[37] Tosio Kato, Perturbation theory for linear operators , Second, Springer-Verlag, Berlin, 1976.Grundlehren der Mathematischen Wissenschaften, Band 132. [38] Michel A. Kervaire,
A manifold which does not admit any differentiable structure , Comment.Math. Helv. (1960), 257–270. MR0139172 (25 D´eveloppement asymptotique de la densit´e d’une diffusion d´eg´en´er´ee , ForumMath. (1992), no. 1, 45–75. MR1142473[40] Zhiqin Lu and Julie M. Rowlett, The sound of symmetry , Amer. Math. Monthly (2015),no. 9, 815–835. MR3418203[41] ,
One can hear the corners of a drum , Bull. Lond. Math. Soc. (2016), no. 1, 85–93.MR3455751[42] Alan McIntosh, On the comparability of A / and A ∗ / , Proc. Amer. Math. Soc. (1972),430–434. MR0290169[43] Henry P. McKean and Isadore M. Singer, Curvature and the eigenvalues of the Laplacian , J.Differential Geometry (1967), no. 1, 43–69. MR0217739[44] Andrew J. Morris, The Kato square root problem on submanifolds , J. Lond. Math. Soc. (2) (2012), no. 3, 879–910. MR3000834[45] James R. Norris, Heat kernel asymptotics and the distance function in Lipschitz Riemannianmanifolds , Acta Math. (1997), no. 1, 79–103. MR1484769 (99d:58167)[46] Alexander M. Polyakov,
Quantum geometry of fermionic strings , Phys. Lett. B (1981),no. 3, 211–213. MR623210[47] Grigori V. Rozenblum,
Distribution of the discrete spectrum of singular differential operators ,Izv. Vysˇs. Uˇcebn. Zaved. Matematika (1976), 75–86. MR0430557[48] Grigori V. Rozenblum and Mikhail Z. Solomyak,
On principal eigenvalues for indefinite prob-lems in Euclidean space , Math. Nachr. (1998), 205–223. MR1626340[49] Laurent Saloff-Coste,
Uniformly elliptic operators on Riemannian manifolds , J. DifferentialGeom. (1992), no. 2, 417–450. MR1180389 (93m:58122)[50] Miles Simon, Deformation of C Riemannian metrics in the direction of their Ricci curvature ,Comm. Anal. Geom. (2002), no. 5, 1033–1074. MR1957662 (2003j:53107)[51] Friedrich Stummel, Rand- und Eigenwertaufgaben in Sobolewschen R¨aumen , Lecture Notes inMathematics, Vol. 102, Springer-Verlag, Berlin-New York, 1969. MR0463907[52] Dennis Sullivan,
Hyperbolic geometry and homeomorphisms , Geometric topology (Proc. Geor-gia Topology Conf., Athens, Ga., 1977), 1979, pp. 543–555. MR537749 (81m:57012)[53] Hermann Weyl,
Ueber die asymptotische verteilung der eigenwerte , Nachrichten von derGesellschaft der Wissenschaften zu G¨ottingen, Mathematisch-Physikalische Klasse (1911), 110–117.[54] Hui-Chun Zhang and Xi-Ping Zhu,
Weyl’s law on rcd*(k,n) metric measure spaces . Lashi Bandara, Institut f¨ur Mathematik, Universit¨at Potsdam, D-14476, PotsdamOT Golm, Germany
URL : E-mail address : [email protected] Medet Nursultanov, Mathematical Sciences, Chalmers University of Technologyand University of Gothenburg, SE-412 96, Gothenburg, Sweden
E-mail address : [email protected] Julie Rowlett, Mathematical Sciences, Chalmers University of Technology andUniversity of Gothenburg, SE-412 96, Gothenburg, Sweden
URL : E-mail address ::