Spectral analysis of a self-similar Sturm-Liouville operator Christophe Sabot 4 0 Laboratoire de Probabilit´es et mod`eles al´eatoires, Universit´e Paris 6, 0 2 4, Place Jussieu, 75252 Paris cedex 5, and n a Ecole Normale Sup´erieure, J DMA, 45, rue d’Ulm, 75005 Paris, 1 0 3 1 v 6 Abstract: In this text we describe the spectral nature (pure point or continuous) 5 0 of a self-similar Sturm-Liouville operator on the line or the half-line. This is motivated 1 by the more general problem of understanding the spectrum of Laplace operators on 0 unbounded finitely ramified self-similar sets. In this context, this furnishes the first 4 0 example of a description of the spectral nature of the operator in the case where the / so-called ”Neumann-Dirichlet” eigenfunctions are absent. h p - h t AMS classification: 34L10(34L20,82B44) a m : v Key words: Spectral theory, Eingenfunctions expansion, Sturm-Liouville operators, i X dynamics in several complex variables, analysis on self-similar sets, fractals. r a 1E-mail address: [email protected] 1 In this text we consider a family of self-similar Sturm-Liouville operators, indexed by a parameter ω called the blow-up, and describe the nature of the spectrum (pure point or continuous). This study is motivated by the problem of understanding the nature of the spectrum for the class of self-similar Laplacians on finitely ramified self- similar sets. In [11], [12], [14], we proved that the spectral properties of these operators are related to the dynamics of a certain renormalization map, which is a rational map of a smooth complex projective variety. The main question emerging from these works is the nature of the spectrum of these operators in the non-degenerate case d = N ∞ (cf [12], d is the asymptotic degree of the renormalization map, N is the number ∞ of subcells of the set at level 1). As far as the author knows, there is no non-trivial example where the nature of the spectrum is understood in this case. The family of self-similar Sturm-Liouville operators we consider here corresponds to the case d = N, and we are able to determine, for all values of the blow-up ω, ∞ whether the spectrum is pure point or continuous (we are not able to distinguish the singular continuous and the absolute continuous part). This example also illustrates [13], since it shows that for atypical value of the blow-up ω the spectral nature of the operator can be radically different. The proofs of the results suggest that the spectral nature of the operators may be related to the way the iterates of the renormalization map approach certain curves (cf remark 4.1). We hope that this example will help to understand the general case of finitely ramified self-similar sets. The strategy we adopt here to obtain the renormalization equation is a bit differ- ent than in [12], and much more straightforward: we write a renormalization equation directly on the propagator of the differential equation associed with the spectral prob- lem. This simplification comes from the 1-dimensional nature of the problem in this example. 1 Definitions and results Let I be the interval I = [0,1] and α a real such that 0< α< 1. We set δ = α . We 1 α define the two homotheties Ψ ,Ψ by : − 1 2 Ψ (x) = αx, Ψ (x) = 1 (1 α)(1 x). 1 2 − − − So, Ψ (I) = [0,α], Ψ (I) = [α,1] and the interval I is self-similar with respect to 1 2 (Ψ ,Ψ ). 1 2 Let b be a real number such that 0 < b < 1. It is classical that there exists a unique probability measure m on I such that 1 1 1 fdm = b f Ψ dm+(1 b) f Ψ dm, f C(I). (1) 1 2 ◦ − ◦ ∀ ∈ Z0 Z0 Z0 N.B.: For α = b, the measure m is singular with respect to the Lebesgue measure, for 6 α = b it is the Lebesgue measure. We denote by H+ the operator d d with Neumann boundary conditions on I, i.e. dmdx 2 it is the operator defined on the domain: f L2(I,m), g L2(I,m),f(x) = ax+b+ x yg(z)dm(z)dy, { ∈ ∃ ∈ 0 0 Rf′(R0) = f′(1) = 0 , } by H+f = g. We denote by H the corresponding operator with Dirichlet boundary conditions on − I, i.e. the operator defined on the domain f L2(I,m), g L2(I,m),f(x) = ax+b+ x yg(z)dm(z)dy, { ∈ ∃ ∈ 0 0 Rf(R0)= f(1) = 0 , } by H f = g. − Remark 1.1 : These operators belong to the class of self-similar Laplacians as defined for example in [7] or [10]. Indeed, it is clear that H+ is the infinitesimal generator associated with the classical Dirichlet form a(f,g) = 1f g dx defined on = f 0 ′ ′ D { ∈ L2(I,m), f L2(I,dx) and with the measure m. By a change of variables we easily ′ ∈ } R see that a satisfies the following self-similarity relation a(f,g) = α 1a(f Ψ ,g Ψ )+(1 α) 1a(f Ψ ,g Ψ ), f,g . − 1 1 − 2 2 ◦ ◦ − ◦ ◦ ∀ ∈ D N Let Ω = 1,2 and let us fix a sequence ω = (ω ,...,ω ,...) in Ω. We call this 1 k { } sequence the blow-up. We extend the set I to a set I (ω) by scaling by <n> I (ω) = Ψ 1 Ψ 1(I). <n> −ω1 ◦···◦ −ωn (We write I (ω) toshow thedependenceof I in ω, butwe willsimplywriteI <n> <n> <n> when no ambiguity is possible). Clearly, I (ω) I (ω) and ω determines <n> <n+1> n+1 ⊂ the position of I (ω) in I (ω) (it is the left subinterval of I (ω) when <n> <n+1> <n+1> ω = 1 and the right otherwise). Then we set n+1 I (ω) = I (ω). <∞> ∪∞n=0 <n> Weclearly seethatI (ω)iseitherahalf-lineboundedfromtheleftifω isstationary < > ∞ to 1, a half-line bounded from the right if ω is stationary to 2, or the real line if ω is not stationary. We denote by ∂I = 0,1 the boundary points of I and by ∂I (ω), <n> { } ∂I (ω), the boundary points of I (ω), I (ω). Of course, ∂I (ω) is empty < > <n> < > < > ∞ ∞ ∞ when ω is not stationary, and contains a unique point when ω is stationary. We extend the measure m to a measure m on I (ω) by scaling by <n> <n> fdm = b 1 b 1 f Ψ 1 Ψ 1dm, (2) ZI<n> <n> −ω1 ··· −ωn ZI ◦ −ω1 ◦···◦ −ωn where we set b = b, b = 1 b. Clearly b is a compatible sequence of measures in 1 2 <n> − the sense that if suppf I then fdm = fdm , for all p 0. ⊂ <n> I<n+p> <n+p> I<n> <n> ≥ Hence, it can be extended to a measure m on I . R < > < > R Then we define H+ (ω), H (ω) (res∞p. H+ ∞(ω), H (ω)) as the operators <n> <−n> < > <− > d d withNeumannorDirichletboundarycond∞itionsonI∞ (ω)(resp. I (ω)). dm<n> dx <n> <∞> Of course, when ∂I (ω) = then H+ (ω) = H (ω) and we simply write < > < > <− > ∞ ∅ ∞ ∞ H (ω). < > ∞ 3 From now on, we make the following assumption (H) We choose b = 1 α, and we set γ = α 1(1 α) 1. − − − − This choice for b is crucial, it implies that the operator is locally invariant by transla- tion, i.e. that in each subcell of I of level < p >, p n (i.e. in each subinterval <n> ≤ of the type Ψ 1 Ψ 1(Ψ Ψ (I)) the restriction of the operator H −ω1 ◦ ··· ◦ −ωn j1 ◦ ··· ◦ jp <n> is the same. This seems to be the natural counterpart of the statistical invariance by translation for random Schro¨dinger operators (cf [12] for details). Thanks to the hypothesis (H), we can easily check that the operators H (ω) are <±n> isomorphic for different ω. Indeed, if Ψ˜ is the right composition of Ψ and Ψ 1 that i −i sends I (ω) to I (ω ), then H (ω)(f Ψ˜) = (H (ω )(f)) Ψ˜. But this is no <n> <n> ′ <±n> <±n> ′ ◦ ◦ longer true for H (ω) (in fact, H (ω) and H (ω ) are isomorphic if ω and < > < > <± > ′ ∞ ∞ ∞ ω are equal after a certain level, but this is a priori not true otherwise, cf [13] for ′ precisions). Let us denote by ν+ and ν the counting measures of the spectrum of H+ <n> <−n> <n> and H , respectively. By the previous remark, ν does not depend on ω. We <−n> <±n> define the density of states as the limit 1 µ = lim ν n 2n <±n> →∞ which exists and does not depend on the boundary condition (cf [3], [8] or [12]). ± We denote byΣ (ω)thetopological spectrumof H (ω), andby Σ (ω), Σ (ω), ± <± > ±ac ±sc ∞ Σ (ω)respectivelytheabsolutelycontinuous,thesingularcontinuous,thepurelyponc- ±pp tual part of the Lebesgue decomposition of the spectrum of H (ω). We also denote <± > ∞ byΣ (ω)theessentialspectrumofH (ω). Werecallfrom[13]thefollowingresults ±ess <± > ∞ (proved in the general setting of finitely ramified self-similar sets). Proposition 1.1 ([13], proposition 1) i) If the boundary set ∂I (ω) is empty, i.e. < > ∞ if ω is not stationary, then suppµ = Σ(ω) = Σ (ω). ess ii) Otherwise we only have suppµ = Σ+ (ω) = Σ (ω). Moreover, the eigenvalues ess −ess eventually lying in Σ (ω) supp(µ) have multiplicity 1. ± \ Proposition 1.2 ([13], proposition 2) There exists deterministic sets Σ,Σ ,Σ and ac sc Σ such that for almost all ω in Ω (for the product of the uniform measure on 1,2 ) pp { } we have Σ (ω)= Σ . ± • • In [11], we gave an explicit expression for the density of states in terms of the Green function of a certain rational map defined on the complex projective plane, and we proved that the spectrum is a Cantor subset of R for α = 1. The aim of this text is − 6 2 to go further and to describe the spectral type of the operator H (ω). The main <± > ∞ result is the following Theorem 1.1 i) If I (ω) is a half-line bounded from the left, i.e. if ω is stationary < > ∞ to 1, then for the Neumann boundary conditions we have If δ > 1, the spectrum of H+ is pure point and the eigenvalues lie in the < > • ∞ complement of suppµ. 4 If δ < 1, then Σ+(ω) = suppµ and the spectrum of H+ (ω) is continuous, i.e. < > • Σ+(ω) = . ∞ pp ∅ For the Dirichlet boundary condition we have If δ > 1, then Σ (ω) = suppµ and the spectrum of H (ω) is continuous, i.e. − <− > • ∞ Σ (ω) = . −pp ∅ If δ < 1, the spectrum of H is pure point and the eigenvalues lie in the <− > • ∞ complement of suppµ. ii) If I (ω) is the half-line bounded from the right, i.e. if ω is stationary to 2, < > then the sam∞e results hold, just replacing δ by δ 1. − iii) If the boundary set ∂I (ω) is empty, i.e. if the blow-up ω is not stationary, < > ∞ then Σ(ω) = suppµ and the spectrum of H (ω) is continuous, that is Σ (ω) = . < > pp ∞ ∅ 2 The renormalization equation of the propagator. Ex- pression of the density of states We suppose in this section that ω = (1,...,1,...) so that I (ω) = R . We denote < > + by Γ (s,t) the propagator on [s,t] R of the equation ∞ λ + ⊂ d d f = λf, (E ) λ dm dx < > ∞ for λ in C. This means that if f is a solution of (E ) on [s,t] then λ f(t) f(s) = Γ (s,t) . f (t) λ f (s) ′ ! ′ ! Classically, Γ (s,t) is an element of Sl (C) and the coefficients of Γ (s,t) are analytic λ 2 λ in λ. We set Γ = Γ (0,1) and Γ = Γ (0,α n). λ λ <n>,λ λ − We use the scaling invariance of the operator and the self-similarity to deduce two fondamental relations. Let us introduce the following notation: for a real β we denote by D the 2 2 matrix β × 1 0 D = . β 0 β ! The scaling relation of m, cf (2), implies Γ<n>,λ = Dαn Γγnλ Dα−n, (3) ◦ ◦ where γ = (α(1 α)) 1, cf hypothesis (H). Indeed, if f is a solution of (E ) on I , − λ <n> − then f ◦Ψ−1n is a solution of (Eγnλ) on I, and f(Ψ n(t)) f Ψ n(t) f (Ψ−1 n(t)) =Dαn (f ◦Ψ−1n)(t) ′ −1 ! ◦ −1 ′ ! 5 for all t in I. From the self-similarity, and hypothesis (H), we can deduce Γ = Γ (1,α 1) Γ (0,1) <1>,λ λ − λ ◦ = Dδ Γλ Dδ−1 Γλ, (4) ◦ ◦ ◦ where we recall that δ = α . Indeed, if f is solution on I of (E ), then f˜ = 1 α <1> λ f|[1,α−1]◦Ψ−11◦Ψ2 is solutio−n of (Eλ) on [0,1] and f˜(t) f(Ψ 1 Ψ (t)) f˜′(t) ! = Dδ−1 f′(Ψ−1−11◦◦Ψ22(t)) !, ∀t ∈ [0,1]. We set a (λ) b (λ) <n> <n> Γ = <n>,λ c (λ) d (λ) <n> <n> ! and simply a = a, b = b, c = c, d = d. <0> <0> <0> <0> Using (3) and (4) an easy computation gives a (λ) b (λ) a(γλ) α 1b(γλ) <1> <1> = − c (λ) d (λ) αc(γλ) d(γλ) <1> <1> ! ! a(λ)(a(λ)+δ 1d(λ)) δ 1 b(λ)(a(λ)+δ 1d(λ)) = − − − − (5) δc(λ)(a(λ)+δ 1d(λ)) δd(λ)(a(λ)+δ 1d(λ)) δ − − ! − We introduce the polynomial map f : C2 C2 → (x,y) (x(x+δ 1y) δ 1,δy(x+δ 1y) δ) (6) − − − → − − and we see from (5) that the map φ :C C2 given by → a(λ) φ(λ) = d(λ) ! satisfies f φ(λ) = φ(γλ), λ C. (7) ◦ ∀ ∈ Hence, φ(λ),λ C is an invariant complex curve for f. { ∈ } Proposition 2.1 (i) The Neumann and Dirichlet spectrum of d d are given by dm<n> dx 1 ν+ = ∆ln c (λ), <n> 2π | <n> | 1 ν = ∆ln b (λ), <−n> 2π | <n> | where ∆ denotes the distributional Laplacian on C. (ii) The infinite product ∞ α(a(γ kλ)+δ 1d(γ kλ)) − − − k=1 Y 6 is convergent and we have b(λ) = α(a(γ kλ)+δ 1d(γ kλ)), ∞k=1 − − − c(λ) = λ α(a(γ kλ)+δ 1d(γ kλ)). Q ∞k=1 − − − Proof: (i) The eigenvalues of HQ+ and H are simple and it is clear that the sets <n> <−n> of zeroes of c and b coincide respectively with the sets of eigenvalues of H+ <n> <n> <n> and H . Thus, the only thing to prove is that the zeroes of c and the zeroes of <−n> <n> b are simple. By scaling it is enough to prove it for c and b. Consider a complex <n> number λ such that Imλ > 0 and denote by f the solution of (E ) on I with initial λ condition f(0) = 1, f (0) = 0. By the Lagrange identity we have ′ [Im(ff )]1 = Im(λ) f 2dm, ′ 0 I| | Z which in our case gives Im(a(λ)c(λ)) = Im(λ) f 2dm. (8) I| | Z In particular, this means that a(λ)c(λ) sends the upper-half plane to itself and hence a(λ)c(λ) cannot have a multiple zero on the real axis. (ii) Remark first that a(0) = b(0) = d(0) = 1 and c(0) = 0. Hence, there exists K such that c(λ) = Kλ+O(λ2) for λ small. Considering equation (8) for λ small, we deduce that K = 1. Indeed, a(0) = 1 and thus we have K = f 2dm, where f is the I| | solution of d d f = 0, with initial condition f(0) = 1, f (0) = 0. But this solution is dmdx ′ R f = 1. From relation (5) we deduce n b(λ) = b(γ nλ) α(a(γ kλ)+δ 1d(γ kλ)) − − − − k=1 Y n α(a(γ kλ)+δ 1d(γ kλ)) n∼ − − − →∞ k=1 Y and n c(λ) = c(γ nλ) (1 α) 1(a(γ kλ)+√δ−1d(γ kλ)) − − − − − k=1 Y n λ α 1(a(γ kλ)+δ 1d(γ kλ)) n∼ − − − − →∞ k=1 Y This immediately implies that the product is convergent and the formulas for b(λ) and c(λ). ♦ 2.1 The Green function of f. Elements of the dynamics of f. The map f has a natural extension to the 2-dimensional projective space P2, given in homogeneous coordinates by f([x,y,z]) = [x(x+δ 1y) δ 1z2,δy(x+δ 1y) δz2,z2]. (9) − − − − − 7 (The point [x,y,z] represents the image of (x,y,z) in C3 by the canonical projection π : C3 P2). Equivalently, this means that f is the map induced on P2 by the → homogeneous polynomial map R :C3 C3 given by → R((x,y,z)) = (x(x+δ 1y) δ 1z2,δy(x+δ 1y) δz2,z2), − − − − − the relation being f(π(x)) = π(R(x)). We can see that f is not definedwhereR is null. We remark that R((1, δ,0)) = (0,0,0) and that C(1, δ,0) is the unique complex line − − on which R is null. We denote by l = [1, δ,0] the associated point in the projective − space and we say that l is a point of indeterminacy. The map f is then a map from P2 l to P2 and is holomorphic on P2 l (in fact the image of the point l by f can \{ } \{ } be defined as a compact Riemann surface called the blow-up of l, cf ([11]). Therefore, the map f is called a rational map of P2. Its degree is 2 in relation with the degree of the homogeneous polynomials appearing in R (cf [15]). We set D = [x,y,z], x+δ 1y = 0 . (10) − { } The line D is sent by f to a unique point, [ δ 1, δ,1], and the orbit of D is − − − f(D l ) = [ δ 1, δ,1] , and fn+1(D l ) = [δ 2n,δ2n,1], (11) − − \{ } {− − } \{ } forn 0. ThesetDiscalledaf-constantcurve(line),sinceitissenttoauniquepoint. ≥ It is a general phenomenon that a f-constant curve contains a point of indeterminacy (cf proposition 1.2 of [15]). Another important property of the map f is that it has no degree lowering curve: a degreeloweringcurveisaf-constantcurvesentbyfn toapointofindeterminacy. When such a phenomenon appears, a common factor, which can be divided out, appears in Rn and the degree of the map fn drops. Here, we remark that the orbit of D does not contain the indeterminacy point l, so D is not a degree lowering curve. Hence, degree(fn) = 2n (this means that Rn is represented by 3 homogeneous polynomials of degree 2n with no common factor). Following the terminology in [15], f is said to be algebraically stable, cf definition 4.4. We set I˜= f n( l ) = [1, δ (n 1),0] the set of preimages of the point of n 0 − − − n 0 ∪ ≥ { } { − } ≥ indeterminacy l . { } The Fatou set of f is defined to be the union of all open balls U P2 I˜ on which ⊂ \ the family fn is normal. The Fatou set is denoted by and its complement, the n 0 Julia set, b{y }=≥ P2 . Of particular interest to us is tFhe fact that the attractive J \F basin of an attractive fixed point is in the Fatou set. A function, useful to study the dynamics of f, is the Green function defined as the limit of the sequence of functions G : C2 R : n → ∪{−∞} 1 G (x) = log(1+ fn(x) ), x C2. (12) n 2n k k ∈ where denotes the usual norm of C2, and where we considered f as a map on C2, k k by (6). We will use the following result (cf [15], proposition 2.11 or [2], theorem 1.6.1): 8 Proposition 2.2 (i) The limit G(x) = lim G (x) n n →∞ exists for all x C2. The function G is plurisubharmonic and satisfies ∈ G f = 2G. (13) ◦ (ii) G is pluriharmonic on C2 . ∩F NB: In general for a rational map on Pk the Green function is defined on Ck+1, cf [15]. In our case since f is polynomial it is easier to adopt the previous definition. Denotebyζ(λ)theLyapounovexponentofthepropagatoroftheO.D.E.(E )given λ by 1 ζ(λ)= lim ln Γ (0,α n) n 2n k λ − k →∞ when this limit exists. In [11], theorem 3.1 and proposition 3.6, we proved Theorem 2.1 (i) The Lyapounov exponent exists and is given by ζ(λ)= G φ(λ) ◦ for all λ in C. (ii) The density of states is given by 1 µ = ∆ζ, 2π where ∆ denotes the distributional Laplacian on C. Rm: Thepoint (ii) correspondsto the Thouless formula in our context. In [12] we gave similar formulas in the general case of finitely ramified self-similar sets for the density of states in terms of the Green function of a certain renormalization map. In [11] we deduced from the analysis of the dynamics of f that supp(µ) is a Cantor subset of R for δ = 1, and for certain values of the parameter δ we proved the local − 6 2 H¨older regularity of the Lyapounov exponent. 2.2 Some details about the dynamics of the map f We give some information about the dynamics of the map f that will be useful in the sequel. We set x = [0,1,0] and x = [1,0,0]. The first important point is that for δ > 1 + − x have one attractive direction and one repulsive (with eigenvalues 0 and δ). • − x is attractive (with eigenvalues 0 and δ 1). + − • (and the reverse for δ < 1). We denote by P2 the hypersurface: C ⊂ = [x,y,z], xy = z2 . (14) C { } 9 Therestriction of toR2 consists oftwo branchesofhyperbolas(cf picture1). Remark C that the restriction of f to is given by C f([x,y,z]) = [x2,y2,z2], [x,y,z] , ∈ C and thus f(C) C. ⊂ We set = [x,y,z] , x < y , = [x,y,z] , x > y . (15) + C { ∈ C | | | |} C− { ∈C | | | |} Clearly, any point in converges to x . Hence, the set is in the Fatou set of f for + C± ± C δ > 1 since it is in the attractive basin of x . The same is true for the set D l since + \{ } fn(D l ) converges to x (cf formula (11). + \{ } We set K (resp. K )= [x,y,z], (x,y,z) R3, xy z2 0 (resp. xy z2 0). (16) + − { ∈ − ≥ } − ≤ Thefollowing propertiesare direct consequences of the forthcoming formula24 (cf [11], formula 3.17) f(K ) K , φ(R ) K . (17) ± ⊂ ± ± ⊂ ± 3 Spectral analysis of the operator on a half-line In this section, we consider the case where ω is stationary, i.e. where ∂I (ω) = . < > ∞ 6 ∅ By the symmetric role played by α and (1 α) it is enough to analyse the case where − I (ω) is a half-line bounded from the left, i.e. when ω is stationary to 1. By scaling < > ∞ it is enough to consider the case ω = (1,...,1,...). Thus, in this section, we only consider the case ω = (1,...,1,...), so that I (ω) = R . < > + Denote by S the set of intersection times∞of the curve φ(γ 1λ) with the line D − defined in (10), i.e. S = λ, φ(γ 1λ) D . − { ∈ } For p in Z we set S = γpS. p We know from proposition 2.1 that the spectrum of H+ and H are given by <0> <−0> suppν+ 0 = suppν = S . (18) <0>\{ } <−0> ∪∞p=0 p Hence, we know that S is non-empty and included in R . The set S is also infinite. Indeed, if λ is in S then φ(γ 1λ ) is in D and φ(λ )− = [ δ 1, δ,1], φ(γλ ) = 0 − 0 0 − 0 − − [δ 2,δ2,1], cf (11). Hence, φ(γ 1λ) must necessarily cross the line D between γλ and − − 0 γ2λ . We denote by 0 > λ > λ > ... > λ > ... the ordered set of points of S. We 0 1 2 k also denote by λ = γpλ the points of S . k,p k p For any k > 0 we denote by f the solutions of the equation (E ) with initial k± λk condition f+(0) = 1, (f+)(0) = 0 and f (0) = 0, (f )(0) = 1. We denote by f , for k k ′ k− k− ′ k±,p p in Z, the scaled copy of f k± p fk±,p = fk±◦Ψ−1 . By scaling fk±,p is solution of (Eγpλk) on R+. We now denote by fk±,p,<n> the restriction of f to I . k±,p <n> 10