HAL Id: hal-00022000
https://hal.archives-ouvertes.fr/hal-00022000
Submitted on 31 Mar 2006
HAL
is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire
HAL, estdestinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.
Bounded eigenfunctions in the real Hyperbolic space
Sandrine Grellier, Jean-Pierre Otal
To cite this version:
Sandrine Grellier, Jean-Pierre Otal. Bounded eigenfunctions in the real Hyperbolic space. interna-
tional mathematical research notices, 2005, 62, pp.3867-3897. �hal-00022000�
2005,No. x
Bounded Eigenfunctions in the Real Hyperbolic Space
Sandrine Grellier and Jean-Pierre Otal
1 Introduction
LetHn be the real hyperbolic space of dimensionn, that is,the complete and simply connected Riemannian manifold of constant curvature−1. Using the Poincar ´e model,we identifyHnwith the unit ball ofRnand its boundary at infinity∂Hnwith the unit sphere Sn−1ofRn.
We denote byΔthe Laplace-Beltrami operator acting on functions onHn.
Definition 1.1. Forλ ∈ C,say thatf : Hn → Cis a λ-eigenfunction whenf is C2and satisfiesΔf+λf=0. Denote byEλthe space of theλ-eigenfunctions which are bounded.
The purpose of this paper is to describe the spaceEλ. The Harnack inequality(see,e.g., [6,page 199])implies that the gradient of a boundedλ-eigenfunctionfis bounded by a constant which only depends onλand on the sup norm off. This implies,using the Ascoli theorem,that the spaceEλwith the sup norm is a Banach space.
Fors ∈ C,we setλs = s(n−1−s). The basic example of aλs-eigenfunction is provided byx→ Ps(x, ξ),whereP(x, ξ)is the Poisson kernel ofHnandξis any point of the boundary at infinity∂Hn ofHn. Anyλs-eigenfunctions can be described in terms of hyperfunctions on∂Hn. Namely, for anyλs-eigenfunction, there is a hyperfunction on
∂Hnsuch that
f(x)=Ps(T)(x)=
T, Ps(x,·)
; (1.1)
Received 25 March 2005. Revision received 26 October 2005.
Communicated by Peter Sarnak.
this is a theorem of Helgason and Minemura(cf.[7,9]),which was also proven later by Agmon[1].
Definition 1.2. Call the functionx → T, Ps(x,·)the Poisson-Helgason transform of T. Denote it byPs(T).
Furthermore,one knows that the hyperfunctionT is unique,except for the cases whens= −1,−2, . . . ,−k, . . .(see[9,Proposition 2.1]). Moreover,iffgrows slowly,that is, at most exponentially with the hyperbolic distance,thenT is a distribution. This result has been proven initially by ¯Oshima and Sekiguchi[10]using microlocal techniques. In [15],the authors give a new proof within the framework of asymptotic expansions(see [15,Theorem 2-2]). In particular,the hyperfunction which represents a bounded eigen- function is indeed a distribution. Thus,to describe the space of bounded eigenfunctions is equivalent to describe the distributionsT onSn−1such thatPs(T)is bounded.
WhenTis equal to the normalized Lebesgue measuredσon∂Sn−1,then the eigen- functionPs(dσ)is the spherical λs-eigenfunction: this is the unique eigenfunction hs such thaths(o)=1,and which is invariant under the subgroup(O(n))of the isometries ofHnwhich fix the origino.
First of all,it is easy to see that the only values ofssuch thatEλs ={0}are con- tained in the strip{s∈C; (s)∈[0, n−1]}. Indeed,letf∈Eλsbe a nonzero function and letx∈Hnbe a point wheref(x)=0. Up to precomposingfwith an isometry which sends otox,we can assume thatf(o) =0. Then,by averagingfover the subgroup(O(n))of all rotations which fixo,we obtain aλs-eigenfunction which is bounded,nonzero,and invariant under the orthogonal groupO(n). But the unicity of the spherical functionhs implies that this function is a multiple ofhs;therefore the spherical functionhshas to be bounded. It is now easy to see that the only values ofsfor which this can occur are exactly those such thats∈[0, n−1].
We denote byHthe vertical strip{s ∈C | s∈ ]0, n−1[}. The method we use in this paper does not allow to deal with the values ofssuch thats=0orn−1. Therefore, we will focus on the values ofs∈H.
Whenn= 2,there is a characterization of the distributionsT such thatPs(T)is bounded[11]: fors = 0,they are exactly the distributional derivatives of(s)-H ¨older functions on∂H2. In higher dimension,the notion of derivatives depends on the data of a vector field,and so does not have easily an intrinsic meaning. One way to avoid this dif- ficulty is to use the spherical LaplacianΔσacting on functions onSn−1. If a distribution TsatisfiesT,1=0,there is an integerkand a continuous functionHsuch that the dis- tributionTcan be written asΔkσ(H). Then the regularity ofTcan be measured in terms of the integerkand of the regularity ofH. We use the Lipschitz spaces of orderαforα > 0
in order to measure the regularity of a functionH:Sn−1→C. Let us recall the definition and the basic properties of Lipschitz spaces.
Letθ∈[−π, π]andRθbe a rotation of angleθacting onSn−1. LetHbe a function onSn−1. We denote byDθ(H)=D1θ(H)the functionH◦Rθ−H,and forn≥1,we define inductivelyDnθ(H)=Dθ◦Dnθ−1(H).
Letα > 0and let[α]be its integer part. Denote byΛα(Sn−1)the space of(equiv- alence classes of)bounded measurable functions onSn−1for which there exists a con- stantC > 0,such that for any rotationRθof angleθ,one has
D[α]θ +1(H)
∞ ≤C|θ|α. (1.2)
One can define a norm on the spaceΛα(Sn−1),by settingHα equal to the infimum of theL∞-norm ofHand of the constantsCwith this property. With this norm,the space Λα(Sn−1)becomes a Banach space: it is called the Lipschitz space of orderα.
Ifα ∈]0, 1[,the spaceΛα(Sn−1)is the classical H ¨older space of orderα,and the norm is equivalent to the classical H ¨older norm. In fact,a smoothing argument shows that any function inΛα(Sn−1)is almost everywhere equal to a H ¨older-continuous func- tion of orderα. Any function inΛ1(Sn−1)admits a continuous representative as well(see [14,page 331]).
Ifα > 1,the following holds. Let Xbe any vector field onSn−1. Then,forH ∈ Λα(Sn−1),the Lie derivativeLXHis contained inΛα−1(Sn−1),with a norm bounded in terms ofHαand the norm of the vector fieldX. By induction,one can characterize,for α > 1,α /∈N,the spaceΛα(Sn−1)as the space of functions which are[α]-times differen- tiable and which have all mixed derivatives of order[α]−1inΛα−[α]+1(Sn−1).
We will use the following important property of the Lipschitz spaces. For any integerm≥ α,one can replace in the definition ofΛα(Sn−1)the quantityD[α]θ +1(H)∞ byDmθ+1(H)∞. The space defined in that way is equal toΛα(Sn−1)and the two norms are equivalent(see,e.g., [14,page 331]). In particular,forα∈]0, n−1[,one can define the spaceΛα(Sn−1)as the space of functionsH∈L∞(Sn−1)such thatDnθ(H)∞ ≤C|θ|α,for some constantCindependent ofθ.
In the following,we denote by Λ0α(Sn−1) the space of functionsH ∈ Λα(Sn−1) such that
Sn−1H dσ=0.
Given an integerj≥0,we denote byΔjσΛ0α(Sn−1)the space of those distributions onSn−1which can be written asΔjσ(H)for a functionH∈Λ0α(Sn−1).
The paper is devoted to prove the following theorem,which describes the spaces Eλsfors∈H.
Theorem 1.3. Lets=δ+it∈H.
If n is odd, then Ps induces an isomorphism between the Banach spaces Δ(nσ−1)/2Λ0δ(Sn−1)⊕CdσandEλs;
If n is even, then Ps induces an isomorphism between the Banach spaces
Δn/2σ Λ0δ+1(Sn−1)⊕CdσandEλs.
In this theorem,theλs-eigenfunctions which vanish at the origin are exactly the image ofΔ(nσ−1)/2Λ0δ(Sn−1)by the Poisson-Helgason transform. The image of the complex lineCdσis equal to the multiples of the spherical functionhswhich equals1at the origin o.
Note thatTheorem 1.3does not include the bounded harmonic functions onHn; they correspond to the cases = n−1. The boundary values of those are characterized by the Fatou theorem: they are exactly the functions inL∞(Sn−1). However,derivatives of order n−1 of functions in Λn−1(Sn−1)are not in L∞(Sn−1), in general. Therefore, Theorem 1.3cannot hold in the cases=n−1.
The proof ofTheorem 1.3contains two parts. The first one amounts to show that Psrestricted toΔ(nσ−1)/2Λ0δ(Sn−1)⊕Rdσor toΔn/2σ Λ0δ+1(Sn−1)⊕Rdσaccording to the par- ity ofntakes range inEλsand that it is a continuous operator. This is done inSection 2.
If the real part ofsis not of the formn−1−2jwithj∈N,the proof is a direct computation using easy estimates on the Poisson kernel. The other cases are obtained by interpolation of analytic families of operators.
The second part amounts to show that the restriction ofPsis surjective. This is proven in two steps. First,we consider the case whenδ = s ≥ (n−1)/2. Under this assumption,we assign inSection 4 to any functionf ∈ Eλs a “boundary value”: it is a distributionT onSn−1which is contained in the expected space,that is,it satisfiesT = Δ(nσ−1)/2Hfor some Lipschitz functionHinΛδ(Sn−1)ifnis odd orT = Δn/2σ Hfor some Lipschitz functionHinΛδ+1(Sn−1)ifnis even. The proof relies on estimates of a certain differential operatorDλacting on functions on the real line;this operator is the radial part of the Laplace operator onHn, written in polar coordinates. These estimates are obtained inSection 3;we use in an essential way the assumption thatδ≥(n−1)/2.
InSection 5,one proves,under the assumptionδ =s≥(n−1)/2,that the dis- tributionT reproducesf,that is, thatPs(T) = f. This shows thatPs is onto under the assumptions ofTheorem 1.3ands≥(n−1)/2.
InSection 6,we extend this to the caseδ <(n−1)/2by using a relation between the(n−1−s)th power of the Poisson kernel and itssth power: they are related by a fractional integration(resp.,derivation)operator which is studied in Proposition 6.1.
We study some continuity properties of this operator acting on Lipschitz spaces
(Proposition 6.1(4)). This result is similar to the one for fractional integration(resp., derivation)operator acting on functions onRn,but as we did not find any reference,we propose an elementary proof of it.
Finally,sincePsis continuous,onto,and injective,the isomorphism statement in Theorem 1.3follows from the Banach closed graph theorem.
The computations in Section 3to Section 5 use heavily estimates on the Green function and on the spherical function of the operatorΔ+λsI. These estimates are well known,but in order to make the paper self-contained,we give the proofs in an appendix.
In all the sequel,we assume thatn≥3.
2 Continuity properties of the Poisson-Helgason transformPs
In order to prove the direct part ofTheorem 1.3,we use the following estimate on the derivatives of the Poisson kernel.
Claim 2.1. Letj ∈ Nands = δ+itinC. There exists a polynomialQin one complex variable so that for anyζandξinSn−1,for any pointx∈Hnat distancerfromoon the ray[oζ[,one has
ΔjσPs(x, ξ)≤C(s) e−δr
e−r+|ζ−ξ|2δ+2j, (2.1)
whereC(s)=|Q(s)|.
Proof. In the ball model,the Poisson kernel can be written as
P(x, ξ)= 1−|x|2
|x−ξ|2. (2.2)
In Euclidean polar coordinates,one hasx=(tanh(r/2))ζ. Therefore,
P(x, ξ)=
1−
tanh r
2 2
tanh
r 2
2
+1−2tanh r
2 ζ, ξ
. (2.3)
The statement follows after differentiating this formula.
Proposition 2.2. Letn≥3. Lets∈H.
Ifnis odd,thenPsmaps continuouslyΔ(nσ−1)/2Λδ(Sn−1)toL∞(Hn). Ifnis even, thenPsmaps continuouslyΔn/2σ Λδ+1(Sn−1)toL∞(Hn).
As it was said inSection 1,this result is false fors = n−1since,by the Fatou theorem,the Poisson-Helgason transformPn−1is an isomorphism fromL∞(Sn−1)toE0. Proof. Assume first thatnis odd. Letf∈Λδ(Sn−1). We choose the integerj≥1such that n−1−δ=2j−αwithα∈]0, 2]. Suppose also thatn−1−δis not an even integer,that is, thatα=2.
The Green formula implies that for anyx∈Hn,one has
I:=Ps
Δ(nσ−1)/2f (x)=
ΔjσPs(x,·), Δ(nσ−1)/2−jf
. (2.4)
But the functionh=Δ(nσ−1)/2−jfbelongs toΛα(Sn−1)with norm bounded byCfδ. Asj≥1,for anyζ∈Sn−1,one has
I=
Sn−1ΔjσPs(x,·)h(·)dσ=
Sn−1ΔjσPs(x,·)
h(·)−h(ζ)
dσ. (2.5)
Let nowζbe the end of the rayoxon the sphere and denote byrthe hyperbolic distance fromxtoo. Forθ∈[0, π],denote byB(ζ, θ)⊂Sn−1the spherical ball of radiusθ aroundζand setP(x, θ)=P(x, ξ)for anyξ∈∂B(ζ, θ).
Using geodesic polar coordinates onSn−1around the pointζ,this last integral also equals
π
0
ΔjσPs(x, θ)
∂B(ζ,θ)
h(·)−h(ζ)
dσθ dθ, (2.6)
wheredσθis the Riemannian measure on the sphere∂B(ζ, θ)for the induced Riemannian metric. Denoting byξthe symmetric ofξ∈∂B(ζ, θ)with respect to the centerζ,one has
∂B(ζ,θ)
h(·)−h(ζ) dσθ
= 1 2
∂B(ζ,θ)
h(ξ)+h(ξ)−2h(ζ) dσθ(ξ)
. (2.7) Sinceh∈Λα(Sn−1)withα∈]0, 2[,|h(ξ)+h(ξ)−2h(ζ)|≤ hα|ξ−ζ|α. So,the left integral above is bounded byChαθn−2+α.
Sincen−2+α=δ+2j−1,it follows fromClaim 2.1that for anyx∈Hn,
|I|≤ hα
c1
e−r
0
e−δr
e−(2δ+2j)rθδ+2j−1dθ+c2 π
e−r
e−δrθδ+2j−1
θ2δ+2j dθ . (2.8) Therefore,|I|≤Chαfor a constantCwhich does not depend onh. So,Ps(Δ(nσ−1)/2f)∞
≤CΔ(nσ−1)/2−jfα≤Cfδ. The result of the proposition is proved except whenn−1−δ is an even integer.
To deal with the caseα=2,that is,when(n−1)−δis even,we use interpolation of analytic families of operators. We give first a complement to the last proof. Letδ0 = n−1−2j. Chooseδ2in]0, n−1[withδ0< δ2< δ0+2. Lets=δ+it,withδ∈]0, δ2].
Then,for f ∈ Λδ2, one hasPs(Δ(nσ−1)/2f)∞ ≤ C(s)fδ2 for a constantC(s) which is independent offand is the absolute value of some polynomialRat points. In- deed,writing δ2 = n−1−2j+α2, the argument above gives, for f ∈ Λδ2(Sn−1)and h=Δ(nσ−1)/2−jf,
Ps
Δ(nσ−1)/2f (x)
≤C(s)hα2
c1
e−r
0
e−δr
e−(2δ+2j)rθδ2+2j−1dθ+c2 π
e−r
e−δrθδ2+2j−1 θ2δ+2j dθ ,
(2.9)
forr=d(o, x). Therefore,ifδandδ2satisfyδ2< 2δ,one has
Ps
Δ(nσ−1)/2f
(x)≤C(s)fδ2e−(δ2−δ)r, (2.10)
whereC(s)can be chosen to be the modulus of a polynomialRwhich does not vanish in the strips≥0.
Choose now δ1 andδ2 with δ1 < δ0 < δ2 and sufficiently close so that δ2 <
2δ1. Consider the Lipschitz spacesΛδ1(Sn−1)andΛδ2(Sn−1);it is known that the spaces Λδ(Sn−1)for δ ∈ [δ1, δ2]form an interpolating family[2]. On the closed stripB = {s | s∈[δ1, δ2]},consider the analytic family of operatorsTs =(1/R(s))Ps◦Δ(nσ−1)/2. Those operators are defined onΛδ2(Sn−1)and satisfy the following properties:
(1) each operator Ts maps continuously Λδ2(Sn−1) to the Banach (separable) spaceCb(Hn)of bounded continuous functions onHn(with the uniform norm);
(2) for allf∈ Λδ2(Sn−1),for allx ∈Hn,the functions→ Ts(f)(x)is continuous, analytic in the interior ofBand bounded overB;
(3) forf∈Λδ2(Sn−1),one hasTδ1+itf ≤ fδ1andTδ2+itf ≤ fδ2.
Then the hypotheses of the interpolation theorem(see,e.g., [3,Theorem 1])are satisfied.
This theorem gives that fors=δ∈B,Tsis continuous fromΛδ(Sn−1)toCb(Hn). Making δ=δ0finishes the proof ofProposition 2.2for odd values ofn.
The proof for even values ofnis the same.
3 Estimates of the solutions ofDλf=g
In this section,we study the differential operatorDλwhich is defined as acting on func- tions on]0,∞[by
Dλ=(sinhr)2 ∂2
∂r2+(n−1)cothr ∂
∂r+λId . (3.1)
Recall that the hyperbolic Laplacian has the following form in polar hyperbolic coordinates(r, ζ)aroundo. For aC2functionu=u(r, ζ):Hn\ {o}→C,one has
Δu=∂2u
∂r2 +(n−1)cothr∂u
∂r + 1
(sinhr)2Δσu(r,·). (3.2)
From this expression,the solutions ofDλu=0are precisely the radialλ-eigenfunctions.
In Propositions7.1and7.2,we will recall the behaviour near∞of the two basic solutions of the eigenfunction equations,that is,the spherical functionhsand the Green function gs. The next result studies similarly the behaviour of a solutionuofDλu=vin terms of the behaviour ofv.
To simplify notations,we writeλ=λs=s(n−1−s).
Throughout this section,we make the assumption thatδ=s≥(n−1)/2.
Proposition 3.1. Letr0 > 0and letl be an integer such that l ≤ n−3−δ. Letv be a smooth function on[r0,∞[such that|v(r)| ≤ Ce−lr forr ≥ r0. Letu : [r0,∞[→ Cbe a smooth function such thatDλu=v. Then,ifl < n−3−δ,one has
u(r)≤κ C+u
r0+u
r0e−(l+2)r. (3.3)
Ifl=n−3−δ,one has u(r)≤κ
C+u
r0+u
r0re−(n−1−δ)r. (3.4)
In both cases,κis a constant which depends only onr0and ons.
Proof. The solutionuofDλu= vis uniquely determined by its behaviour up to first or- der atr0. LetaandbinCbe such thatags(r0)+bhs(r0) = 0andags(r0)+bhs(r0)= 1.
This solution(a, b)exists and is unique,since the determinant of this system equals the Wronskian ofgs andhs atr0which does not vanish. Then,by substracting fromuthe combinationags+bhs,one reduces to the case whenu(r0)=u(r0)=0. This linear com- bination grows less thanκ(|u(r0)|+|u(r0)|)e−(n−1−δ)r,whereκis a constant depending only onr0. Therefore,to proveProposition 3.1,we can assume thatu(r0)=u(r0)=0.
LetU(r)be the vector with coordinatesu(r)andu(r)and letV(r)be the vector with coordinatesv(r)/(sinhr)2and0. Then
U(r)=A(r)U(r)+V(r) on r0,∞
, (3.5)
where
A(r)=
−(n−1)cothr −λ
1 0
. (3.6)
LetM(r)be the solution ofM(r)= −M(r)A(r)which vanishes atr0;then
U(r)= r
r0
M−1(r)M(t)V(t)dt. (3.7)
For any solutionY(r)of the equation Y(r) = A(r)Y(r),the vectorM(r)Y(r)is constant.
This holds in particular for the spherical vectorHs(r)with coordinates(hs(r), hs(r))and for the Green vectorGs(r)=(gs(r), gs(r)). Now,since the WronskianW(t)does not van- ish,we may writeV(t)as a linear combination of these two vectors:V(t) = α(t)Hs(t)+ β(t)Gs(t). It comes out that
U(r)= r
r0
α(t)dt Hs(r)+ r
r0
β(t)dt Gs(r), (3.8)
whereα(t)andβ(t)are given,respectively,by
α(t)=cn−1v(t)gs(t)(sinht)n−3, β(t)= −cn−1v(t)hs(t)(sinht)n−3. (3.9) Using the classical estimates onhs andgs recalled in the appendix,one obtains easily
the required estimates ofProposition 3.1.
Corollary 3.2. Letr0> 0. Letj∈N,l∈Zbe such thatl+2j≤n−1−δ. Letvbe a smooth function on[r0,∞[such that|v(r)|≤Ce−lr. Assume thatuis a smooth function on[r0,∞[
which solves the differential equationDjλu=v. Then ifl+2j < n−1−δ,one has u(r)≤κ
C+
2j−1
p=0
u(p)
r0 e−(l+2j)r. (3.10)
Ifl+2j=n−1−δ,one has u(r)≤κ
C+
2j−1
p=0
u(p)
r0 re−(n−1−δ)r, (3.11)
where the constantκdepends only onr0and ons.
Proof. By decreasing induction onm,1 ≤ m≤ j−1,one proves,usingProposition 3.1, that
Dmu(r)≤κm
C+
2(j−m)−1 0
u(p) r0
e−(l+2(j−m))r (3.12)
forr≥r0,where the constantκmdepends only onr0. Form=0,it gives the result. Notice that one needs to use the degenerate case ofProposition 3.1only at the last step of the
induction and whenl+2j=n−1−δ.
4 Construction of a distribution onSn−1associated to a bounded eigenfunction Now,we begin the proof of the fact thatPsis onto. Letf∈Eλs. We are going to construct an explicit distributionT such thatf=Ps(T). Notice first that we may assume thatf(o)= 0,since up to replacingfbyf−f(o)hs,we get a bounded eigenfunction inEλswhich has L∞-norm less than2f∞. In this section,we associate tofa distributionTsonSn−1and study its Lipschitz regularity. In the next section,we will prove that it satisfiesPs(Ts)=f.
As in the preceding section,we assume thats = δ ∈ [(n−1)/2, n−1[and we writeλ=λs=s(n−1−s)to simplify notations. Consider forr > 0the function
T(r, ξ)=(sinhr)n−1
f∂gs
∂r −∂f
∂rgs (r, ξ), (4.1)
wheregsis the Green function: notice that the Green function depends onsand not only onλs(cf.Section 6).
Proposition 4.1. Lets = δ+itwithδ ∈ [(n−1)/2, n−1[. Then the distributionsT(r,·) converge to a distributionT =TsonSn−1. Furthermore,ifδ=n−1−2j+α,withα∈]0, 2], there is a functionH∈Λ0α(Sn−1)such thatT = Δjσ(H). This functionHsatisfiesHα ≤
Cf∞ for a constantCwhich depends only ons.
Remark 4.2. As a corollary,we obtain that the limit distributionTcan always be written asT = Δ(nσ−1)/2hwithh ∈ Λ0δ(Sn−1)ifnis odd or thatT = Δn/2σ hwithh ∈ Λ0δ+1(Sn−1) ifnis even. Indeed,sinceH has mean-value zero overSn−1,one may write it as H = Δ(nσ−1)/2−jhifnis odd andH = Δn/2σ −jhifnis even. As powers of Laplacians give iso- morphisms between Lipschitz spaces of different parameters(see[12]),we get thath ∈ Λ0δ(Sn−1)ifnis odd and thathinΛ0δ+1(Sn−1)ifnis even sinceH∈Λ0α(Sn−1).
Proof. The following properties ofT(r,·)are easy to establish.
(1)
Sn−1T(r,·)dσ=0: this comes from the Green formula sincef(o)=0and since in the distributional sense,Δgs+λsgsequals the Dirac mass at the origin δo.
(2) (∂T/∂r)(r, ξ)= −(sinhr)n−3gs(r)Δσf(r, ξ): this comes by differentiating the expression ofT(r, ξ)and using the expression for the Laplacian in polar coordinates.
In order to illustrate the method,we consider first the easiest case δ = s ∈ ]n−3, n−2[. By(1),there is a unique smooth functionξ →H(r, ξ)which has zero mean and such thatT(r,·)=ΔσH(r,·). By(2),we have
∂
∂rH(r,·)= −(sinhr)n−3gs(r)f(r,·). (4.2)
Fix anr0> 0. Using the estimates ongsand the boundedness off,one obtains that there existsC > 0which depends only onsand such that for anyr≥r0,
(sinhr)n−3gs(r)f(r,·)≤Cf∞e(n−3−δ)r. (4.3)
Therefore, H(·)=H
r0,·
− ∞
r0
(sinht)n−3gs(t)f(t,·)dt (4.4)
defines a continuous function onSn−1. In the distributional sense,we have ΔσH(·)= lim
r→∞T(r,·)=:T(·). (4.5)
We are going to prove thatH ∈ Λ0δ−(n−3)(Sn−1). By assumptionδ−(n−3) ∈ ]0, 1[so it suffices to show thatDθH∞ ≤ Cf∞|θ|δ−(n−3),for a constantCwhich only depends ons. Recall thatDθis the first-order difference operatorDRθ associated to the rotation Rθ∈O(n),that is,the operator which assigns to a functionFthe functionDRθF=F◦Rθ−F.
We have
DθH(·)=DθH r0,·
− ∞
r0
(sinht)n−3gs(t)Dθf(t,·)dt=I+II,
|II|≤ ∞
r0
(sinht)n−3gs(t)Dθf(t,·)dt
≤
−log|θ|
r0
(sinht)n−3gs(t)Dθf(t,·)dt +
∞
−log|θ|(sinht)n−3gs(t)Dθf(t,·)dt
=(1)+(2).
(4.6)
By the Harnack inequalities,sincefis bounded,the gradient offis bounded in terms of f∞ ands;it follows that
Dθf(r, ξ)≤Cf∞|θ|er (4.7)
for some constantCwhich only depends ons. This gives
(1)≤ f∞
−log|θ|
r0
et(n−2−δ)|θ|dt≤Cf∞|θ|δ−(n−3). (4.8)
Clearly,one has also,for anyr,Dθf(r,·)∞ ≤2f∞,so that one gets
(2)= ∞
−log|θ|(sinht)n−3gs(t)Dθf(t,·)dt
≤Cf∞ ∞
−log|θ|et(n−3−δ)dt≤Cf∞|θ|δ−(n−3).
(4.9)
So |II| ≤ Cf∞|θ|δ−(n−3). It remains to estimate the term I = DθH(r0,·). We have ΔσDθH(r0,·) = Dθf(r0,·). By the Harnack inequality,and due to the compacity ofSn−1, Δ−σ1is a continuous operator for theL∞-norm,so one has
|I|≤DθH r0,·
∞ ≤Cf∞|θ|≤Cf∞|θ|δ−(n−3), (4.10) for a constantCwhich only depends ons. It follows thatH∈Λ0δ−(n−3)(Sn−1),with a norm smaller thanCf∞.
We now show how this method can be extended to handle the general case. We are going to use the results ofSection 3on the operatorDλ;this is why we need to assume thats=δ ∈[(n−1)/2, n−1[. Let us denote byjthe integer such thatδ=n−1−2j+α withα∈]0, 2].
Sincefis aλ-eigenfunction,it satisfies
Dλf(r, ξ)= −Δσf(r, ξ). (4.11)
From the study of the Green functiongs in the appendix,there is anr0depending only onssuch thatgs(r)does not vanish forr ≥ r0. Therefore, by iterating the differential equation satisfied byT(r,·),we see that forr≥r0,one has
Djλ−1
(sinhr)3−n gs(r)
∂T
∂r(r,·) =(−1)jΔjσf(r,·). (4.12)
SinceT(r,·)has zero mean,there exists a unique smooth functionH(r,·)onSn−1which has zero mean and such thatT(r,·) = ΔjσH(r,·). This uniqueness property ofH(r,·)im- plies that
Djλ−1
(sinhr)3−n gs(r)
∂H
∂r(r,·) =(−1)jf(r,·). (4.13)
Sincefis bounded and since2(j−1)≤n−1−δ,Corollary 3.2can be applied with l=0. This gives
(sinhr)3−n gs(r)
∂H
∂r(r,·) ≤C
f∞ +
2j−3
0
∂pH
∂rp r0,·
e−2(j−1)r ifα < 2, (sinhr)3−n
gs(r)
∂H
∂r(r,·) ≤C
f∞ +
2j−3
0
∂pH
∂rp r0,·
re−2(j−1)r ifα=2,
(4.14)
for a constantC,which only depends onsand onr0. One also has
Δjσ ∂pH
∂rp
r0,· = ∂pT
∂rp r0,·
. (4.15)
Going back to the definition ofT(r,·)and using the Harnack inequality,one finds ∂pT
∂rp r0,·
∞ ≤Cf∞, (4.16)
forp≤2j−3,the constantConly depending onr0andj. Therefore,we have ∂pH
∂rp
r0,·
∞
≤Cf∞, ∂H
∂r(r,·)
∞ ≤C(sinhr)n−3gs(r)e−2(j−1)rf∞ ≤Cf∞e−αr ifα < 2,
≤C(sinhr)n−3gs(r)re−2(j−1)rf∞ ≤Cf∞re−αr ifα=2,
(4.17)
for a constantCwhich only depends onr0and ons. In any cases,since the last term is integrable over[r0,∞[,uniformly onSn−1,we may define a continuous functionHonSn−1 by setting
H(·)= lim
r→∞H(r,·)=H r0,·
+ ∞
r0
∂H
∂t(t,·)dt. (4.18) In the distributional sense,this function satisfies
ΔjσH(·)= lim
r→∞ΔjσH(r,·)=:T(·). (4.19)
We are now going to prove thatH∈Λ0α(Sn−1). LetRθ∈O(n)be a rotation of angle θ∈]−π, π[. Sinceα∈]0, 2],we just need to prove thatD3θH∞ ≤Cf∞|θ|αfor a constant Cwhich does not depend onθnor onRθ(see[14,page 331]). We have
D3θH(·)=D3θH r0,·
+ ∞
r0
∂D3θH
∂t (t,·)dt. (4.20) For the termD3θH(r0,·), we note that Δjσ−1D3θH(r0,·) = D3θT(r0,·). By the Harnack in- equality,D3θT(r0,·)∞ is bounded byCf∞|θ|2,for a constantCwhich depends only on r0. Therefore,sinceΔ−σ1is continuous for theL∞-norm,a similar bound holds also for D3θH(r0,·): itsL∞-norm is smaller thanCf∞|θ|2which is smaller thanCf∞|θ|α.
It remains to deal with the term corresponding to the integral. Arguing as in the caseδ ∈]n−3, n−2[,we split this integral into the sum ofI=−log|θ|
r0 andII=∞
−log|θ|. We have
Djλ−1
(sinhr)3−n gs(r)
∂D3θH
∂r (r,·) =(−1)jD3θf(r,·). (4.21) SinceD3θf(r,·)∞ ≤Cf∞|θ|3e3r,for a constantCindependent ofrandθ,we may apply Corollary 3.2withl= −3: it comes out that forα∈]0, 2],
∂D3θH
∂r (r,·)
∞ ≤Cf∞|θ|3e(3−α)r, (4.22) so that
|I|≤Cf∞|θ|3
−log|θ|
r0
e(3−α)tdt≤Cf∞|θ|α. (4.23) Now,ifα < 2,we already proved during the construction ofHthat(∂H/∂r)(r,·)∞
≤Cf∞e−αr;we therefore have also(∂D3θH/∂r)(r,·)∞ ≤Cf∞e−αr. This gives
|II|≤Cf∞|θ|α. (4.24)
Adding the estimates above,we deduce thatD3θH∞ ≤Cf∞|θ|α. This finishes the proof in the caseα < 2.
It remains to consider the caseα=2. There we are in the critical case for applying Corollary 3.2with l = 0 and we argue differently. By the Harnack inequality,we have Dθf(r,·)∞ ≤Cf∞|θ|erfor a constantCindependent offandθ. ApplyingCorollary 3.2 withl= −1,we deduce that(∂DθH/∂r)(r,·)∞ ≤Cf∞|θ|e−r. But,we have
∂D3θH
∂r (r,·) ∞ ≤4
∂DθH
∂r (r,·)
∞ ≤Cf∞|θ|e−r, (4.25)
for a constantCwhich is independent ofRθandf. It follows that ∞
−log|θ|
∂D3θH
∂t (t,·)dt
≤Cf∞|θ|2. (4.26) It ends the proof of the caseα=2and ofProposition 4.1also.
5 Proof ofTheorem 1.3whenδ≥(n−1)/2
Letf ∈ Eλs. By a theorem of Helgason and Minemura([7,9],see also[1]),anyλs-eigen- function can be written asPs(D),whereDis a hyperfunction onSn−1. Also,T is unique except whens∈−N∗[9].
To prove Theorem 1.3 when s ≥ (n−1)/2, we show that the distribution T we constructed in the previous section is equal toD,that is, that one hasf = Ps(T).
Theorem 1.3will then follow fromProposition 4.1. We suppose first thatf = Ps(φ),for aC1-functionφonSn−1. ByProposition 2.2,fis bounded and we can consider the cor- responding distributionTs. We need to show thatTs = D. To simplify the notations,we denoteTsbyT.
Recall that the distributionTis the limit asr→ ∞of the distributions
T(r,·)=(sinhr)n−1
∂f(r,·)
∂r gs−∂gs
∂r f(r,·) . (5.1)
With these notations,we have the following result.
Proposition 5.1. Letφbe aC1function onSn−1andf=Ps(φ). Then,one has
r→∞lim
T(r,·), Ps(z,·)
=f(z). (5.2)
Proof. We begin by studying the kernel functionK(s)r (ζ, ξ)defined by
K(s)r (ζ, ξ)=(sinhr)n−1 ∂Ps
∂r (rζ, ξ)gs(r)−Ps(rζ, ξ)∂gs
∂r (r)
. (5.3)
Lemma 5.2. The kernelK(s)r (ζ, ξ)has the following two properties:
(1) for anyξ∈Sn−1,one has
Sn−1K(s)r (ζ, ξ)dσ(ζ)=1;
(2) for(s)=δ≥(n−1)/2,andr0> 0,there is a constantC > 0such that for all r≥r0,for allζ,ξinSn−1,
K(s)r (ζ, ξ)≤C e(n−1−2δ)r
e−r+|ζ−ξ|2δ, K(s)r (ζ, ξ)≤Cer(n−3−2δ)
|ζ−ξ|2δ+2. (5.4)
Proof. Property(1)follows from the Green formula applied to the functionsPs(·, ξ)and gsand from the fact thatPs(o, ξ)=1for allξ∈Sn−1.
To prove(2),we assume first thatδ > (n−1)/2. For anyr ≥ r0 > 0,we know from the appendix that|gs(r)|≤Ce−δrfor a constantCwhich is uniform assremains in a compact of the set{s | (s) ∈ [(n−1)/2, n−1[}. By the Harnack inequality,it follows that|∂gs(r)/∂r|≤Ce−δr,where the constantCis uniform assremains in a compact set.
This leads first to the inequality K(s)r (ζ, ξ)≤C e(n−1−2δ)r
e−r+|ζ−ξ|2δ, (5.5)
where the constantCis independent ofr ≥ r0and ofsas long as it stays in a compact set. In particular,for allζ=ξ,K(s)r (ζ, ξ)→0asr→ ∞whenδ >(n−1)/2. The derivative with respect torofK(s)r (ζ, ξ)can be computed as we did in the previous section for the construction of the distributionT. This gives
∂K(s)r (ζ, ξ)
∂r =(sinhr)n−3gs(r)ΔσPs(rζ, ξ). (5.6)
Using the same estimates as above ongsand its derivative,and onΔσPs(cf.Claim 2.1), it comes out that
∂K(s)r (ζ, ξ)
∂r
≤C er(n−3−2δ)
e−r+|ζ−ξ|2δ+2. (5.7)
Integrating with respect torand using thatKt(ζ, ξ)→0ast→ ∞,it gives that K(s)r (ζ, ξ)≤Ce(n−3−2δ)r
|ζ−ξ|2δ+2, (5.8)
for a constantCwhich does not depend onsas long as it stays in a compact set ofs≥ (n−1)/2.
Since the kernelK(s)r is continuous with respect tos,the last estimate holds also for anyswiths=(n−1)/2(for a constant which depends onr0and ons): it suffices to approximates=(n−1)/2+itbysk+itfor a sequenceskwhich tends to(n−1)/2from
above.
Proof ofProposition 5.1. Asf=Ps(φ),we have
T(r, ζ)=
Sn−1K(s)r (ζ, ξ)φ(ξ)dσ(ξ). (5.9)